Preface

Transportation sustains economic and social activity and is central to oper-

ations research and management science. When operations research emerged
as a structured field during World War II, some of the first problems inves-
tigated arose from the need to optimize military logistics and transportation
activities. After the war ended, the scope of operations research applications
broadened but transportation problems always occupied a central place. It is
now widely recognized that some of the most successful applications of op-
erations research are encountered in transportation, most significantly in the
airline industry where it underlies almost every aspect of strategic, tactical,
and operational planning. This success story may be explained by a number of
factors, the first being the economic importance of transportation. Also, the
complexity and large scale of transportation problems call for powerful analyt-
ical techniques, and the high volumes involved imply that substantial savings
can often be achieved through the use of optimization. Furthermore, trans-
portation problems are highly structured, making them amenable to the use
of efficient solution methods based on network optimization techniques and
mathematical programming.
This book contains eleven chapters describing some of the most recent
methodological operations research developments in transportation. It is
structured around the main transportation modes, and each chapter is writ-
ten by a group of well-recognized researchers. Because of the major impact
of operations research methods in the field of air transportation over the past
forty years, it is befitting to open the book with a chapter on airline operations
management. While many past publications have focused on airline strategic
and tactical planning, Ball, Barnhart, Nemhauser, and Odoni have chosen to
address the organization and control of recovery operations in the event of
disturbances. This line of research is relatively new and of major importance
to the airline industry. The second chapter, by Desaulniers and Hickman, sur-
veys the planning of public transit operations. The problems addressed and
the methods employed in transit planning, for example, those arising in net-
work design, passenger assignment, scheduling, and fleet and crew assignment,
are often similar to those of the airlines. The railway optimization chapter,
by Caprara, Kroon, Monaci, Peeters, and Toth, covers the realm of planning
problems encountered in railway planning, with an emphasis on European pas-
senger railways. Again, several of these issues are similar to those observed in
other modes, but some problems are specific to the railway industry, such as
train platforming, rolling stock circulation, and train unit shunting. The fourth

v
vi Preface

chapter, by Christiansen, Fagerholt, Nygreen, and Ronen, contains an exten-

sive survey of maritime transportation problems, methods, and applications.
Compared with other modes, maritime transportation has received relatively
little attention from operations researchers. Yet this field is rapidly expanding
with the consolidation of major shipping companies and the development of
large container ports.
The next three chapters cover a variety of planning problems arising in vehi-
cle fleet management. The chapter by Powell, Bouzaïene-Ayari, and Simão ad-
dresses truck transportation planning in contexts where information processes
are dynamic. The focus is on the development of models that capture the
flow of information and decisions. The vehicle routing chapter, by Cordeau,
Laporte, Savelsbergh, and Vigo, concerns what is arguably the most central
problem in distribution management. It surveys several families of vehicle
routing problems, including classical models, inventory routing, and stochastic
routing. In the transportation on demand chapter, Cordeau, Laporte, Potvin,
and Savelsbergh consider the planning of pickup and delivery operations
made at the request of users, such as those encountered in courier services,
dial-a-ride operations, dial-a-flight systems, and ambulance fleet deployment.
The eighth chapter, by Crainic and Kim, is devoted to intermodal trans-
portation and ties in some planning issues encountered in railway, maritime,
and trucking operations. This chapter describes methodologies relevant to the
solution of system design and operations planning problems from the perspec-
tive of a carrier, or from that of an intermodal transfer facility operator. It also
addresses problems encountered at the regional or national level. The next
chapter, by Erkut, Tjandra, and Verter, concerns the transportation of haz-
ardous materials and includes a broad description of the issues encountered in
this field, as well as methodological contributions on risk assessment, routing
and scheduling, and facility location.
The last two chapters of the book cover the area of automobile transporta-
tion. Marcotte and Patriksson first survey the broad field of traffic equilibrium.
Their chapter contains a rich account of the main equilibrium concepts, as
well as subproblems and mathematical algorithms encountered in this area.
This chapter provides an informative bibliographical note at the end of each
section. Finally, in the last chapter, Papageorgiou, Ben-Akiva, Bottom, Bovy,
Hoogendoorn, Hounsell, Kotsialos, and McDonald summarize some of the
most important issues and recent developments encountered in ITS and traffic
management. These include traffic flow models, route guidance and informa-
tion systems, as well as urban and highway traffic control.
We are confident that this book will prove useful to researchers, students,
and practitioners in transportation, and we hope it will stimulate further re-
search in this rich and fascinating area. We are grateful to Jan Karel Lenstra
and George L. Nemhauser who invited us to edit this volume. While the
process took longer than we had expected, we found the experience highly
rewarding. Our deep thanks go to all authors for the quality of their contri-
 Preface vii

butions, to the anonymous referees for their time, effort, and valuable sugges-
tions, and to Gerard Wanrooy of Elsevier for his support.

Cynthia Barnhart
Massachusetts Institute of Technology
Gilbert Laporte
HEC Montréal
C. Barnhart and G. Laporte (Eds.), Handbook in OR & MS, Vol. 14
Copyright © 2007 Elsevier B.V. All rights reserved
DOI: 10.1016/S0927-0507(06)14001-3

Chapter 1
Air Transportation: Irregular
Operations and Control
Michael Ball
Decision and Information Technologies Robert H. Smith School of Business,
University of Maryland, College Park, MD 20742, USA
E-mail: [email protected]

Cynthia Barnhart
Civil and Environmental Engineering Department and Engineering Systems Division,
Massachusetts Institute of Technology, Cambridge, MA 02139, USA
E-mail: [email protected]

George Nemhauser
Department of Industrial and Systems Engineering, Georgia Institute of Technology,
Atlanta, GA 30332, USA
E-mail: [email protected]

Amedeo Odoni
Department of Aeronautics and Astronautics, Massachusetts Institute of Technology,
Cambridge, MA 02139, USA
E-mail: [email protected]

1 Introduction

Commercial aviation operations are supported by what is probably the most

complex transportation system and possibly the most complex man-made sys-
tem in the world. Airports make up the fixed “nodes” on which the system is
built. Aircraft represent the very valuable assets that provide the basic trans-
portation service. Passengers demand transportation between a multitude of
origins and destinations, and request specific travel dates and times. Crews
of pilots and flight attendants operate the aircraft and provide service to pas-
sengers. These disparate entities are coordinated through a flight schedule,
comprised of flight legs between airport locations. The flight schedule itself
defines three other layers of schedules, namely the aircraft schedule, the crew
schedule, and passenger itineraries. The aircraft schedule is an assignment of the
legs in the flight schedule, with each aircraft assigned to a connected sequence
of origin to destination flight legs. When an aircraft carries out a flight between
an origin and destination airport, it follows a flight plan that defines a sequence
of points in the airspace through which it proceeds. The crew schedule is an
assignment of the legs in the flight schedule to pilots and flight attendants,
ensuring that all crew movements and schedules satisfy collective bargaining

1
2 M. Ball et al.

agreements and government regulations. Passenger schedules, which repre-

sent the end-customer services, define the familiar itineraries consisting of lists
of origin and destination airports together with scheduled arrival and depar-
ture times. Typically, pilots and flight attendants have distinct schedules. The
itineraries of a specific crew and a specific aircraft may coincide for several
flight legs, but they, like passengers and aircraft, often separate at some point
during a typical day’s operations.
The fact that a single flight leg is a component of several different types
of schedules implies that a perturbation in the timing of one leg can have
significant “downstream” effects leading to delays on several other legs. This
“fragility” is exacerbated by the fact that most of the largest carriers rely heav-
ily on hub-and-spoke network configurations that tightly inter-connect flights
to/from many different “spokes” at the network’s hubs. Thus, any significant
disturbance at a hub, rapidly leads to disruptions of extensive parts of the
carrier’s schedules. Notable categories of events leading to such disruptions
include:
1. Airline resource shortages stemming from aircraft mechanical problems,
disrupted crews due to sickness, earlier upstream disruptions, longer
than scheduled aircraft turn times caused by lack of ground resources
to operate the turn, longer than expected passenger embarking and dis-
embarking times, or delayed connecting crews or connecting passengers.
2. Airport and airspace capacity shortages due to weather or to excessive traf-
fic. Inclement weather is cited as the source of 75% of airline disruptions
in the United States (Dobbyn, 2000).
In 2000, about 30% of the jet-operated flight legs of one major US airline
were delayed, and about 3.5% of these flight legs were canceled. Yu et al.
(2003) report for another major US airline that, on average, a dozen crews are
disrupted every day. The effects of these disruptions are exacerbated when ap-
plied to optimized airline schedules, for which cost minimization and intensive
resource utilization tend to go hand-in-hand. Nonproductive resources, such
as idle aircraft and crew on the ground, are costly. Hence, optimized sched-
ules have minimal nonproductive, or slack, time between flight legs. In these
finely tuned, optimized schedules, delay often propagates with no slack to ab-
sorb it, making it very difficult to contain disruptions and to recover from their
effects. A mechanical delay affecting a single aircraft can result in delays to
passengers and crews assigned to aircraft other than those delayed, due to the
interconnectivity of passengers, crews, and aircraft. This network propagation
phenomenon explains why weather delays in one geographical area, delaying
flights in and out of that area, can result in aircraft, crew, and passenger de-
lays and cancellations in locations far removed from the weather delay. In fact,
such local delays can impact network operations globally.
The significance of the delay propagation effect is illustrated in Figure 1
(reprinted from Beatty et al., 1998). This graphic is based on an analysis of
American Airlines passenger and aircraft schedule information. The x-axis
 Ch. 1. Air Transportation: Irregular Operations and Control 3

Fig. 1. Estimation of delay propagation multiplier (from Beatty et al., 1998).

tracks time of day from early morning to evening. The y-axis tracks increas-
ing values of an initial flight delay. The color of each box in the x–y plane
corresponds to the multiplier that can be applied to an initial delay to estimate
the impact of delay propagation. For example, an initial delay of 1.5 hours at
8:00 is colored dark green indicating a delay multiplier of 2.5. This means that
an original delay of 1.5 hours on a particular flight induces 25 · 15 = 375
hours in total flight delay. Note that the delay multiplier increases with the size
of the original delay and is greatest during the peak morning periods.
The economic impact of disruptions is great. According to Clarke and
Smith (2000), disruption costs of a major US domestic carrier in one year
exceeded $440 million in lost revenue, crew overtime pay, and passenger hos-
pitality costs. Moreover, the Air Transport Association (http://www.airlines.org/
econ/files/zzzeco32.htm) reported that delays cost consumers and airlines about
$6.5 billion in 2000. These costs are expected to increase dramatically, with
air traffic forecast to double in the next 10–15 years. The MIT Global Airline
Industry Program (http://web.mit.edu/airlines/industry.html) and Schaefer et al.
(2005) indicate that, at current demand levels, each 1% increase in air traffic
will bring about a 5% increase in delays.
In this chapter we consider problems related to the management of air traf-
fic and airline operations for the purpose of minimizing the impact and cost of
disruptions. The considerable system complexity outlined above makes these
problems challenging and has motivated a vibrant and innovative body of re-
search. We start in Section 2 by providing background which is essential to
understanding the fundamental issues and motivating the subsequent material.
We first review the “physics” and characteristics of airspace system elements
and airspace operations in order to explain why capacity constraints are so un-
predictable and variable from day to day. Of critical importance are the arrival
4 M. Ball et al.

and departure capacities of airports, which depend on weather, winds, and the
number of active runways and their configuration.
Providers of air traffic control services, such as the Federal Aviation Ad-
ministration (FAA) and EUROCONTROL, have responsibility for overall
airspace management and as such are interested in achieving high levels of
system-wide performance. The two broad classes of “tools” at their disposal
include restricting schedules and air traffic flow management (ATFM). The
former tool, which is treated in Section 3, is strategic in nature. It seeks to
control or influence the airline schedule-planning process by ensuring that the
resultant schedules do not lead to excessive levels of system congestion and
delays. Restricting schedules is particularly challenging in that the competing
economic interests of multiple airlines must be balanced. In fact, recent re-
search in this area has been investigating the potential use of market-based
mechanisms for this purpose, including auctions and peak-period pricing. The
second tool, ATFM, is tactical in nature and is treated in Section 4. ATFM
encompasses a broad range of techniques that seek to maximize the perfor-
mance of the airspace system on any given day of operations, while taking into
account a possibly broad range of disruptive events. Many ATFM actions and
solutions involve an allocation of decision-making responsibilities between the
air traffic control service provider and an airline. This is most notably the case
for solutions employing the Collaborative decision making (CDM) paradigm,
which has the explicit goal of assigning decision making responsibility to the
most appropriate stakeholder in every case (Section 4.4).
A brief Section 5 describes some simulation models that can be useful sup-
port tools in understanding and visualizing the impact of certain types of dis-
ruptive events on airport, airspace, and airline operations and on air traffic
flows, as well as in testing the effectiveness of potential responsive actions.
Sections 6 and 7 address schedule planning and operations problems from
the airline perspective. The introductory paragraphs of this section described
several factors, which lead to the high complexity of these problems. This com-
plexity is compounded by two additional important considerations:
1. The predominant concern with safety and the significant unionization of
crews that, in combination, have led to the imposition of a very large set
of complicated constraints defining feasibility of flight plans and of flight
aircraft and crew schedules.
2. The size of airline networks and operations, including, in the United States
alone, over 5000 public-use airports serving over 8000 (nongeneral avi-
ation) aircraft transporting approximately 600 million passengers on
flights covering more than 5 billion vehicle miles annually (http://www.
bts.gov/publications/pocket_guide_to_transportation/2004/pdf/entire.pdf ).
The aircraft- and crew-scheduling problem, also referred to as the airline
schedule planning problem, involves designing the flight schedule and assign-
ing aircraft, maintenance operations and crews to the schedule. The typical
size of this problem is so large that it is impossible to solve it directly for large
 Ch. 1. Air Transportation: Irregular Operations and Control 5

airlines. Instead, airlines partition it into four subproblems, namely: (i) sched-
ule generation; (ii) fleet assignment; (iii) maintenance routing; and (iv) crew
scheduling. The subproblems are solved sequentially, with the solutions to
the earlier, higher-level subproblems serving as the fixed inputs to subsequent
ones. The schedule generation problem is to determine the flight legs, with
specified departure times, comprising the flight schedule. These legs, which
define the origin–destination markets served and the frequency and timing of
service, have significant effects on the profitability of airlines. Given the flight
schedule, the fleet assignment problem is to find the profit maximizing assign-
ment of aircraft types to flight legs in the schedule. Where possible, the goal is
to match as closely as possible seat capacity with passenger demand for each
flight leg. With the fleeted flight schedule and the size and composition of the
airline’s fleet as input, the maintenance routing problem is to find for each
aircraft, a set of maintenance-feasible rotations, or routes that begin and end
at the same place and satisfy government- and airline-mandated maintenance
requirements. Finally, given all the schedule design and aircraft assignment de-
cisions, the crew scheduling problem is to find the cost minimizing assignment
of cockpit and cabin crews to flights. Crew costs, second only to fuel costs,
represent a significant operating expense. A detailed description of the airline
schedule planning problem is provided in Barnhart et al. (2003a).
From the airline perspective, the focus of this chapter is motivated by the
fact that, despite advances in aircraft and crew schedule planning, optimized
plans are rarely, if ever, executed. Thus, we shall not provide broad coverage
of airline schedule planning but rather focus on the topics that address the
development of schedules and operating practices and policies that provide
operational robustness. In Section 6, we cover the theme of optimizing airline
schedule recovery. The associated tools are designed for use in a near real-time
mode to adjust operations in response to a variety of disruptions. In Section 7,
we address, by contrast, the more strategic topic of developing schedules that
provide operational robustness. This more recent area of study builds upon the
long-standing and well-known body of research on aircraft and crew scheduling
described in the previous paragraph. Finally, in Section 8, we conclude with a
very general assessment of the state of research and implementation in this
subject area.

2 Flow constraints in the infrastructure of commercial aviation

The airspace systems of developed nations and regions consist of a set of

often extremely expensive and scarce nodes, the airports, and of air traffic
management (ATM) systems that provide aircraft and pilots with the means
needed to fly safely and expeditiously from airport to airport. The essential
components of ATM systems are a skilled workforce of human air traffic con-
trollers; organization of the airspace around airports (“terminal airspace”) and
6 M. Ball et al.

between airports (“en-route airspace”) into a complex network of airways, way-

points, and sectors of responsibility; procedures and regulations according to
which the ATM system operates; automation systems, such as computers, dis-
plays, and decision support software; and systems for carrying out the functions
of communications, navigation, and surveillance (CNS) which are critical to
ATM. Any flow constraints encountered during a flight may result from an
obvious cause (e.g., the closing down of a runway) or from a set of complex in-
teractions involving failure or inadequacy of several of the components of the
ATM system.
A controlled flight is one for which an approved flight plan has been filed with
the air traffic management (ATM) system. Airline and general aviation oper-
ators prepare and file flight plans usually based on criteria that consider each
flight in isolation. Air carriers typically employ sophisticated software, includ-
ing advanced route optimization programs, for this purpose. Far from being
just a “shortest path” problem, the selection of an optimal route for a flight
typically involves a combination of criteria, such as minimum time, minimum
fuel consumption, and best ride conditions for the passengers. Midkiff et al.
(2004) provide a thorough description of air carrier flight planning. By accept-
ing a flight plan, the ATM system agrees to take responsibility for the safe
separation of that aircraft from all other controlled aircraft in the airspace and
to provide many other types of assistance toward the goal of completing the
flight safely and expeditiously. Practically all airline flights and a large number
of general aviation flights are controlled. The focus of this entire chapter is on
flow constraints that such flights often face and on how ATM service providers,
airport operators, and airlines attempt to deal with them. Figure 2 illustrates
schematically the fact that such constraints or “bottlenecks” occur when flights
are departing from and arriving at airports and when they seek to access certain
parts of the airspace.
This section will seek to review the “physics” of the constraints, with em-
phasis on explaining the causes of two of their most distinctive characteris-
tics – variability and unpredictability. It is these characteristics that make the
constraints so difficult to deal with in practice, as well as so interesting for many
researchers. Emphasis will be given to the specific topic of capacity-related flow

Fig. 2. Airspace flow constraints.

 Ch. 1. Air Transportation: Irregular Operations and Control 7

constraints at major commercial airports, due to the enormous practical signif-

icance and cost consequences of these constraints.

2.1 The “physics” of airport capacity

Airports consist of several subsystems, such as runways, taxiways, apron

stands, passenger and cargo terminals, and ground access complexes, each with
its own capacity limitations. At major airports, the capacity of the system of
runways is the most restricting element in the great majority of cases. This is
particularly true from a long-run perspective. While it is usually possible – al-
beit occasionally very expensive – to increase the capacity of the other airport
elements through an array of capital investments, new runways, and associated
taxiways require great expanses of land and have environmental and other im-
pacts that necessitate long and complicated approval processes, often taking a
couple of decades or even longer, with uncertain outcomes.
The capacity of runway systems is also the principal cause, by far, of the
most extreme instances of delays that lead to widespread schedule disruptions,
flight cancellations, and missed flight connections. There certainly have been
instances when taxiway system congestion or unavailability of gates and air-
craft parking spaces have become constraints at airports, but these are more
predictable and stable. The associated constraints can typically be taken into
consideration in an ad hoc way during long-range planning or in the daily de-
velopment of ATFM plans (Section 4). By contrast, the capacity of the runway
system can vary greatly from day to day and the changes are difficult to predict
even a few hours in advance. This may lead to an unstable operating envi-
ronment for air carriers: on days when an airport operates at its nominal,
good-weather capacity, flights will typically operate on time, with the excep-
tion of possible delays due to “upstream” events; but, with the same level of
demand at the same airport, schedule reliability may easily fall apart on days
when weather conditions are less than ideal.
The capacity of a runway, or of a set of simultaneously active runways at an
airport, is defined as the expected number of movements (landings and take-
offs) that can be performed per unit of time in the presence of continuous
demand and without violating air traffic control (ATC) separation require-
ments. This is often referred to as the maximum throughput capacity, C. Note
that this definition recognizes that the actual number, N, of movements that
can be performed per unit of time is a random variable. The capacity C is sim-
ply defined as being equal to E[N], the expected value of N. The unit of time
used most often is one hour.
To understand better the multiple causes of capacity variability, especially
its strong dependence on weather and wind conditions, it is necessary to look
at the “physics” of the capacity of runway systems. It is convenient to consider
first the case of a single runway and then (Section 2.1.4) the case of a system of
several runways.
8 M. Ball et al.

2.1.1 Factors affecting the capacity of a single runway

The capacity of a single runway depends on many factors, the most impor-
tant of which are:
1. The mix of aircraft classes using the airport.
2. The separation requirements imposed by the ATM system.
3. The type (high speed or conventional) and location of exits from the run-
way.
4. The mix of movements on each runway (arrivals only, departures only, or
mixed) and the sequencing of the movements.
5. Weather conditions, namely visibility, cloud ceiling, and precipitation.
6. The technological state and overall performance of the ATM system.
The impacts of 1–4 are summarized below, while 5 and 6 are discussed in the
more general context of multi-runway systems in the next subsection.

Mix of aircraft. The FAA and other Civil Aviation Authorities around the
world classify aircraft into a small number of classes for terminal area ATC
purposes. For example, the FAA defines four classes, based on maximum take-
off weight (MTOW): “Heavy” (H), “Large” (L), the Boeing 757 (a class by
itself), and “Small” (S). Most other Civil Aviation Authorities have adopted
the same or very similar classifications. Roughly speaking, the H class includes
all wide-body jets, and the L class practically all narrow-body commercial jets –
including many of the larger, new generation, regional jets – as well as some of
the larger commercial turbo-props. Most general aviation airplanes, including
most types of private jets, as well as the smaller commercial turboprops and
regional jets with about 35 seats or fewer comprise the S class. The aircraft
mix indicates the composition of the aircraft fleet that is using any particular
runway (e.g., 20% S, 60% L, 5% B757, and 15% H).

Separation requirements and high-speed exits. The single most important fac-
tor in determining runway capacity is the separation requirements, which im-
pose safety-related separations between aircraft that limit the service rate of
the runway, i.e., its maximum throughput capacity. For every possible pair of
aircraft using the same runway consecutively, the FAA and other Civil Avia-
tion Organizations specify a set of separation requirements in units of distance
or of time. These requirements depend on the classes to which the two air-
craft belong and on the types of operation involved: arrival followed by arrival,
“A–A”, arrival followed by departure, “A–D”, etc. Table 1 shows the separation
requirements that currently apply at most of the busiest airports in the United
States for the case in which a runway is used only for arrivals under instrument
flight rules (IFR). Pairs of consecutive landing aircraft must maintain a sepa-
ration equal to or greater than the distances indicated in Table 1 throughout
their final approach to the runway, with the exception of the cases marked with
an asterisk, where the required separation must exist at the instant when the
leading aircraft reaches the runway. The 4, 5, and 6 nautical mile separations
 Ch. 1. Air Transportation: Irregular Operations and Control 9

Table 1.
FAA IFR separation requirements in nautical miles (nmi) for “an
arrival followed by an arrival”. Asterisks indicate separations that
apply when the leading aircraft is at the threshold of the runway.

Leading aircraft Trailing aircraft

H L + B757 S

H 4 5 6*
B757 4 4 5*
L 25 25 4*
S 25 25 2.5

shown in Table 1 are intended to protect the lighter trailing aircraft in the pair
from the hazards posed by the wake vortices generated by the heavier leading
aircraft. These are therefore often referred to as “wake vortex separations”.
In addition to the “airborne separation” requirements of Table 1, a further re-
striction is applied: the trailing aircraft of any pair cannot touch down on the
runway before the leading aircraft is clear of the runway. In other words, the
runway can be occupied by only one arriving aircraft at any time.
The more restrictive of the two requirements – “airborne separation” and
“single occupancy” – is the one that applies for each pair of aircraft. When
arrivals take place in instrument meteorological conditions (IMC), “airborne
separation” is almost always the most restrictive. However, “single occupancy”
may become the constraint when visual airborne separations on final approach
are allowed (instead of the distance requirements of Table 1), as is often done
in the United States under visual meteorological conditions (VMC). In this
case, high-speed runway exits and well-placed runway exits (the third of the
factors identified above), which reduce runway occupancy times for arriving
aircraft, can be helpful in increasing runway capacity. High-speed exits can
also be useful when the runway is used for both arrivals and departures: if
landing aircraft can exit a runway quickly, air traffic controllers may be able to
“release” a following takeoff sooner.

Mix of movements. Separation requirements, analogous to those in Table 1,

are also specified for the other three combinations of consecutive operations,
A–D, D–A, and D–D. Because the separation requirements for each combi-
nation are different – see, e.g., Chapter 10 of de Neufville and Odoni (2003)
for details – the capacity of a runway during any given time period depends on
the mix of arrivals and departures during that period, as well as on how exactly
arrivals and departures are sequenced on the runway. This also suggests that
there is an important tradeoff between the maximum arrival and departure
rates that an airport can achieve.
10 M. Ball et al.

2.1.2 Capacity envelope and its computation

The runway capacity envelope (Figure 3) is convenient for displaying the ar-
rival and departure capacities and associated tradeoffs. The capacity envelope
of a single runway is typically approximated by a piecewise linear boundary that
connects four points (Gilbo, 1993). Points 1 and 4 indicate the capacity of the
runway, when it is used only for arrivals and only for departures, respectively.
Point 2 is known as the “free-departures” point because it has the same ca-
pacity for arrivals as Point 1 and a departures capacity equal to the number of
departures that can be inserted into the arrivals stream without increasing the
separations between successive arrivals – and, thus, without reducing the num-
ber of arrivals from what can be achieved in the all-arrivals case. These “free”
departures are obtained by exploiting large inter-arrival gaps such as the ones
that arise between a “H-followed-by-S” pair of landing aircraft. Point 3 can be
attained, in principle, by alternating arrivals and departures, i.e., by perform-
ing an equal number of departures and arrivals through an A–D–A–D–A–· · ·
sequence. This sequencing strategy can be implemented by “stretching”, when
necessary, inter-arrival (inter-departure) gaps by an amount of time just suf-
ficient to insert a departure (arrival) between two successive arrivals (depar-
tures). Because it is difficult for air traffic controllers to sustain this type of
operation for extended periods of time, Point 3 can be viewed as somewhat
theoretical. However, it provides a useful upper limit on the total achievable
capacity (landings plus takeoffs) when arrivals and departures share a runway
in roughly equal numbers.
Several mathematical models have been developed over the years for
computing the capacity of a single runway under different sets of condi-
tions, beginning with Blumstein’s (1959) classical model of a single runway
used for arrivals only. The models have become increasingly sophisticated

Fig. 3. A capacity envelope for a single runway.

 Ch. 1. Air Transportation: Irregular Operations and Control 11

over the years and include treatment of some of the input parameters as
random variables. Barnhart et al. (2003a) provide a literature review. The
most recent of these models (Long et al., 1999; Stamatopoulos et al., 2004;
EUROCONTROL, 2001) incorporate most of the best features of earlier
models and generate capacity envelopes, such as the one in Figure 3.

2.1.3 The sequencing problem

As a result of the airborne separation requirements shown in Table 1, cer-
tain aircraft pairs require longer separation distances than others and thus
the total time needed for the landing of any set of aircraft on a runway de-
pends on the sequencing of the aircraft. For example, the “H followed by S”
sequence will consume much more time than “S followed by H”. Given a
number n of aircraft, all waiting to land on a runway, the problem of “de-
termining the sequence of landings such as to minimize the time when the last
aircraft lands” is a Hamiltonian path problem with n points (Psaraftis, 1980;
Venkatakrishnan et al., 1993). This is a problem entirely analogous to several
well-known job-sequencing problems in manufacturing.
However, the Hamiltonian path approach addresses only a static version
of a problem. In truth, the problem of sequencing aircraft on a runway is
dynamic: over time, the pool of aircraft available to land changes, as some
aircraft reach the runway while new aircraft join the arrivals queue. Moreover,
minimizing the “latest landing time” (or maximizing “throughput”) should not
necessarily be the objective of optimal sequencing. Many alternative objec-
tive functions, such as minimizing the average waiting time per passenger, are
just as reasonable. A further complication is that the very idea of “sequenc-
ing” runs counter to the traditional adherence of ATM systems to a first-come,
first-served (FCFS) discipline, which is perceived by most as “fair” (see also
Section 4).
These observations have motivated a great deal of research on the runway
sequencing problem with the objective of increasing operating efficiency while
ensuring that all airport users are treated equitably. Dear and Sherif (1991) de-
veloped the concept of constrained position shifting (CPS), i.e., of a limit in the
number of positions by which an aircraft can deviate from its FCFS position in
a queue. For instance, an aircraft in the 16th position in an FCFS queue, would
have to land in one of the positions 14–18, if the specified maximum position
shift (MPS) is 2. Through many numerical examples and for several reason-
able objective functions, Dear showed that, by setting MPS to a small number,
such as 2 or 3, one can obtain most (e.g., 60–80%) of the potential benefits
offered by unconstrained optimal sequences and, at the same time, ensure rea-
sonable fairness in accessing runways. Several researchers (e.g., Psaraftis, 1980;
Venkatakrishnan et al., 1993; Beasley et al., 2001) have investigated a number
of increasingly complex and realistic versions of the sequencing problem. Two
advanced terminal airspace automation systems, CTAS and COMPAS, that
have been implemented in the US and in Germany, respectively, incorporate
sequencing algorithms based on CPS (Erzberger, 1995).
12 M. Ball et al.

Gilbo (1993), Gilbo and Howard (2000), and Hall (1999) have gone beyond
the sequencing of arrivals only, by considering how available capacity can best
be allocated in a dynamic way between landings and take-offs to account for
the distinct peaking patterns in the arrival and departure streams at airports
over the course of a day. They propose the application of optimization algo-
rithms that use capacity envelopes (Figure 3) within the context of ATFM to
achieve an optimal trade-off between arrival and departure rates and, by im-
plication, between delays to arrivals and to departures.

2.1.4 Factors affecting the capacity of multi-runway systems

Most (but certainly not all) major airports typically operate with two or more
simultaneously active runways. The term runway configuration refers to any
set of one or more runways, which can be active simultaneously at an airport.
Multi-runway airports may employ more than ten different runway configura-
tions. Which one they will operate on at any given time will depend on weather
and wind conditions, on demand levels at the time and, possibly, on noise con-
siderations, as will be explained below.
The six factors listed in Section 2.1.1 clearly continue to affect the capacity
of each individual active runway in multi-runway cases. In addition, at least
four other factors may now play a major role:
7. The interactions between operations on different runways, as de-
termined by the geometric layout of the runway system and other
considerations.
8. The allocation of aircraft classes and types of operations (arrivals,
departures, mixed) among the active runways.
9. The direction and strength of winds.
10. Noise-related and other environmental considerations and con-
straints.

Interactions between operations on different runways. The influence of the geo-

metric layout on the interactions among runways can most simply be illustrated
by looking at situations involving two parallel runways. Depending on the dis-
tance between their centerlines, operations on the two runways may have to
be coordinated all the time, or may be dependent in some cases and indepen-
dent in others, or may be completely independent. For example, in the United
States, two parallel runways separated by less than 2500 ft (762 m) must be
operated with essentially single runway separations in instrument meteorolog-
ical conditions (IMC). This means, for instance, that, if two arriving aircraft
are landing, one on the left runway and the other on the right runway, they
are subject to the same set of airborne separation requirements as shown in
Table 1 for a single runway. At the opposite extreme, if the centerlines are
separated by more than 4300 ft (1310 m) the two runways may be operated
independently and can accept simultaneous parallel approaches. (With special
instrumentation, the FAA will consider authorizing independent parallel ap-
proaches with centerline separations as small as 3000 ft (915 m).) Finally, for
 Ch. 1. Air Transportation: Irregular Operations and Control 13

intermediate cases, contemporaneous arrivals on the two parallel runways are

treated as “dependent”, i.e., must be coordinated, but an arrival on one of the
runways and a contemporaneous departure from the other can be handled in-
dependently. It follows that the combined capacity of the two runways will be
highest in the case of independent operations, intermediate in the “partially
dependent” operations case, and lowest when every pair of operations on the
two runways must be coordinated. In a similar way, the combined capacity of
pairs of intersecting runways or of runways that do not intersect physically but
intersect at the projections of their centerlines depends on many geometry-
related parameters such as the location of the intersection or of the projected
intersection; the direction of operations on the two runways; the types of op-
erations and the mix of aircraft; and obviously, the separation requirements
for the particular geometric configuration at hand. Systems of three or more
active, nonparallel runways typically involve even more complex interactions.

Allocation of aircraft and operations. With more than one active runway, there
is some opportunity to “optimize” operations by judiciously assigning opera-
tions and/or aircraft classes to different runways. For example, in the case of
intermediately spaced parallel runways (centerline separations of 2500–4300 ft
in the United States) it may be advisable to use one runway primarily for ar-
rivals and the other primarily for departures. Since, in this case, arrivals on
one runway can operate independently of departures on the other, this alloca-
tion strategy minimizes interactions between runways and reduces controller
workload. Similarly, when two or more runways are used for arrivals, air traffic
managers often try to assign relatively homogeneous mixes of aircraft to each
of the runways, e.g., keep the “Small” aircraft on a separate runway from the
“Heavy” and “Large”, to the extent possible. In this way, air traffic controllers
can avoid the extensive use of the 5 and 6 nautical mile wake-vortex separa-
tions that are required when a Small aircraft is landing behind a Large or a
Heavy (Table 1).

Weather-related factors. It is easy to infer from what has been said so far that
weather-related factors (numbers 5 and 9 in our list) are critical in determining
the variability of the capacity of any system of runways. First, for individual
runways, the actual separations between consecutive operations are strongly
influenced by visibility, cloud ceiling, and precipitation. This is especially true
in the United States where, in good weather, pilots are usually requested to
maintain visual separations during the final approach phase from the aircraft
landing ahead of them. This practice results in somewhat closer spacing of
landing aircraft than suggested by the IFR separations of Table 1. It also means
smaller deviations from the required minima, as pilots can adjust spacing as
they approach the runway. This second effect is also present at airports where
the practice of “visual separations on final” in VMC has not yet been adopted.
The overall effect is that, with the same aircraft mix and the same nominal
14 M. Ball et al.

(IFR) separation requirements, the capacity of individual runways in VMC is

typically greater than in IMC, occasionally by a significant margin.
Second, when it comes to multi-runway configurations, good visibility con-
ditions have a similar effect. For example, in the case of two parallel runways,
the FAA generally authorizes simultaneous parallel approaches in IMC, when
the separation between runway centerlines is 4300 ft or more, as noted earlier.
But in VMC, simultaneous parallel approaches can be performed to parallel
runways separated by only 1200 ft (366 m) when Heavy aircraft are involved
and by only 700 ft (214 m) when they are not. As a consequence, San Francisco
International (SFO), one of the most delay-prone airports in the world, has
an arrival capacity of about 54 per hour in VMC, when simultaneous parallel
approaches are performed to a pair of closely-spaced parallel runways, and of
only 34 per hour in IMC, when the same two runways are operated essentially
as a single runway, as described earlier. The impact of VMC is similar when
it comes to operations on a pair of intersecting runways, as illustrated by New
York’s LaGuardia Airport (LGA). In VMC, LGA has a nominal capacity of
81 movements per hour and, with the same two runways, a capacity of 63 (or
22% less) in IMC.
Wind direction and wind strength are just as critical in determining which
runways will be active at a multi-runway airport at any given time. First, land-
ings and takeoffs are conducted into the wind – the maximum allowable tail-
wind is generally of the order of 5 knots. Thus, the direction of the wind
determines the direction in which the active runways are used. Equally impor-
tant, there are limits on the strength of the crosswinds that aircraft can tolerate
on landing and on takeoff. For any runway, the crosswind is the component of
the wind vector whose direction is perpendicular to the direction of the run-
way. The crosswind tolerance limits vary according to type of aircraft and to the
state of the runway’s surface (dry or wet, slippery due to icy spots, etc.). Thus,
airports are often forced by crosswinds and tailwinds to utilize configurations
that offer reduced capacity. For example, with strong westerly winds, Boston
Logan (BOS) is forced to operate with two main runways even in VMC, in-
stead of the customary three. This reduces capacity by about 30 movements
per hour from the VMC norm!

State and performance of the ATM system. An obvious underlying premise to

all of the above is that a high-quality ATM system with well-trained and ex-
perienced personnel is a prerequisite for achieving high runway capacities. To
use a simple example, tight separations between successive aircraft on final
approach (i.e., separations which are as close as possible to the minimum re-
quired in each case) cannot be achieved unless (a) accurate and well-displayed
information is available to air traffic controllers regarding the positions of the
leading and trailing aircraft, and (b) the controllers themselves are skilled in
the task of spacing aircraft accurately during final approach. Major differences
exist in this respect between ATM systems in different countries.
 Ch. 1. Air Transportation: Irregular Operations and Control 15

Environmental considerations. Finally, runway usage and, by extension, air-

port capacity at some major airports may be strongly affected by noise-
mitigation and other measures motivated by environmental considerations. In
the daily course of airport operations, noise is one of the principal criteria used
by air traffic controllers to decide which one among several usable alternative
runway configurations to activate. (A choice among two or more alternative
configurations may exist whenever weather and wind conditions are sufficiently
favorable.) Environmental considerations act, in general, as a constraint on
airport capacity since they tend to reduce the frequency with which certain
high-capacity configurations may be used.

The capacity envelope for multi-runway systems and its computation. The com-
plexity of computing the capacity envelopes of multi-runway airports depends
on the complexity of the geometric layout of the runway system and the extent
to which operations on different runways are interdependent. The simplest
cases, involving two parallel or intersecting runways, can still be addressed
through analytical models, because they are reasonably straightforward exten-
sions of single-runway models (Stamatopoulos et al., 2004). Analytical models
also provide good approximate estimates of true capacity in cases involving
three or more active runways, as long as the runway configurations can be
“decomposed” into semi-independent parts, each consisting of one or two run-
ways. This is possible at the majority of existing major airports and at practically
every secondary airport.
When such decomposition is not possible or when a highly detailed repre-
sentation of runway and taxiway operations is necessary, simulation models can
be used. General-purpose simulation models of airside operations first became
viable in the early 1980s and have been vested with increasingly sophisticated
features since then. Two models currently dominate this field internationally:
SIMMOD and the Total Airport and Airspace Modeler (TAAM). A report by
Odoni et al. (1997) contains detailed reviews of these and several other airport
and airspace simulation models and assesses the strengths and weaknesses of
each. At their current state of development and with adequate time and per-
sonnel resources, they can be powerful tools not only in estimating the capacity
of runway systems, but also in studying detailed airside design issues, such as
figuring out the best way to remove an airside bottleneck or estimating the
amount by which the capacity of an airport is reduced due to the crossing of
active runways by taxiing aircraft. However, these simulation models still in-
volve considerable expense, as well as require significant time and effort and,
most importantly, expert users.

2.1.5 The variability and unpredictability of the capacity of runway systems

The variability of the capacity of runway systems at major airports can
now be easily explained with reference to the previous discussion. There are
three main causes of drastic reductions in airport capacity. Two are related to
weather: severe events, like thunderstorms or snowstorms; and more routine
16 M. Ball et al.

weather events, like fog or very strong winds. The third major cause is techni-
cal or infrastructure problems, such as air traffic control equipment outage or
the temporary loss of one or more runways, due to an incident or accident or
to maintenance work. For this last case, it should be noted that major airports
schedule runway maintenance carefully, so as to minimize impact on airport
traffic.
Thunderstorms and snowstorms are events that pose hazards to aviation.
Thus, they impede severely the flow of air traffic into and out of airports and
through major portions of affected airspace. They carry the potential for even
shutting down airports completely for several hours and, occasionally, for a few
days at a time in the case of snowstorms. The more routine events can be much
more frequent, such as heavy fog at the San Francisco, Milan, and Amsterdam
airports or strong winds in Boston. These typically cause a severe reduction
of capacity, from the best levels achievable in VMC to levels associated either
with IMC or with nonavailability of some runways due to winds. The FAA in
a 2001 study compared the maximum throughput capacities of the 31 busiest
commercial airports in the United States under optimum weather conditions,
with the capacity of the most frequently used configuration in IMC (FAA,
2001). The study found that, on average, the capacity was reduced by 22%
in the latter case, with 8 of the 31 airports experiencing a capacity reduction
of 30% or more! Note that other, less frequently used IMC configurations at
these airports often have even lower capacities.
The overall effect of weather on an airport’s capacity can be summarized
conveniently through the capacity coverage chart (CCC), which is essentially
a plot of the probability distribution of available capacity over an extended
period of time such as a year. An example for Boston’s Logan International
Airport (BOS) is shown in Figure 4. (The CCC is somewhat simplified to indi-
cate only five principal levels of capacity.) It indicates that the capacity varies

Fig. 4. The capacity coverage chart for Boston Logan International Airport.
 Ch. 1. Air Transportation: Irregular Operations and Control 17

from a high capacity of 115 movements or more per hour, available for about
77% of the time – the leftmost two levels of capacity – and associated with the
most favorable VMC, to a low of about 55–60 movements per hour for about
6% of the time and associated with low IMC. (The airport also has capacity of
zero, meaning it is closed down due to weather conditions, about 1.5% of the
time.) One of the two intermediate levels of capacity (third from left) of about
94 movements per hour is associated with the presence of strong westerly winds
in VMC. As mentioned in the previous section, these force the airport to oper-
ate with only two active main runways. To prepare the CCC, it is necessary to
examine historical hourly weather records (visibility, cloud ceiling, precipita-
tion, winds) for a long period of time (e.g., five years) and identify the capacity
available at each of these hours.
The CCC is drawn under the simplifying assumptions that (a) the mix of
arrivals and departures is 50% and 50% and (b) the airport is operated at
all times with the highest-capacity configuration that can be used under the
prevailing weather conditions. While neither of these assumptions is exactly
true in practice, the CCC nonetheless provides a good indication of the overall
availability of capacity during a year, as well as of the variability of this capacity.
Obviously a CCC that stays level for the overwhelming majority of time – as
one might expect to find at airports that enjoy consistently good weather –
implies a more predictable operating environment than the “uneven” CCCs of
BOS, SFO, LGA, and other airports where weather is highly variable.
Although the associated technology is improving, meteorological forecasts
still have not attained the level of accuracy and detail needed to eliminate
uncertainty from predictions of airport and airspace weather, even for a time-
horizon as short as one or two hours. When it comes to impact on the opera-
tions of any specific airport, the challenge is twofold: predicting the severity of
an anticipated weather event at a quite microscopic level; and equally impor-
tant, determining narrow windows for the forecast starting and ending times of
the event. For example, a few hundred feet of difference in the cloud ceiling or
the presence or absence of “corridors” for the safe conduct of approaches and
departures in convective weather may make a great difference in the amount
of capacity available at an airport. Similarly, over- or under-predicting by just
one hour the ending time of a thunderstorm may have major implications on
a Ground delay program (see Section 4) and, as a result, on the costs and dis-
ruption caused by the associated delays and flight cancellations.

2.2 The capacity of airspace sectors

Within the airspace itself, safety concerns and the need to separate aircraft
leads to yet another set of constraints. The most prevalent of these is associ-
ated with a sector. A sector is a volume of airspace for which a single air traffic
control team of one or two individuals has responsibility. The principal con-
straint on the number of aircraft that can safely occupy a sector simultaneously
is controller workload. Both in North American and in European airspace, it is
18 M. Ball et al.

generally accepted that this number should not exceed the 8–15 range, depend-
ing on a number of factors. This limitation, in turn, translates to typical upper
limits of the order of 15–20 on the number of aircraft that can be scheduled
to traverse a sector during a 15-minute time interval in US en-route airspace.
This capacity may be reduced significantly in the presence of severe weather.
Because of its heavy dependence on controller workload, it is difficult to
compute the capacity of a sector, in terms of either the number of simulta-
neously present aircraft or the number of aircraft traversing the sector per
unit of time (Wyndemere, 1996; Sridhar et al., 1998). Numerous factors af-
fect the complexity of the controller’s task. Hinston et al. (2001) classify these
into three major categories:
(a) Airspace factors: sector dimensions (physical size and shape, area that
the controller must effectively oversee); spatial distribution of airways and of
navigational aids within the sector; number and location of standard ingress
and egress points for the sector; configuration of traffic flows (number and ori-
entation relative to the shape of the sector, complexity of aircraft trajectories,
crossing points and/or merging points of the flows); and complexity of required
coordination with controllers of neighboring sectors (e.g., for “hand-offs” of
aircraft from one sector to the next).
(b) Traffic factors: number and spatial density of aircraft; range of aircraft
performance (homogeneous traffic vs. many different types of aircraft with di-
verse performance characteristics); complexity of resolving aircraft conflicts
(which depends on many variables); sector transit time.
(c) Operational constraints: restrictions on available airspace, e.g., due to
the presence of convective weather or of special use airspace; limitations of
communications systems; and procedural flow restrictions at certain waypoints
(see Section 4.1.2) or noise abatement procedures in place.
Several attempts have been made in recent years to develop quantitative re-
lationships between some of these factors and controller workload – see, e.g.,
Manning et al. (2002). To deal with this complexity and handle a large number
of aircraft, controllers attempt to introduce a “structure” to the traffic patterns
they handle. Examples include (Hinston et al., 2001): spatial standardization
of the flows of aircraft within sectors along specific paths; consideration of air-
craft in groups, with members of each group linked by common attributes; and
concentration around a few “critical points” of the location of potential aircraft
encounters or of other occurrences requiring controller intervention.
In order to effect such structuring, flow may usually be directed through
a few waypoints or fixes, which have an associated maximum flow rate. These
maximum flow rates can be derived from minimum separation standards or
alternatively the maximum rates can be specified by the ATM system itself in
order to limit the amount of flow passing downstream. In this latter case, the
maximum flow rate can be viewed as a control variable.
Finally, it is noted that airspace constraints typically have a less severe ef-
fect on airline operations than airport constraints. This is especially true in
the more flexible ATM environment of the United States. The reason, quite
 Ch. 1. Air Transportation: Irregular Operations and Control 19

simply, is that a capacity-constrained en-route sector can often be bypassed at

limited cost by selecting an alternative route, whereas a flight has no choice but
to eventually end up at its destination airport.

3 Restricting schedules

3.1 Options and current practice on airport scheduling

With the background of Section 2, we can now proceed to review how the
two principal types of airspace system stakeholders attempt to deal with unpre-
dictable and variable capacity constraints in daily operations. This and the next
section will discuss strategies associated primarily with ATM service providers
(Civil Aviation Authorities and other national and international organizations)
while Sections 6 and 7 will present the strategic and tactical options available
to the airlines.
To ATM service providers (e.g., the FAA) two approaches are essentially
available. One, restricting schedules (RS), is of a static and “pro-active” nature
as it places, in advance, limits on the maximum number of aircraft movements
that can be scheduled during a unit of time at an airport or other airspace
element. The second, ATFM, is dynamic and reactive: its goal is to prevent air-
port and airspace overloading by adjusting in “real time” the flows of aircraft
on a national or a regional basis in response to actual conditions. In essence,
the focus of RS is on controlling the number of scheduled operations through
airspace elements and of ATFM on controlling the number of actual opera-
tions through these elements, given a schedule. This section reviews briefly the
RS approach, while the next deals more extensively with ATFM.
A far more frequently used term for RS, especially among aviation policy-
makers, is “demand management”. It refers to any set of administrative and/or
economic policies and regulations aimed at constraining the demand for access
to airspace elements during certain times when congestion would otherwise be
experienced. This term is avoided here, because it may cause confusion with a
major aspect of ATFM, which is also concerned with “managing demand” in a
dynamic way in order to match it with available capacity.
RS is not used currently in the United States, with the exception of four air-
ports (New York LaGuardia and Kennedy, Chicago O’Hare and Washington
Reagan) where limits on the number of movements that can be scheduled per
hour – the so-called “high-density rules” (HDR) – have existed since 1968. The
HDR will be phased out by 2007 according to the so-called AIR-21 legislation
of 2000 and, in fact, in some cases, e.g., Chicago, the restrictions have already
been relaxed. However, RS is widely practiced outside the US: about 140 of the
world’s busiest airports are “fully coordinated”, meaning that they place strict
limits on the number of movements that the airlines can schedule there. These
airports serve the great majority of air travelers outside the US every year.
20 M. Ball et al.

The concept underlying RS is the declared capacity, i.e., a declared limit

on the maximum number of air traffic movements that can be scheduled at
an airport per unit of time. At a few airports, separate limits are specified
for the number of landings, the number of takeoffs and the total number of
movements. The typical unit of time is one hour, but some airports use finer
subdivisions of time. At a few airports, the declared capacity may also vary by
time of day, e.g., more departures than arrivals may be allowed during certain
hours and vice versa for other hours.
The declared capacity is determined by the capacity of the most restricting
element of the airport – the so-called “bottleneck element”. In most cases, this
is the runway system of the airport. However, the bottleneck element can also
be the passenger terminal, or the apron area, or some other part of the airport.
Even in such cases, the computed limit (e.g, the number of passengers that can
be scheduled per hour at the passenger terminal) is converted to a declared
limit on the number of air traffic movements.
The RS approach – and the concept of declared capacity – can be extended
to air traffic control sectors. The declared capacity in this environment is pri-
marily determined by workload considerations, as noted in Section 2.2, taking
into consideration traffic patterns, traffic mix, route configuration, etc. EURO-
CONTROL, the agency that coordinates air traffic management systems over
Europe, in effect uses such capacity figures for en-route sectors in its six-month
advance planning of traffic loads in European airspace.

3.2 Critical issues regarding restricting schedules

If traffic volumes at airports and airspace sectors are restricted to levels that
can be handled comfortably all the time, the RS approach can clearly be effec-
tive in reducing major delays and important schedule disruptions. However,
the approach is also characterized by several fundamental problems, three of
which are described here. First, implicit in the approach is the need to make
a tradeoff between delays, on the one hand, and resource utilization, on the
other, on the basis of only very aggregate information. Consider, for example,
the case of Boston’s Logan International Airport (BOS). As suggested by Fig-
ure 4, the maximum achievable arrival rate at BOS under favorable weather
conditions (visibility, cloud ceiling, winds, precipitation) is around 60 per hour.
Such conditions prevail about 77% of the time. During the other 22%, the
maximum arrival rate is lower – and can be as low as 30 per hour for about
6% of the time. Were BOS to declare an arrival capacity of 60 (and assuming
that the airlines and general aviation operators actually scheduled that many
movements), delays at BOS would reach high to unacceptable levels during
about 22% of the peak traffic hours over any extended span of time, such as a
year. On the other hand, declaring an arrival capacity of 30 would practically
ensure the absence of serious delays, but would result in gross underutilization
of the airport’s resources most of the time. In general, one should note that any
choice of the value of the declared capacity must be made on the basis of very
 Ch. 1. Air Transportation: Irregular Operations and Control 21

aggregate statistical information: when an airport “declares a capacity” for the

next six months (as is currently done under the process organized by IATA, see
below), all it has to go on is historical statistics about weather conditions at the
airport and the resulting available capacity. By contrast, as described in Sec-
tion 4, the ATFM approach uses real-time capacity forecasts with a maximum
time-horizon of about 12 hours. In this light, it is not surprising that no single,
internationally accepted methodology exists today for determining and setting
the declared capacity of airports and airspace sectors. Practices vary greatly
from country to country and, in some cases, from airport to airport within
the same country. Some major international airports in Europe and in Asia
clearly opt for allowing for a considerable margin of “comfort” by declaring
very “conservative” capacities, i.e., capacities near the low end of the available
range. This results in wasting significant amounts of extremely valuable capac-
ity.
The second problem with RS is the way it is currently practiced. The de-
clared capacities of the major airports, as updated at six-month intervals, are
communicated to international aviation authorities and to the airlines and
serve as the basis for scheduling airline operations at busy airports during the
international Schedule Coordination Conferences (SCC) that are organized by
the International Air Transport Association (IATA) and take place in Novem-
ber and June every year. During the SCC, a “schedule coordinator” allocates
the available capacity (“slots”) among the airlines that have requested access
to each fully coordinated airport. If the number of available slots is insuffi-
cient to satisfy demand, then some requests are simply denied. The criteria
used to allocate slots are described in detail in IATA (2000) and discussed in
de Neufville and Odoni (2003). For our purposes, it is sufficient to note that
the dominant and overriding criterion is historical precedent: an airline which
was assigned a slot in the same previous season (“summer” or “winter”) and
utilized that slot for at least 80% of the time during that previous season is
automatically entitled to the continued use of that “historical slot”. No eco-
nomic criteria are used for slot allocation and, in fact, buying and selling of
slots is prohibited – at least, under the official rules. The net result is that
some of the older, traditional airlines maintain in this way a “lock” on most
of the prime slots at the world’s most economically desirable airports. Air-
lines that wish to compete in these markets and may be willing to pay high
fees for the right to operate at the airports in question are effectively “frozen
out”.
Third, and perhaps most important, the RS approach, as currently practiced,
distorts the functioning of the marketplace and suppresses potential demand
by placing an arbitrary cap on the number of operations at some of the world’s
busiest airports. This, in turn, creates an illusionary equilibrium, i.e., an artifi-
cial balance between demand and capacity. Thus, current RS practices do not
provide decision-makers with true information about the economic value that
airspace users and the traveling public may attach to additional capacity at the
schedule-coordinated airports or at other elements of the airspace.
22 M. Ball et al.

A great deal of research has been performed over the years on these and
related issues, focusing primarily on the second of the problems described
above and, to a lesser extent, on the third. Several classical and more recent pa-
pers have dealt with the application of “market-based” mechanisms, typically
in combination with administrative measures, to improve the current capac-
ity allocation process of IATA – see Vickrey (1969), Carlin and Park (1970),
Morrison (1987), Daniel (1995), DotEcon Ltd. (2001), and Ball et al. (2005),
for a sample. Such papers examine the use of congestion pricing and of slot
auctioning at major airports. Interestingly, some ongoing research on real-time
capacity allocation and slot-exchange mechanisms in connection with ATFM –
see, e.g., Vossen and Ball (2006b) and Ball et al. (2005) – has also focused on
market-based approaches (see also Section 4.4).
Other recent work has investigated less generic issues related to compu-
tational and implementation issues associated with such approaches. For ex-
ample, advances in the application of queuing theory have facilitated greatly
the estimation of external delay costs (Andreatta and Odoni, 2003). Fan and
Odoni (2002) and Hansen (2002) report applications of queuing methodolo-
gies to the detailed estimation of external delay costs at New York/LaGuardia
and Los Angeles International, respectively. These independent studies come
up with strikingly similar results: many flights at these airports impose external
delay costs which exceed by at least an order of magnitude the landing fees
these flights pay. For example, Fan and Odoni (2002) estimate that the exter-
nal delay cost per movement for much of the day at New York/LaGuardia was
about $6000 in August 2001, whereas the average fee per movement amounted
to about $300. The implication is that access to many busy airports may be
greatly under-priced, thus attracting excessive demand, which exacerbates con-
gestion. This is especially true in the United States where landing fees are
comparatively very low. A third area of recent research addresses some diffi-
cult problems that set the application of market-based mechanisms at airports
apart from analogous applications in other contexts. For example, Bruckner
(2002) and Fan (2003) point out that each airline operates a (possibly large)
number of flights at airports, in contrast to highway traffic where each user op-
erates a single vehicle. When any single airline operates a large fraction of the
total movements at an airport, it also automatically absorbs (“internalizes”)
a similarly large fraction of the external delay costs that its flights generate.
That airline should therefore be charged only that portion of the external costs
that it does not internalize. But a pricing system under which different airlines
would pay different landing fees for the same type of aircraft would be both
controversial and technically difficult to implement. This and other complica-
tions, along with the presence of many social policy objectives (service to small
communities, access to airports by regional carriers and by general aviation,
etc.) suggest that any future RS schemes that incorporate market-based fea-
tures, will be governed by a complex set of rules that may include exemptions,
subsidies for certain carriers, slots reserved for certain types of flights, etc.
 Ch. 1. Air Transportation: Irregular Operations and Control 23

4 Air traffic flow management

4.1 Background and ATFM controls

4.1.1 Basic premises

Air traffic management (ATM) is now viewed as consisting of a tactical com-
ponent and a strategic component. The tactical component, Air Traffic Control
(ATC), is concerned with controlling individual aircraft on a time horizon rang-
ing from seconds to 30 minutes for the purpose of ensuring safe separation
from other aircraft and from terrain. The strategic component, Air Traffic Flow
Management (ATFM) works at a more aggregate level and on a time horizon
of up to about 12 hours in the United States and 48 hours in Europe. Its objec-
tive is to ensure the overall unimpeded flow of aircraft through the airspace, so
as to avoid congestion and delays and, when delays are unavoidable, to reduce
as much as possible their impact on airspace users.
The primary airports and primary airspace are dominated by the opera-
tions of scheduled air carriers. Air carrier schedules generate airspace demand
and also serve as the basis for measuring the performance of ATFM systems.
The fundamental premise of ATFM is that, roughly speaking, if all operations
occurred at their scheduled times and if all airspace elements were in their
“normal” operating states, then there would be little need for any flow man-
agement. Under such ideal conditions – and with the possible exception of
some brief periods at a few of the busiest airports – demand on all airspace
elements would be less than capacity and operations would generally proceed
as if there were no constraints. But ATFM recognizes that the complexity of
the airspace system, its susceptibility to weather conditions and the large num-
ber of possible ways in which equipment and/or operations can fail to operate
as planned, all imply that the probability that the entire system will operate
exactly according to schedule on any given day is essentially zero. ATFM proce-
dures are therefore in greatest demand when there are significant imbalances
between capacity and demand on airspace elements, usually caused by capac-
ity reductions due to weather or equipment failures. Demand surges can also
cause capacity-demand imbalances. Such surges may occur, for example, when
problems early in the day cause the postponement of scheduled operations into
a time interval that already contained significant numbers of operations. Addi-
tionally, the number of unscheduled flights is growing and is having a greater
impact on overall system performance. Finally, as the number of scheduled
operations has increased, there have been instances where scheduled demand
actually exceeded the capacity of network elements for extended time periods
each day. Probably the most notable example along these lines occurred at La-
Guardia airport between May 2000 and February 2001, i.e., during the period
between enactment of the so-called Air-21 legislation – which opened access
to LaGuardia for certain types of flights – and the imposition of slot lotteries
aimed at relieving the resulting congestion. In such cases, even the routine op-
eration of the air transportation system requires the use of ATFM procedures.
24 M. Ball et al.

4.1.2 Controls
ATFM performance is measured primarily with reference to deviations from
schedules. That is, ATFM systems generally seek to minimize the amount of
time by which actual operations (most importantly, the arrivals of aircraft at
their destination airports) deviate from scheduled operations. Thus, the key
performance indicators usually involve measures of delay. The next fundamen-
tal question to consider in describing and analyzing ATFM systems is what
controls can be used to impact system performance. Before directly addressing
this question we discuss two basic characteristics of controls: who makes and
implements the control decisions and what are the timing constraints associ-
ated with the control. The two critical entities involved in ATFM decisions are
the air traffic service providers and the airspace users. Airspace users could
range from the owner/operator of a general aviation aircraft to a large air car-
rier. Within a large air carrier there can be multiple decision-makers: most
central to this discussion are the airline operational control centers (AOCC)
and the pilots. On the air traffic service provider side, the units involved are
the regional air traffic control centers (Air Route Traffic Control Centers –
ARTCC in the US), as well as the central traffic flow management units (Air
Traffic Control System Command Center – ATCSCC in the US). Properly dis-
tributing decision-making among all these entities can be critical to the success
of ATFM systems and is at the heart of the recent emergence of the Collabora-
tive decision making (CDM) paradigm, which is described in Section 4.3. The
time constraints associated with control actions can greatly impact the manner
in which they can be applied and the manner in which multiple actions can be
coordinated. The key issue, in this respect, is the length of time that elapses
between a control decision and the implementation of that decision. A rough
categorization is that strategic decisions are typically made hours in advance
of implementation, whereas tactical decisions involve shorter time scales. Gen-
erally, the appropriateness and effectiveness of any particular control action,
strategic or tactical, depend on system dynamics and on the level of uncertainty
associated with future states of the airspace system.
We now describe the most important types of ATFM control actions.

Ground holding (including ground stops). Ground holding involves delaying

the departure of a flight in order to avoid overloading a capacitated system
element. Ground holding is most often implemented in the US through a
Ground delay program (GDP), which is put into effect when the demand for
arrivals into an airport is predicted to exceed significantly the arrival capacity.
In Europe, ground holding is commonly used to avoid overloading of en-route
sectors. Ground holding is generally considered a strategic decision. A “ground
stop” is a more tactical and extreme form of ground holding, whereby all de-
partures of aircraft bound for a particular destination airport are temporarily
postponed. A ground stop typically applies only to a specified set of airports,
usually ones that are proximate to the affected destination airport. In the past,
air traffic service providers were the sole decision makers when it came to de-
 Ch. 1. Air Transportation: Irregular Operations and Control 25

cisions regarding ground holding. More recently, however, CDM procedures

have led to shared decision making with airspace users.

Airborne speed control and airborne holding. Airborne speed control and air-
borne holding are tactical controls used to avoid overloading (en-route or
terminal) airspace elements by adjusting the time at which flights arrive at
those elements. Speed controls can involve simply slowing down or speeding up
aircraft or aircraft vectoring implemented through minor directional detours.
Airborne holding is usually implemented by having aircraft fly in oval-shaped
patterns. The providers of air traffic services usually make these control deci-
sions.

Route choice and route adjustments. Generally speaking airspace users con-
trol route choice. As discussed at the beginning of Section 2, airline and general
aviation operators choose and file flight plans based on several criteria in-
cluding winds and other weather conditions, fuel usage, en-route turbulence
predictions, safety constraints, etc. Flight plans may be filed several hours in
advance of departure but in many cases are not filed until shortly before de-
parture (even within an hour of departure) in order to take advantage of the
most recent information on weather conditions and congestion. For purposes
of managing congestion, air space managers can reject flight plans leading to
new filings. In some cases, standard reroute strategies are specified, e.g., the air
space manager designates that all flights originally filed along one airspace path
should re-file along a second alternate path. A variety of more tactical route
adjustments are possible. For example, alternative “departure routes” might
be specified by an airline; then, immediately before departure, one would be
chosen based on a negotiation process between the airspace manager and user.
Once airborne, major or minor route adjustments might be made. These de-
cisions generally involve some level of collaboration between the regional air
traffic service provider and the pilot or AOCC. A very extreme form of route
adjustment is the diversion of a flight, which involves altering its destination
airport.

Flight cancellations. The cancellation of a flight is another drastic, although

not particularly unusual, ATFM measure. Responsibility for flight cancella-
tions always rests with the airline or the general aviation operator.

Arrival sequencing. The sequencing of flights can be very important as the

maximum arrival rates into airports depend on the sequence and mix of aircraft
types (Section 2.1.3). Due to the uncertainty of en-route operations this should
be considered a tactical decision. Primary responsibility in this respect rests
with the air traffic service provider, although CDM concepts are now being
applied to this setting, leading to some air carrier participation in sequencing
decisions.
26 M. Ball et al.

Airport arrival or departure rate restrictions. As discussed in Section 2, there

can be a tradeoff between the maximum airport arrival and departure rates.
Evaluating this tradeoff and setting these rates should be considered a tacti-
cal decision that traditionally is made by regional air traffic service providers
(Gilbo, 1993). Proposals to allow air carrier participation in this decision have
been made (Hall, 1999).

Waypoint flow restrictions. An important and widely used control within the
US airspace system is the so-called miles-in-trail (MIT) restriction. Regional
air traffic service providers impose such restrictions in order to ensure that
the flow of aircraft into a region of airspace is kept at a safe level. When such
a restriction is put in place, the adjacent “upstream” regional air traffic ser-
vice provider has responsibility for maintaining a traffic flow at or below the
restricted level. This can be done in a variety of ways, including the use of air-
borne holding, rerouting of some traffic and issuing similar flow restrictions
on flights further upstream. In this way, it is possible for a flow restriction to
propagate through much of the airspace system, possibly eventually leading to
ground holds at airports of origin.

4.2 Deterministic models

Using as a starting point the capacity constraints described in Section 4.1,

one can formulate large-scale optimization models that map airspace demand
onto the various elements of the airspace in such a way that all capacity
constraints are respected. Such models trace flights through both space and
time and seek to minimize some global demand function. In Section 4.2.1 we
describe a large-scale comprehensive modeling approach and then, in Sec-
tion 4.2.2, specialize this model to the ground holding problem.

4.2.1 Global models

There are two broad classes of global air traffic flow models. The first as-
sumes that the trajectory (route) of each flight is fixed and optimizes the timing
of the flight’s operations. The second allows the route of each flight to vary, as
well. Clearly, the second type of model has a much larger decision space.
We start with models of the first type. The modeling approach chooses a
time horizon of interest and decomposes it into a discrete set of time intervals.
A geographic scope is also selected. This determines the set of capacitated
elements under consideration, including airports, sectors, and waypoints. The
combination of the model’s temporal and geographic scope determines the set
of flights to be considered. The basic decision variable specifies the airspace
element occupied by a flight at each time interval, i.e.,

1 if flight f occupies airspace element e
xf te = during time interval t;
0 otherwise
 Ch. 1. Air Transportation: Irregular Operations and Control 27

For airports, the capacitated airspace element would be either the airport’s ar-
rival or departure component. For example, LaGuardia Airport (LGA) would
be represented by an arrival element LGA1 and a departure element LGA2,
with xf 1tLGA1 indicating the time interval t of a flight f 1’s arrival to LGA and
xf 2tLGA2 indicating the time interval t for flight f 2’s departure from LGA.
The capacity constraint associated with an element e and time interval t is
of the form:

xf te  cap(t e) for all t and e
f
where cap(t e) is the capacity of element e during time interval t. For airport
arrival and departure capacities and for waypoints, cap(t e) is equal to the
maximum number of flights that could flow through that element during time
interval t. For a sector, it is equal to the maximum number of flights that can
occupy the sector simultaneously.
The remaining constraints define temporal restrictions on the manner in
which each flight can progress through the airspace. For example, they might
specify that, once a flight enters a sector, it must remain in the sector for 3, 4,
or 5 time intervals. In this case, 3 time intervals would correspond to travers-
ing the sector on a direct path at maximum speed and 5 time intervals might
correspond to a longer traversal time based on application of some type of
speed control. We note that since the flight’s route is an input, the progression
from departure airport through a specific sequence of sectors to a destination
airport is a fixed model input, as well.
Bertsimas and Stock Paterson (1998) have shown that models of this type
can be solved very efficiently. Of particular note is their use of an alternative
set of decision variables. Specifically, the xf te variables are replaced with a set,
wf te , defined by

wf te = 1 if flight f arrives at airspace element e by time interval t;
0 otherwise
While the w variables can be obtained
 from the x variables through a simple
linear transformation (wf te = ts=1 xf se ), the w variables are much easier to
work with because they produce very simple and natural temporal progression
constraints. Further, the associated linear programming relaxations are very
“strong” in the sense that they lead to the fast solution of the associated inte-
ger programs. Bertsimas and Stock Patterson show that a variety of additional
features can be included in the model, including the propagation of delays that
occurs when a delay in the arrival of a flight arrival causes the delay of an
outbound flight that uses the same aircraft. The model can also capture the
dependence between an airport’s arrival and departure capacities and choose
the appropriate combination of the two for each time interval.
A second type of model also allows for flight routes to vary. Bertsimas and
Stock Paterson (2000) extend the model described above to this case. Since
the decision space becomes much larger, aggregate variables and approximate
methods must be employed in order to solve problems of realistic size.
28 M. Ball et al.

4.2.2 Deterministic ground holding models

A simple, yet very important special case of the model just described is the
deterministic ground holding problem. For this problem, en-route constraints
and airport departure constraints are ignored leaving only a constraint on ar-
rival capacity. These assumptions allow a large multi-airport problem to be
decomposed into separate problems, one for each arrival airport. This model
is very important to the US ATFM environment since it underlies the Ground
delay programs (GDP) as currently operated. Although the control variable for
a GDP is ground delay at the origination airport of each flight, the problem can
be modeled as one of assigning flights to arrival time intervals at the destina-
tion airport. For each flight a (constant) en-route time is then subtracted from
the arrival time to obtain a departure time, which in turn implies an amount of
ground delay at the origination airport.
Define the following inputs: for a set of time intervals t (t = 1 2     T )
and set of flights f (f = 1 2     F), let
bt = the constrained airport’s arrival capacity at time interval t
e(f ) = the earliest time interval at which flight f
can arrive at the constrained airport
cf t = the cost of assigning flight f to arrive at time interval t
and the variable

xf t = 1 if flight f is assigned to time interval t;
0 otherwise
Then, the deterministic ground holding problem can be formulated as

min cf t xf t
f t

subject to xf t  bt for all t
f

xf t = 1 for all f
te(f )
xf t  0 and integer for all f and t
As can be seen, this is a simple transportation model that can be solved very
efficiently. This model was first described in Terrab and Odoni (1993). In Ball
et al. (1993), various issues related to its practical use were investigated. In
particular, it was suggested that the definition cf t = rf (t−e(f ))1+ε is attractive
since flight delay costs tend to grow with time at a greater than linear rate. In
addition, solutions produced using this objective function are attractive from
the standpoint of equity or fairness – see Vossen and Ball (2006a) for a detailed
discussion of this issue. Hoffman and Ball (2003) investigate generalizations of
this model that preserve the proximity of banks of flights associated with airline
 Ch. 1. Air Transportation: Irregular Operations and Control 29

hubbing operations. Vranas et al. (1994) describe a multi-airport version of

the problem where both airport arrival and departure rates are constrained
and flights are “connected” to each other, in the sense that the late arrival
of a flight f at an airport may also imply the late departure of one or more
subsequent flights from that airport. This can happen if the departing flight
must use the same aircraft as flight f or, less obviously, if flight f carries many
transfer passengers who must board subsequent departing flights.

4.3 Uncertainty and stochastic models

Uncertainty on multiple levels has led to the failure of many attempts at

practical implementation of various air traffic flow management models. This
is particularly true for models that attempt to optimize over a broad geographic
area and/or extended periods of time. To be effective, models must include
stochastic components explicitly or they must address problems restricted to
limited geographic and time domains for which available information is less
subject to uncertainty. We use the term demand uncertainty to describe the pos-
sibility that, due to random or unforeseen events, flights may deviate from their
planned departure or arrival times or from their planned trajectories. Similarly,
capacity uncertainty refers to the possibility that random or unforeseen events
will cause changes to the maximum achievable flow rates into and out of air-
ports or through airspace waypoints or to the maximum number of flights that
can occupy simultaneously a portion of the airspace. Examples of factors con-
tributing to demand uncertainty include problems in loading passengers onto
an aircraft, mechanical problems, queues on the departure airport’s surface or
in the air, and en-route weather problems. Examples of factors contributing to
capacity uncertainty include weather conditions at an airport or in the en-route
airspace, failures or degradation of air traffic control equipment, unavailability
of air traffic control personnel and random changes in flight sequences, and
aircraft mix that require alterations of anticipated flight departure or arrival
spacing.
The largest body of work in this area has focused on ground holding models
that explicitly take into account uncertainty in airport arrival capacity. These
models are covered in Section 4.3.1. There is a less extensive body on work on
capacity uncertainty for GDPs, which is covered in Section 4.3.2.
The performance of a GDP planning and control system can be evaluated
based on three measures: the total assigned ground delay, the total airborne
delay and the utilization of the available arrival capacity. Figure 5 illustrates
a generic model (Odoni, 1987) that is convenient for evaluating the tradeoffs
among these three performance criteria and for understanding how they are
impacted by demand and capacity uncertainty. The airport’s arrival compo-
nent (terminal airspace and runways) is viewed as a server subject to demand
uncertainty. The assigned ground delay (the control mechanism) and the ran-
dom events that perturb the planned arrival stream of flights determine the
rate of arrivals at the server. Thus, demand uncertainty impacts the rate of
30 M. Ball et al.

Fig. 5. Schematic representation of a single-airport GDP.

arrivals at the server and capacity uncertainty impacts the service rate. The
planning and control of a GDP must balance the possibility of too low an ar-
rival rate, which leads to underutilization of the server, with too high an arrival
rate, which leads to a large airborne queue and excessive airborne delay. Gen-
erally speaking because of demand and capacity uncertainty, the “best” policy
calls for some degree of airborne delay in order to ensure an acceptably inten-
sive utilization of the available arrival capacity. The starting point for defining
airport arrival capacity is specifying the number of flights that can land within
a time interval (this quantity is referred to as the airport acceptance rate –
AAR). A key observation in light of this discussion is that the presence of
demand and capacity uncertainty makes it necessary to distinguish between
the planned AAR (PAAR) and the actual AAR. In the following two sections
we discuss two types of models: those that assign ground delay to individual
flights and those only concerned with determining a PAAR vector. The second
objective (determining a PAAR vector) is appropriate in the context of Col-
laborative Decision Making (CDM), where the assignment of ground delays
to individual flights results from a complex set of distributed decision making
procedures. Thus, the second set of models is the one more directly geared
to the CDM environment. Interestingly, some models in the first set can also
generate the same PAAR vectors through a simple “post-processing” step, as
has been shown by Kotnyek and Richetta (2004).

4.3.1 Ground holding models with stochastic airport capacity

As discussed in the previous section, optimization models for the ground
holding problem subdivide time into an arbitrary number of discrete intervals.
Typical time intervals might be 10 or 15 minutes or even as much as 1 hour
for the most aggregate models. One would then characterize the AAR over a
period of time, e.g., four hours, as a vector that provides the AAR in each of
the constituent time intervals. In a typical GDP caused by a weather distur-
bance that moves through a region, the AAR would start at its normal level,
e.g., 60 arrivals per hour, decrease to one or more degraded levels, e.g., 30 ar-
rivals per hours, for several time intervals, e.g., 4 hours, and then return to its
original level. If it were known that such a scenario would occur with certainty,
 Ch. 1. Air Transportation: Irregular Operations and Control 31

then a deterministic ground holding model (Section 4.2.2) could obviously be

used with this scenario providing the capacity constraints input. By contrast,
stochastic models treat the AAR as a random variable and use as input several
such scenarios together with associated probabilities. For example, an “opti-
mistic” scenario indicating no capacity degradation would consist of a vector
of hourly AARs of 60 throughout the period of interest, whereas, a more pes-
simistic scenario might assume that the AAR will be 30 during every hour in
the period. The input can thus be characterized as

Dtq = AAR for time interval t under scenario q

for t = 1     T and q = 1     Q
pq = probability of the occurrence of scenario q for q = 1     Q
The xf t variables are then defined as in the deterministic ground holding
model, but the capacity constraints are replaced with a new set of scenario-
based constraints and associated variables. The new variable set is defined by

ytq = the number of flights held in the air from time period t to t + 1
under scenario q for q = 1     Q
The new set of capacity constraints then

xf t + yt−1q − ytq  Dtq for t = 1     T and q = 1     Q
f =1F

Thus, under these constraints, there is a separate capacity for each scenario.
However, the y variables allow for flow between time intervals, so the number
of flights assigned to land in an interval under a particular scenario can exceed
the AAR by allowing excess flights to “flow” to a future time interval. Note
that this set of capacity constraints defines a small network flow problem for
each q, with flights “flowing” from earlier time intervals to later ones. To be
feasible, for each given q, the total arrival capacity for the entire period of
interest, t Dtq , must be at least as large as the total number of flights (F).
The objective function for the model minimizes the sum of the cost of
ground delay plus the expected cost of airborne delay. If we define c a as the
cost of holding one flight in the air for one time period then we can define the
objective function as
   
min cf t xf t + pq c a yt−1q 
f =1F t=1T q=1Q t=1T

This class of model was first described in Richetta and Odoni (1993) and a
dynamic version was given in Richetta and Odoni (1994). Ball et al. (2003) de-
fines a simpler version of this model that computes a PAAR vector (i.e., the
total number of flights assigned to arrive during each time interval in the pe-
riod) rather than an assignment of ground delay to individual flights or groups
32 M. Ball et al.

of flights. This model was created in order to be compatible with CDM pro-
cedures, which will be discussed in Section 4.4. The constraint matrix of the
underlying integer program is the transpose of a network matrix, allowing
the integer program to be solved using linear programming or network flow
techniques. Inniss and Ball (2002) and Wilson (2004) describe recent work
on estimating arrival capacity distributions. Willemain (2002) develops simple
GDP strategies in the presence of capacity uncertainty and also takes into ac-
count the possibility of flight cancellations. Inniss and Ball (2002) also presents
a dynamic approach to this problem.

4.3.2 Modeling demand uncertainty

Recent experience with new decision support tools for planning and con-
trolling GDPs has strongly suggested that demand uncertainty is the reason
that these tools have not performed quite as strongly as expected. In this con-
text, demand uncertainty refers to differences between the planned and actual
characteristics of the stream of flights arriving at the destination airport. These
differences can be attributed to three categories of causes:
(1) cancellations – flights in the planned arrival stream that are absent from
the actual arrival stream during the planning time period;
(2) “pop-ups” – flights that arrive at the destination airport during the plan-
ning time period, but were not in the planned arrival stream;
(3) “drift” – a discrepancy between the actual arrival time of a flight and
the flight’s planned arrival time.
Ball et al. (2001c) present a demand uncertainty model, where each flight
can be canceled with a known probability, pop-ups arrive in each time interval
according to a binomial distribution and flights drift within predefined intervals
according to a uniform distribution. Considering only cancellations and pop-
ups, the authors provide an optimization model that determines a PAAR that
minimizes airborne delay subject to a constraint on the minimum allowable air-
port capacity utilization. This integer programming model contains embedded
binomial distributions. The authors also present simulation results covering all
three forms of demand uncertainty. The results not only provide new PAAR
setting policies, but also give an approach to estimating the benefits of reduc-
ing demand uncertainty. For additional details see Bhogadi (2002) and for a
related model Willemain (2002).

4.4 Collaborative decision making

The Collaborative decision making (CDM) (effort Ball et al., 2001a, 2001b;
Chang et al., 2001; Wambsganss, 1996) grew out of a desire on the part of
both the airlines and the FAA for improvements in the manner in which GDPs
were planned and controlled. The FAA and, more specifically, the ATCSCC
had realized the need for more up-to-date information on the status of flights
currently delayed due to mechanical or other problems or even canceled unbe-
knownst to the ATCSCC. Equally important, the success of GDPs also depends
 Ch. 1. Air Transportation: Irregular Operations and Control 33

vitally on timely information regarding airline intentions vis-à-vis flight cancel-

lations and delays over the next few hours. At the same time, the airlines did
not feel the allocation procedures used by the ATCSCC were always fair and
efficient. In addition, each airline wished to gain more control over how de-
lays were allocated among its own flights. Thus, both the airlines and the FAA
had specific (although different) objectives that motivated their participation
in CDM.
The CDM “philosophy”, broadly speaking, represents an application of the
principles of information sharing and distributed decision making to ATFM.
Specific goals are:
• generating a better “knowledge base” by merging information pro-
vided by the airspace users with the data that are routinely collected
by monitoring directly the airspace;
• creating common situational awareness by distributing the same infor-
mation to both traffic managers and airspace users;
• creating tools and procedures that allow airspace users to respond di-
rectly to capacity/demand imbalances and to collaborate with traffic
flow managers in the formulation of flow management actions.

4.4.1 CDM resource allocation procedures for ground delay programs

CDM brings new modeling requirements to ATFM resource allocation
problems including: considering allocation at a more aggregate level, e.g., by
allocating resources to airlines rather than to individual flights; integrating eq-
uity criteria into model objectives and/or constraints; ensuring that resource
allocation mechanisms provide incentives for information sharing; and devel-
oping inter-airline resource exchange mechanisms. These concepts evolved in
the process of developing new allocation methods for GDPs. We describe the
GDP procedures here, as well as related recent research and efforts to extend
these ideas to en-route ATFM problems.
As discussed in Section 4.2.2, the GDP planning problem can be viewed as
one of assigning each flight in the GDP to an arrival time interval (or time slot).
Figure 6 illustrates the GDP resource allocation process under CDM. The
FAA, using Ration-by-schedule (RBS), provides an initial assignment of slots
to flights. Each airline, using the cancellations and substitution process, may
then cancel flights and modify slot-to-flight assignments for its own flights (in-
tra-airline exchange). Thus, although RBS, in concept, allocates slots to flights,
the cancellation and substitution process effectively converts the slot-to-flight

Fig. 6. CDM resource allocation procedures.

34 M. Ball et al.

assignment into a slot-to-airline assignment. The final step, compression, which

is carried out by the FAA, maximizes slot utilization by performing an inter-
airline slot exchange in order to ensure that no slot goes unused.
The assignment of time slots by RBS can be viewed as a simple priority rule.
Using the scheduled arrival order as a priority order, each flight is assigned the
next available arrival slot. If this rule were applied to all flights and there were
no cancellations or substitutions, then the flights would arrive in their original
sequence, but generally later in time. Two groups of flights are exempted from
this basic allocation scheme:
(1) flights that are currently airborne (clearly these cannot be assigned any
ground delay) and
(2) a set of flights selected according to the distance of their departure air-
ports from the GDP (arrival) airport (Ball and Lulli, 2004).
The motivation for (2) is to include in the allocation scheme flights origi-
nating from airports that are close to the GDP airport and to exempt flights
from airports further away from the GDP airport. The reasoning is that flights
a greater distance away must be assigned ground delays well in advance of their
actual arrival at the GDP airport, e.g., 4 or 5 hours in advance. At such a long
time before arrival, there tends to be a greater level of uncertainty regarding
weather and, as a consequence, airport arrival capacity. If these distant flights
are assigned ground delay, there is a significant likelihood that this ground
delay might prove unnecessary. Thus, distance-based exemptions constitute a
mechanism for avoiding unrecoverable delay, as well as for increasing expected
airport throughput.
After the round of substitutions and cancellations, the utilization of slots
can usually be improved. The reason for this is that an airline’s flight cancel-
lations and delays may create “holes” in the current schedule; that is, there
will be arrival slots that have no flights assigned to them. The purpose of the
compression step is to move flights up in the schedule to fill these slots. The
basic premise behind the algorithm currently used to perform compression is
that airlines should be “paid back” for the slots they release, so as to encourage
airlines to report cancellations.
To illustrate the compression algorithm, consider the example shown in Fig-
ure 7. The left side of the figure represents the flight-slot assignment prior to
the execution of the compression algorithm. Associated with each flight is an
earliest time of arrival, and each slot has an associated slot time. Note that
there is one canceled flight. The right side of the figure shows the flight sched-
ule after execution of the compression algorithm: as a first step, the algorithm
attempts to fill AA’s open slot at 1602–1603. Since, there is no flight from AA
that can use it, the slot is allocated to UA, and the process is repeated with the
next open slot, which, using the same logic, is assigned to US. The process is
repeated for the next open slot, which is now assigned to AA. AA thus receives
the earliest slot that it can use. The net result of compression in this case can
be viewed as an exchange among airlines of the slots distributed through the
initial RBS allocation.
 Ch. 1. Air Transportation: Irregular Operations and Control 35

Fig. 7. Execution of the CDM compression algorithm.

4.4.2 CDM concepts, philosophy, and research directions

We now investigate the special properties of the CDM resource allocation
procedures and describe how these are being extended and enhanced.

RBS, compression and information sharing. One of the principal initial mo-
tivations for CDM was that the airlines should provide updated flight status
information. It was quickly discovered that the existing resource allocation
procedures, which prioritized flights based on the current estimated time of
arrival, actually discouraged the provision of up-to-date information. This was
because, by updating flight status, the airlines would change the current esti-
mate of flight arrival (almost always to a later time), which in turn changed
their priority for resource allocation. On the other hand, the RBS priority,
which is based on the (fixed) schedule, does not vary with changes in flight sta-
tus. Further, as in the example of Figure 7, compression most often provides an
airline with another usable slot whenever an announced cancellation generates
a slot that the airline cannot fill via the substitution process.

Properties of RBS. We summarize some basic properties of RBS derived in

Vossen and Ball (2006a). RBS assigns to each flight, f , a controlled time of
arrival, CTA(f ). This is equivalent to assigning a delay, d(f ), to flight, f , given
by d(f ) = CTA(f ) − OAG(f ), where OAG(f ) is the scheduled arrival time
for f . All time values are rounded to the nearest minute under RBS, hence,
each delay value d(f ) is integer. If we let D equal the maximum delay assigned
to any flight and ai = |{f : d(f ) = i}| for i = 0 1 2     D, then the impor-
tant properties of unconstrained RBS (RBS with no flight exemptions) can be
defined by the following properties.
36 M. Ball et al.

Property 1. RBS minimizes total delay = f d(f ).

Property 2. RBS lexicographically minimizes (aD      a1  a0 ). That is, aD is

minimized; subject to aD being fixed at its lexicographically minimum value,
aD−1 is minimized; subject to (aD  aD−1 ) being fixed at its lexicographic mini-
mum value, aD−2 is minimized; and so on.

Property 3. For any flight f , the only way to decrease a delay value, d(f ), set
by RBS is to increase the delay value of another flight g to a value greater than
d(f ).

CDM and equity. Property 3, which follows directly from Property 2, expresses
a very fundamental notion of equity that has been applied in a number of con-
texts (Young, 1994). It is remarkable that procedures, such as RBS, which
were developed in very practical war-gaming and consensus-building exer-
cises have such elegant and desirable properties. On the other hand, this may
not be surprising in that these properties represented the basis for reaching
consensus, in the first place. The properties show that unconstrained RBS pro-
duces a fair allocation. However, one should consider whether the exemption
policies described earlier, in fact, introduce bias. Vossen et al. (2003) show
that exemptions do introduce a bias and also describe procedures for miti-
gating these biases. The approach taken initially computes the unconstrained
RBS solution and defines it as the “ideal” allocation. Optimization procedures
are then described that minimize the deviation of the actual allocation from
the ideal. These procedures build upon approaches developed in connection
with just-in-time (JIT) manufacturing for minimizing the deviation of an ac-
tual production schedule from an ideal schedule – see, for example, Balinski
and Shihidi (1998). The resulting approaches maintain the exemption policies,
but take into account the “advantages” provided to an airline by its exempted
flights when allocating delays to its other flights.

Compression as trading. Although the initial interpretation of compression is

as a slot reallocation procedure that maximizes slot utilization, there is also a
natural interpretation of compression as an inter-airline trading or bartering
process (Vossen and Ball, 2006a). For example, in Figure 7, American Airlines
“traded” the 1602–1603 slot, which it could not use for the 1606–1607 slot,
which it could use, and United Airlines reduced its delay by trading the
1604–1605 slot for the earlier 1602–1603 slot. Vossen and Ball show that a bar-
tering process can be structured so as to produce a result essentially equivalent
to compression. This view of compression suggests many possible extensions.
For example, Vossen and Ball (2006b) define a more complex 2-for-2 bartering
mechanism and show that using this mechanism offers a substantial potential
for improved economic performance. Probably the most intriguing enhance-
ment is allowing “side payments” with any exchange as well as the buying and
 Ch. 1. Air Transportation: Irregular Operations and Control 37

selling of slots. Ball et al. (2005) provide a discussion of this and other aviation-
related market mechanisms.

4.5 Air carrier response options

As the implementation of the CDM concept expands, it will become increas-

ingly difficult to designate certain controls as exclusive to the traffic managers
or exclusive to the users (airlines). In this section we describe the types of ac-
tions and types of problems that air carriers can take in response to ATFM
initiatives based on current practices. We start by noting that the decision on
whether or not a flight takes place on a given day rests with the airlines, with the
exception of extreme circumstances. Further, the airlines generally control the
route choice for a flight subject to constraints on access to portions of airspace
exercised by ATFM. The airlines can, of course, at their own discretion choose
to delay the departure of flights. On the other hand, ATFM exercises strong
control over the timing of operations. As described in the previous section, for
GDPs in the US, ATFM also exercises control over flight timing by allocating
arrival time slots to airlines, which in turn allocate the slots to individual flights.
For the case of GDPs, we can model an airline’s response problem as one
of assigning flights to the set of time slots that airline “owns”. A first-level ap-
proach to this problem would select an assignment that minimizes the “cost”
of the associated flight delays. Such models should also consider the possibil-
ity and cost of canceling flights. Second-level models would also consider the
downstream effects of the delayed flights. For example, the immediate down-
stream effect involves the delays on flights and passengers outbound from the
GDP airport (Niznik, 2001). Broader airline network models would consider
the airport arrival slots as a resource to be allocated as part of a larger opti-
mization model.
The problem of choosing a route is traditionally addressed on a flight-
by-flight basis. There are a variety of safety regulations that constrain this
problem. Typically, weather conditions, including winds, turbulence, convec-
tive activity, etc., play a strong role in solving this problem. Airspace con-
gestion considerations, as well as ATFM control actions are also playing an
increasingly important role. There can be a high degree of uncertainty in
many of these factors leading to the need for stochastic, dynamic problem
solving approaches – see, for example, Nilim et al. (2001). Furthermore, as
CDM concepts become pervasive, the air carriers may be allocated limited
airspace resources that in turn need to be allocated to individual flights. Un-
der such a scenario, flight-by-flight route planning may no longer be viable.
Berge et al. (2003) describe a comprehensive optimization model that includes
decision variables for the three controls described earlier, i.e., flight cancella-
tions, aircraft route choice, and ground delay. It assumes an environment in
which airspace and airport resources have been allocated to each airline. This
model was developed for integration into a comprehensive Boeing airspace
38 M. Ball et al.

simulation. It can, however, also be viewed as a prototype for future opera-

tional airline decision models.
Sections 6 and 7 describe models for planning adjustments to airline opera-
tions based on day-to-day changes in airspace conditions and airline resource
status. Such models typically are only invoked in the case of reasonably signif-
icant disturbances to “normal” conditions and the underlying environment is
called irregular operations. While these models take into account some of the
concepts described in this section, a full integration of the emerging ATFM
CDM philosophy with airline planning models represents a research challenge.

5 Simulation models

Simulation models are useful support tools in understanding and visualizing

the impact of certain types of disruptive events on airport, airspace, and airline
operations and on air traffic flows, as well as in testing the effectiveness of
potential responsive actions. This section reviews briefly some of these tools.
For analyzing impacts at the local airport or airspace level or for testing the
effectiveness of tactical ATFM actions, simulation models need to be highly
detailed. When it comes to questions at a regional or strategic level, models
of a more aggregate nature are often more appropriate. Detailed (or “mi-
croscopic”) simulation models of airport and airspace operations first became
viable in the early 1980s and have been vested with increasingly sophisticated
features since then. The models that currently dominate this field interna-
tionally are the Airport and Airspace Simulation Model (better known as
SIMMOD), the Total Airport and Airspace Modeler (TAAM), and the Re-
organized ATC Mathematical Simulator (RAMS). The first two are widely
used in studies dealing with the detailed planning and design of airports and/or
of volumes of airspace. SIMMOD is publicly available through the FAA, but
more advanced proprietary versions can be obtained through private vendors
(see, e.g., ATAC, 2003). TAAM is a proprietary model (Preston, 2002). Fi-
nally, RAMS (Eurocontrol, undated) is also a proprietary model limited to
detailed simulation of airspace operations and procedures and is used either
as a training tool for air traffic controllers or for studies of controller workload
in airspace sectors. To our knowledge, none of these, or any other, detailed
simulation models is currently utilized on a routine basis as a “real time” sup-
port tool for the types of dynamic tactical or strategic ATFM actions described
in Section 4. However, the models have reached a state of development where
such use is technically feasible. For example, in a case where a runway at an air-
port is temporarily out of use (e.g., due to weather conditions or to a disabled
aircraft), one could simulate and compare the effectiveness of alternative allo-
cation schemes for the assignment of arrivals and departures to other runways
or of various runway-use sequences. Odoni et al. (1997) provides detailed –
and somewhat dated, by now – descriptions of several microscopic simulation
models, including the three mentioned above.
 Ch. 1. Air Transportation: Irregular Operations and Control 39

Another class of models has been developed to support more aggregate

analysis typically involving a broader scope than the problems addressed by
SIMMOD or TAAM. For example, NASA has sponsored the development of
FACET (Future ATM Concepts Evaluation Tool – Bilimoria et al., 2000). Typ-
ical uses might involve analyzing the impact over the entire national airspace
of new traffic flow management initiatives, air-ground distributed control ar-
chitectures, and decision support tools for controllers. FACET models system-
wide en-route airspace operations over the contiguous United States. It strikes
a balance between flexibility and fidelity enabling the simultaneous represen-
tation of over 5000 active flights on a desktop computer.
The CDM activities have also required the extensive use of human-in-the-
loop (HITL) experiments in order to test new distributed decision-making
ideas. These experiments were initially supported by one-of-a-kind computer-
communications architectures. More recently the FAA has funded the devel-
opment by Metron Aviation of the Jupiter Simulation Environment (JSE). JSE
can emulate the message stream generated by the FAA’s Enhanced Traffic
Management System (ETMS). ETMS provides real-time information on the
status of all flights operating within the US airline operational control cen-
ters and FAA facilities can connect their traffic monitoring systems to the JSE
and participate in HITL experiments involving new decision support tools or
operational concepts. For example, the JSE allowed for the rapid testing and
evaluation of the slot credit substitution (SCS) concepts and tools prior to their
release.
Simulation models are equally important on the airline operations side. To
evaluate the recovery procedures and plans for fleets and crews under op-
erational conditions, it is necessary to have them simulated by a model of
airline operations. MEANS (Clarke et al., 2004) and SIMAIR (Rosenberger
et al., 2002) are such simulations. The MIT Extensible Air Network Simula-
tion (MEANS) can be used to predict the effects of air traffic control, traffic
flow management, airline operations control, and airline scheduling actions on
air transportation system performance, measured in terms of airport conges-
tion and throughput, and aircraft, crew, and passenger delay, and disruption.
MEANS has a modular architecture and interface, with each module corre-
sponding to a set of operational decisions made by air transportation coordina-
tors and controllers. For example, flight cancellation and re-routing decisions
are made in the Airline Module; the amount of traffic allowed between airports
is determined by the Traffic Flow Management Module; and the sequencing
and spacing of aircraft is performed by the Tower/TRACON Module. This
modular structure provides flexibility for implementing additional and/or more
complex modules without requiring changes to the core interfaces.
Another stochastic model of airline operations, SIMAIR, uses an Event
Generator module to generate events such as arrivals, departures and repaired
planes. The generator samples random ground time delays, block time delays,
and unscheduled maintenance. SIMAIR contains two modules for decision-
making. The Controller module maintains the state of the system. It emulates
40 M. Ball et al.

an Airline Operational Control Center in the sense that it recognizes disrup-

tions and implements recovery policies. If a disruption prevents a leg from
being flown as planned, the Controller requests a proposed reaction from the
Recovery Module, which it can accept or request another, time permitting.
SIMAIR has been used to evaluate the recovery procedures and robust
plans presented in several of the papers discussed in Sections 6 and 7.
A SIMAIR Users Group is in place, consisting of several research groups and
airlines that have used or contributed to the development of SIMAIR.

6 Airline schedule recovery

When disruptions occur, airlines adjust their flight operations by delaying

flight departures, canceling flight legs, rerouting aircraft, reassigning crews or
calling in new crews, and re-accommodating passengers. The goal is to find
feasible, cost-minimizing plans that allow the airline to recover from the dis-
ruptions and their associated delays. To address this challenge, airlines have
established Airline Operational Control Centers (AOCC) to control safety of
operations, exchange information with regulatory agencies, and manage air-
craft, crew, and passenger operations. The AOCC is comprised of (Bratu, 2003;
Clarke et al., 2000; Filar et al., 2000):
• Airline operations controllers who, at the helm of the AOCC, are respon-
sible for aircraft re-routing, and flight cancellation and delay decisions
for various types of aircraft.
• Crew planners who find efficient recovery solutions for crews and co-
ordinate with airline operations controllers to ensure that considered
operations decisions are feasible with respect to crews.
• Customer service coordinators who find efficient recovery solutions for
passengers and coordinate with airline operations controllers to pro-
vide an assessment of the impact on passengers of possible operations
decisions.
• Dispatchers who provide flight plans and relevant information to pilots.
• An air traffic control group that collects and provides information, such
as the likelihood of future ground delay programs, to airline operations
controllers.
The AOCC is complemented by Station Operations Control Units, located
at airport stations, responsible for local decisions, such as the assignment of
flights to gates, ground workforce to aircraft, and personnel for passenger ser-
vice.
Airline operations recovery is replete with challenges, including:
1. The recovery solution must take into account the recent flying history
of the aircraft, crews, and passengers to ensure that crew work rules are
satisfied, aircraft maintenance and safety regulations are met, passengers
 Ch. 1. Air Transportation: Irregular Operations and Control 41

are transported to their desired destinations, and aircraft are positioned

appropriately at the end of the recovery period.
2. The recovery solution can utilize additional resources, namely reserve
crews and spare aircraft (Sohoni et al., 2002, 2003).
3. There are multiple recovery objectives, namely: minimizing the cost of re-
serve crews and spare aircraft used; minimizing passenger recovery costs;
minimizing the amount of time to resume the original plan; and minimiz-
ing loss of passenger goodwill.
4. The recovery problem often must be solved quickly, often within minutes.
To meet these challenges, most airline recovery processes are sequential
(Rosenberger et al., 2003). The first step in the process is to recover aircraft,
with decisions involving flight leg cancellation or delay, and/or aircraft re-
routing. The second step is to determine crew recovery plans, assigning crews
to uncovered flight legs by reassigning them or utilizing reserve crews. Finally,
the third step is for customer service coordinators to develop passenger re-
accommodation plans for disrupted passengers. A disrupted passenger is one
whose planned itinerary is broken and impossible to execute during operations
because: (1) at least one of the flight legs in the itinerary is canceled; or (2) the
connection time between consecutive flight legs in the itinerary is too short due
to flight delays. While the AOCC decision process is hierarchical in nature, air-
line operations controllers, crew planners, and passenger service coordinators
consult with one another during the process to assess the feasibility and impact
of possible decisions.
This sequential decision process, first aircraft, then crew, and finally passen-
ger recovery, is reflected in the research on airline recovery performed to date.
In the following subsections, we present selected airline recovery research.

6.1 Aircraft recovery

When schedule disruptions occur, the aircraft recovery problem is to deter-

mine flight departure times and cancellations, and revised routings for affected
aircraft. Re-routing options include ferrying (repositioning an aircraft without
passengers to another location, where it can be utilized); diverting (flying to an
alternate airport); over-flying (flying to another scheduled destination); and
swapping (flight legs are re-assigned among different aircraft). These modifi-
cations must satisfy maintenance requirements, station departure curfew re-
strictions and aircraft balance requirements, especially at the start and end of
the recovery period. At the end of the period, aircraft types should be posi-
tioned to resume operations as planned.
The aircraft balance requirements add complexity to cancellation decisions.
Normally, to ensure that aircraft are positioned where needed to fly down-
stream flight legs, cancellations involve cycles of 2 or more flight legs. To cancel
only a single flight leg l and still be able to execute the remaining schedule, it
is necessary to deploy a spare aircraft of the type assigned to the destination of
42 M. Ball et al.

leg l. Because spare aircraft are typically in very limited supply, canceling only
a single flight leg is not usually an option.
Beyond the inherent complexities of re-routing aircraft, scheduling delayed
flight departures and making cancellation decisions, an effective aircraft recov-
ery solution approach accounts for the downstream costs and impacts on crew
and passengers. The extent to which these complexities are captured in models
varies, with increasing sophistication achieved over time.
Arguello et al. (1997) present an integrated aircraft delays and cancellations
model and generate sequentially, for each fleet type, a set of aircraft routes that
minimize delays, cancellations, and re-routing costs. Their model ensures air-
craft balance by matching aircraft assignments with the actual aircraft locations
at the beginning of the recovery period and with the planned aircraft locations
at the end of the period (that is, the end of the day).
The model includes two types of binary decision variables; namely, mainte-
nance-feasible aircraft routes and schedules, and flight cancellation decisions.
An aircraft route is a sequence of flight legs spanning the recovery period, with
the origin of a flight leg the same as the destination of its predecessor in the
sequence, and the elapsed time between two successive legs at least as great as
the minimum aircraft turn time. Routes for aircraft with planned maintenance
within the recovery period are not altered to ensure that the modified routes
satisfy maintenance requirements.
Let P be the set of aircraft routes, Q be the set of aircraft, and F be the
set of flight legs. Aircraft route variable xkj equals 1 if aircraft k is assigned to
route j and 0 otherwise. Its cost, denoted djk , equals the sum of the delay costs
associated with flight delays implied by assigning aircraft k to route j. Note
that djk is infinite for each aircraft route j commencing at an airport location
other than that of aircraft k at the start of the recovery period. A flight cancel-
lation variable, denoted yi , is set to 1 if flight leg i is canceled and 0 otherwise.
The approximate cost associated with the cancellation of each flight leg i is ci ;
ht equals the number of aircraft needed at airport location t at the end of the
recovery period to ensure that the next-day plan can be executed; δij is equal
to 1 if flight leg i is covered by route j; and btj is equal to 1 if route j ends at
the airport t.
The aircraft recovery model is
 
min djk xkj + ci yi
k∈Q j∈P i∈F
subject to

δij xkj + yi = 1 for all flight legs i (1)
k∈Q j∈P

btj xkj = ht
k∈Q j∈P
for all airports t at the end of the recovery period (2)
 Ch. 1. Air Transportation: Irregular Operations and Control 43

xkj = 1 for all aircraft k (3)
j∈P

xkj ∈ {0 1} for all routes j and aircraft k (4)

yi ∈ {0 1} for all flights i in F (5)
Constraints (1), together with constraints (4) and (5), require each flight to be
included in an assigned route or to be canceled. Constraints (2) ensure that at
the end of the day (that is, at the end of the recovery period), aircraft are repo-
sitioned so that the plan can be resumed at the start of the next day. Finally,
constraints (3) enforce the requirement that each aircraft be assigned to ex-
actly one route, commencing at its location at the start of the recovery period.
The objective is to minimize flight cancellation and delay costs.
A challenge in formulating this model is to estimate the objective function
costs. Because passengers and crews often travel on more than one aircraft
route, the costs of delays and cancellations cannot be expressed exactly as a
function of a single flight-leg, or as a function of a single aircraft routing. In-
stead, these costs depend on the pairs or subsets of flight legs comprising the
passenger and crew connections. Hence, approximate delay and cancellation
costs are used in the model.
The Arguello, Bard, and Yu model and heuristic solution approach is ap-
plied to a relatively small data set representing the Continental Airlines flight
schedule for Boeing 757 aircraft, with 42 flights, 16 aircraft, and 13 airport lo-
cations. They report that for over 90% of the instances tested, their approach
produces a solution within 10% of the lower bound within 10 CPU seconds.
Rosenberger et al. (2002) extend the Arguello, Bard, and Yu model to in-
clude supplementary slot constraints. Let A equal the set of allocated arrival
slots, Rk (a) equal the set of routes for aircraft k that include legs landing in
arrival slot a, and let |H(j a)| represent the number of flight legs in route j
using slot a. Then, the slot constraints are of the form:
  
H(j a)xk  αa for all a in A (6)
j
j∈Rk (a) k∈Q

These constraints ensure that the number of aircraft arriving in each allocated
time slot in the recovery period does not exceed the airport’s restricted capac-
ity, as mandated by ground delay programs, described in Section 4. Additional
work on recovering airline operations under conditions of insufficient airport
capacity are reported in Vasquez-Marquez (1991), Richetta and Odoni (1993),
Hoffman (1997), Luo and Yu (1997), Andreatta et al. (2000), Carlson (2000),
Chang et al. (2001), and Metron Inc. (2001).
The body of literature on aircraft recovery is growing as information tech-
nology capabilities expand. Selected additional references include Teodorovic
and Guberinic (1984), Teodorovic and Stojkovic (1990), Jarrah et al. (1993),
Teodorovic and Stojkovic (1995), Yu (1995), Mathaisel (1996), Rakshit et al.
(1996), Talluri (1996), Yan and Yang (1996), Yan and Young (1996), Cao and
44 M. Ball et al.

Kanafani (1997), Clarke (1997), Lettovsky (1997), Yan and Lin (1997), Yan
and Tu (1997), and Thengvall et al. (2000).

6.2 Expanded aircraft recovery

Bratu and Barnhart (2005) analyze the operations of a major US airline for
the months of July and August 2000, and report that:
(a) Flight delays are not indicative of the magnitude of delay experienced
by disrupted passengers. On the same day that disrupted passengers experi-
enced average delays of 419 minutes, the average delay of nondisrupted pas-
sengers was only 14 minutes, nearly matching the average flight delay that day.
(b) Disrupted passenger delays and associated costs are significant. Bratu
and Barnhart estimate for the airline they study that disrupted passengers rep-
resent just about 4% of passengers but account for more than 50% of the total
passenger delay. Associated with these disrupted passengers are direct and in-
direct costs, which can include lodging, meals, re-booking (possibly on other
airlines), and loss of passenger goodwill.
Bratu and Barnhart conclude that delay cost estimates that do not take
into consideration the costs of disruption cannot be accurate. Recognizing this,
Rosenberger et al. (2003) expand their aircraft recovery model to identify dis-
rupted crews and passengers, and their associated costs, by adding constraints
and variables to:
(i) compute the delay of each flight leg that is not canceled;
(ii) determine if a connection is disrupted; and
(iii) identify disrupted crews and passengers.
They then estimate delay costs, separately, for disrupted passengers and
crews, and for nondisrupted passengers and crews. These, in turn, are included
in the objective function of their extended model to achieve a more accurate
estimate of delay costs.
To solve their model, Rosenberger, Johnson, and Nemhauser limit the num-
ber of aircraft routes considered using an aircraft selection heuristic in which
routes are generated only for a selected subset of aircraft. They evaluate
their approach using a stochastic model (Rosenberger et al., 2002) to simulate
500 days of airline operations. Simulated disruptions include two-day unsched-
uled maintenance delays and severe weather disruptions at hub airports. They
compare the results of their extended model that accounts for crew and passen-
ger disruptions with those of the simplified model. They report that, compared
to the simplified model’s solutions, those generated with the extended model
exhibit significant reductions in passenger inconvenience and disruptions, at
the expense of on-time schedule performance degradation, increased overall
delay, and increased incidence of flight cancellation.
Bratu and Barnhart (2006) report similar findings. They also consider dis-
rupted passengers and crews, and develop an aircraft recovery model to de-
termine flight departure times and cancellations that minimize recovery costs,
 Ch. 1. Air Transportation: Irregular Operations and Control 45

including the costs of re-accommodating disrupted passengers and crews, re-

routing aircraft, and canceling flight legs. Unlike many of the more recent
models, their aircraft routing decision variables are flight-leg based, rather
than route-based. This reduces the number of decision variables significantly,
allowing them to generate recovery solutions for aircraft, crew, and passen-
gers simultaneously. To ensure the satisfaction of maintenance requirements,
they do not allow modification of routes for maintenance-critical aircraft, that
is, aircraft for which maintenance is scheduled that day. They apply their ap-
proach to problem instances containing 303 aircraft, 74 airport locations (3 of
which are hubs), 1088 flight legs per day on average, and 307,675 passenger
itineraries. They achieve solutions within 30 CPU seconds on a PC and report
expected reductions of more than 40% in the number of disrupted passengers,
more than 45% in the number of passengers required to overnight at a destina-
tion other than that planned, and more than 33% in the total delay minutes of
disrupted passengers. To achieve this, total delay minutes of nondisrupted pas-
sengers increased by 3.7% and the airline’s on-time performance, as measured
by the US DOT 15-minute on-time performance metric, worsened. This is an
expected result when one considers that intentionally delaying aircraft that oth-
erwise would be on-schedule can reduce passenger misconnections and hence,
reduce overall passenger delays.

6.3 Crew recovery

Although aircraft recovery decisions repair broken aircraft schedules, they

often result in the disruption of crew schedules. Flight cancellations, delays,
diversion, and swap decisions, together with crew illness, all result in the un-
availability of crews at the locations needed.
Crew recovery options include deadheading of crews (i.e., repositioning
crews by flying them as passengers) from their point of disruption to the lo-
cation of a later flight leg to which they are assigned. Once repositioned, the
crews can then resume their original work schedule. Another option is to as-
sign a reserve crew to cover the flight legs left unassigned by the crew disruption.
Reserve crews are back-up crews, not originally assigned to the flight sched-
ules, but pre-positioned at certain locations and available to report to duty, if
needed. They are guaranteed a minimum monthly salary, whether or not they
are called into work, and they are limited to a maximum number of flying hours
per month. In addition to possibly incurring additional reserve crew costs when
using reserve crews, airlines usually must also pay the replaced crew the entire
amount originally planned, even if the work was not performed. A third re-
covery option is to reassign a crew from its original schedule to an alternative
schedule. In this case, the new assignment must satisfy all collective bargaining
agreements and work rule regulations, including maximum crew work time,
minimum rest time, maximum flying time, maximum time-away-from-home,
etc. When reassigned, crews are typically paid the maximum of the pay asso-
46 M. Ball et al.

ciated with the original schedule or with the new schedule to which they are
assigned.
The crew recovery problem then is to construct new schedules for disrupted
and reserve crews to achieve coverage of all flights at minimum cost. Because
crew costs constitute a significant portion of airline operating costs, second
only to fuel costs, crew planning has garnered significant attention. Crew re-
covery, however, has received much less attention. One reason is that the
crew recovery problem is significantly more difficult. First, because of the time
horizon associated with recovery operations, recovery solutions must be gen-
erated quickly, in minutes instead of the hours or weeks allowed for planning
problems. Moreover, information pertaining to the location and recent flying
history of each crew member must be known at all times in order to gen-
erate recovery plans for the crew that satisfy the myriad of crew rules and
collective bargaining agreements. Finally, the objective function of the crew
recovery problem is multidimensional. Researchers often cast the crew recov-
ery objective as a blend of minimizing the incremental crew costs to operate
the modified schedule, while returning to the plan as quickly as possible and
minimizing the number of crew schedule changes made to do so. By limiting
the number of crews affected, the quality of the original crew plans will be pre-
served to the greatest extent possible. Moreover, returning to plan as quickly
as possible helps to avoid further downstream disruptions to aircraft, crew, and
passengers.
Due to these challenges, the crew recovery literature is relatively limited.
Although both cabin and cockpit crews are disrupted and must be recovered,
most recovery research focuses on cockpit crews, who are both more costly and
more constrained than cabin crews. Pilots have fewer recovery options because
they are qualified to fly only aircraft types with the same crew qualifications.
Yu et al. (2003) focus on cockpit crews and present a crew recovery model
and solution approach. They consider a set of aircraft types with the same crew
qualifications and a set of crews who (i) are qualified to fly these aircraft; and
(ii) are disrupted or are candidates who are likely to improve the crew recovery
solution through swaps. For each of these crews, they construct a set of feasi-
ble pairings, each beginning at the crew’s current location and commencing at
or later than the time at which the crew is available. Moreover, the generated
pairings satisfy all work rules and regulations, considering the amount of work
completed by the crew up to the point of disruption. Pairings selected in the
recovery solution satisfy cover constraints ensuring that each flight leg is either
canceled or assigned to one or more crews. When more than one crew is as-
signed, the additional crews are deadheaded and repositioned to their home
location or to another location where they can resume work. The objective
is to minimize the sum of (1) deadheading costs; (2) modified crew schedule
costs; and (3) cancellation costs due to leaving flight legs uncovered.
Yu et al. define the following sets and parameters:
e equipment type (consisting of one or more crew compatible aircraft
types)
 Ch. 1. Air Transportation: Irregular Operations and Control 47

I set of active flights to be covered by crews of equipment type e

K set of active and reserve crews available for equipment type e
Jk set of potential feasible pairings for crew k
cjk cost of assigning crew k to pairing j
ui cost of not covering flight leg i
qk cost of not assigning a pairing to crew k
di cost of each crew deadheading on flight i
aij equal to 1 if flight leg i is included in pairing j; and 0 otherwise.
The variables are:
xkj equal to 1 if crew k is assigned to pairing j; and 0 otherwise
zk equal to 1 if crew k has no pairing assigned; and 0 otherwise
yi equal to 1 if flight leg i is not covered (is canceled); and 0 otherwise
si equal to the number of crews deadheading on flight leg i.
They then formulate the crew recovery problem as
   
min cjk xkj + ui yi + qk zk + di si
k∈K j∈Jk i∈I k∈K i∈I
subject to

aij xkj + yi − si = 1 for all i ∈ I (7)
k∈Q j∈Jk

xkj + zk = 1 for all k ∈ K (8)
j∈Jk

xkj ∈ {0 1} for all k ∈ K all j ∈ Jk  (9)

yf ∈ {0 1} for all f ∈ I (10)
z k ∈ {0 1} for all k ∈ K (11)
sf ∈ {0 1 2   } for all f ∈ I (12)
Constraints (7) ensure that all flight legs are canceled or covered at least once,
with si representing the number of crews deadheading on flight leg i. Con-
straints (8) determine whether crew k is assigned to a pairing or must be
deadheaded to its crew base, that is, its domicile. Integrality of the solution
is guaranteed by constraints (9)–(12).
Yu et al. state that, for typical instances, there are millions of potential crew
pairings and hence, the size of the crew recovery model renders exact solution
approaches impractical. Using a procedure of Wei et al. (1997), they search
heuristically for solutions. They modify or generate a few pairings at a time
and test the quality of the solution, and then repeat the process if necessary.
Using data provided by Continental Airlines, they evaluate their heuristic ap-
proach on instances corresponding to disruptions affecting 1–40 flight legs of
the airline’s Boeing 737 fleet. Within 8 minutes at most, they generate near-
optimal solutions, achieving at most an average 5% optimality gap.
48 M. Ball et al.

Lettovsky (1997) and Lettovsky et al. (2000) present a similar model, but
they include additional constraints restricting the number of crews deadhead-
ing on each flight leg to the maximum available capacity. Moreover, their flight
cancellation costs include costs of re-assigning passengers to other flights, as-
sociated hotel, and meal costs, and estimates of the loss of passenger goodwill.
They design a heuristic solution approach for their model that keeps intact as
many as possible of the crew schedules, altering only those of disrupted crews
and of a few additional crews who greatly facilitate the recovery of crew oper-
ations. In restricting the set of crews for which new schedules are generated,
optimality of the original crew schedules is preserved for crews not affected by
modifications to the plan, and the size of the problem is contained, improv-
ing tractability and allowing quicker solution times. Lettovsky and Lettovsky,
Johnson and Nemhauser describe heuristics to select the crews whose sched-
ules might be altered.
Stojkovic et al. (1998) also address the operational crew scheduling prob-
lem, and present a set partitioning model and a branch-and-price algorithm
to determine modified monthly schedules for selected crew members. The ob-
jective is to cover all tasks at minimum cost while minimizing the number of
changes to the planned crew schedules. They generate test problems from pair-
ings of a US airline and report that quality solutions are obtained in reasonable
run times.

6.4 Passenger recovery

Just as aircraft recovery decisions result in crew disruptions, aircraft and

crew recovery decisions lead to passenger disruptions. The next step of
the recovery process then is to reassign disrupted passengers to alternative
itineraries, commencing at the disrupted passenger locations after their avail-
able times, and terminating at their destination, or a location nearby. Disrupted
passengers can be assigned to itineraries beginning at least some minimum
connection time after the time of their disruption. Only disrupted passen-
gers can be reassigned, and nondisrupted passengers cannot be displaced
by reassigned passengers. Clarke (2005) presents modeling strategies for re-
accommodating passengers who are disrupted by operations, or by schedule
changes resulting from considerations such as revenue management. Barnhart
et al. (2002) cast this problem as a multicommodity network flow problem.
They let xrp represent the number of disrupted passengers originally sched-
uled on itinerary p who are re-accommodated on itinerary r. In addition
to other itineraries offered by the airline, passengers can be accommodated
on itineraries offered by other airlines, or itineraries on different modes of
transportation. In fact, an alternate itinerary might be the null itinerary rep-
resenting canceled trips, a valid choice for passengers who are disrupted
before departing their origin. The planned arrival time at the destination of
itinerary p is l(p), and a(r) represents the actual arrival time at the destina-
tion of itinerary r. The set of flight legs is F; df is the number of seats available
 Ch. 1. Air Transportation: Irregular Operations and Control 49

for disrupted passengers, that is, the total number of seats less the number of
seats occupied by nondisrupted passengers, on flight f ; δrf equals 1 if flight f
is contained in itinerary r, and equals 0 otherwise; and np is the total number
of disrupted passengers of type p. The passenger re-assignment model is then
formulated as
   
min a(r) − l(p) xrp (13)
p∈P r∈R(p)k
subject to
 
δrf xrp  df for all f ∈ F (14)
p∈P r∈R(p)k

xrp = np for all p ∈ P (15)
r∈R(p)
xrp ∈ {0 1} for all p ∈ P all r ∈ R(p) (16)
The objective (13) is to find the disrupted passenger reassignments that mini-
mize the total delay experienced by disrupted passengers. Flight capacity con-
straints (14) ensure that only seats not occupied by nondisrupted passengers
are assigned to disrupted passengers. Constraints (15) and (16) ensure that
each disrupted passenger is reassigned, albeit perhaps to the null itinerary.
The passenger re-assignment model can be solved exactly, but its solution
time can become prohibitive for real-time operations as the number of dis-
rupted passengers grows, thus causing the need for column generation solu-
tion approaches to be employed. Bratu and Barnhart (2005), however, solve
this problem using a flexible heuristic, termed the Passenger Delay Calcula-
tor, that allows passenger recovery policies (such as frequent flyers first, or
first-disrupted-first-recovered) to be enforced. Using their Passenger Delay
Calculator, Bratu and Barnhart analyze two months of airline operations and
recovery data for a major airline, as well as numerous simulated scenarios.
They conclude that:
1. Connecting passengers are almost three times more likely to be disrupted
than passengers without connections. However, connecting passengers
who miss their connections are often re-accommodated on their best al-
ternative itineraries, that is, on itineraries that arrive at their destinations
at the earliest possible time, given the timing of the disruptions. In com-
parison, only about one-half of the passengers disrupted by flight leg
cancellations are re-accommodated on their best itineraries. This occurs
because the number of misconnecting passengers per flight leg is small
relative to the number of passengers disrupted by a flight leg cancella-
tion.
2. The inability to re-accommodate disrupted passengers on their best
itineraries is exacerbated by high load factors, with average delay for dis-
rupted passengers increasing exponentially with load factor.
50 M. Ball et al.

3. Alternative metrics measuring schedule performance, namely flight can-

cellation rates and the percentage of flights delayed by more than
45 minutes, are better indicators of passenger disruptions than the US
DOT 15-minute on-time performance metric.

7 Robust airline scheduling

Robust planning attempts to deal with data uncertainty in a planning model.

In airline schedule and resource allocation planning, there are two primary
sources of uncertainty: passenger demand and schedule execution. Here we
only deal with uncertainties in the execution of the planned schedule.
In the traditional stochastic programming approach to robust planning, it is
necessary to estimate the probability of each possible outcome. One then min-
imizes the expected cost of the planning decisions plus the cost of the recovery
that takes place as a result of the decision and outcome. Unfortunately, this
approach is completely intractable in the case of airline planning at present
because one has to deal with literally millions of very low probability events.
Moreover, there is no obvious way to aggregate meaningfully these events in a
way that would simplify the analysis. One could consider planning for only the
worst possible outcome, but this would be far too conservative and costly.
Nevertheless, the basic idea of including the anticipated costs of recovery
into the planning model can be very useful. A planner needs to think of an
optimal plan as being one for which the combined planned and recovery costs,
that is, the realized costs, are minimized. This definition of optimality is at
odds with the one typically employed by airline optimizers, who have histor-
ically excluded recovery costs and optimized only planned costs. In doing so,
resource utilization is maximized, with nonproductive time on the ground, i.e.,
slack time, minimized. Lack of slack, however, makes it difficult for disruption
to be absorbed in the schedule and limits the number of options for recovery.
One should not think of this omission as an oversight on the part of the air-
lines. Rather, it is based on recognition of the inherent difficulty of including
recovery costs in a planning model.
Enhancing schedule planning models to account for recovery costs presents
both modeling and computational challenges. A number of researchers have
begun to consider this challenge, recognizing that robust planning is a problem
rich in opportunity and potential impact. To facilitate the generation of robust
plans, they have developed various proxies of robustness, mainly focused on
finding flexible plans that provide a rich set of recovery options for passengers,
crews, and aircraft; or plans that isolate the effects of disruptions, requiring
only localized plan adjustments.
In the following subsections, we highlight selected work on robust airline
planning, while outlining briefly the various modeling and algorithmic ap-
proaches employed.
 Ch. 1. Air Transportation: Irregular Operations and Control 51

7.1 Robust schedule design

In this section, we describe some recent work that represents a step towards
achieving flight schedule designs that are resilient when it comes to passenger
disruption. It extends the cost minimization approaches described in Simpson
(1966), Chan (1972), Soumis et al. (1980), Etschmaier and Mathaisel (1985),
Berge (1994), Marsten et al. (1996), Erdmann et al. (1999), Armacost et al.
(2002), and Lohatepanont and Barnhart (2004).
Lan et al. (2005) develop a new approach to minimize the number of pas-
senger misconnections by re-timing flight departures, while keeping all fleet-
ing and routing decisions fixed. Moving flight leg departure times provides
an opportunity to re-allocate slack time to reduce passenger disruptions and
maintain aircraft productivity. Levin (1971) proposed the idea of adding time
windows to fleet routing and scheduling models. Related research can be found
in Desaulniers et al. (1997), Klabjan et al. (2002), Rexing et al. (2000), and
Stojkovic et al. (2002).
To illustrate the idea, consider the example in Figure 8. In the planned
first-in-first-out aircraft routing, flight leg f 1 is followed by leg f 2 in one air-
craft’s rotation, and leg f 3 is followed by leg f 4 in another aircraft’s rotation.
Assume that f 1 is typically delayed, as indicated in Figure 8. Because insuffi-
cient turn time results from the delay, some of the delay to f 1 will propagate
downstream to f 2, as shown in Figure 7. Then, assuming that f 3 is typically
on-schedule, expected delays are reduced by changing the planned aircraft ro-
tations to f 1 followed by f 4, and f 3 followed by f 2, as shown in Figure 8.
Lan, Clarke, and Barnhart apply their re-timing model to the flight schedule
of a major US airline and compare passenger delays and disruptions in the
original schedules with those expected from the solutions to their re-timing
model. They find that a 30-minute time window, allowing each flight leg to
depart at most 15 minutes earlier or later than in the original schedule, can
result in an expected reduction in passenger delay of 20% and a reduction in
the number of passenger misconnections of about 40%; a twenty-minute time
window can reduce passenger delays by about 16% and reduce the number of
passenger misconnections by over 30%; and finally, a ten-minute time window
can reduce passenger delays by roughly 10% and passenger misconnections
by 20%.

7.2 Robust fleet assignment

Rosenberger et al. (2004) present a robust fleet assignment approach,

building on the work reported in Barnhart et al. (1998a). They identify hub
inter-connectivity as an important indicator of schedule robustness. Because
schedules are sensitive to disruptions at hubs, a more robust schedule is one in
which hubs are isolated to the greatest extent possible. They quantify the de-
gree to which a hub is isolated using a hub connectivity metric; the smaller the
value of hub connectivity, the more isolated the hub.
52 M. Ball et al.

(a)

(b)

(c)
Fig. 8. (a) First-in-first-out routings and delayed flight leg f 1. (b) Delay propagation due to delay of
flight leg f 1. (c) Revised routings minimizing delay propagation.

Rosenberger, Johnson, and Nemhauser characterize a robust fleet assign-

ment as one with limited total hub connectivity and many short cycles (cycles
with a small number of flight legs). Short cycles allow an airline to limit the
number of flights canceled when a cancellation is necessary, thereby lessening
the impact of disruptions and facilitating recovery. They let J denote the set
of fleets and S be the set of strings, or sequences of flight legs beginning at a
hub, ending at a hub, and flown by the same aircraft. The hub connectivity met-
ric hjs for string s ∈ S and aircraft of type j ∈ J equals the number of legs in s
if s begins and ends at different hubs, and equals 0 otherwise. The maximum
value of hub connectivity is specified by a threshold value ς. For each fleet type
j ∈ J and string s ∈ S, cjs is the cost of flying s with aircraft of type j and
decision variable xjs equals 1 if j is assigned to s, and 0 otherwise. The set of
feasible fleet assignment solutions is χ. Their model to determine robust fleet
 Ch. 1. Air Transportation: Irregular Operations and Control 53

assignments with hub isolation and short cycles is then


min cjs xjs
s∈S j∈J

subject to hjs xjs  ς
s∈S j∈J
x ∈ χ
The objective is to find the minimum cost fleet assignment with total value
of hub connectivity not greater than ς. They propose a related model in which
the objective is to maximize hub isolation and limit total fleeting costs to some
pre-specified threshold. Using SIMAIR, they compare the solutions to their
robust fleet assignment models with those obtained solving a traditional FAM
model. They report that with small increases in planned costs it is possible to
reduce cancellations significantly and also improve on-time performance.
Smith (2004) focuses on the revenue aspects of robustness in fleet assign-
ment. He adds purity to fleet assignment solutions, which means that he limits
the number of fleet types at spokes to at most one or two. This, of course, in-
creases planned cost, but purity adds robustness in operations by enhancing the
possibility of crew swaps. It also decreases maintenance cost because the need
for spare parts is reduced. Interestingly, adding the upper bound constraints
on fleet types at spokes makes the FAM model much harder to solve. Smith
introduces a station decomposition approach to solve this more difficult FAM
model.

7.3 Robust aircraft routing

Maintenance routing, as surveyed in Klabjan (2003), provides an attractive

opportunity for adding robustness because modifying routes has a minimal im-
pact on planned cost. Therefore it is not necessary to make an explicit tradeoff
between planned cost and robustness.

7.3.1 Degradable airline scheduling

Kang and Clarke (2002) attempt to achieve robustness by isolating the ef-
fects of disruptions. They partition the legs of the flight schedule into inde-
pendent subnetworks, which are determined through alternative models and
approaches, each applicable at a particular step of schedule planning, such
as schedule design, fleet assignment, or aircraft maintenance routing. The
model solutions are constrained to ensure that aircraft (and ultimately, crew)
are assigned only to flight legs within a single subnetwork, prohibiting them
to operate between subnetworks. (Passengers, on the other hand, can travel
within multiple subnetworks.) The subnetworks are prioritized based on the
total revenue of the flights legs they contain, with the maximum-revenue sub-
network having the highest priority. When disruptions occur, the top priority
subnetworks are recovered first, shielding the associated crew, aircraft, and
54 M. Ball et al.

passengers to the greatest extent possible from the resulting delays. This has
the effect of relegating disruptions to the low priority subnetworks, and mini-
mizing the revenue associated with delayed and disrupted passengers.
An advantage of the approach of Kang and Clarke is that it can simplify
recovery. Because delays and propagation effects are contained within a single
subnetwork, the recovery process needs only to take corrective action on the
flights in the affected subnetwork, and not on the entire airline network.

7.3.2 Robust aircraft routing: Allocating slack to minimize delays

Lan et al. (2005) propose aircraft routing models and algorithms aimed at
minimizing delay propagation and passenger delay and disruption. They par-
tition flight leg delays into two categories, namely: propagated delay, that is,
delay occurring when the aircraft to be used for a flight leg is delayed on its
preceding flight leg; and nonpropagated or independent, delay. Propagated de-
lay is a function of an aircraft’s routing, while nonpropagated delay is not. The
premise underlying their approach is that propagated delay can be reduced
through intelligent aircraft routing, that is, by allocating slack optimally to ab-
sorb delay propagation. By reducing propagated delay, they expect to achieve
a corresponding reduction in passenger delays.
Using historical delay data, they first estimate the expected independent de-
lay for each flight leg, and then, using these estimates, compute the expected
propagated delay for each possible aircraft routing by sequentially computing
the earliest departure time for each subsequent flight leg in the routing, given
the expected independent delay for that flight leg plus the resulting propa-
gated delay accumulated to that point. Next, for each fleet type, they solve a
daily model (that is, one that assumes the flight schedule repeats daily) to select
aircraft routes that satisfy maintenance requirements while minimizing prop-
agated delay. In their model, they let S be the set of feasible strings, where a
string is a sequence of connected flight legs (that is, the departure station for
flight leg f is the same as the arrival station of f ’s predecessor and the de-
parture time of flight leg f is not earlier than the arrival time plus minimum
turn time of flight leg f ’s predecessor) beginning and ending at a maintenance
station and with elapsed time not greater than the maximum time between
maintenance checks. The set of daily flight legs is F, F + is the set of flight legs
originating at a maintenance station, and F − is the set of flight legs terminat-
ing at a maintenance station. The set of ground arcs (including the overnight
or wrap-around arcs to ensure that the flight schedule can repeat daily) is de-
noted by G. The set of strings ending with flight leg i is Si− , and the set of
strings beginning with flight leg i is Si+ . They include one binary decision vari-
able xs for each feasible string s, and ground variables y to count the number
of aircraft on the ground at maintenance stations. The delay propagated from
flight leg i to flight leg j if flight leg i and flight leg j are in string s is pdijs . If
flight leg i is in string s, ais equals 1, otherwise it equals 0. Ground variable
−
yid equals the number of aircraft on the ground before flight leg i departs, and
 Ch. 1. Air Transportation: Irregular Operations and Control 55

+
ground variable yid equals the number of aircraft on the ground after flight leg
−
i departs. Similarly, ground variable yia equals the number of aircraft on the
+
ground before flight leg i arrives, and ground variable yia equals the number
of aircraft on the ground after flight leg i arrives. The count time is a point in
time when aircraft are counted. The number of times string s crosses the count
time is rs ; pg is the number of times ground arc g crosses the count time; and
N is the number of planes available.
The model to determine robust aircraft routes is
   
min E pdijs xs = E xs pdijs
s∈S (ij)∈sk s∈S (ij)∈sk
 
= min xs E pdijs (17)
s∈S (ij)∈sk
subject to

ais xs = 1 for all i ∈ F (18)
s∈S

− +
xs − yid + yid =0 for all i ∈ F +  (19)
s∈Si+

− +
xs − yia + yia =0 for all i ∈ F −  (20)
s∈Si−
 
rs xs + pg yg  0 (21)
s∈S g∈G

yg  0 for all g ∈ G−  (22)

−
xs ∈ {0 1} for all s ∈ S  (23)
The objective (17) is to select strings that minimize the expected total prop-
agated delay. Constraints (18) ensure that each flight leg is contained in exactly
one string, while constraints (19) and (20) guarantee that the number of air-
craft arriving at a location equal the number departing. Constraint (21) ensures
that the total number of aircraft in the solution does not exceed the number
available. Constraints (22) and (23) guarantee a nonnegative number of air-
craft on the ground at all times, and ensure that the number of aircraft assigned
to a string is either 0 or 1, respectively. Because each ground variable can be
expressed as a sum of binary string variables, the integrality constraints on the
ground variables can be relaxed (Hane et al., 1995).
Their solution approach applies the branch-and-price algorithm (Barnhart
et al., 1998b), with column generation to enumerate a relevant subset of string
variables. They apply their approach to four different networks, each corre-
sponding to a different fleet type operated by a major US network carrier.
They compare their robust routing solution with the routing solution generated
by the airline and estimate that their solution can yield average reductions of
56 M. Ball et al.

11% in the number of disrupted passengers, and 44% in total expected propa-
gated delay minutes. They further report that their solution corresponds to an
expected improvement of 1.6% in the airline’s Department of Transportation
(DOT) on-time arrival rate. This is significant because a 1.6% improvement
would allow the airline to improve its position in the DOT’s airline on-time
rankings, which are publicly available and are often cited as an important indi-
cator of airline performance.

7.3.3 Robust routing through swap opportunities

Ageeva and Clarke (2004) use the constraints of the string-based routing
model of Lan et al. (2005) but change the objective function to optimize a
different robustness criterion. They attempt to build flexible aircraft routings
with maximal potential for modification during recovery by adding a reward for
each opportunity to swap aircraft. Their objective is to maximize the number
of swap opportunities in the routing solution. Aircraft swapping is possible
when the routings of two aircraft intersect at least twice. To understand how
swaps can mitigate the impact of a delayed or unavailable aircraft, consider
an example in which aircraft a1 is scheduled to depart station s at time t1
but is delayed until time t2. Further, assume that aircraft a2 is scheduled to
depart s at time t3 (with t3 > t2 > t1) but is available for departure at t1.
Without swapping, the flight legs assigned to aircraft a1 experience delays as
great as t2 − t1, while the flight legs assigned to aircraft a2 are not delayed.
With swapping, however, none of the flights legs originally assigned to a1 or a2
is delayed.
Ageeva and Clarke measure the robustness of their solutions using an oppor-
tunity index, defined as the ratio of the number of actual to potential intersect-
ing partial rotations. They report that using their approach, optimal costs are
maintained and robustness of the aircraft routing solution, as measured by the
opportunity index, are improved up to 35% compared to solutions generated
by a basic routing model devoid of robustness considerations.

7.4 Robust crew scheduling

The crew pairing problem, with a focus on minimizing planned crew-related

cost, has been studied by many researchers. Survey papers on the subject
include Yu (1997), Desaulniers et al. (1998), Clarke and Smith (2000), and
Barnhart et al. (2003b). The focus of the more recent body of research on
robust crew scheduling has instead been on minimizing realized cost. The un-
derlying motivation stems from the observation that the realized cost associ-
ated with a crew pairing solution often differs significantly in practice from the
planned cost. Large additional crew costs are incurred, for example, when re-
serve crews are called in to complete work assigned to disrupted crewmembers
no longer able to perform their originally assigned work. The causes might
include crewmembers in the wrong location due to one or more flight leg can-
cellations, or crewmembers reaching the limit on the maximum allowable duty
 Ch. 1. Air Transportation: Irregular Operations and Control 57

time before completing their work due to flight delays. The resulting cost in-
creases can be of the order of many millions of dollars for a large airline. To
address this issue, researchers, such as Ehrgott and Ryan (2002), Schaefer et al.
(2005), Yen and Birge (2000), and Chebalov and Klabjan (2002) have devel-
oped approaches to minimize the sum of planned and unplanned crew cost.

7.4.1 Robust crew pairing: The role of crew connections between

different aircraft
Ehrgott and Ryan (2002) propose an approach to balance costs and robust-
ness in generating crew pairing solutions. For each pairing, they compute its
value of nonrobustness by approximating downstream effects of delays within
the pairing. In their approximation, the value of nonrobustness is zero if crews
do not change planes, but equals the potential disruptive effects of delays if
the plan requires crews to connect between different aircraft. The objective is
to minimize the value of nonrobustness, while maintaining the cost of the cor-
responding crew pairing solution to less than that of the minimum-cost crew
pairing solution plus some pre-specified positive value.
Ehrgott and Ryan report that small increases in cost allow considerable ro-
bustness gains. In one instance, by increasing costs by less than 1%, they were
able to reduce their metric of “nonrobustness” by more than 2 orders of mag-
nitude. Their more robust solutions are characterized by longer ground times
between successive flights on different aircraft within a pairing; fewer aircraft
changes within pairings; slightly longer duty times; and a slight increase in the
number of pairings in the solution.

7.4.2 Minimizing expected crew costs

Yen and Birge (2000) and Schaefer et al. (2005) develop approaches that in-
clude both planned and unplanned costs in the objective function of the crew
model. Yen and Birge develop a stochastic crew scheduling model and corre-
sponding solution approach, while Schaefer et al. solve a deterministic crew
pairing problem with an objective to minimize expected crew pairing costs. In
Schaefer et al., expected costs are approximated for each pairing under the
assumptions that there are no interactions between the pairings and recovery
is achieved simply by delaying the next flight in the pairing until the crew is
available (pushback). Pushback recovery helps to justify the no interactions
assumption, but much more sophisticated recovery procedures are used in
practice at hubs. Even with these simplifying assumptions, it is still not possible
to calculate the realized cost of pairings analytically. Thus the cost of each pair-
ing considered in the optimization is determined by Monte Carlo simulation.
Schaefer et al. denote J as the set of feasible pairings, F the set of flight legs,
c̃j the expected cost of pairing j, and let aij equal 1 if flight leg i is covered by
pairing j. The decision variable xj , for all j ∈ P, equals 1 is pairing j is included
in the solution, and equals 0 otherwise.
58 M. Ball et al.

The robust crew model of Schaefer et al. is


min c̃j xj (24)
j∈J
subject to

aij xj = 1 for all i ∈ F −  (25)
j∈J

xj ∈ {0 1} for all j ∈ P −  (26)

The objective (24) is to minimize the expected crew costs associated with the
pairings in the solution. The set of selected pairings must contain each flight
leg exactly once (25).
Crew schedules obtained with these expected pairing costs are compared
with those obtained by using the standard deterministic costs and also a set of
penalty costs whereby attributes of pairings that might lead to poor operational
performance are penalized. The attributes considered are sit times between
flights when a crew changes planes, flying and elapsed times of duties, and rest
time between duties. The operational performance of crew schedules are eval-
uated using SIMAIR, with only mild disruptions considered, that is, individual
delays rather than major disruptions such as those which reduce airport capac-
ity. The standard cost measure of FTC, which is total cost in minutes of pay
minus minutes of flying time divided by flying time, is used to compare sched-
ules. For the fleets considered from a major airline, the planned FTCs were
typically in the 2–4 % range and increased only slightly when either the ex-
pected costs or penalty costs were used. The operational FTCs ranged from 4%
to 9% and were lowest for the expected cost method and very close to a lower
bound. However, in an absolute sense the expected cost solutions performed
only marginally better than the deterministic solutions perhaps because of the
assumption of mild disruptions only. Note that more severe disruptions would
have invalidated the pushback assumption.

7.4.3 Move-up crews

In an approach analogous to the idea in Ageeva and Clarke (2004) of provid-
ing flexibility through aircraft swapping, Chebalov and Klabjan (2002) develop
the concept of enhancing recovery flexibility through move-up crews. A move-
up crew for flight i is a crew, not actually assigned to i, but capable of being
assigned to i, if necessary. For feasibility of this potential assignment, the
move-up crew must have the same crew base, or domicile, as the crew cur-
rently assigned to i, must be ready to operate i before the departure time of i,
and must end the pairing on the same day as the pairing currently covering i.
Chebalov and Klabjan consider move-up crews only for flights i that depart hub
locations and do not begin a pairing. The objective is to maximize the number
of move-up crews so that recovery is more likely to be effected by swapping the
assigned pairings of the delayed crew and an available, alternative crew.
 Ch. 1. Air Transportation: Irregular Operations and Control 59

Let J represent the set of feasible pairings; F be the set of flight legs;
aij equal 1 if flight leg i is covered by pairing j; HL designate the set of hub
locations; CB be the set of crew base locations; D be the set of the possible
number of days remaining in a pairing; Jcbd represent the set of pairings start-
ing at the crew base cb with d days remaining after flight leg i to the end of the
pairing; Ji be the set of pairings whose first leg is i; Jicbd be the set of pairings
that yields a move-up crew for flight i covered by a crew originating at cb with
d days remaining from flight leg i to the end of the pairing; r be the robustness
factor denoting the maximum allowable percentage increase in the cost of the
solution to allow increased robustness; and M be an arbitrary number (usually
2 or 3). Chebalov and Klabjan solve the standard crew pairing problem mini-
mizing operating costs with constraints (25) and (26) to obtain the minimum
planned crew pairing costs, copt . They then include in their model four sets of
decision variables. If pairing j is included in the solution, xj , for all j ∈ P, is
equal to 1, otherwise it equals 0. If flight leg i is covered by a pairing starting
at the crew base cb with d days remaining after flight leg i to the end of the
cb is equal to 1, otherwise it equals 0. If flight leg i is covered by a
pairing, yid
pairing whose first leg is i, wi is equal to 1, otherwise it equals 0. The number
of move-up crews for flight leg i, if i is a leg originating at a hub h ∈ HL, is
cb .
denoted by zid
The Chebalov and Klabjan model is
  
cb
max zid (27)
i∈F cb∈CB d∈D
subject to

cb
aij xj = yid for all i ∈ F originating at any hub (28)
j∈Jcbd

aij xj = 1 for all i ∈ F originating at any spoke (29)
j∈J

xj = wi for all i ∈ F originating at any crew base (30)
j∈Ji
 
cb
wi + yid =1
cb∈CB d∈D
for all i ∈ F originating at any crew base (31)
 
cb
yid =1
cb∈CB d∈D
for all i ∈ F originating at a hub but not a crew base (32)

cb
xj  zid for all i ∈ F originating at any hub (33)
j∈Jicbd
cb cb
zid  Myid for all i ∈ F originating at any hub (34)
60 M. Ball et al.

cj xj  (1 + r)copt  (35)
j
xj ∈ {0 1} for all j ∈ J (36)
wi ∈ {0 1} for all i ∈ F (37)
cb
yid ∈ {0 1} for all i ∈ F d ∈ D cb ∈ CB (38)
cb
zid ∈ {0 1     M} for all i ∈ F d ∈ D cb ∈ CB (39)
Constraints (29) and (36) ensure that each flight leg i is covered by exactly
one pairing. With constraints (37)–(39), constraints (30) identify flight legs that
are the first leg in a pairing and constraints (28), (31), and (32) identify flight
legs that are candidates for move-up crews. The number of move-up crews for
flight leg i is bounded by 0 (constraints (33)) if i originates at a spoke location
or is the first leg in a pairing, and otherwise it is bounded by the minimum
of M (usually 2 or 3), and the number of eligible move-up crews for i (con-
straints (34)). Constraint (35) ensures that the pairing solution has cost no
greater than an allowable tolerance above the minimum cost pairing. The ob-
jective (27) is to maximize the total number of move-up crews for all flight
legs.
Chebalov and Klabjan present a Lagrangian decomposition method to solve
their model. They perform computational experiments and report, that for cer-
tain instances, there are crew solutions characterized by only slightly higher
planned costs and by a 5- to 10-fold increase in the number of move-up crews,
compared to the optimal solutions to the more conventional, cost-minimizing
crew pairing problem.

8 Conclusions

Flight and crew schedules and passenger itineraries have become increas-
ingly “fragile” as a result of the growing complexity of the air transportation
system and the tight coupling of its various elements. The resulting direct and
indirect economic costs are very large, certainly amounting to several billion
dollars annually. The airline industry has a vital stake in research aimed at
mitigating the effects of severe weather and other disruptive events and at ex-
pediting recovery from “irregular” operations.
As this chapter has indicated, a very significant body of recent and ongo-
ing work has led to major progress toward these objectives. Two breakthrough
developments have been the primary drivers behind this progress. First, Col-
laborative decision making has made it possible to apply the principles of
information sharing and distributed decision making to ATFM, by expand-
ing the databases available to airline and FAA (and, soon, European) traf-
fic flow managers, creating common situational awareness, and introducing
shared real-time tools and procedures. And second, there is growing recog-
nition in the airline industry of the fact that planning for schedule robustness
 Ch. 1. Air Transportation: Irregular Operations and Control 61

and reliability may be just as important as planning for minimizing costs in

the complex, highly stochastic and dynamic environment of air transportation.
Specific achievements that have been described herein include: improved un-
derstanding and better modeling of the physics of airport and airspace capacity
and delays (Section 2); realization of the need for market-based mechanisms
to supplement widely-used administrative methods for allocating scarce airport
capacity among prospective airport users (Section 3); development of models
and optimization tools to support GDP decision-making under a wide range of
conditions, including the presence of uncertainty regarding forecast capacity
and demand (Section 4); development and implementation of airline-based
models for efficient “recovery” of aircraft, crews, and passengers following
schedule disruptions (Section 6); and the nascent appearance of increasingly
viable models for introducing robustness in airline route design and in the
scheduling of aircraft, crews, and passengers (Section 7).
At the same time, it is fair to describe all this work as still being in its early
stages in many respects – an assessment that applies equally well to the do-
mains of both the airlines and the providers of air traffic management services.
For example, in the case of the latter, approaches for dealing with uncertainty –
an altogether critical issue in the ATM and ATFM context – are still quite far
removed from being applied in practice. Integrated consideration and opti-
mization of both arrival and departure schedules at GDP airports could also
offer significant improvements over the existing approaches that focus solely
on arrivals. Research on collaborative routing is still in its infancy. On the
side of the airlines, decision support software for recovery is perhaps at the
stage where planning software was 15 years ago. While research is active and
hardware and data support have improved substantially, optimization-based
decision support tools for rapid recovery are still at an early stage of imple-
mentation at the major airlines. Finally, and most important, a full integration
of the emerging ATFM CDM philosophy and associated models with airline
recovery planning models and robust scheduling models has not even begun.
This represents a difficult, but crucial future research challenge.

Acknowledgements

The work of the first and fourth authors was supported in part by NEXTOR,
the National Center for Excellence in Aviation Operations Research, under
Federal Aviation Administration cooperative agreement number 01CUMD1.
The work of the second and fourth authors was supported in part by the Alfred
P. Sloan Foundation as part of the MIT Global Airline Industry Program.

References

Ageeva, Y., Clarke, J.P. (2004). Incorporating robustness into airline scheduling. Working paper, De-
partment of Aeronautics and Astronautics, Massachusetts Institute of Technology, Cambridge, MA.
62 M. Ball et al.

Andreatta, G., Odoni, A.R. (2003). Analysis of market-based demand management strategies. In:
Ciriani, T., Fasano, G., Gliozzi, S., Tadei, R. (Eds.), Operations Research in Space and Air. Kluwer
Academic, pp. 257–278.
Andreatta, G., Brunetta, L., Guastalla, G. (2000). From ground holding to free flight: An exact ap-
proach. Transportation Science 34, 394–401.
Arguello, M.F., Bard, J.F., Yu, G. (1997). Models and methods for managing airline irregular opera-
tions. In: Yu, G. (Ed.), Operations Research in the Airline Industry. Kluwer Academic, Boston, MA,
pp. 1–45.
Armacost, A., Barnhart, C., Ware, K. (2002). Composite variable formulations for express shipment
service network design. Transportation Science 36, 1–20.
ATAC Corporation (2003). Airport and airspace simulation model. Available at http://www.atac.com/
prodsvs/simmod.htm.
Balinski, M.L., Shihidi, N. (1998). A simple approach to the product rate variation problem via ax-
iomatics. Operations Research Letters 22, 129–135.
Ball, M.O., Lulli, G. (2004). Ground delay programs: Optimizing over the included flight set based on
distance. Air Traffic Control Quarterly 12, 1–25.
Ball, M., Dahl, R., Stone, L., Thompson, T. (1993). OPTIFLOW build-I design document, version 1.5.
Optiflow project report to the FAA, December.
Ball, M.O., Chen, C., Hoffman, R., Vossen, T. (2001a). Collaborative decision making in air traffic man-
agement: Current and future research directions. In: Bianco, L., Dell’Olmo, P., Odoni, A. (Eds.),
New Concepts and Methods in Air Traffic Management. Springer-Verlag, Berlin, pp. 17–30. A preprint
appeared in Proceedings of ATM’99, Workshop on Advanced Technologies and Their Impact on Air
Traffic Management in the 21st Century, 1999.
Ball, M.O., Hoffman, R., Knorr, D., Wetherly, J., Wambsganss, M. (2001b). Assessing the benefits of
collaborative decision making in air traffic management. In: Donohue, G., Zellweger, A. (Eds.),
Air Transportation Systems Engineering. American Institute of Aeronautics and Astronautics, Inc.,
Reston, VA, pp. 239–250. A preprint appeared in Proceedings of 3rd USA/Europe air Traffic Manage-
ment R&D Seminar, 2000.
Ball, M.O., Vossen, T., Hoffman, R. (2001c). Analysis of demand uncertainty in ground delay programs.
In: Proceedings of 4th USA/Europe Air Traffic Management R&D Seminar, Santa Fe, NM.
Ball, M.O., Hoffman, R., Odoni, A., Rifkin, R. (2003). Efficient solution of a stochastic ground holding
problem. Operations Research 51, 167–171.
Ball, M.O., Donohue, G., Hoffman, K. (2005). Auctions for the safe, efficient and equitable alloca-
tion of airspace system resources. In: Cramton, P., Shoham, Y., Steinberg, R. (Eds.), Combinatorial
Auctions. MIT Press, Cambridge, pp. 507–538.
Barnhart, C., Boland, N., Clarke, L., Johnson, E., Nemhauser, G., Shenoi, R. (1998a). Flight string
models for aircraft fleeting and routing. Transportation Science 32, 208–220.
Barnhart, C., Johnson, E., Nemhauser, G., Savelsbergh, M., Vance, P. (1998b). Branch-and-price: Col-
umn generation for solving huge integer programs. Operations Research 46, 316–329.
Barnhart, C., Kniker, T., Lohatepanont, M. (2002). Itinerary-based airline fleet assignment. Transporta-
tion Science 36, 199–217.
Barnhart, C., Belobaba, P., Odoni, A.R. (2003a). Applications of operations research in the air transport
industry. Transportation Science 37, 368–391.
Barnhart, C., Cohn, A.M., Johnson, E.L., Klabjan, D., Nemhauser, G.L., Vance, P.H. (2003b). Airline
crew scheduling. In: Hall, R.W. (Ed.), Handbook of Transportation Science, 2nd edition. Kluwer
Academic, Norwell, MA, pp. 517–560.
Beasley, J., Sonander, J., Havelock, P. (2001). Scheduling aircraft landings at London Heathrow using
a population heuristic. Journal of the Operational Research Society 52, 483–493.
Beatty, R., Hsu, R., Berry, L., Rome J. (1998). Preliminary evaluation of flight delay propagation
through an airline schedule. In: Proceedings of the Second USA/Europe Air Traffic Management R&D
Seminar, Orlando, Fl.
Berge, M. (1994). Timetable optimization: Formulation, solution approaches, and computational is-
sues. In: AGIFORS Proceedings, pp. 341–357.
 Ch. 1. Air Transportation: Irregular Operations and Control 63

Berge, M., Hopperstad, C., Heraldsdottir, A. (2003). Airline schedule recovery in collaborative flow
management with airport and airspace capacity constraints. In: Proceedings of the Fifth USA/Europe
Air Traffic Management R&D Seminar, Budapest, Hungary.
Bertsimas, D., Stock Paterson, S. (1998). The air traffic flow management problem with en route capac-
ities. Operations Research 46, 406–422.
Bertsimas, D., Stock Paterson, S. (2000). The traffic flow management rerouting problem in air traffic
control: A dynamic network flow approach. Transportation Science 34, 239–255.
Bhogadi, N. (2002). Modeling demand uncertainties during ground delay programs. MSSE thesis, Uni-
versity of Maryland.
Bilimoria, K., Sridhar, B., Chatterji, G., Sheth, K., Grabbe, S. (2000). FACET: Future ATM concepts
evaluation tool. In: Proceedings of 3rd USA/Europe Air Traffic Management R&D Seminar, Napoli,
Italy.
Blumstein, A. (1959). The landing capacity of a runway. Operations Research 7, 752–763.
Bratu, S. (2003). Airline customer service network design and operations. PhD thesis, Department
of Aeronautics and Astronautics and the Center for Transportation and Logistics, Massachusetts
Institute of Technology, Cambridge, MA.
Bratu, S., Barnhart, C. (2005). An analysis of passenger delays using flight operations and passenger
booking data. Air Traffic Control Quarterly 13, 1–27.
Bratu, S., Barnhart, C. (2006). Flight operations recovery: New approaches considering passenger re-
covery. Journal of Scheduling 9 (3), 279–298.
Bruckner, J.K. (2002). Airport congestion when carriers have market power. American Economic Re-
view 92, 1357–1375.
Cao, J., Kanafani, A. (1997). Real-time decision support for integration of airline flight cancellations
and delays, Part I: Mathematical formulation. Transportation Planning and Technology 20, 183–199.
Carlin, A., Park, R. (1970). Marginal cost pricing of airport runway capacity. American Economic Re-
view 60, 310–318.
Carlson, P.M. (2000). Exploiting the opportunities of collaborative decision making: A model and effi-
cient solution algorithm for airline use. Transportation Science 34, 381–393.
Chan, Y. (1972). Route network improvement in air transportation schedule planning. Flight Trans-
portation Laboratory Report R72-3, Massachusetts Institute of Technology, Cambridge, MA.
Chang, K., Howard, K., Oiesen, R., Shisler, L., Tanino, M., Wambsganss, M.C. (2001). Enhancements
to the FAA ground-delay program under collaborative decision making. Interfaces 3, 57–76.
Chebalov, S., Klabjan, D. (2002). Robust airline crew scheduling: Move-up crews. In: Proceedings of the,
2002 NSF Design, Service, and Manufacturing Grantees and Research Conference.
Clarke, J.-P., Melconian, T., Bly, E., Rabbani, F. (2004). MEANS – The MIT extensible air network
simulation. Simulation: Transactions of the Society International for Computer Simulation, in press.
Clarke, M. (1997). Development of heuristic procedures for flight rescheduling in the aftermath of
irregular airline operations. ScD dissertation, Massachusetts Institute of Technology, Cambridge
MA.
Clarke, M. (2005). Passenger reaccommodation a higher level of customer service. Presented at AGI-
FORS Airline Operations Study Group Meeting, Mainz, Germany.
Clarke, M., Smith, B. (2000). The impact of operations research on the evolution of the airline industry:
A review of the airline planning process. Research paper, Sabre Inc., Dallas, TX.
Clarke, M., Lettovsky, L., Smith, B. (2000). The development of the airline operations control cen-
ter. In: Butler, G., Keller, M.R. (Eds.), Handbook of Airline Operations. Aviation Week Group of
McGraw-Hill, Washington, DC.
Daniel, J. (1995). Congestion pricing and capacity at large hub airports: A bottleneck model with sto-
chastic queues. Econometrica 63, 327–370.
de Neufville, R., Odoni, A. (2003). Airport Systems: Planning, Design and Management. McGraw-Hill,
New York.
Dear, R.G., Sherif, Y.S. (1991). An algorithm for computer assisted sequencing and scheduling of ter-
minal area operations. Transportation Research A 25, 129–139.
Desaulniers, G., Desrosiers, J., Solomon, M.M., Soumis, F. (1997). Daily aircraft routing and schedul-
ing. Management Science 43, 841–855.
64 M. Ball et al.

Desaulniers, G., Desrosiers, J., Gamache, M., Soumis, F. (1998). Crew scheduling in air transportation.
In: Crainic, T., Laporte, G. (Eds.), Fleet Management and Logistics. Kluwer Academic, pp. 169–185.
Dobbyn, T. (2000). US prepares plan to cut summer flight delays. Reuters Wire, March 3.
DotEcon Ltd. (2001). Auctioning airport slots. Report for HM Treasury and the Department of the
Environment, Transport and the Regions, United Kingdom. Available at http://www.dotecon.com.
Ehrgott, M., Ryan, D.M. (2002). Constructing robust crew schedules with bicriteria optimization. Jour-
nal for Multi-Criteria Decision Analysis 11, 139–150.
Erdmann, A., Nolte, A., Noltemeier, A., Schrader, R. (1999). Modeling and solving the airline schedule
generation problem. Technical Report 99-351, ZAIK, University of Cologne, Germany.
Erzberger, H. (1995). Design principles and algorithms for automated air traffic management. AGARD
Lecture Series 200, Brussels. Available at http://www.ctas.arc.nasa.gov/.
Etschmaier, M.M., Mathaisel, D.F.X. (1985). Airline scheduling: An overview. Transportation Sci-
ence 19, 127–138.
EUROCONTROL (2001). CAMACA: The commonly agreed methodology for airside capacity assess-
ment. Brussels. Available at http://www.eurocontrol.int/camaca/.
EUROCONTROL (undated). RAMS homepage. Available at http://www.eurocontrol.fr/projects/
rams/mainwin.html.
Fan, T.P. (2003). Market-based airport demand management – theory, model and applications. PhD
dissertation, Department of Aeronautics and Astronautics, Massachusetts Institute of Technology,
Cambridge, MA.
Fan, T.P., Odoni, A.R. (2002). A practical perspective on airport demand management. Air Traffic Con-
trol Quarterly 10, 285–306.
Federal Aviation Administration (FAA) (2001). Airport capacity benchmark report, US Government
Printing Office, Washington, DC. Available at http://www.faa.gov/events/benchmarks/.
Filar, J., Manyem, P., White, K. (2000). How airlines and airports recover from schedule perturbations:
A survey. Annals of Operations Research 108, 315–333.
Gilbo, E.P. (1993). Airport capacity: Representation, estimation, optimization. IEEE Transactions on
Control Systems Technology 1, 144–154.
Gilbo, E., Howard, K. (2000). Collaborative optimization of airport arrival and departure traffic flow
management strategies for CDM. In: Proceedings of 3rd USA/Europe Air Traffic Management R&D
Seminar.
Hall, W. (1999). Information flows and dynamic collaborative decision-making architecture: Increasing
the efficiency of terminal area operations. PhD dissertation, Operations Research Center, Massa-
chusetts Institute of Technology, Cambridge, MA.
Hane, C.A., Barnhart, C., Johnson, E.L., Marsten, R.E., Nemhauser, G.L., Sigismondi, G. (1995). The
fleet assignment problem: Solving a large-scale integer program. Mathematical Programming 70,
211–232.
Hansen, M. (2002). Micro-level analysis of airport delay externalities using deterministic queuing mod-
els: A case study. Journal of Air Transport Management 8, 73–87.
Hinston, J.M., Aigoin, G., Delahaye, D., Hansman, R.J., Puechmorel, S. (2001). Introducing struc-
tural considerations into complexity metrics. In: Proceedings of 4th USA/Europe ATM R&D Seminar,
Santa Fe, NM.
Hoffman, R. (1997). Integer programming models for ground-holding in air traffic flow management.
PhD thesis, The National Center of Excellence for Aviation Operations Research, University of
Maryland, College Park, MD.
Hoffman, R., Ball, M.O. (2003). A comparison of formulations for the single-airport ground-holding
problem with banking constraints. Operations Research 48, 578–590.
Inniss, T., Ball, M.O. (2002). Estimating one-parameter airport arrival capacity distributions for air
traffic flow management. Air Traffic Control Quarterly 12, 223–252.
International Air Transport Association (IATA) (2000). Worldwide Scheduling Guidelines, 1st edition.
Montreal, Canada.
Jarrah, A., Yu, G., Krishnamurthy, N., Rakshit, A. (1993). A decision support framework for airline
flight cancellations and delays. Transportation Science 27, 266–280.
 Ch. 1. Air Transportation: Irregular Operations and Control 65

Kang, L.S., Clarke, J.P. (2002). Degradable airline scheduling. Working paper, Global Airline Industry
Program, Massachusetts Institute of Technology, Cambridge, MA.
Klabjan, D. (2003). Large scale models in the airline industry. Working paper, Department of Mechan-
ical and Industrial Engineering, University of Illinois at Urbana-Champaign, IL.
Klabjan, D., Johnson, E.L., Nemhauser, G.L., Gelman, E., Ramaswamy, S. (2002). Airline crew
scheduling with time windows and plane count constraints. Transportation Science 36, 337–348.
Kotnyek, B., Richetta, O. (2004). Equitable models for the stochastic ground holding problem under
collaborative decision-making. Working paper, MSIS Department, University of Massachusetts-
Boston, Boston, MA.
Lan, S., Clarke, J.P., Barnhart, C. (2005). Planning for robust airline operations: Optimizing aircraft
routings and flight departure times to minimize passenger disruptions. Transportation Science 40,
15–28.
Lettovsky, L. (1997). Airline operations recovery: An optimization approach. PhD thesis, Georgia In-
stitute of Technology, Atlanta, GA.
Lettovsky, L., Johnson, E.L., Nemhauser, G.L. (2000). Airline crew recovery. Transportation Science 34,
337–348.
Levin, A. (1971). Scheduling and fleet routing models for transportation systems. Transportation Sci-
ence 5, 232–255.
Lohatepanont, M., Barnhart, C. (2004). Airline schedule planning: Integrated models and algorithms
for schedule design and fleet assignment. Transportation Science 38, 19–32.
Long, D., Lee, D., Johnson, J., Gaier, E., Kostiuk, P. (1999). Modeling air traffic management technolo-
gies with a queuing network model of the national airspace system. Report NASA/CR-1999-208988,
NASA Langley Research Center, Hampton, VA.
Luo, S., Yu, G. (1997). On the airline schedule perturbation problem caused by the ground delay pro-
gram. Transportation Science 31, 298–311.
Manning, C.A., Mills, S.H., Fox, C.M., Mogilka, H.J. (2002). Using air traffic control taskload mea-
sures and communications events to predict subjective workload. Report DOT/FAA/AM-02/4, FAA
Office of Aerospace Medicine, Washington, DC.
Marsten, R., Subramanian, R., Gibbons, L. (1996). Junior analyst extraordinaire (JANE). In: AGIFORS
Proceedings, pp. 247–259.
Mathaisel, D.F.X. (1996). Decision support airline system operations control and irregular operations.
Computers & Operations Research 23, 1083–1098.
Metron Inc. (2001). Collaborative decision making. Available at http://www.metsci.com/cdm.
Midkiff, A.H., Hansman, R.J., Reynolds, T.G. (2004). Air carrier flight operations. Report ICAT-2004-3,
MIT International Center for Air Transportation, Massachusetts Institute of Technology, Cam-
bridge, MA.
Morrison, S. (1987). The equity and efficiency of runway pricing. Journal of Public Economics 34, 45–60.
Nilim, A., El Ghaoui, L., Duong, V., Hansen, M. (2001). Trajectory based air traffic management un-
der weather uncertainty. In: Proceedings of the Fourth USA/Europe Air Traffic Management R&D
Seminar, Santa Fe, NM.
Niznik, T. (2001). Optimizing the airline response to ground delay programs. Presented at AGIFORS
Symposium, Ocho Rios, Jamaica.
Odoni, A.R. (1987). The flow management problem in air traffic control. In: Odoni, A.R., Bianco, L.,
Szego, G. (Eds.), Flow Control of Congested Network. Springer-Verlag.
Odoni, A.R., Deyst, J., Feron, E., Hansman, R.J., Khan, K., Kuchar, J.K., Simpson, R. (1997).
Existing and required modeling capabilities for evaluating ATM systems and concepts. Inter-
national Center for Air Transportation, Massachusetts Institute of Technology. Available at
http://web.mit.edu/aeroastro/www/labs/AATT/aatt.html.
Preston Aviation Solutions (2002). TAAM solutions. Available at http://www.preston.net/products/
TAAM.htm.
Psaraftis, H.N. (1980). A dynamic programming approach for sequencing groups of identical jobs. Op-
erations Research 28, 1347–1359.
Rakshit, A., Krishnamurthy, N., Yu, G. (1996). System operations advisor: A real-time decision support
system for managing airline operations at United Airlines. Interfaces 26, 50–58.
66 M. Ball et al.

Rexing, B., Barnhart, C., Kniker, T., Jarrah, A., Krishnamurthy, N. (2000). Airline fleet assignment with
time windows. Transportation Science 34, 1–20.
Richetta, O., Odoni, A.R. (1993). Solving optimally the static ground-holding policy problem in air
traffic control. Transportation Science 27, 228–238.
Richetta, O., Odoni, A.R. (1994). Dynamic solutions to the ground-holding problem in air traffic con-
trol. Transportation Research A 28, 167–185.
Rosenberger, J., Schaefer, A.J., Goldsman, D., Johnson, E.L., Kleywegt, A.J., Nemhauser, G.L. (2002).
A stochastic model of airline operations. Transportation Science 36, 357–377.
Rosenberger, J., Johnson, E., Nemhauser, G. (2003). Rerouting aircraft for airline recovery. Transporta-
tion Science 37, 408–421.
Rosenberger, J., Johnson, E.L., Nemhauser, G.L. (2004). A robust fleet-assignment model with hub
isolation and short cycles. Transportation Science 38, 357–368.
Schaefer, A.J., Johnson, E.L., Kleywegt, A.J., Nemhauser, G.L. (2005). Airline crew scheduling under
uncertainty. Transportation Science 39, 340–348.
Simpson, R.W. (1966). Computerized schedule construction for an airline transportation system. Re-
port FT-66-3, Flight Transportation Laboratory, Massachusetts Institute of Technology, Cambridge,
MA.
Smith, B.C. (2004). Robust airline fleet assignment. PhD dissertation, The Logistics Institute, Georgia
Institute of Technology, Atlanta, GA.
Sohoni, M.G., Johnson, E.L., Bailey, T.G., Carter, H. (2002). Operational airline reserve crew planning.
Working paper, The Logistics Institute, Georgia Institute of Technology, Atlanta GA.
Sohoni, M.G., Johnson, E.L., Bailey, T.G. (2003). Long range reserve crew manpower planning. Work-
ing paper, The Logistics Institute, Georgia Institute of Technology, Atlanta GA.
Soumis, F., Ferland, J.A., Rousseau, J.-M. (1980). A model for large scale aircraft routing and schedul-
ing problems. Transportation Research B 14, 191–201.
Sridhar, B., Seth, K.S., Grabbe, S. (1998). Airspace complexity and its application in air traffic manage-
ment. In: The 2nd USA/Europe ATM R&D Seminar, Orlando, FL.
Stamatopoulos, M., Zografos, K., Odoni, A.R. (2004). A decision support system for airport strategic
planning. Transportation Research C 12, 91–118.
Stojkovic, G., Soumis, F., Desrosiers, J., Solomon, M. (2002). An optimization model for a real-time
flight scheduling problem. Transportation Research A 36, 779–788.
Stojkovic, M., Soumis, F., Desrosiers, J. (1998). The operational airline crew scheduling problem. Trans-
portation Science 32, 232–245.
Talluri, K. (1996). Swapping applications in a daily airline fleet assignment. Transportation Science 30,
237–284.
Teodorovic, D., Guberinic, S. (1984). Optimal dispatching strategy on an airline network after a sched-
ule perturbation. European Journal of Operational Research 15, 178–183.
Teodorovic, D., Stojkovic, G. (1990). Model for operational daily airline scheduling. Transportation
Planning and Technology 14, 273–286.
Teodorovic, D., Stojkovic, G. (1995). A model to reduce airline schedule disturbances. Journal of Trans-
portation Engineering ASCE 121, 324–331.
Terrab, M., Odoni, A.R. (1993). Strategic flow management for air traffic control. Operations Re-
search 41, 138–152.
Thengvall, B., Yu, G., Bard, J. (2000). Balancing user preferences for aircraft schedule recovery during
irregular airline operations. IIE Transactions 32, 181–193.
Vasquez-Marquez, A. (1991). American airlines arrival slot allocation system (ASAS). Interfaces 21,
42–61.
Venkatakrishnan, C.S., Barnett, A.I., Odoni, A.R. (1993). Landing time intervals and aircraft sequenc-
ing in a major terminal area. Transportation Science 27, 211–227.
Vickrey, W. (1969). Congestion theory and transport investment. American Economic Review 59, 251–
260.
Vossen, T., Ball, M.O. (2006a). Optimization and mediated bartering models for ground delay pro-
grams. Naval Research Logistics 53, 75–90.
 Ch. 1. Air Transportation: Irregular Operations and Control 67

Vossen, T., Ball, M.O. (2006b). Slot trading opportunities in collaborative ground delay programs.
Transportation Science 40, 29–43.
Vossen, T., Ball, M.O., Hoffman, R., Wambsganss, M. (2003). A general approach to equity in traffic
flow management and its application to mitigating exemption bias in ground delay program. Air
Traffic Control Quarterly 11, 277–292. A preprint was published in Proceedings of 5th USA/Europe
Air Traffic Management R&D Seminar, 2003.
Vranas, P.B., Bertsimas, D.J., Odoni, A.R. (1994). The multi-airport ground holding problem in air
traffic control. Operations Research 42, 249–261.
Wambsganss, M.C. (1996). Collaborative decision making through dynamic information transfer. Air
Traffic Control Quarterly 4, 107–123.
Wei, G., Yu, G., Song, M. (1997). Optimization model and algorithm for crew management during
irregular operations. Journal of Combinatorial Optimization 1, 80–97.
Willemain, T.R. (2002). Contingencies and cancellations in ground delay programs. Air Traffic Control
Quarterly 10, 43–64.
Wilson, F.W. (2004). A stochastic air traffic management model that incorporates probabilistic forecasts.
In: Proceedings of 20th International Conference on Interactive Information and Processing Systems
(IIPS) for Meteorology, Oceanography, and Hydrology.
Wyndemere, I. (1996). An evaluation of air traffic control complexity. Report NASA-2-14284, NASA
Ames Research Center, Moffett Field, CA.
Yan, S.Y., Lin, C.-G. (1997). Airline scheduling for the temporary closure of airports. Transportation
Science 31, 72–82.
Yan, S.Y., Tu, Y.P. (1997). Multifleet routing and multistop flight scheduling for schedule perturbation.
European Journal of Operational Research 103, 155–169.
Yan, S.Y., Yang, D.H. (1996). A decision support framework for handling schedule perturbation. Trans-
portation Research B 30, 405–419.
Yan, S.Y., Young, H.F. (1996). A decision support framework for multi-fleet routing and multi-stop
flight scheduling. Transportation Research A 30, 379–398.
Yen, J., Birge, J. (2000). A stochastic programming approach to the airline crew scheduling problem.
Technical report, University of Washington, Seattle, WA.
Young, H.P. (1994). Equity in Theory and Practice. Princeton Univ. Press, Princeton, NJ.
Yu, G. (1995). An optimization model for airlines’ irregular operations control. In: Proceedings of the
International Symposium on Operations Research with Applications in Engineering, Technology, and
Management, Beijing, China.
Yu, G. (Ed.) (1997). Operations Research in the Airline Industry. Kluwer Academic, Boston, MA.
Yu, G., Arguello, M., Song, G., McCowan, S., White, A. (2003). A new era for crew recovery at Conti-
nental Airlines. Interfaces 33, 5–22.
C. Barnhart and G. Laporte (Eds.), Handbook in OR & MS, Vol. 14
Copyright © 2007 Elsevier B.V. All rights reserved
DOI: 10.1016/S0927-0507(06)14002-5

Chapter 2
Public Transit
Guy Desaulniers
Department of Applied Mathematics and Industrial Engineering, École Polytechnique and
GERAD, Montréal, Québec, H3T 2A7 Canada
E-mail: [email protected]

Mark D. Hickman
Department of Civil Engineering and Engineering Mechanics, The University of Arizona,
Tucson, AZ 85721, USA
E-mail: [email protected]

1 Introduction

For several decades now, operations research has been successful for solv-
ing a wide variety of optimization problems in public transit. Several com-
mercial software systems based on operations research techniques have been
designed and used by the transit agencies to help them plan and run their
operations. Operations researchers have been attracted by the public tran-
sit problems because of their size and complexity. Indeed, some of them are
huge in practice. For instance, the New York City Transit Authority employs
more than 12,000 drivers to operate approximately 4500 buses that serve over
240 bus routes. Furthermore, these problems are complex because they in-
volve passengers, buses, and drivers that are subject to individual preferences
and constraints, and interact with each other according to a set of prescribed
relationships.
The main goal of most transit agencies is to offer to the population a ser-
vice of good quality that allows passengers to travel easily at a low fare. The
agencies thus have a social mission which aims at reducing pollution and traf-
fic congestion, as well as increasing the mobility of the population. In most
cases, the goal is usually not to make profits, as is the case for almost all other
transportation organizations such as airlines, railroads, and trucking compa-
nies. They are, however, subject to budgetary restrictions that force them to
manage expensive resources such as buses, drivers, maintenance facilities, and
bus depots as efficiently as possible. Briefly stated, the global problem faced by
the agencies consists of determining how to offer a good-quality service to the
passengers while maintaining reasonable asset and operating costs.
Addressed as a whole, this global problem is not tractable. Hence, it is
divided into a set of subproblems that are usually solved sequentially at var-
ious stages of the planning process (strategic, tactical, and operational), and
even during operations (real-time control). Strategic planning problems con-

69
70 G. Desaulniers and M.D. Hickman

cern long-term decisions such as the design of the transit routes and networks.
Most of these problems fall within the category of network design problems
and require solving passenger assignment problems as subproblems or for eval-
uation purposes. These strategic problems aim at maximizing service quality
under budgetary restrictions. Tactical planning problems concern the decisions
related to the service offered to the public, namely the frequencies of ser-
vice along the routes and the timetables. These problems are usually solved
on a seasonal basis, with occasional updates. These problems also focus on
the quality of service. Operational planning problems relate to how the oper-
ations should be conducted to offer the proposed service at minimum cost.
They include a wide variety of problems such as vehicle scheduling, driver
scheduling, bus parking and dispatching in garages, and maintenance schedul-
ing. These problems are solved at various intervals that range from once per
month for driver scheduling to once per day for bus parking and dispatching.
In contrast to the objectives of the previous problems, the objective for the
operational planning problems is clearly one of minimizing total cost. Finally,
real-time control problems manage perturbations to the plan using several con-
trol strategies. These problems are solved in real time during operations and
aim at minimizing passenger inconvenience. Usually, minimal or no operating
costs are involved since they consider minor perturbations to the scheduled
service.
The goal of this chapter is to review state-of-the-art models and approaches
for solving these public transit problems. This review is not exhaustive as it
mostly covers the recent contributions that have been applied or have the
potential to be applied from our viewpoint. Readers interested on earlier
works are referred to the survey paper by Odoni et al. (1994), as well as the
series of books arising from the Computer-Aided Transit Scheduling confer-
ences that have contributed tremendously to the practice and growth of op-
erations research in public transit. These books are listed at the beginning of
the references, i.e. Wren (1981), Rousseau (1985), Daduna and Wren (1988),
Desrochers and Rousseau (1992), Daduna et al. (1995), Wilson (1999), Voss
and Daduna (2001), and Hickman et al. (in press).
In certain cases, computational results are briefly reported to give an idea of
the problem sizes that can be solved. However, these results are not intended
to compare the different approaches. In fact, they can hardly be compared
because they most of the times have been obtained using different computers
or on different datasets that did not necessarily exhibit the same characteristics.

2 Strategic planning

At the strategic level, transit planning is concerned with the design of tran-
sit routes and networks. This involves designing a network of routes to meet
passenger demand. Since the demand is based in large part on the network
design, the network design problem relies heavily on methods to determine
 Ch. 2. Public Transit 71

passengers’ route choices (or “assignment” to routes) serving their origins and
destinations. This section includes a description of the transit network design
problem and a discussion of research in passenger assignment.

2.1 Network design

The public transit network design problem is somewhat more complicated

than the traditional network design problem. In addition to determining what
links to include in the network, the transit network design includes assembling
these links into fixed routes, and determining the frequency of service on each
route. The result of the network design, then, should include a set of routes and
their frequencies. Most commonly, the problem is formulated on a graph with
nodes, links, and (subsequently) routes. Let G = (N A) be a graph with N,
the set of nodes, and A, the set of links, and R represents the set of routes.
Nodes represent intersections (e.g., road intersections), but can also represent
zone centroids where a geographic zone is represented by a single point (the
centroid). A link between nodes represents a particular mode of transport be-
tween nodes, and a route represents a sequence of nodes and links of a single
mode.
As input to this problem, the formulation typically assumes that there is an
existing origin–destination (O–D) matrix, covering the demand between a set
of nodes or zones, either on a daily basis or for a specific period within the day.
Alternately, it is reasonable (although complicated) to assume that demand
is endogenous and determined as an equilibrium problem, in which the flows
are a function of the network design. With the origin–destination flows and
an assignment of these flows to routes, the set of routes and their frequencies
must be determined.
One of the challenges of network design is in the specification of the ob-
jective function. Most commonly, the objective is to minimizing the total travel
time or the generalized cost of travel. The generalized cost may be found by ap-
plying different weights in the objective function to the different components
of travel time such as walking (or access) time, initial waiting time, in-vehicle
time, transfer time, and egress time separately. Some formulations also include
the number of transfers as a component in the generalized cost.
In addition, the costs to the transit operator may also be considered, either
explicitly in the objective function or through a constraint on the total budget
(or operating profit or loss). Such costs can include the operating cost, given
as a function of the route length (in distance or time) and frequency, and the
fixed cost of a bus fleet and/or infrastructure along the route network (for rail
transit networks). If operator costs are included, a composite objective may be
formulated, or the problem can be specified as a multiobjective programming
problem.
In addition to this traditional formulation of the objective, several other
constraints often enter into the problem in practice; these include: (1) ensuring
adequate coverage in the network to provide access to specific nodes or zones
72 G. Desaulniers and M.D. Hickman

in the service area; (2) ensuring minimum frequencies of service to specific

nodes or links in the network; and (3) any other design considerations such as
the availability of infrastructure or right-of-way for routes.
In practice, the network design problem incorporates each of these objec-
tives and constraints through a more interactive formulation of the problem,
where sample routes may be constructed by computer methods, but are ulti-
mately selected in conjunction with manual review by network designers and
planners. In this review, we discuss some of the more significant methods that
rely heavily on mathematical programming techniques.
The network design problem is known to be NP-hard (Magnanti and Wong,
1984). As a result, approaches to the problem rely on heuristic techniques to
solve problems of reasonable size. Much of the initial work decomposed the
problem into two stages: in the first, the set of routes is constructed; in the
second, the set of frequencies for these routes are determined. This includes
the work of Lampkin and Saalmans (1967) and Silman et al. (1974). In the first
stage, heuristic methods are used to construct “skeleton” routes, and these
skeleton routes are expanded to cover the full set of nodes in the network.
Once the routes are defined, the frequencies are determined by minimizing
the total passenger travel time, calculated as the sum of the O–D demand Dij
multiplied by the travel time Tij , subject to a constraint on the total fleet size
(as a budget constraint):

minimize Dij Tij (f) (1)
i∈N j∈N
subject to:

RT r fr   total fleet size (2)
r∈R

In this formulation, RT r is the round-trip time on route r and fr is the fre-

quency on route r. The travel time Tij includes the expected waiting time, as a
function of frequencies f of routes serving the origin i and any transfer node k
on the shortest path serving the O–D pair i, j. Lampkin and Saalmans (1967)
used a random gradient-based search procedure to determine the final fre-
quency values. In Silman et al. (1974), a penalty is added to the objective for the
estimated number of standees on the bus; this penalty is given as a piecewise
differentiable function of the route frequency. Their approach uses a gradient
projection method to minimize the total travel time.
Dubois et al. (1979) decomposed the network design problem into three
subproblems. The first involves determining the links in the street network on
which to operate the service; the second determines the routes themselves;
and the third determines the optimal frequencies on each route. In the first
step, a traditional network design problem is formulated, where the objective
is given in (1), using only in-vehicle times as the travel times, subject to a bud-
get constraint on the cost of operating on a street, and binary decision variables
indicating whether a street segment is in the final solution. A heuristic is used
 Ch. 2. Public Transit 73

to solve this problem, beginning with an initial spanning tree to minimize the
total travel time and adding links to minimize this total. A simple all-or-nothing
assignment of the O–D flow on the shortest paths is used to estimate the ob-
jective function. In the second step, a maximal set of routes is generated from
the street network. This set of routes is then subject to heuristic rules to deter-
mine the final route structure: (1) routes with heavy transfer flow are joined;
(2) route segments are deleted where the demand is effectively served by other
routes; and (3) routes that overlap are joined. In the third step, the optimal
route frequencies are found using a gradient-based search heuristic, similar to
Lampkin and Saalmans (1967). Waiting times are explicitly considered in this
last step.
A more formal presentation of the transit network design problem, from
a mathematical programming approach, is given by Hasselström (1981). Has-
selström proposed a two-stage process of network design in which routes and
frequencies are determined simultaneously. In the first stage, an initial route
network is generated; in the second stage, this route network is refined and a
detailed evaluation of the routes and the passenger assignment is performed.
In addition to solving the route and frequency problem simultaneously, the ad-
vantage of Hasselström’s method is in its implementation and application for
realistic network sizes.
In the first stage, Hasselström’s (1981) formulation includes a direct demand
function, allowing the demand to be determined endogenously. The form of
the direct demand model is based on a traditional gravity model with parame-
ter β, where all terms not dependent on the route structure or the frequencies
are rolled into a constant term Kij for each O–D pair i, j. Remaining ele-
ments of the generalized cost are given by Cij , which is a function of the set of
frequencies f. The frequency of each route r is denoted fr . The objective func-
tion, maximizing consumer surplus, is equivalent to maximizing the number of
passengers with this demand function. As constraints, Hasselström (1981) con-
sidered a budget constraint C  that includes a cost per vehicle on each route cr .
Also, there is a required minimum frequency of service for a zone s ∈ S, given
as Δs . If a route r serves zone s, this is indicated with a binary parameter δrs .
Finally, the frequencies must be included in a feasible set F (e.g., nonnegative
integers).
The network design problem in the first stage is then formulated as fol-
lows (in this formulation, the notation is simplified to deal with a single transit
mode; multiple modes may also be considered in the network):

maximize Dij (3)
i∈N j∈N
subject to:
Dij = Kij e−βCij (f)  ∀i j ∈ N (4)


cr fr  C (5)
r∈R
74 G. Desaulniers and M.D. Hickman

fr δrs  Δs  ∀s ∈ S (6)
r∈R
fr ∈ F  ∀r ∈ R (7)
In (4), the generalized cost term Cij , a function of the frequencies f, includes
waiting and transfer times for the O–D pair, as determined in the passenger
assignment. The demand is therefore given as a function of the service fre-
quencies.
To solve this model in the first stage, an initial network is generated by
enumerating all possible routes serving a pair of terminals. A set of heuristic
rules is then used to prune clearly inferior routes from this set. Then, the final
routes and frequencies are constructed by solving a mathematical program:
one method uses a linear program to maximize the passenger flow; a second
method uses a convex nonlinear program to maximize the consumer surplus.
In this first stage, the assignment uses all common routes (see Section 2.2.1) to
determine waiting and transfer times. The decision variables in both formula-
tions are the frequencies of each route; routes with very low frequencies can
be pruned from the solution space.
In the second stage, a detailed assignment is performed, and the routes
are refined. Passenger assignment on existing routes is performed with the
heuristic of Andreasson (1977) (discussed in Section 2.2.1). With this new as-
signment, several route refinements are considered. First, optimization of the
connection of route segments at route intersections is proposed; this problem
is formulated as a maximum weighted matching problem of combining route
segments at the point of intersection. Also, a nonlinear program is formulated
for re-optimization of frequencies, using the vehicle fleet size constraint. This
optimization is solved by Lagrangian relaxation.
Hasselström (1981) reports on a case study with 50 local bus routes, 10 tram
routes, and express bus service. Since this time, the methodology has been de-
veloped as commercial software, and hence is clearly able to solve realistic
problem sizes.
In the work of Ceder and Wilson (1986), two different mathematical formu-
lations of the bus network design problem are suggested. The first formulation
considers the passenger objective of minimizing excess travel time upon board-
ing, expressed as the sum of “excess” travel time (larger than the shortest travel
time with a direct route) plus the transfer time (if any), summed across all
O–D pairs. This objective is minimized subject to constraints on the maximum
O–D travel time (as a percentage above the shortest path), lower and upper
bounds on the route length (expressed in units of running time), and a con-
straint on the maximum number of routes. A second formulation adds the
passenger waiting time and vehicle operating and capital costs to the objec-
tive function; it also includes constraints on the minimum frequency for each
route and a constraint on the maximum fleet size. To generate a large set of
feasible routes, Ceder and Wilson (1986) proposed a heuristic in which each
designated terminal node is processed separately. A (topological) breadth-first
 Ch. 2. Public Transit 75

search is conducted, in which potential routes that do not meet the constraints
on the maximum travel time are eliminated. In addition, the total “excess” pas-
senger hours are also calculated for the route. This set of feasible routes can
then be used for further screening and passenger assignment.
The mathematical formulation of Ceder and Wilson (1986) was extended
more recently by Israeli (1992) and related papers (Israeli and Ceder, 1989,
1995; Ceder and Israeli, 1998). In this research, the transit network design
problem is formulated as a multiobjective programming problem, with two ob-
jectives: the total passenger cost (Z1 ) and the operator fleet size (Z2 ). The
total passenger costs Z1 includes the in-vehicle passenger hours spent between
the origin and destination PH ij , the waiting and transfer time spent traveling
from the origin to the destination WH ij , and the empty seat-hours on a route r
denoted EH r . These three terms are weighted (weights a1 , a2 , and a3 ) in the
objective. Formally, the objectives are described as
  
minimize Z1 = a1 PH ij + a2 WH ij + a3 EH r  (8)
i∈N j∈N i∈N j∈N r∈R
minimize Z2 = fleet size (9)
Constraints in the formulation include the passenger assignment from a fixed
demand matrix, and minimum frequencies on each route. This problem is
solved with the following heuristic:
1. The full set of feasible routes is enumerated, in a manner similar to Ceder
and Wilson (1986).
2. Additional direct routes are added to the network between O–D pairs
with high demand, where the origin and destination nodes are not termi-
nals. Second, the number of transfers required for each O–D pair for the
given route structure is calculated, up to a maximum (e.g., 1 or 2 trans-
fers).
3. A minimal set of routes is obtained through a heuristic procedure. This
problem is set up as a set covering problem with the full set of feasi-
ble routes. Each column is defined as a route or a feasible combina-
tion of routes meeting the maximum number of transfers; the rows are
O–D pairs. The objective minimizes the deviation from shortest paths,
while maintaining constraints on connectivity for each O–D pair (i.e.,
reachable within the maximum number of allowable transfers).
4. The assignment of flow to paths and the frequencies on each path are de-
termined iteratively. The frequencies are determined based on the peak
load segment on each route, and these in turn are used in calculating
the waiting and transfer times in the assignment. The assignment proce-
dure is loosely based on that of Marguier and Ceder (1984), described in
Section 2.2.1. With this information, Z1 is calculated.
5. The minimum fleet size Z2 is determined using the method of Stern and
Ceder (1983).
76 G. Desaulniers and M.D. Hickman

6. New routes are considered to explore other points in the solution space
for the two objectives. In this procedure, a column generation technique
is used to avoid re-evaluating previously accepted route sets. These new
route sets are re-evaluated by repeating steps 3–5.
7. The route sets in the efficient frontier are evaluated and presented to the
decision-maker.
The example problem and solution in Israeli and Ceder (1995) is a problem
with 8 nodes and 14 links (based on a similar problem from Ceder and Wilson
(1986)). It is unknown how the solution method would perform on larger, more
realistic networks.
Another mathematical programming approach to the transit network design
problem was proposed by van Nes et al. (1988). In this approach, the routes
and frequencies are determined simultaneously. The objective function max-
imizes the number of direct trips (i.e., trips without transfers) served in the
network, for a given fleet size. A direct demand model is proposed to esti-
mate the origin–destination trips by public transit; trips are proportional to the
attraction of the origin zone i, Oi , and the destination zone j, Dj , and is an
exponential function of the cost, similar to that of Hasselström (1981). The
objective function is formulated as

maximize aOi Dj e−βCij (f)  (10)
i∈N j∈N

The generalized cost term Cij is defined as an explicit function of the frequen-
cies of the optimal subset of routes serving the passenger’s origin i, R∗i (Chriqui
and Robillard, 1975):
60α
Cij (f) = Kij +  + c (11)
r∈R∗i fr

In this equation, α and c are constants. This equation assumes, for direct ser-
vice, a constant Kij for access time, egress time, and in-vehicle travel time, and
adds a term for the waiting time as a function of the frequencies on acceptable
routes.
A variety of constraints are used in this formulation. For these, consider a set
M = {1     m} of vehicle types, with Nm the total number of vehicles avail-
able of type m. Operating a vehicle of type m incurs a cost factor km . A total
of Nr vehicles are assigned to route r, with the indicator bmr equal to 1 if vehi-
cle type m is assigned to route r. With these variables, the constraints include:
a budget constraint of C,  the vehicle availability Nm , the set of feasible fre-
quencies F , and the allocation of buses among routes based on the frequency
and the round-trip time on the route RT r . Mathematically, these constraints
are formulated as:
 
km 
Nr bmr  C (12)
m∈M r∈R
 Ch. 2. Public Transit 77

Nr br  Nm  ∀m ∈ M (13)
r∈R
fr ∈ F  ∀r ∈ R (14)
RT r
Nr − 1 < fr  Nr  ∀r ∈ R (15)
60
The solution technique adopted by van Nes et al. (1988) is a heuristic in
which all frequencies on proposed routes are set to 0. Each route is evalu-
ated with respect to its potential to improve “efficiency”, defined as the ratio
of passengers added by the direct service to the additional cost of increasing
the frequency, evaluated in (12)–(15). The route with the highest efficiency
is selected and the frequency on that route is increased, until the budget and
the available vehicles are consumed. It is shown that this heuristic is similar to
evaluating the Kuhn–Tucker conditions for the problem when the budget con-
straint is included in the objective with a Lagrange multiplier. The Lagrange
multiplier is identical to this “efficiency” measure, and these should be approx-
imately equal across routes in the final solution.
van Nes et al. (1988) report testing this solution technique on a network
from the Netherlands with 182 nodes and 115 zones, for a network of 8 routes.
The paper also reports that the modeling system is capable of solving instances
up to 250 nodes, 150 zones, and 750 possible routes.
More recent work has also included a number of metaheuristic methods.
Baaj and Mahmassani (1995) and related work (Baaj and Mahmassani, 1990,
1992) decompose the network design problem into three elements: a route
generation step, in which routes and frequencies are constructed; a network
analysis procedure, defining measures of effectiveness at the network-, route-,
and stop-level; and a route improvement algorithm to improve the route de-
sign. The heuristic proposed by Baaj and Mahmassani (1995) begins by gen-
erating additional skeleton routes connecting the highest O–D pairs in the
demand matrix with direct service. With these skeletons, a set of possible
node selection and insertions strategies are used to generate full routes. Then,
passenger assignment follows the method of Han and Wilson (1982) (see Sec-
tion 3.1) and determines the frequency and number of buses on each route
according to a pre-specified maximum loading factor. Once this assignment is
completed, the network evaluation tool is used to identify the number of trips
satisfied by direct, one-transfer, and two-transfer trips, and the total waiting
time, in-vehicle travel time, and transfer time in the network. The route im-
provement procedures then are called to improve the route structure through
heuristics that: (1) prune off low ridership routes and/or route segments and
joining these with other routes; and (2) consider improvements to routes by
splitting routes into two parts or by exchanging route legs between routes at
points of intersection.
A recent study by Fan and Machemehl (2004) examined simulated anneal-
ing, tabu search, genetic algorithms, local search, and random search tech-
niques to solve the network design problem. As with other previous methods,
78 G. Desaulniers and M.D. Hickman

all the techniques begin with a set of skeleton routes. These metaheuristics are
used to generate additional routes; the output is run through a network eval-
uation tool. Subsequent iterations between the network analysis tool and the
metaheuristics are used to improve the quality of the solution.
Other authors have investigated the use of genetic algorithms for the transit
network design problem; these include (among many others) recent works by
Pattnaik et al. (1998), Bielli et al. (2002), Tom and Mohan (2003), and Verma
and Dinghra (2005). These methods involve two steps. First, all feasible routes
are generated, typically in a method similar to that proposed by Ceder and
Wilson (1986). A set of routes are coded into the genetic algorithm as a string,
containing a certain number of routes. For this technique, a fixed number of
routes are required by the algorithm, although the number of potential routes
can vary so as to determine the optimal number of routes. These strings are
then evaluated using the assignment and network evaluation techniques of
Baaj and Mahmassani (1995), and consistent with the methods of genetic algo-
rithms, the pool of strings to be evaluated is evolved to a new population, and
the process iterates. The size of networks in the genetic algorithm approach
can be somewhat larger than with the analytic methods; Bielli et al. (2002) re-
port solving an instance of 1134 nodes, 3016 arcs, 459 stops, and 22 routes. Tom
and Mohan (2003) report solving an instance with 1332 nodes (bus stops) and
4076 arcs.

2.2 Passenger assignment

One of the critical issues in strategic planning is determining the demand

on each route and other measures of service consumption. Most of the more
common strategic measures of performance, from the perspective of the pas-
senger, relate to the amount of time and money spent traveling in the network;
i.e., elements of the passenger’s path from the origin to destination. Hence, the
passenger assignment is critical to determining system performance.
The passenger assignment problem can be defined as follows. Given an
origin-to-destination flow, what are the flows on paths through the transit net-
work, taken by the passengers? In formulating this problem, the passenger’s
objective is assumed to be minimizing travel time or generalized cost. The
travel time may consist of some or all of the following variables, with perhaps
different weights: the time to access a stop, the waiting time in the stop, the
in-vehicle time, the transfer time, the number of transfers, egress time, and
any monetary cost. The passenger then faces the task of selecting a route or
set of routes that may be able to get from the origin to the destination with
the minimum time or generalized cost. In the literature, this problem is ad-
dressed within the network design problem (Section 2.1), as part of the task of
determining frequencies on routes (Section 3.1), and also as a unique problem
itself.
In contrast to the simplicity of the problem definition, there are a number of
aspects of the problem that have led to several different research approaches.
 Ch. 2. Public Transit 79

One important concept in passenger assignment is the determination of the

“minimum cost” path. Important elements in defining the attributes of cost
include:
1. The characterization of time-dependence and stochastic attributes in the
minimum cost path.
2. The characterization of a solution as: (1) a single path, including only a
route or combination of routes; (2) a path that can include a set of com-
mon lines, including cases where multiple routes may overlap on some
part of the shortest path; or (3) a strategy, allowing passengers to choose
their own boarding rules as they travel from origin to destination.
3. The effect of capacity and crowding in the transit network.
In the characterization of the minimum cost path, a traditional approach
assumes that deterministic values can be used for travel times on each link. Tra-
ditional shortest path techniques have been easily modified to solve these prob-
lems. More recent approaches have included stochastic and time-dependent
features of the travel time: passenger arrivals, vehicle arrivals, and travel times
may be stochastic. As might be expected, these stochastic processes will greatly
affect the path assignment approach (Nuzzulo, 2003). There is considerable
evidence that passenger arrivals appear to be Poisson for higher-frequency
(lower-headway) service, with headways up to 10–15 minutes. In these cases,
the assumption of Poisson passenger arrivals appears to be common. However,
at longer headways (lower frequencies), some fraction of passengers may actu-
ally time their arrivals with the schedule, which may again significantly compli-
cate the analysis (Turnquist, 1978; Bowman and Turnquist, 1981). Moreover,
the treatment of vehicle arrivals may be considered deterministic (according
to schedule or at the given headway) or stochastic. If vehicle arrivals are as-
sumed to be Poisson, many of the calculations in the path assignment simplify
considerably.
The element of time-dependence, relating to the transit schedules, can also
affect the modeling approach. In its simplest case, with perfect adherence to
the schedule, the choice of a “minimum cost” path decomposes into a time-
dependent shortest path problem. This is usually well solved using variants
of existing shortest path techniques; see, for example, Tong and Richardson
(1984). However, when some combination of both time-dependence and sto-
chastic travel times are introduced, the problem is not so well behaved. As was
shown by Hall (1986), the problem of finding a stochastic and time-dependent
shortest path suffers from the fact that subpaths do not necessarily concate-
nate; instead, a possibly exponential number of paths must be evaluated to
ensure an optimal path is found. When vehicle arrivals are random but some-
what correlated with a schedule, the assignment becomes significantly more
complicated (Hickman and Bernstein, 1997).
A second complication for transit networks is that there may not be a sin-
gle route or set of routes which has the minimum cost. This may occur in
cases where multiple transit routes may overlap on some part of the origin–
80 G. Desaulniers and M.D. Hickman

destination path. This problem is commonly referred to as the common lines

problem, in which a passenger may take one of many routes for at least part
of the path from the origin to the destination. The more general case of multi-
ple origin–destination paths has led to the term strategies (Spiess and Florian,
1989), reflecting possible boarding rules the passenger may use in traveling
from an origin to a destination. In a graph-theoretic model, the subnetwork of
eligible paths from the origin to the destination in a strategy is characterized
as a hyperpath (Nguyen and Pallottino, 1988).
The final characterization that may be made is based on the treatment of
capacity and crowding (de Cea and Fernández, 1996, 2000). Much of the early
literature in the passenger assignment assumed that vehicle capacity was not
typically exceeded, and as a result, capacity and crowding effects could safely
be ignored. This allowed certain simplifications of the problem, although this is
clearly not applicable in all circumstances. Rather, if passenger volumes are as-
sumed to run close to or over the capacity of a route, it might be expected that
passengers may not be able to board the first vehicle to arrive. Hence, waiting
time and transfer times may be directly affected by the volume of passengers
on the route, creating congestion effects. This congestion affects the problem
formulation and solution techniques. As a result, the discussion that follows in
Sections 2.2.1 and 2.2.2 is decomposed into the passenger assignment under
“uncongested” conditions and “congested” conditions, respectively.

2.2.1 Uncongested assignment

The earliest methods of transit assignment used variants of well-known
shortest path algorithms; examples include Dial (1967), Lampkin and Saalmans
(1967), le Clercq (1972), Silman et al. (1974), and Last and Leak (1976). In
these cases, the full demand of each O–D pair is assigned to a shortest path.
The variations from existing shortest path methods are based on two excep-
tions: waiting times and common lines in the network. These two issues are
intertwined. A general formulation of the waiting time, as a function of the
frequency, suggests that waiting time is related to the inverse of the frequency
of routes serving the passenger. If Ri is the set of routes serving the stop i that
also serve the passenger’s destination or an intermediate (transfer) node, then
the expected waiting time is given as
α
E[WT] =   (16)
r∈Ri fr

where α is a parameter, such that α = 1 for Poisson vehicle arrivals, α ≈ 05

with deterministic arrivals. With the expression in (16), the shortest path prob-
lem can be solved by creating additional links in the network representing the
corresponding waiting time. Moreover, if more than one route serves a node i
and also serves a given intermediate node or destination node for an O–D pair,
the assignment is made to each route on the basis of the frequency share. That
 Ch. 2. Public Transit 81

is, the fraction of passengers served by route r , Pr , is given as

fr
Pr =   (17)
r∈Ri fr

This formulation assumes that the passenger takes the first bus to arrive at a
stop, among all routes serving that stop.
Chriqui and Robillard (1975) provided a more rigorous treatment of the
problem when “common lines” serve some part of an O–D path. Specifically,
it may not be to the passenger’s advantage to choose the first among all routes
serving the pair of stops. The most notable case is where one or more of the
routes has a shorter travel time to the destination than the others. In this case,
it may be advantageous not to board a bus on a slow route, if it is the first to ar-
rive. Chriqui and Robillard (1975) formulated this problem as follows. Assume
the passenger will choose a subset of routes, and will board the first route of
this subset to arrive at the origin stop. One may also assume that the travel time
to the destination after boarding is constant for each route, but may vary across
the set of routes serving the stop. Finally, we assume that the passenger desires
to minimize the sum of waiting time and time after boarding. Then, the selec-
tion of routes becomes a hyperbolic programming problem of selecting routes
to include in this subset. Practically, this problem can be solved to optimality
by enumerating the possible route subsets. The result is an optimal route sub-
set R∗i that can be used to define the waiting time at the node and frequency
shares for each route in the subset, using the subset R∗i in the summations of
(16) and (17). Chriqui and Robillard (1975) also derived expressions for wait-
ing times and frequency shares for both uniform- and exponentially-distributed
bus arrival times. These results were extended by Marguier and Ceder (1984)
and Israeli and Ceder (1996), in which routes can be grouped into slow routes
and fast routes.
A related heuristic approach was suggested by Andreasson (1977), in which
a route is considered in the desirable subset if the travel time upon boarding a
bus on that route is less than or equal to the waiting time plus the travel time
upon boarding of the minimum time route. Waiting times and route shares
are then determined based on this route set. Andreasson’s method was also
extended by Jansson and Ridderstolpe (1992), in which they present an iter-
ative heuristic for calculating waiting times and route proportions for transit
networks with multiple routes and modes between an O–D pair. Jansson and
Ridderstolpe (1992) showed that the functions (16) and (17) can create poor
approximations of the waiting time and route shares under deterministic head-
ways; instead, these functions depend heavily on the exact timetable.
The work of Spiess (1983) and Spiess and Florian (1989) introduced the
concept of passenger strategies. In their formulation, the passenger is assumed
to minimize the sum of waiting time and on-board time, where these may vary
based on the passenger boarding rules. They formulate this assignment as one
to minimize the total travel time in the network, taken as the product of arc
82 G. Desaulniers and M.D. Hickman

flows and their associated travel times, plus the total waiting time in the net-
work. This waiting time depends on the routes in each strategy. For a given
node i, let A+ −
i be the set of arcs leaving i and Ai be the set of arcs enter-
ing i. Let va be the flow on arc a, ta be the travel time on arc a, and fa is the
frequency of service on arc a. Also, ωi is the total waiting time experienced
by passengers boarding at node i. The demand generated at node i is gi , and
Vi (a parameter) is the total flow entering node i. The formulation of the pas-
senger assignment problem is given as an integer linear program, in which the
decision variables xa are binary variables indicating if an arc a is in the strategy.
The relaxation of this problem is
 
minimize ta va + ωi (18)
a∈A i∈N
subject to:
 
va − va = gi  ∀i ∈ N (19)
a∈A+
i a∈A−
i

Vi
ωi =   ∀i ∈ N (20)
a∈A+ fa xa
i

va  fa wi  ∀a ∈ A+
i  ∀i ∈ N (21)
va  0 ∀a ∈ A (22)
0  xa  1 ∀a ∈ A (23)
The dual of this linear program has the form of a shortest path problem, re-
sulting in an optimal label-setting algorithm for its solution.
A similar formulation of the assignment problem is presented by de Cea and
Fernández (1989), with the use of nonlinear equality constraints for (21). The
resulting nonlinear optimization problem is solved by incorporating the nonlin-
ear constraints into the objective function. The solution technique can then be
decomposed into three parts: the selection of common lines (a hyperbolic pro-
gramming problem) for each O–D pair; the assignment of the O–D volumes
to links representing common routes in the hyperpath; and the assignment of
common route flows to specific routes by frequency share.
More recent study of passenger assignment has focused on the assignment
of passengers to specific scheduled vehicle trips. That is, the assignment iden-
tifies a particular vehicle trip to which a passenger is assigned. In this formula-
tion, it is necessary to have a time-dependent origin–destination matrix Dij (t).
An extensive discussion of various approaches to schedule-based transit as-
signment, both for uncongested and congested transit networks, is found in
the recent volume edited by Wilson and Nuzzolo (2004).
In Tong and Wong (1999), the time-dependent shortest path technique of
Tong and Richardson (1984) is used to assign trips to the network. Additional
complications are added using stochastic weights on the various components
 Ch. 2. Public Transit 83

of travel time (walk time, waiting time, and transfer time), relative to in-vehicle
time.
A combination of stochastic and time-dependent travel time attributes are
included in Hickman and Bernstein (1997). This model characterizes the fol-
lowing passenger behavior: upon arriving at a stop, the passenger waits until
a bus arrives, and then determines whether or not to board the bus based on
the time spent waiting and the set of additional bus arrivals expected in the fu-
ture. The formulation of this problem essentially requires full enumeration of
all paths from the origin to the destination, accompanied by the derivation of
the probability distributions of travel times on all paths, at the time the board-
ing decision is made. Such “clever” passenger behavior can be used to simulate
passenger behavior under real-time information, as illustrated by Hickman and
Wilson (1995). A similar framework for passenger boarding strategies with in-
formation at the stop was also presented by Gentile et al. (2005).

2.2.2 Congested assignment

The area of congested transit assignment has evolved in examining the ef-
fects of capacity limitations on passenger path assignment. With vehicle capaci-
ties, it may be the case that demand is sufficiently high that a passenger desiring
to board a given vehicle may be unable to. This results in higher waiting times
when in such crowded conditions. This has led to formal specification of equi-
librium transit assignment in which the delay that each passenger imposes on
other passengers is explicitly included in the model. In general, the effect of
crowding and vehicle capacity is incorporated in the modeling through the
impact of additional waiting and/or transfer time caused by passengers being
unable to board a desired vehicle because it is full. It may also be included as
additional “discomfort” experienced by passengers while on board the vehicle.
However, in these models it is important to note that the effect of congestion
is clearly not symmetrical by arc flows.
Mathematically, the most common approach to include capacity and crowd-
ing has been to formulate the waiting time as an increasing function of the
volume on a particular line, both in terms of the passengers on-board and the
passengers waiting to board at a stop. A graph-theoretic structure for transit
equilibrium assignment was developed by Nguyen and Pallottino (1988). These
authors proposed the graph concept of a hyperpath, defined as an acyclic, di-
rected subgraph of routes connecting the passenger’s origin–destination pair.
Such a hyperpath may be defined using a particular passenger strategy (Spiess
and Florian, 1989). Passengers are assumed to travel on the shortest hyperpaths
connecting their origin and destination.
When equilibrium conditions occur, the problem formulation and solution
methods from traditional traffic assignment can be used (refer to Chapter 10 in
this volume). Nguyen and Pallottino (1988) suggested using traditional traffic
assignment techniques, with the adaptation of these techniques to calculate
shortest hyperpaths (or strategies) rather than shortest paths, when calculating
a direction for improvement in the assignment.
84 G. Desaulniers and M.D. Hickman

In de Cea and Fernández (1993), the waiting time is considered to be a

function of both the passengers desiring to board at the given stop and those
passengers traveling through the stop on board the route. Mathematically, the
passenger travel time is calculated as a power function of the conflicting volume
divided by the capacity to approximate the cost of congestion. The conflicting
volume is the sum of the boarding volume and the passengers on board. With
congestion on the waiting or boarding arcs, a common approach is to define an
effective frequency f , determined as the average frequency observed by the pas-
senger, assuming he/she may be denied boarding on the first vehicle in his/her
strategy. This effective frequency, calculated as the reciprocal of the expected
waiting time for a given route, can be sensitive to the level of crowding upon
boarding. In turn, the effective frequency f can be used in the more tradi-
tional waiting time and frequency-based assignment approaches found in (16)
and the proportional assignment in (17); for a discussion of these issues, see
Bouzaïene-Ayari et al. (2001).
de Cea and Fernández (1993) determined the congestion cost using a linear
function of the conflicting volume divided by the capacity. The model uses a
variational inequality formulation with nonlinear constraints, with assignment
to routes based on the effective frequencies. If the assignment to routes is made
on the basis of actual route frequencies, rather than the effective frequencies,
the problem has linear constraints. In either case, this problem can be solved
using diagonalization, making it susceptible to traditional traffic assignment
algorithms.
Wu and Florian (1993) and Wu et al. (1994) extended this work to include a
formal strategy formulation of the congested assignment problem. The prob-
lem is formulated as an asymmetric network equilibrium problem, but with a
variable transformation to solve in the space of hyperpath flows rather than
route flows. The solution method uses a symmetric linearization (similar to di-
agonalization), called the linearized Jacobi method, for solving the resulting
variational inequality problem.
Moreover, Cominetti and Correa (2001) extended the models of Wu et al.
(1994) to consider an effective frequency model of transit assignment, consider-
ing possible congestion on the boarding links. The principle finding in this work
is the definition of some necessary and sufficient conditions that an equilibrium
transit assignment has been reached, both in terms of arc flows and in terms
of the strategy. A straightforward cost function is established as an objective.
The proposed algorithm performs shortest-hyperpath assignment on the inner
loop, while using the method of successive averages to update the flows after
each iteration.
Lam et al. (1999) presented a stochastic user equilibrium model for passen-
ger assignment. This model assumes a simple bottleneck model for congestion
on a link; i.e., the additional delay is linear with the ratio of the conflicting
volume to the capacity. A multinomial logit model is used for the selection of
paths. The problem is formulated as a nonlinear programming problem with
linear constraints. It is shown that conditions on the Lagrange multipliers in the
 Ch. 2. Public Transit 85

Kuhn–Tucker conditions can be specified such that the route capacities are not
exceeded. Rather, when route capacities are reached, the bottleneck delays are
proportional to the Lagrange multipliers. With this observation, the problem
is solved using existing solution techniques for the stochastic user equilibrium
problem in traffic assignment (see Chapter 10).
The work of Lam et al. (2002) extended this model to a more disaggregate
model of route operations. In a more detailed model of stop operations, the
total route travel time is affected by the delay in boarding and alighting at
the stop. The frequency on the route, in turn, is determined by the number of
vehicles divided by the round-trip travel time. In this case, the waiting time is
a function of the frequency, but the frequency itself is a function of the delay
caused by crowding. As a result, the assignment is then a fixed point in both
the space of frequency and boarding and alighting volumes. This fixed-point
problem forms an outer iteration on the assignment and is solved using the
method of successive averages.
A more general network model has been presented in the work of Lo
et al. (2003). This model accommodates nonlinear fare structures and transfers
through the use of a state-augmented multimodal (SAM) network. In this net-
work representation, the node itself is augmented based on the opportunities
to transfer at the node, the number of transfers made by the passenger in the
network, and explicit representation of direct links in the network to represent
nonlinear costs. This network is then applied in a stochastic user equilibrium
assignment, using a logit model. This framework has also been extended to a
nested logit structure by Lo et al. (2004).
Nielsen (2000) considered a stochastic user equilibrium model that is based
on congestion both for waiting times as well as in-vehicle discomfort. Both
measures are functions of the flow on the associated link and the on-board
capacity of the vehicle. The stochastic user equilibrium is based on the pro-
bit model for the selection of paths, among common lines (line aggregation).
Existing methods for probit-based traffic assignment are used to solve the pas-
senger assignment (see Chapter 10). An application using a nested logit model
for the stochastic user equilibrium assignment on a large regional transit net-
work is presented in Nielsen (2004).
As with the uncongested assignment, recent work has been examining as-
signment to specific vehicle trips in the schedule. A transit equilibrium assign-
ment model that explicitly considers schedules was formulated by Nguyen et al.
(2001). In their model, the capacity constraints on boarding and transfer links
are hard constraints. The additional delay is a function of the available capac-
ity of the vehicle as it arrives. The passenger assignment to routes is based on
the passengers’ desired arrival times at the destination; a penalty term (sched-
ule delay cost) is added to the passenger cost for arriving at the destination
at a time other than the desired arrival time. This problem is set up as a
nonmonotonic variational inequality problem and solved using simplicial de-
composition, in the form of a column generation technique that generates new
extreme points in the arc flow solution space.
86 G. Desaulniers and M.D. Hickman

A separate line of thinking has been developed by Nuzzulo et al. (2001). In

these models, the transit passenger is assumed to make a discrete choice of
route and trip, based on the attributes of the trip as well as the desired arrival
time at the destination. The discrete choice model uses for the utility function
typical variables related to travel times, transfers, and related variables. The
choice of run is based on trading off these terms in both a within-day assign-
ment and a day-to-day assignment based on learning, which is accomplished
through an exponential filter of run attributes. Congestion is included in the
model using a measure of delay in the passenger boarding and in-vehicle time.
Also in the area of trip-based assignment, Poon et al. (2004) have extended
the work of Tong and Wong (1999) to congested assignment. In this work,
a time-dependent origin–destination demand is given, and the assignment is
made in a dynamic network based on specific vehicle trips. Congestion is con-
sidered through queuing delay at the stop or station as passengers prepare to
board, and through the available space on the transit vehicle when it arrives
at a stop. The assignment in a single iteration follows Tong and Richardson
(1984) using the latest network travel times, and is performed by moving pas-
sengers incrementally in time through the network. The final assignment is
solved iteratively by the method of successive averages, with the queuing delay
and vehicle loading being updated after each iteration.
Finally, a multiagent approach to transit assignment has been developed by
Wahba and Shalaby (2005). In this approach, passengers are represented as
agents in the transit network. The passenger behavior is described in terms of
their route, stop, and departure time choice, based on a desired arrival time
at the destination. This behavior is simulated in the network on a given day,
and reinforcement learning is used to represent day-to-day adaptation of pas-
sengers to their experiences in the network. This is repeated for many days to
achieve a final transit network assignment.

3 Tactical planning

In tactical planning, we are concerned with intermediate steps in the plan-

ning process in which the frequencies of routes are constructed and the service
schedule is determined. We include these two parts together in the tactical
level because they are predominantly oriented toward structuring and improv-
ing service to the passenger. In this context, Section 3.1 describes the selection
of frequencies on the set of transit routes, and Section 3.2 describes methods
to construct timetables.

3.1 Frequency setting

The process of determining frequencies for transit routes has already been
introduced in the process of network design (Section 2.1). While a set of fre-
quencies are a necessary product of the network design, it is also true that
 Ch. 2. Public Transit 87

a transit agency will evaluate and determine frequencies on routes more of-
ten than this. Variations in passenger demand patterns and smaller changes
in route design may precipitate a need to adjust frequencies. In this section,
we begin with the problem of frequency setting for typical routes, and then we
briefly describe methods of determining frequencies for some other types of
routes that are commonly used.
The problem of setting frequencies can be approached from several differ-
ent ways. The primary goal is to select frequencies that maximize the passenger
service, which can be defined in a number of different ways, subject to a num-
ber of possible constraints. These include constraints on the overall fleet size
(which is assumed fixed for this process), a constraint that capacity on a route
must be sufficient for the demand, and any policy constraints on minimum
desirable frequencies. Other input data include the round-trip and required
layover time on each route. With this information, a transit agency will choose
an allocation of the fleet to particular routes; this allocation will then directly
indicate the frequency of service on each route. Finally, these frequencies may
be specified by time of day and day of week.
The most common practical approach is to design frequencies to meet the
maximum passenger demands without exceeding the capacity, or without ex-
ceeding some threshold value of bus utilization, the ratio of demand to capacity
(Ceder, 1984). In cases where this produces unreasonably low frequencies,
minimum frequencies (or maximum headways) are also commonly applied.
More rigorous mathematical programming models have been developed and
are outlined below, but these have rarely been applied because of their com-
plexity.
Early work on this problem focused on determining frequencies with com-
mon route structures. Scheele (1980) formulated the problem of determining
route frequencies as a nonlinear program, based on minimizing the total gen-
eralized passenger travel time, with decision variables being the frequency of
each route. Simultaneously, this model solves for the flow on each O–D path
(the passenger assignment problem). An O–D path is defined strictly as a
sequence of route segments. The formulation includes constraints that the de-
mand cannot exceed the available capacity on a route, flow conservation in
the assignment, and fleet size constraints. An entropy constraint is also in-
cluded to distribute trips in the transit network and to ensure accessibility
between all origins and destinations. An iterative solution methodology is pro-
posed in which the set of frequencies is fixed, and the O–D and path flows are
determined through a Lagrangian function. From the Lagrangian, a descent
direction for the frequencies is determined, and the frequencies are updated.
The new frequencies are used to iterate on the assignment, until the frequen-
cies converge.
Similarly, Han and Wilson (1982) formulated the problem of solving for fre-
quencies on each route as an allocation of vehicles among routes in a network.
The problem is formulated as one of solving for frequencies, with the con-
straints being the passenger assignment to individual links, the capacity of each
88 G. Desaulniers and M.D. Hickman

route, and the total fleet size. In contrast to Scheele (1980), however, the objec-
tive is to minimize the maximum “occupancy level” at the maximum load point
for each route in the network. To solve this problem, Han and Wilson (1982)
proposed a two-stage heuristic as follows. In the first stage, a base allocation
is achieved that guarantees that all routes have sufficient frequency so that all
passengers are served, but there is 100% utilization on at least one route seg-
ment for each route. In this first stage, the passenger assignment, O–D pairs
are decomposed into those with only a single path (so-called “captive” flow)
and those with multipath (“variable” flow) assignment. For the multipath as-
signment, a simple frequency-share model is adopted for both direct paths and
transfer paths. In the base allocation, the captive flow is assigned in the net-
work, and a lower bound on the frequency for each route is determined based
on setting the frequency equal to the maximum link flow divided by the ve-
hicle capacity. Second, an iterative procedure is used in which the “variable”
flow is assigned and the frequencies are updated to equal the maximum link
flow on each route divide by the vehicle capacity. This process iterates until
the “variable” flow on each route segment converges. In the second stage, any
remaining vehicles in the fleet are allocated to routes to reduce the utiliza-
tion uniformly. In the example in the paper, this is achieved by increasing the
frequency of all routes directly in proportion to the remaining vehicles in the
fleet.
Furth and Wilson (1982) presented a model to determine route headways
that maximize the consumer surplus (measured in terms of waiting time) plus
the total ridership, as a function of the headway. In this formulation, the de-
mand is a function of the headway, making the total ridership and waiting time
dependent on the headway. The formulation includes constraints on the total
subsidy, the total fleet size, and maximum headway values (as a policy device).
The problem is solved through an algorithm using the Kuhn–Tucker conditions
on a relaxation where the maximum headway and fleet size constraints are re-
laxed. Violations of the maximum headway constraint are projected back to
the maximum headway, with associated reductions in the available subsidy.
Violations of the fleet size constraint are accommodated with a new set of
Kuhn–Tucker conditions where this constraint is binding. The algorithm it-
erates through these conditions until all routes have similar multipliers. The
result is an optimal allocation of buses to routes.
A more complex model, minimizing passenger waiting cost, operating cost,
and the cost of vehicle “crowding” was introduced by Koutsopoulos et al.
(1985). In their formulation, demand is assumed to be fixed, and constraints
on the available subsidy, the maximum fleet size, and available capacity on
each route. This is formulated as a nonlinear program, which under certain
simplifying assumptions is formulated and solved as a linear program.
Recent work by Gao et al. (2004) has explored a bi-level model for de-
termining line frequencies and the corresponding network assignment. The
upper level solves for the optimal frequency of each route, minimizing the to-
tal passenger cost. This solution is then iterated with the lower-level passenger
 Ch. 2. Public Transit 89

assignment, which is based on the congested assignment model of de Cea and

Fernández (1993). The assignment is solved using a diagonalization approach.
Additional work has been done in consideration of special scheduling
cases, particularly in high-demand corridors. These include short-turning, zone
scheduling (or express services), and deadheading. In short-turning, some ve-
hicles on a route will serve only one segment of the route before returning to
the terminal. The objective is to reduce the total number of vehicles serving
a route, while still meeting passenger demand and/or minimum levels of pas-
senger service. Furth (1987) presented a model to consider short-turn design
in which the objective is to minimize the total fleet size serving a route, with
constraints that the load cannot exceed capacity at any point on either the full
route or on the short-turn segment. For a given short-turn segment, the prob-
lem is cast separately for different multiples of the full route: a 1:1 strategy
implies one vehicle on the full route for every one on the short-turn segment;
a 1:2 implies one on the full route for every two on the short-turn route, etc.
This problem is formulated and solved as a linear program with two decision
variables: the frequency on the full route; and the relative offset of the dispatch
times on the short-turn route versus the full route. The offset is used to balance
the loads on the full route and the short-turn route, depending on the loading
pattern over the common route segment. From the continuous linear program
solution, a simple rounding technique can be used to find offsets at the near-
est minute. Additional deadheading and interlining options are considered in
a heuristic technique to reduce the total fleet requirement.
The problem description by Ceder (1989) considered a variety of possible
short-turn segments on a route. As an input, a full route schedule is assumed.
With this information, Ceder (1989) presented a method to determine which
trip segments in the schedule could be eliminated by including short-turn trips.
This uses a heuristic to minimize the maximum headway for a route segment
when the short-turn trip is introduced. Once the short-turn trips are sched-
uled, a new estimate of the fleet size is found using the technique of Stern and
Ceder (1983), based on the creation of deadheading trips and interlining of
trips. Since it works with an existing (initial) route schedule, the technique re-
sults in both the definition on the short-turn segments as well as the schedule
of the short-turn trips.
A more recent investigation by Site and Filippi (1998) posed the short-
turning problem as one of determining the short-turn segment, the types of
vehicles to operate on both the full route and the short-turn segment, and the
frequency of service on both the full route and the short-turn segment. The
objective function is to maximize the net benefits, given as the passengers’ con-
sumer surplus less the net subsidy (total costs minus fare revenues) for the
operator. Costs for the operator include both capital and operating costs. In
this model, passenger demand is endogenously determined as a function of the
selected frequency. The model includes constraints to ensure demand does not
exceed capacity, and a constraint on the maximum available subsidy. The prob-
lem is formulated as a nonlinear program. Site and Filippi (1998) decomposed
90 G. Desaulniers and M.D. Hickman

the full problem into smaller subproblems; each subproblem is solved for the
optimal frequencies, for a given short-turn segment and set of vehicle types.
These subproblems are solved heuristically using random search methods, and
the global solution is found as the subproblem that maximizes the objective of
net benefits.
The zone scheduling (or express service) problem was introduced by Jordan
and Turnquist (1979) as a strategy to improve service reliability on bus routes.
In the zone scheduling concept, a bus serves only selected segments of a route
as it travels from one terminus to the other. In their simplified route model,
Jordan and Turnquist (1979) assume that all passengers are destined for a sin-
gle terminus. Their decision variables are the number of zones, the first stop in
each zone, and the number of buses allocated to serve each zone. The problem
is formulated as one of minimizing the passenger utility, comprising the ex-
pected value and variance of the waiting time and on-board travel time for all
passengers. The problem is formulated using a dynamic programming recur-
sion, in which the stages are the number of zones and the state variables are
the combination of beginning stop and the number of buses allocated to that
zone. The means and variances of waiting times and on-board running times
are formulated using a stochastic model of transit service calibrated from data
in Chicago.
The work of Furth (1986) extended the model of Jordan and Turnquist
(1979) for several additional cases: (1) zone scheduling where additional stops
outside the zone can be served for alighting (inbound) and for boarding (out-
bound); (2) zone scheduling where the zone boundaries are asymmetric, de-
pending on the direction of the trip during any given period; and (3) zone
scheduling for branching corridors. Again, dynamic programming is used to
solve these cases of the zone scheduling problem.
Finally, Furth (1985) presented a model for deadheading, in which the desir-
able headway in the off-peak direction is higher than that in the peak direction.
In providing service in the off-peak direction, vehicles not in service are dead-
head back to the initial terminal in the peak direction. This strategy, while
offering less service in the off-peak direction, may allow sufficient time savings
for fleet size reduction. The problem of determining the minimum fleet size,
for given (fixed) maximum headways in both the peak and off-peak direction,
is found by solving a maximum flow problem on a time-space network. When
the headway constraints are changed to inequalities (i.e., with only a maximum
headway), it is possible that shorter headways could yield fleet size reductions,
due to the scheduling requirements for these trips. Furth (1985) presented an
algorithm to solve for the minimum number of vehicles, using the ratio of the
number of peak trips per outbound trip in service. Related techniques are used
to solve for the optimal headways in the peak and off-peak directions for a
given fleet size, for two cases: (1) minimizing the total passenger waiting time;
and (2) minimizing a combination of passenger waiting time and operating
costs.
 Ch. 2. Public Transit 91

3.2 Timetabling

Timetabling is the process of converting the desired frequency of service on

each fixed route into a schedule. The inputs to this process include the route
structure, including running times between major timepoints, the frequency of
service, and any necessary layover times at terminals or schedule slack (extra
time built into the schedule) at the major timepoints on the route. The result
is a set of trips and the scheduled times at the terminals and major timepoints
on the route.
In most treatments of transit planning methods, timetabling is included
within operational planning, since it occurs frequently, with every service ad-
justment (e.g., every 3–6 months). Also, it is from the timetables that vehicle
and crew schedules are constructed. In this chapter, it is included as an element
of tactical planning in the sense that it involves determining the schedule to
maximize passenger service. This is in contrast to the vehicle and crew schedul-
ing problems that are typically associated with operations planning, which are
intended to minimize transit operating costs.
In many cases, the timetabling is relatively straightforward (Ceder, 1986),
primarily working on the assumption that headways are constant and that de-
mand is relatively uniform over the time period of interest. Some clock time is
specified for the first vehicle trip of each period, and vehicle departures from
a terminal or a maximum load point are set as multiples of the desired head-
way. Estimated running times between timepoints are used to determine the
schedule, and layover times are used to schedule return trips. The simplicity of
this task may explain in part why it has received relatively little attention from
researchers. In this section, some methods for timetabling are described. How-
ever, a major complication in timetabling occurs when schedules are intended
to be coordinated at a transfer stop or terminal; methods to create timetables
under these circumstances are also presented in this section.
Initial work into the timetabling problem for time-dependent passenger ar-
rivals was suggested by Newell (1971), Salzborn (1972), and Hurdle (1973a,
1973b). These works formulated the timetable problem for a single route with
the objective of minimizing passenger waiting time. These results suggest that
the optimal rate at which vehicles are dispatched is proportional to the square
root of the passenger arrival rate, but with the constraint that vehicle capac-
ity cannot be exceeded. These models were extended by Sheffi and Sugiyama
(1982) to consider multiple origin–destination pairs (complicating the deriva-
tion of the capacity constraint) and to the case of boarding-dependent dwell
times.
The work of Wirasinghe and Liu (1995) considers the problem of determin-
ing optimal schedule slack times at intermediate timepoints on a route. Given
the running time distributions on the route and a schedule-based holding pol-
icy, the research gives a model for determining the slack time by minimizing
the sum of the passenger waiting time, the passenger schedule delay, and the
92 G. Desaulniers and M.D. Hickman

operating cost. The problem is solved using first-order conditions on the ob-
jective function.
Perhaps the most pressing challenge in timetabling is the synchronization of
vehicle timetables so that transfers within the network are well timed. Specif-
ically, one would like to time the arrival of a vehicle on one route with that
on another route so that passengers transferring between routes can make the
connection with the minimum waiting time. Much of the early work on this
problem focused on methods for synchronization at a single timepoint; more
recent methods have used heuristics for larger network problems. However,
the combinatorial nature of the problem indicates that it is NP-hard, and the
computational issues of exact solutions are still vexing. This limitation is im-
portant to note, in that only the more recent approaches have considered more
practical problems.
The work of Salzborn (1980) considers two related problems. The first prob-
lem is determining the feasibility of scheduling a single transfer route through
a series of transfer locations (“interchanges”). Inputs to this analysis include
the running times on the transfer route, the slack time built into the transfer
route and all connecting services (“feeder routes”) at these interchanges, and
the minimum time required for passengers to make the transfer. In this analy-
sis, the feasibility of the schedule for the transfer route depends on the ability
of the schedule to meet the necessary time windows at each interchange. For
the second problem, Salzborn (1980) considers the scheduling of the feeder
routes at the interchange, and derives conditions under which a feasible feeder
route schedules can be constructed. In this case, the feasible scheduling of the
feeder routes requires that the headway on these routes be a multiple of the
headway on the transfer route. In addition, the scheduling of the feeder routes
requires that departures and arrivals at the interchange be balanced (the total
number of departures equals the total number of arrivals) over one headway
on the feeder routes, so that time slots at the interchange can be scheduled.
Hall (1985) considers the more specific problem of scheduling at the inter-
change when the feeder route may be delayed. Under an assumed exponential
distribution of this delay, Hall (1985) derives equations for the optimal slack
time, based on the objective of minimizing passenger delay. In this case, “slack
time” is defined as the time between the scheduled arrival on the feeder route
and the scheduled departure on the transfer route. A similar approach was
used by Knoppers and Muller (1995) to characterize the optimal slack time
on the transfer route, with a normal distribution of arrival times on the feeder
route. Knoppers and Muller (1995) also investigate the possible reductions in
average waiting time if a holding policy is used: a vehicle on the transfer route
is held until the feeder bus arrives. In this case, the optimal slack time can be
reduced, with corresponding reductions in the average waiting time. However,
the work does not examine the implications of additional waiting at down-
stream stops from the holding strategy.
There has also been a number of more analytic studies to optimize both the
slack time and the headways of connecting routes at a transfer point. Purely
 Ch. 2. Public Transit 93

analytic optimization models for a single transfer point have been developed
by a number of researchers, including Lee and Schonfeld (1991) and Chien and
Schonfeld (1998). More recently, these analytic models have been extended to
cover multiple transfer points and multiple modes. Heuristics for this prob-
lem, to solve for the slack times and any common headways among routes,
have been proposed by Chowdhury and Chien (2002) and Ting and Schonfeld
(2005). Both of these studies use a combination of operator costs, from the
added slack time, and user costs, including waiting time, transfer time, and
in-vehicle time.
A more rigorous treatment of the transfer synchronization problem was pre-
sented originally by Klemt and Stemme (1988) for a completely deterministic
problem, and later by Bookbinder and Désilets (1992) for the case of random
delay. The synchronization problem is defined as one to determine the ideal
“offsets” for the schedule for each route r in the set R. Here, an offset tr for a
given route r is defined as the minutes after some given time at which the first
departure from the terminal is scheduled. If the headway on route r is hr , the
possible offsets are assumed to be in the set of integers Tr = {0 1     hr − 1}.
Suppose the set of transfer opportunities between routes is given as the set
K = {1 2     k}, with the set Aij describing the complementary set of pairs
of routes i and j representing transfer opportunity k (i.e., at the intersection of
routes i and j). The utility of a given transfer opportunity is given as Dk (ti  tj ),
and the number of passengers making this transfer is given as nk . Then, the
optimization problem to determine the optimal offsets is given as
 
minimize nk Dk (ti  tj ) (24)
i∈R j∈R k∈Aij

subject to: ti ∈ Ti  ∀i ∈ R (25)

To solve this model, Bookbinder and Désilets (1992) use a heuristic de-
veloped by Rapp and Gehner (1976), in which the offset of each route is
determined iteratively. In Bookbinder and Désilets (1992), several utility func-
tions Dk (ti  tj ) based on the passenger waiting time are evaluated, using a
simulation model to generate random transfer arrivals. These are evaluated
initially for a single transfer connection and also two small networks with mul-
tiple transfer connections.
Ceder and Tal (1999) and Ceder et al. (2001) introduced a model to maxi-
mize the number of synchronized connections between routes. The objective is
simply to maximize the number of potential connections that can be made. In
this formulation, the decision variables include the offset of the initial trip in
the schedule (as before), and the headway of each vehicle trip is permitted to
vary from some minimum to a maximum headway. This creates more flexibility
in the construction of schedules, as each vehicle on the route may have a dif-
ferent headway. The problem is formulated as a mixed integer linear program,
and solved using a heuristic that processes the nodes sequentially, using some
selection criteria. For each node, the headways of all routes at the node are
94 G. Desaulniers and M.D. Hickman

matched if possible, and offsets of all connecting routes are set so as to create
a simultaneous arrival of all routes at the given node. The heuristic proceeds
through the set of interchanges until all vehicle departure times are set.

4 Operational planning

Given a set of timetabled trips to be operated, a set of available resources

(buses and drivers) and their distribution among the transit agency depots, the
operational planning phase aims at constructing vehicle and crew schedules
that minimize total costs while respecting all operational constraints and work
regulations. This phase also includes planning the assignment of the buses to
parking slots in garages and their dispatch to bus schedules, as well as estab-
lishing bus maintenance schedules.
As mentioned in the Introduction, these various problems are usually solved
sequentially. Figure 1 illustrates the usual solution sequence and the possible
types of feedback. The frequency setting and timetabling tactical problems de-
fine the main input for the operational planning phase, namely the timetable.
After this, the vehicle scheduling problem is solved first. This stage is crucial
to assess whether the proposed timetable can be operated with the available
fleet of vehicles and how costly it will be. When the vehicle scheduling prob-
lem is infeasible or its operating cost is excessive, a revision of the timetable
and possibly the route frequencies is performed in order to facilitate vehi-
cle scheduling. Once a vehicle schedule is determined, driver duty scheduling,
which consists of constructing anonymous work days for the drivers, is per-
formed to ensure a complete coverage of the vehicle schedule at a reasonable
cost. When this problem is not feasible or too costly, the vehicle schedule must

Fig. 1. Relationships between the tactical and operational problems.

 Ch. 2. Public Transit 95

be updated to ease the construction of the driver duties. Once the driver duties
are known, a rostering problem is solved to establish the personalized driver
schedules over a given time horizon (e.g., monthly). This stage rarely requires
one to revise the previous decisions since part-time drivers are usually avail-
able to provide additional flexibility in the rosters. In parallel to the driver
scheduling problems, the bus parking and dispatching problem as well as the
bus maintenance scheduling problem are tackled once the vehicle schedule
is known. These two problems are more or less solved on a daily basis since
they are heavily impacted by the perturbations of the planned vehicle sched-
ule. These two problems are also quite dependent.
Looking at Figure 1, one can see that feedback may frequently occur dur-
ing the operational planning process. This shows that there is an opportunity
for integrating some of these steps. In this regard, Section 4.3 discusses the
integration of vehicle and duty scheduling.

4.1 Vehicle scheduling

Vehicle scheduling plays an important role in the management of a pub-

lic transit agency since it is the first planning step where the primary focus
is put on minimizing costs, namely, the acquisition, and operational costs of
the buses. Previous steps put a large emphasis on passenger service, which is
fixed for vehicle scheduling. Indeed, at this stage, the service offered to the
customers is completely determined by the fixed trip timetable. In general, the
same daily timetable applies for the weekdays, while a different timetable is de-
fined for days on the weekend. These timetables are usually valid for a certain
period of time (a season). Thus vehicle scheduling at a planning level needs to
be performed once per timetable and season. It should be noted that in large
cities buses operate 24 hours per day, which makes it more difficult to define a
daily problem. However, given the low volume of activities during night-time,
a 24-hour timetable is split into a day timetable and a night timetable in prac-
tice.
The vehicle scheduling problem faced by the public transit agencies cor-
responds to the single-depot vehicle scheduling problem (SDVSP) when the
agency operates its fleet of buses out of a single depot, and to the multidepot
vehicle scheduling problem (MDVSP) when several depots are used or when
several vehicle types are available. This problem can be stated as follows. Let
T = {1 2     n} be a set of n timetabled trips where trip i ∈ T starts at
time si and ends at time ei . These trips are qualified as active since passengers
travel along them. Denote by τij the travel time (possibly including some lay-
over time) between the end location of trip i and the start location of trip j.
We assume that this travel time is the same for all vehicles. Two trips i and j
are said to be compatible if and only if they can be covered consecutively by
the same vehicle (j immediately follows i), that is, if and only if ei + τij  sj .
The traveling between two such trips is called a deadhead trip since there are
no passengers on board.
96 G. Desaulniers and M.D. Hickman

Let K = {n + 1 n + 2     n + m} be the set of m depots housing the buses

that must be assigned to cover the active trips. Depot k ∈ K manages vk iden-
tical buses which must start and end their schedule at this depot. A bus leaving
a depot to reach the start location of an active trip is said to be performing a
pull-out trip, while it performs a pull-in trip when it returns to the depot from
the end location of an active trip. A feasible schedule for a bus housed in de-
pot k is composed of a pull-out trip starting at k, a sequence of active trips
separated by deadhead trips, and a pull-in trip ending at k. Consecutive active
trips must be pairwise compatible.
The cost structure is as follows. A cost is incurred each time that a vehicle
performs a deadhead, pull-out or pull-in trip. This cost is denoted by cij for the
deadhead trip connecting trips i to j, by ckj for the pull-out trip linking depot k
to the start location of trip j, and cik for the pull-in trip returning to depot k
from the end location of trip i. Note that the active trips bear no cost since
they represent a fixed amount for any feasible solution. Note also that vehicle
fixed costs can be added to the pull-out or the pull-in trip costs. The cost of a
schedule is simply the sum of the costs of the trips it contains.
The MDVSP can be defined as the problem of finding a set of feasible
vehicle schedules such that each active trip i ∈ T is covered by exactly one
schedule, at most vk schedules are defined for each depot k ∈ K, and the sum
of the schedule costs is minimized. The SDVSP simply corresponds to the case
where |K| = 1. In certain versions of the MDVSP additional constraints are
considered. For instance, there may exist trip–depot compatibility constraints
which restrict the set of depots that can provide a vehicle to perform an active
trip, especially when the depots are associated with different vehicle types (see
Costa et al., 1995; Löbel, 1998). A soft version of these constraints, when they
take the form of preferences rather than strict constraints, can also be handled
by defining depot-dependent deadhead costs. Another example of additional
constraints consists of imposing a maximum duration or length to every vehi-
cle schedule (Freling and Paixão, 1995). Such a constraint may be needed in
a extra-urban context where the potential for driver exchanges is restricted or
when fueling considerations must be taken into account.
It should be noted that, in the context of public transit, a deadhead trip
that involves a long waiting time before the start of the next active trip is of-
ten replaced by a pull-in trip, an idle period at the depot, and a pull-out trip.
The vehicle schedules are then seen as sequences of vehicle blocks, where each
block consists of a sequence of trips that starts and ends at the same depot
without returning to it in the middle of the sequence.
The SDVSP arises for small to medium-size transit agencies that rely on a
single depot. It can also appear as a subproblem of the MDVSP. The SDVSP
is solvable in polynomial time. In fact, it can be modeled as a minimum-cost
network flow problem. It has also been formulated as a linear assignment prob-
lem, a transportation problem, a quasiassignment problem, and a matching
problem. Surveys on the SDVSP and its extensions can be found in Daduna
and Paixão (1995) and Desrosiers et al. (1995). Recently, Freling et al. (2001b)
 Ch. 2. Public Transit 97

proposed an efficient auction algorithm for solving the quasiassignment formu-

lation of the SDVSP. Based on this algorithm, they also developed a two-phase
approach where blocks are built first and combined afterwards, and a core-
oriented approach that starts with a network containing a subset of the arcs
(the core) and adjusts it iteratively according to the reduced cost of the arcs not
considered. The two-phase approach is only valid under certain cost assump-
tions. Computational experiments showed that these approaches outperform
the algorithms previously exposed in the literature.
The MDVSP is common in medium-size transit agencies, and inevitable in
larger ones. As proposed by Ribeiro and Soumis (1994), it can be modeled
using an integer multicommodity flow formulation as follows. Associate with
each depot k ∈ K a directed graph Gk = (N k  Ak ), where N k and Ak denote
its sets of nodes and arcs, respectively. The node set is defined by N k = T ∪{k}.
The arc set is given by Ak = C ∪ ({k} × T ) ∪ (T × {k}), where C is a subset
of T × T that contains an arc (i j) ∈ T × T if and only if trips i and j are
compatible. Then, define a binary variable Xijk for each k ∈ K and each arc
(i j) ∈ Ak that indicates the flow (0 or 1) of buses originating from depot k
on the arc (i j).
Using this notation Ribeiro and Soumis (1994) formulated the MDVSP as
follows:
 
minimize cij Xijk (26)
k∈K (ij)∈Ak
subject to:
 
Xijk = 1 ∀j ∈ T  (27)
k∈K i:(ij)∈Ak

k
Xkj  vk  ∀k ∈ K (28)
j∈T
 
Xijk − Xjik = 0 ∀k ∈ K j ∈ T ∪ {k} (29)
i:(ij)∈Ak i:(ji)∈Ak

Xijk ∈ {0 1} ∀k ∈ K (i j) ∈ Ak  (30)

The objective function (26) aims at minimizing total costs. Constraints (27)
ensure that each active trip is covered exactly once, while constraints (28) limit
the number of buses that can be used from each depot. Flow conservation and
binary constraints are given by (29) and (30), respectively.
The MDVSP has been studied for more than twenty-five years. Given that
it is an NP-hard problem when m  2 (Bertossi et al., 1987), the early work on
this problem focused on heuristic algorithms (for reviews on these methods,
see Dell’Amico et al., 1993; Odoni et al., 1994). Since the end of the 1980s,
several exact algorithms have been proposed in the literature: Carpaneto et al.
(1989), Ribeiro and Soumis (1994), Forbes et al. (1994), Bianco et al. (1994),
98 G. Desaulniers and M.D. Hickman

Löbel (1998), Mesquita and Paixão (1999), Hadjar et al. (2006), and Kliewer
et al. (2006). The results in the last three papers clearly show that real-world
large-scale instances can be solved efficiently. In the following we present the
methodologies introduced in these three papers.
The approach proposed by Löbel (1998) consists of solving the linear re-
laxation of the multicommodity network flow model (26)–(30) using a column
generation method directly on this formulation; that is, the generated variables
are the Xijk . Before starting the column generation process, a heuristic proce-
dure is used to find a feasible solution. The positive-valued variables of this
solution are added to the initial restricted master problem in order to speed up
the solution process. Then, at each column generation iteration, the restricted
master problem is solved by the dual simplex algorithm and the columns are
generated based on a so-called Lagrangian pricing strategy (discussed below)
and the standard reduced cost criterion. When the progress in the objective
value becomes too small, only the standard reduced cost criterion is used and
the restricted master problem is re-optimized by the primal simplex algorithm.
Using this methodology, the optimal linear relaxation solution found by Löbel
(1998) was already integer for most of the instances treated. When this was not
the case, a simple rounding procedure was used to derive an integer solution
that was often proved to be optimal.
Lagrangian pricing is a strategy which allows the simultaneous generation
of negative reduced cost variables and nonnegative reduced cost variables
that complement well the former set of variables. In this way, the column
generation process does not require additional iterations to identify these com-
plementary variables as is often the case with the traditional pricing strategy.
Given a Lagrangian relaxation of the linear relaxation of the model (26)–(30),
the Lagrangian pricing proposed by Löbel (1998) consists of solving the La-
grangian subproblem for a given set of multipliers, namely the values of the
dual variables associated with constraints (27)–(29) in the current restricted
master problem. All the variables taking a positive value in the solution of this
subproblem are then candidates that can be added to the restricted master
problem, even if their reduced costs are nonnegative.
Löbel (1998) considered at each iteration two Lagrangian relaxations. In
the first, the trip covering constraints (27) are relaxed in the objective function
to obtain m independent minimum-cost flow problems that are easily solvable.
For the second Lagrangian relaxation, the redundant covering constraints
 
Xijk = 1 ∀i ∈ T  (31)
k∈K j:(ij)∈Ak

are added to the formulation (26)–(30) before relaxing constraint sets (28)
and (29). The resulting Lagrangian subproblem is also a minimum-cost flow
problem that can be solved by inspection. Its solution can however produce
bus schedules containing deadhead trips assigned to different depots.
Using this column generation approach, Löbel (1998) reports solving real-
world instances from German public transit companies involving up to 49 de-
 Ch. 2. Public Transit 99

pots and 24,906 trips. It should be noted however that trip–depot compatibility
constraints were considered, yielding a maximum average of 4 depots per trip
among these large instances.
Recently, Hadjar et al. (2006) developed a branch-and-bound approach for
the MDVSP that combines column generation, variable fixing, and cutting
planes. As introduced in Ribeiro and Soumis (1994), traditional column gener-
ation is used to compute a lower bound at each node of the branch-and-bound
search tree. This column generation process is executed on a set partitioning
type reformulation of the MDVSP that can be derived from model (26)–(30)
by applying the Dantzig–Wolfe decomposition principle (Dantzig and Wolfe,
1960). In contrast to Löbel (1998) approach, columns in this set partitioning
model correspond to vehicle schedules. They are generated by solving shortest
path problems.
The variable fixing strategy used by Hadjar et al. (2006) is similar to the
one developed by Bianco et al. (1994) and consists of fixing to zero the vari-
ables Xijk that satisfy the following criterion. To simplify notation, we rewrite
η η
model (26)–(30) as min{cx | Ax = b x ∈ Z+ }, where x = (xi )i=1 is a vector

of η (= k∈K |Ak |) variables, Z+ is the set of nonnegative integers, and the
equality Ax = b is, in fact, an inequality (Ax  b) for constraint set (28).
Denoting by x̄ a feasible solution to this problem and by π̄ a feasible solution
to the dual of its linear relaxation, a variable xi can be set to zero if its reduced
cost is greater than or equal to c x̄ − π̄b. Hadjar et al. (2006) computed a first
feasible solution x̄ by performing a depth-first search without backtracking in
the branch-and-bound tree and imposing multiple decisions at each branch-
ing node. At each node of the branching tree, part of the dual solution π̄ is
provided by the dual solution produced by the column generation method at
this node and the remainder is found by solving shortest path problems. This
variable fixing strategy, which is performed at each node of the tree, can fix
over 90% of the Xijk variables in most instances treated by Hadjar et al. (2006).
Hadjar et al. (2006) proposed to add at each node of the search tree cutting
planes that are related to the odd cycles in the MDVSP underlying network.
These valid inequalities are lifted through a heuristic procedure. The authors
showed that, under certain conditions, the lifted inequalities define facets of
the convex hull of the feasible solution set. With this approach, Hadjar et al.
(2006) succeeded in solving randomly generated MDVSP instances that in-
volve up to 6 depots and 750 trips. It should, however, be mentioned that these
results are difficult to compare with the results obtained by Löbel (1998) or
Kliewer et al. (2006) (see below) because the characteristics of the test prob-
lems differ significantly from one paper to the other.
Instead of using model (26)–(30), Kliewer et al. (2006) developed a multi-
commodity network flow model based on a time-space network. In fact, when
several trips start and end at the same terminus, a substantial reduction in the
number of variables can be obtained by using a sequence of waiting arcs at each
terminus instead of representing explicitly all possible connections. Kliewer
100 G. Desaulniers and M.D. Hickman

et al. (2006) also applied an aggregation procedure for reducing the number
of arcs representing potential deadhead trips. The resulting model is solved to
optimality with the CPLEX MIP solver. With this approach, the authors re-
port solving large real-world instances. In particular, they solved one instance
that involves 7068 trips, 5 depots, and 124 termini in approximately 3 hours of
computational time.
An interesting extension to the MDVSP is the possibility of changing the
scheduled departure times of the trips within certain time intervals, called time
windows. Such flexibility on departure times, which can be considered when
the frequency on a route is not too high, can often yield significant savings by
providing additional possible deadhead trips that would be infeasible other-
wise. To our knowledge, this extension has been tackled first by Mingozzi et al.
(1995) who adapted the methodology they have developed for the MDVSP in
Bianco et al. (1994). Their approach consists of solving by branch-and-bound a
set partitioning model that contains a reduced set of columns. The reduction is
obtained by variable fixing as described above. More recently, Desaulniers et
al. (1998b) have proposed a branch-and-price approach for this extension that
generalizes the work of Ribeiro and Soumis (1994). They also showed that,
with a slight modification, this approach is capable of handling an exact cost
on the waiting time occurring between two consecutive trips. Given the time
windows, such waiting times and their ensuing waiting costs cannot be com-
puted a priori; they must be computed during the solution process.

4.2 Duty scheduling

Duty scheduling, also known as driver scheduling, is the second step in the
process of planning the operations for a public transit agency. As with the vehi-
cle scheduling problem, the duty scheduling problem (DSP) is important from
an economic point of view since it determines most of the wages paid to the
drivers. The DSP is separable by depot and consists of determining the work
days (also called duties) of the drivers based at a depot in order to cover all
the vehicle blocks assigned to this depot. Since a driver exchange can occur
at various points along a vehicle block, all blocks are divided into a sequence
of segments according to these relief points. The consecutive segments along a
block assigned to the same driver are collectively called a piece of work. Duties
are therefore composed of pieces of work that are usually separated by breaks.
Different duty types, differing, for instance, by the number of pieces of work
they can contain and their possible starting times and durations, can be consid-
ered. As examples, there may exist straight duties that contain a single piece
of work, and split duties containing two pieces of work. Duties are subject to a
wide variety of safety regulations and collective agreement rules such as a max-
imum duty spread, a maximum duration of a piece of work, and a predefined
time interval in which a break must be awarded. These rules vary according to
the duty type.
 Ch. 2. Public Transit 101

In general, the objective of the DSP is twofold and consists of minimizing

first the total number of duties and second the total number of worked hours.
Using duty fixed costs and an hourly rate for the worked hours, this objective
is usually transformed into one that minimizes total cost. In summary, the DSP
can be stated as follows. Given the segments of a set of vehicle blocks, find
a set of valid duties that covers all these segments and minimizes total cost.
Additional constraints such as a limit on the number of duties of a certain type
can also be taken into account.
The DSP can be formulated as a set partitioning problem that relies on the
following notation. Let S be the set of block segments to cover and D the set of
valid duties. Denote by cd the cost of duty d (including fixed costs and wages)
and by asd a binary parameter that takes value 1 if duty d covers segment s,
and 0 otherwise. Finally, define a binary variable Yd for each duty d ∈ D that
indicates if duty d is retained in the solution. Using this notation the DSP is
formulated as

minimize cd Yd (32)
d∈D
subject to:

asd Yd = 1 ∀s ∈ S (33)
d∈D
Yd ∈ {0 1} ∀d ∈ D (34)
The objective function (32) consists of minimizing total duty costs. Con-
straint set (33) ensures that a driver is assigned to each block segment. Finally,
binary requirements on the Yd variables are expressed by (34). It should be
noted that all work rules defining the validity of the duties are taking into
account in the definition of set D. In some cases, this set or part of it can be enu-
merated a priori using an enumeration algorithm that considers these rules (for
instance, see Smith and Wren, 1988). Otherwise, it can be defined implicitly as
a set of constrained paths in one or several networks, where the constraints
are used to model the complex work rules (for instance, see Desrochers and
Soumis, 1989).
Several authors formulate the DSP as a set covering model that allows the
over-covering of each segment (the equalities in (33) are replaced by greater-
than-or-equal-to inequalities). This over-covering is usually not acceptable, but
solving this model often produces a solution that contains very little or no over-
covering at all, especially when assigning one driver to a segment is cheaper
than assigning several drivers to it. In this case, over-covering can usually be
easily eliminated a posteriori using a heuristic procedure. The main advantage
of a set covering model over a set partitioning model is its flexibility which
allows more rapid computation of a feasible continuous solution and good
heuristic integer solutions.
The DSP, modeled as a set partitioning or a set covering problem, is much
more difficult to solve than the MDVSP due to the complexity of the work
102 G. Desaulniers and M.D. Hickman

rules and the huge number of duties that are valid in real-world DSP instances.
Indeed, since some of these rules can only be modeled using nonlinear rela-
tionships, one has to rely on a formulation similar to (32)–(34) to avoid explicit
nonlinear constraints. This formulation however contains a huge number of
variables in practice, making it very difficult to solve to optimality.
As surveyed in Wren and Rousseau (1995), several heuristic approaches
were proposed before the 1990s. One of the most successful consists essentially
of generating a priori a subset of the valid duties and solving a set partition-
ing/covering model (32)–(34) restricted to this subset (for instance, see Smith
and Wren, 1988). The subset of valid duties is composed of promising duties
that offer various possibilities for covering each block segment. The restricted
model is generally solved by a heuristic integer linear programming method.
Research on this type of approach is still ongoing. In 1999, an attempt was
made by Curtis et al. (1999) to solve the restricted set partitioning model using
a hybrid constraint programming/linear programming heuristic method where
the linear programming solutions are used to guide variable and value ordering
in the constraint programming algorithm. One nice feature of this method-
ology is that its constraint programming component is capable of handling
nonlinear constraints which can arise from certain work rules. However, the
authors could not solve instances involving more than 203 segments and 26 du-
ties. More recently, Fores et al. (2002) have incorporated a column generation
strategy to the solution process of the restricted set covering model. Column
generation, which is applied only at the root node of the branch-and-bound
search tree, consists of generating as needed negative reduced cost columns
from a superset of a priori enumerated valid duties. This novelty improves the
quality of the solutions while slightly increasing solution times. For instance,
they can solve medium-size instances involving close to 90 duties in one hour
of computational time.
Column generation for the DSP was introduced by Desrochers and Soumis
(1989). In their approach the master problem corresponds to the linear re-
laxation of the set covering model. The subproblems are constrained shortest
path problems which are solved by a generalized version of the dynamic pro-
gramming algorithm of Desrochers and Soumis (1988) (see Desaulniers et al.,
1998a; Irnich and Desaulniers, 2005). Integer solutions are found by a branch-
and-bound scheme, where column generation is used at each node of the
search tree to compute a lower bound. The overall approach obtains optimal
solutions for small instances and near-optimal solutions for larger instances. In
1995, Desrosiers and Rousseau (1995) reported solving DSP instances involv-
ing 156 duties using a commercial version of this branch-and-price approach.
In the last fifteen years, branch-and-price approaches have been applied
with success to a wide variety of vehicle routing and crew scheduling problems.
When the size of these problems are huge as in real-world DSPs, accelerating
strategies, various heuristics, and stabilization techniques, such as the ones re-
ported in Barnhart et al. (1998), du Merle et al. (1999), and Desaulniers et al.
(2002), must be included in these approaches to produce good quality solutions
 Ch. 2. Public Transit 103

in acceptable computational times. In a recent paper on the DSP, Borndörfer

et al. (2003) pursued this direction by presenting a heuristic branch-and-price
approach that calls upon several speed-up strategies. Firstly, instead of solving
the master problem by the simplex or barrier algorithm, they proposed to solve
its dual using a heuristic coordinate ascent method combined with a boxstep
stabilization method. Secondly, the constrained shortest path subproblems are
solved by an enumerative algorithm that relies on lower bounds computed by
Lagrangian relaxation. Finally, they suggested a heuristic branching scheme
without backtracking that fixes at each branching node a duty variable Yd
to one. This variable is selected from a set of twenty candidate variables us-
ing a probing strategy. For each candidate variable, this strategy fixes it to
one and evaluates the impact of this decision on the master problem solu-
tion value without generating additional columns. The selected variable is the
one yielding the smallest deterioration of the master problem solution value.
After branching, when this deterioration is less than a predetermined thresh-
old value, no columns are generated and variable fixing is performed again.
Using this customized column generation approach, Borndörfer et al. (2003)
solved large real-world DSP instances involving close to 2000 segments and
over 110 duties.
Freling et al. (1999, 2003) proposed to solve the linear relaxation of the DSP
by a column generation approach where the master problem is approximately
solved by Lagrangian relaxation. At each column generation iteration, a con-
strained shortest path subproblem is solved in two main steps: first, pieces of
work are generated by solving an all-pairs shortest path problem; and second,
duties are generated from these pieces of work by solving a constrained short-
est path problem. To avoid generating the same column twice, the Lagrange
multipliers are modified before generating new columns in such a way that all
columns already generated get a nonnegative reduced cost. Once the linear re-
laxation is solved, a set covering problem involving all columns generated along
the way is heuristically solved to find a feasible integer solution. As reported
in Freling et al. (2003), this approach can easily solve small-size DSP instances
with 238 segments and 24 duties. No results on larger instances are reported,
however, since the DSP was solved only for comparison with an integrated ve-
hicle and crew scheduling approach that is discussed in the next section.
One drawback of the column generation approach that relies on constrained
shortest path subproblems is the inability to model all work rules that may de-
fine the validity of a duty. One way of overcoming this drawback is to ignore
these rules at the subproblem level and simply to reject all generated columns
that violate them. Another way that was suggested by Borndörfer et al. (2003)
consists of defining an infeasible path constraint for each identified illegal col-
umn and including these constraints in the subproblems. A third alternative has
been proposed by de Silva (2001), who suggested formulating the subproblems
as flexible constraint programming models and solving them using constraint
programming tools. He succeeded in solving real-world DSP instances with
complex work rules involving up to 495 segments.
104 G. Desaulniers and M.D. Hickman

Recently, Gintner et al. (2004) proposed a DSP approach that benefits from
the fact that several vehicle schedules may be optimal. Indeed, given a feasible
vehicle schedule, the segments derived from the blocks of this schedule can
be rearranged into different blocks yielding the same vehicle costs. Their DSP
model allows for this possibility by enumerating all feasible pieces of work and
duties that can be obtained from the segments. Their approach solves the lin-
ear relaxation of this model using column generation and computes, using the
CPLEX MIP solver, an integer solution for the restricted set covering model
containing only the columns generated. This approach was tested on random
datasets involving up to 400 trips with one segment per trip. These tests showed
that substantial savings in the number of duties can be achieved from this ad-
ditional flexibility.
Research investigating the use of metaheuristics for solving the DSP has
also been carried out recently. Kwan et al. (1999, 2001) proposed a genetic al-
gorithm that relies on the linear relaxation solution of a restricted set covering
model to identify important traits that should appear in the optimal integer so-
lution. Shen and Kwan (2001) developed a tabu search approach that involves
multiple neighborhoods and an appropriate memory scheme. Lourenço et al.
(2001) introduced a genetic and a tabu search algorithm to solve DSPs involv-
ing multiple objectives such as minimizing the number of duties, minimizing
the number of duties with a single piece of work, minimizing the number of ve-
hicle changes, and minimizing the over-covering when allowed. Both of these
algorithms use for large instances a greedy randomized adaptive search pro-
cedure (GRASP) as an intensification tool. All these metaheuristics are fast
and produce solutions that are comparable (in terms of quality) to the solu-
tions produced by an approach based on a restricted set partitioning/covering
model, similar to that of Smith and Wren (1988). For instance, Shen and Kwan
(2001) reported solving a DSP involving 859 segments and 106 duties in less
than 18 minutes with their tabu search algorithm.

4.3 Integrated vehicle and duty scheduling

In general, vehicle scheduling is performed before duty scheduling in the

operational planning process of a public transit agency. Since driver relief op-
portunities are numerous in most contexts, an efficient duty schedule can often
be obtained from a near-optimal bus schedule to yield an overall high-quality
solution. On the other hand, when these relief opportunities are rare, as is
the case in extra-urban mass transit systems or for a line-by-line scheduling
process, a very efficient vehicle schedule may lead to a poor duty schedule or
even to an infeasible DSP. Integrating vehicle scheduling and duty scheduling
is therefore essential in these situations, and research on this topic has been
conducted recently.
The integrated vehicle and duty scheduling problem (IVDSP) can be stated
as follows. Given a set of timetabled trips and a fleet of vehicles assigned to
several depots, find minimum-cost vehicle blocks and valid driver duties such
 Ch. 2. Public Transit 105

that each active trip is covered by one block, each active trip segment is covered
by one duty, and each deadhead, pull-in, and pull-out trip (hereafter called an
inactive trip) used in the vehicle schedule is also covered by one duty. As in
the MDVSP, each block must start and end at the same depot and, as in the
DSP, driver duties must comply with a set of work rules and each duty must be
composed of trips that are covered by buses originating from the same depot.
This last requirement is often mandatory since drivers are usually assigned to a
depot. Additional constraints such as vehicle availability can also be imposed.
Next, we propose a formulation for the IVDSP that combines the models
presented above for the MDVSP and the DSP. Besides the notation introduced
in the previous two sections, the following notation is required. Let Dk be
the set of valid duties for a driver assigned to depot k, and bdij be a binary
parameter equal to 1 if duty d ∈ Dk covers the trip associated with arc (i j) ∈
Ak and to 0 otherwise. For each depot k ∈ K and each duty d ∈ Dk , we define
a binary variable Ydk that takes the value 1 if duty d is selected and the value 0
otherwise.
The proposed formulation for the IVDSP is as follows:
   
minimize cij Xijk + cd Ydk (35)
k∈K (ij)∈Ak k∈K d∈Dk
subject to:
 
Xijk = 1 ∀j ∈ T  (36)
k∈K i:(ij)∈Ak

k
Xkj  vk  ∀k ∈ K (37)
j∈T
 
Xijk − Xjik = 0 ∀k ∈ K j ∈ T ∪ {k} (38)
i:(ij)∈Ak i:(ji)∈Ak

Xijk ∈ {0 1} ∀k ∈ K (i j) ∈ Ak  (39)

 
asd Ydk = 1 ∀s ∈ S (40)
k∈K d∈Dk

bdij Ydk − Xijk = 0 ∀k ∈ K (i j) ∈ Ak  (41)
d∈Dk
Ydk ∈ {0 1} ∀k ∈ K d ∈ Dk  (42)
The objective (35) minimizes the sum of the vehicle and duty costs. Con-
straints (36)–(39) define the vehicle scheduling problem. They are identical
to (27)–(30). Constraint sets (40) and (42) are the counterparts of (33) and (34)
for the multidepot case. Finally, constraints (41) establish the link between the
vehicle schedule and the duty schedule; that is, each inactive trip covered by a
bus must also be covered by a duty assigned to the depot from which this bus
originates.
106 G. Desaulniers and M.D. Hickman

In comparison with the DSP, the IVDSP is highly combinatorial since the
inactive trips are unknown a priori; they have to be determined by the opti-
mization process. In consequence, the number of possible valid duties is very
large especially when multiple depots are considered. Given its complexity and
its lesser importance, the IVDSP has not been addressed in the literature as
much as the MDVSP and the DSP. Indeed, as surveyed in Freling et al. (1999),
only a few heuristic approaches have been proposed in the 1980s and the early
1990s. However, it seems that this problem has lately attracted the attention of
several researchers who developed solution approaches based on mathemati-
cal programming decomposition techniques.
Freling et al. (1999, 2003) addressed the single-depot IVDSP where no bus
availability constraints are considered but the main objective consists of min-
imizing the overall number of buses and duties required to cover all active
trips. Similar to the approach they proposed for the DSP, they developed a col-
umn generation approach where the master problem is solved by Lagrangian
relaxation. In this case, all constraints involving duty variables are relaxed in
the Lagrangian function, yielding a Lagrangian subproblem that corresponds
to pricing out the duty variables and solving a single-depot vehicle schedul-
ing problem. Thus, a feasible vehicle schedule is computed each time that the
Lagrangian subproblem is solved. When the linear relaxation of the IVDSP is
satisfactorily solved using this process, the last computed vehicle schedule is
kept and used to define a DSP that is solved by their DSP column generation
approach (see Section 4.2) to derive a feasible duty schedule. In Freling et al.
(2003), the authors report solving real-world IDVSP instances involving up to
148 segments and 23 duties in reasonable computation times. Their results also
show that small gains in the total number of buses and duties can be attained by
solving the IVDSP instead of solving the vehicle scheduling problem and the
duty scheduling problem sequentially. These gains are more substantial when
drivers are not allowed to change buses after a break (see also Freling et al.,
2001a).
In 2001, Haase et al. (2001) introduced a formulation that only involves
duty variables and one bus counter variable which is used to apply a fixed cost
per bus. This model can be partially derived from model (35)–(42), adapted
to the single-depot case and without availability constraints, by substituting
the X variables according to their definition in constraint set (41). Bus-count
constraints, similar to the plane-count constraints of Klabjan et al. (2002), are
added to complete the model. These constraints provide lower bounds on the
number of buses required at specific times of the horizon, namely each time
that a bus can leave the depot to reach just in time the beginning location of an
active trip. Solving this model provides optimal duties and ensures that an op-
timal vehicle schedule can be obtained a posteriori using a simple polynomial-
time procedure. To do so, Haase et al. (2001) proposed a branch-and-price
approach that relies on several accelerating strategies such as dynamically gen-
erating the bus-count constraints and reducing the average number of nonzero
elements in the constraint coefficient matrix by an appropriate constraint sub-
 Ch. 2. Public Transit 107

stitution. Two versions of this approach are presented: an exact version where
branching is performed at the subproblem level, and a heuristic version where
multiple branching decisions on the duty variables are made at every branch-
ing node. With the exact version of the algorithm, randomly generated IVDSP
instances involving up to 400 segments and 60 duties were solved in less than
3 hours, while the heuristic version succeeded in solving instances with 700 seg-
ments and 121 duties within the same time frame.
Recently, Elhallaoui et al. (2005) developed a dynamic constraint aggrega-
tion algorithm for speeding up the solution process of set partitioning type
problems solved by a column generation approach. This exact algorithm ag-
gregates and disaggregates, as needed, the set partitioning constraints in order
to reduce the size of the master problem and degeneracy. They tested this
new approach on the single-depot IVDSP instances of Haase et al. (2001).
They report reducing the time needed for solving the linear relaxation by up
to 80% on instances involving up to 1280 segments. Furthermore, they ob-
served that the number of fractional-valued variables in a linear relaxation
solution decreases considerably with this methodology, yielding high expec-
tations to compute rapidly optimal integer solutions.
The multidepot version of the IVDSP has been investigated in Huisman
et al. (2005), where the authors presented two formulations for this problem
that are generalizations of the single-depot models developed in Freling et al.
(2003) and in Haase et al. (2001). Hereafter, we refer to these formulations
as the MD-FHW model and the MD-HDD model, respectively. Two similar
solution approaches, that are adaptations of the approach proposed for the
single-depot case in Freling et al. (2003), are also proposed. Both approaches
contain two phases: the first phase computes a lower bound on the optimal
value, while the second one finds a feasible solution. The lower bound is com-
puted by approximately solving a linear relaxation using a combined column
generation/Lagrangian relaxation method. The first approach relies on the lin-
ear relaxation of the MD-FHW model while the second one uses that of the
MD-HDD model. The second approach also includes a special treatment of
the bus-count constraints which are added one at a time. The second phase is
identical for both approaches. A heuristic feasible vehicle schedule is found by
applying Lagrangian relaxation on the MD-FHW model, where only the duty
variables generated during the first phase are considered. Once this schedule
is established, a duty schedule is computed for each depot using the DSP ap-
proach proposed in Freling et al. (1999, 2003). A series of comparative tests on
real-life and randomly generated datasets involving up to 653 segments showed
that both integrated approaches can solve these instances to yield substantial
savings when compared to the traditional bus-first, duty-second sequential ap-
proach. Furthermore, neither of the integrated approaches could clearly out-
perform the other one, even though the second one regularly provided weaker
lower bounds than those produced by the first approach. To reduce solution
times and solve larger instances, de Groot and Huisman (2004) devised and
108 G. Desaulniers and M.D. Hickman

compared different heuristic strategies for splitting an instance into smaller

ones which are thereafter solved individually by an integrated approach.
For the multidepot IVDSP, Borndörfer et al. (2004) used an integer pro-
gramming formulation that essentially combines together model (26)–(30) for
the MDVSP and model (32)–(34) for the DSP and adds synchronization con-
straints between the buses and the drivers on the deadhead, pull-in and pull-
out trips. They proposed a heuristic solution approach based on a Lagrangian
relaxation of these synchronization constraints. The Lagrangian dual is solved
by a proximal bundle method and integer solutions are obtained through a
heuristic branch-and-bound procedure. With this approach, they report solv-
ing large real-world instances.

4.4 Crew rostering

Given a set of anonymous duties defined over a certain time horizon (typ-
ically, a week or a month) for the drivers assigned to a particular depot, crew
rostering consists of assigning these duties to the available drivers to form their
work schedules (called rosters). As with the duties, the validity of the rosters is
restricted by safety regulations and collective agreement rules. For instance,
a driver cannot work more than a certain number of consecutive days. In most
North American public transit agencies, drivers build their own rosters in or-
der of seniority, leaving no place for optimization. On the other hand, in many
European agencies, the main objective of the crew rostering problem is to
distribute the work load evenly among the drivers, yielding an interesting opti-
mization problem.
As surveyed in Odoni et al. (1994), the common practice for solving transit
rostering problems consists of first solving a sequence of assignment problems
to build an initial solution and then using a local improvement procedure to
better this solution. In the first phase, for each day of the horizon, an assign-
ment problem is defined to assign the duties of the corresponding day to the
partial rosters that were built by the previous assignment problems. The cost
structure aims at balancing the workload among the drivers. It can also incor-
porate bonuses to account for the preferences of the drivers for certain duties.
An iteration of the second-phase heuristic procedure can be, for example, to
select a day, divide all the rosters into two parts according to that day, and
solve an assignment problem to match the first parts of the current rosters with
possibly different second parts. Such heuristic approaches were developed in
the mid-1980s and are still in use due to their computational speed and their
flexibility with regards to the work rules.
From a mathematical programming point of view, the crew rostering prob-
lem can be formulated as a set partitioning or a set covering problem where
a row is defined for each duty and a column is associated with each valid
roster. Solving such a model is however not popular for crew rostering prob-
lems encountered in public transit systems. This is in contrast to the air and
rail contexts where various mathematical programming approaches based on a
 Ch. 2. Public Transit 109

set partitioning/covering formulation of the crew rostering problem have been

proposed lately in the literature. One major difference between transit and
air/rail rostering problems appears to be in the size of the real-world instances,
which is larger for transit problems. Indeed, the transit problems are not sepa-
rable per vehicle type since the drivers are usually allowed to drive all the buses.
Furthermore, in these problems, the tasks to cover correspond to individual
duties, while they correspond to tours of duties (also known as pairings) that
may span up to six days in air/rail rostering problems. Nevertheless, we think
that most methodological advances in air/rail crew rostering can be adapted
for public transit rostering (at least for small- and medium-size problems). We
thus refer the interested reader to Chapters 1 and 2 of this book for a review
of the latest advances in air and rail crew rostering.

4.5 Parking and dispatching

An operational planning problem that has attracted little attention in the

literature is the management of the parking area in vehicle depots. In con-
gested cities, depots are often restrained in space and quite crowded from late
evening to early morning. They also contain different types of buses that are
needed by particular bus routes. Therefore, when a bus of a particular type has
to leave the depot in the morning, several other buses might need to be moved
to clear the way out, resulting in a delay. Two alternatives can be considered to
avoid delays. The first one consists of always assigning a directly accessible bus
to every morning pull-out even if the bus type is not the one requested for the
corresponding bus schedule. In this case, we say that a mismatch occurs. The
second alternative is to reorder the buses during the night so that they all are
properly positioned for the morning pull-outs. An exchange of parking slots
between two buses is called a crossing or a maneuver.
Given a sequence of bus arrivals during the evening, a set of timetabled
pull-outs in the morning, and a required bus type for each pull-out, the ve-
hicle parking and dispatching problem consists of parking the buses in the
depot upon their arrival and dispatching them to the pull-outs such that the
number of mismatches or crossings is minimized while satisfying the following
constraints. For safety reasons, buses are not allowed to go backwards in the
depot. Therefore, assuming that the depot is made up of lanes that operate as
queues, buses enter the lanes at one end and exit them at the other. Obviously,
lane capacity must not be exceeded. Finally, given the limited space available
to perform crossings, they are only permitted between vehicles of the same
lane.
The vehicle parking and dispatching problem is an operational planning
problem that usually needs to be solved on a daily basis due to the high
variability of the bus availability per type. This variability arises from regular
maintenance requirements and unexpected breakdowns. To our knowledge,
Winter and Zimmermann (2000) were the first to introduce this problem,
110 G. Desaulniers and M.D. Hickman

which was defined for tram operations. They showed that it is an NP-hard prob-
lem and formulated it as a quadratic assignment model with side constraints.
By linearizing this model and adding valid inequalities, they could only solve
small-size instances to optimality using the CPLEX MIP solver. Consequently,
they proposed heuristics for solving larger instances.
In 2001, Gallo and Di Miele (2001) addressed the vehicle parking and
dispatching problem in the context of mass transit buses. They proposed an
integer programming model, suitable for both objectives stated above, that
relies on three variable types: a first type to assign arriving buses to lanes,
a second type to assign morning pull-outs to lanes, and a third type to iden-
tify the matchings between the buses and the pull-outs. To solve this model,
they developed a three-step heuristic approach. In the first step, Lagrangian
decomposition is applied to fix the values of the first two types of variables.
In this decomposition, these two types of variables are duplicated to yield two
generalized assignment subproblems (one for the arrivals and the other for
the pull-outs) that are solved by the CPLEX MIP solver, and a set of design
noncrossing matching subproblems (one for each lane) which can be solved in
polynomial time. In this context, a design noncrossing matching problem con-
sists of matching arriving buses with morning pull-outs with no crossings while
selecting, as design decisions, the subsets of buses and pull-outs to consider
in the associated lane. A bundle method is used for solving the Lagrangian
dual problem. In the second step, after assigning the buses and pull-outs to
the lanes according to the last computed solutions in the first step, a simplified
design noncrossing matching problem (the design decisions are fixed) is solved
for each lane to obtain a complete solution that may contain undesirable cross-
ings or mismatches. A heuristic procedure is then invoked in the third step in
an attempt to improve this solution. Using this three-step approach, Gallo and
Di Miele (2001) reported solving real-life instances involving up to 4 bus types,
12 lanes, and 77 buses in a few minutes.
Very recently, Hamdouni et al. (2006) argued that an optimal solution
to the vehicle parking and dispatching problem, as stated in Winter and
Zimmermann (2000) and in Gallo and Di Miele (2001), may be difficult to
use in practice due to the randomness of the bus arrival times. Indeed, such a
solution may contain a large number of pairs of consecutive slots in a lane to
which buses of different types are assigned. Each such pair of slots is likely to
lead to a mismatch during the operations if the buses planned for these slots ar-
rive in reverse order. In order to increase the solution robustness, Hamdouni
et al. (2006) proposed a restricted definition for the vehicle parking and dis-
patching problem in which a lane can contain a maximum of two bus types,
each bus type being confined to a single block of consecutive parking slots. In
this case, a maximum of one pair of consecutive slots per lane is susceptible to
mismatches. They also suggest replacing the objective of minimizing the num-
ber of crossings by the objective of minimizing the number of lanes that need
to be reordered. This suggestion is motivated by the fact that all the buses in
a lane must be moved out of the depot when a crossing must be performed
 Ch. 2. Public Transit 111

in that lane. Therefore, performing crossings in a lane costs approximately the

same independently of the number of crossings to perform. For this version
of the problem, Hamdouni et al. (2006) have presented an integer program-
ming model that is based on an enumeration of the possible patterns that can
be used to divide a lane into a maximum of two blocks. This model is solved
by the CPLEX MIP solver after adding a series of cuts that reduce the fea-
sible region without hindering the search for an optimal solution. Real-world
instances involving up to 4 bus types, 16 lanes, and 144 buses were solved to
optimality in less than twenty seconds of computation time.

4.6 Maintenance scheduling

Another area where operations research can be helpful in planning public

transit operations is bus maintenance scheduling. Since maintenance costs are
one of the largest expense categories in a typical transit system, some transit
agencies are now investing in maintenance scheduling systems to help them re-
duce maintenance costs while maintaining a reliable, safe, and attractive transit
system. Unfortunately, maintenance systems are not applicable everywhere.
When parking depots have limited space, bus assignment is performed on a
daily basis according to the day’s parking pattern and it is impossible to pre-
dict how many miles each bus will travel on the following days. In this case,
buses are simply withdrawn from the fleet when they are due for maintenance
and maintenance resources are available. These last-minute withdrawals can
often reduce service quality. On the other hand, for spacious parking depots
where almost all parking slots are directly accessible at any time, it might be
desirable to assign buses to specific vehicle schedules that are valid for a long
period of time. The rationale behind this strategy is that drivers can then be
assigned to the same bus every day in hope that they will be more sensitive
to mechanical anomalies of their bus and report minor problems before they
become major ones. In this context where vehicle schedules are fixed for a
long time horizon, a maintenance system can be devised for scheduling main-
tenance activities when buses are not supposed to be in service, and in such a
way to maximize maintenance resource utilization.
As stated in Haghani and Shafahi (2002), the bus maintenance scheduling
problem can be defined as follows. Given the buses’ operating schedules, their
maintenance requirements, and maintenance resource and crew availability,
schedule buses for maintenance and assign them to existing facilities such that
each bus is maintained in time, while the amount of time that the buses are out
of service is minimized. Maintenance requirements, involving various types of
maintenance, are expressed in terms of a maximum mileage or number of days
in between two consecutive maintenance activities. Note that maintenance fa-
cilities cannot all be used for all types of maintenance.
To our knowledge, the bus maintenance scheduling problem has only been
addressed by Haghani and Shafahi (2002). These authors presented three inte-
ger programming models for this problem. The first model is very general but
112 G. Desaulniers and M.D. Hickman

unsolvable in practice. The second one relies on the assumption that the buses
requiring maintenance for each type are identified and sorted in the order they
should be maintained. This assumption often can be held in practice. The third
model is a network model with side constraints that also relies on assumptions
that are usually valid in small to medium-size agencies, namely: regular inspec-
tions can be performed in all maintenance bays; and, when a bus is due for
more than one inspection in a planning period, it is possible to compute for
these inspections nonoverlapping maintenance intervals in which these activ-
ities will be scheduled. Haghani and Shafahi (2002) proposed three heuristic
approaches for solving the second model and a fourth heuristic for the third
model. The first two heuristics are simple branch-and-bound methods, while
the third heuristic fixes a large number of variables to zero based on the linear
relaxation solution, before solving the resulting problem by an exact branch-
and-bound scheme. Finally, the fourth heuristic is a network-based algorithm.
Using each of these four approaches, Haghani and Shafahi (2002) solved a
series of instances arising from simulated operations involving 181 buses and
five maintenance types. The results show that all heuristics can solve these in-
stances in less than five minutes on average to yield acceptable solutions. As
mentioned by these authors, future research on this problem can focus on de-
veloping better heuristic procedures as well as exact solution approaches based
on decomposition methods.

5 Real-time control

In actual operations, a wide variety of exogenous and endogenous factors

can affect service delivery, such as weather, incidents, variations in traffic con-
ditions, vehicle breakdowns, etc. These factors may degrade the level of service
experienced by transit passengers. For this discussion, we differentiate minor
and major service disruptions: minor disruptions are those that create small
perturbations from the schedule (e.g., 5–10 minutes), and major disruptions
cause longer breaks in the schedule. The distinction is made in order to differ-
entiate the typical responses to these service problems.
To address these challenges, a transit operator may employ a variety of
operations control techniques (Turnquist, 1981; Levinson, 1991). These will
generally vary depending on the magnitude of the perturbation to service. In
normal service with only minor perturbations from the schedule and small
service disruptions, vehicle holding and transit signal priority are the most com-
mon techniques that are applied. In holding, a vehicle may be held at a stop to
improve passenger service. The first part of this section (Section 5.1) addresses
the vehicle holding problem. For transit signal priority, transit vehicles may be
given preferential treatment in signal timing in order to move more rapidly
through a signalized intersection. For reasons of scope, signal priority is not
discussed in this chapter.
 Ch. 2. Public Transit 113

When major service disruptions occur, more serious control measures may
be considered. These can include skipping stops on a route, including expressing
over parts of the route or skipping particular stops, in order to catch up on the
schedule. It may also be useful to re-position a vehicle on the route. In short-
turning, a vehicle on a route is emptied and placed in service traveling in the
other direction on the route, in order to accommodate passenger demand in
the other direction. Real-time deadheading may also be employed to re-locate
a vehicle to another part of the route (or to another route entirely) where it
may be of greater service. In addition, extra vehicles and drivers can be made
available, to be inserted into service as the need arises. Models and methods
for these types of control measures are outlined in Section 5.2.

5.1 Vehicle holding

The transit vehicle holding problem has been explored by many researchers
over the past 30 years. Early approaches to this problem have generally focused
on either threshold-based holding or schedule-based holding. The threshold
technique involves holding a vehicle only if the preceding headway is below
a certain amount of time (e.g., the desired headway or some other thresh-
old value). In this case, the vehicle is held only until the threshold time and
then dispatched. If the vehicle arrives after the threshold value, it is dispatched
immediately. On the other hand, schedule-based holding involves holding a
vehicle only until its scheduled departure time; if it arrives later than the sched-
uled time, it is dispatched immediately. More recently, in contrast to such
holding policies, models have been developed to determine optimal holding
times for each vehicle individually. In all of the modeling approaches, the ob-
jective is to minimize the total passenger delay (or waiting time), as measured
by the delay or waiting time for passengers waiting to board, passengers al-
ready on board, and passengers at downstream locations. Generally, for cases
where passengers may arrive at stops according to a printed schedule, the delay
is measured in terms of the deviation from the schedule. In cases where pas-
sengers arrive randomly at the stops, the average passenger waiting time was
given by Welding (1957) in the following expression:
 
E[H] Var[H]
E[WT] = 1+  (43)
2 E[H]2
where E[WT] is the expected waiting time per person, E[H] is the expected
headway, and Var[H] is the variance of the headway.
Analytic approaches, considering idealized routes with stochastic service
characteristics, were studied in the 1970s. The analytic work at this time fo-
cused on optimal threshold policies for simplistic networks, as analytic models
for optimal threshold-based holding policies were not easily found for more
realistic problems. Partly as a consequence, most subsequent analysis schedule-
based and threshold-based holding from the 1970s through the 1990s has been
conducted using simulation, rather than analytic techniques.
114 G. Desaulniers and M.D. Hickman

In contrast to methods to find an optimal threshold value, Barnett (1974) in-

troduced a model to solve directly for the optimal holding time of each vehicle
at a control stop. Barnett (1974) derived approximations for optimal holding
times at a single control point along a transit route, using simplified two-point
discrete distributions of the vehicle lateness to that point. The optimal holding
time for each vehicle is found by minimizing the expected waiting time, given
as a quadratic function of the holding time. The optimal holding times are a
function of the mean and variance of the headway distribution, the ratio of the
passenger load at the control stop to the load downstream, and the covariance
of vehicle arrivals at the control stop.
Turnquist and Blume (1980) extended the analysis by Barnett (1974) to con-
sider the effectiveness of holding decisions. They show that holding will only
serve to reduce the passenger waiting time if
05γ
COV[H] >  (44)
1−γ
where COV[H] is the coefficient of variation of headways at the control stop
and γ is the ratio of passengers on board at the control stop to those at down-
stream stops. The obvious implications of this formula are that control is best
implemented when the coefficient of variation of headways is large, and/or
when the ratio of on-board passengers to downstream passengers is small. This
can be used to select whether and at what locations holding is implemented.
More rigorous analytic methods to solve for individual vehicle holding times
have only emerged more recently. Most of the recent holding models include
detailed models of transit operations, such as dwell times, passenger board-
ing and alighting processes, and minimum headway and capacity constraints.
For holding decisions, most models assume that vehicle dwell times at stops
are modeled explicitly as a linear function of the number of boarding and/or
alighting passengers, plus any holding time added at the control stop. More
critically, this level of detail allows a certain realism in the modeling of the ef-
fect of holding decisions: once a hold is effected, the dwell time of subsequent
vehicles at the same stop, and their trajectories downstream, will be changed.
The model formulation of Adamski and Turnau (1998) addressed the prob-
lem of minimizing schedule deviations on route. The problem is formulated as
an optimal control problem, and the operating dynamics are explained through
a set of linear difference equations as a vehicle moves across stops on a route.
Using these linear difference equations and a quadratic objective function in
the vehicle departure times, the determination of a control (a holding time)
can be solved through traditional control methods. This approach appears to
be most applicable for maintaining schedule adherence on lower-frequency
transit lines, where schedule adherence may be more important than main-
taining regular headways.
For higher-frequency service, headway regularity becomes the dominant
factor in minimizing passenger waiting time. Holding in this case becomes a
problem of adjusting vehicle headways to minimize this variability. Here, we
 Ch. 2. Public Transit 115

present the vehicle holding problem formulation based on the work of Eberlein
(1995) and Eberlein et al. (2001). A transit route has stations K = {1     kt },
where kt is a terminus with sufficient layover time to recover from a minor ser-
vice disruption. A control is exerted only at stop k, and affects only downstream
stops from k to kt . Let K = {k + 1     kt } be the set of downstream stops.
Vehicle trips affected by the hold are in the set Im = {i i + 1     i + m − 1}
where trip m is not controlled. The decision variables are the departure times
for each vehicle at the control stop (djk for vehicle j at the control stop k).
The headways, then, are measured as the time between consecutive departures
from each stop.
In the model of operations, passengers arrive at the stop k at a rate of λk .
The load on board vehicle j after leaving stop k is Ljk , and the percentage
alighting at any stop k is given by qk . Also, ajk is the arrival time of vehicle j
at stop k , sjk is the dwell time of vehicle j at stop k to allow passengers to
alight and board, Rk is the running time from stop k − 1 to k . The dwell time
is based on a linear function of the alighting and boarding passengers, where
c0 is a constant term, c1 is the incremental time necessary for one passenger
to board, and c2 is the incremental time for one passenger to alight. Also, de-
lay may propagate at a downstream timepoint kc if the departure time djkc
after passenger alighting and boarding is greater than the scheduled departure
0
time tjkc . The departure time djk is the departure time at this timepoint if
c
no control action is taken. Finally, no vehicle may enter a stop for a period
of time hmin after a vehicle has departed, and the minimum headway upon
entering the stop must be at least h0 .
The formulation is as follows, where θ is a weight on the delay to on-board
passengers compared to waiting passengers:

 
kt
λk 2
minimize h + θLik (dik − aik − sik ) (45)
2 jk
j∈Im k =k
subject to:
djk − ajk − sjk  0 ∀j ∈ Im  (46)
djk − ajk − sjk = 0 ∀j ∈ Im  ∀k ∈ K  (47)
di+mk − ai+mk − si+mk = 0 (48)
ajk − dj−1k  hmin  ∀j ∈ Im  ∀k ∈ K  (49)
 0

djkc  max tjkc  djkc
 ∀j ∈ Im  (50)
ajk = max{djk −1 + Rk  dj−1k + h0 }
∀j ∈ Im  k ∈ K ∪ {k} (51)
sik = c0 + c1 λk hik + c2 qk Lik −1 
∀j ∈ Im  k ∈ K ∪ {k} (52)
hjk = djk − dj−1k  ∀j ∈ Im  k ∈ K ∪ {k} (53)
116 G. Desaulniers and M.D. Hickman

Ljk = λk hjk + (1 − qk )Ljk −1 

∀j ∈ Im  k ∈ K ∪ {k} (54)
The first term in (45) gives the total waiting time experienced by passengers at
the current stop and all downstream stops, and the second term is the weighted
objective value of delay to passengers on board during the hold. The constraint
sets (46)–(54) account for the most commonly used dynamics of operations
across the different vehicles and stops. In this formulation, (46)–(48) account
for arrival and dwell times at stops; (49)–(50) account for maintaining mini-
mum time separation between vehicles at all stops and at the next timepoint,
respectively; (51) accounts for run time between stops ensuring a minimum
headway; (52) gives a linear accounting of the dwell time for boarding and
alighting passengers at a stop; (53) defines the headways in terms of consecu-
tive departure times; and (54) gives the load on board the vehicle upon leaving
a stop.
The formulation in Eberlein et al. (2001) excluded the second term in the
objective function (the delay imposed to those on board). The formulation
includes as decision variables the departure times djk for all vehicles j ∈ Im
at the control stop k. Note that once these departure times are determined,
the system evolution is automatic using (47)–(54). Also, this formulation of
the vehicle holding problem is a quadratic program but generally not convex.
To solve the problem, a heuristic to optimize each departure time, sequentially
across vehicles from i to i + m and iterating until convergence, is proposed.
A similar mathematical formulation is presented by Zhao et al. (2001), but
using a more general formulation of the cost function. However, the formu-
lation includes only one decision variable, the holding time of the current
vehicle i. The proposed solution technique uses a multiagent system approach,
in which the vehicle and the set of impacted stations engage in negotiation us-
ing the marginal costs of the proposed holding time. This method is proved to
be optimal for convex cost functions.
An extension of these models is described by Sun and Hickman (2004). This
work considers the use of multiple holding stations along a route, in order
to minimize a weighted sum of passenger waiting costs and on-board delay.
A heuristic is proposed which solves for the optimal holding times of each ve-
hicle at its next holding station. The models are solved sequentially, beginning
at the holding station furthest downstream and moving upstream along the
route. The solution at each holding station is solved using a steepest descent
method.
The vehicle holding problem formulation by Hickman (2001) differs from
these previous formulations in that the analysis explicitly includes stochastic
elements, in contrast to these strictly deterministic models. In Hickman (2001),
the passenger arrival and alighting processes and the vehicle running times are
considered stochastic. The objective function and the operational dynamics
incorporate these processes through the expected values and variances of the
vehicle headways and loads. The total passenger waiting time, based on (43),
 Ch. 2. Public Transit 117

yields the following objective function, minimizing over the holding time t  0:

kt
λk  
minimize Var[hjk |t] + E[hjk |t]2 + θE[Lik ]t (55)
2
k =k j∈Im

where the variables are defined as before. This objective is minimized, sub-
ject to the operational dynamics including both expectations and variances of
headway and load. To this end, the operations model of Marguier (1985) was
used, in which the expectations and variances of vehicle headways and loads
are formulated as linear difference equations. As a result, the model becomes
a (convex) quadratic optimization problem with linear constraints, in the single
decision variable of the holding time t. A simple line search is proposed to find
the optimal holding time, while accounting for these operational dynamics.
Another version of the holding problem considers holding strategies for ve-
hicles at a timed transfer terminal. In this condition, passengers arriving late
on one route may not be able to make a transfer to a connecting route. The
purpose of terminal holding is to hold a vehicle so that transfer passengers
can make a connection. In this situation, the objective includes the delay to
passengers on board or downstream on the held route, and the delay to pas-
sengers wishing to transfer. A hold reduces the delay to transfer passengers
while increasing the delay to passengers on board or downstream. The criti-
cal variables include the lateness of vehicles at the terminal and the volume of
transfer, boarding, and downstream passengers on each route.
In a deterministic operating environment where the passenger boarding and
transfer passenger loads are known, and the lateness of any vehicle is known
with certainty, the problem reduces to the situation where either the vehicle
is not held, or it is held until the moment when a vehicle on another route
arrives. In Hall et al. (2001), this model was extended to stochastic vehicle
arrivals, giving a distribution of vehicle lateness on each route. In this case,
analytic methods may be used to determine the optimal holding time for each
route. The objective function reaches a global minimum either with no holding
or at one of the local minima in the neighborhood of each expected vehicle
arrival time. An extension of this model to the case of technology that allows
one to forecast vehicle arrivals at a transfer terminal is described in Dessouky
et al. (2003).

5.2 Other strategies

There have been a variety of other control strategies that have been ana-
lyzed using mathematical programming techniques. In all the cases cited here,
deterministic models of transit operations are used. This means that the objec-
tive function uses the waiting time as a function of the square of the headway.
Li et al. (1991, 1992) presented a model in which stop-skipping and holding
are considered simultaneously in order to bring a route back on schedule after
a service disruption. In this case, skipping stops will reduce dwell times for a
118 G. Desaulniers and M.D. Hickman

vehicle, reducing the preceding headway. This may reduce the average passen-
ger waiting time, but this effect is weighed against the extra waiting time for
passengers whose stops are skipped. With this in mind, the objective function
is to minimize the total passenger waiting time along the route, by selecting for
each vehicle which stops to include in its trip. The decision variables include
binary variables indicating if vehicle j is to stop at stop k, and the continuous
variable of the departure time of each vehicle at each stop. Constraints include
the vehicle operating dynamics and the vehicle capacity. Three solution heuris-
tics are proposed that iterate among local improvement techniques, with each
iteration considering changes in only a subset of decision variables.
Fu et al. (2003) proposed a variation on stop-skipping in which every second
vehicle is considered for stop-skipping. In this way, at most one vehicle will pass
a stop before it is served. The problem is formulated as a nonlinear 0–1 pro-
gramming problem, and is solved separately for each dispatched vehicle. The
objective function includes passenger waiting time, passenger in-vehicle time,
and the bus travel time. Constraints include the typical operations dynamics of
passenger boarding and alighting processes and bus running times. The prob-
lem is solved using explicit enumeration for each bus. Results suggest that this
can be solved in real-time for routes with a small number of potential stops;
there were 14 stops in their case study.
These previous studies assume that a stop-skipping decision is made before
the vehicle is dispatched from a terminal. Sun and Hickman (2005) extended
this concept to consider a real-time stop-skipping policy, made while the vehi-
cle is traveling on the route. Based on the latest vehicle location and disruption
information, the problem is formulated as solving for a skipping segment along
the route, considering passenger waiting time and in-vehicle time. The model
solves for the start and end point of the skipped segment using explicit enumer-
ation. A route with 41 stops was analyzed in their case study, with the model
being solved in real time (i.e., in seconds of CPU time).
Eberlein (1995) examined the real-time control actions of expressing and
deadheading. The expressing problem is defined as determining if a vehicle
should skip over a route segment. In this definition, both the starting stop and
ending stop of the express segment are determined. However, only one express
segment is considered per vehicle. In the deadheading problem, a vehicle is
to be dispatched from a terminal, and the decision faced is whether to begin
revenue service at the terminal or to begin revenue service further down the
route. If deadheading is preferred, a deadhead segment is created over which
the vehicle runs empty. The deadheading work was later published in a sepa-
rate paper (Eberlein et al., 1998). Because of similarities in the formulation of
these problems, only the deadheading work is presented here.
In Eberlein (1995) and Eberlein et al. (1998), this problem was formulated
as a nonlinear integer program in order to identify at what downstream stop to
resume revenue service. The complication of vehicle dynamics make an ana-
lytic solution impossible and significantly complicates the solution using math-
ematical programming techniques. However, under simplifying assumptions
 Ch. 2. Public Transit 119

about the passenger boarding and alighting processes and the vehicle running
times in the deadhead segment, the problem becomes analytically tractable.
With this simplification, the objective function (minimizing total passenger
waiting time) is convex in the number of stops to skip. Using a continuous re-
laxation, the end of the deadhead segment is calculated as a real value, which is
then rounded up or down to the nearest integer to find the stop that minimizes
the objective function.
In Eberlein (1995) and Eberlein et al. (1999), a set of models are pro-
posed to simultaneously examine holding, deadheading, and expressing. It is
observed that the strategy of holding has the opposite effect of deadheading
and expressing: holding a vehicle will lengthen the preceding headway, while
deadheading and expressing a vehicle will shorten the preceding headway. As
a result, at most one control strategy will be applied to a single vehicle i. This
means the control problems are separable, and an efficient heuristic is ap-
plied. In the first step, a holding time is determined, following the heuristic
from Eberlein et al. (2001). If station skipping is feasible, deadheading and ex-
pressing are considered for vehicle i, and holding is considered for subsequent
vehicles i + 1 to i + m to minimize the total passenger waiting time.
Two other recent studies have examined additional operations control
strategies under service disruptions, particularly for rail lines. O’Dell and
Wilson (1999) performed a comparison of holding and short-turning. The
holding model formulation has as its objective minimizing the passenger wait-
ing time along the route, but uses a piecewise linear function as an approxima-
tion of the traditional quadratic objective function. Hard capacity constraints
are included, resulting in integer constraints and, as a result, a mixed integer
linear program. The short-turning model extends the holding model to accom-
modate a vehicle that may be turned around on the route at a control stop; it
is also formulated as a mixed-integer linear program. Commercial software is
used to solve both the holding and short-turning models for a set of disruption
scenarios.
Shen and Wilson (2001) extended the model of O’Dell and Wilson (1999)
to include expressing, in addition to short-turning and holding. The objec-
tive function again includes a piecewise linear approximation to the quadratic
function of passenger waiting time, but includes a large number binary vari-
ables for whether or not to skip a stop (during expressing) and whether or
not to short-turn a vehicle. Additional linear approximations of several nonlin-
ear constraints are used to create a mixed-integer linear program. Commercial
mixed-integer programming software is used to solve these models.
Finally, a model has been presented by Li et al. (2004) which considers bus
re-routing to accommodate vehicle breakdowns. Such a model can be used
to insert a replacement vehicle or re-assign existing service vehicles when a
vehicle must unexpectedly be removed from service. The model uses an auc-
tion heuristic for the multidepot vehicle scheduling problem (MDVSP) to solve
practical instances in real time.
120 G. Desaulniers and M.D. Hickman

6 Conclusion

This survey chapter has reviewed the operations research literature applied
to the domain of public transit, with a focus on recent contributions. It has
highlighted a fruitful cooperation between the public transit agencies and the
operations research community. Indeed, public transit has provided interest-
ing and challenging problems to operations research, while operations research
has been successful at solving efficiently several important public transit prob-
lems (for instance, network design, timetabling, vehicle scheduling, and crew
scheduling). Research on these problems is still going on with the aim of de-
veloping new solution approaches or improving existing ones that will allow to
solve larger instances and to address additional complexities such as stochas-
ticity and complicated operational rules that were previously ignored.
This survey has also shown that new problems (integration of vehicle and
crew scheduling, bus parking and dispatching, as well as a wide variety of real-
time control problems), presenting new challenges to the operations research
community, have also been studied recently. Research on these problems has
already suggested innovative models and solution methodologies which might
be applicable in practice in a near future.
This fruitful collaboration between transit agencies and operations research
will certainly continue for a long time as transit agencies continue to strive
to provide a good quality service at minimum cost, with continual pressure
from budgetary restrictions. Operations research should therefore remain an
essential tool for helping the agencies plan and run their operations efficiently.

References

Books from CASPT conferences in chronological order

Wren, A. (Ed.) (1981). Computer Scheduling of Public Transport. North-Holland, Amsterdam.

Rousseau, M. (Ed.) (1985). Computer Scheduling of Public Transport 2. North-Holland, Amsterdam.
Daduna, J.R., Wren, A. (Eds.) (1988). Computer-Aided Transit Scheduling. Lecture Notes in Economics
and Mathematical Systems, vol. 308. Springer-Verlag, Heidelberg.
Desrochers, M., Rousseau, M. (Eds.) (1992). Computer-Aided Transit Scheduling. Lecture Notes in Eco-
nomics and Mathematical Systems, vol. 386. Springer-Verlag, Heidelberg.
Daduna, J.R., Branco, I., Paixão, J.M.P. (Eds.) (1995). Computer-Aided Transit Scheduling. Lectures
Notes in Economics and Mathematical Systems, vol. 430. Springer-Verlag, Heidelberg.
Wilson, N.H.M. (Ed.) (1999). Computer-Aided Transit Scheduling. Lecture Notes in Economics and
Mathematical Systems, vol. 471. Springer-Verlag, Heidelberg.
Voss, S., Daduna, J.R. (Eds.) (2001). Computer-Aided Scheduling of Public Transport. Lecture Notes in
Economics and Mathematical Systems, vol. 505. Springer-Verlag, Heidelberg.
Hickman, M., Mirchandani, P., Voss, S. (Eds.) (in press). Computer-Aided Scheduling of Public Transport.
Lecture Notes in Economics and Mathematical Systems, Springer-Verlag, Heidelberg, in press.

Cited references

Adamski, A., Turnau, A. (1998). Simulation support tool for real-time dispatching control in public
transport. Transportation Research – Part A 32 (2), 73–87.
 Ch. 2. Public Transit 121

Andreasson, I. (1977). A method for the analysis of transit networks. In: Roubens, M. (Ed.), Advances
in Operations Research. North-Holland, pp. 1–8.
Baaj, M.H., Mahmassani, H.S. (1990). TRUST: A LISP program for the analysis of transit route con-
figurations. Transportation Research Record 1283, 125–135.
Baaj, M.H., Mahmassani, H.S. (1992). Artificial intelligence-based system representation and search
procedures for transit route network design. Transportation Research Record 1358, 67–70.
Baaj, M.H., Mahmassani, H.S. (1995). Hybrid route generation heuristic algorithm for the design of
transit networks. Transportation Research – Part C 3 (1), 31–50.
Barnett, A. (1974). On controlling randomness in transit operations. Transportation Science 8, 102–116.
Barnhart, C., Johnson, E.L., Nemhauser, G.L., Savelsbergh, M.W.P., Vance, P.H. (1998). Branch-and-
price: Column generation for solving huge integer programs. Operations Research 46, 316–329.
Bertossi, A.A., Carraresi, P., Gallo, G. (1987). On some matching problems arising in vehicle scheduling
models. Networks 17, 271–281.
Bianco, L., Mingozzi, A., Ricciardelli, S. (1994). A set partitioning approach to the multiple depot
vehicle scheduling problem. Optimization Methods and Software 3, 163–194.
Bielli, M., Caramia, M., Carotenuto, P. (2002). Genetic algorithms in bus network optimization. Trans-
portation Research – Part C 10 (1), 19–34.
Bookbinder, J.H., Désilets, A. (1992). Transfer optimization in a transit network. Transportation Sci-
ence 26 (2), 106–118.
Borndörfer, R., Grötschel, M., Löbel, A. (2003). Duty scheduling in public transit. In: Jäger, W., Krebs,
H.-J. (Eds.), Mathematics – Key Technology for the Future. Springer-Verlag, New York, pp. 653–674.
Borndörfer, R., Löbel, A., Weider, S. (2004). A bundle method for integrated multi-depot vehicle and
duty scheduling in public transit. ZIB-Report 04-14, Konrad-Zuse-Zentrum für Informationstch-
nik, Berlin, Germany. In: Hickman, M., Mirchandani, P., Voss, S. (Eds.), Computer-Aided Transit
Scheduling. Lecture Notes in Economics and Mathematical Systems. Springer-Verlag, Heidelberg, in
press.
Bouzaïene-Ayari, B., Gendreau, M., Nguyen, S. (2001). Modeling bus stops in transit networks: A sur-
vey and new formulations. Transportation Science 35 (3), 304–321.
Bowman, L.A., Turnquist, M.A. (1981). Service frequency, schedule reliability and passenger wait times
at transit stops. Transportation Research – Part A 15 (6), 465–471.
Carpaneto, G., Dell’Amico, M., Fischetti, M., Toth, P. (1989). A branch and bound algorithm for the
multiple vehicle scheduling problem. Networks 19, 531–548.
Ceder, A. (1984). Bus frequency determination using passenger count data. Transportation Research –
Part A 18 (5–6), 439–453.
Ceder, A. (1986). Methods for creating bus timetables. Transportation Research – Part A 21 (1), 59–83.
Ceder, A. (1989). Optimal design of transit short-turn trips. Transportation Research Record 1221, 8–22.
Ceder, A., Israeli, Y. (1998). User and operator perspectives in transit network design. Transportation
Research Record 1623, 3–7.
Ceder, A., Tal, O. (1999). Timetable synchronization for buses. In: Wilson, N.H.M. (Ed.), Computer-
Aided Transport Scheduling. Springer-Verlag, Heidelberg, pp. 245–258.
Ceder, A., Wilson, N.H.M. (1986). Bus network design. Transportation Research – Part B 20 (4), 331–344.
Ceder, A., Golany, B., Tal, O. (2001). Creating bus timetables with maximal synchronization. Trans-
portation Research – Part A 35, 913–928.
Chien, S., Schonfeld, P. (1998). Joint optimization of a rail transit line and its feeder bus system. Journal
of Advanced Transportation 32 (3), 253–284.
Chowdhury, S., Chien, S. (2002). Intermodal transit system coordination. Transportation Planning and
Technology 25, 257–287.
Chriqui, C., Robillard, P. (1975). Common bus lines. Transportation Science 9 (1), 115–121.
Cominetti, R., Correa, J. (2001). Common-lines and passenger assignment in congested transit net-
works. Transportation Science 35 (3), 250–267.
Costa, A., Branco, I., Paixão, J.M.P. (1995). Vehicle scheduling problem with multiple type of vehi-
cles and a single depot. In: Daduna, J.R., Branco, I., Paixão, J.M.P. (Eds.), Computer-Aided Transit
Scheduling. Lecture Notes in Economics and Mathematical Systems, vol. 430. Springer-Verlag, Hei-
delberg, pp. 115–129.
122 G. Desaulniers and M.D. Hickman

Curtis, S.D., Smith, B.M., Wren, A. (1999). Forming bus driver schedules using constraint program-
ming. In: Proceedings of the 1st International Conference on the Practical Application of Constraint
Technologies and Logic Programming (PACLP99). The Practical Application Company, London, pp.
239–254.
Daduna, J.R., Paixão, J.M.P. (1995). Vehicle scheduling for public mass transit – an overview. In:
Daduna, J.R., Branco, I., Paixão, J.M.P. (Eds.), Computer-Aided Transit Scheduling. Lectures Notes
in Economics and Mathematical Systems, vol. 430. Springer-Verlag, Heidelberg, pp. 76–90.
Dantzig, G.B., Wolfe, P. (1960). Decomposition principle for linear programs. Operations Research 8,
101–111.
de Cea, J., Fernández, E. (1989). Transit assignment to minimal routes: An efficient new algorithm.
Traffic Engineering and Control 30 (10), 491–494.
de Cea, J., Fernández, E. (1993). Transit assignment for congested public transport systems: An equi-
librium model. Transportation Science 27 (2), 133–147.
de Cea, J., Fernández, E. (1996). An empirical comparison of equilibrium and non-equilibrium transit
assignment models. Traffic Engineering and Control 37 (7), 441–445.
de Cea, J., Fernández, E. (2000). Transit-assignment models. In: Hensher, D.A., Button, K.J. (Eds.),
Handbook of Transport Modelling. Handbooks in Transport. Elsevier, Amsterdam, pp. 497–508.
de Groot, S.W., Huisman, D. (2004). Vehicle and crew scheduling: Solving large real-world instances
with an integrated approach. Report EI2004-13, Econometric Institute, Erasmus University of Rot-
terdam, The Netherlands. In: Hickman, M., Mirchandani, P., Voss, S. (Eds.), Computer-Aided Transit
Scheduling. Lecture Notes in Economics and Mathematical Systems. Springer-Verlag, Heidelberg, in
press.
de Silva, A. (2001). Combining constraint programming and linear programming on an example of bus
driver scheduling. Annals of Operations Research 108, 277–291.
Dell’Amico, M., Fischetti, M., Toth, P. (1993). Heuristic algorithms for the multiple depot vehicle
scheduling problem. Management Science 39 (1), 115–125.
Desaulniers, G., Desrosiers, J., Ioachim, I., Solomon, M.M., Soumis, F., Villeneuve, D. (1998a). A uni-
fied framework for deterministic time constrained vehicle routing and crew scheduling problems.
In: Crainic, T.G., Laporte, G. (Eds.), Fleet Management and Logistics. Kluwer Academic, Norwell,
MA, pp. 57–93.
Desaulniers, G., Lavigne, J., Soumis, F. (1998b). Multi-depot vehicle scheduling with time windows and
waiting costs. European Journal of Operational Research 111, 479–494.
Desaulniers, G., Desrosiers, J., Solomon, M.M. (2002). Accelerating strategies for column generation
methods in vehicle routing and crew scheduling problems. In: Ribeiro, C.C., Hansen, P. (Eds.),
Essays and Surveys in Metaheuristics. Kluwer Academic, Norwell, MA, pp. 309–324.
Desrochers, M., Soumis, F. (1988). A generalized permanent labeling algorithm for the shortest path
problem with time windows. INFOR 26, 191–212.
Desrochers, M., Soumis, F. (1989). A column generation approach to the urban transit crew scheduling
problem. Transportation Science 23, 1–13.
Desrosiers, J., Rousseau, J.-M. (1995). Results obtained with crew-opt: A column generation method
for transit crew scheduling. In: Daduna, J.R., Branco, I., Paixão, J.M.P. (Eds.), Computer-Aided
Transit Scheduling. Lectures Notes in Economics and Mathematical Systems, vol. 430. Springer-Verlag,
Heidelberg, pp. 349–358.
Desrosiers, J., Dumas, Y., Solomon, M.M., Soumis, F. (1995). Time constrained routing and scheduling.
In: Ball, M.O., Magnanti, T.L., Monma, C.L., Nemhauser, G.L. (Eds.), Network Routing. Handbooks
in Operations Research and Management Science, vol. 8. Elsevier, Amsterdam, pp. 35–139.
Dessouky, M., Hall, R., Zhang, L., Singh, A. (2003). Real-time control of buses for schedule coordina-
tion at a terminal. Transportation Research – Part A 37, 145–164.
Dial, R.B. (1967). Transit pathfinder algorithm. Highway Research Record 205, 67–85.
du Merle, O., Villeneuve, D., Desrosiers, J., Hansen, P. (1999). Stabilized column generation. Discrete
Mathematics 194, 229–237.
Dubois, D., Bel, G., Llibre, M. (1979). A set of methods in transportation network synthesis and analy-
sis. Journal of the Operational Research Society 30 (9), 797–808.
 Ch. 2. Public Transit 123

Eberlein, X.-J. (1995). Real-time control strategies in transit operations: Models and analysis. PhD the-
sis, Massachusetts Institute of Technology, Cambridge, MA.
Eberlein, X.-J., Wilson, N.H.M., Barnhart, C., Bernstein, D. (1998). The real-time deadheading prob-
lem in transit operations control. Transportation Research – Part B 32 (2), 77–100.
Eberlein, X.-J., Wilson, N.H.M., Bernstein, D. (1999). Modeling real-time control strategies in public
transit operations. In: Wilson, N.H.M. (Ed.), Computer-Aided Transit Scheduling. Lecture Notes in
Economics and Mathematical Systems, vol. 471. Springer-Verlag, Heidelberg, pp. 325–346.
Eberlein, X.-J., Wilson, N.H.M., Bernstein, D. (2001). The holding problem with real-time information
available. Transportation Science 35 (1), 1–18.
Elhallaoui, I., Villeneuve, D., Soumis, F., Desaulniers, G. (2005). Dynamic aggregation of set partition-
ing constraints in column generation. Operations Research 53 (4), 632–645.
Fan, W., Machemehl, R.B. (2004). Optimal transit route network design problem: Algorithms, imple-
mentations, and numerical results. Report SWUTC/04/167244-1, Center for Transportation Re-
search, University of Texas at Austin.
Forbes, M.A., Holt, J.N., Watts, A.M. (1994). An exact algorithm for multiple depot bus scheduling.
European Journal of Operational Research 72 (1), 115–124.
Fores, S., Proll, L., Wren, A. (2002). TRACS II: A hybrid IP/heuristic driver scheduling system for
public transport. Journal of the Operational Research Society 53, 1093–1100.
Freling, R., Paixão, J.M.P. (1995). Vehicle scheduling with time constraint. In: Daduna, J.R., Branco,
I., Paixão, J.M.P. (Eds.), Computer-Aided Transit Scheduling. Lecture Notes in Economics and Math-
ematical Systems, vol. 430. Springer-Verlag, Heidelberg, pp. 130–144.
Freling, R., Wagelmans, A.P.M., Paixão, J.M.P. (1999). An overview of models and techniques for
integrating vehicle and crew scheduling. In: Wilson, N.H.M. (Ed.), Computer-Aided Transit Schedul-
ing. Lecture Notes in Economics and Mathematical Systems, vol. 471. Springer-Verlag, Heidelberg,
pp. 441–460.
Freling, R., Huisman, D., Wagelmans, A.P.M. (2001a). Applying an integrated approach to vehicle and
crew scheduling in practice. In: Voss, S., Daduna, J.R. (Eds.), Computer-Aided Scheduling of Public
Transport. Lecture Notes in Economics and Mathematical Systems, vol. 505. Springer-Verlag, Heidel-
berg, pp. 73–90.
Freling, R., Wagelmans, A.P.M., Paixão, J.M.P. (2001b). Models and algorithms for single-depot vehicle
scheduling. Transportation Science 35 (2), 165–180.
Freling, R., Huisman, D., Wagelmans, A.P.M. (2003). Models and algorithms for integration of vehicle
and crew scheduling. Journal of Scheduling 6, 63–85.
Fu, L., Liu, Q., Calamai, P. (2003). A real-time optimization model for dynamic scheduling of transit
operations. In: The 82nd Annual Meeting of the Transportation Research Board, Washington, DC,
January.
Furth, P.G. (1985). Alternating deadheading in bus route operations. Transportation Science 19 (1), 13–
28.
Furth, P.G. (1986). Zonal route design for transit corridors. Transportation Science 20 (1), 1–12.
Furth, P.G. (1987). Short turning on transit routes. Transportation Research Record 1108, 42–52.
Furth, P.G., Wilson, N.H.M. (1982). Setting frequencies on bus routes: Theory and practice. Transporta-
tion Research Record 818, 1–7.
Gallo, G., Di Miele, F. (2001). Dispatching buses in parking depots. Transportation Science 35 (3), 322–
330.
Gao, Z., Sun, H., Shan, L.L. (2004). A continuous equilibrium network design model and algorithm for
transit systems. Transportation Research – Part B 38 (3), 235–250.
Gentile, G., Nguyen, S., Pallottino, S. (2005). Route choice on transit networks with online information
at stops. Transportation Science 39 (3), 289–297.
Gintner, V., Kliewer, N., Suhl, L. (2004). A crew scheduling approach for public transit enhanced with
aspects from vehicle scheduling. DSOR Working Paper WP0407, University of Paderborn, Ger-
many. In: Hickman, M., Mirchandani, P., Voss, S. (Eds.) (2004). Computer-Aided Transit Scheduling.
Lecture Notes in Economics and Mathematical Systems. Springer-Verlag, Heidelberg, in press.
Haase, K., Desaulniers, G., Desrosiers, J. (2001). Simultaneous vehicle and crew scheduling in urban
mass transit systems. Transportation Science 35 (3), 286–303.
124 G. Desaulniers and M.D. Hickman

Hadjar, A., Marcotte, O., Soumis, F. (2006). A branch-and-cut algorithm for the multiple depot vehicle
scheduling problem. Operations Research 54 (1), 130–149.
Haghani, A., Shafahi, Y. (2002). Bus maintenance systems and maintenance scheduling: Model formu-
lations and solutions. Transportation Research – Part A 36, 453–482.
Hall, R.W. (1985). Vehicle scheduling at a transportation terminal with random delay en route. Trans-
portation Science 19 (3), 308–320.
Hall, R.W. (1986). The fastest path through a network with random time-dependent travel times. Trans-
portation Science 20 (3), 182–188.
Hall, R., Dessouky, M., Lu, Q. (2001). Optimal holding times at transfer stations. Computers & Indus-
trial Engineering 40, 379–397.
Hamdouni, M., Desaulniers, G., Soumis, F., Marcotte, O., Van Putten, M. (2006). Parking and dispatch-
ing buses in depots using block patterns. Transportation Science 40 (3), 364–377.
Han, A.F., Wilson, N.H.M. (1982). The allocation of buses in heavily utilized networks with overlapping
routes. Transportation Research – Part B 16 (3), 221–232.
Hasselström, D. (1981). Public transportation planning – a mathematical programming approach. Doc-
toral dissertation, University of Göteborg, Sweden.
Hickman, M. (2001). An analytic stochastic model for the transit vehicle holding problem. Transporta-
tion Science 35 (3), 215–237.
Hickman, M.D., Bernstein, D.H. (1997). Transit service and path choice models in stochastic and time-
dependent networks. Transportation Science 31 (2), 129–146.
Hickman, M.D., Wilson, N.H.M. (1995). Passenger travel time and path choice implications of real-time
transit information. Transportation Research – Part C 3 (4), 211–226.
Huisman, D., Freling, R., Wagelmans, A.P.M. (2005). Multiple-depot integrated vehicle and crew
scheduling. Transportation Science 39 (4), 491–502.
Hurdle, V.F. (1973a). Minimum cost schedules for a public transportation route – I. Theory. Transporta-
tion Science 7 (2), 109–137.
Hurdle, V.F. (1973b). Minimum cost schedules for a public transportation route - II. Examples. Trans-
portation Science 7 (2), 138–157.
Irnich, S., Desaulniers, G. (2005). Shortest path problems with resource constraints. In: Desaulniers, G.,
Desrosiers, J., Solomon, M.M. (Eds.), Column Generation. Springer-Verlag, New York, pp. 33–65.
Israeli, Y. (1992). Transit route and scheduling design at the network level. Doctoral dissertation, Tech-
nion Israel Institute of Technology, Haifa, Israel.
Israeli, Y., Ceder, A. (1989). Designing transit routes at the network level. In: Proceedings of the First
Vehicle Navigation and Information Systems Conference. IEEE Vehicular Technology Society, pp.
310–316.
Israeli, Y., Ceder, A. (1995). Transit route design using scheduling and multiobjective programming
techniques. In: Daduna, J.R., Branco, I., Piaxão, J. (Eds.), Computer-Aided Transit Scheduling. Lec-
ture Notes in Economics and Mathematical Systems, vol. 430. Springer-Verlag, Heidelberg, pp. 56–75.
Israeli, Y., Ceder, A. (1996). Public transportation assignment with passenger strategies for overlapping
route choice. In: Lesort, B. (Ed.), Transportation and Traffic Theory: Proceedings of the 13th Interna-
tional Symposium on Transportation and Traffic Theory. Pergamon, pp. 561–588.
Jansson, K., Ridderstolpe, B. (1992). A method for the route choice problem in public transport sys-
tems. Transportation Science 26 (3), 246–251.
Jordan, W.C., Turnquist, M.A. (1979). Zone scheduling of bus routes to improve service reliability.
Transportation Science 13 (3), 242–268.
Klabjan, D., Johnson, E.L., Nemhauser, G.L., Gelman, E., Ramaswamy, S. (2002). Airline crew
scheduling with time windows and plane-count constraints. Transportation Science 36, 337–348.
Klemt, W.D., Stemme, W. (1988). Schedule synchronization for public transit networks. In: Daduna,
J.R., Wren, A. (Eds.), Computer-Aided Transit Scheduling. Springer-Verlag, New York, pp. 327–335.
Kliewer, N., Mellouli, T., Suhl, L. (2006). A time-space network based exact optimization model for
multi-depot bus scheduling. European Journal of Operational Research 175 (3), 1616–1627.
Knoppers, P., Muller, T. (1995). Optimized transfer opportunities in public transport. Transportation
Science 29 (1), 101–105.
 Ch. 2. Public Transit 125

Koutsopoulos, H.N., Odoni, A., Wilson, N.H.M. (1985). Determination of headways as a function of
time varying characteristics on a transit network. In: Rousseau, J.M. (Ed.), Computer Scheduling of
Public Transport 2. North-Holland, Amsterdam, pp. 391–414.
Kwan, A.S.K., Kwan, R.S.K., Wren, A. (1999). Driver scheduling using genetic algorithms with em-
bedded combinatorial traits. In: Wilson, N.H.M. (Ed.), Computer-Aided Transit Scheduling. Lecture
Notes in Economics and Mathematical Systems, vol. 471. Springer-Verlag, Heidelberg, pp. 81–102.
Kwan, A.S.K., Kwan, R.S.K., Wren, A. (2001). Evolutionary driver scheduling with relief chains. Evo-
lutionary Computing 9, 445–460.
Lam, W.H.K., Gao, Z.Y., Chan, K.S., Yang, H. (1999). A stochastic user equilibrium assignment model
for congested transit networks. Transportation Research – Part B 33, 351–368.
Lam, W.H.K., Zhou, J., Sheng, Z.-H. (2002). A capacity restraint transit assignment with elastic line
frequency. Transportation Research – Part B 36, 919–938.
Lampkin, W., Saalmans, P.D. (1967). The design of routes, service frequencies, and schedules for a
municipal bus undertaking: A case study. Operational Research Quarterly 18 (4), 375–397.
Last, A., Leak, S.E. (1976). Transept: A bus model. Traffic Engineering and Control 18 (1), 14–20.
le Clercq, F. (1972). A public transport assignment method. Traffic Engineering and Control 14 (2), 91–
96.
Lee, K.T., Schonfeld, P. (1991). Optimal slack time for timed transfers at a transit terminal. Journal of
Advanced Transportation 25 (3), 281–308.
Levinson, H. (1991). Supervision strategies for improved reliability of bus routes. In: Synthesis of Transit
Practice 15, National Cooperative Transit Research and Development Program.
Li, Y., Rousseau, J.-M., Gendreau, M. (1991). Real-time scheduling on a transit bus route: A 0–1 sto-
chastic programming model. Publication 772, Centre de Recherche sur les Transports, Université
de Montréal.
Li, Y., Rousseau, J.-M., Wu, F. (1992). Real-time scheduling on a transit bus route. In: Desrochers, M.,
Rousseau, M. (Eds.), Computer-Aided Transit Scheduling. Lecture Notes in Economics and Mathe-
matical Systems, vol. 386. Springer-Verlag, Heidelberg, pp. 213–235.
Li, J., Mirchandani, P., Borenstein, D. (2004). Parallel auction algorithm for bus rescheduling. In: Hick-
man, M., Mirchandani, P., Voss, S. (Eds.), Computer-Aided Scheduling of Public Transport. Lecture
Notes in Economics and Mathematical Systems. Springer-Verlag, Heidelberg, in press.
Lo, H., Yip, C.W., Wan, K.H. (2003). Modeling transfers and nonlinear fare structure in multi-modal
network. Transportation Research – Part B 37 (2), 149–170.
Lo, H., Yip, C.W., Wan, Q.K. (2004). Modeling competitive multi-modal transit services: A nested logit
approach. Transportation Research – Part C 12, 251–272.
Löbel, A. (1998). Vehicle scheduling in public transit and Lagrangean pricing. Operations Research 44
(12), 1637–1649.
Lourenço, H.R., Paixão, J.P., Portugal, R. (2001). Multiobjective metaheuristics for the bus-driver
scheduling problem. Transportation Science 35 (3), 331–343.
Magnanti, T.L., Wong, R.T. (1984). Network design and transportation planning: Models and algo-
rithms. Transportation Science 18 (1), 1–55.
Marguier, P.H.J. (1985). Bus route performance evaluation under stochastic conditions. PhD thesis,
Massachusetts Institute of Technology, Cambridge, MA.
Marguier, P.H.J., Ceder, A. (1984). Passenger waiting strategies for overlapping bus routes. Transporta-
tion Science 18 (3), 207–230.
Mesquita, M., Paixão, J. (1999). Exact algorithms for the multi-depot vehicle scheduling problem based
on multicommodity network flow type formulations. In: Wilson, N.H.M. (Ed.), Computer-Aided
Transit Scheduling. Lecture Notes in Economics and Mathematical Systems, vol. 471. Springer-Verlag,
Heidelberg, pp. 221–243.
Mingozzi, A., Bianco, L., Ricciardelli, S. (1995). An exact algorithm for combining vehicle trips. In:
Daduna, J.R., Branco, I., Paixão, J.M.P. (Eds.), Computer-Aided Transit Scheduling. Lecture Notes in
Economics and Mathematical Systems, vol. 430. Springer-Verlag, Heidelberg, pp. 145–172.
Newell, G.F. (1971). Dispatching policies for a transportation route. Transportation Science 5 (1), 91–
105.
126 G. Desaulniers and M.D. Hickman

Nguyen, S., Pallottino, S. (1988). Equilibrium traffic assignment for large scale transit networks. Euro-
pean Journal of Operational Research 37, 176–186.
Nguyen, S., Pallottino, S., Malucelli, F. (2001). A modeling framework for passenger assignment on a
transport network with timetables. Transportation Science 35 (3), 238–249.
Nielsen, O.A. (2000). A stochastic transit assignment model considering differences in passengers utility
functions. Transportation Research – Part B 34, 377–402.
Nielsen, O.A. (2004). A large scale stochastic multi-class schedule-based transit model with random co-
efficients. In: Wilson, N.H.M., Nuzzolo, A. (Eds.), Schedule-Based Dynamic Transit Modeling: Theory
and Applications. Kluwer Academic, Boston, pp. 53–77.
Nuzzulo, A. (2003). Transit path choice and assignment model approaches. In: Lam, W.H.K., Bell,
M.G.H. (Eds.), Advanced Modeling for Transit Operations and Service Planning. Pergamon, Amster-
dam, pp. 93–124.
Nuzzulo, A., Russo, F., Crisalli, U. (2001). A doubly dynamic schedule-based assignment model for
transit networks. Transportation Science 35 (3), 268–285.
O’Dell, S., Wilson, N.H.M. (1999). Optimal real-time control strategies for rail transit operations during
disruptions. In: Wilson, N.H.M. (Ed.), Computer-Aided Transit Scheduling. Lecture Notes in Eco-
nomics and Mathematical Systems, vol. 471. Springer-Verlag, Heidelberg, pp. 299–323.
Odoni, A.R., Rousseau, J.-M., Wilson, N.H.M. (1994). Models in urban and air transportation. In:
Pollock, S.M., Rothkopf, M.H., Barnett, A. (Eds.), Operations Research and the Public Sector. Hand-
books in Operations Research and Management Science, vol. 6. North-Holland, Amsterdam, pp. 107–
150.
Pattnaik, S.B., Mohan, S., Tom, V.M. (1998). Urban bus transit network design using genetic algorithm.
Journal of Transportation Engineering 124 (4), 368–375.
Poon, M.H., Wong, S.C., Tong, C.O. (2004). A dynamic schedule-based model for congested transit
networks. Transportation Research – Part B 38, 343–368.
Rapp, M.H., Gehner, C.D. (1976). Transfer optimization in an interactive graphic system for transit
planning. Transportation Research Record 619, 27–33.
Ribeiro, C.C., Soumis, F. (1994). A column generation approach to the multiple depot vehicle schedul-
ing problem. Operations Research 42 (1), 41–52.
Salzborn, F.J.M. (1972). Optimum bus scheduling. Transportation Science 6 (2), 137–148.
Salzborn, F.J.M. (1980). Scheduling bus systems with interchanges. Transportation Science 14 (3), 211–
220.
Scheele, S. (1980). A supply model for public transit services. Transportation Research – Part B 14, 133–
146.
Sheffi, Y., Sugiyama, M. (1982). Optimal bus scheduling on a single route. Transportation Research
Record 895, 46–52.
Shen, S., Wilson, N.H.M. (2001). An optimal integrated real-time disruption control model for rail tran-
sit systems. In: Voss, S., Daduna, J. (Eds.), Computer-Aided Scheduling of Public Transport. Lecture
Notes in Economics and Mathematical Systems, vol. 503. Springer-Verlag, Heidelberg, pp. 335–363.
Shen, Y., Kwan, R.S.K. (2001). Tabu search for driver scheduling. In: Voss, S., Daduna, J.R. (Eds.),
Computer-Aided Scheduling of Public Transport. Lecture Notes in Economics and Mathematical Sys-
tems, vol. 505. Springer-Verlag, Heidelberg, pp. 121–135.
Silman, L.A., Barzily, Z., Passy, U. (1974). Planning the route system for urban buses. Computers &
Operations Research 1, 210–211.
Site, P.D., Filippi, F. (1998). Service optimization for bus corridors with short-turn strategies and vari-
able vehicle size. Transportation Research – Part A 32 (1), 19–38.
Smith, B.M., Wren, A. (1988). A bus crew scheduling system using a set covering formulation. Trans-
portation Research – Part A 22, 97–108.
Spiess, H. (1983). On optimal route choice strategies in transit networks. Publication 285, Centre de
Recherche sur les Transports, Université de Montréal.
Spiess, H., Florian, M. (1989). Optimal strategies: A new assignment model for transit networks. Trans-
portation Research – Part B 23 (2), 83–102.
Stern, H.I., Ceder, A. (1983). An improved lower bound to the minimum fleet size problem. Transporta-
tion Science 17 (4), 471–477.
 Ch. 2. Public Transit 127

Sun, A., Hickman, M. (2004). The holding problem at multiple holding stations. In: Hickman, M.,
Mirchandani, P., Voss, S. (Eds.), Computer-Aided Scheduling of Public Transport. Lecture Notes in
Economics and Mathematical Systems. Springer-Verlag, Heidelberg.
Sun, A., Hickman, M. (2005). The real-time stop-skipping problem. Journal of Intelligent Transportation
Systems 9 (2), 91–109.
Ting, C.-J., Schonfeld, P. (2005). Schedule coordination in a multiple hub transit network. Journal of
Urban Planning and Development 131 (2), 112–124.
Tom, V.M., Mohan, S. (2003). Transit route network design using frequency coded genetic algorithm.
Journal of Transportation Engineering 129 (2), 186–195.
Tong, C.O., Richardson, A.J. (1984). A computer model for finding the time-dependent minimum path
in a transit system with fixed schedules. Journal of Advanced Transportation 18 (2), 145–161.
Tong, C.O., Wong, S.C. (1999). A stochastic transit assignment model using a dynamic schedule-based
network. Transportation Research – Part B 33, 107–121.
Turnquist, M.A. (1978). A model for investigating the effects of service frequency and reliability on bus
passenger waiting times. Transportation Research Record 663, 70–73.
Turnquist, M.A. (1981). Strategies for improving reliability of bus service. Transportation Research
Record 818, 7–13.
Turnquist, M.A., Blume, S.W. (1980). Evaluating potential effectiveness of headway control strategies
for transit systems. Transportation Research Record 746, 25–29.
van Nes, R., Hamerslag, R., Immers, B.H. (1988). Design of public transport networks. Transportation
Research Record 1202, 74–83.
Verma, A., Dinghra, S.L. (2005). Feeder bus routes generation within integrated mass transit planning
framework. Journal of Transportation Engineering 131 (11), 822–834.
Wahba, M., Shalaby, A. (2005). A multi-agent learning-based approach to the transit assignment prob-
lem: A prototype. Transportation Research Record 1926, 96–105. Paper taken from CD-ROM of
Conference Proceedings.
Welding, P.I. (1957). The instability of a close-interval service. Operational Research Quarterly 8 (3),
133–148.
Wilson, N.H.M., Nuzzolo, A. (Eds.) (2004). Schedule-Based Dynamic Transit Modeling: Theory and Ap-
plications. Kluwer Academic, Boston.
Winter, T., Zimmermann, U.T. (2000). Real-time dispatch of trams in storage yards. Annals of Opera-
tions Research 96, 287–315.
Wirasinghe, S.C., Liu, G. (1995). Optimal schedule design for a transit route with one intermediate
time point. Transportation Planning and Technology 19, 121–145.
Wren, A., Rousseau, J.-M. (1995). Bus driver scheduling – an overview. In: Daduna, J.R., Branco, I.,
Paixão, J.M.P. (Eds.), Computer-Aided Transit Scheduling. Lecture Notes in Economics and Mathe-
matical Systems, vol. 430. Springer-Verlag, Heidelberg, pp. 173–187.
Wu, J.H., Florian, M. (1993). A simplicial decomposition method for the transit equilibrium assignment
problem. Annals of Operations Research 44, 245–260.
Wu, J.H., Florian, M., Marcotte, P. (1994). Transit equilibrium assignment: A model and solution algo-
rithms. Transportation Science 28 (3), 193–203.
Zhao, J., Dessouky, M., Bukkapatnam, S. (2001). Distributed holding control of bus transit operations.
In: Proceedings of the IEEE Intelligent Transportation Systems Council (ITSC) Conference, Oakland,
CA, August.
C. Barnhart and G. Laporte (Eds.), Handbook in OR & MS, Vol. 14
Copyright © 2007 Elsevier B.V. All rights reserved
DOI: 10.1016/S0927-0507(06)14003-7

Chapter 3
Passenger Railway Optimization
Alberto Caprara
DEIS, Università di Bologna, Viale Risorgimento 2, 40136 Bologna, Italy
E-mail: [email protected]

Leo Kroon
NS Reizigers, Department of Logistics, P.O. Box 2025, 3500 HA Utrecht, The Netherlands
and
Erasmus University Rotterdam, Rotterdam School of Management,
P.O. Box 1738, 3000 DR Rotterdam, The Netherlands
E-mail: [email protected]

Michele Monaci
DEI, Università di Padova, Via Gradenigo 6/A, 35131 Padova, Italy
E-mail: [email protected]

Marc Peeters
Electrabel, Strategy R&D, Avenue Einstein 2A, 1348 Louvain-la-Neuve, Belgium
E-mail: [email protected]

Paolo Toth
DEIS, Università di Bologna, Viale Risorgimento 2, 40136 Bologna, Italy
E-mail: [email protected]

1 Introduction

Railway systems are highly complex systems. Therefore, planning and oper-
ational processes related to railway systems are fields that are rich in interesting
combinatorial optimization problems. Well-known examples of these are op-
erational planning problems such as line planning, timetabling, platforming,
rolling stock circulation, shunting, and crew planning.
However, in the railway industry it was recognized only recently that such
problems can be analyzed and solved through the application of mathematical
models and optimization techniques, and that this kind of innovation may lead
to an improvement in the performance of the railway system as a whole, as well
as to a reduction in the time required for solving these problems. The latter
aspect is important, because it increases the flexibility of the railway system:
the system can adapt in a faster way to changes in the environment.
Railway transportation can be split into passenger transportation and cargo
transportation. In this chapter, we mainly focus on the European situation,
where the major part of railway transportation consists of passenger trans-

129
130 A. Caprara et al.

portation, without addressing important problems in cargo transportation such

as car blocking, train makeup, train routing, empty car distribution. Whereas
in the recent past many European railway companies were state-owned, many
of them are currently operating (partly) on a commercial basis, due to the new
regulations of the European Commission, which specify that the management
of the infrastructure should be the responsibility of the governments, but op-
erating trains should be carried out by independent Train Operators on a com-
mercial basis. This introduces the separate organizations of the Infrastructure
Manager, who is responsible for train planning and real-time traffic control,
and the Train Operators, who provide their preferred timetable, rolling stock,
and transport services. The Train Operators can be split in operators of pas-
senger trains and operators of cargo trains.
The recognition that combinatorial optimization problems arising in railway
applications can be solved through the application of mathematical models and
of the corresponding solution techniques was caused by several factors. Indeed,
the ability to undertake infrastructure planning in a very timely, smooth and ef-
ficient way is becoming one of the most important tasks of the Infrastructure
Manager, who at the same time has to optimize the use of the infrastructure
and to provide line allocation as well as time slot allocation, through ratio-
nal and transparent procedures. Moreover, the Train Operators increased the
demand for improved performance and for higher speed and flexibility in the
planning and in the operations. This also stimulated the search for innovative
decision support tools, which were recognized much earlier already in the air-
line industry. On the other hand, the increased power of the currently available
hardware as well as the recent developments in mathematical modeling and so-
lution techniques allow one to find solutions close to optimality for real-world
combinatorial optimization problems that, until the end of the last century, one
did not dare to dream about. Nevertheless, a lot of research is still to be carried
out, since in general the combinatorial optimization problems in the railway in-
dustry are highly complex and usually also much larger than the corresponding
ones in the airline industry.
As a consequence of the above discussion, real-world applications of math-
ematical optimization techniques in the railway industry are not yet as wide-
spread as in the airline industry. In this chapter, we focus attention on the
optimization techniques developed in each of the operational areas mentioned
in the beginning, devoting a section to each of them and considering the prob-
lems in the order in which they have to be faced in the planning phase.
We start in Section 2 with the Line Planning Problem, which amounts to
deciding the routes for the passenger trains as well as the types and frequen-
cies of the trains on each route. Afterwards, the actual timetable of each train
has to be fixed, solving the Train Timetabling Problem of Section 3. Once their
timetable is fixed, trains have to be assigned to platforms in the stations they
visit, leading to the Train Platforming Problem of Section 4. The definition of
train units (locomotives and train carriages) to be assigned to the trains with
known timetable and platforms is the Rolling Stock Circulation Problem of
 Ch. 3. Passenger Railway Optimization 131

Section 5, whereas the parking of these train units outside the stations when
they are not used or are in maintenance is the Train Unit Shunting Problem of
Section 6. Finally, the definition of the workload of train drivers and conduc-
tors to operate a given timetable is the Crew Planning Problem addressed in
Section 7.
Each section contains an introduction, a survey of the available literature,
a formal statement of the problem along with one or more (Mixed) Integer
Linear Programming ((M)ILP) formulations, and for some cases, experimen-
tal results on specific case studies. The (M)ILP formulations presented have
been used to solve the associated problems in practice through the solution
of the associated Linear Programming (LP) relaxation, either directly by some
general-purpose solver or (heuristically) by using Lagrangian relaxation and
heuristic methods. Huisman et al. (2005) and Bussieck et al. (1997) also pro-
vide overviews of the application of mathematical optimization techniques in
passenger railway systems.
Other areas where optimization techniques could be used effectively, such
as long term capacity planning of infrastructure, rolling stock, and personnel
did not yet get a lot of attention in the literature. The same applies to real-time
control of the railway processes: in these areas the development of effective
model based decision support tools is still in its infancy.
Nevertheless, we conclude the introduction with a brief description of a
number of strategic issues, in order to provide some background information
on the relevant problems there. We hope that this will inspire researchers to
study these subjects as well, since, in the long run, adequate models and so-
lution techniques for solving these problems may be even more relevant for
the quality of the railway system than the effective solution of the operational
scheduling problems.

Strategic planning

Strategic issues in railway systems are related to the desirable service level
to be provided to the customers (in terms of number of direct connections, fre-
quencies, and reliability), and the capacities of the resources that are required
to accomplish these services. The most important resources are the railway in-
frastructure, the rolling stock, and the crews. Usually enormous amounts of
money are involved with the management of these resources.
Strategic issues are far more difficult to handle by optimization approaches
than more operational issues due to the extremely long planning horizons that
are involved. For example, building a new infrastructure usually takes several
years, and once the new infrastructure has been built, it will be operational
for several decades. For rolling stock, the situation is similar. These long plan-
ning horizons imply a lot of uncertainty, e.g., about the demand for railway
transportation in the long run. Therefore, the availability of dependable de-
mand forecasting models is highly important. Models that are to be used for
supporting strategic decisions on long term capacities of resources preferably
132 A. Caprara et al.

take into account this uncertainty explicitly, e.g., by applying scenario analysis
or stochastic optimization techniques.
In the following, we briefly address the cases of rolling stock and crew man-
agement.

Rolling stock management

Railway passenger transportation is characterized by a demand that is usu-

ally concentrated in two peak periods per workday. Therefore, when determin-
ing the required capacities of the railway system, not only the total demand
for railway transportation has to be taken into account, but also the difference
between peak demand and off-peak demand. The demand for passenger trans-
portation during the peak periods is usually not symmetric (e.g., in the morning
peak period, most passengers travel towards the large cities). This implies that
the utilization rate of the rolling stock cannot be as high as one would like.
The required capacity of the rolling stock for passenger transportation is
also influenced by the requested service level. Also the difference between first
and second class demand is to be taken into account. Another issue in rolling
stock management is the maintenance policy to be applied: a more intensive
maintenance strategy requires more rolling stock than a less intensive strategy
but, on the other hand, will probably lead to more reliable rolling stock, thereby
resulting in less disturbances and a higher punctuality in the operations. Al-
though it is clear that such relations exist in practice, it is hard to model them
explicitly, since issues like the reliability of railway systems have hardly been
quantified so far.
Options that have to be decided upon in rolling stock management are
the following: selection of the types of rolling stock to be used, acquisition
of new rolling stock, temporarily hiring or leasing of rolling stock, upgrading
of existing rolling stock, life time extension of existing rolling stock, selling of
redundant rolling stock, destruction of rolling stock that has completed its life
cycle, and, as mentioned, the maintenance policy to be applied. For rolling
stock to be acquired, the appropriate types and first and second class capac-
ities per unit are relevant issues besides all kinds of technical specifications.
Typically, large rolling stock units may be inflexible in the operations and small
units may be relatively expensive.
Usually, in strategic issues the objective is to minimize the expected costs
over the life cycle of the rolling stock, given the fact that certain criteria with
respect to the service for the passengers have to be met. The latter implies that
usually a shortage of rolling stock capacity is worse than a surplus, although
both are preferably to be avoided. One of the few papers describing the non-
operational problem of allocating the available rolling stock capacity among
the lines to be operated, based on the passenger demand during the morning
rush hours, is Abbink et al. (2004). This chapter also takes into account some
robustness aspects, by minimizing the number of different rolling stock types
per line.
 Ch. 3. Passenger Railway Optimization 133

Crew management

Crew management for train drivers and conductors deals with strategic is-
sues related, e.g., to the locations and the capacities of the crew depots. Also
the balance between the capacity for drivers and the capacity for conductors
per depot is important, in particular if drivers and conductors are assumed to
operate in teams. Moreover, a balance per depot with respect to age, skills, and
gender is to be pursued.
The objective of crew management is to establish a matching between the
required and the available capacities of the depots. The required capacities of
the depots depend on the timetable and the rolling stock circulation, but also
on the agreements between management and crews about the structure of the
crew workload. The required capacities of the depots can also be influenced by
shifting certain amounts of work from one depot to another.
However, the main measures to be taken for creating a matching involve the
capacities of the depots. Relevant measures include: hiring additional crews,
training crews so that they become more flexible, or moving crews from one
depot to another. Financial measures can be taken to stimulate people to re-
tire earlier. In extreme cases, crews can be fired. Agreements about working
conditions, workloads, and salaries should be such that the crews are satisfied
in their work, since the crews are an important factor for providing services of
a high quality.
An important aspect to be taken into account in crew management is the
relatively long throughput time of the process of hiring additional crews until
they are fully operational. This time varies from about one year for conductors
to about two years for drivers. The length of the throughput time is mainly due
to the required training for new employees, both theoretical and on-the-job.
Obviously, hiring new employees has a long lasting effect on the capacities of
the involved depots.

2 Line planning

Many European operators of passenger trains operate a cyclic (also called

“clockfaced”) timetable. That is, a timetable in which the scheduled trains can
be partitioned into a number of subsets, where the trains in each subset have
the same routes and the same stop stations. The only difference between the
trains in each subset are the arrival and departure times: for two consecutive
trains in the same subset and for each station in their common set of stop sta-
tions, the difference of the departure times of the trains in the same direction
is one cycle time (or half a cycle time). A similar statement also holds for the
arrival times of the trains.
Each subset of trains in such a partitioning is called a line. The trains in a line
differ from each other only in their arrival and departure times. The frequency
of a line denotes the number of trains that run in each direction per cycle on
134 A. Caprara et al.

their common route. A cycle time that one encounters frequently in practice is
the hour.
Usually, the stop stations of a line are based on the status of the line and on
the status of the stations: there are m types of stop stations and m types of lines,
and the trains of a line of type τl dwell at the stations of type τs if and only if
τl  τs . For example, there may be Intercity, Interregional, and Regional sta-
tions, and the same line types. Then the trains of an Intercity line dwell at the
Intercity stations only, and the trains of a Regional line dwell at all stations
underway. A line system based on the foregoing assumptions is called a het-
erogeneous line system. Recently, also other line systems have been designed,
such as homogeneous line systems. However, these will not be discussed in this
chapter.
The Line Planning Problem (LPP) is the problem of designing a line system
such that all travel demands are satisfied and certain objectives are met. There
are two main conflicting objectives to be pursued when planning a line system,
namely:
(i) maximizing the service towards the passengers, and
(ii) minimizing the operational costs of the railway system.
In the following we define as direct passengers the passengers that can travel
from their origin station to their destination station without having to change
trains. Maximizing the number of direct passengers usually results in long lines:
the longer a line, the more direct connections are provided. However, long
lines may transfer delays more easily over a wider geographical area, and they
may prohibit an efficient allocation of rolling stock: the rolling stock is usually
allocated according to the peak demand along the line. Therefore, in a robust
and cost-optimal line system, the lines are relatively short, which may force the
passengers to transfer from one train to another relatively often. Thus an ac-
ceptable trade-off between the two objectives has to be made. Both objectives
have been studied, but the combination of the two in one model has not been
described yet.
When designing a line system, one has several options for providing suf-
ficient capacity to transport all passengers: lines can be operated with a high
frequency and with trains with a relatively low capacity, or with a low frequency
with trains with a relatively high capacity. Here also the choice between the
allocation of Single Deck or Double Deck rolling stock is relevant. A high fre-
quency on a line also adds to the service towards the passengers.
Finally, it should be noted that in practice there is a bidirectional relation
between the travel demand and the operated line system. On the one hand, the
operated line system should be such that the full travel demand can be accom-
modated. On the other hand, a line system that provides a high service towards
the passengers may also attract additional passengers, thereby increasing the
original travel demand. Thus, the travel demand also depends on the operated
line system.
 Ch. 3. Passenger Railway Optimization 135

Literature review

There is a limited set of papers dealing with LPP. Most of them deal with
one line type only. None of the papers distinguishes between the utilization of
Single Deck or Double Deck train units on a line, although this seems to be a
relevant choice.
Dienst (1978) describes a branch-and-bound method to find a line system
with a maximum number of direct passengers. He assumes that all trains have
an infinite capacity, i.e., if there is a line between two nodes, then this line
can accommodate all passengers that want to use this line. Thereby the need
for a frequency greater than one on a line is eliminated, but it is questionable
whether the capacity of the obtained line plan will be sufficient to accommo-
date all passengers in practice.
Bussieck et al. (1996) also search for a line system with a maximum number
of direct travelers. They use decision variables denoting the frequency of each
line and assume that all trains have the same fixed capacity. In order to re-
duce the number of decision variables, they use an aggregation of the decision
variables. This aggregation requires the capacity constraints of the trains to be
relaxed. The model is solved by first applying several preprocessing techniques,
and then by applying a general-purpose ILP solver. The authors describe sev-
eral valid inequalities to improve the LP lower bound. Bussieck (1998) further
extends these methods.
Claessens et al. (1998) study the problem of finding a minimum cost line
system. They start with a nonlinear mixed integer model involving binary de-
cision variables for the selection of the lines and additional variables for the
frequencies and the train lengths. Since the nonlinearity of the model leads to
computational problems, the authors switch to an ILP. The model is solved by
a general-purpose ILP solver, after applying several preprocessing techniques.
Goossens et al. (2004) also focus on the design of a minimum cost line sys-
tem. Their model is similar to the model described by Claessens et al. (1998).
In Goossens et al. (2004) a branch-and-cut approach to solve LPP is described.
The main ingredients of the algorithm are preprocessing techniques combined
with several classes of valid inequalities to improve the LP lower bounds and
a number of strategies for variable selection and branching during the branch-
ing process. Goossens et al. (2005) consider the problem of designing a line
system for several line types simultaneously. In order to reduce the number of
decision variables, they combine and disaggregate the origin/destination flows
of the passengers. This is justified by the fact that they only consider the objec-
tive of cost minimization and that the consequences for the passengers are not
taken into account.
Lindner (2000) also studies the minimization of the costs of the line system.
He develops a branch-and-bound method for finding a cost-optimal line sys-
tem. His model also integrates the minimum cost LPP with a model for finding
a cyclic timetable. The timetabling part of the model is based on the model
136 A. Caprara et al.

for the Periodic Event Scheduling Problem described by Serafini and Ukovich
(1989), see Section 3.2.
Recently, also the objective of minimizing the number of transfers from one
train to another has been studied by Scholl (2005). Note that minimizing the
number of transfers from one train to another is more complex than maximiz-
ing the number of direct passengers, since basically for each origin/destination
pair the associated path through the network has to be followed. Scholl (2005)
develops a so-called “Switch-and-Ride” network. For real-life instances such
networks are huge. Scholl (2005) uses Lagrangian relaxation for obtaining
lower bounds, and several heuristic methods for generating feasible solutions.

Model formulation

LPP is a strategic planning problem, since the line system is the basis of
the railway services provided by a Train Operator: the variable costs of a Train
Operator are determined to a large extent by the design of the line system.
The same holds, to a smaller extent, for the Train Operator’s income. As a
consequence, LPP is a highly complex problem in which many uncertainties
have to be dealt with. For modeling LPP many assumptions have to be made,
which are described next.

Assumptions
The basic input for LPP consists of the global structure of the railway in-
frastructure and the expected demand for railway transportation, e.g., per
hour. We assume that the forecasted travel demand is represented by a given
origin/destination matrix. The bi-directional relationship mentioned earlier
may be handled by applying an optimization model iteratively.
Furthermore, in practice the travel demand is usually not stationary nor
symmetric: there are peak hours and off-peak hours, where especially during
the peak hours the travel demand has a dominating direction. Nevertheless,
line systems are usually symmetric in practice. That is, for each pair of stations
s1 and s2 , the number of direct trains from s1 to s2 is more or less the same as
the number of direct trains from s2 to s1 . This is due to the fact that the rolling
stock is usually operated on a line-by-line basis and circulates along its line.
Due to the principle of flow conservation for the rolling stock, such a system
does not allow for large deviations from a symmetric line system.
Since usually the peak hours are the bottlenecks in a railway system, Train
Operators are mainly interested in designing a line system that can accommo-
date the travel demand during the peak hours. Therefore, the capacities of
the trains should be such that they can accommodate both the passenger flows
during the morning peak hours and the passenger flows during the afternoon
peak hours. Hence, if dsm1 s2 and dsm2 s1 denote the passenger flows per hour be-
tween stations s1 and s2 during the morning peak, and dsa1 s2 and dsa2 s1 denote
the passenger flows per hour between stations s1 and s2 during the afternoon
 Ch. 3. Passenger Railway Optimization 137

peak, then the lines and the corresponding line capacities should be such that
 
ds1 s2 = ds2 s1 = max dsm1 s2  dsm2 s1  dsa1 s2  dsa2 s1
passengers can be transported between stations s1 and s2 during the peak
hours. This property should hold for all pairs of stations s1 and s2 simultane-
ously. Note that the models may be adapted to differentiate between the travel
demand during the morning peak and during the afternoon peak, and possibly
also between the travel demand during the off-peak hours.
Furthermore, it is usually assumed that each passenger uses a pre-specified
path through the network. This is not a very strong assumption, since usually
each passenger’s path through the network is specified by the ticket regula-
tions: each passenger is supposed to travel along the shortest-distance path
through the network from her/his origin to her/his destination. This assump-
tion allows one to compute a priori the number of passengers on each edge in
the railway network based on the origin/destination matrix.
Next, although it is also possible to deal with different line types simultane-
ously, see Goossens et al. (2005), the passenger flows are usually assumed to be
split per line type. A consequence is that the different line types can be consid-
ered apart from each other. For generating such a split of the passenger flows
several methods can be used, e.g., the System Split procedure described by
Oltrogge (1994) and by Bouma and Oltrogge (1994). This procedure assumes
that each passenger switches to a faster train as soon as possible, and switches
back to a slower train as late as possible. For example, if a passenger travels
from Regional station s1 to Intercity station s3 , where Intercity station s2 is the
first Intercity station on the shortest path from s1 to s3 , then this passenger is
assumed to travel from station s1 to station s2 in one or more Regional trains,
and from station s2 to station s3 in one or more Intercity trains.
Finally, it is assumed that the lines are simple lines. That is, each line is de-
fined as a simple path in the railway network. In practice, lines may have a more
complex structure, for example, due to underway splitting and combining. An
advantage of this structure is that splitting and combining of trains increases
the number of direct connections without leading to an additional utilization
of the infrastructure on the common part of the route. Note that these more
complex lines may also be handled by the model described in the following.

The model
We describe a formulation for a single line type, taking into account both
the objective of maximizing the number of direct passengers and the objective
of minimizing the operational costs.
We assume that the railway network is represented by an undirected graph
G = (V  E), where the nodes v ∈ V represent the stations and the edges e ∈ E
represent the tracks between the stations. Since it is assumed that only a single
line type is considered, the trains are assumed to dwell at all stations under-
way. The set of stations, however, can be subdivided into the set of stations
138 A. Caprara et al.

where a line may start and end, and the set of stations where this is impos-
sible due to a lack of facilities. The travel demand is given by the symmetric
origin/destination matrix defined as follows. The set P denotes the unordered
pairs of stations with a positive demand. If p = (p1  p2 ) ∈ P is such a pair,
then let dp be the number of passengers that want to travel between stations
p1 and p2 . Furthermore, for each p ∈ P, let Ep be the set of edges on the
shortest path between stations p1 and p2 , and with some abuse of notation, we
write de for the total number of passengers that want to travel along edge e,
namely de = p:e∈Ep dp .
We assume that a set L of potential lines is given a priori. The set of edges
of line l ∈ L is denoted by El . The objective is to select an appropriate subset
from the given set of potential lines. For each line, also a certain frequency
and a certain capacity per train are to be selected. This is necessary, because
the costs of a line are strongly related to the provided capacity per train. This
capacity is a combination of a rolling stock type (Single Deck or Double Deck)
and a number of carriages. The different options for the capacity per train are
represented by the set C. The set F denotes the set of potential frequencies.
Usually, a line can be operated one or two times per cycle time. The capacity
of a line equals the capacity per train multiplied by the line’s frequency. The
number of rolling stock units that are required to operate a line depends on
the number of trains on the line and on the lengths of these trains. Further-
more, the number of trains on a line depends on the line’s frequency and on its
circulation time.
Each feasible combination of a line l ∈ L, a frequency f ∈ F, and a capacity
option c ∈ C is denoted by an index i. The set of indices i is denoted by I. The
line index, frequency index and capacity index corresponding to index i are de-
noted by li , fi , and ci , respectively. Thus, the capacity of the line corresponding
to index i in terms of the number of passengers that can be accommodated by
this line equals fi ci . The operational cost associated with option i is denoted
by ki . These costs are mainly determined by the variable train costs (e.g., the
train driver) and by the variable carriage costs (e.g., the conductor(s) and the
carriage kilometers).
The main decision variables in the formulation are the binary variables xi ,
equal to 1 if and only if line li is to be operated with frequency fi and ca-
pacity ci . Additional decision variables are the variables dlp , representing the
number of direct passengers that travel on line l between the pair of stations p.
Now the model for LPP reads as follows:
 
max w1 dlp − w2 ki xi (1)
l∈L p∈P i∈I
subject to

xi  1 l ∈ L (2)
i∈I:li =l
 Ch. 3. Passenger Railway Optimization 139

fi ci xi  de  e ∈ E (3)
i∈I:e∈Eli
 
dlp  fi ci xi  l ∈ L e ∈ El  (4)
p∈P:e∈Ep i∈I:li =l

dlp  dp  p ∈ P (5)
l∈L:Ep ⊂El
xi ∈ {0 1} i ∈ I (6)
dlp  0 l ∈ L p ∈ P (7)
The objective function (1) describes the fact that we want to obtain a balance
between maximizing the number of direct passengers and minimizing the oper-
ational costs. Here w1 and w2 are weights that describe the relative importance
of the two partial objectives. Constraints (2) specify that for each potential line
l ∈ L, at most one appropriate frequency and one appropriate option for the
capacity per train are to be selected. Constraints (3) describe that on each track
e ∈ E the provided capacity should be sufficient to accommodate all passen-
gers that travel on this track e. Next, constraints (4) provide a link between the
two sets of decision variables: for each line l ∈ L and for each edge e ∈ El ,
the total number of direct passengers that travel on line l should not exceed
the provided capacity fi ci on line l. Finally, constraints (5) specify that the to-
tal number of direct passengers between the pair of stations p cannot exceed
the total travel demand between these stations. Note that the integrality of the
variables dlp is not imposed, since this constraint generally does not influence
the obtained results.
There are several additional constraints that may be taken into account. For
example, one may wish to specify that only fixed numbers of Single Deck and
Double Deck rolling stock units are to be allocated to the lines. Another ex-
ample involves capacity constraints for the tracks or the stations: for each track
or station upper bounds on the number of passing trains can be specified. Sim-
ilarly, for each station an upper bound on the number of starting and ending
trains may have to be respected. Such capacity constraints may positively influ-
ence the robustness of the railway system. Service towards the passengers may
be improved by also specifying for each track a lower bound on the number
of passing trains, or by specifying a lower bound on the total number of direct
passengers. These additional constraints can be represented in terms of the
available decision variables in a straightforward way.

Comments
Note that the line system is designed at a very early stage in the planning
process. Therefore, only rough estimates of the travel demand and of the costs
can be made. For example, the cost coefficients ki in the objective function (1)
are mainly based on the characteristics of the peak hours, which may give an
overestimate of the real costs. This may be overcome to some extent by taking
140 A. Caprara et al.

into account a “known” relationship between the number of carriage kilome-

ters during peak hours and during off-peak hours. Also the required number
of train units determined based on the line plan cannot be more than a rough
estimate.
With respect to model (1)–(7), several comments are in place. First, note
that a disadvantage of the model is its size: the numbers of decision variables
and constraints grow very fast with the numbers of origins and destinations and
with the number of potential lines. Therefore, all models that have been stud-
ied in the literature consider either the objective of maximizing the number
of direct passengers, or the objective of minimizing the operational costs. The
size of the model also leads to long computation times: most literature reports
that only approximate solutions could be found within reasonable amount of
computation time.
Second, note that in model (1)–(7) the capacities of the lines may be insuffi-
cient to handle all direct passengers on a direct connection. As a consequence,
the model has to “select” the direct passengers from the available ones. How-
ever, this may lead to an overestimate of the number of direct passengers.
For example, suppose there are 100 passengers between stations s1 and s2 ,
100 passengers between s2 and s3 , and 100 between s1 and s3 . If the direct lines
between stations s1 and s3 have a total capacity of 100 passengers, then the
model’s result will be that there are 200 direct passengers, namely 100 direct
passengers between s1 and s2 and 100 direct passengers between s2 and s3 .
There are 0 direct passengers from s1 to s3 , since such passengers would dete-
riorate the objective. This overestimate of the number of direct passengers is
caused by the fact that the passenger behavior is not represented in the model
in detail. On the other hand, a detailed description of the passenger behav-
ior would make the model even more complex to solve than it is already in its
current form.
In order to overcome this problem to some extent, one may take into ac-
count the lengths of the direct connections in the objective function. That is,
a long direct connection is more valuable than a short one. Alternatively, one
may argue that the capacities of the lines should be such that, if there is at
least one direct connection between a certain pair of stations, then all direct
connections between these stations should be able to accommodate all pas-
sengers that want to travel between these stations. This would reflect the fact
that passengers prefer to have direct connections, and that Train Operators try
to facilitate these through their line capacities as well. In order to model this
alternative set of constraints, for each pair of stations p ∈ P, the decision vari-
ables dlp are to be replaced by a binary variable yp equal to 1 if and only if
there is a direct connection between the two stations of pair p. Now the model
reads as follows:

 
max w1 dp yp − w 2 ki xi (8)
p∈P i∈I
 Ch. 3. Passenger Railway Optimization 141

subject to

fi ci xi  de  e ∈ E (9)
i∈I:e∈Eli

yp  xi  p ∈ P (10)
i∈I:Ep ⊂Eli

xi  yp  p ∈ P l ∈ L Ep ⊂ El  (11)
i∈I:li =l
 
dp yp  fi ci xi  p ∈ P (12)
p ∈P:Ep ⊂Ep i∈I:Ep ⊂Eli

xi ∈ {0 1} i ∈ I (13)

yp ∈ {0 1} p ∈ P (14)
The objective function (8) describes that the passengers between each
pair of stations are either counted completely as direct passengers, or they
are not counted at all. Constraints (10) and (11) provide the link between
the x-variables and the y-variables. Note that constraints (11) imply con-
straints (2). Finally, constraints (12) specify that the capacities of the lines
should be such that, if there is a direct connection between a certain pair of
stations, then all passengers between these stations should be accommodated
on a direct line.

3 Train timetabling

The general aim of the Train Timetabling Problem (TTP) is to provide a

timetable for a number of trains on a certain part of the railway network. As
was indicated already in Section 2, one may distinguish between cyclic timeta-
bles and noncyclic timetables.
An advantage of a cyclic railway system is the fact that such a system’s
timetable is easy to remember for the passengers. For example, at a certain
station, the trains heading for a certain direction always leave at x:12 and x:42.
On the other hand, a drawback is that such a system is expensive to operate.
Even in the periods between the peak hours with low travel demand, more or
less the same timetable is operated as during the peak hours. The only way
to differentiate the system’s capacity between the peak hours and the off-peak
hours is to modify the lengths of the trains. The latter impacts both the vari-
able rolling stock costs and the variable crew costs (shorter trains require less
conductors).

3.1 Noncyclic timetabling

The noncyclic TTP is especially relevant on heavy-traffic, long-distance cor-

ridors, where the capacity of the infrastructure is limited due to greater traffic
142 A. Caprara et al.

densities, and competitive pressure among the Train Operators is expected to

increase in the near future. This allows the Infrastructure Manager to allocate
“optimally” the train paths requested by all Train Operators and proceed with
the overall timetable design process, possibly with final local refinements and
minor adjustments, as in the tradition of railway planners. In brief, this allows
each Train Operator to submit requests for paths on the given railway line,
and allows the Infrastructure Manager to collect all the requests, run the opti-
mization algorithm to allocate (if possible) all of them at maximum profit, and
eventually respond to the Train Operators with the proposed plan of the time
slot allocation and the relative “access fees”.
The essential characteristics of the process can be summarized as follows:
• Each Train Operator has associated with each train a profit (i.e., a pri-
ority), an ideal timetable, with an ideal departure time, and tolerances
within which it can be changed.
• The optimal allocation is found by maximizing the overall profit for the
Infrastructure Manager, i.e., the difference between the profits of the
scheduled trains and a cost penalty function, which takes into account
the deviations of the final timetables with respect to the ideal ones.
In addition, under the assumption of a competitive market, the process can
be iterated if some Train Operator does not accept the solution and asks for
a re-evaluation by the Infrastructure Manager, e.g., by using modified path
profits.

Literature review

TTP has received considerable attention in the literature. Many references

consider MILP formulations in which the arrival and departure times are rep-
resented by continuous variables and there are binary variables expressing the
order of the train departures from each station.
Szpigel (1973) considers a variant of these models in which the order of the
train departures from a station is not represented by binary variables but by
disjunctive constraints. Small instances of the problem are solved by branch-
and-bound by computing bounds through the relaxation of these disjunctive
constraints. Jovanovic and Harker (1991) solve, by branch-and-bound tech-
niques, a version of these models that calls for a feasible schedule rather than
for the optimization of a suitable objective function. Cai and Goh (1994) illus-
trate a constructive greedy heuristic driven by one of these models. Carey and
Lockwood (1995) define a heuristic that considers the trains one at a time (in
appropriate order), and for each train solves a MILP analogous to these mod-
els in order to schedule the train optimally, keeping the path of the previously
scheduled trains partially fixed. More precisely, the relative order of the train
departures for these trains is kept fixed, whereas their arrival and departure
times may be changed. Higgings et al. (1997) define local search, tabu search,
 Ch. 3. Passenger Railway Optimization 143

genetic and hybrid heuristics, finding a feasible solution by using a model in

the family above.
Brännlund et al. (1998) discretize the time into one-minute time slots and
subdivide the line into blocks. Operational constraints impose that two trains
cannot be in the same block in the same time slot. They define an ILP model
with a binary variable xsbt each time the timetable constraints allow train t to
be in block b in time slot s. This model is not suited for large size instances as
those arising for the main European corridors.
Oliveira and Smith (2000) model TTP as a special case of the Job-Shop
Scheduling Problem, considering trains as jobs to be scheduled on lines re-
garded as resources, and present a hybrid algorithm devised under the Con-
straint Programming paradigm, showing how to adapt this framework in some
special real-life applications.
Different ILP models based on a graph representation of the problem were
presented by Caprara et al. (2002, 2006). In both papers, time is discretized
(i.e., expressed in minutes from 1 to 1440) and Lagrangian relaxation is used
to derive bounds on the optimal solution value as well as to drive a heuristic
procedure. This approach, whose main features are outlined in the following,
produced good relaxations and heuristic solutions also for large-size instances.

Model formulation

In this section we consider the basic version of noncyclic TTP, which con-
siders a single, one-way line linking two major stations, with a number of
intermediate stations in between. Let S represent the set of stations, ordered
according to the sequence in which they appear along the line for the running
direction considered, and T denote the set of trains.
A timetable defines, for each train t ∈ T , the departure time from its first
station ft ∈ S, the arrival time at its last station lt ∈ S, and the arrival and
departure times for the intermediate stations ft +1     lt −1. Each train t ∈ T
is assigned on input an ideal timetable with departure time dts for each station
s ∈ {ft      lt − 1} and arrival time ats for each station s ∈ {ft + 1     lt },
which would be the most desirable timetable for the train, that may however
be modified in order to satisfy the line capacity constraints. In particular, one
is allowed to slow down each train with respect to its ideal timetable, and/or to
increase the stopping time interval at the stations. Moreover, one can modify
the departure time of each train from its first station, or even cancel the train.
The final solution for the problem will be referred to as the actual timetable.
The line capacity constraints impose that overtaking between trains occurs
only within a station. To this end, a train is allowed to stop in any intermediate
station (even if the ideal timetable does not include a stop in that station) to
give the possibility to some other train to overtake it. Furthermore, for each
station, there are lower bounds on the time interval between two consecutive
arrivals and two consecutive departures, respectively. Assuming the speed of
a train on a line segment to be constant, this last constraint implicitly imposes
144 A. Caprara et al.

a minimum time interval between two consecutive trains in the line segment
connecting two consecutive stations. Note that line capacity constraints, along
with the fact that the actual timetable has to be repeated every day, may force
some trains to be canceled to obtain a feasible solution.
The objective is to maximize the sum of the profits of the scheduled trains,
defined as follows. The profit achieved for each train t ∈ T depends on the
train’s ideal profit πt , on the shift νt , defined as the absolute difference between
the departure times from the first station in the ideal and actual timetables, and
on the stretch μt , defined as the (nonnegative) difference between the total
travel times in the actual and the ideal timetables. According to the charg-
ing rules generally adopted by the Infrastructure Manager, the profit for each
train t is given by
πt − φt (νt ) − γt μt  (15)
where φt (·) is a user-defined nondecreasing function penalizing the train shift
(with φt (0) = 0), and γt is a given nonnegative parameter (i.e., the function
penalizing the train stretch is assumed to be linear). Typically, the profit func-
tion is identical for trains of the same type (e.g., intercity trains) running in the
same time interval of the day. If the profit of train t turns out to be nonpositive,
it is better not to schedule train t, i.e., to cancel it.

The model
We next outline a mathematical formulation of the problem, calling for a
maximum-profit set of paths in a multigraph, as proposed in Caprara et al.
(2002). We refer to that paper for a detailed description. Let G = (V  A)
be the directed acyclic multigraph in which nodes correspond to arrivals and
departures from the stations along the line at some instant (recall that times
are discretized) and arcs correspond both to train stops within a station and to
train trips from a station to the next one. More specifically, for each station s
except the first one, the nodes associated with an arrival at s at some instant are
denoted by U s , and for each station s except the last one, the nodes associated
with a departure from s at some instant are denoted by W s . The arc set is
partitioned into arc sets At associated with each train t ∈ T . Arcs in At from
a node w ∈ W s−1 to a node u ∈ U s model train t departing from station
s − 1 at the time instant associated with w and arriving at station s at the time
instant associated with u. Moreover, arcs in At from a node u ∈ U s to a node
w ∈ W s model train t arriving at station s at the time instant associated with
u and departing at the time instant associated with w. In addition, there is one
artificial source node σ, whose leaving arcs represent train departures from
their first station, and one artificial sink node τ, whose entering arcs represent
train arrivals at their last station. The definition of the graph G guarantees that
every path from σ to τ using arcs in At corresponds to a feasible timetable for
train t and vice versa.
The objective function can be modeled by associating, for each train t,
a (possibly negative) profit pa with each arc a ∈ At . The profit associated
 Ch. 3. Passenger Railway Optimization 145

with an overall train timetable is given by the sum of the profits of the arcs in
the paths of G corresponding to the scheduled trains.
Line capacity constraints impose that certain pairs of arcs, associated
with different trains, cannot be selected in the overall solution. In particu-
lar, two arcs are called incompatible if they either correspond to train ar-
rivals/departures too close in time, or if they correspond to an overtaking
between two consecutive stations.
An ILP formulation of TTP is the following. For each t ∈ T and each arc
a ∈ At , introduce a binary variable xa equal to 1 if and only if arc a is selected
in an optimal solution, i.e., the path in the solution associated with train t con-
tains arc a. For notational convenience, for each node v ∈ V and each train
t ∈ T , let δ+ − t
t (v) and δt (v) denote the (possibly empty) sets of arcs in A leav-
ing and entering node v, respectively. Finally, let C denote the (exponentially
large) family of maximal subsets C of pairwise incompatible arcs. Then the
model reads as follows:

max pa xa (16)
t∈T a∈At
subject to

xa  1 t ∈ T (17)
a∈δ+
t (σ)
 
xa = xa  t ∈ T v ∈ V \ {σ τ} (18)
a∈δ−
t (v) a∈δ+
t (v)

xa  1 C ∈ C (19)
a∈C
xa ∈ {0 1} a ∈ A (20)
The objective function (16) is defined as the sum of the profits of the arcs
associated with each path in the solution. Constraints (17) impose that at most
one arc associated with a train is selected among those leaving the source
node σ, while constraints (18) impose equality on the number of selected arcs
associated with a train entering and leaving each arrival or departure node.
Consequently, the set of selected arcs associated with a train can either be
empty, or define a path from the source to the sink. Finally, clique constraints
(19) forbid the simultaneous selection of incompatible arcs, imposing the line
capacity constraints.
Note that (16)–(20) is indeed a Multicommodity Flow formulation, in which
the commodity index is hidden in the multigraph definition. It can be shown
that the associated LP relaxation can be solved efficiently through a cutting
plane approach. On the other hand, the very large number of variables for
real-life instances of the problem suggests to use an alternative approach. The
one used in Caprara et al. (2002) is based on a reformulation in which the
line capacity constraints are modeled using variables associated with the nodes
146 A. Caprara et al.

of G, and then relaxed in a Lagrangian way. The main advantage of the alter-
native formulation is that the Lagrangian relaxation leads to a relaxed problem
in which the profits of the x variables are unchanged, whereas Lagrangian
penalties are associated with the nodes of G, which are much easier to han-
dle than penalties associated with the arcs. The resulting Lagrangian problem
then calls for a set of paths for the trains, each having maximum Lagrangian
profit, given by the sum of the original profits for the arcs in the path, minus
the sum of the Lagrangian penalties of the nodes visited by the path. If the
maximum Lagrangian profit of a path for a train is nonpositive, then the train
is not scheduled by the optimal solution of the relaxed problem.
The approach in Caprara et al. (2002) determines near-optimal Lagrangian
multipliers through subgradient optimization, and applies a constructive
heuristic procedure to determine feasible solutions at each subgradient iter-
ation. In this procedure, trains are scheduled according to decreasing values
of the Lagrangian profits of their paths, and each train is assigned a timetable
corresponding to the maximum Lagrangian profit path compatible with the
previously scheduled trains. The very large number of line capacity constraints
that are relaxed in a Lagrangian way is handled according to a so-called relax-
and-cut framework, explicitly considering each constraint only when it turns
out to be violated by the relaxed solution at some iteration of the subgradient
procedure.
A discussion of how the model can be modified to handle additional fea-
tures of real-world applications can be found in Caprara et al. (2006). These
features include manual block signaling for managing at most one train on a
line segment between two consecutive stations, a maximum number of trains
that can be present in a station at the same time, a prescribed timetable for a
subset of the trains, and maintenance operations that occupy a line segment
for a given period.

Experimental results

We report some computational experiments obtained by applying the basic

model described above on a set of instances from FS Rete Ferroviaria Italiana,
the Italian Infrastructure Manager. The program was run on a Digital Ultimate
Workstation 533 MHz.
The function penalizing the shift in time of train t is defined as φt (νt ) :=
αt νt , i.e., the penalty is linear in the shift. The profit coefficients (πt , αt , and γt )
are identical for the trains of the same type and are reported in Table 1. The
main characteristics of the instances are outlined in Table 2, showing the num-
ber of trains on input (# trains) followed by an array with the number of
Eurostar, Euronight, Intercity, Express, Direct, Local, and Freight trains.
Table 3 reports the results obtained with a limit of 1000 subgradient iter-
ations, recording the best heuristic solution value and the best upper bound
found before this limit. In the table, we indicate: the sum of the profits achiev-
able by scheduling each train according to its ideal timetable (ideal prof.); the
 Ch. 3. Passenger Railway Optimization 147

Table 1.
Train profit coefficients depending on the train type

Train type πt αt γt

Eurostar 200 7 10
Euronight 150 7 10
Intercity 120 6 9
Express 110 5 8
Direct 100 5 8
Local 100 5 6
Freight 100 2 3

Table 2.
Characteristics of the instances

Instance First stat. Last stat. # stat. # trains

BN-BO Brennero Bologna 40 68 (1, 0, 5, 13, 11, 38, 0)

MU-VR Munich Verona 48 54 (0, 0, 0, 7, 0, 47, 0)
CH-RM Chiasso Rome 73 36 (15, 0, 9, 0, 0, 7, 5)
MO-MI-1 Modane Milan 39 16 (1, 0, 3, 0, 4, 5, 3)
MO-MI-2 Modane Milan 39 23 (0, 3, 1, 11, 0, 2, 6)
PC-BO-1 Piacenza Bologna 16 221 (28, 3, 52, 35, 28, 14, 61)
PC-BO-2 Piacenza Bologna 16 93 (12, 3, 18, 10, 20, 6, 24)
PC-BO-3 Piacenza Bologna 16 60 (12, 1, 14, 0, 12, 7, 14)
PC-BO-4 Piacenza Bologna 16 40 (6, 0, 10, 0, 12, 2, 10)
BZ-VR Bolzano Verona 20 128 (34, 0, 1, 10, 11, 38, 34)
CH-MI Chiasso Milan 15 194 (20, 1, 29, 8, 19, 66, 51)

best upper bound found by the subgradient optimization procedure (best UB)
with, in brackets, the percentage improvement over ideal prof.; the solution
value found by scheduling the trains by decreasing values of πj (breaking ties
arbitrarily) and assigning to each train the timetable corresponding to the max-
imum profit path compatible with the previous trains (greedy sol.); the value of
the best solution found by the Lagrangian heuristic (best sol.) with, in brack-
ets, the percentage improvement over greedy sol.; the percentage gap between
the value of the best heuristic solution and the best upper bound (GAP%); the
number of trains scheduled in the best solution (# sched); the average shift
and the average stretch (in minutes) for the trains scheduled in the best solu-
tion (avg. ν and avg. μ); the overall running time in seconds needed to find the
best solution (time).
The manual methods proceed in a way similar to the one used to derive the
solution whose value is reported in the greedy sol. column, therefore the quality
of the solution provided by the practitioners is typically close to the value given
in this entry.
 148
Table 3.
Results on real-world instances

Instance Ideal prof. Best UB Greedy sol. Best sol. GAP% # sched. Avg. ν Avg. μ Time

BN-BO 7130 6891 (3.4%) 6746 6779 (0.5%) 17% 68 06 03 34

A. Caprara et al.
MU-VR 5470 4991 (8.8%) 3332 4208 (20.8%) 186% 48 14 11 92
CH-RM 5280 5129 (2.9%) 4844 4871 (0.6%) 53% 35 19 13 48
MO-MI-1 1760 1713 (2.7%) 1648 1684 (2.1%) 17% 16 06 04 0
MO-MI-2 2580 2533 (1.8%) 2486 2520 (1.3%) 06% 23 09 03 78
PC-BO-1 25740 24,142 (6.2%) 19535 21,397 (8.7%) 128% 192 08 08 672
PC-BO-2 11010 10,947 (0.6%) 10848 10,882 (0.3%) 06% 93 00 02 118
PC-BO-3 7450 7222 (3.1%) 6709 7119 (5.8%) 14% 60 06 08 271
PC-BO-4 4800 4130 (14.0%) 3350 3656 (8.4%) 130% 35 32 09 266
BZ-VR 16300 16,101 (1.2%) 15902 16,003 (0.6%) 06% 127 03 00 266
CH-MI 21930 21,467 (2.1%) 20923 21,215 (1.4%) 12% 193 07 01 285
 Ch. 3. Passenger Railway Optimization 149

The method succeeded in scheduling almost all trains for most instances,
the exceptions being instances MU-VR, PC-BO-1, and PC-BO-4. Moreover,
for most instances, the final percentage gap is quite small (less than 2%). For
the remaining cases, the best heuristic solution is considerably better than the
greedy one (with the only exception of instance CH-RM).

3.2 Cyclic timetabling

In a cyclic timetable, each trip is operated in a cyclic way. That is, each
period of the timetable is the same. If the cycle time of the timetable is denoted
by T , then this means that, if a trip between stations s1 and s2 leaves at time t1
and arrives at time t2 , then analogous trips are carried out with departure and
arrival times t1 + kT and t2 + kT for all integer values of k.
The first ones to develop a model for generating cyclic timetables were
Serafini and Ukovich (1989); in that paper, a mathematical model for the so-
called Periodic Event Scheduling Problem (PESP) is presented. In PESP, a set
of repetitive events is scheduled under cyclic time window constraints. Conse-
quently, the events are scheduled for one cycle in such a way that the cycle can
be repeated. Most models for cyclic TTP are based on PESP.
A timetable consists of a number of processes, such as running between two
stations, dwelling at a station, passenger connections etc. The start and end
times of these processes are the events of the timetable. The set of events is
denoted by E. In PESP, for each e ∈ E, the decision variable ve represents the
time instant at which event e has to be scheduled. All constraints that have to
be satisfied by these decision variables specify minimum and maximum process
times for the corresponding processes between the events. As a consequence,
all constraints have the following form:

lef  (ve − vf ) mod T  uef  (21)

Here e and f are two events in the timetable, and lef and uef are appropri-
ate lower and upper bounds for the process time of the process between the
events e and f . For example, if e and f are the departure and arrival of a train
at two consecutive stations, then lef and uef denote the minimum and maxi-
mum running time of the train on this trip, respectively. The modulo operator
models the cyclicity of the timetable. For instance, if the cycle time equals 60
minutes, then the process time from t = 55 to t = 5 equals 10 minutes.
Since the modulo operation in (21) is relatively hard to handle in optimiza-
tion methods, it can be replaced by introducing a binary variable qef for each
constraint of the form (21), and by replacing (21) by the following constraint:

lef  ve − vf + T qef  uef  (22)

Variables qef make the PESP quite hard to solve by standard branch-and-
bound methods: due to the relatively large coefficient T in (22), the LP relax-
ations of models based on this formulation of PESP are quite weak.
150 A. Caprara et al.

Therefore Schrijver and Steenbeek (1994) develop a constraint propaga-

tion algorithm for solving PESP. Their algorithm has been implemented in the
DONS system, that has become an indispensable tool in the Dutch long term
railway timetabling process of Netherlands Railways and ProRail. Schrijver
and Steenbeek also develop local optimization techniques to improve a fea-
sible solution for fixed values of the variables qef in (22). Instances with up to
250 trains (all trains running in one hour of the Dutch timetable) can be solved
usually within reasonable computing time.
In order to cope with the weak LP relaxation of models based on con-
straints (22), Nachtigall (1999), Lindner (2000), and Peeters (2003) also
describe a formulation of PESP based on cycle bases. This formulation is some-
what easier to solve than the formulation based on constraints (22), because of
the lower number of integer variables and the somewhat better LP relaxation.
Nachtigall and Voget (1996) use PESP to generate cyclic timetables with
minimal passenger waiting times. Odijk (1996) uses the PESP at a strategic
level to determine the capacity of the infrastructure around railway stations.
Kroon and Peeters (2003) describe a PESP model including variable trip times.
This may result in additional flexibility leading to a higher probability of obtain-
ing a feasible solution.
Finally, Kroon et al. (2005) describe a stochastic optimization variant of
PESP. Their model explicitly takes into account stochastic disturbances of
the railway processes, distinguishing between a planned timetable and several
realizations of the timetable under pre-determined stochastic disturbances.
The model can be used to allocate time supplements and buffer times to the
processes in the planned timetable in such a way that the average delay of the
realizations of the trains is minimal. In order to keep the computation times at
an acceptable level, they start with an existing timetable and they fix the vari-
ables qef in (22). They show that, by taking into account stochastic disturbances
in the design of the timetable, an increase in robustness can be achieved.

4 Train platforming

The definition of the optimal timetables discussed in Section 3 does not take
into account the actual routing of the trains within the stations considered.
As briefly mentioned in that section, only an upper bound on the maximum
number of trains that can be simultaneously present in a station is imposed.
Routing a train in a railway station means finding for each train a path from
the point where it enters the station to the point where it leaves the station.
Thereby a train usually passes through (and possibly stops at) a platform within
the station. For this reason, the problem is generally referred to as the Train
Platforming Problem (TPP).
While the problem is very easy to solve for relatively small stations, in which
there is a very small number of alternative paths to route the trains, it becomes
challenging for major stations, that typically have very complex topologies.
 Ch. 3. Passenger Railway Optimization 151

The customary problem input for a given station contains a set of directions
for train arrivals and departures, and a set of platforms for train stops. There
may also be dummy platforms representing the possibility to traverse the sta-
tion without stopping at any real platform. Moreover, there is a set of trains,
each associated with scheduled (or ideal) arrival and departure times, arrival
and departure directions, and a set of platforms to which the trains may be
assigned (possibly without stopping). If a train comes from (or goes to) the
shunting area, there is generally no scheduled arrival time, but there may be a
maximum time by which the train should arrive at (or depart from) its platform.
Finally, the (complex) station topology is generally represented by defining a
set of (bidirectional) routes joining each direction to the set of platforms that
can be reached by that direction. The physical overlap between two routes, that
may correspond to the routes sharing either a line segment or a node (or cross-
ing point) is indicated by an incompatibility relation between the two routes. It
is assumed that each train occupies its associated route for a given time inter-
val, corresponding to the time required by the train to go from its direction to
its platform, or vice versa.
There are two types of operational constraints. The first imposes that, for
each platform, a minimum time interval must elapse between the departure of
a train from the platform and the arrival of the next train at the platform. The
second forbids the occupation of an incompatible pair of routes by two trains
in the same time instant. In some cases, for major stations with great traffic
densities, this latter type of constraints is relaxed, allowing two incompatible
routes to be used for an interval whose length does not exceed a given upper
bound, at the cost of paying a suitable penalty. However, in the version of the
problem discussed here we will not consider this possible relaxation.
The problem requires the specification, for each train, of the platform to
which it is assigned along with the associated arrival and departure routes and
the actual arrival and departure times that may be different from the ideal
ones at the cost of paying a penalty. Generally, using a terminology analogous
to that used in Section 3, a maximum (positive or negative) shift with respect to
these ideal times is allowed. Moreover, note that the choice of the arrival and
departure routes implicitly defines the platform to which the train is assigned.
As different platforms within those feasible for the train may have different
priorities, a penalty is paid if the train is not assigned to its highest-priority
platform.

Literature review

TPP has not received considerable attention in the literature so far. De Luca
Cardillo and Mione (1998) consider the simplified version in which, for each
train, the scheduled arrival and departure times cannot be changed, and the
arrival and departure routes are uniquely determined by the choice of the plat-
form. In this case, one may avoid considering the routes explicitly, implicitly
152 A. Caprara et al.

representing their incompatibilities by defining a list of incompatible train-

platform pairs of the form (j a k b), stating that it is infeasible to assign
train j to platform a as well as train k to platform b. This version of the problem
is modeled as a Graph Coloring Problem with additional constraints, for which
a heuristic algorithm is proposed and applied to real-world instances. The same
version of TPP is addressed by Billionnet (2003), who considers a classical ILP
formulation for Graph Coloring and shows how to incorporate into the model
the clique constraints associated with the list of incompatible train-platform
pairs. Randomly-generated instances are solved by using a general-purpose
ILP solver.
A more general version of the problem, in which arrival and departure
times and arrival and departure routes are not fixed a priori is addressed
in Zwaneveld (1997), Zwaneveld et al. (1996), Zwaneveld et al. (2001), and
Kroon et al. (1997). Actually, these authors distinguish among inbound, plat-
form, and outbound routes, meaning the path followed by a train to enter the
station (without going to a platform), the path inside the station stopping at
a platform, and the path to exit the station, respectively. Although in practice
each train has to be assigned all these three paths, the authors consider the
possibility that only some of these (possibly none) are found for the train. This
version is modeled as a Stable Set Problem on a graph in which each node
corresponds to a choice for a train, with an associated cost, and edges join
node pairs associated with the same train as well as node pairs correspond-
ing to incompatible choices for different trains. This model is analogous to
that presented in the next section. The problem is naturally formulated as an
ILP with one binary variable for each node and clique inequality constraints.
The above mentioned references describe heuristic and exact (branch-and-cut)
algorithms based on the solution of the corresponding LP relaxation. The ap-
proach is embedded within the STATIONS system used by the Dutch railway
company to solve the problem. Results for real-world instances corresponding
to the main Dutch stations are reported in Zwaneveld (1997).
The version addressed in Carey and Carville (2003) is intermediate between
the two versions above, in that the arrivals and departure times can be changed
but the assignment of a train to a platform uniquely determines the routes that
the train will follow on its way to and from the platform. The paper discusses
in great detail the conflicts that may arise from the assignments and the pro-
cedures that are followed to evaluate the costs. Rather than a mathematical
formulation, the authors describe a constructive heuristic that is applied to a
real-world instance from the British railway company concerning the station of
Leeds.

Model formulation

As anticipated, we describe a version of TPP analogous to the one consid-

ered by Zwaneveld et al. (1996). As for TTP in Section 3, times are discretized
and expressed in minutes.
 Ch. 3. Passenger Railway Optimization 153

Let P denote the set of platforms in the station, D the set of possible di-
rections for train arrivals and departures, and R the set of all (bidirectional)
routes. For each direction d ∈ D and platform p ∈ P, let Rdp ⊆ R denote the
set of routes linking direction d to platform p and vice versa. For each route
r ∈ R, a list Ir ⊆ R of incompatible routes is given, specifying the routes that
cannot be occupied by a train when r is occupied.
Let T denote the set of trains which have to be assigned to a platform in
the time horizon considered. Each train t ∈ T is associated with an arrival
direction dta ∈ D, a departure direction dtd ∈ D, and a set Ct ⊆ P of candidate
platforms; for each candidate platform p ∈ Ct a nonnegative (possibly null)
penalty cpt for the assignment of train t to platform p is given. If train t is
assigned a platform p ∈ / Ct , the penalty is c0t .
In addition, each train t ∈ T has an associated ideal arrival time uat at a
platform, along with a maximum arrival shift sta , and an associated ideal depar-
ture time udt from the platform, along with a maximum departure shift std . As
the values of these shifts are typically small, often any combination of actual
arrival and departure times that do not differ from the ideal ones by more than
the corresponding maximum shift is feasible (also allowing, e.g., an arrival sta
minutes after the ideal time and a departure std minutes before the ideal time).
On the other hand, the mathematical model presented in the following can be
adapted to handle any constraint on the combination of actual arrival and de-
parture times. Penalties cta and ctd for each minute of shift in the arrival and
departure time, respectively, are incurred. As already mentioned, there are
trains that arrive from the depot or depart for the depot, for which the values
cta or ctd are “small” (or null) and the values of sta and std are “large”.
The problem requires to define, for each train t ∈ T , a path, which is spec-
ified by a platform p ∈ Ct , an arrival route r1 ∈ Rdta p , a departure route
r2 ∈ Rdd p and the corresponding arrival and departure times, within sta and std
t
from the ideal ones, respectively, so as to minimize the corresponding penal-
ties. Let Pt denote the set of all feasible paths for train t and, for each path
P ∈ Pt , qtP be the associated penalty, computed according to the above rules.
Even for large size instances, the number of possible paths for a train is rela-
tively small, and all these paths can be generated and considered explicitly, due
to the relatively small number of routes joining directions to platforms and the
limited maximum shifts from the ideal arrival and departure times, that are
associated with a timetable determined in previous phases. According to the
operational constraints, two paths are incompatible if they either occupy the
same platform for time periods that overlap or are too close in time, or use
incompatible routes for the same time instant, with the convention that each
arrival route r ∈ R is occupied by a train t ∈ T for a time wrt a ending at the

arrival of the train at the platform, and each departure route r ∈ R is occupied
by a train t ∈ T for a time wrt d starting from the departure of the train from the

platform.
154 A. Caprara et al.

As customary, the incompatibilities between train paths can be represented

by an incompatibility graph G = (V  E) having one node (t P) for each train
t ∈ T and path P ∈ Pt and edges joining nodes associated either with the same
train or with paths that are incompatible. Clearly, each feasible solution of the
problem corresponds to a stable set of G that contains one node associated
with each train in T , and vice versa. Let C denote the (exponentially large)
family of maximal cliques C of G.
The natural ILP model is obtained by introducing for each node (t P) of G
a binary variable xtP whose value is 1 if and only if train t is assigned path P.
The model reads as follows:

min qtP xtP (23)
t∈T P∈Pt
subject to

xtP = 1 t ∈ T (24)
P∈Pt

xtP  1 C ∈ C (25)
(tP)∈C
xtP ∈ {0 1} t ∈ T P ∈ Pt  (26)
The objective function (23) simply minimizes the overall cost. Constraints
(24) ensure that each train is assigned a path, while constraints (25) forbid the
selection of incompatible paths.
The model considered by Zwaneveld et al. (1996) and Zwaneveld et al.
(2001) is analogous to the one above, with the difference that a train path
is split into three parts as discussed above and a train may not be assigned
all these three parts (very loosely speaking, constraint (24) is an inequality).
The above references illustrate preprocessing rules that successfully reduce
the number of variables and a procedure to generate a relevant subset of
polynomially-many clique inequalities (25). All studied instances related to
several stations in the Netherlands with up to 50 trains per hour can be solved
within small computing time.

5 Rolling stock circulation

The Rolling Stock Circulation Problem (RSCP) is an important problem

for Train Operators, since the acquisition of rolling stock is expensive and a
long-term investment. Also the operational costs of rolling stock are usually
substantial. These costs include maintenance costs and power supply (electric-
ity or diesel). Both are positively correlated to the number of kilometers that
the rolling stock travels in the circulation. For these reasons, a Train Operator
has to decide carefully on the type and the number of rolling stock units per
scheduled train. Other important concerns in the planning of the rolling stock
 Ch. 3. Passenger Railway Optimization 155

circulation are the provided service to the passengers, which can give rise to
higher revenues, and the robustness of the circulation.
In order to obtain a better match between the available rolling stock and the
passengers’ seat demand, the compositions of the trains usually can be changed
at several stations by adding equipment to or removing it from the trains. These
coupling and uncoupling operations are usually penalized with switching costs.
The removed equipment can later be used for another train departing from the
same station. Several restrictions must be taken into account when changing
the composition of a train. These restrictions are related to the time required to
carry out the shunting operations to change the composition and the available
time at the station, which is the waiting time between two consecutive trips of
the train. For this reason, the order of the equipment in the train composition
matters, since switching equipment situated in the body of the train requires
more time than switching equipment situated at the tail or the head.
Various versions of RSCP arise depending on the equipment that is used
and on the nature of the railway network. Concerning the equipment, we dis-
tinguish two cases:
(i) locomotives and train carriages, and
(ii) aggregated modules, subsequently called train units.
The latter are composed of a number carriages in a fixed composition, and
can move in both directions, without the need of an extra locomotive. An ex-
ample of the latter is the well known French Train à Grande Vitesse (TGV).
A TGV train unit consists of six to ten passenger carriages and two power
units, each including a driver’s cabin, situated at both ends of the train unit,
see Ben-Khedher et al. (1998). Train units can move individually in both direc-
tions. A scheduled train can then be composed of several coupled train units.
In the first case, for each trip scheduled during the time horizon, one must
determine, for the associated train, the locomotive types and their number,
and the carriage types and their number. These numbers are not independent
of each other, since the number of carriages determines the type and the num-
ber of locomotives in order to provide sufficient pulling capacity. Turning a
locomotive-hauled train at the endpoint of a line is quite complex, since the lo-
comotive(s) must be uncoupled and driven to the other side of the train, where
it must be coupled again. It also occurs that a reserve locomotive is available
for the return trip, or that the entire train is turned around, if the so-called
Y-shaped tracks are available, see Lingaya et al. (2002). In the second case
(train units), one has to determine only the type and the number of train units
per type that are deployed. The order of the train units in a train may also be
important. Since a train unit can move in both directions, quick turn-around
times at the end points are possible.
Concerning the nature of the network, we distinguish two cases as well:
(i) a sparse network with long distances, and hence, long travel times and
relatively low frequencies of trains, and
156 A. Caprara et al.

(ii) a dense network with relatively short-distances and high frequencies of

trains.
This difference is important, in particular in connection with the preventive
maintenance of the rolling stock.
In a sparse network the rolling stock circulation usually describes in detail
the circulation of the individual rolling stock units over a longer period of time,
thereby taking into account the fact that each individual rolling stock unit
should reach a maintenance center sufficiently often. In contrast, in a dense
network the rolling stock circulation is usually anonymous. Moreover, there is
a certain capacity reservation for preventive maintenance of the rolling stock.
Routing the units that need preventive maintenance to a maintenance center
is handled in the operations by exchanging duties of rolling stock units. This is
possible because, in a dense network, there are usually sufficient possibilities
for exchanging units, i.e., time intervals where two identical rolling stock units
are located at the same station.
A second difference between dense and sparse networks is that in a dense
network with a lot of commuter trains, a seat reservation system usually does
not exist and only the expected numbers of passengers are known. On the other
hand, in a sparse network, a seat reservation system often exists, such that the
Train Operator has detailed knowledge of the number of passengers and can
change its circulation based on the actual reservations. Also revenue manage-
ment becomes possible in such a network.

5.1 Dense network

Literature review

For the dense network case, Schrijver (1993) describes a model to determine
the minimum number of train units that must be deployed on a single line
in order to avoid seat shortages. A line is defined by two endpoints between
which several trains run up and down according to the timetable. In the case
considered, the railway company runs an hourly service and one type of train
units is deployed there. There exist, however, two subtypes of train units that
differ in length and in capacity of first and second class seats. A train can consist
of several units of these subtypes. For every trip, the required numbers of first
and second class seats are known. Train units can be coupled to or uncoupled
from a train at several stations along the line. Obviously, a train unit can only
be coupled to a train at a certain station if the unit is available there at the right
moment. The model is basically an Integer Multicommodity Flow model with
several additional constraints.
The model proposed by Schrijver (1993) assumes that a train composition
can change to any other composition at a station between two subsequent trips.
In practice, however, several (un)coupling constraints must be taken into ac-
count. Therefore, not only the number of units deployed on a trip is important,
but also their order in the train composition. In this particular case, only one
 Ch. 3. Passenger Railway Optimization 157

operation is allowed, i.e., coupling or uncoupling, but not both. In addition, the
position where train units are coupled and uncoupled is fixed, i.e., in front or at
the rear of the train, which depends on the station. Moreover, Schrijver (1993)
assumes that the allocated rolling stock capacities per train should be such that
all passengers should obtain a seat. However, in the absence of a reservation
system, the passengers’ demand has a stochastic nature. Therefore, a seat for
all passengers cannot be guaranteed. A final issue not dealt with by Schrijver
(1993) is that, if a train arrives at its endpoint, it leaves the station usually as
soon as possible. That is, the train carries out the first trip of the same line
leaving the station, possibly after coupling or uncoupling some units. Given
that the subsequent trip for a train at an endpoint is known, several sequences
of trips can be distinguished. We refer to such a sequence of trips as a train.
In order to cope with the above issues, Peeters and Kroon (2003) propose
a model that minimizes a weighted function of the number of seat shortages
and cost factors, approximated by the number of carriage kilometers. They also
explicitly take into account the number of changes in the compositions of the
trains, since these may give an indication of the robustness of the rolling stock
circulation. Their approach to deal with the restricted transition possibilities
from one train composition to another is based on the concept of a transition
graph. This concept is further explained later.
Fioole et al. (2006) deal with a more complex version of the RSCP stud-
ied by Peeters and Kroon (2003). Their problem also includes the underway
combining and splitting of trains. Since the concept of transition graph is hard
to apply in such cases, especially in the case that two branches of a split line
do not have the same length, they use a MIP model that can be seen as an
extended version of the model described by Schrijver (1993). Due to several
methods to improve the quality of the model’s LP relaxation, relatively large
and complex instances of the problem can be solved to near optimality in an
acceptable amount of time by CPLEX.
Another work on the dense network case is the one by Brucker et al. (2003),
focusing on rerouting locomotive-hauled carriages. Their objective is to match
supply and demand for carriages in a region in Germany, which can be consid-
ered as a dense network. The routing of the locomotives is performed at a later
stage in the planning process and is not discussed in this chapter. For every trip,
the departure and arrival times and stations are given as well as the regular
composition of the train, i.e., the locomotive and the type and number of car-
riages. The rolling stock flow, imposed by the timetabled trips, is probably not
feasible given the limited availability of rolling stock, i.e., the required rolling
stock cannot always be available at the departure station in time, given the re-
quirements for other trips. To obtain a match between the requested number
of carriages and the available number of carriages, there exist two options:
(i) extending existing trains by coupling empty carriages to the train, and
(ii) introducing empty repositioning trips between two stations.
158 A. Caprara et al.

Obviously, the second option is much more expensive. The resulting model
is a huge Integer Multicommodity Flow model that is solved heuristically using
simulated annealing.

Model formulation

We present the models described by Schrijver (1993) and by Peeters and

Kroon (2003). As was mentioned earlier, the first model is basically an In-
teger Multicommodity Flow model with several additional constraints. The
underlying network is a directed time–space graph. The stations and times,
corresponding to the departures and the arrivals at the stations, characterize
the vertices. Two types of arcs can be distinguished, namely trip arcs, corre-
sponding to a trip between two stations, and inventory arcs, representing the
numbers of train units staying in a station between two events (i.e., a depar-
ture or arrival) at that station. These inventory arcs include an arc from the
last event to the first event at every station, thereby representing the cyclicity
of the circulation.
In order to further describe the model of Schrijver (1993), let M be the set of
rolling stock subtypes and let G = (V  A) be the time–space graph whose set
of arcs is given by the union of the trip arcs, denoted by AT , and the inventory
arcs, denoted by AI . For every trip arc a ∈ AT , the number of first and second
class passengers is denoted by p1a and p2a , respectively, and the maximum train
length as a . With every subtype m ∈ M, we associate the parameters qm 1 , q2 ,
m
cm , and wm , denoting, respectively, the capacities of first and second class seats,
the cost and the length of a train unit of subtype m ∈ M. With every arc a ∈ A
and every subtype m ∈ M, we associate an integer variable xm a , representing
the number of units of subtype m deployed on the trip if a ∈ AT , and staying
in the station if a ∈ AI .
The objective pursued in the model is to minimize the costs of the train
units deployed on the trains. To this end, the number of train units on the night
inventory arcs is minimized. Letting AN be the set of night inventory arcs and
δ+ (v) and δ− (v) be the sets of arcs entering and leaving vertex v, respectively,
the model reads as follows:
 
min cm xm
a (27)
m∈M a∈AN
subject to
 
xm
a = xm
a v ∈ V  m ∈ M (28)
a∈δ− (v) a∈δ+ (v)

k m
qm xa  pka  a ∈ AT  k = 1 2 (29)
m∈M

wm xm
a  a a ∈ AT  (30)
m∈M
 Ch. 3. Passenger Railway Optimization 159

xm
a  0 integer a ∈ A m ∈ M (31)
Constraints (28) are flow conservation constraints in each vertex of G. Con-
straints (29) impose that the capacity on each arc must not be less than the
expected number of first and second class passengers. Constraints (30) impose
that the sum of the lengths of the train units deployed on a trip does not exceed
the maximum train length.
The model described by Peeters and Kroon (2003) can be seen as an ex-
tension of Schrijver’s model. In order to deal with the fact that in each station
the transition possibilities from one train composition to another one are usu-
ally limited, they use the concept of a transition graph. Each train has its own
transition graph. As was mentioned earlier, a train is a sequence of trips to be
carried out by the same rolling stock units.
In a transition graph of a train, the nodes represent the feasible train com-
positions on the trips, and the arcs represent the feasible transitions between
compositions. More specifically, for every trip, the feasible compositions are
enumerated, given the available subtypes, the limits on the train length, and
the maximum allowable first and second class seat shortages. Next, all feasi-
ble transitions between the compositions on subsequent trips are determined,
thereby taking care of coupling and uncoupling restrictions at the stations.
By selecting a path through the transition graph for every train, a feasible
composition for every trip is determined, where also the transitions between
two compositions are feasible. Since a train unit can only be coupled onto a
train if the train unit is available at the right time and station, the interaction
between different trains that run simultaneously is modeled by keeping track
of the inventory positions of the subtypes at all relevant events at the stations
during the considered time period. The inventory position at an event e of a
station equals the initial allocation of the rolling stock to the station augmented
with the uncoupled train units at that station and decreased with the coupled
train units at that station before event e. The model ensures that, at any time
and in all stations, the inventory positions of all subtypes are nonnegative. This
gives rise to a Dantzig–Wolfe reformulation, whose LP relaxation is solved in
Peeters and Kroon (2003) through a column generation algorithm.
Formally, let T be the set of trains, and let P t be the (exponentially large)
set of all paths through the transition graph of train t ∈ T . Associate a binary
variable ytP with every path P ∈ P t , which equals 1 if the path is selected in
the solution, and 0 otherwise. The cost associated with path P ∈ P t , denoted
by ctP , is a weighted function of the first and second class seat shortages and
the number of carriage kilometers. It equals the sum of the cost of the arcs of
the path.
Letting S be the set of stations and Es be the set of events at station s,
the model also contains integer inventory variables xm es , defined for all s ∈ S,
e ∈ Es , and m ∈ M, representing the number of train units of subtype m
present in station s immediately after event e. The set of events consists of the
initial state 0, the departures and arrivals at the station, and a final state f . Fur-
thermore, q(e) ∈ Es denotes the event at station s ∈ S immediately preceding
160 A. Caprara et al.

event e ∈ Es . Finally, the parameters am m

tPes and btPes represent, respectively,
the number of uncoupled and coupled units of subtype m at station s at event
e for path P ∈ P i of train t ∈ T . Recall that each path specifies the subse-
quent compositions of a train at the subsequent trips, so that these numbers of
uncoupled and coupled train units can be derived a priori.
Let nm be the total number of train units of subtype m ∈ M available for the
involved trains, then we can state the Dantzig–Wolfe reformulation as follows:
 
min ctP ytP (32)
t∈T P∈P t
subject to

ytP = 1 t ∈ T (33)
P∈P t

xm m
0s = n  m ∈ M (34)
s∈S
   
xm
q(e)s + am
tPes ytP − bm m
tPes ytP = xes 
t∈T P∈P t t∈T P∈P t
s ∈ S m ∈ M e ∈ Es \ {0} (35)
xm
0s = xm
f s s ∈ S m ∈ M (36)
ytP ∈ {0 1} t ∈ T P ∈ P  t
(37)
xm
es  0 integer s ∈ S e ∈ Es  m ∈ M (38)

Constraints (33) impose that for every train a path must be selected. Con-
straints (34) represent the limited availability of rolling stock, i.e., the sum of
the initial inventories at all stations must be equal to the total available rolling
stock. Constraints (35) are the inventory constraints, i.e., the inventory of sub-
type m ∈ M after event e must equal the inventory after the previous event
q(e) increased with the number of uncoupled train units and decreased with
the number of coupled units between q(e) and e. Finally, constraints (36) imply
cyclicity.
In general, this formulation has a huge number of path variables, and it
would be impossible to consider them all explicitly, even for small RSCP in-
stances. Therefore, the LP relaxation is solved using column generation, lead-
ing to a branch-and-price approach to obtain the optimal integer solution.

Experimental results

After deriving local parts of the convex hull of integral solutions of the
|M|-dimensional polytope defined by constraints (29) and (30) for each trip
arc a ∈ AT in a preprocessing step, real-world RSCP instances of the model
of Schrijver (1993) are solved within a few seconds by CPLEX.
 Ch. 3. Passenger Railway Optimization 161

Table 4.
Results on real-world instances with different availabilities of rolling stock

Line # trains # types # inst. # nodes Time

2100/15 182 2 62 111 6

2100/15 182 3 82 99 91
3000/12 115 2 45 109 1
3000/12 115 3 58 175 10

The algorithm of Peeters and Kroon (2003) for dense networks was tested
on two lines of NS Reizigers, the main Dutch Train Operator. These two lines
are indicated as 2100 and 3000. The experiments are carried out on a 1.6 GHz
IBM NetVista 6343-25G Pentium 4 PC, using the extended LINDO/PC 6.1
optimization library for solving the LP relaxations. The problem characteristics
and the results obtained are summarized in Table 4. For each of the two lines,
several instances with different availabilities of rolling stock were solved. In
the table, we give the number of trains in the series (# trains), the number of
rolling stock subtypes considered (# types), the resulting number of instances
(# inst), and for these instances, the average number of nodes in the branch-
and-price tree (# nodes) and the average computing time in seconds (time).

5.2 Sparse network

Literature review

Since the planning horizon in a sparse network is usually much longer than
in a dense network, and the same holds for the travel time to the maintenance
centers, maintenance requirements must be taken explicitly into account when
determining the circulation of the rolling stock.
Cordeau et al. (2001) present a model for the simultaneous locomotive and
carriage assignment problem. As a result, the type and the number of both
equipment types must be determined for each scheduled train, taking care
that the assigned locomotives provide sufficient pulling capacity for the as-
signed carriages. In general, several combinations of types of locomotives and
carriages are possible, where carriages typically differ in seat capacity and in
class (first or second) and locomotives differ in pulling capacity. The operating
speed of the different equipment types may vary, and the operating speed of a
train equals the speed of its slowest component, thereby requiring some flexi-
bility in the timetable. The solution found is cyclic and can be repeated period
after period for a whole season.
Lingaya et al. (2002) present a model to deal with seasonal cycles, adapt-
ing the model of Cordeau et al. (2001) to short-term demand revisions. Based
on true data of sold and requested tickets, the model tries to find alternative
162 A. Caprara et al.

cycles, seeking to maximize the expected profit, subject to several operational

constraints. The locomotive cycles cannot be changed anymore, because they
are the basis for the crew schedules. The most important constraints are the
maintenance requirements and the minimum switching times, i.e., the mini-
mum time needed to uncouple a carriage from a train or to couple it onto a
train, between two consecutive trips that the train must make. This switching
time depends on the position of the carriages in the consist. An in-body-switch
is more time consuming than a switch at the tail of the train. For every carriage
and day in the planning horizon on which the carriage can begin a cycle, a net-
work is generated, representing all potential cycles for the carriage. These are
based on the fixed locomotive cycles. These networks also reflect the possible
initial and final conditions for the carriage, imposed by its position at the begin-
ning of the planning horizon or by the fact that before a given day the carriage
must be at the maintenance center. The model takes into account the posi-
tions of the units of equipment in the train, such that a tail switch can only be
performed for the carriages positioned at the tail of the consist. A column gen-
eration procedure is proposed to solve the LP relaxation and next a heuristic
branch-and-bound scheme is applied to find an integer solution. The required
computing time is generally small.
Finally, Ben-Khedher et al. (1998) study the problem of allocating train
units to the French TGVs. Their rolling stock allocation system is based on
a capacity adjustment model that is linked to the seat reservation system and
seeks to maximize the expected profit. The TGV fleet consists of several types
of train units with different first and second class seat capacities. In principle,
one train unit is deployed per scheduled trip. However, part of the fleet can be
used to reinforce a scheduled train, because of a high number of reservations.
The model is basically a huge Integer Multicommodity Flow model with side
constraints that is solved with a commercial ILP solver.

Model formulation

In this section we present a simplified version of the formulation of Cordeau

et al. (2001), discussing extensions in the end. Let T be the set of all trips dur-
ing the time horizon, S be the set of stations, M be the set of equipment types,
and D = {1     n} be the set of days of the time period. Then, for each sched-
uled trip t ∈ T , the consist is assumed to be given and the minimum number
of required locomotives and carriages for the various types of equipment of
the consist are denoted by nm t for m ∈ M. The model relies on the concept of
equipment cycles, i.e., a sequence of trips and waiting times in some stations
between two sequences, where each cycle starts and ends in a unique mainte-
nance center. After a number of days, the equipment must spend some time
at the maintenance center for a preventive check and possibly for some re-
pairs. We further assume that maintenance takes place during the night and
that there is enough time to do so.
 Ch. 3. Passenger Railway Optimization 163

We now discuss the conditions for a feasible equipment cycle. Two trips
t1 and t2 requiring the same consist can be covered by the same equipment
cycle if the arrival station of t1 is the departure station of t2 and there is suffi-
cient connection time at the station. The required connection time, however,
depends on whether or not carriages are coupled to or uncoupled from the
train. It can occur that the same train can cover two trips only if the composi-
tion of the train is not changed at the station. To this end, the concept of train
sequence is introduced, which is an ordered set of trips that can be covered by
the same train, if the composition is not changed at an intermediate station. In
addition, the connection time also depends on whether or not the train must
be turned around in order to make the connection. This is the case if trips t1
and t2 have opposite directions.
A time–space network Gm represents all possible cycles that an equipment
type m ∈ M can make. As explained earlier, a cycle essentially consists of con-
secutive train sequences, on which a given equipment type can be deployed,
and the waiting time that is spent at the station between two consecutive se-
quences. The networks are determined so as to respect the connection times.
Each network has several source and sink nodes associated with, respectively,
the start of the day and the end of the day at the maintenance center. For
the other stations, the networks contain Start-Of-Day (SOD) and End-Of-Day
(EOD) nodes for every day of the planning period. The arcs between the EOD
and SOD nodes of two consecutive days allow the equipment to stay during the
night at a station. The EOD nodes of the last day are connected with the SOD
nodes of the first day at all stations to represent the fact that a cyclic solution
that can be replicated is sought.
For each equipment type m ∈ M, let am be the number of available units
and P m be the set of possible paths, corresponding to paths from a source to a
sink node in Gm . For each path P ∈ P m , cP denotes the cost of the path (de-
fined by the length of the path multiplied by the cost per kilometer, associated
with power supply and maintenance), and parameter otP equals 1 if path P
covers trip t, and 0 otherwise. Next, two parameters bdP and edP are defined to
represent, respectively, the start day and end day of path P. That is, bdP (edP )
equals 1 if path P begins (ends) on day d ∈ D. The parameter vP equals 1 if
path P crosses the end of the time horizon, i.e., the equipment stays at a sta-
tion, different from the maintenance center, during the night between day p
and day 1.
To formulate the problem, two sets of decision variables are defined:
(i) the flow yP on path P ∈ P m (m ∈ M) and
(ii) the number of units xmd of equipment type m ∈ M staying at the main-
tenance center during day d ∈ D.
The problem can then be stated as follows:
 
min cP yP (39)
m∈M P∈P m
164 A. Caprara et al.

subject to

otP yP  nm
t  m ∈ M t ∈ T (40)
P∈P m
 
xm
d + edP yP − bd+1P yP = xm
d+1 
P∈P m P∈P m
m ∈ M d ∈ D \ {n} (41)
 
xm
n + enP yP − b1P yP = xm
1  m ∈ M (42)
P∈P m P∈P m
 
vP yP + b1P yP + xm m
1 a  m ∈ M (43)
P∈P m P∈P m
yP  0 integer m ∈ M P ∈ P m  (44)
xm
d  0 integer m ∈ M d ∈ D (45)
Constraints (40) take care that sufficient equipment is deployed on every
trip. Constraints (41) and (42) are flow conservation constraints, i.e., they im-
pose that the number of units of equipment type m staying at the maintenance
center during day d + 1 equals the number of units during day d increased by
the number of units arriving at the maintenance center during day d and de-
creased by the number of units leaving on day d + 1. Constraints (43) ensure
that the total number of units of an equipment type used in the circulation
does not exceed the available number am of equipment type m. To this end,
Cordeau et al. (2001) compute the flow crossing the time horizon at each sta-
tion, except at the maintenance center, and add to this flow the number of units
of an equipment type that must be available at the maintenance center at the
beginning of day 1, i.e., the number of units that stay at the center during day
1 and the number of units of equipment starting a cycle on day 1.
The LP relaxation of this model is solved by column generation. The column
generation problem consists of finding the shortest paths through the equip-
ment networks Gm for all m ∈ M. However, given that the network contains
several source and sink nodes, and given that, depending on which source node
is chosen, some sink nodes become infeasible, Cordeau et al. (2001) solve a
shortest path problem for every source node, making sure that only the feasible
sink nodes can be reached. Next, an integer solution is obtained by heuristically
applying a truncated branch-and-bound procedure.
Cordeau et al. (2001) present several extensions of this model, namely they
allow for:
(i) constraints on the locomotive pulling capacity for a sequence,
(ii) unavailability of equipment type on a given day,
(iii) storage capacity of the stations and the maintenance center,
(iv) substitution between equipment types, which implies, for example, that
a second class carriage can be replaced by a first class carriage,
(v) daytime maintenance, and
 Ch. 3. Passenger Railway Optimization 165

(vi) various consist types for a trip.

In addition, they propose a two-phase method, where in a first phase the
locomotive cycles are determined, imposing the integrality on the locomotive
flows only. In a second phase the assignment of carriages is done, given the
fixed locomotive cycles of the first phase and taking into account not only the
circulation costs but also the switching costs, incurred if carriages are coupled
to or uncoupled from the train.
The algorithm of Cordeau et al. (2001) was tested on six real-life instances of
VIA Rail, a Canadian Train Operator. A weekly circulation is sought for more
than 325 trips and for 130 units of equipment. There are 2 locomotive types
and 4 carriage types, that can be combined into three different consist types
with a different operating speed. The computing time for the first phase of the
algorithm, corresponding to the determination of the locomotive cycles, lies
between approximately 5000 and 50,000 seconds, depending on the instances.
The second phase typically requires only a few seconds. The obtained solutions
considerably improve the manual solutions of VIA Rail.

5.3 Maintenance routing

As was indicated earlier, each rolling stock unit has to visit a maintenance
center regularly in order to be checked and repaired, if necessary. In this sec-
tion, we describe some details of the problem of routing rolling stock units
towards a maintenance center. In the same way as in the previous sections, one
may distinguish here between “sparse” and “dense” systems.
In a “sparse” railway system the maintenance checks of the rolling stock are
incorporated into the basic rolling stock circulation, since there is the risk that
it will not be possible to get a rolling stock unit in time at the maintenance
center in the moment in which this is necessary. In such a system, each rolling
stock unit follows a planned cycle that starts and ends in the maintenance cen-
ter. On the other hand, incorporating the maintenance checks already into
the basic rolling stock plan leads to the risk that, during the operations, the
maintenance plan has to be updated regularly, since, due to disruptions in the
operations, the realized rolling stock circulation differs from the planned one.
Another disadvantage of incorporating the maintenance checks already into
the basic rolling stock plan is the fact that the length of the maintenance cycles
does not fit with the cycle length of the basic rolling stock circulation.
Therefore, especially in “dense” railway systems, the rolling stock units may
be routed to the maintenance center on a more or less ad hoc basis. That is, on
a day-by-day basis, one determines which rolling stock units need to be taken
away from the operations in order to undergo a maintenance check, and how
these are routed towards the maintenance center. The latter is done preferably
with a minimum number of additional train movements, since these are usually
quite expensive. Rolling stock units that need to be routed towards a mainte-
nance check are called urgent units. Here we focus on this routing problem of
rolling stock units towards the maintenance center.
166 A. Caprara et al.

Usually, each rolling stock unit has been assigned to a series of duties. Here
each duty is a set of trips that are to be carried out by the rolling stock unit.
From one day to another, there may also be planned links between consecutive
duties. Some series of consecutive duties pass along the maintenance center
during the next days, and other series do not. Now the problem is to find ap-
propriate swaps in the series of duties such that the urgent units get to serve on
a series of duties passing through the maintenance center at the right time.
In order to solve the problem of efficiently routing the urgent train units to
the maintenance checks, one may use an Integer Multicommodity Flow model.
The underlying network is a time–space network, where the nodes correspond
to the trips to be carried out. Pairs of scheduled trips in the same duty between
which a swap to another duty is not possible are represented by a single node.
The planned connections from one trip to another are represented by arcs
in the network. All these arcs have cost zero and capacity one. Furthermore,
there are also arcs that represent the potential swaps of duties in the network.
If one unit of flow passes such an arc, then this means that the corresponding
rolling stock unit is swapped from one duty to another one, in order to bring
it onto the right track towards the maintenance center. The costs of these arcs
represent the complexity of the involved swap. It should be noted that it is usu-
ally hard to make a detailed estimate of the cost of a swap. However, the cost
structure can be designed in such a way that a distinction is made between easy
swaps, moderately difficult swaps, hard swaps, and nearly impossible swaps. For
example, swapping the duties of two single train units that are standing at the
same time at the same shunting area for more than one hour can be considered
as easy. On the contrary, swapping two train units that are both the middle
train unit of two trains consisting of three train units is nearly impossible, in
particular if only a small amount of time overlap is available.
Each urgent unit is represented by its own commodity in the network. One
unit of such a commodity is to be routed from the start of the duty that is cur-
rently served by the urgent unit to an appropriate maintenance check. Each
commodity corresponding to an urgent unit has its own set of sinks. Further-
more, there is an additional commodity that represents all nonurgent units.
The amount of flow of this commodity to be routed equals the number of
nonurgent units. The nonurgent units have to be routed through the network
in order to check that an overall feasible solution exists.
Now the problem is to find an Integer Multicommodity Flow with minimum
cost in the constructed network. A solution to this problem can be interpreted
as a set of node disjoint paths for the urgent train units that follow as much as
possible the planned duties, and in which the complexity of the required swaps
is as low as possible.
Maróti and Kroon (2007) describe a model for solving the maintenance
routing problem that requires many details of the potential swaps as input.
These details are required in order to be able to evaluate the complexity of
the involved shunting movements. Since the required data may be hard to ob-
tain, Maróti and Kroon (2005) also describe a simplified model, which requires
 Ch. 3. Passenger Railway Optimization 167

less detailed input. Since, usually, the number of urgent units that need to be
routed simultaneously is small (1 up to 5), both models can be solved quickly
by CPLEX.

6 Train unit shunting

Within the rush hours, the rolling stock of a passenger Train Operator is
typically operating the timetable or it is in maintenance. However, outside the
rush hours, and in particular during the night, most of the rolling stock is to
be parked on a shunting area near one of the stations in the railway network.
During the night, one of the objectives is to park the rolling stock on the shunt-
ing area in such a way that the railway operations can start up as smoothly as
possible on the next morning.
The process of parking rolling stock on a shunting area, together with sev-
eral related processes such as routing rolling stock between the station area
and the shunting area, short term maintenance, and inside and outside clean-
ing, is called shunting. A major complicating issue is the fact that rolling stock
is strongly restricted in its movements by the railway infrastructure. Therefore,
units of rolling stock are easily blocking each other in their movements. In ad-
dition, time is also a restrictive resource for shunting. For example, for safety
reasons, it is mandatory to respect a certain minimum headway time between
any two train movements on the same track or switch.
Shunting processes are highly dependent on changes in the timetable and
in the rolling stock circulation of a Train Operator: as soon as the timetable
or the rolling stock circulation changes, the shunting plans have to be updated
as well. Therefore, the planning of the shunting processes is the closing stone
of the logistic plans of a Train Operator: tools that support planners in quickly
generating shunting plans are highly relevant in practice.
In order to define the problem, we call an arriving train unit that has to be
parked on a shunt yard an arriving shunt unit, and similarly, a train unit that has
to be supplied from the shunt yard a departing shunt unit. Arriving shunt units
are uncoupled from through trains or come from complete ending trains, and
departing shunt units are units that are coupled onto through trains or form
complete starting trains. Now the shunting problem can be defined as follows.
Given (i) a railway station and a nearby shunting area, (ii) a timetable with, for
each train, the arrival and/or departure time and platform, and its composition,
and (iii) the routing costs from each platform track to each shunt track and vice
versa, as well as several other cost estimates, the Train Unit Shunting Problem
(TUSP) consists of matching the arriving and departing shunt units, as well as
parking these shunt units on the shunt tracks, such that the total shunting costs
are minimal.
The shunting costs consist of the routing costs, the train unit dependent
penalties for certain shunt tracks, and the penalties for not parking shunt units
that should be parked. Note that the routing costs usually vary over time,
168 A. Caprara et al.

since the (time-dependent) claims on the station infrastructure by the through

trains have to be taken into account. Estimating these routing costs is a diffi-
cult problem in itself, which lies outside the scope of this chapter. For more
details we refer to van’t Woudt (2001). The shunting costs are mainly driven
by the shunting movements. Therefore, minimizing the shunting costs can be
attained by minimizing the number of shunting movements. Especially in the
early morning, with the start-up of the railway operations, the number of shunt-
ing movements should be minimized. Further characteristics of the shunting
problem are the following:
• Arrivals and departures of train units may be mixed in time. This im-
plies that, within the planning horizon, the first departure may take
place before the last arrival has taken place.
• Shunt units may have different types and subtypes. Train units of the
same type, but possibly of different subtypes, may be combined with
each other in one train. The type of a unit may restrict the set of shunt
tracks where the unit can be parked.
• Shunt tracks may have different types and lengths. The type of a track
determines how a unit can approach the track. Some tracks can be ap-
proached from one side only. These tracks will be called Last In First
Out (LIFO) tracks. Other tracks can be approached from both sides.
These tracks will be called free tracks.
In the matching of arriving and departing train units it is required that the
matched units have the same subtype. Furthermore, if several units of differ-
ent subtypes of one arriving train are matched with one departing train, then
the subtypes of the units in both trains have to be in the same order. Finally,
a crossing occurs whenever a train unit i obstructs a train unit j during the
departure or arrival of train unit j. Such crossings are not allowed.

Literature review

A survey on shunting processes of cargo trains is provided by Cordeau et al.

(1998). While literature on TUSP for passenger trains is scarce, there are a few
references dealing with the analogous problem for mass-transit companies.
Freling et al. (2005) present a solution approach consisting of the following
steps:
(i) matching of arriving and departing shunt units,
(ii) parking the shunt units on the tracks, and
(iii) routing the shunt units between the station area and the shunting area.
For several instances from the Dutch passenger Train Operator NS Reizi-
gers, the matching step is solved quickly by CPLEX, the parking step is mod-
eled as a Set Covering Problem and solved by dynamic column generation
techniques, and the routing step is solved through a dynamic programming
approach.
 Ch. 3. Passenger Railway Optimization 169

Tomii et al. (1999) and Tomii and Zhou (2000) propose a genetic algorithm
for solving a version of the problem that takes into account several practical
issues, including routing, maintenance, and duties for shunt personnel. Their
problem is relatively simple, since in their context at most one train unit can
be parked on a shunt track at the same time. Indeed, the latter restriction
eliminates the problem of the crossings.
Di Stefano and Koci (2004) look at the problem of parking trains on the
available shunting tracks in order to avoid shunting movements in the next
morning. They assume that each track is long enough to host the trains as-
signed to it. Their main objective is to minimize the number of shunting tracks
necessary to park all the trains without additional shunting movements. They
consider several variants of their shunting problem, distinguished from each
other by the ends of the shunting tracks that can be used for entering or leaving
these tracks. For example, in the SISO-variant (Single Input Single Output),
each train enters the shunting area along one end of the tracks and each train
leaves the shunting area from one end of the tracks. For several variants of
their problem they provide computational complexity results.
The subject of the paper by He et al. (2000) is the separation of train units
from arriving trains, sorting the trains according to their destination, and finally
combining them to form new departing trains. This resembles the problem of
matching arriving and departing train units as described by Freling et al. (2005).
Dahlhaus et al. (2000) discuss the problem of rearranging carriages of pas-
senger trains in a station in order to group them by destination. Their goal is to
use a minimum number of tracks for this rearrangement. They show that this
problem is NP-hard.
Van den Broek (2002) describes a model that can be used to test whether the
capacity of the railway infrastructure around a station is sufficient for handling
all the shunting movements that are enforced by the rolling stock circulation.
This paper does not focus on the storage of train units, but on the scheduling
of the involved shunting movements. The model assumes that the routes for
the shunting movements are fixed beforehand and verifies that each shunting
movement can be scheduled at a time instant such that each infra-element is
occupied by at most one movement at the same time.
Blasum et al. (2000) focus on dispatching trams in a depot. They prove that
this problem is NP-hard and that a restricted version of the problem can be
solved in polynomial time by a dynamic programming approach. Their analysis
in the context of dispatching trams is extended by Winter and Zimmermann
(2000). Winter (1999) extends this approach to the case of length restrictions
and mixed arrivals and departures, and presents an application at a bus de-
pot. Furthermore, several variants of the studied problems are shown to be
NP-hard.
Gallo and di Miele (2001) discuss the problem of dispatching and parking
buses in a bus depot. Here the dispatching of the buses takes place in First
In First Out (FIFO) order. They model this problem as a Noncrossing As-
170 A. Caprara et al.

signment Problem. They also include an extension of their model taking into
account mixed arrivals and departures of buses.
Another application of dispatching and parking buses in a depot is described
by Hamdouni et al. (2006). Here robust solutions are emphasized by having as
little different bus types as possible in each track, and by grouping the buses of
the same type as much as possible. This makes the plans less sensitive to the
actual arrival times of the buses in the operations.

Model formulation

In this section we describe a model (first proposed by Schrijver (2003)) that

can be used to find an appropriate allocation of train units to shunting tracks.
The model focuses on storing the train units without crossings at the shunt
tracks. In order to keep the presentation as clear as possible, we make the
following simplifying assumptions.
• All tracks can be entered from one side only. This implies that all tracks
are used in a Last In First Out (LIFO) fashion.
• Each arriving and each departing train consists of a single train unit.
This simplifies the problem significantly. Indeed, if this is not assumed,
then the model has to guarantee that train units from the same train
are kept together as much as possible.
• No additional shunting movements related to cleaning or maintenance
of train units have to be carried out. That is, briefly after the arrival of
a train unit at the station, it is parked at a shunting track, and briefly
before its departure from the station it leaves this shunting track again.
Relevant elements for the objective function are the routing costs between
the platform zone and the shunting zone of the station, and the number of
tracks on which more than one type of train units is stored. The latter is rele-
vant for the robustness of the solution: if only one type of train units is stored
on a track, then crossings will not occur at that track.
The sets A and D denote the sets of arriving and departing train units, re-
spectively. The arriving and departing train units have been ordered according
to their arrival and departure time. Each train unit has a certain type τa or τd ,
depending on whether the train unit is arriving or departing. An arriving train
unit can be matched with a departing train unit only if they have the same type.
According to these types and to the arrival and departure times, for each arriv-
ing train unit a ∈ A there is a set Da ⊆ D of departing train units that can be
matched with a. Similarly, for each departing train unit d ∈ D we let Ad ⊆ A
be the set of arriving train units that can be matched with d. The length of each
train unit is denoted by la (a ∈ A) or ld (d ∈ D). Let T denote the set of the
shunt tracks. The length of shunt track t ∈ T is denoted by t . The cost of rout-
ing train unit a ∈ A or d ∈ D to or from track t ∈ T is denoted by cat or cdt ,
respectively.
 Ch. 3. Passenger Railway Optimization 171

The model contains binary variables zat , equal to 1 if and only if arriving
train unit a ∈ A is parked at shunt track t ∈ T , and zdt , equal to 1 if and
only if departing train unit d ∈ D is parked at shunt track t ∈ T . Moreover,
binary variables xadt are equal to 1 if and only if arriving train unit a ∈ A is
matched with departing train unit d ∈ D at shunt track t ∈ T . The model reads
as follows:
   
min cat zat + cdt zdt (46)
t∈T a∈A d∈D
subject to

zat = 1 a ∈ A (47)
t∈T

zdt = 1 d ∈ D (48)
t∈T

zat = xadt  t ∈ T a ∈ A (49)
d∈Da

zdt = xadt  t ∈ T d ∈ D (50)
a∈Ad
 
xad t + xa dt  1
d  ∈Da :d  >d a ∈A d :a <a
t ∈ T a ∈ A d ∈ D a < d τa = τd  (51)
 
la za t − ld zd t  t  t ∈ T a ∈ A (52)
a ∈A:a a d  ∈D:d  <a
zat ∈ {0 1} a ∈ A t ∈ T (53)
zdt ∈ {0 1} d ∈ D t ∈ T (54)
xadt ∈ {0 1} a ∈ A d ∈ D t ∈ T (55)
The objective function (46) expresses the fact that the routing costs for the
shunting movements are to be minimized. (Note that also the other mentioned
objective of minimizing the number of tracks with more than one train unit type
can be expressed easily in the decision variables.) Constraints (47) and (48)
specify that each arriving train unit and each departing train unit, respectively,
is stored at a certain track. Constraints (49) specify that, if an arriving train
unit a is stored at shunting track t, then it is matched there with an appropriate
departing train unit d. Constraints (50) specify the same for each departing
train unit d. Constraints (51) are the crossing constraints and guarantee that,
on a single track, the arriving and departing train units are matched in a LIFO
way. Indeed, constraints (51) guarantee that, if there are other train units than
train unit d at track t at the moment that train unit d wants to depart from
track t, then these train units have arrived earlier at track t than train unit d.
Thus, train unit d is the first train unit at track t and can, thus, depart without
172 A. Caprara et al.

a crossing. Note that constraints (51) are only relevant for τa = τd . Indeed,
τa = τd implies τd = τa = τd = τa , and if all these train units have the same
type, then they cannot create a crossing: identical train units can be exchanged
if necessary. Constraints (52) guarantee that the length of each shunt track is
not exceeded by the train units that are stored on the track. It is sufficient to
take these constraints into account only at the arrival times of the arriving train
units, as they are the only time instants at which the length of a track might be
exceeded. In constraints (52), the first term represents the total length of the
train units that have been stored at shunting track t before (and including)
the arrival of train unit a. The second term represents the total length of the
train units that have departed from track t before the arrival of train unit a.
The difference between these two is the total length of the train units that are
stored at track t just after the arrival of train unit a.
Model (46)–(55) can be extended to take into account also trains that consist
of several train units and tracks that can be approached from both sides. Com-
putational experiments based on station Enschede in the Netherlands show
that the resulting model can be solved in an acceptable amount of time by
CPLEX. For the more complex situation at station Zwolle, CPLEX usually
finds high quality solutions quickly, but closing the gap between the involved
lower and upper bounds may take a lot of time.
Schrijver (2003) also describes several methods for reducing the running
times. One of these methods is based on the concept of so-called virtual tracks.
This concept is based on the fact that the number of crossing constraints is
quite high, and that the crossing constraints are only required for shunting
tracks on which train units of different types are parked. For tracks containing
train units of a single type only, the crossing constraints can be replaced by
simple flow conservation constraints.

7 Crew planning

The Crew Planning Problem (CPP) is the problem to be faced by Train Op-
erators which is concerned with building the work schedules of crews needed
to cover a planned timetable. In CPP, we are given a planned timetable for
the train services (i.e., both the actual journeys with passengers or freight, and
the transfers of empty trains or equipment between different stations) to be
performed every day of a certain planning horizon. Each train service has first
been split into a sequence of trips, defined as segments of train journeys which
must be serviced by the same crew (i.e., driver or conductor) without rest. Each
trip is characterized by a departure time, a departure station, an arrival time,
an arrival station, and possibly by additional attributes. Each daily occurrence
of a trip has to be performed by one crew. In fact, each crew performs a roster,
defined as a sequence of trips whose operational cost and feasibility depend on
several rules laid down by union contracts and company regulations. The prob-
lem consists of finding a set of rosters covering every trip of the given planning
horizon, so as to satisfy all the operational constraints with minimum cost.
 Ch. 3. Passenger Railway Optimization 173

CPP represents a very complex and challenging problem, due to both the
size of the instances to be solved and the type and number of operational con-
straints. Typical figures for the main European Train Operator companies are a
few thousand trains per day and a workforce of several thousand drivers spread
among several crew depots. Usually, CPP is approached in two phases, accord-
ing to the following scheme:
1. Crew Scheduling: The short-term schedule of the crews is considered, and
a convenient set of duties (also called pairings) covering all the trips is
constructed. Each duty represents a sequence of trips to be covered by a
single crew within a given planning horizon overlapping at most one or
two consecutive days.
2. Crew Rostering: The duties selected in the Crew Scheduling phase are
sequenced to obtain the final rosters. Here, trips are no longer taken into
account explicitly, but determine the attributes of the duties which are
relevant for the roster feasibility and cost.
Decomposition is motivated by several reasons. First of all, each crew mem-
ber is located in a given crew depot, which represents the starting and ending
point of its work segments. A natural constraint imposes that each crew must
return to its home depot within one (or two, in case an external rest is assigned)
day, which leads to the concept of a duty as a short-term work segment start-
ing and ending at the home depot and overlapping very few consecutive days.
Secondly, constraints affecting the short-term work segments are different in
nature from those related to the overall crew rosters. For example, the min-
imum time interval between two consecutive trips in a duty is a few minutes
for changing trains, whereas the time interval between two consecutive duties
is several hours for home rest. It is worth noting that in Crew Scheduling ad-
ditional constraints, called depot constraints, typically impose bounds on the
number of duties with given characteristics for each depot. Moreover, Crew
Rostering considers each depot separately, since a roster cannot include duties
of different depots.
A main objective of CPP is the minimization of the global number of crews
needed to perform all the daily occurrences of the trips in the given plan-
ning horizon. In some applications, the Crew Rostering phase plays a minor
role, since the corresponding constraints are rather weak and the number of
crews is easily determined from the solution of the Crew Scheduling phase.
This typically happens, e.g., when the considered trains cover a relatively small
area, running mainly within the day, and the wide majority of the crews leave
from their depot in the morning/afternoon and return back to it in the after-
noon/evening. In this case, the duty performed by a crew in one day puts very
limited restrictions on the duties that it may perform on the next day, and Crew
Rostering is aimed at balancing the workload among the crews as evenly as
possible. As a result, the objective used in the Crew Scheduling phase mainly
calls for the minimization of the number of working days corresponding to the
duties.
174 A. Caprara et al.

In applications involving several trains covering a wide area and/or running

overnight, instead, considerable savings can be obtained through a clever se-
quencing of the duties obtained in the first phase. Therefore, the objective of
the Crew Scheduling phase has to take into account the characteristics of the
duties selected and their implication in the subsequent rostering phase. This
suggests the opportunity of integrating the two phases, as we will discuss later.

Literature review

Several papers on CPP appeared in the literature. For papers concerning

mass-transit and airline transportation, we refer the interested reader to the
surveys in Arabeyre et al. (1969), Wren (1981), Bodin et al. (1983), Rousseau
(1985), Daduna and Wren (1988), Desrochers and Rousseau (1992), Barnhart
et al. (1994), Desrosiers et al. (1995), Wise (1995), Wilson (1999), Daduna and
Voss (2001), and Ernst et al. (2001, 2004b), as well as to other chapters of this
book.
As far as railway applications are concerned, Caprara et al. (1997) present a
survey of the methods used in the literature. Most of these works focus on the
Crew Scheduling phase, that is mainly solved by generating (a suitable subset
of) all duties and then selecting them by solving a Set Covering Problem, pos-
sibly with additional constraints. The exact Set Covering algorithms proposed
in the literature can solve instances with up to a few hundred trips and a few
thousand potential duties, see Beasley (1987), Beasley and Jörnsten (1992),
and Balas and Carrera (1996). At present, the best methods to solve the prob-
lem to proven optimality appear to be the state-of-the-art general-purpose ILP
solvers, see Caprara et al. (2000). When larger instances are tackled, one has to
resort to heuristic algorithms. Classical greedy algorithms are very fast in prac-
tice, but typically do not provide high quality solutions, as reported in Balas
and Ho (1980) and Balas and Carrera (1996). Jacobs and Brusco (1995) and
Beasley and Chu (1996) propose a genetic and a simulated annealing algo-
rithm, respectively, whereas Lorena and Lopes (1994) use an approach based
on surrogate relaxation. However, the most effective heuristic approaches to
the problem appear to be those based on Lagrangian relaxation, following the
seminal work by Balas and Ho (1980), and then the improvements by Beasley
(1990), Fisher and Kedia (1990), Wedelin (1995), Balas and Carrera (1996),
Ceria et al. (1998), Caprara et al. (1999), Kroon and Fischetti (2001), and
Yagiura et al. (2006), the latter making also extensive use of local search and
providing most of the best known solutions so far, even if within computing
times that are considerably larger than those of, e.g., Caprara et al. (1999).
Abbink et al. (2005) describe a railway crew scheduling problem in the Nether-
lands, where one of the objectives is to allocate the total workload as fairly as
possible among the crew depots, thereby taking into account both the attrac-
tive parts of the workload and the less attractive parts.
As to Crew Rostering, the only optimization approaches for the railway case
that we are aware of in the literature refer to the Italian case and are described
 Ch. 3. Passenger Railway Optimization 175

in Caprara et al. (1998). The other published works on the Crew Rostering
Problem concern urban mass-transit systems, where the minimum number of
crews required to perform the duties can easily be determined, and the objec-
tive is to evenly distribute the workload among the crews, and the airline case,
for which Set Partitioning approaches can be used, see Ryan (1992), Gamache
and Soumis (1998), Gamache et al. (1999), as well as the above mentioned
surveys by Bodin et al. (1983), Ernst et al. (2004a, 2004b), and Wren (1981).
Related cyclic staff scheduling problems are dealt with in Tien and Kamiyama
(1982), Balakrishnan and Wong (1990), and Caprara et al. (2003).
The only attempts to integrate the Crew Scheduling and Rostering phases
for railway CPP, to the best of our knowledge, are by Caprara et al. (2001),
Ernst et al. (2001), and Freling et al. (2004). The latter two are based on Set
Covering/Partitioning approaches, whereas the first one is outlined in the fol-
lowing sections.
Other approaches for railway CPP can be found in Morgado and Martins
(1992), Chu and Chan (1998), Kwan et al. (2001), Fores et al. (2001), Freling
et al. (2001), and Constantino et al. (2006).

Model formulation

Crew Scheduling and Rostering problems require finding min-cost se-

quences through a given set of items. Items correspond to trips for Crew
Scheduling, and to duties for Crew Rostering, whereas sequences correspond
to duties for Crew Scheduling, and to rosters for Crew Rostering.
A natural formulation of both problems in terms of graphs associates a node
with each item, and a directed arc with each possible item transition. More
specifically, one can define a directed multigraph G = (V  A) having one node
j ∈ V for each item, and an arc (i j) ∈ A if and only if item j can appear
right after item i in a feasible sequence. In some cases, two or more types of
transition from item i to item j are possible, e.g., two duties in a roster may be
separated by different types of rest (daily, weekly, etc.). In these cases, two or
more arcs from i to j are present in G, one for each type of transition. This
explains why G is a multigraph. In Crew Scheduling and Rostering problems
arising in railway applications the graph G is not acyclic, since the departure
and arrival times of the trips are intended modulo 24 hours. This allows an arc
to connect an item i to an item j even if the end time of i is greater than the
start time of j, meaning that a crew performs duties i and j on different days.
With this representation, the above problems call for a min-cost collection
of circuits of G covering each node once, as discussed in the sequel. There are
two basic ways of modeling the problem of covering the nodes of a directed
graph through a suitable set of circuits as an ILP.
The first model associates a binary variable with each arc (i j) ∈ A, indicat-
ing whether the arc is selected in an optimal solution or not. This is a natural
model that is particularly suitable for cases in which the most relevant con-
straints concern the direct transition of the items within the sequence, hence,
176 A. Caprara et al.

they can be effectively modeled through an appropriate definition of the arc

set A and the associated arc costs. On the other hand, this model can only
be applied when the cost of the solution can be expressed as the sum of the
costs associated with the arcs, and its LP relaxation can be very weak when the
operational constraints not concerning the direct transition of two items are
tight.
The second model has a possibly exponential number of binary variables,
each associated with a feasible circuit of G. The main advantages of this model
are that (i) it allows for circuit costs depending on the whole sequence of items,
and (ii) the feasibility constraints in the model do not have to include restric-
tions concerning the feasibility of a single circuit. This produces a formulation
whose LP relaxation is typically much tighter than in the previous model. How-
ever, the model often requires dealing with a very large number of variables.
In some cases, the explicit generation of all feasible circuits is impractical, and
one has to resort to a column generation approach, provided that an effective
column generation procedure is available to find feasible circuits whose corre-
sponding variable has a negative LP reduced cost.
In practice, the choice of the appropriate model and solution algorithm
strongly depends on the particular structure of the problem in hand. The sec-
ond model is particularly suitable for cases in which feasible circuits cover a
small number of nodes, and the constraints on the circuit feasibility are cum-
bersome and depend on the overall node sequence. This is the situation arising
in railway Crew Scheduling. On the contrary, as already mentioned, the first
model appears attractive for those cases where the main feasibility constraints
concern the direct sequencing of two nodes, since they can be dealt with im-
plicitly by an appropriate definition of the arc costs. This is the case of railway
Crew Rostering.

Crew Scheduling
Crew Scheduling calls for a min-cost collection of paths in G covering all the
nodes (trips) once, each path satisfying a set of constraints related to the feasi-
bility of the corresponding duty (maximum driving time, meal breaks, etc.). As
already mentioned, a basic constraint for Crew Scheduling is that every duty
must start and end at the crew home location (depot). It is then natural to in-
troduce in G a dummy node d for each depot, along with the associated arcs
(d j) (respectively, (j d)) for each node j associated with a trip which can be
the first (respectively, the last) trip in a duty assigned to depot d. This allows
one to convert each path representing a duty into a circuit by connecting the
terminal nodes of the path to the depot node representing the home location
of the crew. Let D denote the subset of nodes corresponding to the depots.
The model that is most frequently used in this case is the following. Let
C = {C1      Cn } denote the collection of all the simple circuits of G corre-
sponding to a feasible duty for one crew member, with n = |C |. Each circuit
Cj has an associated cost cj , and covers the node set Ij . The binary variable
yj takes value 1 if Cj is part of the optimal solution, and 0 otherwise. Letting
 Ch. 3. Passenger Railway Optimization 177

N := {1     n}, we then have the following Set Partitioning Problem with side
constraints:

min cj yj (56)
j∈N

subject to

yj = 1 v ∈ V \ D (57)
j∈N:v∈Ij

yj  |S| − 1 S ∈ S (58)
j∈S

yj ∈ {0 1} j ∈ N (59)

where S denotes the family of all inclusion-minimal sets S ⊆ N with the prop-
erty that no feasible solution contains all circuits Cj for j ∈ S. Constraints (57)
impose that each node not associated with a depot is covered by exactly one
circuit, whereas inequalities (58) model the crew depot constraints.
A main advantage of the Set Partitioning model is that it allows for circuit
costs depending on the whole sequence of arcs. In most cases, the feasibility
constraints (58), that need not take into account restrictions concerning the
feasibility of a single circuit, can be replaced by a compact set of inequalities of
the form By  w, modeling crew depot constraints only.
Due to the nature of the services to be carried out, in most railway applica-
tions a typical crew duty covers only a small number of trips. Moreover, heavy
operational constraints affect duty feasibility. This makes it sometimes practi-
cal to explicitly generate all feasible duties, which are computed and stored in
a preprocessing phase called duty generation. In addition, operational rules al-
low a crew to be transported with no extra cost as a passenger on a trip, hence,
the overall solution can cover a trip more than once, a main difference with re-
spect to airline applications. In this situation, the Set Partitioning formulation
(56)–(59) can profitably be replaced by its Set Covering relaxation obtained
by replacing “=” with “” in (57). As a result, only inclusion-maximal feasi-
ble duties, among those with the same cost, need to be considered in the duty
generation. This considerably reduces the number of variables.
Most of the existing literature is focused on the pure Set Covering Problem
without depot constraints, one of the exceptions being Abbink et al. (2005).
The case study of this section does not take into account depot constraints.
The main complication in this case is the size of the instances, with up to a few
million variables and a few thousand constraints. For this case, the key ideas
for the most successful heuristics by Caprara et al. (1999), Ceria et al. (1998),
Wedelin (1995), and Yagiura et al. (2006) are:
(i) the use of Lagrangian relaxation combined with iterative procedures to
find near-optimal multipliers (e.g., subgradient optimization), and
178 A. Caprara et al.

(ii) the use of Lagrangian costs to drive the construction of feasible solu-
tions at each iteration of these iterative procedures.
In our case study, we used the algorithm proposed in Caprara et al. (1999),
which is based on the selection of a suitable dynamic “core” problem, defined
by a small subset of the variables, so as to avoid working with the full problem.

Crew Rostering
In Crew Rostering, no dummy depot nodes are needed, as all duties refer to
the same crew depot. With an appropriate definition of the arc costs, the prob-
lem calls for a min-cost collection of circuits covering all the nodes once, each
circuit satisfying a set of constraints related to the feasibility of the associated
roster.
The same model used for the Crew Scheduling phase (and for the Crew
Rostering case in some airline applications, see Gamache and Soumis, 1998;
Gamache et al., 1999; Ryan, 1992) does not seem to be effective here, because
of the difficulties in using a column generation technique, probably due to the
combination of the relatively large number of duties in a roster and the very
complicated operational constraints imposed on a roster.
An alternative model associates a binary variable xa with each arc a ∈ A,
where xa = 1 if and only if arc a is used in the optimal solution. Let ca be the
cost of each arc a ∈ A. In case the objective is to minimize the overall length of
the rosters (that often coincides with the number of crews needed to perform
all the rosters, see Caprara et al., 1998), the cost of arc a = (u v) can be set
equal to the time elapsing between the start of duty u and the start of duty v
in case they are consecutive in the same roster. Moreover, let δ+ (v) and δ− (v)
represent the set of arcs of G leaving and entering node v ∈ V , respectively.
The model reads:

min ca xa (60)
a∈A
subject to
 
xa = xa  v ∈ V (61)
a∈δ+ (v) a∈δ− (v)

xa = 1 v ∈ V (62)
a∈δ+ (v)

xa  |P| − 1 P ∈ P (63)
(ij)∈P
xa ∈ {0 1} a ∈ A (64)
where the family P contains the inclusion-minimal arc sequences (paths or cir-
cuits) P which cannot be part of any feasible solution. Note that |P | generally
grows exponentially with |V |.
Constraints (61) and (62) impose that each node is covered by exactly one
circuit. Constraints (63) forbid the choice of all the arcs in any infeasible arc
 Ch. 3. Passenger Railway Optimization 179

subset P. Notice that P contains all the arc sequences which cannot be covered
by a single crew because of operational constraints.
Provided that the objective function has been properly modeled by associ-
ating costs with arcs, as is the case in our case study, the main disadvantage
of the model (60)–(64) is the weakness of constraints (63) when its LP relax-
ation is considered. (On the other hand, their very large number is in principle
possible to handle, as it is easy to show that these constraints can be separated
efficiently whenever a polynomial time procedure is available to tell whether a
given sequence P is in P or not.) For this reason, it is often the case that the
above model is not used directly to derive feasible solutions. Rather, heuristic
algorithms can be driven by its relaxation without inequalities (63), and possi-
bly with additional inequalities that take into account the specific structure of
the problem. Their success is often related to the tightness of this relaxation.
The main idea of the heuristic algorithm presented in Caprara et al. (1998),
to which our case study refers, is to construct the rosters one at a time, and
to choose the next duty j to be sequenced after the last duty i in the current
roster as the one for which the relaxation value increment due to fixing xij = 1
is smallest.

An integrated approach

The main drawbacks of the classical approach that solves the Crew Schedul-
ing and Rostering phases sequentially are the following:
• The construction of the heuristic solutions in the Crew Scheduling
phase takes into account only the duty costs, and not directly the real
objective function, i.e., the minimization of the global cost of the ros-
ters for all the depots in the set D.
• The duty costs only partly reflect the constraints of the Crew Rostering
phase. In particular, it is difficult to find out, a priori and separately for
each depot, which are the constraints that will make the construction
of the rosters for this depot difficult.
• Only one solution is kept among those found in the Crew Scheduling
phase, whereas the Crew Rostering phase could produce much better
rosters starting from the duties selected in some other solution which
was not stored because its duty cost was not the best one.
These observations inspired the design of an integrated CPP approach, which
iteratively performs the Crew Scheduling phase and, within each iteration, calls
the Crew Rostering phase several times, as illustrated in Caprara et al. (2001).
In the sequel, we briefly illustrate the main features of this approach.
The first novelty concerns the possible updating of the best CPP solution
found so far each time a new Crew Scheduling solution is constructed. For each
such solution, given by the duty set S, one computes a simple lower bound L
on the value of the Crew Rostering solution for the duties in S. Letting z be the
value of the best CPP solution found so far (initially, z = ∞), if L < z, one calls
180 A. Caprara et al.

the Crew Rostering phase, as the duty set S may lead to a better CPP solution.
Let z H ( L) be the value of the solution found by the rostering optimization
phase, which is executed with a relatively small time limit. If z H < z, one
updates the best CPP solution obtained so far.
The second novelty concerns the definition of the duty costs for the Crew
Scheduling phase. In the first application of this phase these costs are set to
the same value as in the original approach (e.g., they are equal to the duty
duration). This first application is run with a given time limit (with several
applications of the Crew Rostering phase). At the end of the execution, the
information obtained from the various calls to the Crew Rostering phase, con-
cerning the effect of the duty characteristics on the rostering solution value
(possibly depending on the associated depot), is used to update the duty costs.
For instance, if overnight duties for a given depot turned out to be hard to
schedule in the rostering phase, then the cost of these duties is increased. The
Crew Scheduling phase is then applied again. The overall method terminates
after a prefixed time limit or number of applications of the Crew Scheduling
phase.

Experimental results

In this section, we illustrate the classical and integrated approaches on a

set of real-world instances provided by FS Trenitalia, the main Italian Train
Operator. The programs were run on a Digital Ultimate Workstation 533 MHz.
In Table 5, we report the main characteristics of each instance, namely the
number of trips (# trips), the number of depots (# depots), the number of
duties generated during the duty generation phase (# duties), and the time
required by the duty generation, expressed in seconds (time). The first three
instances refer to trains which have to be covered by crews from a single depot,
namely Mestre, Milan, and Verona, respectively. The fourth instance is asso-
ciated with trains to be covered by crews from the depots of Bolzano, Trieste,
and Udine, whereas the last two instances refer to international trains con-
necting Austria, Switzerland, France, Germany, and Italy, where the various
depots are located. Even if, at the moment, crews are handled separately for

Table 5.
Characteristics of the instances considered

Instance # trips # depots # duties Time

MESTRE 121 1 1024448 166

MILAN 502 1 874416 629
VERONA 86 1 484139 94
BZ_TS_UD 118 3 457021 109
A_CH_F_D_I_1 91 13 796771 136
A_CH_F_D_I_2 309 15 497847 370
 Ch. 3. Passenger Railway Optimization 181

each country, these last instances simulate what would happen if all crews for
the trains considered were handled by a unique European Train Operator.
The standard approach was run with time limits of 9000 seconds for the
Crew Scheduling phase and of 1000 seconds for the Crew Rostering phase.
The time limit for the first phase is much larger than the time limit for the sec-
ond one since the Crew Scheduling phase, having to deal with several hundred
thousand duties for these instances, is typically much more time consuming
than the Crew Rostering phase, which has to deal with a few hundred duties se-
lected in the previous phase, often subdivided among different depots. On the
same instances, we also ran the integrated approach. The overall time limit was
10,000 seconds, and the time limit for each internal execution of the rostering
optimization phase was set to 100 seconds. The condition for the termination of
each application of the Crew Scheduling phase is actually a logical one rather
than a time limit, as explained in Caprara et al. (2001).
The results are reported in Table 6. For each instance and each approach,
we give the number of crews in the final CPP solution (# crews), along with
the associated number of selected duties (# duties). Moreover, for the stan-
dard approach, we report a lower bound (LB) on the optimal value of the CPP
solution computed with respect to the duties selected by the Crew Schedul-
ing phase. For the integrated approach, we report the best overall CPP lower
bound (LB) computed over all Set Covering solutions found during the Crew
Scheduling phase. Note that this bound may not correspond to the set of duties
yielding the best CPP solution. Finally, for both approaches we report the time
required to obtain the best solution (time). This time is reported in seconds.
For the standard approach, this time is equal to the sum of the times spent to
find the best solutions required by the Crew Scheduling phase and, for each
depot, by the Crew Rostering phase.
The table shows the considerable improvement that is achieved by the in-
tegration of the Crew Scheduling and Rostering phases, leading to an average
percentage of saving of about 95%. Note that the time required to find the best
solution is similar for the two approaches, and that the number of duties in the
best solution is in some cases much smaller for the integrated approach (for

Table 6.
Solutions found with the standard and integrated approaches

Instance Standard approach Integrated approach

# crews LB # duties Time # crews LB # duties Time

MESTRE 66 66 37 1118 66 66 21 148

MILAN 300 300 145 6082 288 282 152 7786
VERONA 48 48 20 102 42 42 22 344
BZ_TS_UD 66 66 30 103 60 54 28 138
A_CH_F_D_I_1 72 72 28 812 66 66 24 3434
A_CH_F_D_I_2 294 294 127 1208 246 240 94 633
182 A. Caprara et al.

instances MESTRE and A_CH_F_D_I_2). We also observe that the improve-

ment is particularly significant for instance A_CH_F_ D_I_2, probably because
the integrated approach is able to subdivide the selected duties among the de-
pots in a much more effective way than the standard one.

8 Perspective

In this chapter we have described several mathematical models and opti-

mization techniques that have been developed for effectively supporting tradi-
tional planning processes in passenger railway transportation. A lot of research
has been carried out in this area, both of a practical and theoretical nature.
The results of this research are starting to be applied in practice. For example,
several railway companies are currently using automated crew scheduling and
train platforming systems. Usually, the underlying models and solution tech-
niques of these systems are similar to the ones described in this chapter.
We have also given in the introduction a brief description of the strategic is-
sues of long term rolling stock and crew management. Although these subjects
did not yet receive a lot of attention of researchers in mathematical optimiza-
tion so far, they may be fruitful areas of further research. In particular, the
current solutions to these long term planning problems determine the railway
systems of the future. Being able to carry out these planning processes in a ra-
tional way supported by effective decision support tools may have even more
effect on the quality of future railway systems than being able to find better
solutions to the operational problems in less time.
Real-time control is at the other side of the planning spectrum. The current
trend in the railway industry is a shift from “planning in detail” to “effective
real-time control”. Disturbances and disruptions in the railway operations are
inevitable. Therefore, large parts of the operational plans are never carried
out. In case of a disruption, one needs as soon as possible an alternative plan.
To some extent, several potential alternative plans can be prepared already,
e.g., in the form of disruption scenarios for adapting the timetable and the
rolling stock circulation. The latter may be particularly affective in the case of
a cyclic timetable. However, crew schedules are usually noncyclic. Therefore,
being able to quickly generate alternative crew schedules is highly important
in case of a disruption of the railway system. In order to make this effective in
practice, one needs:
(i) to have detailed information on the status quo of the railway systems
(e.g., the positions of trains and crews),
(ii) to be able to quickly generate alternative crew schedules, and
(iii) to disseminate the alternative plans in a dependable way among all
stake-holders.
Although, from a mathematical point of view, these problems may seem to
be similar to the corresponding operational planning problems, they are quite
 Ch. 3. Passenger Railway Optimization 183

different, mainly due to the dynamic character of real-time control and the
high time pressure.
Another fruitful area of further research is the provision of dependable dy-
namic travel information to the passengers, in particular in case of a disruption
of the railway system. In such cases, railway companies are often blamed for
providing insufficient or incorrect travel information to the passengers. Pro-
viding passengers with dependable dynamic travel information requires both
quickly determining alternative travel paths for the passengers, and dissemi-
nating this information in a dependable way among the passengers.
One of the aims of this chapter is to illustrate the fact that railway systems
are abundant of interesting combinatorial optimization problems. Currently,
many of these problems are still solved manually in practice, partly due to
the traditionally rather conservative character of the railway industry. On the
other hand, the innovative possibilities provided by the effective application
of mathematical models and optimization techniques were also recognized by
the railway industry itself, and software applications based on these techniques
recently started to be implemented. Researchers in mathematical optimization
should grasp the currently available momentum and opportunities in the rail-
way industry by not focusing too much on theoretical results, but by going for
real-world applications of their models and techniques. The latter will lead to
a win–win situation, both for the researchers and for the railway industry.

References

Abbink, E.W.J., van den Berg, B.W.V., Kroon, L.G., Salomon, M. (2004). Allocation of railway rolling
stock for passenger trains. Transportation Science 38, 33–41.
Abbink, E.W.J., Fischetti, M., Kroon, L.G., Timmer, G., Vromans, M.J.C.M. (2005). Reinventing crew
scheduling at Netherlands railways. Interfaces 35, 393–401.
Arabeyre, J., Fearnley, J., Steiger, F., Teather, W. (1969). The airline crew scheduling problem: A survey.
Transportation Science 3, 140–163.
Balakrishnan, N., Wong, R.T. (1990). A network model for the rotating workforce scheduling problem.
Networks 20, 25–42.
Balas, E., Carrera, M.C. (1996). A dynamic subgradient-based branch-and-bound procedure for set
covering. Operations Research 44, 875–890.
Balas, E., Ho, A. (1980). Set covering algorithms using cutting planes, heuristics and subgradient opti-
mization: A computational study. Mathematical Programming Study 12, 37–60.
Barnhart, C., Johnson, E.L., Nemhauser, G.L., Savelsbergh, M.W.P., Vance, P.H. (1994). Branch-and-
price: Column generation for solving huge integer programs. In: Birge, J.R., Murty, K.G. (Eds.),
Mathematical Programming: State of the Art 1994. University of Michigan Press, Ann Arbor, pp. 186–
207.
Beasley, J.E. (1987). An algorithm for set covering problems. European Journal of Operational Re-
search 31, 85–93.
Beasley, J.E. (1990). A Lagrangian heuristic for set covering problems. Naval Research Logistics 37,
151–164.
Beasley, J.E., Chu, P.C. (1996). A genetic algorithm for the set covering problem. European Journal of
Operational Research 94, 392–404.
Beasley, J.E., Jörnsten, K. (1992). Enhancing an algorithm for set covering problems. European Journal
of Operational Research 58, 293–300.
184 A. Caprara et al.

Ben-Khedher, N., Kintanar, J., Queille, C., Stripling, W. (1998). Schedule optimization at SNCF: From
conception to day of departure. Interfaces 28, 6–23.
Billionnet, A. (2003). Using integer programming to solve the train platforming problem. Transportation
Science 37, 213–222.
Blasum, U., Bussieck, M.R., Hochstättler, W., Moll, C., Scheel, H.H., Winter, T. (2000). Scheduling
trams in the morning. Mathematical Methods of Operations Research 49, 137–148.
Bodin, L., Golden, B., Assad, A., Ball, M. (1983). Routing and scheduling of vehicles and crews: The
state of the art. Computers & Operations Research 10, 63–211.
Bouma, A., Oltrogge, C. (1994). Linienplanung und Simulation für Öffentlichen Verkehrswege in Praxis
und Theorie. Eisenbahntechnische Rundschau 43, 369–378 (in German).
Brännlund, U., Lindberg, P.O., Nöu, A., Nilsson, J.E. (1998). Railway timetabling using Lagrangian
relaxation. Transportation Science 32, 358–369.
Brucker, J., Hurink, J.L., Rolfes, T. (2003). Routing of railway carriages: A case study. Journal of Global
Optimization 27, 313–332.
Bussieck, M.R. (1998). Optimal lines in public rail transport. PhD thesis, TU Braunschweig.
Bussieck, M.R., Kreuzer, P., Zimmermann, U.T. (1996). Optimal lines for railway systems. European
Journal of Operational Research 96, 54–63.
Bussieck, M.R., Winter, T., Zimmermann, U.T. (1997). Discrete optimization in public rail transport.
Mathematical Programming 79, 415–444.
Cai, X., Goh, C.J. (1994). A fast heuristic for the train scheduling problem. Computers & Operations
Research 21, 499–510.
Caprara, A., Fischetti, M., Toth, P., Vigo, D., Guida, P.L. (1997). Algorithms for railway crew manage-
ment. Mathematical Programming 79, 125–141.
Caprara, A., Fischetti, M., Toth, P., Vigo, D. (1998). Modeling and solving the crew rostering problem.
Operations Research 46, 820–830.
Caprara, A., Fischetti, M., Toth, P. (1999). A heuristic method for the set covering problem. Operations
Research 47, 730–743.
Caprara, A., Fischetti, M., Toth, P. (2000). Algorithms for the set covering problem. Annals of Operations
Research 98, 353–371.
Caprara, A., Monaci, M., Toth, P. (2001). A global method for crew planning in railway applications.
In: Daduna, J., Voss, S. (Eds.), Computer-Aided Transit Scheduling. Lecture Notes in Economics and
Mathematical Systems, vol. 505. Springer-Verlag, Berlin, pp. 17–36.
Caprara, A., Fischetti, M., Toth, P. (2002). Modeling and solving the train timetabling problem. Opera-
tions Research 50, 851–861.
Caprara, A., Monaci, M., Toth, P. (2003). Models and algorithms for a staff scheduling problem. Math-
ematical Programming 98, 445–476.
Caprara, A., Monaci, M., Toth, P., Guida, P.L. (2006). A Lagrangian heuristic approach to real-world
train timetabling problems. Discrete Applied Mathematics 154, 738–753.
Carey, M., Carville, S. (2003). Scheduling and platforming trains at busy complex stations. Transporta-
tion Research 37, 195–224.
Carey, M., Lockwood, D. (1995). A model, algorithms and strategy for train pathing. Journal of the
Operational Research Society 46, 988–1005.
Ceria, S., Nobili, P., Sassano, A. (1998). A Lagrangian-based heuristic for large-scale set covering prob-
lems. Mathematical Programming 81, 215–228.
Chu, S., Chan, E. (1998). Crew scheduling of light rail transit in Hong Kong: From modeling to imple-
mentation. Computers & Operations Research 25, 887–894.
Claessens, M.T., van Dijk, N.M., Zwaneveld, P.J. (1998). Cost optimal allocation of passenger lines.
European Journal of Operational Research 110, 474–489.
Constantino, A.A., de Mendonca Neto, C.F.X., Novaes, A.G. (2006). Crew rostering problem with
distribution of workload based on preferences. Annals of Operations Research, in press.
Cordeau, J.-F., Toth, P., Vigo, D. (1998). A survey of optimization models for train routing and schedul-
ing. Transportation Science 32, 380–404.
Cordeau, J.-F., Soumis, F., Desrosiers, J. (2001). Simultaneous assignment of locomotives and cars to
passenger trains. Operations Research 49, 531–548.
 Ch. 3. Passenger Railway Optimization 185

Daduna, J.R., Voss, S. (Eds.) (2001). Computer-Aided Scheduling of Public Transport. Lecture Notes in
Economic and Mathematical Systems, vol. 505. Springer-Verlag, Berlin.
Daduna, J.R., Wren, A. (Eds.) (1988). Computer-Aided Transit Scheduling. Lecture Notes in Economic
and Mathematical Systems, vol. 308. Springer-Verlag, Berlin.
Dahlhaus, E., Horak, P., Miller, M., Ryan, J.F. (2000). The train marshalling problem. Discrete Applied
Mathematics 103, 41–54.
De Luca Cardillo, D., Mione, N. (1998). k l-list τ colouring of graphs. European Journal of Operational
Research 106, 160–164.
Desrosiers, J., Dumas, Y., Solomon, M.M., Soumis, F. (1995). Time constrained routing and scheduling.
In: Ball, M.O., Magnanti, T.L., Monma, C.L., Nemhauser, G.L. (Eds.), Handbooks in Operations
Research and Management Science, vol. 8. Elsevier, Amsterdam, pp. 35–139.
Desrochers, M., Rousseau, J.M. (Eds.) (1992). Computer-Aided Transit Scheduling. Lecture Notes in
Economic and Mathematical Systems, vol. 386. Springer-Verlag, Berlin.
Dienst, H. (1978). Linienplanung in Spurgeführten Personenverkehr mit Hilfe eines Heuristischen Ver-
fahrens, PhD thesis, TU Braunschweig (in German).
Di Stefano, G., Koci, M.L. (2004). A graph theoretical approach to the shunting problem. In: Gerards,
B. (Ed.), Proceedings of ATMOS Workshop 2003. Electronic Notes in Theoretical Computer Science,
vol. 92. Elsevier, Amsterdam, pp. 16–33.
Ernst, A.T., Jiang, H., Krishnamoorthy, M., Nott, H., Sier, D. (2001). An integrated optimization model
for train crew management. Annals of Operations Research 108, 211–224.
Ernst, A.T., Jiang, H., Krishnamoorthy, M., Sier, D. (2004a). Staff scheduling and rostering: A review
of applications, methods and models. European Journal of Operational Research 153, 3–27.
Ernst, A.T., Jiang, H., Krishnamoorthy, M., Owens, B., Sier, D. (2004b). Annotated bibliography of
personnel scheduling and rostering. Annals of Operations Research 127, 21–144.
Fioole, P.-J., Kroon, L.G., Maróti, G., Schrijver, A. (2006). A rolling stock circulation model for com-
bining and splitting of passenger trains. European Journal of Operational Research 174, 1281–1297.
Fisher, M.L., Kedia, P. (1990). Optimal solutions of set covering/partitioning problems using dual
heuristics. Management Science 36, 674–688.
Fores, S., Proll, L., Wren, A. (2001). Experiences with a flexible driver scheduler. In: Daduna, J., Voss, S.
(Eds.), Computer-Aided Transit Scheduling. Lecture Notes in Economics and Mathematical Systems,
vol. 505. Springer-Verlag, Berlin, pp. 137–152.
Freling, R., Lentink, R.M., Odijk, M. (2001). Scheduling train crews: A case study for the Dutch rail-
ways. In: Daduna, J., Voss, S. (Eds.), Computer-Aided Transit Scheduling. Lecture Notes in Economics
and Mathematical Systems, vol. 505. Springer-Verlag, Berlin, pp. 153–166.
Freling, R., Lentink, R.M., Wagelmans, A.P.M. (2004). A decision support system for crew planning
in passenger transportation using a flexible branch-and-price algorithm. Annals of Operations Re-
search 127, 203–222.
Freling, R., Lentink, R.M., Kroon, L.G., Huisman, D. (2005). Shunting of passenger train units in a
railway station. Transportation Science 39, 261–272.
Gallo, G., di Miele, F. (2001). Dispatching buses in parking depots. Transportation Science 35, 322–330.
Gamache, M., Soumis, F. (1998). A method for optimally solving the rostering problem. In: Yu, G.
(Ed.), Operations Research in Airline Industry. Kluwer Academic, Boston, pp. 124–157.
Gamache, M., Soumis, F., Marquis, G., Desrosiers, J. (1999). A column generation approach for large
scale aircrew rostering problems. Operations Research 47, 247–263.
Goossens, J.H.M., van Hoesel, C.P.M., Kroon, L.G. (2004). A branch-and-cut approach for solving
railway line-planning problems. Transportation Science 38, 379–393.
Goossens, J.H.M., van Hoesel, C.P.M., Kroon, L.G. (2005). On solving multi-type railway line planning
problems. European Journal of Operational Research 168, 403–424.
Hamdouni, M., Desaulniers, G., Marcotte, O., Soumis, F., Van Putten, M. (2006). Dispatching buses in
a depot using block patterns. Transportation Science 40, 364–377.
He, S., Song, R., Chaudry, S.S. (2000). Fuzzy dispatching model and genetic algorithms for railyard
operations. European Journal of Operational Research 124, 307–331.
Higgings, A., Kozan, E., Ferreira, L. (1997). Heuristic techniques for single line train scheduling. Jour-
nal of Heuristics 3, 43–62.
186 A. Caprara et al.

Huisman, D., Kroon, L.G., Lentink, R.M., Vromans, M.J.C.M. (2005). Operations research in passen-
ger railway transportation. Statistica Neerlandica 59, 467–497.
Jacobs, L.W., Brusco, M.J. (1995). A local search heuristic for large set-covering problems. Naval Re-
search Logistics 52, 1129–1140.
Jovanovic, D., Harker, P.T. (1991). Tactical scheduling of rail operations: The SCAN I system. Trans-
portation Science 25, 46–64.
Kroon, L.G., Fischetti, M. (2001). Crew scheduling for the Netherlands railways destination: Customer.
In: Daduna, J., Voss, S. (Eds.), Computer-Aided Transit Scheduling. Lecture Notes in Economics and
Mathematical Systems, vol. 505. Springer-Verlag, Berlin, pp. 181–201.
Kroon, L.G., Peeters, L.W.P. (2003). A variable trip time model for cyclic railway timetabling. Trans-
portation Science 37, 198–212.
Kroon, L.G., Romeijn, H.E., Zwaneveld, P.J. (1997). Routing trains through railway stations: Complex-
ity issues. European Journal of Operational Research 98, 485–498.
Kroon, L.G., Dekker, R., Vromans, M.J.C.M. (2005). Cyclic railway timetabling: A stochastic optimiza-
tion approach. Technical Report ERS-2005-051-LIS, Erasmus University Rotterdam.
Kwan, A., Kwan, R., Parker, M., Wren, A. (2001). Producing train driver schedules under different
operating strategies. In: Wilson, N. (Ed.), Computer-Aided Transit Scheduling. Lecture Notes in Eco-
nomics and Mathematical Systems, vol. 471. Springer-Verlag, Berlin, pp. 129–154.
Lindner, T. (2000). Train schedule optimization in public rail transport. PhD thesis, TU Braunschweig.
Lingaya, N., Cordeau, J.-F., Desaulniers, G., Desrosiers, J., Soumis, F. (2002). Operational car assign-
ment at VIA rail Canada. Transportation Research 36, 755–778.
Lorena, L.A.N., Lopes, F.B. (1994). A surrogate heuristic for set covering problems. European Journal
of Operational Research 79, 138–150.
Maróti, G., Kroon, L.G. (2005). Maintenance routing for train units: The transition model. Transporta-
tion Science 39, 518–525.
Maróti, G., Kroon, L.G. (2007). Maintenance routing for train units: The scenario model. Computers &
Operations Research 34, 1121–1140.
Morgado, E., Martins, J. (1992). Scheduling and managing crew in the Portuguese railways. Expert
Systems with Applications 5, 301–321.
Nachtigall, K. (1999). Periodic network optimization and fixed interval timetables. Habilitation thesis,
Deutsches Zentrum für Luft-und Raumfahrt, Braunschweig.
Nachtigall, K., Voget, S. (1996). A genetic algorithm approach to periodic railway synchronization.
Computers & Operations Research 23, 453–463.
Odijk, M. (1996). A constraint generation algorithm for the construction of periodic railway timetables.
Transportation Research 30, 455–464.
Oliveira, E., Smith, B.M. (2000). A job-shop scheduling model for the single-track railway scheduling
problem. School of Computing Research Report 2000.21, University of Leeds.
Oltrogge, C. (1994). Linienplanung für Mehrstufige Bedienungssysteme in Öffentlichen Personen-
verkehr. PhD thesis, TU Braunschweig (in German).
Peeters, L.W.P. (2003). Cyclic railway timetable optimization. PhD thesis, Erasmus University Rotter-
dam.
Peeters, M., Kroon, L.G. (2003). Circulation of railway rolling stock: A branch-and-price approach.
Computers & Operations Research, in press.
Rousseau, J.M. (Ed.) (1985). Computer Scheduling of Public Transport 2. North-Holland, Amsterdam.
Ryan, D.M. (1992). The solution of massive generalized set partitioning problems in aircrew rostering.
Journal of the Operational Research Society 43, 459–467.
Scholl, S. (2005). Customer-oriented line planning. PhD thesis, University of Kaiserslautern.
Schrijver, A. (1993). Minimum circulation of railway stock. CWI Quarterly 6, 205–217.
Schrijver, A. (2003). Rangeren op Opstelsporen. Technical report, CWI (in Dutch).
Schrijver, A., Steenbeek, A. (1994). Dienstregelingontwikkeling voor Railned. Technical report, CWI
(in Dutch).
Serafini, P., Ukovich, W. (1989). A mathematical model for periodic event scheduling problems. SIAM
Journal on Discrete Mathematics 2, 550–581.
 Ch. 3. Passenger Railway Optimization 187

Szpigel, B. (1973). Optimal train scheduling on a single track railway. In: Ross, M. (Ed.), Operation
Research’72. North-Holland, Amsterdam, pp. 343–351.
Tien, J.M., Kamiyama, A. (1982). On manpower scheduling algorithms. SIAM Review 24, 275–287.
Tomii, N., Zhou, L.J. (2000). Depot shunting scheduling using combined genetic algorithms and PERT.
In: Brebbia, C.A., Allan, J., Hill, R.J., Sciutto, G., Sone, S. (Eds.), Computers in Railways VII. Ad-
vances in Transport, vol. 7. WIT Press, Southampton, pp. 437–446.
Tomii, N., Zhou, L.J., Fukumara, N. (1999). Shunting scheduling problem at railway stations. In: Imam,
I.F., Kodratoff, Y., El-Dessouki, A., Ali, M. (Eds.), Lecture Notes in Artificial Intelligence, vol. 1611.
Springer-Verlag, Berlin, pp. 790–797.
van den Broek, J.J.J. (2002). Toets op Inplanbaarheid van Rangeerbewegingen. MSc thesis, Eindhoven
University of Technology (in Dutch).
van’t Woudt, C. (2001). Shunting of passenger train units. MSc thesis, Erasmus University Rotterdam.
Wedelin, D. (1995). An algorithm for large scale 0-1 integer programming with application to airline
crew scheduling. Annals of Operational Research 57, 283–301.
Wilson, N. (Ed.) (1999). Computer-Aided Transit Scheduling. Lecture Notes in Economic and Mathemat-
ical Systems, vol. 471. Springer-Verlag, Berlin.
Winter, T. (1999). Online and real-time dispatching problems. PhD thesis, TU Braunschweig.
Winter, T., Zimmermann, U.T. (2000). Real-time dispatch of trams in storage yards. Annals of Opera-
tions Research 96, 287–315.
Wise, T.H. (1995). Column generation and polyhedral combinatorics for airline crew scheduling. PhD
thesis, Cornell University.
Wren, A. (Ed.) (1981). Computer Scheduling of Public Transport. North-Holland, Amsterdam.
Yagiura, M., Kishida, M., Ibaraki, T. (2006). A 3-flip neighborhood local search for the set covering
problem. European Journal of Operational Research 172, 472–499.
Zwaneveld, P.J. (1997). Railway planning and allocation of passenger lines. PhD thesis, Rotterdam
School of Management.
Zwaneveld, P.J., Kroon, L.G., Romeijn, H.E., Salomon, M., Dauzère-Pérès, S., van Hoesel, C.P.M., Am-
bergen, H.W. (1996). Routing trains through railway stations: Model formulation and algorithms.
Transportation Science 30, 181–194.
Zwaneveld, P.J., Kroon, L.G., van Hoesel, C.P.M. (2001). Routing trains through a railway station based
on a node packing model. European Journal of Operational Research 128, 14–33.
C. Barnhart and G. Laporte (Eds.), Handbook in OR & MS, Vol. 14
Copyright © 2007 Elsevier B.V. All rights reserved
DOI: 10.1016/S0927-0507(06)14004-9

Chapter 4
Maritime Transportation
Marielle Christiansen
Department of Industrial Economics and Technology Management, Norwegian University
of Science and Technology, Trondheim, Norway
Department of Applied Economics and Operations Research, SINTEF Technology and
Society, Trondheim, Norway
E-mail: [email protected]

Kjetil Fagerholt
Department of Industrial Economics and Technology Management, Norwegian University
of Science and Technology, Trondheim, Norway
Department of Marine Technology, Norwegian University of Science and Technology,
Trondheim, Norway
Norwegian Marine Technology Research Institute (MARINTEK), Trondheim, Norway
E-mail: [email protected]

Bjørn Nygreen
Department of Industrial Economics and Technology Management, Norwegian University
of Science and Technology, Trondheim, Norway
E-mail: [email protected]

David Ronen
College of Business Administration, University of Missouri-St. Louis, St. Louis, MO, USA
E-mail: [email protected]

1 Introduction

Maritime transportation is the major conduit of international trade, but the

share of its weight borne by sea is hard to come by. The authors have surveyed
the academic members of the International Association of Maritime Econo-
mists and their estimates of that elusive statistic range from 65% to 85%.
Population growth, increasing standard of living, rapid industrialization, ex-
haustion of local resources, road congestion, and elimination of trade barriers,
all of these contribute to the continuing growth in maritime transportation. In
countries with long shorelines or navigable rivers, or in countries consisting of
multiple islands, water transportation may play a significant role also in do-
mestic trades, e.g., Greece, Indonesia, Japan, Norway, Philippines, and USA.
Table 1 demonstrates the growth in international seaborne trade during the
last couple of decades (compiled from UNCTAD, 2003, 2004).
Since 1980 the total international seaborne trade has increased by 67% in
terms of weight. Tanker cargo has increased modestly, but dry bulk cargo has

189
190 M. Christiansen et al.

Table 1.
Development of international seaborne trade (millions of tons)

Year Tanker cargo Dry cargo Total

Main bulk Other
commodities1

1980 1871 796 1037 3704

1990 1755 968 1285 4008
2000 2163 1288 2421 5872
2001 2174 1331 2386 5891
2002 2129 1352 2467 5948
20032 2203 1475 2490 6168

1 Iron ore, grain, coal, bauxite/alumina, and phosphate.

2 Estimates.

Table 2.
World fleet by vessel type (million dwt)

Year Oil Bulk General Container Other Total

tankers carriers cargo ships

1980 339 186 116 11 31 683

1990 246 234 103 26 49 658
2000 286 281 103 69 69 808
2001 286 294 100 77 69 826
2002 304 300 97 83 60 844
2003 317 307 95 91 47 857

increased by 85%. The “Other” dry cargo, which consists of general cargo (in-
cluding containerized cargo) and minor dry bulk commodities, has more than
doubled.
The world maritime fleet has grown in parallel with the seaborne trade. Ta-
ble 2 provides data describing the growth of the world fleet during the same
period (compiled from UNCTAD, 2003, 2004).
The cargo carrying capacity of the world fleet has reached 857 million tons
at the end of 2003, an increase of 25% over 1980. It is worth pointing out the
fast growth in the capacity of the container ships fleet with 727% increase dur-
ing the same period. These replace general cargo ships in major liner trades. To
a lesser extent we see also a significant growth in the bulk carriers fleet. The
gap between the increase in total trade (67%) and in the world fleet (25%)
is explained by two factors. First, the boom in construction of tankers during
the 1970s that resulted in excess capacity in 1980, and second, the increas-
ing productivity of the world fleet, as demonstrated in Table 3 (compiled from
UNCTAD, 2003, 2004).
 Ch. 4. Maritime Transportation 191

Table 3.
Productivity of the world fleet

Year World fleet Total cargo∗ Total ton-miles Tons carried Thousands of
(million dwt) (million tons) performed (thousands of per dwt ton-miles
millions of ton-miles) performed per dwt

1980 6828 3704 16,777 54 255

1990 6584 4008 17,121 61 260
2000 8084 5871 23,016 73 285
2001 8257 5840 23,241 71 281
2002 8442 5888 23,251 70 275
2003 8570 6168 24,589 72 287

∗ Inconsistencies between these data and the Total in Table 1 are in the source. However, they do not
affect the productivity statistics presented in this table.

The utilization of the world fleet has increased from 5.4 tons carried per
deadweight ton in 1980 to 7.2 in 2003. At the same time the annual output per
deadweight ton has increased from 25.5 thousand ton-miles to 28.7.
These statistics demonstrate the dependence of the world economy on
seaborne trade. A ship involves a major capital investment (usually millions
of US dollars, tens of millions for larger ships) and the daily operating cost of
a ship may easily amount to thousands of dollars and tens of thousands for the
larger ships. Proper planning of fleets and their operations has the potential
of improving their economic performance and reducing shipping costs. This is
often a key challenge faced by the industry actors in order to remain competi-
tive.
The purpose of this chapter is to introduce the reader who is familiar with
Operations Research (OR), and may be acquainted with other modes of trans-
portation, to maritime transportation. The term maritime transportation refers
to seaborne transportation, but we shall include in this chapter also other
water-borne transportation, namely inland waterways. The chapter discusses
various aspects of maritime transportation operations and presents associated
decision making problems and models with an emphasis on ship routing and
scheduling models. This chapter focuses on prescriptive OR models and as-
sociated methodologies, rather than on descriptive models that are usually of
interest to economists and public policy makers. Therefore we do not discuss
statistical analysis of trade and modal-split data, nor ship safety and casualty
records and related topics. To explore these topics the interested reader may
refer to journals dealing with maritime economics, such as Maritime Policy
and Management and Maritime Economics and Logistics (formerly International
Journal of Maritime Economics).
The ocean shipping industry has a monopoly on transportation of large vol-
umes of cargo among continents. Pipeline is the only transportation mode that
is cheaper than ships (per cargo ton-mile) for moving large volumes of cargo
over long distances. However, pipelines are far from versatile because they can
192 M. Christiansen et al.

Table 4.
Comparison of operational characteristics of freight transportation modes

Operational characteristic Mode

Ship Aircraft Truck Train Pipeline

Barriers to entry small medium small large large

Industry concentration low medium low high high
Fleet variety (physical & large small small small NA
economic)
Power unit is an integral part yes yes often no NA
of the transportation unit
Transportation unit size fixed fixed usually fixed variable NA
Operating around the clock usually seldom seldom usually usually
Trip (or voyage) length days–weeks hours–days hours–days hours–days days–weeks
Operational uncertainty larger larger smaller smaller smaller
Right of way shared shared shared dedicated dedicated
Pays port fees yes yes no no no
Route tolls possible none possible possible possible
Destination change while possible no no no possible
underway
Port period spans multiple yes no no yes NA
operational time windows
Vessel-port compatibility yes seldom no no NA
depends on load weight
Multiple products shipped yes no yes yes NA
together
Returns to origin no no yes no NA

NA – not applicable.

move only fluids in bulk over fixed routes, and they are feasible and econom-
ical only under very specific conditions. Other modes of transportation (rail,
truck, air) have their advantages, but only aircraft can traverse large bodies of
water, and they have limited capacity and much higher costs than ships, thus
they attract high-value low-volume cargoes. Ships are probably the least reg-
ulated mode of transportation because they usually operate in international
water, and very few international treaties cover their operations.
Ship fleet planning problems are different than those of other modes of
transportation because ships operate under different conditions. Table 4 pro-
vides a comparison of the operational characteristics of the different freight
transportation modes. We wish also to point out that ships operate mostly in
international trades, which means that they are crossing multiple national juris-
dictions. Actually, in many aspects aircraft are similar to ships. In both modes
each unit represents a large capital investment that translates into high daily
cost, both pay port fees and both operate in international routes. However,
most aircraft carry mainly passengers whereas most ships haul freight. Even
aircraft that transport freight carry only packaged goods whereas ships carry
mostly liquid and dry bulk cargo, and often nonmixable products in separate
 Ch. 4. Maritime Transportation 193

compartments. Since passengers do not like to fly overnight most aircraft are
not operated around the clock whereas ships are operated continually. In ad-
dition, aircraft come in a small number of sizes and models whereas among
ships we find a large variety of designs that result in nonhomogeneous fleets.
Both ships and aircraft have higher uncertainty in their operations due to their
higher dependence on weather conditions and on technology, and because they
usually straddle multiple jurisdictions. However, since ships operate around
the clock their schedules usually do not have buffers of planned idleness that
can absorb delays. As far as trains are concerned, they have their own dedi-
cated right of way, they cannot pass each other except for at specific locations,
and their size and composition are flexible (both number of cars and number
of power units). Thus the operational environment of ships is different from
other modes of freight transportation, and they have different fleet planning
problems.
The maritime transportation industry is highly fragmented. The web site of
Lloyd’s Register boasts of listing of “   over 140,000 ship and 170,000 ship
owner and manager entries”. In order to take advantage of differences among
national tax laws, financial incentives, and operating rules, the control struc-
ture of a single vessel may involve multiple companies registered in different
countries.
Although ships are the least regulated mode of transportation, there are
significant legal, political, regulatory, and economic aspects involved in mar-
itime transportation. The control structure of a ship can be designed to hide
the identity of the real owner in order to minimize liability and taxes. Liability
for shipping accidents may be hard to pinpoint, and damages may be impos-
sible to collect, because numerous legal entities from different countries are
usually involved, such as: owner, operator, charterer, flag of registration, ship-
yards, classification society, surveyors, and contractors. That is in addition to
the crew that may have multiple nationalities and multiple native languages.
Only a small share of the world fleet competes directly with other modes of
transportation. However, in certain situations such competition may be im-
portant and encouraged by government agencies. In short haul operations,
relieving road congestion by shifting cargo and passengers to ships is often de-
sirable and even encouraged through incentives and subsidies. A central policy
objective of the European Union for the upcoming years is to improve the
quality and efficiency of the European transportation system by shifting traf-
fic to maritime and inland waterways, revitalizing the railways and linking up
the different modes of transport. For further information regarding the Euro-
pean transport policy see the European Commission’s white paper European
Transport Policy for 2010: Time to Decide (European Commission, 2004). This
source provides information about many of the European Union’s programs
where maritime transportation plays a prominent role.
Transportation planning has been widely discussed in the literature but most
of the attention has been devoted to aircraft and road transportation by trucks
and buses. Other modes of transportation, i.e., pipeline, water, and rail, have
194 M. Christiansen et al.

attracted far less attention. One may wonder what the reason is for that lower
attention, especially when considering the large capital investments and op-
erating costs associated with these modes. Pipeline and rail operate over a
dedicated right of way, have major barriers to entry, and relatively few op-
erators in the market. These are some issues that may explain the lower level
of attention. It is worth mentioning that research on rail planning problems
has increased considerably during the last fifteen years. However, the issues
mentioned for pipeline and rail do not hold for water transportation. Several
explanations follow for the low attention drawn in the literature by maritime
transportation planning problems:
Low visibility. In most regions people see trucks, aircraft, and trains, but not
ships. Worldwide, ships are not the major transportation mode. Most cargo
is moved by truck or rail. Moreover, research is often sponsored by large
organizations. Numerous large organizations operate fleets of trucks, but
few such organizations operate ships.
Maritime transportation planning problems are less structured. In maritime
transportation planning there is a much larger variety in problem struc-
tures and operating environments. That requires customization of decision
support systems, and makes them more expensive. In recent years we
see more attention attracted by more complex problems in transportation
planning, and this is manifested also in maritime transportation.
In maritime operations there is much more uncertainty. Ships may be delayed
due to weather conditions, mechanical problems and strikes (both on
board and on shore), and usually, due to their high costs, very little slack is
built into their schedules. This results in a frequent need for replanning.
The ocean shipping industry has a long tradition and is fragmented. Ships have
been around for thousands of years and therefore the industry may be con-
servative and not open to new ideas. In addition, due to the low barriers
to entry there are many small, family owned, shipping companies. Most
quantitative models originated in vertically integrated organizations where
ocean shipping is just one component of the business.
In spite of the conditions discussed above we observe significant growth
in research in maritime transportation. The first review of OR work in ship
routing and scheduling appeared in 1983 (Ronen, 1983), and it traced papers
back to the 1950s. A second review followed a decade later (Ronen, 1993),
and recently a review of the developments over the last decade appeared
(Christiansen et al., 2004). Although these reviews focused on ship routing
and scheduling problems, they discussed also other related problems on all
planning levels. A feature issue on OR in water transportation was published
by the European Journal of Operational Research (Ronen, 1991), and a spe-
cial issue on maritime transportation was published by Transportation Science
(Psaraftis, 1999). A survey of decision problems that arise in container termi-
nals is provided by Vis and de Koster (2003). The increasing research interest
in OR-based maritime transportation is evidenced by the growing number of
 Ch. 4. Maritime Transportation 195

references in the review papers. The first review paper had almost forty refer-
ences covering several decades. The second one had about the same number of
references most of which were from a single decade, and the most recent one
has almost double that number of references for the last decade. It is worth
mentioning that a large share of the research in transportation planning does
not seem to be based on real cases but rather on artificially generated data. The
opposite is true for maritime transportation, where the majority of problems
discussed are based on real applications.
We focus our attention on planning problems in maritime transportation,
and some related problems. With the fast development of commercial aircraft
during the second half of the 20th century, passenger transportation by ships
has diminished to ferries and cruises. Important as they are, these are small
and specialized segments of maritime transportation. Therefore we shall focus
here on cargo shipping. Related topics that are discussed in other chapters of
this volume are excluded from this chapter, namely maritime transportation
of hazardous materials (Erkut and Verter, 2007) and operations of the land-
side of port terminals (Crainic and Kim, 2007). We try to confine ourselves to
discussion of work that is relatively easily accessible to the reader. This chapter
is intended to provide a comprehensive picture, but by no means an exhaustive
one.
This chapter is organized around the traditional planning levels, strategic,
tactical, and operational planning. Within these planning levels we discuss the
three types of operations in maritime transportation (liner, tramp, industrial)
and additional specialized topics. Although we try to differentiate among the
planning levels, one should remember the interplay among them. On the one
hand, the higher-level or longer-term decisions set the stage for the lower-level
decisions. On the other hand, one usually needs significant amount of details
regarding the shorter-term decisions in order to make good longer-term deci-
sions. We focus here on OR problems in maritime transportation, the related
models, and their solution methods. Due to the fast development of computing
power and memory, information regarding the computing environment be-
comes obsolete very quickly, and such information will only occasionally be
presented.
The rest of the chapter is organized as follows: Section 2 defines terms used
in OR-applications in maritime transportation and describes characteristics of
the industry. Sections 3–5 are dedicated to strategic, tactical, and operational
problems in maritime transportation, respectively. In these sections we present
problem descriptions, models and solution approaches for the three modes
of operations in maritime transportation, namely liner, industrial, and tramp.
We also address in these sections naval operations, maritime supply chains,
ship design and management, ship loading, contract evaluation, booking or-
ders, speed selection, and environmental routing. The issue of robustness in
maritime transportation planning is addressed in Section 6. Important trends
and perspectives for the use of optimization-based decision support systems in
196 M. Christiansen et al.

maritime transportation and suggestions for future research are presented in

Section 7, and some concluding remarks follow in Section 8.

2 Characteristics and terminology of maritime transportation

Maritime transportation planning problems can be classified in the tradi-

tional manner according to the planning horizon into strategic, tactical and
operational problems.
Among the strategic problems we find:
• market and trade selection,
• ship design,
• network and transportation system design (including the determina-
tion of transshipment points for intermodal services),
• fleet size and mix decisions (type, size, and number of vessels), and
• port/terminal location, size, and design.
The tactical problems include:
• adjustments to fleet size and mix,
• fleet deployment (assignment of specific vessels to trade routes),
• ship routing and scheduling,
• inventory ship routing,
• berth scheduling,
• crane scheduling,
• container yard management,
• container stowage planning,
• ship management, and
• distribution of empty containers.
The operational problems involve:
• cruising speed selection,
• ship loading, and
• environmental routing.
Handling of hazardous materials poses additional challenges. However, this
chapter concentrates on the water-side of maritime transportation. Land-side
operations and hazardous materials are discussed in other chapters in this
volume. Before diving into discussion of OR models in maritime transporta-
tion it is worthwhile to take a closer look at the operational characteristics
of maritime transportation and to clarify various terms that are used in this
area. Figure 1 relates the demand for maritime transportation to its supply,
provides a comprehensive view of these characteristics and ties them together
(adapted from Jansson and Shneerson, 1987). The following three sections de-
scribe these characteristics, starting on the supply side.
 Ch. 4. Maritime Transportation 197

Fig. 1. Characteristic of maritime transportation demand and supply.

2.1 Ship and port characteristics

In this chapter we use the terms ship and vessel interchangeably. Although
vessel may refer to other means of transportation, we shall use it in the tradi-
tional sense, referring to a ship.
Ships come in a variety of sizes. The size of a ship is measured by its weight
carrying capacity and by its volume carrying capacity. Cargo with low weight
per unit of volume fills the ship’s volume before it reaches its weight capacity.
Deadweight (DWT) is the weight carrying capacity of a ship, in metric tons.
That includes the weight of the cargo, as well as the weight of fuels, lube oils,
supplies, and anything else on the ship. Gross Tons (GT) is the volume of the
enclosed spaces of the ship in hundreds of cubic feet.
Ships come also in a variety of types. Tankers are designed to carry liquids
in bulk. The larger ones carry crude oil while the smaller ones usually carry
oil products, chemicals, and other liquids. Bulk carriers carry dry bulk com-
modities such as iron ore, coal, grain, bauxite, alumina, phosphate, and other
minerals. Some of the bulk carriers are self-discharging. They carry their own
198 M. Christiansen et al.

unloading equipment, and are not dependent on port equipment for unloading
their cargo. Liquefied Gas Carriers carry refrigerated gas under high pressure.
Container Ships carry standardized metal containers in which packaged goods
are stowed. General Cargo vessels carry in their holds and above deck all types
of goods, usually packaged ones. These vessels often have multiple decks or
floors. Since handling general cargo is labor intensive and time consuming,
general cargo has been containerized during the last four decades, thus reduc-
ing the time that ships carrying such cargo spend in ports from days to hours.
Refrigerated vessels or reefers are designed to carry cargo that requires refrig-
eration or temperature control, like fish, meat, and citrus, but can also carry
general cargo. Roll-on–Roll-off (Ro–Ro) vessels have ramps for trucks and cars
to drive on and off the vessel. Other types of vessels are ferries, passenger ships,
fishing vessels, service/supply vessels, barges (self propelled or pushed/pulled by
tugs), research ships, dredgers, naval vessels, and other, special purpose vessels.
Some ships are designed as combination of the above types, e.g., ore-bulk-oil,
general cargo with refrigerated compartments, passenger and Ro–Ro.
Ships operate between ports. Ports are used for loading and unloading cargo
as well as for loading fuel, fresh water, and supplies, and discharging waste.
Ports impose physical limitations on the dimensions of the ships that may
call in them (ship draft, length and width), and usually charge fees for their
services. Sometimes ports are used for transshipment of cargo among ships,
especially when the cargo is containerized. Major container lines often oper-
ate large vessels between hub ports, and use smaller vessels to feed containers
to/from spoke ports.

2.2 Types of shipping services

There are three basic modes of operation of commercial ships: liner, tramp,
and industrial operations (Lawrence, 1972). Liners operate according to a pub-
lished itinerary and schedule similar to a bus line, and the demand for their
services depends among other things on their schedules. Liner operators usu-
ally control container and general cargo vessels. Tramp ships follow the avail-
able cargoes, similar to a taxicab. Often tramp ships engage in contracts of
affreightment. These are contracts where specified quantities of cargo have to
be carried between specified ports within a specific time frame for an agreed
upon payment per unit of cargo. Tramp operators usually control tankers and
dry bulk carriers. Both liner and tramp operators try to maximize their profits
per time unit. Industrial operators usually own the cargoes shipped and control
the vessels used to ship them. These vessels may be their own or on a time
charter. Industrial operators strive to minimize the cost of shipping their car-
goes. Such operations abound in high volume liquid and dry bulk trades of
vertically integrated companies, such as: oil, chemicals, and ores. When any
type of operator faces insufficient fleet capacity the operator may be able to
charter in additional vessels. Whereas liners and tramp operators may give up
the excess demand and related income, industrial operators must ship all their
 Ch. 4. Maritime Transportation 199

cargoes. In cases of excess fleet capacity, vessels may be chartered out (to other
operators), laid-up or even scrapped. However, when liners reduce their fleet
size they must reshuffle their itineraries and/or schedules, which may result in
reduced service frequency or withdrawal from certain markets. In both cases
revenues may drop. An interesting historical account of the development of
liner services in the US is provided by Fleming (2002, 2003).
Industrial operators, who are usually more risk-averse and tend not to char-
ter-out their vessels, size their fleet below their long-term needs, and comple-
ment it by short-term (time or voyage/spot) charters from the tramp segment.
Seasonal variations in demand, and uncertainties regarding level of future de-
mand, freight rates, and cost of vessels (both newbuildings and second-hand)
affect the fleet size decision. However, when the trade is highly specialized
(e.g., liquefied gas carriers) no tramp market exists and the industrial operator
must assure sufficient shipping capacity through long-term commitments. The
ease of entry into the maritime industry is manifested in the tramp market that
is highly entrepreneurial. This results in long periods of oversupply of shipping
capacity and the associated depressed freight rates and vessel prices. However,
certain market segments, such as container lines, pose large economies of scale
and are hard to enter.
Naval vessels are a different breed. Naval vessels alternate between deploy-
ment at sea and relatively lengthy port periods. The major objective in naval
applications is to maximize a set of measures of effectiveness. Hughes (2002)
provides an interesting personal perspective of naval OR.

2.3 Cargo characteristics

Ships carry a large variety of goods. The goods may be manufactured con-
sumer goods, unprocessed fruits and vegetables, processed food, livestock,
intermediate goods, industrial equipment, processed materials, and raw ma-
terials. These goods may come in a variety of packaging, such as: boxes, bags,
drums, bales, and rolls, or may be unpackaged, or even in bulk. Sometimes car-
goes are unitized into larger standardized units, such as: pallets, containers, or
trailers. Generally, in order to facilitate more efficient cargo handling, goods
that are shipped in larger quantities are shipped in larger handling units or in
bulk. During the last several decades packaged goods that required multiple
manual handlings, and were traditionally shipped by liners, have been con-
tainerized into standard containers. Containerization of such goods facilitates
efficient mechanized handling of the cargo, and thus saves time and money,
and also reduces pilferage. Shipping containers come in two lengths, 20 feet
and 40 feet. A 20 container carries up to approximately 28 tons of cargo with a
volume of up to 1000 cubic feet. Most containers are metal boxes with an 8 ×8
cross-section, but other varieties exist, such as: refrigerated containers, open
top, open side, and half height. In addition there are containers of nonstan-
dard sizes. Large containerships can carry thousands of Twenty feet Equivalent
Units (TEUs), where a 40 container is counted as two TEUs.
200 M. Christiansen et al.

In addition, goods that are shipped in larger quantities are usually shipped
more often and in larger shipment sizes. Cargoes may require shoring on the
ship in order to prevent them from shifting during the passage, and may require
refrigeration, controlled temperature, or special handling while on board the
ship. Different goods may have different weight density, thus a ship may be full
either by weight or by volume, or by another measure of capacity.

2.4 Geographical characteristics

Shipping routes may be classified according to their geographical character-

istics (and the corresponding type and size of vessel used): deep-sea, short-sea,
coastal, and inland waterways. Due to economies of scale in shipping larger size
vessels are employed in deep-sea trades between continents whereas smaller
size vessels usually operate in short-sea and coastal routes, where voyage legs
are relatively short. As mentioned above, smaller containerships are used on
short-sea routes that feed cargo to larger vessels that operate on long deep-sea
routes. A similar picture can sometimes be observed with tankers where large
crude carriers used for long routes are lightered at an off shore terminal to
smaller vessels (often barges). Due to draft restrictions inland waterways are
used mainly by barges. Barges are used to move cargoes between the hinter-
land and coastal areas, often for transshipments to/from ocean-going vessels,
or to move cargoes between inland ports.

2.5 Terms used in maritime transportation planning

• Shipping refers to moving cargoes by ships.

• The shipper is the owner of the transported cargo.
• A shipment is a specified amount of cargo that must be shipped to-
gether from a single origin to a single destination.
• Routing is the assignment of a sequence of ports to a vessel. Environ-
mental routing or weather routing is the determination of the best path
in a body of water that a vessel should follow.
• Scheduling is assigning times (or time windows) to the various events
on a ship’s route.
• Deployment refers to the assignment of the vessels in the fleet to trade
routes. The differentiation between deployment and scheduling is not
always clear cut. Deployment is usually used when vessels are desig-
nated to perform multiple consecutive trips on the same route, and
therefore is associated with liners and a longer planning horizon. Lin-
ers follow a published sailing schedule and face more stable demand.
Scheduling does not imply allocation of vessels to specific trade routes,
but rather to specific shipments, and is associated with tramp and
industrial operations. Due to higher uncertainty regarding future de-
mand in these operations, their schedules usually have a shorter plan-
ning horizon.
 Ch. 4. Maritime Transportation 201

• A voyage consists of a sequence of port calls, starting with the port

where the ship loads its first cargo and ending where the ship unloads
its last cargo and becomes empty again. A voyage may include multiple
loading ports and multiple unloading ports. Liners may not become
empty between consecutive voyages, and in that case a voyage starts at
the port specified by the ship operator (usually a primary loading port).
Throughout this chapter we use also the following definitions:
• A cargo is a set of goods shipped together from a single origin to a
single destination. In the vehicle routing literature it is often referred
to as an order. The terms shipment and cargo are used interchangeably.
• A load is the set of cargoes that is on the ship at any given point in time.
• A load is considered a full shipload when it consists of a single cargo
that for practical and/or contractual reasons cannot be carried with
other cargoes.
• A product is a set of goods that can be stowed together in the same
compartment. In the vehicle routing literature it is sometimes referred
to as a commodity.
• A loading port is a pickup location (corresponds to a pickup node).
• An unloading port is a delivery location (corresponds to a delivery
node).

3 Strategic planning in maritime transportation

Strategic decisions are long-term decisions that set the stage for tactical and
operational ones. In maritime transportation strategic decisions cover a wide
spectrum, from the design of the transportation services to accepting long-term
contracts. Most of the strategic decisions are on the supply side, and these
are: market selection, fleet size and mix, transportation system/service network
design, maritime supply chain/maritime logistic system design, and ship design.
Due to characteristics discussed earlier maritime transportation markets are
usually competitive and highly volatile over time, and that complicates strategic
decisions.
In this section we address the various types of strategic decisions in maritime
transportation and present models for making such decisions. Section 3.1 that
discusses ship design is followed by Section 3.2 that deals with fleet size and mix
decisions. Section 3.3 treats network design in liner shipping, and Section 3.4
handles transportation system design. Finally, Section 3.5 addresses evaluation
of long-term contracts.
In order to be able to make strategic decisions one usually needs some tacti-
cal or even operational information. Thus there is a significant overlap between
strategic and tactical/operational decisions. Models used for fleet size and mix
decisions and network design decisions often require evaluation of ship rout-
ing strategies. Such routing models usually fall into one of two categories, arc
202 M. Christiansen et al.

flow models or path flow models. In arc flow models a binary variable is used to
represent whether a specific vessel v travels directly from port (or customer) i
to port (or customer) j. The model constructs the routes that will be used by
the vessels, and the model has to keep track of both travel time and load on
each vessel. In path flow models the routes are predefined, one way or another,
and a binary variable represents whether vessel v performs route r. A route is
usually a full schedule for the vessel that specifies expected arrival times and
load on the vessel along the route. Such a model can focus on the set of ports
or customers to serve, and only feasible routes are considered.

3.1 Ship design

A ship is basically a floating plant with housing for the crew. Therefore, ship
design covers a large variety of topics that are addressed by naval architects and
marine engineers, and they include structural and stability issues, materials,
on-board mechanical and electrical systems, cargo handling equipment, and
many others. Some of these issues have direct impact on the ship’s commercial
viability, and we shall focus here on two such issues, ship size and speed.
The issue of the optimal size of a ship arises when one tries to determine
what is the best ship for a specific trade. In this section we deal with the opti-
mal size of a single ship regardless of other ships that may be included in the
same fleet. The latter issue, the optimal size and composition of a fleet, is dis-
cussed in Section 3.2. The optimal ship size is the one that minimizes the ship
operator’s cost per ton of cargo on a specific trade route with a specified cargo
mix. However, one should realize that in certain situations factors beyond costs
may dictate the ship size.
Ships are productive and generate income at sea. Port time is a “necessary
evil” for loading and unloading cargo. Significant economies of scale exist at
sea where the cost per cargo ton-mile decreases with increasing the ship size.
These economies stem from the capital costs of the ship (design, construc-
tion, and financing costs), from fuel consumption, and from the operating costs
(crew cost, supplies, insurance, and repairs). However, at port the picture is dif-
ferent. Loading and unloading rates are usually determined by the land-side
cargo handling equipment and available storage space. Depending on the type
of cargo and whether the cargo handling is done by the land-side equipment or
by the equipment on the ship (e.g., pumps, derricks), the cargo handling rate
may be constant (i.e., does not depend on the size of the ship), or, for dry cargo
where multiple cranes can work in parallel, the cargo handling rate may be ap-
proximately proportional to the length of the ship. Since the size of the ship
is determined by its length, width, and draft, and since the proportions among
these three dimensions are practically almost constant, the size of the ship is
approximately proportional to the third power of its length. Therefore, in the
better case, cargo-handling rates will be proportional to the 1/3 power of the
ship size. However, when the cargo is liquid bulk (e.g., oil) the cargo-handling
rate may not be related to the size of the ship.
 Ch. 4. Maritime Transportation 203

A ship represents a large capital investment that translates into a large cost
per day. Port time is expensive and presents diseconomies of scale. Thus the
time of port operations caps the optimal size of ship. Generally, the longer a
trade route is, the larger the share of sea-days in a voyage, and the larger the
optimal ship size will be. Other factors that affect the optimal ship size are the
utilization of ship capacity at sea (the “trade balance”), loading and unloading
rates at the ports, and the various costs associated with the ship. On certain
routes there may be additional considerations that affect the size of the ship,
such as required frequency of service and availability of cargo.
A ship is a long-term investment. The useful life of a ship spans 20–30 years.
Thus, the optimal ship size is a long-term decision that must be based on ex-
pectations regarding future market conditions. During the life of a ship a lot
of market volatility may be encountered. Freight rates may fluctuate over a
wide range, and the same is true for the cost of a ship, whether it is a second
hand one or a newbuilding. When freight rates are depressed they may not
even cover the variable operating costs of the ship, and the owner has very few
alternatives. In the short run the owner may either reduce the daily variable
operating cost of the ship by slow steaming, that results in significant reduction
in fuel consumption, or the owner may lay up the ship till the market improves.
Laying up a ship involves a significant set-up cost to put the ship into lay up,
and, eventually, to bring it back into service. However, laying up a ship signifi-
cantly reduces its daily variable operating cost. When the market is depressed,
owners scrap older ships. The value of a scrapped ship is determined by the
weight of its steel (the “lightweight” of the ship), but when there is high sup-
ply of ships for scrap the price paid per ton of scrap drops. Occasionally, in a
very depressed market, a newly built vessel may find itself in the scrapping yard
without ever carrying any cargo.
In the shorter run ship size may be limited by parameters of the specific
trade, such as availability of cargoes, required frequency of service, physical
limitations of port facilities such as ship draft, length, or width, and available
cargo handling equipment and cargo storage capacity in the ports. In the longer
run many of these limitations can be relaxed if there is an economic justifica-
tion to do so. In addition there are limitations of ship design and construction
technology, as well as channel restrictions in canals in the selected trade routes.
The issue of long-run optimal ship size has been discussed mainly by econo-
mists. Jansson and Shneerson (1982) presented a comprehensive model for the
determination of long-run optimal ship size. They separated the ship capacity
into two components:
• the hauling capacity (the ship size times its speed), and
• the handling capacity (cargo loaded or unloaded per time unit).
This separation facilitated the division of the total shipping costs into cost
per ton of cargo carried in the voyage that does not depend on the length of the
voyage, and cost per time unit. These two cost components are combined into a
cost model that conveys the cost of shipping a ton of cargo a given distance. The
204 M. Christiansen et al.

model requires estimation of output and cost elasticities. These elasticities,

combined with the route characteristics and input prices, allow estimation of
the optimal size of the ship. This model requires estimation of its parameters
through regression analysis. However, high shipping market volatility over time
results in low reliability of such estimates. They demonstrated the use of the
model by calculating the optimal size of a coal bulk carrier for a specific trade.
This work also inspects the sensitivity of the optimal ship size to four route
characteristics: distance, port productivity, trade balance, and fuel costs. Most
of the elasticities that are necessary for this model were estimated from several
datasets in their earlier work (Jansson and Shneerson, 1978). However, that
work calculated a single ship size elasticity of operating costs for each ship
type. In a later study, Talley et al. (1986) analyzed short-run variable costs of
tankers and concluded that the ship size elasticity of operating costs may vary
according to the size of the ship of the specific type.
Modern cargo handling equipment that is customized for the specific cargo
results in higher loading and unloading rates, and shorter port calls. Such
equipment is justified where there is a high volume of cargo. That is usually the
case in major bulk trades. Garrod and Miklius (1985) showed that under such
circumstances the optimal ship size becomes very large, far beyond the capac-
ity of existing port facilities. In addition, with such large ships the frequency of
shipments drops to a point where inventory carrying costs incurred by the ship-
per start playing a significant role (the shipment size is the ship capacity). When
one includes the inventory costs in the determination of the optimal ship size,
that size is reduced significantly. The resulting ship sizes are still much larger
than existing port facilities can accommodate, and thus the main limit on ship
sizes is the draft limitation of ports. However, for a higher value cargo, or for
less efficient port operations, smaller vessel sizes are optimal (see, for example,
Ariel, 1991). In short-sea operations competition with other modes may play a
significant role. In order to compete with other modes of transportation more
frequent service may be necessary. In such cases frequency and speed of ser-
vice combined with cargo availability may be a determining factor in selecting
the ship size.
In liner trades, where there are numerous shippers, multiple ports, and a
wide variety of products shipped, the inclusion of the shippers’ inventory costs
in the determination of the optimal ship size is more complex. Jansson and
Shneerson (1985) presented the initial model for this case. In addition to the
costs incurred by the ship owner/operator they included the costs of inventory
that are incurred by the shipper (including the cost of safety stocks). The size
(and cost) of the safety stocks is a function of the frequency of sailings on the
route, and that frequency is affected by the ship size and the volume of trade.
Numerous assumptions regarding the trade and the costs were necessary, and
the inclusion of the shippers’ costs reduced very much the optimal ship size.
One could argue with the assumptions of the model, but the conclusions make
sense.
 Ch. 4. Maritime Transportation 205

Whereas Jansson and Shneerson (1985) considered a continuous review in-

ventory control system by the shippers, Pope and Talley (1988) looked at the
case of a periodic review system that is more appropriate when using a (sched-
uled) liner service. They found that “   optimal ship size is highly sensitive
to the inventory management model selected, the treatment of stockouts and
safety stocks, and the inventory management cost structure that prevails”, and
concluded that “rather than computing optimal ship size, it may be more ap-
propriate to compute the optimal load size”. As far as liner operations are
concerned we agree with this conclusion. The optimal ship size is a long-term
decision of the ship owner/operator who serves a large number of shippers.
Each shipper may face different circumstances that may change over time,
and therefore should be concerned with the optimal load (shipment) size. The
optimal load size is a short-term decision that may change with the changing
circumstances.
A historical perspective on the development of size, speed, and other char-
acteristics of large container ships is provided by Gillman (1999). Cullinane and
Khanna (1999) present a more recent detailed study of the economies of scale
of large container ships. They take into account the considerable increase in
port productivity, and take a closer look at the time in port. They find smaller
diseconomies of scale (in port) than earlier studies, and show that the opti-
mal size of a container ship continues to increase with improvements in port
productivity. Taking advantage of these economies of scale to reduce shipping
costs per unit while maintaining frequency of service, requires larger volumes
on the trade route. This is one of the major catalysts for industry consolidation.
However, McLellan (1997) injects a dose of reality to the discussion and points
out that there are practical limits to the size of large containerships imposed
by port draft, container handling technology, space availability, and required
investments in port and transportation infrastructure.
Whereas cargo ships come in a large variety of sizes, from under 1000 DWT
up to more than 500,000 DWT, their designed speed varies in a much narrower
range. When one excludes outliers the ratio between the designed speed of a
fast ship and a slow ship is about 2. The designed speed of a ship is a long-
term decision that affects it’s hauling capacity and is part of optimal ship size
considerations. As a general rule the design speed of a ship increases by the
square root of its length. This implies that the design speed is proportional to
the 1/6 power of the size of the ship. This relationship was confirmed statis-
tically by Jansson and Shneerson (1978), and more recently by Cullinane and
Khanna (1999).

3.2 Fleet size and mix

One of the main strategic issues for shipping companies is the design of an
optimal fleet. This deals with both the type of ships to include in the fleet, their
sizes, and the number of ships of each size.
206 M. Christiansen et al.

In order to support decisions concerning the optimal fleet of ships for an

operator, we have to consider the underlying structure of the operational plan-
ning problem. This means that fleet size and mix models very often include
routing decisions. For the various fleet size and mix problem types discussed in
this section we can develop models that are based on the tactical models de-
scribed thoroughly in Section 4.1. The objective of the strategic fleet size and
mix problem is usually to minimize the fixed (setup) costs of the ships used and
the variable operating costs of these ships. In a tactical routing and scheduling
problem one usually minimizes only the operating costs of the ships. However,
the routing decisions made in a strategic model can be later changed during
tactical planning.
In addition, the fleet size and mix decisions have to be based on an esti-
mate of demand for the transportation services. The demand forecast is highly
uncertain, and stochastic techniques for considering the uncertain information
are relevant for solving such strategic planning problems. Issues of robust plan-
ning are discussed in Section 6. In the literature, various demand patterns are
considered where either the size of the cargoes or the frequency of sailing is
specified.
In tramp shipping, contract evaluation and fleet size issues are closely re-
lated. A shipping company has to find the best split between fixed long-term
cargo contracts and spot cargoes. This split should be based on estimation of
future prices and demand. When considering the fleet size and mix these issues
should be included. This topic is further discussed in Section 3.5.
In Section 3.2.1 we describe the fleet planning problem for a homogeneous
fleet where all the vessels are of the same type, size, and cost, while the fleet
size and mix for a heterogeneous fleet is the topic of Section 3.2.2.

3.2.1 Homogeneous fleet size

In this section, we want to focus on a simple industrial fleet size problem
for a fleet consisting of ships of the same type, size, and cost. In the end of the
section some comments regarding other studies are given.
In the fleet size planning problem considered here, a homogeneous fleet of
ships is engaged in transportation of full shipload cargoes from loading ports to
unloading ports. This means that just one cargo is onboard a ship at a time, and
each cargo is transported directly from its loading port to it’s corresponding
unloading port.
All the required ship arrival times at the loading ports are fixed and known.
Further, we also assume that the loading times and sailing times are known,
such that the arrival times at the unloading ports can be easily calculated. The
unloading times and the sailing time from each unloading port to all loading
ports are also known.
The demand is such that all cargoes, given by specified loading and unload-
ing ports, have to be serviced. The ships should be routed from the unloading
ports to the loading ports in a way that minimizes the total cost of their ballast
legs. Since the fleet is homogeneous and all cargoes must be transported, the
 Ch. 4. Maritime Transportation 207

cost of the loaded legs is constant and we can leave it out. In addition, we want
to minimize the number of necessary ships, and we assume that the number of
ships needed dominates the sailing costs.
In the mathematical description of the problem, let N be the set of cargoes
indexed by i and j. Cargo i is represented by a node in the network, and this
node includes one loading port and one unloading port for cargo i. Since we
have full information about activity times, we can determine the feasible cargo
pairs (i j). If cargo i can be serviced just before cargo j by the same ship, such
an (i j)-pair is feasible and represents an arc in the network. However, if the
time between the loads is too long, the arc may be eliminated since using such
arcs would result in unacceptable high waiting times. Similarly, if the departing
time at node i plus the sailing time to j is greater than the given arrival time
at j there will be no arc connecting the two cargoes. Let Ni− and Ni+ be the set
of all cargoes a ship can service immediately before and after servicing cargo i,
respectively. Further, let V be the set of ships in the fleet indexed by v, and
this set includes an assumption on the upper bound on the number of ships
necessary. For each possible ship, we define an artificial origin cargo o(v) and
an artificial destination cargo d(v).
The operational cost of sailing from the unloading port for cargo i to the
loading port of cargo j is denoted by Cij .
In the mathematical formulation, we use the following types of variables: the
binary flow variable xij , i ∈ N , j ∈ Ni+ , equals 1, if a ship services cargo i just
before cargo j, and 0 otherwise. In addition, we define flow variables for the
artificial origin and artificial destination cargoes: xo(v)j , v ∈ V , j ∈ N ∪ {d(v)},
and xid(v) , v ∈ V , i ∈ N ∪{o(v)}. If a ship v is not operating, then xo(v)d(v) = 1.
The arc flow formulation of the industrial ship fleet size problem for one
type of ships and full ship loads is as follows:
   
min Cij xij − xo(v)d(v) (3.1)
i∈N j∈N + v∈V
i

subject to

xo(v)j = 1 ∀v ∈ V  (3.2)
j∈N ∪{d(v)}

xid(v) = 1 ∀v ∈ V  (3.3)
i∈N ∪{o(v)}
 
xij + xid(v) = 1 ∀i ∈ N  (3.4)
j∈Ni+ v∈V
 
xij + xo(v)j = 1 ∀j ∈ N  (3.5)
i∈Nj− v∈V
   
xij ∈ {0 1} ∀v ∈ V  i ∈ N ∪ o(v)  j ∈ Ni+ ∪ d(v)  (3.6)
208 M. Christiansen et al.

In the first term of the objective function (3.1), we minimize the costs of the
ballast legs of the ships. Since xo(v)d(v) = 1 if ship v is not operating, the second
term in the objective function minimizes the number of ships in operation. The
first term is scaled in a manner that its absolute value is less than one. This
means that the objective (3.1) first minimizes the number of ships in use and
then as a second goal minimizes the operating costs of the ships. The second
term in the objective function could easily be incorporated in the first term.
However, the present form of the objective function is chosen to highlight the
twofold objective. Constraints (3.2) ensure that each ship leaves its artificial
origin cargo and either services one of the real cargoes or sails directly to its
artificial destination cargo. In constraints (3.3) each ship in the end of its route
has to arrive at its artificial destination cargo from somewhere. Constraints
(3.4) ensure that the ship that services cargo i has to either service another
cargo afterward or sail to its artificial destination cargo, while constraints (3.5)
say that the ship servicing cargo j has to come from somewhere. Finally, the
formulation involves binary requirements (3.6) on the flow variables.
We can easily see that the formulation (3.1)–(3.6) has the same structure as
an assignment problem. Therefore the integrality constraints (3.6) are not a
complicating factor. The problem is easily solved by any version of the simplex
method or by a special algorithm for the assignment problem.
When applying a simplex method, it would be possible to have just one com-
mon artificial origin, o, and one common artificial destination, d, cargo. Then
xo(v)j , v ∈ V , j ∈ N ∪ {d(v)}, and xid(v) , v ∈ V , i ∈ N ∪ {o(v)}, can be trans-
formed into xoj , j ∈ N ∪ {d}, and xid , i ∈ N ∪ {o}. While the xoj and xid
variables remain binary the variable xod becomes integer.
For some problems, some of the cargoes may have a common loading port
and/or a common unloading port. If the given starting times are such that sev-
eral cargoes are loaded or unloaded in the same port at the same time, we
assume that if this has any effect on the (un)loading times it is already ac-
counted for in the specified data.
In a case with the same starting times in the same ports, we might change the
formulation slightly. Constraints (3.4) can be considered as the constraints for
leaving the unloading port for cargo i, and (3.5) as the constraints for arriving
at the loading port for cargo j. We can then aggregate constraints for cargoes
with the same ports and starting times. This will give more variables at the
left-hand side of the constraints and a right-hand side equal to the number
of aggregated constraints. The corresponding flow variables from and to the
artificial cargoes will become integers rather than binary.
If some of the cargoes have the same loading and unloading ports and the
same starting times then we can switch from indexing the variables by cargo
numbers to indexing them by loading port, unloading port, and both loading
and unloading times. Then the variables can be integer rather than binary, and
their number will be reduced. Dantzig and Fulkerson (1954) pioneered such
a model using a different notation for a problem with naval fuel oil tankers.
 Ch. 4. Maritime Transportation 209

They solved a problem with 20 cargoes by using the transportation model. The
number of ships was minimized and 6 ships were needed.
Later Bellmore (1968) modified the problem. An insufficient number of
tankers and a utility associated with each cargo were assumed. The problem
was to determine the schedules for the fleet that maximized the sum of the
utilities of the carried cargoes, and it was shown to be equivalent to a trans-
shipment problem.
Another homogeneous fleet size problem is considered in Jaikumar and
Solomon (1987). Their objective is to minimize the number of tugs required to
transport a given number of barges between different ports in a river system.
They take advantage of the fact that the service times are negligible compared
with the transit times, and of the geographical structure of the port locations
on the river, and develop a highly effective polynomial exact algorithm. This
problem has a line (or tree) structure, and this fact is exploited in the model
definition.
Recently Sambracos et al. (2004) addressed the fleet size issue for short-
sea freight services. They investigate the introduction of small containers for
coastal freight shipping in the Greek Aegean Sea from two different aspects.
First, a strategic planning model is developed for determining the homoge-
neous fleet size under known supply and demand constraints where total fuel
costs and port dues are minimized. Subsequently, the operational dimension of
the problem is analyzed by introducing a vehicle routing problem formulation
corresponding to the periodic needs for transportation using small containers.
Many simplifying assumptions are made in this study. They conclude that a 5 %
cost saving may be realized by redesigning the inter-island links.

3.2.2 Heterogeneous fleet size and mix

In this section we extend the planning problem discussed in Section 3.2.1
and include decisions about the mix of different ship sizes.
We study here one particular fleet size and mix problem, where a liner
shipping company wants to serve several customers that have a demand for
frequent service. The problem consists of determining the best mix of ships to
serve known frequencies of demand between several origin–destination port
pairs. Many feasible routes are predefined, and just some of them will be used
in the optimal solution. The demand is given as a minimum required number
of times each port pair has to be serviced. The underlying real problem is a
pickup and delivery problem. However, with predefined routes in the model,
the loading and unloading aspects are not visible but hidden in the routes.
Since this is a pickup and delivery problem, the frequency demand applies to
a pair of ports. The ships are heterogeneous so not all ships can sail all routes.
The capacity of a ship determines, among other factors, which routes it can
sail. A ship is allowed to split its time between several routes.
The planning problem consists of deciding: (1) which ships to operate and
(2) which routes each ship should sail and the number of voyages along each
route. The first part is a strategic fleet mix and size problem and the second
210 M. Christiansen et al.

part is a tactical fleet deployment problem. Fleet deployment problems are

discussed in Section 4.4. The second part is used here only to find the best
solution to the first part. If the demand pattern changes later, the second part
can be resolved for the then available fleet.
In the mathematical description of the problem, let V be the set of ships
indexed by v and Rv the set of routes that can be sailed by ship v indexed by r.
The set of origin–destination port pairs is called N indexed by i, and each such
pair needs to be serviced at least Di times during the planning horizon.
The cost consists of two parts. We define the cost of sailing one voyage with
ship v on route r as CVvr . The fixed cost for ship v during the planning horizon
is called CFv . Each voyage with ship v on route r takes TVvr time units, and Air
is equal to 1 if origin–destination port pair i is serviced on route r. The length
of the planning horizon is T , and we assume that the ships are available for the
whole horizon. Let Uv be an upper bound on the number of voyages ship v can
sail during the planning horizon.
Here we use the following types of decision variables: uvr , v ∈ V , r ∈ Rv ,
represents the number of voyages along route r with ship v during the planning
horizon, and sv , v ∈ V , is equal to 1 if ship v is used.
The model for the strategic fleet size and mix problem with predefined
routes can then be written as
   
min CVvr uvr + CFv sv (3.7)
v∈V r∈Rv v∈V
subject to

uvr − Uv sv  0 ∀v ∈ V  (3.8)
r∈Rv

Air uvr  Di  ∀i ∈ N  (3.9)
v∈V r∈Rv

TVvr uvr  T ∀v ∈ V  (3.10)
r∈Rv
uvr  0 and integer ∀v ∈ V  r ∈ Rv  (3.11)
sv ∈ {0 1} ∀v ∈ V  (3.12)
Here (3.7) is the cost of sailing the used routes together with the fixed cost of
the ships in operation. Constraints (3.8) ensure that the fixed costs for the ships
in operation are taken into account. Constraints (3.9) say that each port pair
is serviced at least the required number of times, and constraints (3.10) ensure
that each ship finishes all its routes within the planning horizon. Finally, the
formulation involves integer and binary requirements on the variables.
Fagerholt and Lindstad (2000) presented this model with different notation
and gave an example where the model was used to plan deliveries to Norwe-
gian petroleum installations in the North Sea. Their problem had one loading
port and seven unloading installations. They managed to pre-calculate all the
 Ch. 4. Maritime Transportation 211

feasible routes and their integer program was solved by CPLEX. The model
does not ensure that services for a given port pair are properly spaced during
the planning horizon. This aspect was treated manually after the model so-
lutions were generated. Fagerholt and Lindstad (2000) report that the model
solution implemented gave annual savings of several million US dollars.
Another study regarding fleet size and mix for liner routes was done by Cho
and Perakis (1996). The study was performed for a container shipping com-
pany. The type of model and solution method is similar to the one used by
Fagerholt and Lindstad (2000). Xinlian et al. (2000) consider a similar prob-
lem. They present a long-term fleet planning model that aims at determining
which ships should be added to the existing fleet, ship retirements, and the op-
timal fleet deployment plan. Another study regarding the design of an optimal
fleet and the corresponding weekly routes for each ship for a liner shipping
system along the Norwegian coast was presented by Fagerholt (1999). The so-
lution method is similar to the one used by Fagerholt and Lindstad (2000). In
Fagerholt (1999) the solution method handled only instances where the dif-
ferent ships that could be selected have the same speed. This is in contrast to
the work in Fagerholt and Lindstad (2000), where the ships can have different
speeds. Yet another contribution within fleet size and mix for liner shipping is
given by Lane et al. (1987). They consider the problem of deciding a cost effi-
cient fleet that meets a known demand for shipping services on a defined liner
trade route. The solution method has some similarities to the approach used
by Fagerholt and Lindstad (2000), but the method gives no proven optimal so-
lution since only a subset of the feasible voyage options are selected and the
user determines the combination of vessel and voyage. The method has been
applied on the Australia/US West coast route. Finally, resource management
for a container vessel fleet is studied by Pesenti (1995). This problem involves
decisions on the purchase and use of ships in order to satisfy customers’ de-
mand. A hierarchical model for the problem has been developed, and heuristic
techniques, which solve problems at different decision levels, are described.
A rather special problem regarding the size of the US destroyer fleet is de-
scribed in Crary et al. (2002), which illustrates the use of quantitative methods
in conjunction with expert opinion. These ideas are applied to the planning
scenario for the “2015 conflict on the Korean Peninsula”, one of two key sce-
narios the Department of Defense uses for planning.

3.3 Liner network design

On all three planning levels the challenges in liner shipping are quite differ-
ent from those of tramp or industrial. Liner ships are employed on more or less
fixed routes, calling regularly at many ports. In contrast to industrial or tramp
ships a liner ship serves demand of many shippers simultaneously, and its pub-
lished route and frequency of service attract demand. The major challenges
for liners at the strategic level are the design of liner routes and the associated
frequency of service, fleet size and mix decisions and contract evaluation for
212 M. Christiansen et al.

long-term contracts. The fleet size and mix decisions for the major market seg-
ments, including liner operations, are discussed in Section 3.2, while contract
evaluation will be treated in Section 3.5. Here we focus on the design of liner
routes. We split this section into three parts, where traditional liner operations
are discussed in Section 3.3.1, and the more complex hub and spoke networks
are considered in Section 3.3.2. Finally, we comment upon shuttle services in
Section 3.3.3.

3.3.1 Traditional liner operations

Liner routes and schedules are usually set up in a manner similar to bus
schedules. Before entering a particular market a liner shipping company has
to thoroughly estimate the demand, revenue and cost of servicing that market.
Based on this information, the company has to design its routes and to publish
a sailing schedule.
Most liner companies are transporting containers, so we use here the term
container(s) instead of cargo units or cargoes. We focus here on a problem
where a liner container company is going to operate several different routes
among a set of ports ordered more or less along a straight line. Meaning that
even if a route skips a port in a contiguous sequence of ports the ship passes
fairly close to the skipped port. This is usually the situation faced by longer
container lines. The demands, as upper bounds on the number of transported
containers, are given between all pairs of ports. The fleet of ships is heteroge-
neous and the planning problem consists of designing a route for each ship in
a manner that maximizes the total net revenue of the fleet. One route is con-
structed for each ship and the ship sails as many voyages along that route as it
can during the planning horizon.
The mathematical model is based on an arc flow formulation. The ports
are numbered from 1 to N, and there are some strict constraints on how the
routes can be constructed. Each route must have two end ports i and j, where
1  i < j  N. A route then starts in i and travels outbound to ports with
higher and higher number until the route reaches j, where it turns around
and starts its inbound travel to ports with lower and lower number until the
route ends in i. A ship with i and j as end ports, does not necessarily call at
all the ports between i and j, and it does not need to visit the same ports on
the outbound and inbound legs of the route. See Figure 2 for an illustration of
such routes.
When a ship arrives at one of its end ports it unloads all containers that are
on board before it starts loading all the containers that it should load in that
port. This means that each container is loaded in its loading port and stays on
board the ship while the ship either sails a part of the outbound or inbound
route before it is unloaded in its unloading port.
In the mathematical description of the problem, let V be the set of ships
indexed by v and N the set of linearly ordered ports indexed by i, j, k, i ,
or j  . In addition we need the subsets Ni+ = {i + 1     N} of ports in the
 Ch. 4. Maritime Transportation 213

Fig. 2. Liner network design for traditional liner operations including some but not all routes.

line numbered after i and Ni− = {1     i − 1} of ports in the line numbered
before i.
The revenue for transporting one container from port i to port j is RTij and
the cost of sailing directly from port i to port j with ship v is CTijv . Ship v has
a capacity that is measured in number of containers when it sails directly from
port i to port j, and it is represented by QTijv . Most often it will be sufficient
not to let capacity depend on the sailing leg (i j), but in rare cases capacity
may depend on weather conditions or other factors. The ship spends TTijv time
units on that trip including the time for unloading and loading in port i. It is
meaningful to assume that this time does not vary with the number of contain-
ers loaded and unloaded only if the number of such containers does not vary
from call to call or that the unloading and loading time is very short compared
to the sailing time. The demand as an upper bound on the number of contain-
ers transported from port i to port j during the planning horizon is denoted
by DTij . The constant Sv is the maximum time ship v is available during the
planning period.
We use the following types of decision variables: eijv , v ∈ V , i ∈ N , j ∈ N ,
represents the number of containers transported from port i to port j by ship v
on each voyage during the planning horizon. Ship v does not necessarily sail
directly from port i to port j. If ship v sails directly from port i to port j on its
route, then the binary variable xijv , v ∈ V , i ∈ N , j ∈ N , is equal to 1. The
integer variable wv , v ∈ V , gives the number of whole voyages ship v manages
to complete during the planning horizon. The binary variable yijv , ∀v ∈ V ,
i ∈ N \{N}, j ∈ Ni+ , is equal to 1 if ship v is allocated to a route that starts in
port i and turns around in port j. These two ports i and j are called end ports
for ship v.
A route design model for traditional liner operators can then be written as
 
max wv (RTij eijv − CTijv xijv ) (3.13)
v∈V i∈N j∈N
subject to
   
xijv ei j  v − QTijv  0
− +
i ∈Ni+1 j  ∈Nj−1

∀v ∈ V  i ∈ N \{N} j ∈ Ni+  (3.14)

214 M. Christiansen et al.
   
xijv e
i j  v − QTijv  0
+ −
i ∈Ni−1 j  ∈Nj+1

∀v ∈ V  i ∈ N \{1} j ∈ Ni−  (3.15)


wv eijv  DTij xij  v 
j  ∈Ni+ \Nj+

∀v ∈ V  i ∈ N \{N} j ∈ Ni+  (3.16)


wv eijv  DTij xij  v 
j  ∈Ni− \Nj−

∀v ∈ V  i ∈ N \{1} j ∈ Ni−  (3.17)


wv eijv  DTij xi jv 
i ∈Nj− \Ni−

∀v ∈ V  i ∈ N \{N} j ∈ Ni+  (3.18)


wv eijv  DTij xi jv 
i ∈Nj+ \Ni+

∀v ∈ V  i ∈ N \{1} j ∈ Ni−  (3.19)


wv eijv  DTij  ∀i ∈ N  j ∈ N  i = j (3.20)
v∈V
  
wv TTijv xijv  Sv  ∀v ∈ V  (3.21)
i∈N j∈N
 
yijv  1 ∀v ∈ V  (3.22)
i∈N \{N} j∈N +
i
  
yijv xij  v − 1 = 0
j  ∈Ni+ \Nj+

∀v ∈ V  i ∈ N \{N} j ∈ Ni+  (3.23)

  
yijv xj  iv − 1 = 0
j  ∈Ni+ \Nj+

∀v ∈ V  i ∈ N \{N} j ∈ Ni+  (3.24)

   
yijv xi kv − xkj  v = 0
i ∈Nk− \Ni− j  ∈Nk+ \Nj+

∀v ∈ V  i ∈ N \{N} j ∈ Ni+  k ∈ Ni+ \Nj−1

+
 (3.25)
 Ch. 4. Maritime Transportation 215
   
yijv x
i kv − x
kj  v = 0
i ∈Nk+ \Nj+ j  ∈Nk− \Ni−

∀v ∈ V  i ∈ N \{N} j ∈ Ni+  k ∈ Ni+ \Nj−1

+
 (3.26)
xijv ∈ {0 1} ∀v ∈ V  i ∈ N  j ∈ N  i = j (3.27)
eijv  0 ∀v ∈ V  i ∈ N  j ∈ N  i = j (3.28)
wv  0 and integer ∀v ∈ V  (3.29)
yijv ∈ {0 1} ∀v ∈ V  i ∈ N \{N} j ∈ Ni+  (3.30)
The objective function (3.13) maximizes the difference between the revenue
from transporting containers and the cost of operating the ships. The capacity
of the ship might vary from leg to leg of the voyage, and (3.14) and (3.15)
represent the capacity constraints for the possible outbound and inbound legs.
To be able to transport containers from port i to port j on ship v, the ship needs
to depart from i, either directly to j or to a port between them. In addition the
ship needs to arrive in j either directly from i or from a port between them.
The four constraints, (3.16)–(3.19), express these issues. The constraints for the
outbound and inbound parts of the voyage had to be given separately. Each of
these constraints ensures that if none of the binary flow variables, xi jv or xij  v ,
is equal to 1, the number of containers transported by ship v from port i to
port j during the planning horizon is zero. When the binary flow variables are
equal to 1, the corresponding constraint is redundant. The demands as upper
bounds on the number of transported containers are expressed in (3.20), and
the upper bound on the number of voyages for each ship is expressed in (3.21).
The connectivity of each route is expressed by (3.22)–(3.26). Constraints (3.22)
ensure that each ship can have only one pair of end ports (one starting port i
and one turning port j). A ship that starts in port i and turns around in port j,
needs to leave i for a port not farther away than j and it needs to arrive in i
from a port not farther away than j. This is expressed in (3.23) and (3.24). For
each port, k, numbered between i and j, the same ship must arrive in k the
same number of times, 0 or 1, as the number of times it departs from k, both
on the outbound part and on the inbound part of the route. This is taken care
of by (3.25) and (3.26). The turning around in port j is taken care of by the
fact that if port k is the last port ship v visits before it reaches port j, then
constraints (3.25) say that the ship has to travel directly from port k to port j.
And if port k is the first port ship v visits on the inbound part of its voyage after
leaving port j, then constraints (3.26) say that the ship has to travel directly
from port j to port k .
Rana and Vickson (1988) presented a model for routing of one ship. Later
(Rana and Vickson, 1991) they enhanced the model to a fleet of ships, and this
latter model is the same as the one presented here with a different notation,
and with constraints (3.14) and (3.15) written linearly. The solution method
used by Rana and Vickson can be summarized as follows. They started with
reducing the nonlinearities in the model. If we look carefully at constraints
216 M. Christiansen et al.

(3.14)–(3.26) we see that constraints (3.20) are the only type of constraints that
is summed over v. All the other constraints are written separately for each
ship. The authors exploited this fact to apply Lagrangian relaxation to con-
straints (3.20). Then the problem decomposes into one problem for each ship.
However, they needed to iterate or optimize over the Lagrangian multipliers.
In solving the problem for each ship they solved it for different fixed values for
the number of voyages. In this way, they got mixed linear integer subproblems,
which they solved to near optimality by using Bender’s decomposition. They
give results for problems with 3 ships and between 5 and 20 ports. On average
their solutions are about 2% from the upper bounds.
All the nonlinearities in (3.13)–(3.26) consist of products of two variables or
one variable and a linear expression in other variables. Apart from the terms
with wv eijv , all the nonlinear terms consist of products where at least one vari-
able is binary. So by first expressing wv by binary variables, we can remove the
product terms by defining one new variable and three new constraints for each
product term as described by Williams (1999) in Chapter 9.2. We might then,
over a decade after the publication of that paper (Rana and Vickson, 1991),
be able to solve small instances of the underlying problem by using standard
commercial software for mixed integer programming.
A rather special liner shipping problem is described by Hersh and Ladany
(1989). However, the structure of the problem has some similarities to the
problem described here. A company leasing a luxury ocean liner for Christ-
mas cruises from Southern Florida is confronted with the problem of deciding
upon the type of cruises to offer. The decision variables in the problem include
the routing of the ship, the duration of the cruises, the departure dates, and
the fare schedules of the cruises.

3.3.2 Hub and spoke networks

Containers are usually both faster and cheaper to load and unload than
the general cargo that is stuffed in them. This means that containers can ef-
ficiently be loaded and unloaded several times between their origin and their
final destination. One type of maritime transportation systems for containers
is the so-called hub and spoke network or a trunk line and feeder system. In such
systems we have a trunk line operating between the major ports (hubs) and
a system of feeder ships working in the geographical region around each hub
port visited by the trunk line. The ports feeding containers to a hub are the
spokes. Thus, a container is typically loaded and unloaded three times. First
a feeder ship transports the container from its initial loading port to a trunk
line hub port. Then a trunk line ship transports the container to another trunk
line hub port, and finally another feeder ship takes the container to its final
unloading port. Such networks are further described in the chapter by Crainic
and Kim (2007) on intermodal transportation in this handbook.
Here we study a short-sea application of a feeder system around one trunk
line hub port with a homogeneous fleet of feeder ships. We model the trans-
portation of containers between one hub port and a set of feeder ports (spokes)
 Ch. 4. Maritime Transportation 217

in one geographical region. Each container is either loaded or unloaded in the

hub.
The demands both to and from a spoke port are assumed to increase with
the number of visits in the port during the planning horizon. These demands
are upper bounds on the number of containers available for transportation, but
the shipping company is not obliged to satisfy the total demand.
The planning problem consists of choosing which of a possible huge set of
predefined routes to use and how many voyages to sail along the chosen routes,
while maximizing the net revenue. Figure 3 illustrates the problem with one
hub and several spokes. The designed routes might be overlapping.
In the mathematical description of the problem, let R be the set of pre-
defined routes indexed by r and N be the set of ports, excluding the hub,
indexed by i. Further, let Nr be the set of ports, excluding the hub, visited
on route r. The routes that visit port i are given by the set Ri . The ports called
after port i on route r belong to the set Nir+ and the ports called before and
including port i on route r belong to the set Nir− . Let M be the set of possible
calls at the same port during the planning horizon indexed by m.
We assume that there are fixed revenues, RLi and RUi , for carrying one
container to and from port i. The cost consists of three parts. We call the fixed
cost of operating a ship during the planning horizon CF . The cost of sailing one
voyage along route r is CVr and the cost of unloading (loading) one container
in port i on route r is CUir (CLir ). Since the fleet is homogeneous and the unit
costs are specified before we know the loading pattern along the routes, we will
normally have CUir and CLir independent of r. The time each ship is available
during the planning horizon is called the shipping season S. The sailing time
for one voyage along route r is TVr and the capacity measured in number of
containers of a ship is Q. The demand is specified in the following way: DUim

Fig. 3. Liner network design for a hub and spoke system. Example of three overlapping routes.
218 M. Christiansen et al.

(DLim ) is the incremental demand for unloading (loading) in port i when the
number of calls at that port increases from m − 1 to m.
In the mathematical formulation, we use the following types of variables: the
integer variable s represents the number of ships in operation and ur , r ∈ R,
represents the number of voyages along route r during the planning horizon.
The number of containers unloaded and loaded in port i on route r during the
planning horizon is given by qUir and qLir , r ∈ R, i ∈ Nr , respectively. The
integer number of calls at port i is hi , i ∈ N , and finally, the binary variable
gim , i ∈ N , m ∈ M, is equal to 1 if port i is called at least m times during the
planning horizon.
A liner network design model for a network with one hub and several spokes
is as follows:
   
max (RUi − CUir )qUir
r∈R i∈Nr
    
+ (RLi − CLir )qLir − CF s − CVr ur (3.31)
r∈R i∈Nr r∈R
subject to

TVr ur − Ss  0 (3.32)
r∈R

qUir − Qur  0 ∀r ∈ R (3.33)
i∈Nr
 
qLjr + qUjr − Qur  0 ∀r ∈ R i ∈ Nr  (3.34)
j∈Nir− j∈Nir+

ur − hi = 0 ∀i ∈ N  (3.35)
r∈Ri

gim − hi = 0 ∀i ∈ N  (3.36)
m∈M
gi(m−1) − gim  0 ∀i ∈ N  m ∈ M (3.37)
 
qUir − DUim gim  0 ∀i ∈ N  (3.38)
r∈Ri m∈M
 
qLir − DLim gim  0 ∀i ∈ N  (3.39)
r∈Ri m∈M
qUir  qLir  0 ∀r ∈ R i ∈ Nr  (3.40)
hi  s ur  0 and integer ∀r ∈ R i ∈ N  (3.41)
gim ∈ {0 1} ∀i ∈ N  m ∈ M (3.42)
The objective function (3.31) maximizes the net revenue over the planning
horizon. We calculate the number of needed ships in (3.32) in a way that might
 Ch. 4. Maritime Transportation 219

be too simple. The constraints ensure that the total available sailing time for
the total fleet of ships is larger than the sum of the voyages’ times. We have
not verified that the available time of the ships can be split in such a manner
that each ship can perform an integer number of voyages during the planning
horizon. Constraints (3.33) and (3.34) take care of the capacity when the ships
leave the hub and the spokes on the route. Constraints (3.35) and (3.36) use
the number of voyages along the routes to calculate the number of calls at each
port. The precedence constraints (3.37) for the gim variables are not needed
if the incremental increase in the demand diminishes with increasing num-
ber of calls. The numbers of containers unloaded and loaded in the ports are
bounded by the demand constraints (3.38) and (3.39). Finally, the formula-
tion involves binary, integer and nonnegativity requirements on the variables
in (3.40)–(3.42).
Bendall and Stent (2001) presented this model using a different notation
and equal costs for loading and unloading containers. Their paper does not
provide any information regarding how the model is solved. From the size of
their practical example and the lack of information about the solution method,
we conclude that they used some standard software for integer programming.
After solving the stated model, they use heuristic methods to find a schedule
for each ship. They report results for an application with Singapore as the hub
and 6 spokes in East-Asia. The routes are different from the impression that
the mathematical model gives, because they had 6 single spoke routes, one for
each spoke and 2 routes with 2 spokes each. The demand data was for one
week and it was assumed that the transportation pattern would be replicated
for many weeks.
If we cannot guarantee that the incremental demand diminishes with in-
creasing number of visits, then (3.35)–(3.39) can be reformulated in the fol-
lowing way. Some of the symbols will be redefined to avoid defining too many
new ones. Now, let DUim (DLim ) be the unloading (loading) demand in port i
when the number of calls in port i is m, and gim is equal to 1 if port i is called
exactly m times during the planning horizon.
These changes result in the following new or revised constraints:
 
mgim − ur = 0 ∀i ∈ N  (3.43)
m∈M r∈Ri

gim = 1 ∀i ∈ N  (3.44)
m∈M
 
qUir − DUim gim  0 ∀i ∈ N  (3.45)
r∈Ri m∈M
 
qLir − DLim gim  0 ∀i ∈ N  (3.46)
r∈Ri m∈M

Here (3.43) has replaced (3.35) and (3.36) and (3.44) is used instead of
(3.37). After changing the meaning of the symbols, the last two constraints
220 M. Christiansen et al.

above, (3.45) and (3.46), are unchanged from the original formulation. This
reformulation might be useful when branching on gim for one value of i and all
values of m as one entity. Some solvers include this possibility, and this set of
variables is then defined as a special ordered set of type one (SOS1 or S1). For
a definition of such sets, see Chapter 9.3 in Williams (1999). For such sets some
solvers will do binary branching by setting some of the variables equal to zero
in one branch and setting the other variables equal to zero in the other branch.
Such branching often results in a more evenly balanced branching tree. This in
turn usually results in fewer branches to investigate.

3.3.3 Shuttle services

Ferries are often used to provide a shuttle service between a pair of ports.
The ferries are often custom built to serve a particular route, fitting comfort-
ably into available berths. Ferries may carry passengers, and usually can carry
cars or trucks that are driven on and off board. Larger ferries that are designed
to carry trucks or cars are called roll-on roll-off vessels. Very little research
has been devoted to this area. A simulation model for ferry traffic among the
Aegean Islands is described by Darzentas and Spyrou (1996). The model is
used for decision support on a “what if” basis for regional development. By
using the simulation model, they were able to evaluate the appropriateness of
existing ferry routes, as well as new transportation scenarios, including the use
of new technology vessels and changes in port capacities.

3.4 Design of maritime transport systems

In a maritime transport system, sea transport constitutes at least one vital

link. An important strategic planning issue is the design of such systems. In
the literature such systems are also referred to as maritime logistics systems or
maritime supply chains. Reported research in the literature on such systems is
scarce. We shall briefly discuss here one optimization-based application and a
couple of simulation studies.
A real strategic and tactical industrial ocean-cargo shipping problem was
studied by Mehrez et al. (1995). The problem involves the shipping of dry bulk
products from a source port to transshipment ports, and then distribution of
the products from the transshipment ports to the customers over land. The
decisions made include the number and size of ships to charter in each time
period during the planning horizon, the number and location of transshipment
ports to use, and transportation routes from the transshipment ports to the
customers. The problem is modeled and solved using a MIP model. Recom-
mendations from this study were implemented by the client company.
Richetta and Larson (1997) present a problem regarding the design of New
York City’s refuse marine transport system. Waste trucks unload their cargo at
land-based stations where refuse is placed into barges that are towed by tug-
boats to the Fresh Kills Landfill on Staten Island. They developed a discrete
 Ch. 4. Maritime Transportation 221

event simulation model incorporating a complex dispatching module for deci-

sion support in fleet sizing and operational planning. This work is an extension
of an earlier study by Larson (1988).
Another simulation study regarding maritime supply chain design can be
found in Fagerholt and Rygh (2002). There, the problem is to design a seaborne
system for transporting freshwater from Turkey to Jordan. The fresh water was
to be transported by sea from Turkey to discharging buoy(s) off the coast of Is-
rael, then in pipeline(s) to a tank terminal ashore and finally through a pipeline
from Israel to Jordan. The study aimed at answering questions regarding the
required number, capacity and speed of vessels, capacity and number of dis-
charging buoys and pipelines, and the necessary capacity of the tank terminal.
Sigurd et al. (2005) discuss a problem where a group of companies, that need
transport between locations on the Norwegian coastline and between Norway
and The European Union, is focusing on reducing costs and decreasing trans-
port lead-time by combining their shipments on the same ships. The companies
need to analyze if there is a realistic possibility to switch some of their demand
for transportation from road to sea. New transport solutions would need faster
ships in order to substantially decrease the existing travel time. The underlying
planning problem consists of finding recurring liner routes. These routes need
to fit both with the quantity and frequency demanded by the companies.

3.5 Contract evaluation

This section discusses another important strategic problem faced by most

shipping companies, namely contract evaluation. This problem is to some ex-
tent related to the fleet size and composition issue, and it consists of deciding
whether to accept a specified long-term contract or not. The characteristics of
this problem differ between tramp and liner operations, and this problem is of
little relevance in an industrial operation.
For a tramp shipping company the problem is to decide whether to accept
a Contract of Affreightment (a contract to carry specified quantities of cargo
between specified ports within a specific time frame for an agreed payment per
ton). In this case, the shipping company has to evaluate whether it has sufficient
fleet tonnage to fulfill the contract commitments together with its existing com-
mitments, and if so, whether the contract is profitable. To check if a contract
will be profitable one also has to make assumptions about how the future spot
market will develop for the given contract period. Typically, if a shipping com-
pany anticipates low spot rates, it will prefer to have as large contract coverage
as possible or ‘go short of tonnage’ and vice versa. The authors are not aware
of any published work in this area.
In the liner shipping industry these problems look slightly different. It is com-
mon that shippers buy a certain capacity for a given trade route. For instance
in container freight transportation, which constitutes most of the liner shipping
trade, it is not unusual that some of the bigger ocean carriers do between 80%
and 95% of their business under such contracts. Most contracts between ocean
222 M. Christiansen et al.

carriers and shippers are negotiated once a year, typically one or two months
before the peak season of the major trade covered by the contract. A key pa-
rameter of a contract is the set of prices for the different cargoes between
any pair of ports. The United States Ocean Shipping Reform Act of 1998 for
the first time allows ocean carriers moving freight into and out of the US to
enter into confidential contracts with shippers, and to charge different ship-
pers different prices. This makes the problem of how to structure these prices
relevant. This problem has many similarities with yield management in the air-
line industry. Kleywegt (2003) presents a model that can be used to support
such decisions before and during contract negotiations. A somewhat similar
problem can be found for cruise lines. Ladany and Arbel (1991) present four
models for determining the optimal price differentiation strategy that a cruise
liner should follow in order to maximize its profit for four different situations.
A price differentiation strategy means that customers belonging to different
market segments would pay different prices for identical cabins. Also this prob-
lem is similar to yield management in airlines.

4 Tactical planning in maritime transportation

At the tactical planning level we concentrate on medium-term decisions, and

the focus of this level in maritime transportation is on routing and scheduling.
Therefore, most of this section is devoted to these planning issues. We start
this section by presenting some classical industrial and tramp ship schedul-
ing problems and give arc flow formulations of these problems in Section 4.1.
Then in Section 4.2 we discuss frequently used solution methods for solving
ship routing and scheduling problems. Throughout the presentation of prob-
lems, formulations and solution approaches we refer to important research
in industrial and tramp ship scheduling, as we deem appropriate. In Section
4.3 we present several tactical planning problems and applications in maritime
supply chains, where sea transport constitutes at least one vital part of the
supply chain. Fleet deployment in liner shipping is presented and discussed in
Section 4.4, whereas barge scheduling on inland waterways is presented in Sec-
tion 4.5. Section 4.6 is dedicated to naval vessel scheduling, while in Section 4.7
we briefly discuss ship management.

4.1 Scheduling problems for industrial and tramp shipping

As described in Section 2, in industrial shipping the cargo owner or shipper

controls the ships. Industrial operators try to ship all their cargoes at minimum
cost. Tramp ships follow the available cargoes like a taxi. A tramp shipping
company may have a certain amount of contract cargoes that it is committed
to carry, and tries to maximize the profit from optional cargoes. From an OR
point of view the structure of the planning challenges for these two modes of
operation is very similar regarding the underlying mathematical models and
 Ch. 4. Maritime Transportation 223

solution approaches. Therefore we treat these modes of operations together

in this section. During the last decades there has been a shift from industrial
to tramp shipping (see Christiansen et al., 2004 and Section 7). In Section 7
we discuss some reasons for the shift from industrial to tramp shipping. Per-
haps the main reason is that many cargo owners are now focusing on their core
business and have outsourced other activities like transportation to indepen-
dent shipping companies. From the shipper’s perspective, this outsourcing has
resulted in reduced risk. Most contributions in the OR literature are for indus-
trial shipping, while only a few are in the tramp sector. The main reason for the
minimal attention to tramp scheduling in the literature may be that historically
the tramp market was operated by a large number of small operators, even
though this is not the case anymore.
In this section we present classes of real ship routing and scheduling prob-
lems. We start with the simplest type of problems in Section 4.1.1 dealing with
routing and scheduling of full shiploads. Here just one cargo is onboard the
ship at a time. We extend this problem to multiple cargoes onboard at the
same time, where each of the cargoes has a fixed size. This problem is ad-
dressed in Section 4.1.2. We continue in Section 4.1.3 with similar problems
but where flexible cargo sizes are allowed. In Section 4.1.4 we present routing
and scheduling problems where multiple nonmixable products can be carried
simultaneously, and the ship capacity is split into separate compartments. Typ-
ical tramp shipping characteristics concerning contracted and optional cargoes
are considered in Section 4.1.5. Finally, we discuss the use of spot charters in
Section 4.1.6.
In practice, at the beginning of the planning horizon the ships in the fleet
may be occupied with prior tasks. For all the classes of problems described in
this section we find the first point in time where the ship is available for loading
a new cargo during the planning horizon, and we assume that at that time the
ship is empty.

4.1.1 Full shiploads

In some market segments, the ship is loaded to its capacity in a loading port
and the cargo is transported directly to its unloading port. A typical example is
the transportation of crude oil.
The objective of an industrial ship scheduling problem for full shipload car-
goes is to minimize the sum of the costs for all the ships in the fleet while
ensuring that all cargoes are lifted from their loading ports to their correspond-
ing unloading ports. Time windows are usually imposed for both loading and
unloading the cargoes.
In such an operation, an industrial shipping company usually operates a het-
erogeneous fleet of ships with specific ship characteristics including different
cost structures and load capacities. In the short-term, it is impractical to change
the fleet size. Therefore, we are concerned with the operations of a given num-
ber of ships within the planning horizon. The fixed cost of the fleet can be
disregarded as it has no influence on the planning of optimal routes and sched-
224 M. Christiansen et al.

ules. We consider the case where the fleet has sufficient capacity to serve all
committed cargoes during the planning horizon. The ships are charged port
and channel tolls when visiting ports and passing channels, and these costs de-
pend on the size of the ship. The remaining variable sailing costs consist mainly
of fuel and oil costs, and depend usually on the ship size.
The quantity of a particular cargo is given and the corresponding loading
and unloading port of that cargo are known, so the time from arrival at the
loading port until the time of departure from the unloading port can be easily
calculated.
In the case where a ship can carry only one cargo at a time but the ship is
not necessary filled up each time, the underlying planning problem is identical
to the problem of full shiploads.

Example 4.1. Consider the following simplified example of a route from a so-
lution to a full shipload planning problem. In this planning problem several
ships are going to service a set of cargoes. In the optimal solution, one ship
is going to lift cargoes 1, 2, and 3. In Table 5, information about the loading
and unloading ports is given for each of the cargoes. In addition, we specify
the quantity of each of the cargoes. Notice that not all cargo sizes are equal to
the capacity of the ship. Two of the cargoes have a quantity equal to half the
capacity of the ship. In reality, the utilization of the ship is too low, but this
case is a basis for another problem presented later on in this section. For the
sake of simplicity, the time windows information is omitted in this example.
The geographical picture of the ports is given in Figure 4(a), while the phys-
ical planned route for the ship is shown in Figure 4(b). The physical planned
route is the shortest route for this set of cargoes. Notice that the sequence of
cargoes in the optimal solution might be different when we consider the time
windows. Finally, in Figure 4(c), we see the load onboard the ship at departure
from the respective ports for the planned route.

In the mathematical description of the problem, let N be the set of cargoes

indexed by i. Cargo i is represented by a node in a network, and this node
includes one loading port and one unloading port for cargo i. Further, let V
be the set of ships in the fleet indexed by v. The set (Nv  Av ) is the network
associated with a specific ship v, where Nv and Av represent the sets of the
nodes and arcs, respectively. Not all ships can visit all ports and take all cargoes,
and Nv = {feasible nodes for ship v}∪{o(v) d(v)}. Here, o(v) and d(v) are an
artificial origin cargo and an artificial destination cargo for ship v, respectively.
If the ship is not used, d(v) will be serviced just after o(v). The set Av contains
all feasible arcs for ship v, which is a subset of {i ∈ Nv } × {i ∈ Nv }. This set will
be calculated based on time constraints and other restrictions. The arc (i j)
connects cargo i and cargo j, where cargo i will be serviced just before cargo j
if the arc is used.
Let us look again at Example 4.1. Figure 5 shows the route of this example
(marked with bold lines) drawn over the underlying network. The ship leaves
 Ch. 4. Maritime Transportation 225

Table 5.
Cargo information for Examples 4.1 and 4.2

Loading port Unloading port Quantity

Cargo 1 A C 1 ship
2

Cargo 2 D E full ship

Cargo 3 B D 1 ship
2

Fig. 4. (a) Geographical picture of the ports for Examples 4.1 and 4.2. (b) Physical route for the ship
for Example 4.1. (c) Load onboard the ship at departure for Example 4.1.

the artificial origin cargo node in the beginning of its route and lifts cargo 1 that
is represented by node Cargo 1. The route is then followed by node Cargo 3,
node Cargo 2, and finally the artificial destination cargo node. The other arcs
are possible precedence combinations between the cargoes given in this exam-
ple.
For each arc, TSijv represents the calculated time for ship v from the arrival
at the loading port for cargo i until the arrival at the loading port for cargo j. It
includes the sum of the time for loading and unloading cargo i, the sailing time
between ports related to cargo i and the sailing time from the unloading port
for cargo i to the loading port for cargo j. Let [TMNiv  TMXiv ] denote the time
window for ship v associated with the loading port for cargo i, where TMNiv
is the earliest time for start of service, while TMXiv is the latest time. In the
226 M. Christiansen et al.

Fig. 5. The route of Example 4.1 drawn over the underlying network.

underlying real problem these data are seldom specified for each ship v but
are appropriate in the mathematical model due to a preprocessing phase. The
variable sailing and port costs are represented by Cijv .
In the mathematical formulation, we use the following types of variables: the
binary flow variable xijv , v ∈ V , (i j) ∈ Av , equals 1, if ship v services cargo i
just before cargo j, and 0 otherwise. This flow variable determines which ship
takes a particular cargo. The time variable tiv , v ∈ V , i ∈ Nv , represents the
time at which service begins at the loading port of cargo i with ship v.
The arc flow formulation of the industrial ship scheduling problem with full
shiploads is as follows:
 
min Cijv xijv (4.1)
v∈V (ij)∈Av
subject to

xijv = 1 ∀i ∈ N  (4.2)
v∈V j∈Nv

xo(v)jv = 1 ∀v ∈ V  (4.3)
j∈Nv
   
xijv − xjiv = 0 ∀v ∈ V  j ∈ Nv \ o(v) d(v)  (4.4)
i∈Nv i∈Nv

xid(v)v = 1 ∀v ∈ V  (4.5)
i∈Nv
xijv (tiv + TSijv − tjv )  0 ∀v ∈ V  (i j) ∈ Av  (4.6)
TMNiv  tiv  TMXiv  ∀v ∈ V  i ∈ Nv  (4.7)
xijv ∈ {0 1} ∀v ∈ V  (i j) ∈ Av  (4.8)
 Ch. 4. Maritime Transportation 227

The objective function (4.1) minimizes the costs of operating the fleet. Con-
straints (4.2) ensure that all cargoes that the shipping company has committed
itself to carry are serviced. Constraints (4.3)–(4.5) describe the flow on the sail-
ing route used by ship v. Constraints (4.3) and (4.5) ensure that ship v services
the artificial origin cargo and the artificial destination cargo once, respectively.
Constraints (4.6) describe the compatibility between routes and schedules. The
time for start of service of cargo j cannot be less than the sum of the start time
of cargo i and the service time for loading, transporting and unloading cargo i
and the sailing time from the unloading port for cargo i to the loading port
for cargo j with ship v, if ship v is really servicing cargo i just before cargo j.
Constraints (4.6) contain an inequality sign because waiting time is permitted
before the start of service in a port. The time window constraints are given by
constraints (4.7). For the artificial origin cargo, this time window is collapsed to
the value when ship v is available for new cargoe(s) during the planning hori-
zon. If ship v is not servicing cargo i, we get an artificial starting time within
the time windows for that (i v)-combination. This means that we get a start-
ing time for each (i v)-combination. However, just the starting time associated
with ship v actually lifting the particular cargo i is real. Finally, the formulation
involves binary requirements (4.8) on the flow variables.
This industrial ship scheduling problem for full shipload cargoes cor-
responds to a multitraveling salesman problem with time windows (see
Desrosiers et al., 1995).
The model (4.1)–(4.8) is still valid if the planning problem involves cargoes
that are not equal to the capacity of the ship but a ship can carry only one cargo
at a time. The set Nv gives the cargoes that can be serviced by ship v. For this
variant of the problem, the set Nv is calculated based on the capacity of the
ship and the load quantity of cargo i.
The quantities of some cargoes might be given in an interval, and the cargo
size is then determined by the ship capacity a priori for each cargo and ship
combination. Relative revenues for loading larger cargo quantities for a cargo i
due to larger ship capacity can be included in Cijv .
The load of the ship might in some cases be first loaded in several loading
ports in the same region and unloaded in one or several ports. The model
(4.1)–(4.8) is also valid for such a situation. However, the calculated sailing
times have to be adjusted such that times in all ports are included. Now, the
time variable tiv represents the time at which service begins at the first loading
port for cargo i with ship v.
In the literature, we find several studies on the industrial ship scheduling
problems with full shipload cargoes. Brown et al. (1987) describe such a prob-
lem where a major oil company is shipping crude oil from the Middle East
to Europe and North America. Fisher and Rosenwein (1989) study a prob-
lem that is conceptually quite similar to the one in Brown et al. (1987). Here,
a fleet of ships controlled by the Military Sealift Command of the US Navy is
engaged in pickup and delivery of various bulk cargoes. Each cargo may have
up to three loading points which are often the same port or nearby ports and
228 M. Christiansen et al.

up to three unloading points that are frequently close to each other. In con-
trast to Brown et al. (1987), each cargo may not be a full shipload. However,
at most one cargo is on a vessel at any time. Therefore, the same model is still
valid. Another similar problem of shipping crude oil is studied by Perakis and
Bremer (1992).

4.1.2 Multiple cargoes with fixed cargo size

Here we present an industrial ship routing and scheduling problem where
several cargoes are allowed to be onboard the ship at the same time. The
objective of the scheduling problem is to minimize the sum of the costs for
all the ships in the fleet while ensuring that all cargoes are lifted from their
loading ports to their corresponding unloading ports. Each cargo consists of a
designated number of units of a product or a commodity. Time windows are
normally imposed for both the pickup and delivery of the cargoes. The ship ca-
pacities, the cargo type and quantities are such that the ships may carry several
cargoes simultaneously. This means that another loading port can be visited
with some cargoes still onboard. We assume that the cargoes are compatible
with each other.

Example 4.2. This example is based on Example 4.1. We have the same cargo
information as given in Table 5, and the geographical picture of the ports is
shown in Figure 4(a). However, multiple cargoes can be carried simultane-
ously. Figure 6(a) shows the physical route for the ship.
Cargo 1 is lifted in port A and the ship sails to port B to load Cargo 3. On
departure the ship is fully loaded with two cargoes. Figure 6(b) shows the load
onboard the ship upon departure from each port.

Fig. 6. (a) Physical route for a ship with multiple cargoes onboard for Example 4.2. (b) Load onboard
the ship at departure for Example 4.2.
 Ch. 4. Maritime Transportation 229

We have the same conditions for the fleet as for the problem described
in Section 4.1.1, concerning a heterogeneous fixed fleet with various variable
costs. In addition, we assume that the sailing costs do not depend on the load
onboard the ship.
In the mathematical description of the problem also here each cargo is rep-
resented by an index i. However, associated with the loading port of cargo i,
there is a node i, and with the corresponding unloading port a node N + i,
where N is the number of cargoes that has to be serviced during the plan-
ning horizon. Note that different nodes may correspond to the same phys-
ical port. Let NP = {1     N} be the set of loading (or pickup) nodes
and ND = {N + 1     2N} be the set of unloading (or delivery) nodes,
and define N = NP ∪ ND . V is the set of ships in the fleet indexed by v.
Then (Nv  Av ) is the network associated with a specific ship v. Here, Nv =
{feasible nodes for ship v} ∪ {o(v) d(v)} is the set of ports that can be visited
by ship v and o(v) and d(v) are the artificial origin depot and artificial desti-
nation depot of ship v, respectively. Geographically, the artificial origin depot
o(v) can be either a port or a point at sea, while the artificial destination depot
d(v) is the last planned unloading port for ship v. If the ship is not used d(v)
will represent the same location as o(v). Here Av contains the set of all feasi-
ble arcs for ship v, which is a subset of {i ∈ Nv } × {i ∈ Nv }. This set will be
calculated based on capacity and time constraints, and other restrictions such
as those based on precedence of loading and unloading nodes for the same
cargo. From these calculations, we can extract the sets NPv = NP ∩ Nv and
NDv = ND ∩ Nv consisting of loading and unloading nodes that ship v may
visit, respectively.
Let us refer back to Example 4.2. In the underlying network for the example,
we introduce two nodes for each of the cargoes. This means that Cargo 1 is
represented by the loading node 1 and the unloading node N + 1. The loading
port for Cargo 2 and the unloading port for Cargo 3 are the same physical
port. That means that both node 2 and node N + 3 represent port D. Figure 7
shows the route of this example (marked with bold lines). The other arcs are
left out of the figure for sake of clarity. In general, there will be arcs from o(v)
to all loading ports and d(v). In addition, we will have arcs into d(v) from o(v)
and all unloading ports. The network for the real loading and unloading ports
will be complete except for arcs from each of the unloading ports N + i to
the corresponding loading port i. The sequence of nodes for this example is as
follows: o(v)–1–3–(N + 1)–(N + 3)–2–(N + 2)–d(v).
The fixed cargo quantity for cargo i is given by Qi , while the capacity of
ship v is given by VCAPv . For each arc, TSijv represents the sum of the calculated
sailing time from node i to node j with ship v and the service time at node i.
Let [TMNiv  TMXiv ] denote the time window associated with node i and ship v.
The variable sailing and port costs are represented by Cijv .
In the mathematical formulation, we use the following types of variables:
the binary flow variable xijv , v ∈ V , (i j) ∈ Av , equals 1, if ship v sails from
node i directly to node j, and 0 otherwise. The time variable tiv , v ∈ V , i ∈ Nv ,
230 M. Christiansen et al.

Fig. 7. The route of Example 4.2.

represents the time at which service begins at node i, while variable liv , v ∈ V ,
i ∈ Nv \{d(v)}, gives the total load onboard ship v just after the service is
completed at node i.
The arc flow formulation of the industrial ship scheduling problem with
fixed cargo sizes is as follows:
 
min Cijv xijv (4.9)
v∈V (ij)∈Av
subject to

xijv = 1 ∀i ∈ NP  (4.10)
v∈V j∈Nv

xo(v)jv = 1 ∀v ∈ V  (4.11)
j∈NPv ∪{d(v)}
 
xijv − xjiv = 0
i∈Nv i∈Nv
 
∀v ∈ V  j ∈ Nv \ o(v) d(v)  (4.12)

xid(v)v = 1 ∀v ∈ V  (4.13)
i∈NDv ∪{o(v)}
xijv (tiv + TSijv − tjv )  0 ∀v ∈ V  (i j) ∈ Av  (4.14)
TMNiv  tiv  TMXiv  ∀v ∈ V  i ∈ Nv  (4.15)
xijv (liv + Qj − ljv ) = 0 ∀v ∈ V  (i j) ∈ Av | j ∈ NPv  (4.16)
 Ch. 4. Maritime Transportation 231

xiN+jv (liv − Qj − lN+jv ) = 0

∀v ∈ V  (i N + j) ∈ Av | j ∈ NPv  (4.17)
lo(v)v = 0 ∀v ∈ V  (4.18)
 
Qi xijv  liv  VCAPv xijv  ∀v ∈ V  i ∈ NPv  (4.19)
j∈Nv j∈Nv

0  lN+iv  (VCAPv − Qi )xN+ijv 
j∈Nv
∀v ∈ V  i ∈ NPv  (4.20)
tiv + TSiN+iv − tN+iv  0 ∀v ∈ V  i ∈ NPv  (4.21)
 
xijv − xjN+iv = 0 ∀v ∈ V  i ∈ NPv  (4.22)
j∈Nv j∈Nv
xijv ∈ {0 1} ∀v ∈ V  (i j) ∈ Av  (4.23)
The objective function (4.9) minimizes the costs of operating the fleet. Con-
straints (4.10) ensure that all cargoes that the shipping company has committed
itself to carry are serviced. Constraints (4.11)–(4.13) describe the flow on the
sailing route used by ship v. Constraints (4.14) describe the compatibility be-
tween routes and schedules. The starting time of the service at node j cannot
be less than the sum of the starting time and the loading time at node i and the
sailing time from i to j with ship v, if ship v is really sailing between these two
nodes. The time window constraints are given by (4.15). If ship v is not visiting
node i, we will get an artificial starting time within the time windows for that
(i v)-combination. Introduction of artificial starting times is practical, due to
constraints (4.21). Constraints (4.16) and (4.17) give the relationship between
the binary flow variables and the ship load at each loading and unloading port,
respectively. The initial load condition for each ship is given by (4.18). The
ship is empty at the beginning of the planning horizon as mentioned in the
opening of Section 4.1. Constraints (4.19) and (4.20) represent the ship capac-
ity intervals at loading and unloading nodes, respectively. Constraints (4.20)
can be omitted from the model since the upper bound can never be exceeded
due to constraints (4.19) and the precedence and coupling constraints (4.21)
and (4.22). The precedence constraints forcing node i to be visited before node
N + i are given in (4.21). For both constraints (4.14) and (4.21), the constraints
appear only if the beginning of the time window for nodes j and N + i, respec-
tively, is less than the earliest calculated arrival time at the node. Along with
the coupling constraints (4.22), constraints (4.21) ensure that the same ship v
visits both node i and N + i, i ∈ NPv . Finally, the formulation involves binary
requirements (4.23) on the flow variables.
We find a few applications for this industrial shipping problem with fixed
cargo quantities in the literature. Fagerholt and Christiansen (2000a) study a
multiproduct scheduling problem. They extend the model presented here, and
include allocation of cargoes to different flexible cargo holds. For more details,
232 M. Christiansen et al.

see Section 4.1.4. Further, Christiansen and Fagerholt (2002) present a real
ship scheduling problem which is based on the model (4.9)–(4.23). In addition,
they focus on two important issues in the shipping industry, namely ports closed
at night and over weekends and long loading or unloading operations. This
study is described in more detail in Section 6.
The multiple cargo with fixed cargo size ship scheduling problem is also
studied by Psaraftis (1988) for the US Military Sealift Command. The ob-
jective is to allocate cargo ships to cargoes so as to ensure that all cargoes
arrive at their destinations as planned. Constraints that have to be satisfied in-
clude loading and unloading time windows for the cargoes, ship capacity and
cargo/ship/port compatibility. The problem is dynamic, because in a military
mobilization situation anything can change in real time. The paper focuses
on the dynamic aspects of the problem and the algorithm that is developed
is based on the “rolling horizon” approach. Later, Thompson and Psaraftis
(1993) applied a new class of neighborhood search algorithms to a variety of
problems, including the problem of the US Military Sealift Command.

4.1.3 Multiple cargoes with flexible cargo size

For many real ship scheduling problems, the cargo quantity is given in an
interval and the shipping company can choose the actual load quantity that
best fits its fleet and schedule. For such problems, the minimum cost problem
is transferred to a maximum profit problem. Apart from these issues, the prob-
lem is identical to the problem described in Section 4.1.2. We use the same
mathematical notation and the same type of network representation as in Fig-
ure 7. However, we need the following additional notation:
The variable quantity interval is given by [QMNi  QMXi ], where QMNi is the
minimum quantity to be lifted, while QMXi is the maximum quantity for cargo i.
The time required to load or unload one unit of a cargo at node i is given
by TQi . The node can either be a loading or unloading node, which means that
the time per unit might be different for loading and unloading. Here TSijv is
just the sailing time between the two ports and does not include the service
time in any of the ports.
We need an additional continuous variable qiv , v ∈ V , i ∈ NPv , that rep-
resents the quantity of cargo i, when cargo i is lifted by ship v and loaded at
node i and unloaded at node N +i. The revenue of carrying a cargo is normally
the cargo quantity qiv multiplied by a revenue per unit of cargo Pi . However,
in some cases the revenue from a cargo may be a lump sum or another func-
tion of the cargo quantity, and then the objective function becomes nonlinear.
In the following mathematical formulation of the objective function we use a
linear term for the revenue from carrying the cargoes.
The ship scheduling problem with flexible cargo sizes is formulated as fol-
lows:
    
max Pi qiv − Cijv xijv (4.24)
v∈V i∈NPv v∈V (ij)∈Av
 Ch. 4. Maritime Transportation 233

subject to

xijv = 1 ∀i ∈ NP  (4.25)
v∈V j∈Nv

xo(v)jv = 1 ∀v ∈ V  (4.26)
j∈NPv ∪{d(v)}
 
xijv − xjiv = 0
i∈Nv i∈Nv
 
∀v ∈ V  j ∈ Nv \ o(v) d(v)  (4.27)

xid(v)v = 1 ∀v ∈ V  (4.28)
i∈NDv ∪{o(v)}
xijv (tiv + TQi qiv + TSijv − tjv )  0
∀v ∈ V  (i j) ∈ Av | i ∈ NPv ∪ o(v) (4.29)
xN+ijv (tN+iv + TQN+i qiv + TSN+ijv − tjv )  0
∀v ∈ V  (N + i j) ∈ Av | i ∈ NPv  (4.30)
TMNiv  tiv  TMXiv  ∀v ∈ V  i ∈ Nv  (4.31)
xijv (liv + qjv − ljv ) = 0
∀v ∈ V  (i j) ∈ Av | j ∈ NPv  (4.32)
xiN+jv (liv − qjv − lN+jv ) = 0
∀v ∈ V  (i N + j) ∈ Av | j ∈ NPv  (4.33)
 
QMNi xijv  qiv  QMXi xijv 
j∈Nv j∈Nv
∀v ∈ V  i ∈ NPv  (4.34)
lo(v)v = 0 ∀v ∈ V  (4.35)

qiv  liv  VCAPv xijv  ∀v ∈ V  i ∈ NPv  (4.36)
j∈Nv

0  lN+iv  VCAPv xN+ijv − qiv 
j∈Nv
∀v ∈ V  i ∈ NPv  (4.37)
tiv + TQi qiv + TSiN+iv − tN+iv  0
∀v ∈ V  i ∈ NPv  (4.38)
 
xijv − xjN+iv = 0 ∀v ∈ V  i ∈ NPv  (4.39)
j∈Nv j∈Nv
xijv ∈ {0 1} ∀v ∈ V  (i j) ∈ Av  (4.40)
The objective function (4.24) maximizes the profit gained by operating the
fleet. The constraints (4.25)–(4.40) are equivalent to (4.10)–(4.23), apart from
234 M. Christiansen et al.

the following constraints. The constraints ensuring feasible time schedules are
split into constrains for loading in port i, (4.29), and unloading in port N + i,
(4.30). These constraints are adjusted for the variable loading time at port i.
Variable qiv is not defined for i = o(v), so the term TQi qiv does not exist
for i = o(v) in constraints (4.29). Here, constraints (4.32) and (4.33) in-
clude a variable load quantity instead of the fixed quantity in constraints (4.16)
and (4.17). In constraints (4.34) the load quantity interval is defined for each
cargo i. The load variable qiv is forced to 0 by (4.34) if cargo i is not lifted by
ship v. Constraints (4.36)–(4.38) are adjusted for the variable load quantity.
A ship scheduling problem with flexible cargo sizes is studied by Brønmo et
al. (2007) for transportation of bulk cargoes by chemical tankers and has many
similarities to the problem described here. The solution method is based on a
set partitioning approach that gives optimal solutions to the problem. Korsvik
et al. (2007) solve the same problem by using a multistart local search heuristic.
There are operations where a ship can carry only one cargo at a time, but
the ship is not necessarily filled up each time and the cargo quantity is given in
an interval. For this situation, we still have variable load quantities and arrival
times as in the model of this section. However, we do not need nodes for both
loading and unloading ports, but just a common node representing the cargo
as we did in the model of Section 4.1.1.

4.1.4 Multiple products

In Sections 4.1.1–4.1.3 we assumed that the cargoes consist of mixable prod-
ucts that can be loaded onboard regardless of the type of product already on-
board. In addition, different cargoes are compatible with each other. However,
often multiple nonmixable products are carried onboard a ship simultaneously.
In such cases the cargo carrying space of the vessel must be divided into sepa-
rate tanks (compartments or holds) that are usually fixed. For example, a large
chemical tanker may have from 20 to 50 tanks. We start with considering the
case where the cargo tanks of the ship are of equal size. In reality, this is seldom
the case. However, it may be possible to separate the tanks into sets that are of
about equal size. If the ship has many tanks, this assumption is reasonable. In
addition, we assume that the cargo consists of mixable products, but different
cargoes have to be stored in different tanks.
In the mathematical description of the problem, we need the following no-
tation: the number of tanks (or cargo holds) of ship v is given by Hv and
the capacity of a tank (hold) of ship v is given by HCAPv = VCAPv /Hv . As the
ship is assumed empty at the first time it is available for scheduling during the
planning horizon, the number of tanks (holds) occupied is also 0. Variable hiv ,
v ∈ V , i ∈ Nv , represents the number of tanks (holds) occupied after servicing
node i by ship v. We still use the continuous variable qiv , v ∈ V , i ∈ NPv , rep-
resenting the quantity of cargo i, when cargo i is lifted by ship v and loaded at
node i and unloaded at node N + i.
 Ch. 4. Maritime Transportation 235

In order to allow several different nonmixable cargoes onboard simultane-

ously, we need the following constraints added to formulation (4.24)–(4.40):
 
qjv
xijv hiv + − hjv = 0
HCAPv
∀v ∈ V  (i j) ∈ Av | j ∈ NPv  (4.41)
 
qjv
xiN+jv hiv − − hN+jv = 0
HCAPv
∀v ∈ V  (i N + j) ∈ Av | j ∈ NPv  (4.42)
 qiv 
xijv  hiv  Hv xijv  ∀v ∈ V  i ∈ NPv  (4.43)
HCAPv
j∈Nv j∈Nv
 qiv
0  hN+iv  Hv xN+ijv −  ∀v ∈ V  i ∈ NPv  (4.44)
HCAPv
j∈Nv
ho(v)v = 0 ∀v ∈ V  (4.45)
hiv ∈ [0 Hv ] and integer ∀v ∈ V  i ∈ Nv  (4.46)
Constraints (4.41) and (4.42) describe the compatibility between routes and
the number of occupied tanks when the arrival node is a loading port and an
unloading port, respectively. The intervals of the number of occupied tanks af-
ter servicing the loading and unloading nodes are given in constraints (4.43)
and (4.44), respectively. Next, constraints (4.45) impose the initial tank occu-
pancy condition for each ship. Finally, the integer requirements for the tank
number variables are given. The integer interval [0 Hv ] in (4.46) can be re-
duced by information from (4.34) and (4.44).
For problems with multiple, nonmixable, products for a cargo, the alloca-
tion of products to the various tanks is normally needed. For transportation of
liquid products, the quantity has to be flexible due to stability considerations
and to prevent product sloshing in partially empty tanks.
In the literature, Scott (1995) presents a problem involving the shipping
of refined oil products from a refinery to several depots. Several types of
tankers/ships with fixed tanks enable different products to be carried on the
same voyage (without mixing them). Another study with multiple products
is given by Bausch et al. (1998). They present a decision support system for
medium-term scheduling where a fleet of coastal tankers and barges are trans-
porting liquid bulk products among plants, distribution centers, and industrial
customers. A set of cargoes has to be conveyed by the available fleet of ves-
sels and each cargo consists of an ordered volume of up to five products. The
vessels may have up to seven fixed tanks, thus allowing a cargo consisting of
several products to be lifted by the same ship. When multiple cargoes are
carried simultaneously, different cargoes of the same product are stowed in
different tanks. Such cargoes are not mixed in order to eliminate the need
for measuring the unloaded quantity at the multiple unloading ports. A similar
236 M. Christiansen et al.

problem is studied by Sherali et al. (1999) describing a ship scheduling problem

where crude oil and a number of refined oil-related products are to be shipped
from ports in Kuwait to customers around the world. Here, each cargo is a full
shipload of a compartmentalized group of products, and is characterized by its
mix (oil, refined products, etc.), loading port, loading date, unloading port, and
unloading date. The ships have multiple tanks of different sizes, so they intro-
duce a flow variable that is 1 if a particular tank carries a particular product on
a particular leg (i j) with ship v. The model is extended compared to the one
presented here and includes the allocation of product quantities to tanks.
Recently, Jetlund and Karimi (2004) presented a similar problem for multi-
compartment tankers engaged in shipping bulk liquid chemicals. They present
a mixed-integer linear programming formulation using variable-length time
slots. They solve real instances of the problem by a heuristic decomposition
algorithm that obtains the fleet schedule by repeatedly solving the base formu-
lation for a single ship.
Fagerholt and Christiansen (2000a, 2000b) extend the model formulated
above and study a ship scheduling problem where each ship in the fleet is
equipped with a flexible cargo hold that can be partitioned into several smaller
compartments in a given number of ways. The scheduling of the ships consti-
tutes a multiship pickup and delivery problem with time windows, while the
partitioning of the ships’ flexible cargo holds and the allocation of cargoes to
the smaller compartments is a multiallocation problem.

4.1.5 Contracted and optional cargoes

A ship scheduling problem for the tramp market boils down to pickup and
delivery of cargoes at maximum profit. A tramp shipping company often en-
gages in Contracts of Affreightment (COA). These are contracts to carry speci-
fied quantities of cargo between specified ports within a specific time frame for
an agreed payment per ton. Mathematically, these cargoes can be handled in
the same way as the cargoes for an industrial shipping problem. Tramp ships
operate in a manner similar to a taxi and follow the available cargoes. They
may also take optional cargoes. These optional cargoes will be picked up at
a given loading port and delivered to a corresponding unloading port if the
tramp shipping company finds it profitable. Thus in tramp shipping each cargo
is either committed or optional and consists of a quantity given in an interval.
In the mathematical description of the problem we need to define two ad-
ditional sets. For the tramp ship scheduling problem we need to partition the
set of cargoes, NP , into two subsets, NP = NC ∪ NO , where NC is the set of
cargoes the shipping company has committed itself to carry, while NO repre-
sents the optional spot cargoes. The mathematical formulation is the same as
(4.24)–(4.40), except for constraints (4.25). These constraints are split into two
types of constraints as follows:

xijv = 1 ∀i ∈ NC  (4.47)
v∈V j∈Nv
 Ch. 4. Maritime Transportation 237

xijv  1 ∀i ∈ NO  (4.48)
v∈V j∈Nv

Constraints (4.47) ensure that all the cargoes that the shipping company
has committed itself to carry are serviced. The corresponding constraints for
the optional cargoes are given in (4.48). Note that the equality sign in (4.47)
is replaced by an inequality in (4.48) since these cargoes do not have to be
carried. When one uses a branch-and-bound algorithm to solve this problem it
may be useful to insert an explicit slack variable in constraints (4.48).
A typical tramp ship scheduling problem with both optional and contracted
cargoes is described in the pioneer work of Appelgren (1969, 1971). The ships
in the fleet are restricted to carry only one cargo at a time, and the cargo quan-
tities are fixed. This type of problem is extended in Brønmo et al. (2006) where
cargoes are of flexible sizes for a tramp ship scheduling application.

4.1.6 Use of spot charters

In some cases the controlled fleet may have insufficient capacity to serve all
cargoes for an industrial ship scheduling problem or all committed cargoes for
a tramp ship scheduling problem during the planning horizon. In such a case
some of the cargoes can be serviced by spot charters, which are ships chartered
for a single voyage.
We extend the formulation for the tramp ship scheduling problem and in-
troduce a variable si , i ∈ NC , that is equal to 1 if cargo i is serviced by a spot
charter and 0 otherwise. In addition, let πi be the profit if cargo i is serviced by
a spot charter. This profit can be either positive or negative. When we take the
spot shipments into account, (4.24) and (4.25) (or (4.47)) become:
     
max Pi qiv − Cijv xijv + πi si (4.49)
v∈V i∈NPv v∈V (ij)∈Av i∈NC
subject to

xijv + si = 1 ∀i ∈ NC  (4.50)
v∈V j∈Nv
si ∈ {0 1} ∀i ∈ NC  (4.51)
Now, the objective function (4.49) maximizes the profit (or actually the mar-
ginal contribution, since fixed costs are excluded from the formulation). The
terms are divided into the profit gained by (a) operating the fleet and (b) ser-
vicing the cargoes by spot charters. Also here it is assumed that the fleet is
fixed during the planning horizon, and it is not possible to charter out some
of the ships during that horizon. Constraints (4.50) ensure that all committed
cargoes are serviced either by a ship in the fleet or by a spot charter. Con-
straints (4.51) impose the binary requirements on the spot variables. According
to (4.50), these variables do not need to be defined as binary since the flow vari-
ables are binary. However, doing so might give computational advantages in a
branch-and-bound process.
238 M. Christiansen et al.

We can find several applications described in the literature for both tramp
and industrial shipping where some of the cargoes might be serviced by spot
charters, see, for instance, Bausch et al. (1998), Christiansen and Fagerholt
(2002), Sherali et al. (1999), and Fagerholt (2004).

4.2 Solution approaches for industrial and tramp scheduling models

Theoretically the models presented in Section 4.1 can be solved directly by

use of standard commercial optimization software for mixed integer linear pro-
gramming after linearization of some nonlinear functions.
For instance, constraints (4.32) are given as follows:

xijv (liv + qjv − ljv ) = 0 ∀v ∈ V  (i j) ∈ Av | j ∈ NPv 

These constraints are linearized as

liv + qjv − ljv + VCAPv xijv  VCAPv 

∀v ∈ V  (i j) ∈ Av | j ∈ NPv  (4.52)
liv + qjv − ljv − VCAPv xijv  −VCAPv 
∀v ∈ V  (i j) ∈ Av | j ∈ NPv  (4.53)
The ship capacity VCAPv is the largest value that (liv + qjv − ljv ) can take, so
constraints (4.52) are redundant if xijv is equal to 0. Similarly, (liv + qjv − ljv )
will never be less than −VCAPv . The schedule constraints (4.29) are linearized
in the same way as constraints (4.32), but, because the original constraints have
a  sign, just one type of constraints is necessary in the linearized version. This
way of linearizing the nonlinear constraint is also presented by Desrosiers et
al. (1995).
Due to the models’ complexity, only small sized data instances can be solved
directly to optimality by using standard commercial optimization software.
Therefore, these models usually require reformulation in order to solve them
to optimality.
By studying the models presented, we see that for each cargo i we have
exactly one constraint linking the ships. This corresponds to constraint types
(4.2), (4.10), and (4.25) for the industrial shipping problems presented in Sec-
tions 4.1.1, 4.1.2, and 4.1.3, respectively, and constraint types (4.47) and (4.48)
for the tramp shipping problems. These constraints ensure that each cargo i is
served by a ship exactly once (or at most once). These constraints are called
here common constraints. All other constraints refer to each ship v and will be
called the ship routing constraints. For example, in the model (4.24)–(4.40), the
constraints (4.26)–(4.40), constitute the routing problem for each ship where
the time windows, load quantity interval and load on board the ship are con-
sidered. This observation is often exploited in the solution methods used for
such type of problems. The exact solution methods are usually based on col-
umn generation approaches, where the ship routes constitute the columns. We
 Ch. 4. Maritime Transportation 239

will therefore concentrate on two such main solution approaches, the Dantzig–
Wolfe decomposition approach in Section 4.2.1, and the set partitioning ap-
proach with columns generated a priori in Section 4.2.2. Finally, in Section
4.2.3 we will briefly discuss some other approaches.

4.2.1 The Dantzig–Wolfe decomposition approach

The common constraints constitute the master problem in the Dantzig–
Wolfe (DW) decomposition approach. None of the ship routing constraints
include interaction between ships, so these constraints can be split into one
subproblem for each ship. For each ship’s subproblem, we need to find a feasi-
ble route with regard to the time windows, quantity intervals and the quantity
on board the ship, so that this quantity does not exceed the capacity of the ship.
Each of the feasible combinations of sailing legs (i j) to geographical routes
for a ship, including the information about starting times and load quantities
at each port, is called a ship schedule and is indexed by r. That means a ship
schedule r for ship v includes information about the values of the flow from
each node i directly to node j in the geographical route, the quantity loaded
or unloaded at each node i, and the starting times at each node i. The con-
stant Xijvr equals 1 if leg (i j) by vessel v in route r and 0 otherwise. Given a
geographical route, it is possible to find the optimal load quantity and starting
time at each port in the route.
Since the ship routing subproblems define path structures, their extreme
points correspond to paths in the underlying networks. Set Rv defines the ex-
treme points for ship v. Any solution xijv satisfying the ship routing constraints
can then be expressed as a nonnegative convex combination of these extreme
points and must consist of binary xijv values, i.e.,

xijv = Xijvr yvr  ∀v ∈ V  (i j) ∈ Av  (4.54)
r∈Rv

yvr = 1 ∀v ∈ V  (4.55)
r∈Rv
yvr ∈ {0 1} ∀v ∈ V  r ∈ Rv  (4.56)
The new variables yvr , v ∈ V , r ∈ Rv , are called the schedule variables, and
equal 1 if ship v chooses to sail schedule r. Let Aivr = j∈Nv Xijvr be equal
to 1 if schedule r for ship v services cargo i and 0 otherwise. The column vector
in the master problem contains information about the actual cargoes in sched-
ule r for ship v. In addition, the optimal geographical route, the arrival times
and the size of the cargoes for the given set of cargoes for a schedule (v r)
determine the profit coefficient in the objective function for the corresponding
column.

The master problem in the DW decomposition approach. Substituting (4.54)–

(4.56) in (4.24) and (4.25), the integer master problem for the industrial ship
240 M. Christiansen et al.

scheduling problem with flexible cargo sizes is transformed into:


max Pvr yvr (4.57)
v∈V r∈Rv
subject to

Aivr yvr = 1 ∀i ∈ NP  (4.58)
v∈V r∈Rv

yvr = 1 ∀v ∈ V  (4.59)
r∈Rv
yvr ∈ {0 1} ∀v ∈ V  r ∈ Rv  (4.60)
The objective function (4.57) maximizes the profit, where Pvr is the profit
of carrying the cargoes on schedule r by ship v, respectively. Constraints (4.58)
ensure that all cargoes are serviced by a ship in the company’s fleet. Constraints
(4.59) assure that each ship in the fleet is assigned exactly one schedule. Con-
straints (4.60) impose the binary requirements on the variables.
The corresponding master problem for the tramp ship routing and schedul-
ing problem with spot charters can be formulated as follows:
   
max Pvr yvr + πi si (4.61)
v∈V r∈Rv i∈NC
subject to

Aivr yvr + si = 1 ∀i ∈ NC  (4.62)
v∈V r∈Rv

Aivr yvr  1 ∀i ∈ NO  (4.63)
v∈V r∈Rv

yvr = 1 ∀v ∈ V  (4.64)
r∈Rv
yvr ∈ {0 1} ∀v ∈ V  r ∈ Rv  (4.65)
si ∈ {0 1} ∀i ∈ NC  (4.66)

Column generation and the subproblems within the DW decomposition approach.

The models (4.57)–(4.60) and (4.61)–(4.66) are based on knowledge of all fea-
sible ship schedules (columns). However, for some real ship scheduling prob-
lems it is time consuming to generate all these schedules, and the number of
such schedules would result in too many columns when solving the models. In-
stead, we solve the LP-relaxation of the restricted master problem which only
differs from the continuous original master problem by having fewer variables.
First, an initial restricted master problem is solved. Then some new columns
are added to the restricted master problem. These columns correspond to ship
schedules with positive reduced costs in the solution of the (maximization)
 Ch. 4. Maritime Transportation 241

master problem. This means that the dual values from the solution of the re-
stricted master problem are transferred to the subproblems. The subproblems
are solved and ship schedules are generated. The restricted master problem is
reoptimized with the added new columns, resulting in new dual values. This
procedure continues until no columns with positive reduced costs exist, and
no improvements can be made. At that point all the feasible solutions in the
original master problem have been implicitly evaluated. A continuous opti-
mal solution is then attained for both the original and the restricted master
problem. This LP-relaxed solution approach can be embedded in a branch-
and-bound search to find an optimal solution.
The subproblems can be formulated as shortest path problems and solved
by specific dynamic programming algorithms on generated networks for each
ship. The underlying network for each ship is specified by nodes, each of which
includes information about the port and the corresponding cargo with time
window for starting service and feasible cargo quantities. The recursive for-
mulas in the dynamic programming algorithms include the expressions for
the reduced costs. Algorithms for solving such problems are thoroughly de-
scribed in Desrosiers et al. (1995) and, for a special ship scheduling problem,
in Christiansen and Nygreen (1998b).
The DW decomposition approach has been used in numerous vehicle rout-
ing applications during the last twenty years. However, Appelgren (1969, 1971)
was the first one to use this approach for a pickup and delivery problem with
time windows, and that application was for the tramp shipping industry. An-
other ship routing application using the DW decomposition approach was
studied by Christiansen (1999) (see also Christiansen and Nygreen, 1998a,
1998b) and is discussed in Section 4.3.1.

4.2.2 The set partitioning approach

Ship scheduling problems are often tightly constrained, and in such a case it
is possible to generate schedules for all cargo combinations for all ships (i.e.,
all columns) a priori. The original arc flow models given in Section 4.1 can
be transformed to path flow models, and these path flow models correspond
to the master problems (4.57)–(4.60) and (4.61)–(4.66) in the Dantzig–Wolfe
(DW) decomposition approach. Both models are set partitioning (SP) models
or can easily be transformed into a SP model by introducing a slack variable
to constraints (4.63). In this approach all column vectors for the set partition-
ing model are generated in advance, and a binary variable yvr is defined for
each column vector generated. We can find numerous ship scheduling applica-
tions where this approach is used, see for instance Brown et al. (1987), Fisher
and Rosenwein (1989), Bausch et al. (1998), Fagerholt (2001), Fagerholt and
Christiansen (2000a, 2000b), Christiansen and Fagerholt (2002), and Brønmo
et al. (2006).
Here, we are generating columns for all feasible cargo combinations for a
particular ship v. For each of the feasible cargo combinations, we have to find
the geographical route, arrival time at the ports and the load quantities of the
242 M. Christiansen et al.

cargoes, such that the sum of the profits in the schedule is maximized. Further,
each node has to be serviced within its specified load interval and time window.
Finally, the loading node has to be visited before its corresponding unloading
node. If the ships in the fleet are equipped with cargo holds or tanks of various
capacities, the optimal allocation of products to tanks has to be determined as
well. All constraints that are exclusive for a particular ship have to be consid-
ered in the column generation phase of this approach. The problem of finding
the optimal route and schedule for a single ship can be solved by using dynamic
programming or by enumerating all feasible combinations of routes for a given
set of cargoes. Both approaches have been used. Fagerholt and Christiansen
(2000b) describe a dynamic programming approach for a combined multiship
pickup and delivery problem with time windows and a multiallocation prob-
lem, while Brønmo et al. (2006) describes an enumeration procedure for a
tramp scheduling problem with flexible cargo sizes.

4.2.3 Other solution approaches

In general, many solution methods, both optimization-based and heuris-
tic ones, were developed to solve routing and scheduling problems for other
modes of transportation. These methods can often be used with some minor
modifications for ship scheduling problems. Here we report several studies in
the ship scheduling literature where solution approaches other than the ones
discussed in Sections 4.2.1 and 4.2.2 were used.
Sherali et al. (1999) presented an aggregated mixed integer programming
model retaining the principal features of the real ship scheduling problem
with various cargo hold capacities and possible spot charters. A rolling hori-
zon heuristic is developed to solve the problem.
The ship scheduling problem studied by Scott (1995) is solved by applying
Lagrangian relaxation to the model to produce a set of potentially good sched-
ules, containing the optimal cargo schedule. A novel refinement of Benders’
decomposition is then used to choose the optimum schedule from within the
set, by avoiding solving an integer LP-problem at each iteration. The method
manages to break a difficult integer programming (IP) problem into two rela-
tively simple steps which parallel the steps typically taken by schedulers.
The tramp ship scheduling problem is studied by Brønmo et al. (2006, 2007),
and two solution approaches are suggested and compared. In addition to a set
partitioning approach, they describe a multistart heuristic consisting of two
phases. First multiple initial solutions are constructed by a simple insertion
method. Then a subset of the best initial solutions is improved by a quick local
search. A few of the best resulting solutions from the quick local search are
improved by an extended local search.

4.3 Maritime supply chains

A maritime supply chain is a supply chain where sea transport constitutes

at least one vital link. Supply chains of companies with foreign sources of raw
 Ch. 4. Maritime Transportation 243

materials or with overseas customers very often include maritime transporta-

tion. Supply chain optimization is an active field of research, and we can see
applications in almost all industries. However, the focus of such applications is
usually not on maritime transportation. At the tactical planning level the sup-
ply chain perspective is missing in ship routing and scheduling studies reported
in the literature.
Fleet scheduling is often performed under tight constraints. The shipper
specifies the cargoes with little or no flexibility in cargo quatnities and the time
widows are unnecessarily tight. The shipping company tries to find an optimal
fleet schedule based on such requirements while trying to maximize the profit
(or minimizing the costs). Realizing the potential of relaxing such constraints,
Brønmo et al. (2006) and Fagerholt (2001) considered flexibility in shipment
sizes and in time windows. The results of their studies show that there might
be a great potential in collaboration and integration along the supply chain, for
instance between the shippers and the shipping company.
Vendor managed inventory (VMI) takes advantage of the benefits of intro-
ducing flexibility in delivery time windows and cargo quantities, and transfers
inventory management and ordering responsibilities completely to the vendor
or the logistics provider. From recent literature and from our active contacts
with the shipping industry we see that an increased number of shipping com-
panies play the role of vendors in such logistics systems.
In this section we emphasize combined ship scheduling and inventory
management problems in the industrial and tramp shipping sectors. Section
4.3.1 discusses such a problem for transportation of a single product, while
Section 4.3.2 considers planning problems with multiple products. Finally, in
Section 4.3.3 we will comment on some other research within supply chain op-
timization that focuses on ship scheduling.

4.3.1 Inventory routing for a single product

In industrial maritime transportation, the transporter has often a twofold
responsibility. In this segment large quantities are transported, and normally
considerable inventories exist at each end of a sailing leg. In some situations,
the transporter has both the responsibility for the transportation and the inven-
tories at the sources and at the destinations. We consider a planning problem
where a single product is transported, and we call this problem the single prod-
uct inventory ship routing problem (s-ISRP). The single product is produced
at the sources, and we call the associated ports loading ports. Similarly, the
product is consumed at certain destinations and the corresponding ports are
called unloading ports. Inventory storage capacities are given in all ports, and
the planners have information about the production and consumption rates of
the transported product. We assume that these rates are constant during the
planning horizon. In contrast to most ship scheduling problems, the number of
calls at a given port during the planning horizon is not predetermined, neither
is the quantity to be loaded or unloaded in each port call. The production or
consumption rate and inventory information at each port, together with ship
244 M. Christiansen et al.

capacities and the location of the ports, determine the number of possible calls
at each port, the time windows for start of service and the range of feasible
load quantities for each port call.
If the product is loaded and unloaded in time at the sources and destina-
tions, respectively, neither production nor consumption will be interrupted.
The planning problem is therefore to find routes and schedules that minimize
the transportation cost without interrupting production or consumption. The
transporter owns both the producing sources and consuming destinations and
controls the inventories at both ends, so the inventory costs do not come into
play. The transporter operates a heterogeneous fleet of ships.
This s-ISRP has many similarities to the ship scheduling problem with flex-
ible cargo sizes. In contrast to the problem described in Section 4.1.3, the
number of cargoes is not given in advance, neither is the number of ship calls
at a port. Further, we have no predetermined loading and unloading port for
a particular cargo. In contrast to the problem described in Section 4.1.3, we
assume that the ship is not necessarily empty in the beginning of the planning
horizon but might have some load onboard. In addition, we have to keep track
of the inventory levels. There must be sufficient product in consumption in-
ventories, and their inventory in production ports cannot exceed the inventory
storage capacity. In addition, storage capacity limits exist for all consumption
inventories.
In the mathematical description of the problem each port is represented by
an index i and the set of ports is given by N . Let V , indexed by v, be the set
of available ships to be routed and scheduled. Not all ships can visit all ports,
and Nv = {feasible ports for ship v} ∪ {o(v) d(v)} is the set of ports that can
be visited by ship v. The terms o(v) and d(v) represent the artificial origin port
and artificial destination port of ship v, respectively. Each port can be visited
several times during the planning horizon, and Mi is the set of possible calls at
port i, while Miv is the set of calls at i that can be made by ship v. The port call
number is represented by an index m, and Mi is the last possible call at port i.
The necessary calls to a port are given by the set MCi and these necessary calls
have similarities to the contracted cargoes in the problems discussed in Section
4.1.5.
The set of nodes in the flow network represents the set of port calls, and
each port call is specified by (i m), i ∈ N , m ∈ Mi . In addition, we specify
flow networks for each ship v with nodes (i m), i ∈ Nv , m ∈ Miv . Finally,
Av contains all feasible arcs for ship v, which is a subset of {i ∈ Nv  m ∈
Miv } × {i ∈ Nv  m ∈ Miv }.
Figure 8 shows an artificial, simplified case consisting of five ports and two
ships. Each potential port call is indicated by a node. We see that port 1 can
be called three times during the planning horizon. We have three loading ports
and two unloading ports. The arrows indicate a solution to the planning prob-
lem where the routes and schedules satisfy the time windows and inventory
constraints.
 Ch. 4. Maritime Transportation 245

Fig. 8. A solution for a single product inventory routing problem with 5 ports and 2 ships.

Port 5 is the initial location for ship 1. The ship loads up to its capacity before
sailing to port call (3 1) and unloading this quantity. The ship continues to port
call (4 1) to load before ending up at port call (1 1). Ship 2 is empty at sea at
the beginning of the planning horizon and starts service at port call (2 1) after
some time. Here the ship loads to its capacity before sailing toward port call
(3 2). At port call (3 2) the ship unloads half of its load before it continues
to port call (1 2) and unloads the rest of the quantity on board. Here, two
unloading ports are called in succession.
Port 3 is called several times during the planning horizon. The solid, gray
line in Figure 9 shows the inventory level for port 3 during the planning hori-
zon. Ship 2 unloads half of its load at port call (3 2) as soon as possible. Here
it is important to ensure that the inventory level does not exceed the maximal
one when the unloading ends. Regardless of the rest of the planning problem,
the broken line in Figure 9 illustrates another extreme situation where ship 2
starts the service at port 3 as late as possible. Here, the inventory level is not
allowed to be under the minimal stock level when the unloading starts. From
these two extreme scenarios for the inventory levels, we can derive the feasible
time window for port call (3 2) given that the rest of the planning problem
remains unchanged.
The variable quantity interval is given by [QMNim  QMXimv ], where QMNim
is the minimum quantity to be (un)loaded at port call (i m) given that the port
is called, while QMXimv is the maximum quantity to be (un)loaded at port call
(i m) for ship v. The capacity of ship v is given by VCAPv .
246 M. Christiansen et al.

Fig. 9. The inventory level at port 3 during the planning horizon.

The time required to load or unload one unit of a cargo at port i is given
by TQi . The term TSijv represents the sailing time from port i to port j with
ship v. Let [TMNim  TMXim ] denote the arrival time window associated with
port call (i m). This time window can be calculated based on other data in
the model, such as the inventory conditions. In addition, for some port calls
the time windows are explicitly given. In a preprocessing phase, it is important
to make efforts to reduce the time window widths. In some ports, there is a
minimum required time, TBi , between a departure of one ship and the arrival
of the next ship, due to small port area or narrow channels from the port to the
pilot station. Let T denote the planning horizon.
The levels of the inventory (or stock) have to be within a given interval at
each port [SMNi  SMXi ]. The production rate Ri is positive if port i is producing
the product, and negative if port i is consuming the product. Further, constant
Ii is equal to 1, if i is a loading port, −1, if i is an unloading port, and 0, if i is
o(v) or d(v). The total variable cost Cijv that includes port, channel, and fuel
oil costs, corresponds to a sailing from port i to port j with ship v.
In the mathematical formulation we use the following types of variables: the
binary flow variable ximjnv , v ∈ V , (i m j n) ∈ Av , equals 1, if ship v sails from
node (i m) directly to node (j n), and 0 otherwise, and the slack variables wim ,
i ∈ N , m ∈ Mi \MCi , is equal to 1 if no ship takes port call (i m), and 0
otherwise. The time variable tim , (i ∈ N  m ∈ Mi ) ∪ (i ∈ o(v) ∀v m = 1),
represents the time at which service begins at node (i m). Variable limv , v ∈ V ,
i ∈ Nv \{d(v)}, m ∈ Miv , gives the total load onboard ship v just after the
service is completed at node (i m), while variable qimv , v ∈ V , i ∈ Nv \{d(v)},
 Ch. 4. Maritime Transportation 247

m ∈ Miv , represents the quantity loaded or unloaded at port call (i m), when
ship v visits (i m). Finally, sim , i ∈ N , m ∈ Mi , represents the inventory (or
stock) level when service starts at port call (i m). It is assumed that nothing is
loaded or unloaded at the artificial origin o(v) and that the ship arrives at o(v)
at a given fixed time; to(v)1 = TMNo(v)1 = TMXo(v)1 . The ships may have cargo
onboard, L0v , at the beginning of the planning horizon; lo(v)1v = L0v . At the
beginning of the planning horizon, the stock level at each port i is S0i .
The arc flow formulation of the single product inventory ship routing prob-
lem (s-ISRP) is as follows:
 
min Cijv ximjnv (4.67)
v∈V (imjn)∈Av
subject to
 
ximjnv + wim = 1 ∀i ∈ N  m ∈ Mi  (4.68)
v∈V j∈Nv n∈Mjv
 
xo(v)1jnv = 1 ∀v ∈ V  (4.69)
j∈Nv n∈Mjv
   
ximjnv − xjnimv = 0
i∈Nv m∈Miv i∈Nv m∈Miv
 
∀v ∈ V  j ∈ Nv \ o(v) d(v)  n ∈ Mjv  (4.70)
 
ximd(v)1v = 1 ∀v ∈ V  (4.71)
i∈Nv m∈Miv

ximjnv (tim + TQi qimv + TSijv − tjn )  0

∀v ∈ V  (i m j n) ∈ Av | j = d(v) (4.72)
to(v)1 = TMNo(v)1 = TMXo(v)1  ∀v ∈ V  (4.73)
TMNim  tim  TMXim  ∀i ∈ N  m ∈ Mi  (4.74)
ximjnv (limv + Ij qjnv − ljnv ) = 0
∀v ∈ V  (i m j n) ∈ Av | j = d(v) (4.75)
qo(v)1v = 0 ∀v ∈ V  (4.76)
lo(v)1v = L0v  ∀v ∈ V  (4.77)
 
qimv  limv  VCAPv ximjnv 
j∈Nv n∈Mjv
∀v ∈ V  i ∈ Nv  m ∈ Miv | Ii = 1 (4.78)
 
0  limv  VCAPv ximjnv − qimv 
j∈Nv n∈Mjv
∀v ∈ V  i ∈ Nv  m ∈ Miv | Ii = −1 (4.79)
248 M. Christiansen et al.
 
qimv  QMXimv ximjnv 
j∈Nv n∈Mjv
 
∀v ∈ V  i ∈ Nv \ o(v) d(v)  m ∈ Miv  (4.80)

qimv + QMNim wim  QMNim  ∀i ∈ N  m ∈ Mi  (4.81)
v∈V
si1 − Ri ti1 = S0i  ∀i ∈ N  (4.82)

si(m−1) − Ii qi(m−1)v + Ri (tim − ti(m−1) ) − sim = 0
v∈V
∀i ∈ N  m ∈ Mi \{1} (4.83)
SMNi  sim  SMXi  ∀i ∈ N  m ∈ Mi  (4.84)

SMNi  sim − Ii qimv + Ri (T − tim )  SMXi 
v∈V
∀i ∈ N  m = Mi  (4.85)
wim − wi(m−1)  0 ∀i ∈ N  m ∈ Mi \MCi  (4.86)

tim − ti(m−1) − TQi qi(m−1)v + TBi wim  TBi 
v∈V
∀i ∈ N  m ∈ Mi \{1} (4.87)
ximjnv ∈ {0 1} ∀v ∈ V  (i m j n) ∈ Av  (4.88)
wim ∈ {0 1} ∀i ∈ N  m ∈ Mi \MCi  (4.89)
The objective function (4.67) minimizes the total costs. Constraints (4.68)
ensure that each port call is visited at most once. Constraints (4.69)–(4.71) de-
scribe the flow on the sailing route used by ship v. Constraints (4.72) take into
account the timing on the route. Initial time conditions for each ship are de-
fined by constraints (4.73). The time windows are given by constraints (4.74).
If no ship is visiting port call (i m), we will get an artificial start time within the
time windows for a “dummy ship”. These artificial start times are used in the
inventory balances. Constraints (4.75) give the relationship between the binary
flow variables and the ship load at each port call. Initial conditions for the load
quantity and the quantity on board are given in constraints (4.76) and (4.77),
respectively. Constraints (4.78) and (4.79) give the ship capacity intervals at
the port calls for loading and unloading ports, respectively. Constraints (4.80)
and (4.81) are the load limit constraints. All constraints (4.68)–(4.81) so far are
similar to constraints (4.25)–(4.37) for the industrial ship scheduling problem
with flexible cargo sizes in Section 4.1.3. In addition, we have some inventory
constraints for this problem. The inventory level at the first call in each port
is calculated in constraints (4.82). From constraints (4.83), we find the inven-
tory level at any port call (i m) from the inventory level upon arrival at the
port in the previous call (i m − 1), adjusted for the loaded/unloaded quan-
tity at the port call and the production/consumption between the two arrivals.
The general inventory limit constraints at each port call are given in (4.84).
 Ch. 4. Maritime Transportation 249

Constraints (4.85) ensure that the level of inventory at the end of the planning
horizon is within its limits. It can be easily shown by substitution that con-
straints (4.85) ensure that the inventory at time T will be within the bounds
even if ports are not visited at the last calls. One or several of the calls in a
specified port can be made by a dummy ship, and the highest call numbers
will be assigned to dummy ships in constraints (4.86). These constraints reduce
the number of symmetrical solutions in the solution approach. For the calls
made by a dummy ship, we get artificial starting times within the time windows
and artificial stock levels within the inventory limits. Constraints (4.87) pre-
vent service overlap in the ports and ensure the order of real calls in the same
port. A ship must complete its service before the next ship starts its service
in the same port. Finally, the formulation involves binary requirements (4.88)
and (4.89) on the flow variables and port call slack variables, respectively.
This s-ISRP can be solved by the Dantzig–Wolfe (DW) decomposition ap-
proach described in Section 4.2.1, where we have a ship routing and scheduling
problem for each ship and an inventory management problem for each port.
However, if we try to decompose the model directly, it does not separate due to
the starting time tim and the load quantity qimv variables. These variables are
needed in both subproblems that we have here, the routing and the inventory
subproblems. This issue is resolved by introducing new time and quantity vari-
ables, such that we get variables for each (i m v)-combination (timv and qimv )
and each port call (tim and qim ) and introducing coupling constraints to the
problem as follows:
  
(1 − wim ) tim − timv = 0 ∀i ∈ N  m ∈ Mi  (4.90)
v∈V

qim − qimv = 0 ∀i ∈ N  m ∈ Mi  (4.91)
v∈V

Now, the constraint set can be split into three independent groups. The first
constraint group consists of ship constraints and constitutes the routing prob-
lem for each ship where the time windows and load on board the ship are
considered. The ship routing constraints are based on constraints (4.69)–(4.81)
with the starting time, timv , and load quantity, qimv , variables. The port in-
ventory constraints describe the inventory management problem for each port,
and here tim and qim are used in the problem and are based on constraints
(4.74) and (4.80)–(4.87). The remaining constraints are the common constraints
(4.68), (4.90), and (4.91).
As described in Section 4.2.1 we introduce a variable yvr for each of the
feasible combinations of sailing legs to geographical routes, starting times and
load quantities at the port calls, and such a combination is called a ship sched-
ule r, r ∈ Rv . The schedule r includes information about the sailed legs in the
route (Ximjnvr equals 0 or 1), number of visits at port call (Aimvr equals 0 or 1),
the load quantity of each port call (QVimvr ), and the starting time of each port
250 M. Christiansen et al.

call (TVimvr ). No quantity and starting time information is given for “dummy
calls”.
At the ports, it is important to determine the load quantity and starting time
at each call in the port such that the inventory level is within its limits during
the entire planning horizon. Each of the feasible combinations of load quanti-
ties, starting times and number of calls at a port i during the planning horizon
is called a port call sequence s, s ∈ Si . The values of QHims and THims repre-
sent the load quantity and starting time for the port call (i m) in sequence s,
respectively. The value of Wims is 1 if sequence s is not visiting port call (i m),
and from this constant we can find the number of calls at port i. Let variable
zis , i ∈ N , s ∈ Si , be 1, if port i selects sequence s and 0 otherwise.
The resulting master problem becomes:

min Cvr yvr (4.92)
v∈V r∈Rv
subject to
 
Aimvr yvr + Wims zis = 1
v∈V r∈Rv s∈Si
∀i ∈ N  m ∈ Mi  (4.93)
 
QVimvr yvr − QHims zis = 0
v∈V r∈Rv s∈Si
∀i ∈ N  m ∈ Mi  (4.94)
 
TVimvr yvr − THims zis = 0
v∈V r∈Rv s∈Si
∀i ∈ N  m ∈ Mi  (4.95)

yvr = 1 ∀v ∈ V  (4.96)
r∈Rv

zis = 1 ∀i ∈ N  (4.97)
s∈Si
yvr  0 ∀v ∈ V  r ∈ Rv  (4.98)
zis  0 ∀i ∈ N  s ∈ Si  (4.99)

Ximjnvr yvr ∈ {0 1} ∀v ∈ V  (i m j n) ∈ Av  (4.100)
r∈Rv

The objective function (4.92) minimizes the transportation costs. No such

costs exist for the inventory problem, so just the route variables with asso-
ciated costs are present. Unlike usual vehicle routing problems solved by a
DW decomposition approach, the master problem includes additional cou-
pling constraints for the load quantities and starting times to synchronize the
port inventory and ship route aspects. These are given in constraints (4.94) and
 Ch. 4. Maritime Transportation 251

(4.95), respectively. The convexity rows for the ships and ports are given in con-
straints (4.96) and (4.97). The integer requirements are defined by (4.100) and
correspond to declaring the original flow variables ximjnv as binary variables.
In the DW decomposition approach, the port call sequences and ship sched-
ules with least reduced costs in the (minimization) master problem are gener-
ated. This procedure is described in Section 4.2.1 for a maximization problem.
We solve subproblems for each port and each ship, and both types of sub-
problems can be solved by dynamic programming algorithms. Christiansen
(1999) studies a real ship scheduling and inventory management problem
for transportation of ammonia. The overall solution approach is described in
Christiansen and Nygreen (1998a), and the method for solving the subprob-
lems is given in detail in Christiansen and Nygreen (1998b).
In the real problem described by Christiansen (1999), the shipper trades am-
monia with other operators in order to better utilize the fleet and to ensure the
ammonia balance at it’s own plants. These traded volumes are determined by
negotiations. The transporter undertakes to load or unload ammonia within a
determined quantity interval and to arrive at a particular external port within a
given time window. For these external ports, no inventory management prob-
lem exists. This is an example of a shipper operating its fleet in both the
industrial and tramp modes simultaneously.
Another solution approach to the same problem was developed by Flatberg
et al. (2000). They have used an iterative improvement heuristic combined with
an LP solver to solve this problem. The solution method presented consists of
two parts. Their heuristic is used to solve the combinatorial problem of finding
the ship routes, and an LP model is used to find the starting time of service
at each call and the loading or unloading quantity. Computational results for
real instances of the planning problem are reported. However, no comparisons
in running time or solution quality of the results in Flatberg et al. (2000) and
Christiansen and Nygreen (1998a) exist.
At the unloading ports ammonia is further processed into different fertil-
izer products, and these products are supplied to the agricultural market. Fox
and Herden (1999) describe a MIP model to schedule ships from such ammo-
nia processing plants to eight ports in Australia. The objective is to minimize
freight, discharge and inventory holding costs while taking into account the
inventory, minimum discharge tonnage and ship capacity constraints. Their
multiperiod model is solved by a commercial optimization software package.

4.3.2 Inventory routing for multiple products

When there are multiple products, the inventory ship routing problem be-
comes much harder to solve. Until recently this problem was scarcely consid-
ered in the literature. However, Hwang (2005) studied this problem in his PhD
thesis and assumed that the various products require dedicated compartments
in the ship. The problem studied is to decide how much of each product should
be carried by each ship from production ports to consuming ports, subject to
the inventory level of each product in each port being maintained between cer-
252 M. Christiansen et al.

tain levels. These levels are set by the production/consumption rates and the
storage capacities of the various products in each port. The problem is formu-
lated as a mixed-integer linear programming problem with a special structure.
Small test problems are randomly generated and solved. Hwang uses a com-
bined Lagrangian relaxation and heuristic approach to solve the test problems.
In this section, we consider a special case of the multiple inventory routing
problem where several products are produced at several plants located adja-
cent to ports, and the same products are consumed at consuming plants in
other ports. In contrast to the single inventory ship routing problem (s-ISRP)
described in Section 4.3.1, we assume that in the problem considered here, the
shipper does not control and operate the fleet of ships. The transportation is
carried out by ships that are chartered for performing single voyages from a
loading to an unloading port at known cost (spot charters). This means that
the focus of the problem is to optimally determine the quantity and timing of
shipments to be shipped, while the routing of the ships is not an important part
of the problem.
As before, we call the production plants loading ports and the consuming
plants unloading ports. Not all the products are produced or consumed at all
the plants. The plants have limited storage capacity for the products that they
produce or consume. Unlike the s-ISRP discussed in Section 4.3.1, the pro-
duction and consumption rates may vary over time. However, total production
and total consumption of each product are balanced over time. It is therefore
possible to produce and consume continuously at all the plants while the in-
ventories are between their lower and upper limits, given that the products are
shipped from the loading ports to the unloading ports frequently enough. Pre-
vention of interruption in production or consumption at all plants due to lack
of materials or storage space is the main goal of our planning, same as for the
s-ISRP.
Ship voyages have a single loading port and a single unloading port. We
assume that the cost of a voyage between two ports consists of two components,
a fixed set-up cost, and a variable cost per unit shipped that is based on the
distance between the two ports. Further, we assume that there is a sufficient
number of ships of different sizes. Figure 10 illustrates the situation modeled,
where the bold arcs are in the model and the stippled ones are not.
There is uncertainty both in the sailing times and in the production and
consumption rates. This is taken into account by the use of safety stocks in the
inventory planning. If a ship arrives late at a loading port, production may stop
at the plant due to shortage of storage space for the products. To reduce the
possibility of such situations, the storage capacities modeled are less than the
actual capacities. This has the same effect as the use of safety stocks. In our
model we set an upper safety stock level that is below the storage capacity, and
a lower safety stock level that is above a specified lower storage capacity. Any
diversion of the inventory from this band of safety stock limits is penalized, as
illustrated in Figure 11.
 Ch. 4. Maritime Transportation 253

Fig. 10. A multiple product inventory routing problem. The bold arcs are in the model, the stippled
ones are not.

Fig. 11. The inventory level during the planning horizon for one port.
254 M. Christiansen et al.

In Figure 11, we see the inventory level during the planning horizon for one
of the products produced at a loading port. The port is visited twice during
the planning horizon. The production rate is lower than the ship loading rate.
Compared with the s-ISRP where time was continuous, we revert here to using
one day (24 hours) time increments. We measure the transportation time in
days such that all products produced in one day can leave the loading port on
the same day, and all products that arrive at an unloading port during a day can
be consumed on the same day. However, introducing a one day lag between
these operations requires only a minor change in the formulation. Further, we
assume at most one ship sailing per day between any loading and unloading
port pair.
The objective of the model is to find a transportation plan that minimizes
the sum of the transportation cost and the inventory penalties.
In the mathematical description of the problem, let N be the set of ports
indexed by i or j. Divide this set into the subset of loading ports NP and the
subset of unloading ports ND . Let K be the set of products indexed by k, and
let T be the set of periods (days) indexed by t.
The time for sailing from loading port i to unloading port j including the
loading and unloading time is Tij . Rikt is production or consumption of prod-
uct k in port i during day t. These rates are positive in loading ports and
negative in unloading ports.
The inventory information is given by the storage capacities and the safety
stock. The absolute lower and upper storage capacities for product k in port i
are 0 and SMXik , respectively. The lower and upper safety stocks for the same
products in the same ports are SSLik and SSUik . The inventory in the beginning
of the planning horizon for product k in port i is given by SSTik .
Uij represents the maximal capacity/size of a ship that can sail between the
loading port i and unloading port j. Due to the setup cost involved with a voy-
age between two ports the transportation cost will be minimized by using the
largest ship possible. By always using ships of maximal size, the model becomes
simple.
The fixed cost for sailing a ship from loading port i to unloading port j is
represented by CFij , while CVij is the variable cost of shipping one ton of a
product between i and j. The penalty cost per day for each ton of lower (upper)
safety stock shortfall (excess) for product k in port i is CLik (CUik ).
In the mathematical formulation we use the following types of variables: the
binary flow variable xijt , i ∈ NP , j ∈ ND , t ∈ T , equals 1, if a ship sails from
port i on day t to port j, and 0 otherwise. The quantity variable qijkt , i ∈ NP ,
j ∈ ND , k ∈ K, t ∈ T , represents the number of tons of product k that leaves
port i on day t on a ship bounded for port j. The inventory of product k at the
end of day t in port i is given by sikt , i ∈ N , k ∈ K, t ∈ T , while the lower
safety stock shortfall and the upper safety stock excess of product k at the end
of day t in port i are sLikt , i ∈ N , k ∈ K, t ∈ T , and sUikt , i ∈ N , k ∈ K, t ∈ T ,
respectively.
 Ch. 4. Maritime Transportation 255

The mathematical formulation of the multiple product inventory ship rout-

ing problem considered here is as follows:
     
min CFij xijt + CVijk qijkt
i∈NP j∈ND t∈T i∈NP j∈ND k∈K t∈T
  
+ CLik sLikt + CUik sUikt (4.101)
i∈N k∈K t∈T i∈N k∈K t∈T
subject to

qijkt − Uij xijt  0 ∀i ∈ NP  j ∈ ND  t ∈ T  (4.102)
k∈K

sikt − sik(t−1) + qijkt = Rikt 
j∈ND
∀i ∈ NP  k ∈ K t ∈ T  (4.103)

sjkt − sjk(t−1) − qijk(t−Tij ) = Rikt 
i∈NP
∀j ∈ ND  k ∈ K t ∈ T  (4.104)
sikt + sLikt  SSLik  ∀i ∈ N  k ∈ K t ∈ T  (4.105)
sikt − sUikt  SSUik  ∀i ∈ N  k ∈ K t ∈ T  (4.106)
sikt  SMXik  ∀i ∈ N  k ∈ K t ∈ T  (4.107)
qijkt  0 ∀i ∈ NP  j ∈ ND  k ∈ K t ∈ T  (4.108)
xijt ∈ {0 1} ∀i ∈ NP  j ∈ ND  t ∈ T  (4.109)
sikt  sLikt  sUikt  0 ∀i ∈ N  k ∈ K t ∈ T  (4.110)
The objective (4.101) minimizes the sum of the transportation and penalty
costs. Constraints (4.102) together with the binary specifications in (4.109)
force the ship usage variables to be equal to one for the ships in operation,
so that we get the full setup cost for the ship voyages. Constraints (4.103) and
(4.104) are the inventory balances at the loading and unloading ports, respec-
tively, while constraints (4.105) and (4.106) calculate the safety stock shortfall
and excess in the ports. The inventory starting level SSTik is used for sik0 in
(4.103) and (4.104). The upper storage capacity constraints at the ports are
given in (4.107). Finally, the formulation involves binary requirements (4.109)
and nonnegativity requirements (4.108) and (4.110).
This model (4.101)–(4.110) is reasonably easy to understand, but it is hard
to solve because the solution of the linear relaxation will often transport small
quantities to avoid penalty costs and just take the “needed” portion of the fixed
sailing costs. Normally we will have CLik > CUik in unloading ports and the
other way around in loading ports.
Ronen (2002) used a model very similar to (4.101)–(4.110) to plan dis-
tribution from refineries. He presented the model without any upper safety
256 M. Christiansen et al.

stock penalties but mentioned the use of such penalties in the discussion. Con-
straints (4.105) were given as equality constraints with explicit variables for
lower safety stock excess. We get the same variables as surplus variables. For-
mulation (4.101)–(4.110) should make the LP marginally faster to solve. Ronen
(2002) used CPLEX 6.5 to solve his model. For very small cases CPLEX man-
aged to find the optimal solution with a user chosen relative tolerance of 1%
for cutting off nodes in the branch-and-bound tree.
To be able to solve larger problems, we need to generate some cuts that
restrict the number of xijt variables that can be 1, so that many xijt variables
will be fixed to 0 after fixing some to 1.
Ronen (2002) added the following constraints to the model (4.101)–(4.110):

qijkt − xijt  0 ∀i ∈ NP  j ∈ ND  t ∈ T  (4.111)
k∈K
If we look at this as a valid cut, it is usually far from sharp enough. But if the
qijkt variables are scaled such that the ship capacities, Uij , have values slightly
greater than 1, then constraints (4.111) will force all ships branched to be used
to be nearly full. If the cost structure is such that we know that it is optimal
to have all ships nearly full, then we can use (4.111) with scaled qijkt variables
or better with a constant slightly less than Uij in front of the xijt variable. This
might make the problem easier to solve.
In addition to solving the model by use of commercial optimization software
for smaller sized problems, Ronen (2002) presented a cost-based heuristic al-
gorithm to assure that acceptable solutions were obtained quickly.

4.3.3 Other maritime supply chain applications

Reported research of more complex maritime supply chains is scarce. How-
ever, we will briefly refer to some existing studies.
A tactical transshipment problem, where coal is transported at sea from sev-
eral supply sources to a port with inventory constraints was studied by Shih
(1997). The coal is then transported from the port to several coal fired power
plants. The objective is to minimize the procurement costs, transportation
costs, and holding costs. Constraints on the system include company procure-
ment policy, power plant demand, port unloading capacity, blending require-
ments, and safety stocks. The study proposes a MIP model for a real problem
faced by the Taiwan Power Company. Kao et al. (1993) present a similar prob-
lem for the same company. They applied inventory theory to determine an
optimal shipping policy. The underlying inventory model is nonlinear where
the procurement costs, holding costs, and shortage costs are minimized subject
to inventory capacity constraints. Liu and Sherali (2000) extended the prob-
lem described by Shih (1997), and included the coal blending process at the
power plants in the mathematical model. They present a MIP model for find-
ing optimal shipping and blending decisions on an annual basis. The solution
procedure employs heuristic rules in conjunction with a branch-and-bound al-
gorithm.
 Ch. 4. Maritime Transportation 257

In a supply chain for oil, several types of models dealing with logistics are
necessary. Chajakis (1997) describes three such models:
(a) crude supply – models for generating optimal short-term marine-based
crude supply schedules using MIP,
(b) tanker lightering – models of tanker lightering, which is necessary in
ports where draft or environmental restrictions prevent some fully loaded ves-
sels from approaching the refinery unloading docks. Both simulation and MIP
based tools are used, and
(c) petroleum products distribution – a simulation model that was devel-
oped for analyzing products distribution by sea.
However, several legs of the supply chain are not included in these models. In
Chajakis (2000) additional models for freight rate forecasting, fleet size and
mix, and crew planning are discussed.
A planning model for shipments of petroleum products from refineries to
depots and its solution method is described by Persson and Göthe-Lundgren
(2005). In the oil refining industry, companies need to have a high utilization of
production, storage and transportation resources to be competitive. Therefore,
the underlying mathematical model integrates both the shipment planning and
the production scheduling at the refineries. The solution method is based on
column generation, valid inequalities and constraint branching.

4.4 Fleet deployment in liner shipping

Liner shipping differs significantly from the other two types of shipping op-
erations, industrial and tramp shipping, discussed so far in Section 4. However,
also liner shipping involves decisions at different planning levels. Strategic
planning issues were discussed in Section 3.2 for liner fleet size and mix and in
Section 3.3 for liner network design. The differences among the types of ship-
ping operations are also manifested when it comes to routing and scheduling
issues. One main issue for liners on the tactical planning level is the assignment
of vessels to established routes or lines and is called fleet deployment.
As discussed in Section 1, during the last four decades general cargo has
been containerized and we have evidenced a tremendous increase in container
shipping. This increased number of containerships almost always falls in the
realm of liner shipping. Despite this fast growth, studies on deployment in liner
shipping are scarce.
In this section, we want to focus on a fleet deployment problem where we
utilize the different cruising speeds of the ships in the fleet. The routes are
predefined, and each route will be sailed by one or more ships several times
during the planning horizon. Each route has a defined common starting and
ending port. A round-trip along the route from the starting port is called a
voyage.
The demand is given as a required number of voyages on each route without
any explicit reference to the quantities shipped. The fleet of ships is heteroge-
neous, so we can reference quantities implicitly by saying that not all ships
258 M. Christiansen et al.

can sail all routes. Such a specification can incorporate needed ship capacity
together with compatibilities between ships and ports. With information about
the feasible ship-route combinations, we do not need to keep track of the loads
on board the ships. Further, the routes do not need to share a common hub.
Figure 12(a) presents a case with one hub, and Figure 12(b) presents one with
several hubs.
The ships in the available fleet have different cruising speeds. Each ship
is assigned to a single route and is not allowed to switch routes during the
planning horizon. The fleet deployment problem consists of determining which
route each ship is going to sail. The planning goal is to minimize the cost of the
ships.
In the mathematical description of the problem each ship type is repre-
sented by an index v and the set of ship types is given by V . Let R be the
set of routes and Rv the set of routes that can be sailed by a ship of type v. The
elements in both sets are indexed by r.
Vv is the number of ships available of type v. The number of voyages during
the planning horizon along route r for a ship of type v is represented by NVYvr .
Normally this is not treated as an integer number of voyages. The demand is
given by DVr which is the required minimal number of voyages along route r
during the planning horizon. T is the length of the planning horizon in days,
and is one year for the underlying real problem. Svr represents the shipping
season for a ship of type v operating on route r. The shipping season Svr is the
total length of the service periods for ship type v during the planning horizon.
This means that if a ship is allocated to a route, it is operating on that route
during its whole shipping season.
Often, the demand requirement is such that we, for instance, are allowed
to combine 3.7 voyages of one ship with 8.4 voyages of another ship to get a
total of 12.1 voyages to meet a demand of 12 voyages. In such cases, it is not
necessary for NVYvr to be integer. This also gives Svr equal to the time a ship
of type v is available during a year independently of route r.
However, if we want to be sure that each port on route r is visited at least
DVr times during the planning horizon, we need to calculate NVYvr as an in-
teger. Then Svr is calculated as the number of whole voyages multiplied by the
time of one voyage. This is the reason for defining the shipping season for a
ship dependent on the route.
The cost consists of two parts. First, the cost of operating a ship of type v
on route r during the planning horizon is given by CYvr . Secondly, we have
the lay-up cost. The days the ship is out of service for maintenance or other
reasons are called lay-up days. The cost for each lay-up day for a ship of type v
is denoted by CEv .
To make the model similar to the models in the literature, we use the fol-
lowing types of decision variables: the fleet deployment variables, svr , v ∈ V ,
r ∈ Rv , represents the integer number of ships of type v allocated to route r,
and dv , v ∈ V , gives the total number of lay-up days for ships of type v.
 Ch. 4. Maritime Transportation 259

(a)

(b)
Fig. 12. (a) Fleet deployment with nonoverlapping routes and a common hub. (b) Fleet deployment
with nonoverlapping routes and several hubs.

The mathematical formulation of this fleet deployment problem for ships

with different operating speeds and capacities is as follows:
   
min CYvr svr + CEv dv (4.112)
v∈V r∈Rv v∈V
260 M. Christiansen et al.

subject to

svr  Vv  ∀v ∈ V  (4.113)
r∈Rv

NVYvr svr  DVr  ∀r ∈ R (4.114)
v∈V

dv + Svr svr = Vv T ∀v ∈ V  (4.115)
r∈Rv
svr  0 and integer ∀v ∈ V  r ∈ Rv  (4.116)
dv  0 ∀v ∈ V  (4.117)
Here (4.112) is the total cost of sailing the routes with the selected ships
and the cost of the lay-up days. Constraints (4.113) prevent the number of
ships in operation from exceeding the number available, while constraints
(4.114) ensure that each route is sailed at least the required number of voyages
demanded. The lay-up days for each ship type are calculated in constraints
(4.115). Finally, the formulation involves integer and nonnegativity require-
ments on the fleet deployment variables and lay-up variables, respectively.
Powell and Perakis (1997) presented this model using a different notation.
The model was tested on a real liner shipping problem and substantial savings
were reported compared to the actual deployment. Powell and Perakis (1997)
used standard commercial software for the formulation (AMPL) and solution
(OSL) of their model. The example they give has 11 types of ships and 7 routes
with an average number of required voyages just below 20. All their data for
the number of voyages for ships of a given type on a given route are noninteger.
We have assumed here that a ship allocated to a route is just operating on
that route during its whole shipping season, even if that results in more voyages
than required on that route. This means that the model does not allow for a
choice between extra voyages or extra lay-up days.
Constraints (4.115) calculate the total number of lay-up days for each ship
type. It is reasonably easy to remove these constraints from the model by a
reformulation. Since each ship is used only on one route, we can pre-calculate
the number of lay-up days for a ship of type v that is used on route r, before the
optimization and add the corresponding lay-up cost to the annual cost of using
the ship on that route. This also removes variable dv . If we want an integer
number of voyages for each ship, we need to divide the lay-up days calculated
by (4.115) into two parts, one part for the real lay-up days for maintenance,
and one part where the ship only waits for the next planning horizon. The cost
per day for each of these parts may be different, and this difference is most
easily taken care of in a pre-calculation phase.
The formulation (4.112)–(4.117) is a tactical planning model. If we want
to use it in a pure strategic planning situation, we will normally assume that
we can buy or build as many ships as we want of each type. Then constraints
(4.113) will not be binding and the optimization problem decomposes into a
 Ch. 4. Maritime Transportation 261

problem with one constraint, (4.114), for each route after pre-calculating the
lay-up cost and removing (4.115).
The work presented by Powell and Perakis (1997) is an extension and im-
provement of the work in Perakis and Jaramillo (1991) and Jaramillo and
Perakis (1991). In the latter two papers, an LP approach is used to solve the
fleet deployment problem. Manipulation of the results is needed to achieve
integer solutions from the continuous ones, which may lead to a suboptimal
solution and even violation of some constraints.
Section 3.2.2 discussed a strategic fleet size and mix model originally given
by Fagerholt and Lindstad (2000). With a fixed fleet that model becomes a
tactical fleet deployment model.
Earlier fleet deployment studies for assigning vessels to origin–destination
port pairs can be found in Papadakis and Perakis (1989), Perakis and Papadakis
(1987a, 1987b) and Perakis (1985). Various models were presented where both
full and ballast speeds and several additional constraints were considered.

4.5 Barge scheduling

Barges usually operate in protected bodies of water, generally in inland wa-

terways. Barges can be self-propelled or they may be towed by a tugboat, or
pushed by a tugboat. Often multiple barges are combined into a single tow
that is pushed or pulled by a single tug. On the Mississippi River system a
barge can load up to 600 tons and a tow is composed of up to 15 barges. Since
barges operate on inland waterways they must follow the navigable waterway
and therefore their routes are linear like coastal routes or, if there are branches
in the waterway, the routes may have a tree shape. Loaded and/or empty barges
can be added to a tow or dropped off from a tow in ports that are passed-by
along the route of the tow. Barges often have to pass through locks on their
way up or down the river. This complicates their scheduling because they may
have to wait for their turn to pass through a lock, and locks may have limited
hours of operation. Research on barge transportation is scarce. Several papers
discussing fleet design were discussed in Section 3.
Scheduling of barges in inland waterways is similar to scheduling ships with
the additional complications that may be posed by locks. Such additional con-
straints may be very important in barge scheduling, but may be hard to incor-
porate in scheduling models similar to those described in Section 4.1.
Very few works are dedicated to barge scheduling. The initial work on
scheduling barges was provided by O’Brien and Crane (1959) who used simula-
tion to evaluate the impact of various scheduling policies on fleet size and mix
requirements. Schwartz (1968) suggested a linear MIP model for scheduling a
fleet of tugs and barges for the delivery of a given set of cargoes at minimal cost.
The size of the model was far beyond the capabilities of solution algorithms at
that time. A special barge scheduling problem that involves high uncertainty in
timing of activities was discussed by Vukadinovic and Teodorovic (1994) and
later by Vukadinovic et al. (1997). The barges are used to move gravel from a
262 M. Christiansen et al.

dredging site and are moved in tows by pushing tugs. The barge loading and
discharging process is subject to significant uncertainty regarding its timing.
The key decision is the assignment of loaded barges to tugs for a planning
horizon of one day. There is a single loading location and multiple discharg-
ing ports, but a loaded barge is unloaded in a single port. The initial paper
used fuzzy logic to suggest a schedule, and the second one proposed a neural
network that learns from examples and can emulate the dispatcher’s decision
making process.

4.6 Scheduling naval vessels

In contrast to commercial vessels that are usually used to transport one type
of cargo or another, the main mission of naval and coast guard vessels is to
perform assigned tasks at sea. Such tasks may include patrols, training, exer-
cises, law enforcement, search and rescue, and others. In smaller navies, naval
vessels usually stay close to home and return to base frequently. However, in
larger navies, naval vessels may spend extended periods of time at sea, and
have to be resupplied at sea. Naval vessels also spend lengthy periods of time
at port or yards for maintenance, renovation, and training. Usually the major
objective in scheduling naval and coast guard vessels is to assign the available
fleet to a set of specified tasks in a manner that maximizes or minimizes a set of
measures of effectiveness. First we discuss scheduling naval combatants, then
we move to scheduling coast guard vessels, and we close with logistical support
at sea.

4.6.1 Scheduling naval combat vessels

Scheduling an available naval fleet to perform a planned set of activities is a
classical naval application. Such activities may include major operations, exer-
cises, maintenance periods, and other events. Brown et al. (1990) considered
the problem of determining the annual schedule of the US Navy’s Atlantic
Fleet combatants. The goal is to assign ships to events in a manner that meets
all the event requirements and minimizes deviations from ideal schedules for
individual ships. Each event requires a given number of units of particular ves-
sel types and weapon systems. A generalized set partitioning model is used
to solve the problem optimally. Intricate realistic schedule constraints can be
incorporated in the schedule generator.
The same problem is addressed by Nulty and Ratliff (1991), but in a some-
what different manner. An integer programming formulation is developed, but
results in a model of prohibitive size. This fact combined with the qualitative
nature of additional secondary objectives and constraints suggest an interactive
optimization approach. The proposed approach allows the user to generate a
good initial fleet schedule by using network algorithms, and improve the so-
lution by interactively addressing the issues that are difficult to quantify. They
also suggest that the method of Brown et al. (1990) could be used to find a
starting solution for the interactive procedure.
 Ch. 4. Maritime Transportation 263

4.6.2 Scheduling coast guard vessels

A problem that is essentially similar to scheduling naval combatants is faced
by coast guards. However, coast guard vessels stay closer to their home base
and generally do not have to be resupplied at sea. A typical problem is to
schedule a fleet of coast guard cutters (vessels) to perform a set of assign-
ments. Each assignment has a given duration, and a desired number of cutters.
Such a problem was addressed by Darby-Dowman et al. (1995). In their model
the requirements are treated as goals, and not meeting a goal is allowed but
penalized. The problem is solved by using a set partitioning model. The objec-
tive is to select the set of schedules that provides a solution that is as close to
meeting the requirements as possible. The system was originally intended for
use in determining operational schedules. However, additional use to address
strategic issues such as future operating policy and fleet mix arose during the
project.
A system for solving similar scheduling problems for the US Coast Guard
cutters was presented by Brown et al. (1996). They developed costs and penal-
ties for the model to mimic the motives and rules of thumb of a good scheduler.
The objective was to minimize the costs, and the elastic MIP model was solved
on a personal computer within a few minutes.
Another type of vessel scheduling problem faced by a coast guard is routing
and scheduling buoy tenders. These vessels are used to service and maintain
navigational buoys. Cline et al. (1992) describe a heuristic algorithm for rout-
ing and scheduling US Coast Guard buoy tenders. Each buoy has a service
time window dictated by the planned maintenance schedule. Since each tender
has the sole responsibility for servicing its set of buoys, the problem is decom-
posed into a set of traveling salesman problems with time windows, one for
each tender. They used a best-schedule heuristic to solve the problem.

4.6.3 Logistical support at sea

Supporting naval vessels at sea poses additional challenges. Schrady and
Wadsworth (1991) described a logistic support system that was designed to
track and predict the use of consumable logistic assets (fuel, ordnance) by a
battle group. The system was tested during fleet exercises and was quite suc-
cessful. Williams (1992) dealt with the replenishment of vessels at sea. He
presented a heuristic algorithm to replenish a group of warships at sea while
the ships carry out their assignments. The heuristic rules were derived from
replenishment experts and are based on experimentation.

4.7 Ship management

Several topics fall into the category of ship management and we shall discuss
briefly the following ones: crew scheduling, maintenance scheduling, position-
ing of spare parts, and bunkering. Deep-sea vessels spend extended periods of
time at sea and the crew lives on board the ship. Short-sea vessels make fre-
quent port calls and the crew may live on shore. This difference has significant
264 M. Christiansen et al.

impact on ship management issues. Crew scheduling for deep-sea vessels is not
a major issue. Crew members spend months on the vessel and then get a long
shore leave. For short-sea vessels the crew may change frequently, and crew
scheduling may be an issue. A special type of such crew scheduling problem is
presented by Wermus and Pope (1994).
Numerous mechanical and electrical systems are installed on board a ship
and they require maintenance. A ship is usually scheduled once a year for
maintenance in a port or a shipyard, and once every several years a ship is sur-
veyed by its classification society in a shipyard. However, some maintenance
is required between such planned maintenance periods, and this includes both
routine/preventive maintenance, and repair of breakdowns, at least temporar-
ily, till the ship reaches the next port. On-board maintenance is usually done
by the crew, but the shrinking size of crews reduces the availability of the crew
for maintenance work. A large ship may have less than two dozen seamen on
board, and that includes the captain and the radio officer. This limited crew
operates the ship around the clock. A specialized analysis of repair and re-
placement of marine diesel engines was presented by Perakis and Inozu (1991).
In order to facilitate maintenance a ship must carry spare parts on board. The
amount of spare parts is determined by the frequency of port calls and whether
spares and equipment are available in these ports. Large and expensive spares
that cannot be shipped by air, such as a propeller, may pose a special problem,
and may have to be prepositioned at a port or carried on board the vessel.
A ship may consume tens of tons of bunker fuel per day at sea, and there
may be significant differences in the cost of bunker fuel among bunkering
ports. Thus one has to decide where to buy bunker fuel. Sometimes it may
be worthwhile to divert the ship to enter a port just for loading bunker fuel.
The additional cost of the ship’s time has to be traded off with the savings in
the cost of the fuel.

5 Operational planning

When the uncertainty in the operational environment is high and the sit-
uation is dynamic, or when decisions have only short-term impact, one re-
sorts to short-term operational planning. In this section we discuss operational
scheduling where only a single voyage is assigned to a vessel, environmental
routing where decisions are made concerning the next leg of the voyage, speed
selection, ship loading, and booking of single orders.

5.1 Operational scheduling

The demarcation between tactical and operational scheduling in industrial

and tramp shipping is fuzzy, and therefore Section 4.1 considered both plan-
ning levels. However, there are some situations that can be placed solely on
 Ch. 4. Maritime Transportation 265

the operational planning level and they are discussed here. In certain circum-
stances it is impractical to schedule ships beyond a single voyage. This happens
when there is significant uncertainty in the supply of the product to be shipped,
or in the demand for the product in the destination markets. The shipped
product may be seasonal and its demand and supply may be affected by the
weather. These factors contribute to the uncertainty in the shipping schedule.
Citrus fruit is an example of such a product. This is a highly seasonal product
that is shipped in large quantities over several months a year, and may require
refrigerated vessels. The shipper has to assure sufficient shipping capacity in
advance of the shipping season, but does not know in advance the exact tim-
ing, quantities, and destinations of the shipments. The shipper, who owns the
cargo, does not have return cargoes for the ships, so the ships are hired under
contracts of affreightment or spot charters, and generally do not return to load
a second voyage. Thus every week the shipper has ships available for loading
in the producing area and either a load is assigned to each ship or demurrage
is paid for the ship. Based on product availability, demand projections, inven-
tory at the markets, and transit times, the shipper builds a shipping plan for the
upcoming week, and has to assign the planned shipments to the available fleet
at minimal cost. Usually the contract of most vessels hired for a single voyage
confines them to a range of unloading ports. In some operations a vessel may
unload in more than one port, and the requirement of a destination port may
exceed the size of the largest vessel and can be split among several vessels.
Ronen (1986) discussed such an operational scheduling problem, presented
a model and a solution algorithm that provided optimal solutions to smaller
size problems, and heuristics for solving larger problem instances. Later Cho
and Perakis (2001) suggested a more efficient formulation to the same problem
that is a generalized version of the capacitated facility location problem.

5.2 Environmental routing

Ships navigate in bodies of water and are exposed to a variety of envi-

ronmental conditions, such as: currents, tides, waves, and winds. Recognizing
these conditions is the first step toward selecting routes that mitigate their ef-
fects, or even take advantage of them. Generally, when a ship moves between
two ports it has to select its route through the body of water. However, such
a choice is very limited in coastal and inland waterway navigation. Proper se-
lection of the route may assure on-time arrival at the destination port, or even
shorten the time of the passage and reduce its cost. The terms environmental
routing and weather routing are often used interchangeably although the sec-
ond one is a subset of the first. Weather is part of the environment in which
ships operate, and it affects the waves encountered by ships. We confine our
short discussion to the impact of waves and ocean currents. Material concern-
ing tides and winds can be found in the naval architecture, navigation, and
meteorology literature.
266 M. Christiansen et al.

5.2.1 Waves
Waves may have a significant impact on route selection. In order to take
waves into account one has first to know their height and direction along
the contemplated route as a function of location (x and y coordinates). Such
knowledge may allow selection of the route and of power setting that minimize
the transit time. However, the waves’ height and direction may change over
time, and may not be known in advance. Papadakis and Perakis (1990) analyzed
a minimal time vessel routing problem under stationary conditions that is ap-
propriate for relatively short passages. Given wave height and wave direction as
a function of location, select the route and power setting of the vessel that min-
imize the transit time. Local optimality considerations combined with global
boundary conditions resulted in piecewise continuous optimal policies. They
used variational calculus and optimal control theory in their analysis. Perakis
and Papadakis (1989) extended their analysis of the minimal time vessel rout-
ing problem to a time-dependent environment, where the sea condition at any
point changes over time. This allows analysis of longer passages. In addition
they considered voyages consisting of multiple legs with port calls of known
length between the legs. Although they provide a numerical example, no esti-
mates of potential benefits (or savings) are available. However, they show that
when the objective is to minimize transit time “wait for a storm to pass” policy
is never optimal. Instead “one should go ahead under the maximum permissi-
ble power setting  ”.

5.2.2 Ocean currents

In most oceans there are regular currents that ships may be able to ex-
ploit for faster passage. Lo et al. (1991) estimated that by exploiting ocean
currents the world commercial fleet could reduce its annual fuel costs by at
least $70 million. They also provide anecdotal evidence that some operators
try to take advantage of prevailing currents. However, this is easier said than
done. Ocean currents are not constant but rather change over time. Thus get-
ting reliable timely information regarding the ocean current at the location
of a vessel poses a major obstacle. Satellite altimetry may provide timely re-
liable estimates of dynamic current patterns that are necessary for routing a
vessel through such currents. McCord et al. (1999) took a closer look at poten-
tial benefits of such data. Their study uses dynamic programming to optimize
ship routes through estimated current patterns in a dynamic area of the Gulf
Stream. They conclude that elimination of data bias and present sampling lim-
itations can produce about 11% fuel savings for a 16-knot vessel. They found
that the contribution of such routing is much better on with-current voyages
than on counter-current voyages. The major question is whether there is a suf-
ficient market to justify development of a system for collection of the necessary
data.
Environmental routing is complicated by the complexity of the continuous
dynamic environment in which it takes place, and the lack of the necessary
 Ch. 4. Maritime Transportation 267

timely reliable data. Due to these reasons environmental routing seems to be

in its infancy and is a fertile ground for further research.

5.3 Speed selection

A ship can operate at a speed slower than its design speed and thus sig-
nificantly reduce its operating cost. However, a ship must maintain a minimal
speed to assure proper steerage. For most cargo vessels the bunker fuel con-
sumption per time unit is approximately proportional to the third power of the
speed (the consumption per distance unit is proportional to the second power
of the speed). Thus, reducing the speed by 20% reduces the fuel consumption
(per time unit) by about 50% (or by about 36% for a given sailing distance).
When bunker fuel prices are high the cost of bunker fuel may exceed all other
operating costs of the ship. Thus there may be a strong incentive to steam at
slower speed and reduce the operating costs. In the wake of the high fuel price
during the 1970s, Ronen (1982) presented three models for the determination
of short-run optimal speed for different types of legs:
• an income generating leg,
• a positioning (empty/ballast) leg, and
• a leg where the income depends on the speed.
When one widens the horizon beyond a single vessel, the perspective may
change. A fleet operator that controls excess capacity can reduce the speed of
the vessels and thus reduce the effective capacity of the fleet, instead of laying-
up, chartering-out or selling vessels.
Under various operational circumstances a scheduler has to assign an avail-
able fleet of vessels to carry a specified set of cargoes among various ports.
Often cruising speed decisions may be an inherent part of such fleet schedul-
ing decisions. Cruising speed decisions affect both the effective capacity of the
fleet and its operating costs.
Under a contract of affreightment (COA) a ship operator commits to carry
specified amounts of cargo between specified loading port(s) and unloading
port(s) at a specific rate over a specific period of time for an agreed upon
revenue per delivered unit of cargo. The term fleet deployment is usually used
for ship scheduling problems associated with liners and with COAs, because
the vessels are essentially assigned to routes that they follow repeatedly, and
the deployment decisions cover longer terms. Perakis and Papadakis (1987a,
1987b) determined the fleet deployment and the associated optimal speed,
both loaded and in ballast, for ships operating under a COA between a sin-
gle loading port and a single unloading port. A more comprehensive version
of this problem was later dealt with by Papadakis and Perakis (1989). They ex-
panded the problem to address multiple loading ports and multiple unloading
ports, but still assumed that each ship returns in ballast to its loading port after
unloading its cargo. They used nonlinear programming to determine the vessel
allocation to the routes and their cruising speed, both loaded and in ballast.
268 M. Christiansen et al.

Tramp and industrial operators usually face shorter term ship scheduling
problems. A set of cargoes has to be carried by the available fleet, and if the
fleet has insufficient capacity some cargoes may be contracted out. The cruis-
ing speed of the vessels in the available fleet can be an inherent part of the
scheduling decisions. Bausch et al. (1998) and Brown et al. (1987) addressed
this situation, and in their work the cruising speed was determined simultane-
ously with the schedule. Whereas the last two papers had hard time windows
for loading and unloading the cargoes, Fagerholt (2001) considered also soft
time windows, a situation that allows more flexibility in determining the cruis-
ing speed of the vessels, and may result in a lower cost schedule.
In addition to cost and schedules, short-term cruising speed decisions should
take into account also the impact of the destination port operating times. If the
destination port is closed over the weekend (or at night) there is no point arriv-
ing there before the port opens. Thus reducing the cruising speed and saving
fuel makes sense. In the case where cargo-handling operations of a vessel that
started when the port was open continue until the vessel is finished, even after
the port closes, it may be worthwhile to speed up and arrive at the destination
port to start operations before it closes. A more detailed discussion of these
tactics is provided in Section 6.2.

5.4 Ship loading

A ship must be loaded in a safe manner in order to prevent loss of the ship
or damage to the cargo. Ships are designed with certain types of cargo in mind.
A crude tanker is designed to carry crude oil, and a containership is designed to
carry containers. A ship floats on water and its stability must be assured during
passage as well as in port. Ballast tanks are built into the hull of a ship in order
to help maintain its stability by filling them with seawater. When a ship is full
with cargo of a uniform density for which it is designed, such as crude oil or
iron ore, usually there are no stability problems. Stability problems arise when
(a) a ship is partially loaded, then the weight distribution of the cargo must be
properly planned and monitored, both while sailing at sea and during loading
or unloading operations in port, or (b) the cargo is not properly secured and
may shift during passage, for example, liquid bulk cargo may slosh in partially
empty tanks, or (c) when the ship is fully loaded with nonuniform cargo, such
as containers or general cargo. In such a case an improper weight distribution
of the cargo may result in excessive rolling or pitching that may lead to loss
of the ship. In extreme cases improper weight distribution may cause excessive
structural stress that may lead to break up of the ship.
Ship stability has several dimensions. The Trim of a ship is the difference
between the forward and aft draft, and must remain within a narrow range.
There also should be balance between weight of the cargo on the port (left)
side and the starboard (right) side of the ship so it will remain horizontal. The
center of gravity of the ship should not be too high in order not to make the
 Ch. 4. Maritime Transportation 269

ship “top heavy” and easy to roll, and not too low so the ship will not snap back
too fast from a roll which may cause on-deck cargo to break loose.
The more complex ship loading problems are encountered in loading con-
tainerships. Not only the stability of the vessel has to be assured but also the
efficiency of cargo handling operations in the current and following ports must
be taken into account. Containers have different weights and that may affect
the vessel stability. Due to the design of containerships access to a specific
container may be obstructed by other containers stowed on top of it. Thus con-
tainer shifting may be necessary to unload a specific container. Therefore, in
order to minimize future container shifting operations one has to take into ac-
count the destination port of the loaded containers when one decides where
to load them onboard the vessel. Moreover, one also has to consider the des-
tination ports of the containers that will be loaded in following ports of call,
and some of these containers may not even be booked yet. There may also be
containers stuffed with dangerous goods. Such containers impose additional
constraints due to spatial separation requirements.
The focus of research on ship loading has been on loading container ships.
A good description of the various considerations involved in containership
loading is provided by Martin et al. (1988). They developed heuristics that em-
ulate strategies used in manual load planning and showed some improvements
in materials handling measures.
Avriel et al. (1998) focused on minimizing container shifting. They formu-
lated a binary linear program for the container stowage planning problem that
minimizes the number of container shifting operations. Since the problem is
NP-complete they designed a “suspensory heuristic” to achieve a stowage plan.
Their work is of limited applicability because it assumes away stability and
strength requirements, accommodates only one size of containers, and ignores
hatch covers.
A comprehensive approach for planning container stowage on board con-
tainerships is provided by Wilson and Roach (2000). Their objective is to find
a stowage plan that assures that no ship stability or stress constraints are vio-
lated, and minimizes container shifts (re-handles). Additional considerations
are reduction of the ballast required by the vessel and efficient use of cranes
when loading and unloading. Wilson and Roach described a computerized
methodology for generating commercially viable stowage plans. All character-
istics of the problem are considered, but optimality is not necessarily sought.
Their stowage planning process is broken down into two phases, (a) “strategic
planning” where “generalized” containers are assigned to “blocks” of cargo
space, and (b) “tactical planning” where specific containers are assigned to
specific slots within the blocks determined earlier. This approach significantly
reduces the combinatorial complexity of the problem. Their objective consists
of a dozen different criteria that are assigned weights. The strategic planning
phase uses a branch-and-bound search, and the tactical planning phase uses a
tabu search. They tested their methodology on commercial data for a 688 TEU
vessel with a mix of container sizes and types, and four destination ports. Com-
270 M. Christiansen et al.

mercially viable solutions were received in a couple of hours on a 166 MHz

computer. These solutions were comparable with those generated by experi-
enced human planners. However, it takes a human planner several days to get
such solutions.
A similar two-stage approach is used by Kang and Kim (2002) to gener-
ate container stowage plans. In the first stage they assign containers to holds
for each port separately by solving a problem similar to a fixed charge trans-
portation problem using a heuristic based on the transportation method. In the
second stage they assign containers to slots for each hold separately using a tree
search procedure. Since the first stage is done for one port at a time the result-
ing stowage plan may be problematic. Therefore they iterate between the two
stages to improve the plan. They tested their approach on randomly generated
problems and compared their results to a couple of earlier suggested mod-
els. However, they admit the limited applicability of their approach because it
considers only one size of containers (40 ), and does not consider refrigerated
containers or ones with hazardous materials.
The container stowage planning problem is very complex and we are far
from finding optimal solutions, or even agreeing on the components of the ob-
jective function. The related problem of stowage sequencing, which represents
the port’s perspective, is discussed at length in the chapter by Crainic and Kim
(2007).

5.5 Booking of single orders

An important operational problem in commercial shipping is booking of sin-

gle orders. Since a shipper expects an acceptance/rejection decision on a single
cargo request more or less immediately, for the shipping company the problem
consists of deciding whether to accept a single cargo or not. This problem is
somewhat different between liner and tramp/industrial shipping. In liner ship-
ping, where a single cargo is usually a small fraction of the vessel’s capacity, it is
usual to accept a cargo if there is space available on the given ship line, and to
reject or suggest another time of departure if not. However, sometimes it may
not be profitable to accept a cargo even if there is space available, as there may
appear requests for better paying cargoes later on. This problem of stochastic
optimization in liner shipping is rarely dealt with in the literature. The authors
are aware of only a single reference on the subject, and it is a rather out-dated
conference contribution (Almogy and Levin, 1970).
In tramp shipping it is also usual to accept a single cargo request if the plan-
ner is able to find space available. To see if there is space available, reschedul-
ing the whole fleet with the existing cargo commitments together with the new
optional cargo may be necessary because a single cargo may take a large share
of a vessel’s capacity, or even be a full shipload. This is thoroughly discussed in
Section 4.1.5. Industrial shipping is similar in this respect to tramp. However,
also in tramp shipping, as for liner shipping, it may sometimes be advantageous
 Ch. 4. Maritime Transportation 271

not to accept a single cargo request as more profitable cargoes may appear
later. The authors are not aware of any published work on this aspect.

6 Robustness in maritime transportation

As discussed in previous sections, there are many uncertain factors in the

ocean shipping industry resulting in delays and lack of timely fulfillment of
plans. Therefore, in order to encourage trust in the planning process, it may
often be important to consider robustness in optimization models used for
planning. Despite this, models that have been developed for the shipping in-
dustry only rarely deal with these aspects.
In this section we discuss a few problems from the shipping industry where
uncertainty and robustness play important roles, as well as approaches for
achieving more robust solutions. It should be emphasized that this section does
not present a comprehensive overview but rather provides several examples. In
Section 6.1 we concentrate on strategic planning problems, while tactical and
operational planning problems are considered in Sections 6.2 and 6.3, respec-
tively. Section 6.4 discusses optimization and persistence.

6.1 Strategic planning and uncertainty

The most important strategic planning problem for all shipping segments
(industrial, tramp, and liner) is probably fleet sizing and composition. How-
ever, the quality of decisions regarding this aspect is strongly influenced by
many uncertainties, probably much more than decisions for any shorter plan-
ning horizon. There are several major reasons for this uncertainty:
• The long time horizon that these decisions span, which can be several
years. In some cases, when the decision involves building new ships, it
may span up to 20–30 years.
• Demand for shipping is a derived demand. It depends on the level of
economic activity, prices of commodities, and other factors.
• There is a significant time lag between changes in demand for maritime
transportation and the corresponding adjustments in the capacity of
such services.
During such a long time horizon one will experience major unpredictable
fluctuations both in the demand for shipping services and on the supply side.
These factors are highly interwoven. For instance, if demand for transport
services within a given market segment increases, we would probably see an
increase in both freight rates and ship prices, and the same is true in the oppo-
site direction.
Another important strategic decision that is relevant to all shipping seg-
ments is whether a shipping company should accept a long-term contract or
not. In such a long-term contract, the shipping company is typically committed
272 M. Christiansen et al.

to carry a specific quantity more or less evenly distributed over the contract
period, and receives a given revenue per unit of cargo lifted. Also here, the
decision should be made only after cautious consideration (or speculation) re-
garding the direction that the market will take in the future. If, for instance, the
spot market experiences a boost and the actual freight rates increase it would
be unfortunate to have most of the fleet tied up in contracts at lower rates. On
the contrary, if the market dips, it would be advantageous to have a substantial
contractual coverage, in order to ensure both income and engagement for the
ships.
There are different approaches for handling uncertainty and robustness,
such as:
• simulation,
• re-optimization for different scenarios or input parameters,
• adding slack to the input parameters (e.g., service speed),
• deterministic models that incorporate penalties, and
• stochastic optimization models.
Simulation is a simple approach that is used to consider stochastic con-
ditions and uncertainty. There are some examples where simulation models
have been used for strategic planning purposes in the shipping industry, see,
for instance, Darzentas and Spyrou (1996), Richetta and Larson (1997), and
Fagerholt and Rygh (2002).
Another simple approach for considering uncertainty is to make several runs
with an optimization model for different scenarios. In this way, one can decide
what is the optimal decision for a given scenario or for a given set of input
parameters. The problem in using this method is that solutions are often not
robust and are strongly affected by the specific set of values used for the input
parameters. Since flexibility is not built into the plans, extreme solutions are
often produced.
Stochastic conditions like the ones mentioned above and other ones can
also be approached both by deterministic and stochastic optimization mod-
els. An example of using deterministic optimization models with penalties to
achieve more robust solutions is discussed in the next section for a tactical ship
scheduling problem. To the authors’ best knowledge there are no published
papers where stochastic optimization models are used for strategic planning
in the shipping industry. The only one discussing the issue is by Jaikumar and
Solomon (1987), where a model for determining the minimum number of tugs
needed to move barges between ports on a river is presented. They discuss how
their model can be extended to incorporate stochastic demands.

6.2 Robust tactical planning

In Section 4 we presented tactical planning problems and models for the

different shipping segments. However, the models presented there and the so-
lutions that can be obtained from them do not handle the uncertainty and
 Ch. 4. Maritime Transportation 273

robustness aspects. Several unpredictable factors influence the fulfillment of

plans and should often be considered in the planning process. The two most
important are probably:
• weather conditions that can strongly influence the sailing time, and
• port conditions, such as strikes and mechanical problems that can af-
fect the time in port.
A ship may often have to reduce its speed in bad weather. This may result
in late arrival for the next planned cargo. In such cases the planner often has
to reschedule the whole fleet. If the planner has built in enough slack in the
schedule, the planned schedule may still be valid. However, since ships have
high costs, very little slack is usually built into their schedules.
In some cases, ships may require high tide to get into the port fully loaded.
In other cases empty barges may not be able to pass under bridges at high
tide. In short-sea shipping applications where sailing times are short relatively
to port times, and tides may have a significant impact on port access, a small
delay may be amplified due to additional waiting for high tide. Many ports are
also closed for cargo handling operations during nights and weekends. Cargo
handling time that is longer than one working day of the port will span multiple
days. This means that the ship will stay idle much of the time in port, and the
total time in port will depend on the ship’s arrival time.
Consider the following example. A ship has to load a cargo at a specified
port. The loading time window starts on Wednesday at 8:00 and ends on the
next Monday at 24:00. The operating hours of the port are between 8:00 and
16:00 from Monday to Friday. It takes 12 operating hours to load the cargo.
Figure 13 shows the necessary time in port as a function of the arrival time of
the vessel. We see that the total time spent in port varies from 28 to 92 hours,
depending on the arrival time. Twenty eight hours is the minimal time spent in
port, while 92 hours is the maximal time and includes a lot of idle time during
the weekend.

Fig. 13. Time spent in port as a function of arrival time.

274 M. Christiansen et al.

A ship arriving on Wednesday morning at 8:00 will be loading for 8 hours

on the first day and 4 hours on the next day. This gives a total of 28 hours
in port. In the other extreme, take a ship arriving at 15:00 on Thursday. It
loads for one hour on that day, stays idle for 16 hours during Thursday night
and continues loading on Friday for 8 hours, but she does not finish loading
before the port closes for the weekend. It has to continue loading on Monday
morning at 8:00 and finishes at 11:00. This means that the ship stays idle in
port for 64 hours during the weekend, giving a total of 92 hours in the port. It
should be emphasized that in practice it may often be possible to negotiate a
few hours extension to the loading/unloading operations, usually at a cost.
In these cases, a delay due to bad weather or port conditions may have even
stronger effect than in other cases, as the delay may result in the ship stay-
ing idle in port during weekends. Christiansen and Fagerholt (2002) deal with
such a problem. There, a deterministic solution method for making the sched-
ules robust is presented. Their solution method is based on the set partitioning
approach described in Section 4.2.2. However, to ensure schedules that are
robust the concept of risky arrival is introduced. A risky arrival is defined as
a planned arrival time in port that with only a moderate delay will result in
the ship staying idle during a weekend. In order to reduce the number and
magnitude of risky arrivals for a fleet schedule, Christiansen and Fagerholt
(2002) calculate a penalty cost depending on how ‘risky’ the arrival time is.
This penalty cost is calculated during the a priori schedule generation and is
added to the other cost elements in the objective function in the set partition-
ing model. The computational results show that the planned fleet schedule’s
robustness can be significantly increased at the sacrifice of only small increases
in transportation costs.
We can also find a few other contributions within ship scheduling where
penalty costs are used in connection with time windows. In Fagerholt (2001),
hard time windows are extended to soft ones by allowing late or early ser-
vice, though at a penalty cost. Christiansen (1999) studies a combined ship
routing and inventory management problem described in Section 4.3.1. The
transported product is produced in some port factories and consumed in oth-
ers. At all factories there are hard inventory limits for the transported product.
In order to reduce the possibility of violating the inventory limits at the port
factories Christiansen and Nygreen (2005) introduce an additional pair of soft
inventory limits within the hard ones. Thus the soft inventory limits can be vio-
lated at a penalty, but it is not possible to exceed the stock capacity or to drop
below the lower inventory limit. They show that the soft inventory constraints
can be transformed into soft time windows.
Another problem regarding uncertainty and robustness in ship routing and
scheduling is that in some cases the planner knows the loading port but the
unloading port is not known at the time of loading. Sometimes just a geograph-
ical region is given for unloading, and the particular unloading port is specified
after the voyage has started. In these cases the planner has several practical
options:
 Ch. 4. Maritime Transportation 275

• the planner can, based on his or her experience, make a qualified guess
regarding the unloading port and use it for planning,
• use a port that is more or less in the middle of the specified unloading
area as an “average”,
• plan for worst-case, i.e., use the port that is farthest away in the area
(e.g., farthest up in the river), and
• run different scenarios regarding the different optional unloading ports
to see how the different alternatives affect the schedule.

6.3 Robust operational planning

Also operational problems in maritime transportation may pose robustness

issues. Delays due to tides and restricted opening hours in ports, as discussed in
the previous section, can often be regarded as operational ones. How to handle
such delays when they occur is often referred to as “disruption management”.
Typically for shipping, it is often possible to increase the ship’s speed to some
extent when a delay occurs. However, this comes at the sacrifice of much higher
fuel consumption, see Section 5.3 on speed selection. Sometimes, it may also
be possible to increase the loading or unloading rate with a proper incentive.
The problem of whether to accept a single cargo request or not is also an
operational problem since the potential customer often requires an answer im-
mediately, see Section 5.5. In practice, a cargo is often accepted if there is
available capacity. However, accepting a new cargo may restrict the possibili-
ties for taking a more profitable cargo that becomes available in the market
later. Therefore, it could be advantageous to introduce the concept of sto-
chastic optimization to such problems. The authors are not aware of such
contributions.

6.4 Persistence

Schedules have often to be changed due to unforeseen delays, changes in

requirements or other events. In such circumstances it may be highly desirable
to minimize changes to already published schedules. Thus, necessary changes
in the schedule of one vessel should have a minimal effect on the schedule of
other vessels. Optimization models have a well-deserved reputation for am-
plifying small input changes into drastically different solutions. A previously
optimal solution may still be nearly optimal in a new scenario and manage-
rially preferable to a dramatically different solution that is mathematically
optimal. Optimization models can be stated so that they exhibit varying de-
grees of persistence with respect to previous values of variables, constraints, or
even exogenous considerations. Brown et al. (1997a) discuss these aspects of
optimization and persistence.
In another paper by Brown et al. (1997b), the persistence aspect is consid-
ered when optimizing submarine berthing plans. Once in port, submarines may
276 M. Christiansen et al.

be shifted to different berthing locations to allow them to better receive ser-

vices that they require, or to clear space for other shifted vessels. Submarine
berth shifting is expensive, labor intensive and may be potentially hazardous.
Brown et al. (1997b) present a mixed-integer programming model for this
berth planning problem with a planning horizon of 1–2 weeks. Once a berthing
plan has been approved, changes are inevitable due to delays, changed requests
for services, and early arrival of inbound submarines. An optimization model
that only minimizes the costly berth shifts is not appropriate in this situation,
because it can amplify minor modifications in service requests into wholesale
revisions in the approved berthing plan. Revisions to the plan and the disrup-
tions they bring must therefore be controlled to encourage trust in the planning
process. Therefore Brown et al. (1997b) have incorporated a persistence in-
centive into the mixed-integer programming model that results in a decreased
number of changes in previously published plans.

7 Perspectives and future research

As mentioned in the Introduction, demand for maritime transport services is

increasing consistently, and there are no signs that the world economy will rely
less heavily on maritime transport in the future. In this section we shortly dis-
cuss some trends in ocean shipping that will probably influence both the need
for optimization-based decision support systems for maritime applications, and
the shipping industry’s acceptance of and benefits from such systems. We also
wish to point out trends that result in a need for researchers to pay attention
to new problem areas in maritime transportation. The focus is on applications
within ship routing and scheduling. Trends in the land-side of maritime trans-
portation operations are discussed in the Perspectives section of Crainic and
Kim (2007). There may be additional trends, but these are the ones that we
deem to be the primary ones, and that may have significant impact on the var-
ious aspects discussed in this chapter. A more detailed discussion of current
trends in ship routing and scheduling is provided in Christiansen et al. (2004).

7.1 Mergers, acquisitions, and collaborations

During the last couple of decades we have witnessed consolidation in the

manufacturing sector resulting in bigger actors on the demand side for mar-
itime transport services. This has given the shippers increased market power
compared to the shipping companies, resulting in squeezed profit margins for
the shipping companies. In order to reduce this imbalance, there have been
many mergers among shipping companies in the last decade. Many shipping
companies have entered into pooling and collaboration efforts in order to in-
crease their market power and gain flexibility in the services that can be offered
(see Sheppard and Seidman, 2001). In such collaboration, a number of ship-
ping companies bring their fleets into a pool and operate them together. The
 Ch. 4. Maritime Transportation 277

income and costs are distributed among the different shipping companies ac-
cording to certain rules that have been agreed upon. The split of income and
costs is an intriguing topic for research.
Traditionally, scheduling in maritime transportation has been done man-
ually by pencil and paper, based on the planners’ knowledge and experi-
ence. The above trends of mergers and pooling collaborations result in larger
controlled fleets. This means that it becomes much harder to determine a
fleet schedule only by manual planning methods. Therefore, the need for
optimization-based decision support systems has increased and will probably
continue to increase in the future.

7.2 New generation of planners

Decision-makers and planners in the shipping industry are traditionally ex-

perienced, often with a sea-going background. As the fleets become larger,
the planning problems focused on in this chapter become much harder to
handle by manual methods. Despite this, planners are often very skeptical
of computers in general and of optimization-based decision support systems
in particular. However, in recent years we have seen that shipping compa-
nies have started employing planners with less practical but more academic
background. This new generation of planners is more used to computers and
software, and therefore is often much more open to new ideas such as using
optimization-based decision support systems for the different applications in
maritime transportation. Even though there is still a gap to bridge between
researchers and planners in the shipping industry, we expect more willingness
and interest from the ocean shipping industry to introduce such systems in the
future.

7.3 Developments in software and hardware

The fast technological development in computers and communications also

weighs heavily for the introduction of optimization-based decision support sys-
tems in shipping companies. Many earlier attempts failed due to restricted
computer power, making it hard to model all the important problem charac-
teristics and to facilitate a good user interface. However, today’s computers
enable an intuitive user interface to be implemented, something that is cru-
cial for acceptance by the planners. In addition, there have been significant
algorithmic developments. This, together with advances in computing power,
has made it feasible to find good solutions to hard problems in a reasonable
amount of time.

7.4 Shift from industrial to tramp shipping

Looking at the literature review on ship routing and scheduling presented

by Christiansen et al. (2004), we observe that most contributions are in indus-
trial shipping, while only a few are in the tramp market. In industrial shipping
278 M. Christiansen et al.

the shipper controls the cargo and the fleet of ships. The purpose of an in-
dustrial operation is usually to provide the required transportation services for
the organization’s cargo requests at minimum cost. Industrial shipping is prac-
ticed by large extracting and manufacturing companies that have their own
division that controls a number of ships for the transportation of their own
cargoes. However, in recent years this has changed. Many such companies are
now focusing on their core business and have outsourced other activities like
transportation to independent shipping companies. Therefore, the emphasis
has shifted somewhat from industrial to tramp shipping. Increasing global com-
petition results in shifting industrial shipping operations from being considered
“cost centers” into “profit centers” and compels them to become more involved
in the spot market. This also brings new opportunities for optimization-based
decision support systems for ship scheduling planners.

7.5 Focus on supply chains

In most ship scheduling studies reported in the literature, the supply chain
perspective is missing. Recently we see an increasing competition between sup-
ply chains even more than between shipping companies. Shipping companies
must consider themselves as total logistics providers, or at least as a part of a
total logistics provider, instead of only a provider of sea transport services. This
means that there must be some sort of collaboration and integration along the
supply chain, for instance, between the shippers and the shipping company.
Vendor managed inventory takes advantage of the benefits of introducing this
integration and transfers inventory management and ordering responsibilities
completely to the vendor or the logistics provider. The logistics provider de-
termines both the quantity and timing of customer deliveries. The customer is
guaranteed not to run out of product, and in return the logistics provider gains
a dramatic increase in flexibility that leads to more efficient use of its resources.
We expect an increasing emphasis on integrating maritime transportation
into the supply chain. This will also bring new interesting challenges to the
research community in routing and scheduling, such as inventory routing, col-
laboration, and cost and/or profit sharing along the supply chain.

7.6 Strategic planning issues and market interaction

Vessel fleet sizing should be given more attention in the future. This strate-
gic problem is extremely important as decisions concerning fleet size and com-
position set the stage for routing and scheduling. Even though there have been
a few studies on this type of problem, the potential for improving fleet size
decisions by using optimization-based decision support systems is probably sig-
nificant. As already discussed, we have seen a trend from industrial to tramp
shipping, with much more interaction with the market. This high degree of
market interaction probably makes the fleet size issue even more important
 Ch. 4. Maritime Transportation 279

and complex, as one now has to make some assumptions on future market
development in order to determine the optimal fleet.
Contract evaluation (discussed in Section 3.5) is yet another important
strategic problem that has only scarcely been considered in the research liter-
ature. This is to a large extent related to the fleet size issue, since the shipping
company has to evaluate whether it has sufficient fleet tonnage to fulfill poten-
tial contract commitments together with its existing commitments. If so, one
has to check whether a contract is profitable or not. In order to do so, one
also has to make some assumptions about how the spot market will develop
for the given contract period. Since both fleet sizing and contract evaluation
decisions are to a large extent dependent on the expectations of how a future
market will develop, concepts of optimization under uncertainty must probably
be considered.

8 Conclusion

Maritime transportation is the backbone of international trade. The volume

of maritime transportation has been growing for many years, and is expected
to continue growing in the foreseeable future. Maritime transportation is a
unique transportation mode possessing characteristics that differ from other
modes of transportation, and requires decision support models that fit the spe-
cific problem characteristics.
Maritime transportation poses a rich spectrum of decision making prob-
lems, from strategic ones through tactical to operational. We also find within
maritime transportation a variety of modes and types of operations with their
specific characteristics: industrial, tramp, liner, deep-sea, short-sea, coastal and
inland waterways, port and container terminals, and their interface with ves-
sels.
Research interest in maritime transportation problems has been increasing
in recent years but still lags behind the more visible modes, namely truck, air,
and rail. In this chapter we have presented a variety of decision making prob-
lems in maritime transportation. For some common problems we presented
models as well as discussed solution approaches, whereas for other problems
we confined ourselves to a general description of the problems and referred
the reader to sources that deal with the problems more extensively. Although
most of the research in maritime transportation stemmed from real-life prob-
lems only a fraction of it has matured into real decision support systems that
are used in practice.
The fast containerization of general and break-bulk cargo combined with
fast development of information technology and telecommunications, and with
competitive pressures, have resulted in a shift of emphasis from ocean trans-
portation to intermodal supply chains. The economies of scale that such supply
chains pose result in industry consolidation and larger controlled fleets, pre-
senting a fertile ground for applying quantitative decision support tools. At the
280 M. Christiansen et al.

same time shippers started to focus on their core operations and to outsource
logistic functions to third party providers who also have significant economies
of scale. Thus, also on the demand side we observe consolidation and higher
potential for applying quantitative decision support tools.
Uncertainty plays a major role in maritime transportation and therefore ro-
bust and stochastic models should take center stage. However, in this respect
the surface has only been scratched.
Maritime transportation poses a wide variety of challenging research prob-
lems, the solutions to which have high potential to improve economic per-
formance and increase profitability in this highly competitive arena. The fast
development of optimization algorithms and computing power facilitate solu-
tion of more realistic problems, and we are confident that more research will
be directed to this crucial transportation mode.

Acknowledgements

This work was carried out with financial support from the Research Coun-
cil of Norway through the TOP project (Improved Optimisation Methods in
Transportation Logistics), the INSUMAR project (Integrated supply chain and
maritime transportation planning) and the OPTIMAR project (Optimization
in Maritime transportation and logistics). We want to thank the Doctoral stu-
dents Roar Grønhaug, Frank Hennig, and Yuriy Maxymovych for a careful
reading of the chapter and for helpful suggestions.

References

Almogy, Y., Levin, O. (1970). Parametric analysis of a multi-stage stochastic shipping problem. In:
IFORS, OR 69: Proc. Fifth International Conference on OR. Tavistock, London, pp. 359–370.
Appelgren, L.H. (1969). A column generation algorithm for a ship scheduling problem. Transportation
Science 3, 53–68.
Appelgren, L.H. (1971). Integer programming methods for a vessel scheduling problem. Transportation
Science 5, 64–78.
Ariel, A. (1991). The effect of inventory holding costs on the optimal payload of bulk carriers. Maritime
Policy & Management 18 (3), 217–224.
Avriel, M., Penn, M., Shipier, N., Witteboon, S. (1998). Stowage planning for container ships to reduce
the number of shifts. Annals of Operations Research 76, 55–71.
Bausch, D.O., Brown, G.G., Ronen, D. (1998). Scheduling short-term marine transport of bulk prod-
ucts. Maritime Policy & Management 25 (4), 335–348.
Bellmore, M. (1968). A maximum utility solution to a vehicle constrained tanker scheduling problem.
Naval Research Logistics Quarterly 15, 403–411.
Bendall, H.B., Stent, A.F. (2001). A scheduling model for a high speed containership service: A hub
and spoke short-sea application. International Journal of Maritime Economics 3 (3), 262–277.
Brown, G.G., Graves, G.W., Ronen, D. (1987). Scheduling ocean transportation of crude oil. Manage-
ment Science 33 (3), 335–346.
Brown, G.G., Goodman, C.E., Wood, R.K. (1990). Annual scheduling of Atlantic Fleet naval combat-
ants. Operations Research 38 (2), 249–259.
 Ch. 4. Maritime Transportation 281

Brown, G.G., Dell, R.F., Farmer, R.A. (1996). Scheduling coast guard district cutters. Interfaces 26 (2),
59–72.
Brown, G.G., Dell, R.F., Wood, R.K. (1997a). Optimization and persistence. Interfaces 27 (5), 15–37.
Brown, G.G., Cormican, K.J., Lawphongpanich, S., Widdis, D.B. (1997b). Optimizing submarine
berthing with a persistence incentive. Naval Research Logistics 44, 301–318.
Brønmo, G., Christiansen, M., Nygreen, B. (2006). Ship routing and scheduling with flexible cargo
sizes. Journal of the Operation Research Society, doi:10.1057/palgrave.jors.2602263. Advance online
publication, 16 August 2006.
Brønmo, G., Christiansen, M., Fagerholt, K., Nygreen, B. (2007). A multi-start local search heuristic
for ship scheduling – a computational study. Computers & Operations Research 34 (3), 900–917.
Chajakis, E.D. (1997). Sophisticated crude transportation. OR/MS Today 24 (6), 30–34.
Chajakis, E.D. (2000). Management science for marine petroleum logistics. In: Zanakis, S.H., Doukidis,
G., Zopounidis, C. (Eds.), Decision Making: Recent Developments and Worldwide Applications.
Kluwer Academic, pp. 169–185.
Cho, S.-C., Perakis, A.N. (1996). Optimal liner fleet routing strategies. Maritime Policy & Manage-
ment 23 (3), 249–259.
Cho, S.-C., Perakis, A.N. (2001). An improved formulation for bulk cargo ship scheduling with a single
loading port. Maritime Policy & Management 28 (4), 339–345.
Christiansen, M. (1999). Decomposition of a combined inventory and time constrained ship routing
problem. Transportation Science 33 (1), 3–16.
Christiansen, M., Fagerholt, K. (2002). Robust ship scheduling with multiple time windows. Naval Re-
search Logistics 49 (6), 611–625.
Christiansen, M., Nygreen, B. (1998a). A method for solving ship routing problems with inventory
constraints. Annals of Operations Research 81, 357–378.
Christiansen, M., Nygreen, B. (1998b). Modeling path flows for a combined ship routing and inventory
management problem. Annals of Operations Research 82, 391–412.
Christiansen, M., Nygreen, B. (2005). Robust inventory ship routing by column generation. In: De-
saulniers, G., Desrosiers, J., Solomon, M.M. (Eds.), Column Generation. Springer-Verlag, New York,
pp. 197–224.
Christiansen, M., Fagerholt, K., Ronen, D. (2004). Ship routing and scheduling: Status and perspectives.
Transportation Science 38 (1), 1–18.
Cline, A.K., King, D.H., Meyering, J.M. (1992). Routing and scheduling of coast guard buoy tenders.
Interfaces 22 (3), 56–72.
Crainic, T.G., Kim, K.H. (2007). Intermodal transportation. In: Barnhart, C., Laporte, G. (Eds.), Trans-
portation. Handbooks in Operations Research and Management Science. North-Holland, Amsterdam,
pp. 467–537. This volume.
Crary, M., Nozick, L.K., Whitaker, L.R. (2002). Sizing the US destroyer fleet. European Journal of
Operational Research 136, 680–695.
Cullinane, K., Khanna, M. (1999). Economies of scale in large container ships. Journal of Transport
Economics and Policy 33 (2), 185–208.
Dantzig, G.B., Fulkerson, D.R. (1954). Minimizing the number of tankers to meet a fixed schedule.
Naval Research Logistics Quarterly 1, 217–222.
Darby-Dowman, K., Fink, R.K., Mitra, G., Smith, J.W. (1995). An intelligent system for US coast guard
cutter scheduling. European Journal of Operational Research 87, 574–585.
Darzentas, J., Spyrou, T. (1996). Ferry traffic in the Aegean Islands: A simulation study. Journal of the
Operational Research Society 47, 203–216.
Desrosiers, J., Dumas, Y., Solomon, M.M., Soumis, F. (1995). Time constrained routing and scheduling.
In: Ball, M.O., Magnanti, T.L., Monma, C.L., Nemhauser, G.L. (Eds.), Network Routing. Handbooks
in Operations Research and Management Science, vol. 8. North-Holland, Amsterdam, pp. 35–139.
Erkut, E., Verter, V. (2007). Hazardous materials transportation. In: Barnhart, C., Laporte, G. (Eds.),
Transportation. Handbooks in Operations Research and Management Science. North-Holland, Ams-
terdam, pp. 539–621. This volume.
European Commission (2004). European Transport Policy for 2010: Time to Decide. White paper.
282 M. Christiansen et al.

Fagerholt, K. (1999). Optimal fleet design in a ship routing problem. International Transactions in Op-
erational Research 6 (5), 453–464.
Fagerholt, K. (2001). Ship scheduling with soft time windows – an optimization based approach. Euro-
pean Journal of Operational Research 131, 559–571.
Fagerholt, K. (2004). A computer-based decision support system for vessel fleet scheduling – experience
and future research. Decision Support Systems 37 (1), 35–47.
Fagerholt, K., Christiansen, M. (2000a). A combined ship scheduling and allocation problem. Journal
of the Operational Research Society 51 (7), 834–842.
Fagerholt, K., Christiansen, M. (2000b). A travelling salesman problem with allocation, time window
and precedence constraints – an application to ship scheduling. International Transactions in Opera-
tional Research 7 (3), 231–244.
Fagerholt, K., Lindstad, H. (2000). Optimal policies for maintaining a supply service in the Norwegian
Sea. OMEGA 28, 269–275.
Fagerholt, K., Rygh, B. (2002). Design of a sea-borne system for fresh water transport – A simulation
study. Belgian Journal of Operations Research, Statistics and Computer Science 40 (3–4), 137–146.
Fisher, M.L., Rosenwein, M.B. (1989). An interactive optimization system for bulk-cargo ship schedul-
ing. Naval Research Logistics 36, 27–42.
Flatberg, T., Haavardtun, H., Kloster, O., Løkketangen, A. (2000). Combining exact and heuristic
methods for solving a vessel routing problem with inventory constraints and time windows. Ricerca
Operativa 29 (91), 55–68.
Fleming, D.K. (2002). Reflections on the history of US cargo liner service (part I). International Journal
of Maritime Economics 4 (4), 369–389.
Fleming, D.K. (2003). Reflections on the history of US cargo liner service (part II). Maritime Economics
& Logistics 5 (1), 70–89.
Fox, M., Herden, D. (1999). Ship scheduling of fertilizer products. OR Insight 12 (2), 21–28.
Garrod, P., Miklius, M. (1985). The optimal ship size: A comment. Journal of Transport Economics and
Policy 19 (1), 83–91.
Gillman, S. (1999). The size economies and network efficiencies of large containerships. International
Journal of Maritime Economics 1 (1), 39–59.
Hersh, M., Ladany, S.P. (1989). Optimal scheduling of ocean cruises. INFOR 27 (1), 48–57.
Hughes, W.P. (2002). Navy operations research. Operations Research 50 (1), 103–111.
Hwang, S.-J. (2005). Inventory constrained maritime routing and scheduling for multi-commodity liquid
bulk. Phd thesis, Georgia Institute of technology, Atlanta.
Jaikumar, R., Solomon, M.M. (1987). The tug fleet size problem for barge line operations: A polyno-
mial algorithm. Transportation Science 21 (4), 264–272.
Jansson, J.O., Shneerson, D. (1978). Economies of scale of general cargo ships. The Review of Eco-
nomics and Statistics 60 (2), 287–293.
Jansson, J.O., Shneerson, D. (1982). The optimal ship size. Journal of Transport Economics and Policy 16
(3), 217–238.
Jansson, J.O., Shneerson, D. (1985). A model of scheduled liner freight services: Balancing inventory
cost against shipowners’ costs. The Logistics and Transportation Review 21 (3), 195–215.
Jansson, J.O., Shneerson, D. (1987). Liner Shipping Economics. Chapman & Hall, London.
Jaramillo, D.I., Perakis, A.N. (1991). Fleet deployment optimization for liner shipping. Part 2. Imple-
mentation and results. Maritime Policy & Management 18 (4), 235–262.
Jetlund, A.S., Karimi, I.A. (2004). Improving the logistics of multi-compartment chemical tankers.
Computers & Chemical Engineering 28, 1267–1283.
Kang, J.-G., Kim, Y.-D. (2002). Stowage planning in maritime container transportation. Journal of the
Operational Research Society 53, 415–426.
Kao, C., Chen, C.Y., Lyu, J. (1993). Determination of optimal shipping policy by inventory theory.
International Journal of Systems Science 24 (7), 1265–1273.
Kleywegt, A. (2003). Contract planning models for ocean carriers. Working paper, Georgia Institute of
Technology, Atlanta, GA.
Korsvik, J.E., Fagerholt, K., Brønmo, G. (2007). Ship scheduling with flexible cargo quantities: A heuris-
tic solution approach. Working paper, Norwegian University of Science and Technology, Trondheim,
Norway.
 Ch. 4. Maritime Transportation 283

Ladany, S.P., Arbel, A. (1991). Optimal cruise-liner passenger cabin pricing policy. European Journal of
Operational Research 55, 136–147.
Lane, D.E., Heaver, T.D., Uyeno, D. (1987). Planning and scheduling for efficiency in liner shipping.
Maritime Policy & Management 14 (2), 109–125.
Larson, R.C. (1988). Transporting sludge to the 106-Mile site: An inventory/routing model for fleet
sizing and logistics system design. Transportation Science 22 (3), 186–198.
Lawrence, S.A. (1972). International Sea Transport: The Years Ahead. Lexington Books, Lexington, MA.
Liu, C.-M., Sherali, H.D. (2000). A coal shipping and blending problem for an electric utility company.
OMEGA 28, 433–444.
Lo, H.K., McCord, M.R., Wall, C.K. (1991). Value of ocean current information for strategic routing.
European Journal of Operational Research 55, 124–135.
Martin, G.L., Randhawa, S.U., McDowell, E.D. (1988). Computerized container-ship loading:
A methodology and evaluation. Computers & Industrial Engineering 14 (4), 429–440.
McCord, M.R., Lee, Y.-K., Lo, H.K. (1999). Ship routing through altimetry-derived ocean currents.
Transportation Science 33 (1), 49–67.
McLellan, R.G. (1997). Bigger vessels: How big is too big? Maritime Policy & Management 24 (2), 193–
211.
Mehrez, A., Hung, M.S., Ahn, B.H. (1995). An industrial ocean-cargo shipping problem. Decision Sci-
ences 26 (3), 395–423.
Nulty, W.G., Ratliff, H.D. (1991). Interactive optimization methodology for fleet scheduling. Naval
Research Logistics 38, 669–677.
O’Brien, G.G., Crane, R.R. (1959). The scheduling of a barge line. Operations Research 7, 561–570.
Papadakis, N.A., Perakis, A.N. (1989). A nonlinear approach to multiorigin, multidestination fleet de-
ployment problem. Naval Research Logistics 36, 515–528.
Papadakis, N.A., Perakis, A.N. (1990). Deterministic minimal time vessel routing. Operations Re-
search 38 (3), 426–438.
Perakis, A.N. (1985). A second look at fleet deployment. Maritime Policy & Management 12, 209–214.
Perakis, A.N., Bremer, W.M. (1992). An operational tanker scheduling optimization system. Back-
ground, current practice and model formulation. Maritime Policy & Management 19 (3), 177–187.
Perakis, A.N., Inozu, B. (1991). Optimal maintenance, repair, and replacement for Great Lakes marine
diesels. European Journal of Operational Research 55, 165–182.
Perakis, A.N., Jaramillo, D.I. (1991). Fleet deployment optimization for liner shipping. Part 1. Back-
ground, problem formulation and solution approaches. Maritime Policy & Management 18 (3), 183–
200.
Perakis, A.N., Papadakis, N.A. (1987a). Fleet deployment optimization models. Part 1. Maritime Policy
& Management 14, 127–144.
Perakis, A.N., Papadakis, N.A. (1987b). Fleet deployment optimization models. Part 2. Maritime Policy
& Management 14, 145–155.
Perakis, A.N., Papadakis, N.A. (1989). Minimal time vessel routing in a time-dependent environment.
Transportation Science 23, 266–276.
Persson, J.A., Göthe-Lundgren, M. (2005). Shipment planning at oil refineries using column generation
and valid inequalities. European Journal of Operational Research 163, 631–652.
Pesenti, R. (1995). Hierarchical resource planning for shipping companies. European Journal of Opera-
tional Research 86, 91–102.
Pope, J.A., Talley, W.K. (1988). Inventory costs and optimal ship size. Logistics and Transportation Re-
view 24 (2), 107–120.
Powell, B.J., Perakis, A.N. (1997). Fleet deployment optimization for liner shipping: An integer pro-
gramming model. Maritime Policy & Management 24 (2), 183–192.
Psaraftis, H.N. (1988). Dynamic vehicle routing problems. In: Golden, B.L., Assad, A.A. (Eds.), Vehicle
Routing: Methods and Studies. North-Holland, Amsterdam, pp. 223–248.
Psaraftis, H.N. (1999). Foreword to the focused issue on maritime transportation. Transportation Sci-
ence 33 (1), 1–2.
Rana, K., Vickson, R.G. (1988). A model and solution algorithm for optimal routing of a time-chartered
containership. Transportation Science 22 (2), 83–95.
284 M. Christiansen et al.

Rana, K., Vickson, R.G. (1991). Routing container ships using Lagrangean relaxation and decomposi-
tion. Transportation Science 25 (3), 201–214.
Richetta, O., Larson, R. (1997). Modeling the increased complexity of New York City’s refuse marine
transport system. Transportation Science 31 (3), 272–293.
Ronen, D. (1982). The effect of oil price on the optimal speed of ships. Journal of the Operational
Research Society 33, 1035–1040.
Ronen, D. (1983). Cargo ships routing and scheduling: Survey of models and problems. European Jour-
nal of Operational Research 12, 119–126.
Ronen, D. (1986). Short-term scheduling of vessels for shipping bulk or semi-bulk commodities origi-
nating in a single area. Operations Research 34 (1), 164–173.
Ronen, D. (1991). Editorial to the feature issue on water transportation. European Journal of Opera-
tional Research 55 (2), 123.
Ronen, D. (1993). Ship scheduling: The last decade. European Journal of Operational Research 71, 325–
333.
Ronen, D. (2002). Marine inventory routing: Shipments planning. Journal of the Operational Research
Society 53, 108–114.
Sambracos, E., Paravantis, J.A., Tarantilis, C.D., Kiranoudis, C.T. (2004). Dispatching of small con-
tainers via coastal freight liners: The case of the Aegean sea. European Journal of Operational
Research 152, 365–381.
Schrady, D., Wadsworth, D. (1991). Naval combat logistics support system. Journal of the Operational
Research Society 42 (11), 941–948.
Schwartz, N.L. (1968). Discrete programs for moving known cargos from origins to destinations on time
at minimum bargeline fleet cost. Transportation Science 2, 134–145.
Scott, J.L. (1995). A transportation model, its development and application to a ship scheduling prob-
lem. Asia-Pacific Journal of Operational Research 12, 111–128.
Sheppard, E.J., Seidman, D. (2001). Ocean shipping alliances: The wave of the future? International
Journal of Maritime Economics 3, 351–367.
Sherali, H.D., Al-Yakoob, S.M., Hassan, M.M. (1999). Fleet management models and algorithms for
an oil tanker routing and scheduling problem. IIE Transactions 31, 395–406.
Shih, L.-H. (1997). Planning of fuel coal imports using a mixed integer programming method. Interna-
tional Journal of Production Economics 51, 243–249.
Sigurd, M.M., Ulstein, N.L., Nygreen, B., Ryan, D.M. (2005). Ship scheduling with recurring visits and
visit separation requirements. In: Desaulniers, G., Desrosiers, J., Solomon, M.M. (Eds.), Column
Generation. Springer-Verlag, New York, pp. 225–245.
Talley, W.K., Agarwal, V.B., Breakfield, J.W. (1986). Economics of density of ocean tanker ships. Journal
of Transport Economics and Policy 20 (1), 91–99.
Thompson, P.M., Psaraftis, H.N. (1993). Cyclic transfer algorithms for multi-vehicle routing and
scheduling problems. Operations Research 41 (5), 935–946.
UNCTAD (2003). Review of Maritime Transport, 2003. United Nations, New York and Geneva.
UNCTAD (2004). Review of Maritime Transport, 2004. United Nations, New York and Geneva.
Vis, I.F.A., de Koster, R. (2003). Transshipment of containers at a container terminal: An overview.
European Journal of Operational Research 147, 1–16.
Vukadinovic, K., Teodorovic, D. (1994). A fuzzy approach to the vessel dispatching problem. European
Journal of Operational Research 76 (1), 155–164.
Vukadinovic, K., Teodorovic, D., Pavkovic, G. (1997). A neural network approach to the vessel dis-
patching problem. European Journal of Operational Research 102 (3), 473–487.
Wermus, M., Pope, J.A. (1994). Scheduling harbor pilots. Interfaces 24 (2), 44–52.
Williams, H.P. (1999). Model Building in Mathematical Programming, 4th edition. Wiley, West Sussex,
pp. 160–165.
Williams, T.M. (1992). Heuristic scheduling of ship replenishment at sea. Journal of the Operational
Research Society 43 (1), 11–18.
Wilson, I.D., Roach, P.A. (2000). Container stowage planning: a methodology for generating comput-
erized solutions. Journal of the Operational Research Society 51 (11), 1248–1255.
Xinlian, X., Tangfei, W., Daisong, C. (2000). A dynamic model and algorithm for fleet planning. Mar-
itime Policy & Management 27 (1), 53–63.
C. Barnhart and G. Laporte (Eds.), Handbook in OR & MS, Vol. 14
Copyright © 2007 Elsevier B.V. All rights reserved
DOI: 10.1016/S0927-0507(06)14005-0

Chapter 5
Dynamic Models for Freight Transportation
Warren B. Powell
Department of Operations Research and Financial Engineering, Princeton University,
Princeton, NJ 08544, USA
E-mail: [email protected]

Belgacem Bouzaïene-Ayari
Department of Operations Research and Financial Engineering, Princeton University,
Princeton, NJ 08544, USA
E-mail: [email protected]

Hugo P. Simão
Department of Operations Research and Financial Engineering, Princeton University,
Princeton, NJ 08544, USA
E-mail: [email protected]

1 Introduction

Dynamic models arise in a vast array of transportation applications as a re-

sult of the need to capture the evolution of activities over time. But exactly
what do we mean by a “dynamic model”? All dynamic models capture the
timing of physical activities which introduces both modeling and algorithmic
challenges and opportunities. In real problems, there are three classes of ac-
tivities that can evolve over time: the physical movement of freight, equipment
and people; the evolution of information; and the movement of financial re-
sources as services are paid for.
In this chapter we focus on modeling the organization and flow of infor-
mation and decisions, in the context of freight transportation problems that
involve the management of people and equipment to serve the needs of cus-
tomers. The timing of the flow of capital is becoming an increasingly important
dimension of freight transportation, but as of this writing, there has been virtu-
ally no formal research on the topic. Modeling the timing of physical activities,
by contrast, dates to the 1950s. These models introduce a range of modeling
and algorithmic challenges that have been studied since the early years of op-
erations research models.
Our decision to focus on modeling the evolution of information (or more
broadly, the organization and flow of information and decisions) reflects the
growing maturity of this class of models, and the importance of questions that
require an explicit model of information processes. Often categorized under
the phrase “stochastic, dynamic models”, there are two issues that arise in this
context:

285
286 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

• Improving the management of physical resources by using a more ac-

curate representation of information processes.
• Designing and controlling information processes to maximize perfor-
mance.
The first class of questions focuses on more realistic models to improve the
design and control of physical systems. Examples of questions include:
• How do we best allocate rail cars to handle future requests?
• Which driver should move a load of freight given the as-yet unknown
loads that will be available at the destination of the first load?
• How many dock workers should be assigned to work a shift given the
uncertain amount of freight to be unloaded?
• How many locomotives should be held at a yard to handle potential
equipment failures of locomotives that are expected to arrive inbound
over the next 12 hours?
All of these are examples of questions concerning the allocation of physical
resources in the presence of uncertainty in the future. Most often, the uncer-
tainty arises around demands that are placed on the system, but there can be
uncertainty in the availability of resources (how many drivers can the trucking
company hire this week? will the locomotive break down?) and the perfor-
mance of the network itself (primarily travel times). A common theme when
designing and controlling systems under uncertainty is producing solutions that
work well over as broad a range as possible of potential outcomes in the future.
Such solutions are referred to as robust, although this term is most often used
in the context of design problems.
The second class of questions arises in the design and control of information
processes. Examples include:
• Should a company invest in wireless communication devices that allow
them to determine the location and status of drivers and equipment?
• Should a company invest in a database that tracks the status of all of its
drivers/crew?
• What is the cost of allowing customers to make requests at the last
minute?
Questions such as these require a model that allows us to explicitly manipulate
the information process. If we consider a scenario which reduces the informa-
tion available when we are making a decision, then we have to be sure that
we are producing the best decisions that we can even if we do not have the
information.

2 Some illustrative applications

It helps to begin our presentation with a discussion of different applications

in freight transportation. Our presentation of models focuses on the resources
 Ch. 5. Dynamic Models for Freight Transportation 287

being managed, and the information and decision processes that govern the
evolution of our systems. Below we list a series of application areas, along with
examples of resource classes, (exogenous) information classes, and decision
classes. For each subcategory, we briefly list typical examples of each class. For
transportation problems, every resource includes as attributes its location and
the estimated time at which the resource will arrive at the destination (later,
we provide a more formal and general characterization of “estimated time of
arrival”). For this reason, we exclude these attributes in the discussion below.
(1) Truckload trucking:
Resource classes:
• Drivers, who are normally modeled as including the tractor. Drivers
are characterized by attributes such as location, home domicile, time
due at home, international experience, and sleeper vs. single driver.
• Trailers: loaded status (loaded or empty), cleanliness, repair status.
• Loads of freight: destination, service requirements, equipment
needs, shipper classification.
Information classes:
• Customer requests for service.
• Drivers entering and leaving the system.
• Equipment status.
• Transit times.
Decision classes:
• Drivers: couple with trailer and/or load, move empty, move loaded,
go on rest (at home or away).
• Trailers: move (empty or loaded), clean, repair.
• Loads: accept/reject initial request, spot pricing of a load, couple
with trailer, move (with trailer), move using a subcontractor (a driver
not managed by the company).
(2) Rail car distribution:
Resource classes:
• Cars: car type, loaded status, maintenance status, cleanliness, ship-
per pool, owner.
• Orders: destination, service requirements, commodity type and sub-
stitutability.
Information classes:
• Customer orders, including updates to customer orders.
• Empty cars becoming available from shippers or other railroads.
• Transit times.
• The destination of an order (revealed only after a car is filled).
• Acceptability of a car to the customer (is it clean? in good repair?).
Decision classes:
• Cars: move (loaded or empty), repair, clean, switch pools, move off-
line to another railroad, spot pricing of an order.
288 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

• Orders: accept/reject, subcontract.

(3) Locomotive management:
Resource classes:
• Locomotives: type, horsepower, high/low adhesion, fuel status, due-
in-shop date, inbound train that the locomotive last pulled (deter-
mines the other locomotives the unit is connected to).
• Trains: train type (grain, coal, merchandise, stack, priority), destina-
tion, departure time, tonnage, horsepower required per ton being
pulled.
Information classes:
• Locomotives: maintenance status, arrivals to system, departures
from system.
• Trains: tonnage, additions to schedule, cancellations (annulments),
delays.
Decision classes:
• Locomotives: assignments to trains, assignment to shop, reposition-
ing (moving without a train).
• Trains: decisions to delay trains.
(4) Less-than-truckload trucking:
Resource classes:
• Shipments: destination, service requirement, size (weight and den-
sity).
• Trailers: type (typically 28 or 48 ), current load (shipments on the
trailer), destination, departure time.
• Drivers: domicile, bid vs. nonbid, sleeper vs. single, days away from
home, hours of service.
• Tractors: type, status.
Information classes:
• Shipments entering the system.
• Changes in driver status.
• Transfer of shipments at sort facilities.
Decision classes:
• Where to run trailers direct.
• What trailer to put a shipment on.
• Coupling of trailers into loads to be moved (typically involves pairing
28 trailers).
• Assignment of drivers to loads.
(5) Intermodal container operations:
Resource classes:
• Containers: type, loaded status, cleanliness, repair status, ownership.
• Loads: destination, weight, service requirements, routing restric-
tions.
 Ch. 5. Dynamic Models for Freight Transportation 289

Information classes:
• Customer orders.
• Transit times.
• Movement of containers.
Decision classes:
• Customer orders: accept/reject, move via subcontractor, spot pricing
of an order.
• Containers: assignment to order, movement (loaded or empty), as-
signment to ship or train departures, intra-port management, rout-
ing and scheduling by truck.
(6) Air cargo:
Resource classes:
• Aircraft: type, configuration, load status, maintenance schedule.
• Requirements (loads of freight): destination, weight, size character-
istics, passengers, pickup and delivery windows.
• Pilots: training, home location, hours worked, work schedule.
Information classes:
• Requests to move freight and passengers.
• Equipment failures (and random changes in status).
• Weather delays.
Decision classes:
• Move aircraft from one location to another.
• Reconfigure aircraft (to handle passengers or different types of
cargo).
• Repair, refuel.
• Outsource customer requests.

3 A resource model

At the heart of our problems is the challenge of modeling resources. Al-

though the focus of this chapter is information, we are interested in informa-
tion in the context of managing resources for freight applications.
We begin in Section 3.1 by addressing the simple question: what is a re-
source? Section 3.2 introduces a basic resource model which gives us a general
notational framework for modeling resource classes and attributes. Section 3.3
introduces the concept of resource layers, which allows us to handle combi-
nations such as driver/trailer/load and locomotive/train; these are composite
resources that take on new behaviors. Section 3.4 notes the interaction between
the design of the attribute space, and the flow of information and decisions.
Section 3.5 shows how our attribute notation can be used to describe six classes
of resource management problems. Finally, Section 3.6 closes with a discussion
of three perspectives of the “state” of our system.
290 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

3.1 What is a resource?

Any freight transportation system can be thought of as a problem in the

management of resources. When we put it this way, we tend to think of re-
sources as people and equipment, and sometimes fixed facilities. When we
formulate optimization models, resources tend to appear as right-hand side
constraints, which suggests that resources are anything that constrains the sys-
tem. From the perspective of models, we can divide all data into information
classes, where “resources” are information classes that constrain the system. It
is useful to distinguish between static resource classes (such as terminals) which
constrain the system, but which may not themselves be decision variables, and
active resource classes (people, equipment) which we are managing.
When we adopt this more formal definition of a resource, we find that
“demands” (also known as customers, tasks, requirements, loads, moves, pick-
ups/deliveries, schedules) are also a form of “resource”. Often, the distinction
between “managing resources” (objects over which we have control) to “serve
customers” (objects over which we have no control) is an artificial one that
does not stand up to scrutiny (and certainly not in the formalism of a mathe-
matical model).
A simple example illustrates why it is useful to use a broader definition of
resources. Consider the problem faced by a truckload carrier of moving a load
from city i to city j. The driver at i needs to return to his home domicile in
city d. For this reason, the carrier assigns the driver to the load but moves both
to an intermediate relay r, where the driver and load separate. The driver is
then assigned to another load that moves him closer to his home, while the load
is assigned to a new driver who will presumably move it toward its final desti-
nation (in industries such as long-haul less-than-truckload trucking, a load can
move through four or five such relays before reaching its destination). During
this time, we are trying to get the driver to his home and the load to its desti-
nation. While the attributes of each are different, the modeling issues are the
same.
A working definition of a resource is an information class that constrains
the system. This broad definition includes not only equipment and people, but
also fixed assets such as truck terminals and railroad track which limits our
ability to move freight. An active resource class is an endogenously controllable
information class that constrains the system. Drivers, trailers, locomotives, and
planes are all examples of active resource classes. A load of freight also limits
our system (it limits how many trucks we can move loaded). In most models,
a load is an example of a passive resource class since it constrains the system,
but we cannot manage it directly (we can move a trailer with a load in it, but not
a load without a trailer). Exceptions arise with companies that can contract out
movements. In this case, we can effectively move a load without using any of
the company-owned drivers or equipment. In such a setting, the load becomes
an active resource class.
 Ch. 5. Dynamic Models for Freight Transportation 291

3.2 A basic resource model

Our resource model, then, begins with a listing of the resource classes:
C R = set of resource classes
We describe a resource using a vector of attributes
a = vector of attributes which describe the state of a resource
Ac = space of possible attribute vectors of a resource in class c ∈ C R 
Attributes can be categorical (home domicile, equipment type, maintenance
status) or numerical (hours worked, days until next maintenance stop, fuel
level). Attribute vectors can be divided between static attributes (for example,
the type of equipment or domicile of a driver) and dynamic attributes (loca-
tion, maintenance status, hours of service). Often, the set of static attributes is
combined to identify a resource type or commodity, while the set of dynamic
attributes is the state. Dynamic resources have at least one dynamic attribute;
static resources are characterized by attributes that are completely static.
We describe the status of all the resources in our system using:
Rcta = the number of resources in class c ∈ C R
with attribute a at time t
 c 
Rct = Rta a∈Ac c∈C R 
Recall that one of our resource classes is the customer demands. It is com-
mon to model customer demands as the only source of exogenous information
(new requests being made on the system) but there are many problems where
there are exogenous updates to the other resources in the system (drivers be-
ing hired, equipment breaking down, equipment arriving exogenously to the
system). We capture all these forms of information using
ta = change in Rta between t − 1 and t
R
due to exogenous information
t = (R
R ta )a∈A 

3.3 Resource layering

It is often the case that resources from different classes will be coupled to
form a layered resource. For example:
• When managing locomotives, a set of locomotives will be assigned to
a train to move cars from one location to the next. When a locomotive
is attached to a train, it can only be removed at a cost, so the behavior
of a locomotive at a location depends on whether it is attached to a
train or not. Furthermore, the ability to route a locomotive back to its
292 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

home shop is very much affected by the attributes of a train that it is

attached to (a railroad might easily decide not to detach a locomotive
from a train that it is pulling, even if the destination of the train takes
the locomotive away from its maintenance base).
• A cargo aircraft loaded with freight that breaks down will have a dif-
ferent behavior from an empty aircraft.
• Decisions about a business jet, with a particular pilot moving a pas-
senger, that has just broken down will depend on the attributes of the
aircraft, pilot and passenger.
We have to define the ways that resources may be coupled, which produces a
set of resource layers. A layer, in general, is a resource class comprised of the
coupling of one or more resource classes. We might have two resource classes
(pilots and aircraft) and two resource layers. The first layer might be just the
pilots (P) while the second might be aircraft with pilots (A|P).
To clearly distinguish between the resource and the resource layer, we use
the same letter as a resource surrounded by a pair of parentheses to denote
that resource layer. We let
(A) = A|P denotes the aircraft layer
(P) = P denotes the pilot layer
L = a layering representing the set of resource layers
 
= (A) (P)
= {A|P P}
l = the index for resource layer l ∈ L i.e. l = A|P or l = P
For this example, we would refer to the pilot layer as a primitive layer, while
the aircraft layer is a composite layer. The attributes of each layer are given by

A(P) = AP  the attribute space of a pilot layer

(P) P
a =a  the attribute vector of a pilot layer
A (A)
=A A|P

= AA × AP  the attribute space of an aircraft layer

(A) A|P
a =a
= aA |aP ∈ A(A)  the attribute vector of an aircraft layer
It is very common in transportation and logistics to model one or two-layer
problems. A good example of a one-layer problem is the assignment of air-
craft to a fixed schedule of flights (where the departure time of the flight is
fixed). A two-layer problem would arise when assigning truck drivers to loads
of freight, where we have to determine when the load should be moved (some
loads have 24 hour windows or longer). This would be an example of a two-
layer problem with a single active layer (the driver) and a single passive layer
 Ch. 5. Dynamic Models for Freight Transportation 293

(the load). In many companies, it is possible to outsource a load, which is equiv-

alent to moving a load without a driver, giving us an instance of a two-layer
problem with two active layers. A three-layer problem might arise when we
have to simultaneously manage drivers, tractors, and trailers, or pilots, aircraft,
and loads of cargo.
Published examples of three-layer problems are fairly rare, but a good ex-
ample is the NRMO model for the military airlift problem (Morton et al., 1996;
Baker et al., 2002) which tracks the flows of aircraft, crews and cargo in a mili-
tary airlift problem.

3.4 Attributes, decisions, and information

The attribute space A should not reflect every possible state that a resource
may occupy, but rather the states where the arrival of new information, and/or
the ability to make a decision (which, after all, represents the arrival of new
information) might affect the behavior of the resource. For example, consider
a military aircraft which has to move from the United States to India through a
series of three intermediate airbases (for refueling and potential repairs). We
would need to represent the attributes of the aircraft at any point where new
information might change the movement of the aircraft. For example, equip-
ment breakdowns, congestion at the airbase or weather delays can affect the
timing of the route. Furthermore, this new information might result in deci-
sions being made to reroute the aircraft. In the simulation package AMOS,
used by the air mobility command, it is common not to model new information
or decisions while an aircraft is enroute, in which case we only would need to
model the aircraft at the end of the route.
The role of information and decisions in the design of the attribute space
is especially important in the presence of layered resources. It is clear the at-
tribute space of a layered resource is much more complex than that of either
primitive resource. The question is: what do we need to model? Consider a
simplified model of a fleet management problem where a vehicle (truck, jet,
container) is assigned to move a request from one location to another. There
are no intermediate stops, and no decisions can be made about the equipment
being managed until it has finished a task. We might refer to this as a “two-
layer problem” but we would never need to actually model a layered resource.
We do not make any decisions about a piece of equipment until it has finished
a task. While there is a period of time when the equipment is coupled with the
customer request, there is no need to explicitly model this.

3.5 Major problem classes

We can use our attribute vector to describe six major problem classes. These
are:
294 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

• Inventory problems – a = {·}. This basic problem arises in classical in-

ventory theory, where there is a single type of product being held in
inventory. The attribute a is a null vector.
• Multiproduct inventory – a = {k}, where k ∈ K is a set of resource types
(product classes). The attribute a consists of a single, static element.
• Single commodity flow problems – a = {i}, where i ∈ I is a set of states
or locations of a piece of equipment. An example is managing a fleet
of identical trucks, where the only attribute of a truck is its current
location.
• Multicommodity flow – a = {k i}, where k ∈ K represents types of
resources and i ∈ I is a set of locations or states.
• Heterogeneous resource allocation problems – a = (a1      an ). Here
we have an n-dimensional attribute vector, where ai is the ith attribute
of a. These applications arise in the management of people and com-
plex equipment.
• Multilayered resource scheduling problems – a = {ac1 |ac2 | · · · |acn }. Now
the attribute vector is a concatenation of attribute vectors.
These six problem classes represent an increasing progression in the size of
the attribute space. The problems with smaller attribute spaces are generally
viewed as flow problems, even if we require integer solutions.

3.6 The states of our system

In dynamic models, it is common to talk of the “state of the system”, but it

is very hard to define a state formally. A review of virtually any of the major
textbooks on dynamic programming will fail to produce a formal definition of
the state of the system.
We find that there are three perspectives of state variables. The first is the
attribute vector a which can be thought of as the state of a single resource. In
fields such as crew scheduling which make use of column generation methods
(Desrosiers et al., 1995; Vance et al., 1997; Desrochers and Soumis, 1989), the
attribute vector is referred to as a label, and the determination of the best
schedule for a crew is formulated as a dynamic program over the state space A.
The second is the vector Rt which represents the resource state of the system
(in most dynamic programming applications, Rt is viewed as the state variable).
While the size of A can range from one to the millions (or larger), the num-
ber of possible values of Rt can be truly huge. Let R be the space of possible
values of Rt (which we assume is discrete), and let N = a∈A Rta be the total
number of resources being managed. Then it is straightforward to show that
 
N + |A | − 1
|R| =  (1)
|A| − 1
Even for small problems, this number can be extremely large (creating prob-
lems where it exceeds 1050 is not that difficult). For practical purposes, it is
useful to refer to such spaces as “effectively infinite”.
 Ch. 5. Dynamic Models for Freight Transportation 295

The third state variable might be referred to as the “information state”,

although a more precise definition is the “knowledge state”. Consider a prob-
lem where there is a series of customer demands D t which vary from period
to period. We might estimate the mean demand using a simple exponential
smoothing model
t 
μ̄t = (1 − α)μ̄t−1 + αD
In addition, we might estimate the variance using

σ̄t2 = (1 − α)σ̄t−1
2 t − μ̄t−1 )2 
+ α(D
At time t, (μ̄t  σ̄t2 ) would be part of what we know at time t. Given (μ̄t  σ̄t2 ), we
have all the information we need from history about customer demands. Our
state variable would then be represented by St = (Rt  μ̄t  σ̄t2 ).

3.7 Bibliographic notes

The mathematical properties of pure networks were discovered in the 1950s

(see Ford and Fulkerson, 1962, for a classic early description of networks),
but it was in the 1970s that Glover, Klingman, and Kennington discovered
and exploited the intersection of network algorithms and emerging computer
science techniques to produce exceptionally fast (even in those days) algo-
rithms for solving large networks (see Glover et al., 1974; Langley et al., 1974;
Glover et al., 1992, summarizes this body of work). Network models required
that all the flows represented a single equipment type (see, for example, White,
1972). Most problems in practice involve different types of equipment (such as
box cars with different configurations) with some degree of substitution al-
lowed to satisfy demands (a shipper might accept different car types). The nu-
merous applications of these multicommodity flow problems produced a vast lit-
erature which sought to exploit the embedded network structure to overcome
the otherwise large size of these problems (see, for example, Lasdon, 1970;
Kennington, 1978; Kennington and Helgason, 1980; Assad, 1978). A significant
development in the multicommodity literature was its adaptation to solving
routing and scheduling problems for vehicles and crews, where the latter are
typically characterized by complex vectors of attributes (Desrosiers et al., 1984;
Lavoie et al., 1988; Desrochers and Soumis, 1989; Vance et al., 1997; Barnhart
et al., 1998). These techniques involve solving a multiattribute dynamic pro-
gram for a single crew; a master program then chooses the best from a set of
potential schedules to produce an optimal overall schedule. The attribute vec-
tor appears only in the dynamic program solved in the subproblem; once the
potential tours are generated, the master problem is typically a set partitioning
or set covering problem which is a very special type of integer multicommod-
ity flow problem. Powell et al. (2002) formally proposes the “heterogeneous
resource allocation problem” which is a generalization of the classical multi-
commodity flow problem. The modeling framework in this chapter is based
296 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

on Powell et al. (2001) which introduces a general problem class termed the
“dynamic resource transformation problem”.

4 Modeling exogenous information processes

The modeling of the flow of information can be surprisingly subtle. In a

deterministic model, we might try to solve a problem that looks like

min ct xt
x
t∈T
subject to
t 
At xt − Bt−1 xt−1 = R (2)
xt  0
In a model such as this, we would understand that xt = (xtij )ij∈I , where xtij
is the flow from city i to city j starting at time t (and arriving at j after some
travel time τij ). In other words, xt is the vector of activities happening at time t.
We would assume that ct are the costs incurred in that time period. In other
words, in a model such as this, the index t refers to when things are happening.
In a model that reflects dynamic information processes, we have to combine
models of physical activities over time with models of information arriving over
time. Consider a problem that arises in freight car distribution. If we were to
look at an order in a historical file, we would simply see a movement from one
location to another at a point in time. But if we were to capture the evolution
of information that occurred before this physical event, we would see a series
of information events as depicted in Figure 1, which reflects: (1) the initial call
that there is an order requiring a car; (2) the amount of time required to move
the empty car to the order; (3) the acceptance of the car by the shipper (who
will sometimes reject a car because it is damaged or dirty); (4) the time re-
quired to fill the car, and (5) the destination of an order (which is not revealed
until the car is loaded).
In our model above, the right-hand side constraint R t in Equation (2) would
normally be interpreted as the new arrivals to the system at time t. However,
we need to distinguish between the “phone call” representing the information
about a new order and the physical arrival of a new order.
In the remainder of this section, we set up the fundamentals for modeling
dynamic information processes. Section 4.1 begins by establishing our conven-
tion for modeling time. Then, Section 4.2 introduces a notational style for mod-
eling information processes. Section 4.3 introduces the important concept of
lagged information processes that arises in almost all transportation problems.
Finally, Section 4.4 provides an introduction to an important mathematical
concept used widely in the probability community for modeling information.
 Ch. 5. Dynamic Models for Freight Transportation 297

Fig. 1. The evolution of information for a single order in a freight car application.

Fig. 2. The relationship between continuous and discrete time.

4.1 Modeling time

For computational reasons, it is common to model problems in transporta-

tion and logistics in discrete time, but it is apparent that in real problems,
activities happen continuously. Surprisingly, the modeling of time (which is
fairly obvious for deterministic problems) is not standard in the stochastic op-
timization community (for example, some authors like to index the first time
period as t = 0 while others choose t = 1).
For computational reasons, we model our decision making process in dis-
crete time (these can be fixed time periods such as hourly or daily, or driven
by information events). However, it is useful to think of information arriving
in continuous time, which resolves any ambiguity about what information is
available when we make a decision. Figure 2 shows the relationship between
discrete and continuous time which is most commonly used in the stochas-
tic process community (Cinlar, 2003). In this convention, time t = 0 refers to
“here and now”, while discrete time period t refers to the time interval (t−1 t].
It is very important to identify the information content of a variable. In a
purely discrete time model, the precise sequencing of decisions and informa-
tion is ambiguous (was a decision at time t made before or after the infor-
mation that arrived at time t?). For this reason, we will view information as
arriving continuously over time, while a decision will be made at a discrete
298 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

point in time. Since a state variable is the information we need to make a deci-
sion, it will also be measured in discrete time.

4.2 The information process

In the math programming community, it is common to use xt as a generic

decision variable (xt might come in different flavors, or we might include
yt and zt ), but there is not a commonly accepted “generic variable” for in-
formation. For our presentation, we adopt the convention that Wt is a variable
representing all the information that arrived during the time interval between
t − 1 and t. This information can be about new customer orders, transit times,
costs, and the status of people and equipment. We may in general define
C I = the set of dynamic, exogenous information classes, representing
different types of information arriving over time (prices,
demands, equipment failures, travel times)
c
Wt = the information arriving for class c
 
Wt = Wtc c∈C I 
In a particular application, it will be common to model specific information
processes and give each of these names such as demands or prices. As a result,
it may not be necessary to use the generic “Wt ” notation. But it can be useful
to have a generic information variable (as it will be in our presentation) since
it will allow us to add new information processes without changing our basic
model. We adopt the notation that with the exception of our generic infor-
mation variable Wt , we use hats to indicate exogenous information processes.
For example, if a trucking company faces random demands and spot prices, we
might let D t and p̂t be the demands and prices that we learn about during time
interval t (between t − 1 and t). Thus we would write Wt = (D t  p̂t ). Given our
convention for labeling time, W1 is our first set of information arriving (W0 has
no meaning since information arrives over time intervals, and the first time
interval is t = 1).
We further adopt the notation that any variable with subscript t uses in-
formation up through time t. This notation is of fundamental importance, and
represents a significant departure from standard deterministic dynamic models
where t represents when a physical activity is happening. Thus, xt is a decision
that uses information up through time t; St might be a state variable at time t
(which uses Wt ), and Ct (xt ) is the cost computed using information up through
time t.
Normally, models that capture dynamic information processes do so be-
cause we do not know what information is going to arrive in the future. This is
fundamental to stochastic models, but we wish to emphasize that there are nu-
merous engineering applications which explicitly model dynamic information,
but model only one instance of information (for example, what might have oc-
curred in history). When we wish to model multiple information processes, we
 Ch. 5. Dynamic Models for Freight Transportation 299

can adopt the standard notation of the probability community. Let ω repre-
sent a particular realization of the information process: W1  W2      Wt    ,
and let Ω be the set of all possible realizations (we now see the motivation for
using Wt as our generic variable for information – it “looks like” ω).
It is typically the case that we will be working with a single realization at
a time. We can think of ω as an indexing of all the possible outcomes (as in
the integers from 1 to |Ω|), or as the literal outcomes of all the possible pieces
of information that become known. When we are using Monte Carlo sampling
techniques, we might think of ω as the random number seed to generate all the
possible random variables. These are all effectively equivalent views, but we
suggest that it is generally best to think of ω as a simple index. When we want
to refer to the information that becomes known during time interval t, we may
write ωt = Wt (ω). In this case, ωt is a sample realization of all the different
sources of exogenous information. If we view ω as the actual information in a
realization, we can write
ω = (ω1  ω2      ωt    )
Some authors use ω or ωt as if they are random variables, while the proba-
bility community insists that ω is strictly a sample realization. Readers new to
the field need to understand that there is an astonishing lack of unanimity in
notational styles. For example, the probability community will insist that an ex-
pression such as Ef (x ω) has no meaning (ω is not random, so the expectation
is not doing anything), while others have no objection to an expression such as
this. The debate can be avoided by writing Ef (x W ) where W is our random
“information variable”. Readers in engineering might feel that such debates
are somewhat pedantic, but the issues become more relevant for researchers
who wish to make fundamental theoretical contributions. Our position is to
adopt the simplest possible notation that is as mathematically correct as possi-
ble.
In time-staged problems, it is often necessary to represent the information
that we know up through time t. For this, we suggest the notation
Ht = the history of the process, consisting of all the information
known through time t
= (W1  W2      Wt )
Ht = the set of all possible histories through time t
= Ht (ω) | ω ∈ Ω 
ht = a sample realization of a history
= Ht (ω)
Later, we are going to need to define the subset of Ω that corresponds to a
particular history. For this purpose, we define
 
Ωt (ht ) = ω | W1 (ω) W2 (ω)     Wt (ω) = ht  ω ∈ Ω  (3)
300 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

The notation for modeling the information up to a particular time period is

not standard, and readers need to be prepared for a range of different nota-
tions in the literature on stochastic optimization.
Our interest is in systems where exogenous information and decisions occur
sequentially. It might be useful for some readers to think of “decisions” as
endogenously controllable information classes. If ωt is a sample realization of
the information arriving during time interval t, then xt is the decision made
at time t (using the information up through time t). Let St be the state of
our system immediately before a decision is made (which means that it also
contains ωt ). It is customary to represent the evolution of the state variable
using
St+1 = S M (St  xt  ωt+1 ) (4)
The superscript “M” refers to “model” since many authors refer to Equa-
tion (4) as the “plant model” or “system model”. We note that some authors
will write Equation (4) as St+1 = S M (St  xt  ωt ). This means that t refers to
when the information is used, rather than the information content (we refer
to this as an actionable representation, while Equation (4) is an informational
representation). For simple problems, either representation works fine, but as
richer problems are addressed, we believe that the informational representa-
tion will prove more effective.
The state variable St is the information available right before we make a
decision, so we might call it the pre-decision state variable. For algorithmic pur-
poses, we sometimes need to use the post-decision state variable (denoted Stx ),
which is the state right after we make a decision, before any new information
arrives. The relationship between pre- and post-decision state variables is de-
scribed using
Stx = S Mx (St  xt ) (5)
 
St+1 = S MW Stx  Wt+1  (6)
This representation becomes particularly useful when using approximate dy-
namic programming methods (see Section 8.3). Finally, it is sometimes conve-
nient to work only with the post-decision state variable. If this choice is made,
we suggest simply defining St to be the post-decision state variable (without
the superscript x), and writing the transition function using
St = S M (St−1  ωt  xt )
Note that we are using the same function S M (·), but we change the order of the
information and decision variables, reflecting the order in which they occur.

4.3 Lagged information processes

When we are solving deterministic models, we assume all information is

known in advance. By contrast, most stochastic models assume that the infor-
mation about an event (in particular a customer demand) becomes known at
 Ch. 5. Dynamic Models for Freight Transportation 301

the same time as the event actually happens. It is often the case, however, that
the information about an event arrives at a different time than the physical
event. The most common example is customers calling in orders in advance,
but there are examples where the information arrives after an event (the dis-
patch of a truck might be entered into the computer hours after a truck has
left).
We refer to these problems as instances of lagged information processes, re-
flecting the more common instance when information arrives before an event.
This is quite common in freight transportation, and represents an important
area of study. For example, carriers may be able to charge extra for last-minute
requests, so they need to be able to estimate the cost of the higher level of un-
certainty.
In general, a customer might call in a request for service, and then further
modify (and possibly cancel) the request before it is finally served. For our
purposes, we are going to treat the information in the initial phone call as
accurate, so that when we receive information about a future event we can
then model that event as being fully known. We introduce two terms that help
us with these problems.
Knowable time – This is the time at which the information about a physical
resource arrives.
Actionable time – This is the time at which a physical resource can be acted on.
We assume throughout that the knowable time arrives at or before the action-
able time. The information process is assumed to bring to us the attributes of
a resource that might not be actionable. This can be a customer booking an
order in advance, or information that a vehicle has departed one location (the
information event) headed for a destination in the future (at which point it
becomes actionable).
The actionable time can be viewed as nothing more than an attribute of a
resource. We can pick one element of the attribute vector a, and give it a name
such as aactionable . However, it is often useful to explicitly index the actionable
time to simplify the process of writing flow conservation constraints, since it
will usually be the case that we can only act on a resource (with a decision) that
is actionable. For this reason, we let
Rtt  a = the number of resources that are known at time t to be
actionable at time t   t with attribute vector a
Rtt  = (Rtt  a )a∈A 
Rt = (Rtt  )t  t 
Thus, we can continue to use Rt as the resource that we know about at time t. If
we write Rta , then we implicitly assume that the actionable time is an attribute.
If we write Rtt  or Rtt  a then we assume that t  is the actionable time.
The double time index (t t  ) provides an explicit model of the information
process (the first time index) and the physical process (the second time in-
302 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

dex). For deterministic models, we could (clumsily) index every variable by R0t ,
whereas if we want to model processes where the information arrives as the
physical event occurs, variables would all be indexed by Rtt . Clearly, in both
cases it is more natural and compact to use a single time index.
The double time indexing can be used elsewhere. We might plan an activity
at time t (implicitly, using the information available at time t) that will happen
at time t  in the future. Such a decision vector would be denoted xtt  , where
xtt is an action (something implemented right away) while xtt  , t  > t, would
represent a plan of an activity that might happen in the future. We can lock in
such plans (normally we might do so for t  within some decision horizon), or
treat them simply as forecasts of activities in the future (similar to forecasts of
customer demands). As we did with Rt , we can let

xt = (xtt  )t  t 
ct = (ctt  )t  t 
This allows us to retain a compact notation with a single time index, as long as
it is understood that these variables can be vectors extending over a range of
actionable times in the future.
Although we have not formally introduced our cost function, we can also
let ctt  be the cost of an activity planned (using the information available) at
time t. We note that the flow of dollars is itself a separate process from the
flow of information and physical flows. Normally, payment for a service comes
at some time after the service is provided (and as with the flow of the physical
resource, there is some uncertainty associated with the financial flows). We are
not going to further address the flow of money, but a complete dynamic model
would separate the timing of the informational, physical and financial flows.

4.4 Information and sigma-algebras

Students of mathematical programming learn that they have to understand

linear algebra and the meaning of equations such as Ax = b. People who work
in the area of dynamic information processes (alternatively, stochastic, dy-
namic models) learn the meaning of statements such as “the function Xt must
be Ft -measurable”. In this section, we undertake a brief explanation of this
concept.
We first need to emphasize that measure theory is an abstract part of the
field of mathematics that was adopted by the probability community (just as lin-
ear algebra was adopted by the math programming community). As a general
rule, it is not necessary to have a strong command of measure theory to work
successfully in the field of stochastic, dynamic models (in contrast with linear
algebra which is essential to work in the field of math programming), but it
will add richness to a presentation (and is essential for some theoretical work).
Most important is the notion of a concept known as a “sigma-algebra” (often
designated σ-algebra) and its role in conveying information. In engineering,
 Ch. 5. Dynamic Models for Freight Transportation 303

information is simply data; to a probabilist, information is always represented

as a sigma-algebra.
If the random variables Wt were discrete, we could describe all the elements
of Ω as “outcomes”, with p(ω) the “probability of outcome ω”. In general, this
is not the case (random variables can be continuous), in which case we cannot
refer to ω as an event. Instead, an event has to be a set of outcomes ω. For
the moment (we refine this concept later), let F be all possible subsets of Ω,
with the property that every union of sets in F is an element of F , and every
complement of F is a member of F (the set Ω and the empty set φ would both
be elements of F ). Since every element of F is actually a set of events, F is,
literally, a “set of sets”. F is called a “sigma-algebra”.
It is clear that for every set of outcomes Ω, we can define a sigma-algebra F ,
and for every element in F , we can define a probability. We let P be a function
(which we will write as p(ω)) defined over Ω that allows us to determine the
probability of an event in F (if Ω is discrete, we just have to add up p(ω) for
each ω in an event E ∈ F ). If there are continuous random variables, then
p(ω) will be a density function that we have to integrate over.
Given the set of outcomes Ω, events F and a probability measure, we have
all the elements of a probability space which we write (Ω F  P ). It is not un-
common to find papers which will start by defining a generic probability space
(Ω F  P ), without actually specifying what these are (nor do they refer to it
again). For engineering applications, the outcomes Ω should be clearly defined
in terms of actual information events.
This generic notation applies to any stochastic problem, but we are in par-
ticular interested in problems which involve information that arrives over time.
Assume that we are at time t (the information during time interval (t −1 t) has
already arrived). We can define a sigma-algebra Ft consisting of all the events
that can be created just using the information available up through time t. This
is a fairly subtle concept that is best illustrated by an example. Assume we have
a single customer which may or may not make a request in time period t. Our
information Wt , then, is the random demand that we might call Dt (if this was
our only source of exogenous information, we would just use the variable Dt
for our information). In our simple example, Dt can be zero or one. If we are
considering three time periods, then our sample space Ω consists of eight po-
tential outcomes, shown in Table 1.
Now assume that we have just finished time interval 1. In this case, we
only see the outcomes of D1 , which can be only 0 or 1. Let E{D1 } be the
set of outcomes ω that satisfy some logical condition on D1 . For example,
E{D1 =0} = {ω|D1 = 0} = {1 2 3 4} is the set of outcomes that corre-
spond to D1 = 0. The sigma-algebra F1 would consist of the set of sets
{E{D1 =0}  E{D1 =1}  E{D1 ∈(01)}  E{D1 ∈(01)}
/ }.
Next assume that we are in time period 2. Now our outcomes consist of the
elementary events in H2 = {(0 0) (0 1) (1 0) (1 1)}, and the sigma-algebra
F2 would include all the unions and complements of events in H2 . For exam-
ple, one event in F2 is {ω|(D1  D2 ) = (0 1)} = {3 4}. Another event in F2 is
304 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

Table 1.
Set of demand outcomes

Outcome Time period

ω 1 2 3

1 0 0 0
2 0 0 1
3 0 1 0
4 0 1 1
5 1 0 0
6 1 0 1
7 1 1 0
8 1 1 1

the {ω|(D1  D2 ) = (0 0)} = {1 2}. A third event in F2 is the union of these
two events, which consists of ω = {1 2 3 4} which, of course, is one of the
events in F1 . In fact, every event in F1 is an event in F2 , but not the other way
around. The reason is that the additional information from the second time
period allows us to divide Ω into a finer set of subsets. Since F2 consists of
all unions (and complements), we can always take the union of events which
is the same as ignoring some piece of information. By contrast, we cannot di-
vide F1 into a finer partition. The extra information in F2 allows us to filter Ω
into a finer set of subsets than was possible when we only had the information
through the first time period. If we are in time period 3, F will consist of each
of the individual elements in Ω, as well as all the unions needed to create the
same events in F2 and F1 .
In other words, additional information allows us to divide our set of out-
comes Ω into finer sets of subsets. This is the reason why a sigma-algebra is
viewed as representing information.
With each additional time period, the information from that time period
allows us to filter Ω into finer and finer subsets. As a result, we can write
F1 ⊆ F2 ⊆ · · · ⊆ Ft . We term this a set of “increasing sub-sigma-algebras”.
When this occurs, we term the sequence (F1  F2      Ft ) a filtration. This is
the reason why the notation Ft is used rather than a more mnemonic notation
such as Ht (as in “history”).
The sequence of sub-sigma-algebras is especially important, even if the
mechanism will appear clumsy (not just at first – even experienced profession-
als feel this way). We can use this notation to specify that a sequential decision
process is not cheating by using information in the future. When this is the
case, we say that the decision rule is nonanticipative. Let Xt be a function that
determines a decision at time t (more precisely, using the information available
at time t). As further illustration, assume that we are faced with the problem
of whether or not to dispatch a single truck with customers or freight waiting
to move over a single link. In this case, Xt ∈ (0 1), where Xt = 1 means to dis-
patch the truck and Xt = 0 means to hold it another time period. We further
 Ch. 5. Dynamic Models for Freight Transportation 305

assume that the function Xt uses some well-defined rule to determine when
the vehicle should be dispatched.
Our exogenous information process consists of At which is the number of
customers arriving to move over our link. This problem is much too complex
to use as a numerical example, but all the same concepts apply. However, this
time we are going to define a set of events in a more meaningful way. Instead
of a sigma-algebra that includes every possible event, we only care about the
events that correspond to specific decisions. For example, for the decision X1 ,
we might assume that
1 if A1  y
X1 =
0 otherwise
where y is a prespecified parameter. We do not care about every possible out-
come of A1 ; we only care about whether it is greater than or equal to y, or less
than y. Let E{A1 y} = {ω | A1 (ω)  y}, with E{A1 <y} defined similarly. We can
define a new sigma-algebra which we will call F1X which consists of the events
{E{A1 y}  E{A1 <y}  Ω φ}. This is much smaller than F1 , but contains the events
that are relevant to our decision. We would say that F1X is the sigma-algebra
generated by the decision function X1 . Clearly, F1X ⊆ F1 . Although it gets more
complicated, we can continue this idea for other time periods. For example, we
can think of a sequence of decisions (X1  X2  X3 ). Each possible sequence, say
(0 1 0), corresponds to a set of realizations of (A1  A2  A3 ). We can build a
set of events F3X around all the possible sequences of decisions.
We have now built two sets of sigma-algebras. The original generic one,
Ft which consists of all the possible events created by the sequence (A1     
At ), and our new one, FtX which identifies the events relevant to our decision
function, where FtX ⊆ Ft . Since we have a probability measure P defined on
(Ω F ), we can compute the probability of any event in Ft . This means that we
can compute the probability of any event in FtX . When this is the case, we say
that “Xt is Ft -measurable” because the sigma-algebra generated by the deci-
sion function Xt creates sets of events that are already elements of the original
sigma-algebra (where we already know the probability of each event). It is com-
mon in stochastic optimization to create a decision function Xt and require it
to be “Ft -measurable”. Equivalent terms are to say that Xt is Ft -adapted, or
that Xt is nonanticipative. All of these mean the same thing: that the decision
function has access only to the information that has arrived by time t.
When would this not be the case? Let’s say we cheated, and built Xt in a way
that it used At+1 . While this is not physically possible (you cannot use infor-
mation that has not arrived yet), it is possible in computer simulations where,
for example, we are solving a problem that happened in history. We might be
running simulations to test new policies on historical data. It is possible in this
situation to “cheat”, which in formal terms could be called a “measurability
violation”. If a decision was able to see information in the future, we would be
able to create a finer grained set Ft . If some readers feel that this is a clumsy
306 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

way to communicate a basic idea, they would have a lot of company. But this is
fairly standard vocabulary in the stochastic optimization community.
Another way to identify a valid set of decisions is to identify, for each out-
come of the exogenous information ω (in our example, the realization of the
number of arrivals in each time period), the corresponding set of decisions.
Just as we can identify a set of events on the sequence (A1  A2      At ), we
can define sets of events on the sequence (X1  X2      Xt ). For each ω ∈ Ω,
there is a sequence of decisions (x1 (ω) x2 (ω)     xt (ω)). For lack of bet-
ter notation, let Gt be the sigma-algebra created by all the possible outcomes
of (X1  X2      Xt ). Then, we have a valid information process (that is, no
cheating), if Gt ⊆ Gt+1 (that is, it forms a filtration). More than an abstract
mathematical concept, the field of stochastic programming uses this relation-
ship to develop an explicit set of constraints to ensure a valid set of decisions
(see Section 8.2.2).

5 Decisions

Dynamic systems evolve because of information that arrives from outside

the system (“exogenous information processes”) and decisions (which can be
viewed as “endogenous information processes”). The challenge with decisions
is determining how they are made, or how they should be made.
For our presentation, we focus purely on decisions that act on resources (this
excludes, for example, pricing decisions). This means that all our decisions are
constrained by resources, and their impact on the future is reflected in the
resource constraints of future decisions.

5.1 Representing decisions

Although transportation problems are often viewed as consisting of deci-

sions to move from one location to another, real problems also include a vari-
ety of other decisions, such as cleaning or repairing equipment, buying/selling
equipment, sending crews home on rest, serving customers, loading/unloading
equipment, reconfiguring equipment, moving equipment between customer-
controlled equipment pools, and so on. To handle this generality we define

C D = set of decision classes (move to a location, clean/repair,

serve a customer request, buy, sell)
D = set of decision types for class c
c

D= Dc 
c∈C D

It is often necessary, for practical purposes, to define decision sets that depend
on the attributes of the resource (which might be a layered resource). For this
 Ch. 5. Dynamic Models for Freight Transportation 307

purpose we let
Da = set of decisions that can be used to act on
a resource with attribute a
An element d ∈ D is a type of decision. We measure the number of decisions
made of each type using
xtad = the number of resources of type a that a decision d is applied
to using the information available at time t
xt = (xtad )a∈Ad∈Da 
It is useful to contrast our indexing of decisions. It is common in determinis-
tic models in transportation to use a variable such as xtij to be the flow from
i to j departing at time t. In a stochastic model, the outcome of a decision
might be random, especially when we are working with vectors of attributes
(for example, one attribute might indicate the maintenance status of a piece
of equipment). The notation xtad indexes the decision variable by information
known when the decision is made, and not on the outcome (the notation for
representing the outcome of a decision is given in Section 6).

5.2 The decision function

The real challenge, of course, is determining how to make a decision. We

assume that we have a decision function that we represent using
Xtπ (It ) = a function that returns a vector xt given information It 
It = the set of functions representing the information available
at time t
Π = a family of decision functions (or policies)
The decision function is required to return a feasible decision vector. We let
Xt = the set of feasible decision vectors given the information
available at time t
For example, we would expect Xt to include constraints such as

xtad = Rtta  ∀a ∈ A (7)
d∈Da
xtad  utad  ∀a ∈ A d ∈ Da  (8)
xtad  0 ∀a ∈ A d ∈ Da  (9)
Of course, other constraints might arise in the context of a particular appli-
cation. We note, however, that it is very common in deterministic models to
treat as constraints serving a customer, possibly within a time window. These
308 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

are rarely expressed as constraints in dynamic models, simply because it is too

easy for the model to be infeasible. We encourage limiting Xt to true physi-
cal constraints (such as flow conservation or hard constraints on flow), and let
other “desirable behaviors” (such as serving a customer on time) be handled
through the objective function.
The information It is all the data needed to make a decision, and only
the data needed to make a decision. There is, of course, a close relationship
between the information available and the type of decision function that is
used. In our presentation, it is useful to identify four major classes of deci-
sion functions that reflect the type of information that is available. Each class
corresponds with a well-established algorithmic strategy. The four classes and
corresponding algorithmic strategies are summarized in Table 2. The first class
uses what we know now, which is a standard myopic model. This covers the vast
array of deterministic models which can be solved by algorithms developed for
this problem class. This class, of course, is the foundation for the three other
classes, which are all assumed to use this information (and these algorithmic
strategies), augmented by the information listed for each class.
The second class includes forecasts of future information (this is not events
in the future, but rather information that has not yet arrived). When we in-
clude information about the future, we can include a point forecast (which is
the most widely used) or a distributional forecast. Point forecasts produce the
rolling horizon procedures that are widely used in engineering practice (algo-
rithmically, these can be solved using the same techniques as myopic models,
but the problems are usually quite a bit larger). Alternatively, we can use a dis-
tributional forecast (where we explicitly represent multiple future scenarios)
which puts us in the realm of stochastic programming. These are widely used
in financial asset allocation, but have not been successfully applied to problems
in transportation and logistics.
The third information class represents a forecast of the impact of a decision
now on the future. This puts us into the framework of dynamic programming,
where the future impact of a decision now is captured through a value function

Table 2.
Four classes of information and corresponding algorithmic strategies

Information class Algorithmic strategy

1 Knowledge: The data that describes what we know Classical deterministic math programming
about the system now (myopic models)
2a Forecasts of exogenous information processes Rolling horizon procedures
(point forecasts)
2b Forecasts of exogenous information processes Stochastic programming
(distributional forecasts)
3 Forecasts of the impacts now on the future (value Dynamic programming
functions)
4 Forecasts of future decisions (patterns) Proximal point algorithms
 Ch. 5. Dynamic Models for Freight Transportation 309

(also known as a cost-to-go function). A value function can be viewed as the

value of being in a particular state in the future, or for some classes, a forecast
of a future dual variable.
The fourth and last information class is a forecast of a future decision. This
can be accomplished in several ways, but a simple example is of the form “we
normally put this type of locomotive on this type of train” or “we normally
put our sleeper teams on long loads”. Each of these examples summarizes
a behavioral pattern for a company, represented in a form of an aggregated
attribute–action pair (it is a type of decision acting on a resource with a subset
of attributes).
It is possible, of course, to mix and match algorithmic strategies, but any
reasonable strategy would start with the first information class and then seek
to improve the solution by adding additional information classes.

6 System dynamics

We next have to model the impact of decisions on the system over time. We
represent these effects using a device called the modify function which works
at the level of an individual resource. The modify function can be represented
as the mapping
M(t a d) → (a  c τ) (10)
Here, we are acting on a resource with attribute vector a with decision type d
at time t (which also defines the information available when the decision is
implemented). a is the attribute of the resource after the decision, c is the
contribution (or cost), and τ is the time required to complete the decision. It is
useful to also write the results of the modify function as functions themselves
 
M(t a d) → aM (t a d) c M (t a d) τM (t a d)  (11)
Here aM (t a d) is the attribute transition function, where the superscript “M”
is used to help identify the difference between the attribute vector a and the
attribute transition function aM (t a d). It is common to use ctad = c M (t a d)
and τtad = τM (tad).
There are many problems where the evolution of the attribute vector is de-
terministic. This is particularly true if the only source of randomness is external
customer demands. But many problems exhibit behaviors such as equipment
failures or random delays due to weather or traffic. For these situations, we
need to model the evolution of the attribute vector at just as we model the
evolution of the state variable St . The attribute vector at in (11) would be a
pre-decision attribute vector. We would change our attribute transition func-
tion so that it parallels our transition function, where we might write

at+1 = aM (at  dt  Wt+1 )

310 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

where Wt+1 is the exogenous information that describes equipment failures,

delays, or the decision by a customer that a vehicle is dirty or unacceptable.
Any model that manages resources over time would have software that
roughly corresponds to the modify function (which might be a single routine
or several performing component tasks). The modify function can be arbitrary,
rule-based logic, and as such is extremely flexible. However, this is very difficult
to use when writing equations, so for this purpose we define

δtt  a (t a d) = 1 if aM (t a d) = a 

0 otherwise
The δ function is simply an indicator function that captures the outcome of
a decision. We note that it is possible to capture uncertainty in the outcome
of a decision. This is one reason why our decision variables xtad are indexed
with information known when a decision is made, and not the outcome of the
decision. While this somewhat complicates tasks such as summing the flows
into a location, it provides an important level of generality (for example, the
time required to complete a decision might easily be random).
To model the resource dynamics, we are going to assume that only the vector
xtt is implementable (that is, if we are choosing a plan xtt  , t  > t, these plans
are ignored and replanned in the next time interval). We also assume that xtt
can only act on actionable resources (that is, those in the vector Rtt , as we did
in Equation (7)). Our resource dynamics are given by
 
Rt+1t  a = Rtt  a + δtt  a (t a d)xtad
a∈A d∈Da

t+1t  a 
+R t  > t (12)

We note that the right-hand side of (12) does not include the case t = t  .
Resources in the vector Rtt are all captured by the decision vector xt (including
the resources where we are doing nothing). Reading across, this equation says:
the resources that we will know about at the end of time period t + 1 which are
actionable at time t   t + 1 consist of (a) the resources that are actionable at
time t  that we knew about at time t (but could not act on), (b) the resources
that were actionable at time t (we assume we act on everything, even if the
action is to “do nothing”), and (c) the new resources entering the system during
time interval t + 1 which are actionable at time t  .
Although it is clearer to write out this equation in this form, it is more com-
pact to write it out in matrix form. Let Δtt  be the matrix where δtt  a (t a d)
is the element in row a , column (a d). As before, xtt  is a vector with element
(a d). We can then rewrite (12) as

t+1t  
Rt+1t  = Rtt  + Δtt  xtt  + R (13)
 Ch. 5. Dynamic Models for Freight Transportation 311

7 An optimization formulation

We are now ready to state our problem in the form of an optimization

model. Our problem is to maximize the total contribution received over all
time periods. Our contribution function in period t is given by

Ct (xt ) = ctad xtad 
d∈D a∈A
We can write our decision model in the general form
Xtπ (Rt ) = arg max Ctπ (xt ) (14)
x t ∈Xt

Note the difference between the actual contributions in a time period, Ct (xt ),
and the contributions that we are optimizing over to obtain a set of decisions,
Ctπ (xt ). Later, we provide specific examples of Ctπ (xt ), but for now it suf-
fices to establish the relationship between a policy and the specific contribution
function being used to make a set of decisions. To establish our objective func-
tion, we first define

T
  

Ftπ (Rt ) = E Ct  Xtπ Rt 
t  =t
Ftπ (Rt )is the expected contribution from following policy π starting at time t
in state Rt . The best policy is found by solving
F0∗ (S0 ) = sup F0π (S0 ) (15)
π
We note that we use the supremum operator to handle the mathematical pos-
sibility that we are searching over a class of policies that might not be bounded.
For example, a policy might be characterized by a parameter that is required
to be strictly greater than zero, but where zero produces the best answer. Or,
the optimal parameter could be infinite. If the set of policies is contained in a
compact set (that is, a set that is closed and bounded), then we can replace the
supremum operator by a simple maximization operator.
Equation (15) is computationally intractable, so we generally find ourselves
looking at different classes of policies, and finding the best within a class. If this
were a deterministic problem (that is, no new information arriving over time),
we could write it in the following form

max ct xt (16)
x
t∈T
subject to

At xt − t 
Δt  t xt  = R (17)
t  <t
xt  0 (18)
312 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

In this formulation, the decision variable is the vector of decisions xt . When

the problem is stochastic, it becomes a lot harder to write down the decision
variables, because the decisions we make will depend on the state of the system
at time t or, more generally, on the particular scenario. For this reason, we
cannot simply maximize the expectation, because the expectation is summing
over many scenarios, and we can make different decisions for each scenario
(this would not be the case if the decision is a static parameter such as the
capacity of a facility or the size of the fleet of vehicles).
An alternative approach is to assume that instead of choosing all the deci-
sions, over all the time periods, over all the scenarios, that we are going to find
a single decision function Xtπ (It ) that returns a decision given the informa-
tion It . The challenge is that this single function has to work well across all the
scenarios (which produces different information sets). However, if the function
is defined to work only with the information available at time t, then it will au-
tomatically satisfy the various measurability/nonanticipativity conditions that
we require. Most people will generally find that this approach is more natural.
When we use this approach, our optimization problem looks like
 
π
max E ct Xt (It )  (19)
π∈Π
t∈T ph

where T ph is the set of time periods that define our planning horizon (say,
t = 0     T ). Unlike classical optimization, we do not specify all the con-
straints, since we already require that the decision function produce decisions
that satisfy the constraints Xt at a point in time. As a result, we have only to
specify constraints that govern the evolution of the system over time, as with
Equation (12).
We are not going to attempt to solve Equation (19) exactly. We quickly give
up the search for optimal solutions (or even optimal policies) if we can find
“good” policies that work well over as many scenarios as possible. Such policies
are often referred to as robust. In the next section, we address the problem of
finding good policies.

8 Algorithmic strategies

Unlike deterministic optimization problems where there is a well-defined

concept of an optimal solution (in the case of integer programs, these can be
difficult to both find and identify, but at least they are well defined). In the
dynamic world, there will be valid disagreements over what even constitutes a
valid policy.
In this section, we summarize four major classes of policies, and offer these
as choices for modelers to consider and adapt to their needs. These are comple-
mentary strategies, since they can be used in various combinations depending
on the characteristics of the problems at hand. Each policy is going to use a
 Ch. 5. Dynamic Models for Freight Transportation 313

new class of information, and open up a new algorithmic strategy. But all the
strategies start with the first strategy.

8.1 Myopic models

The most basic strategy on which all other methods build is a myopic policy,
which optimizes over information that is known now, but extends into the fu-
ture (we might take a driver available three hours from now and assign him to
a load that will be available five hours from now). The resulting model can be
stated as

max c0t  x0t  (20)
x0
t  ∈T ph
subject to
A00 x00 = R00  (21)
 −1
t
A0t  x0t  − 0t  
Δt  t  x0t  = R t   1 (22)
t  =0
x0t   0 and possibly integer (23)
Here, x0t  is a vector of decisions to be implemented at time using the infor- t 
mation available at time 0. The matrix Δt  t  is an indicator matrix that tells us
when a decision implemented at time t  produces a resource that is actionable
at time t  . Technically, we need a third time index to show that Δt  t  is computed
with information available at time 0.
With this generic model we include the entire range of classical optimiza-
tion problems, and algorithms that have been studied in transportation and
logistics. These might be simple assignment problems, vehicle routing prob-
lems, network design problems and multicommodity flow problems. In general,
however, these are much smaller than the types of problems that arise in plan-
ning applications. We assume that this basic myopic problem is solvable to an
acceptable level of optimality. This is the fundamental building block for all
other modeling and algorithmic strategies.
There are some problems where it is possible to provide effective solutions
using a simple myopic model. Assigning taxi drivers to customers, assigning
long-haul truckload drivers to loads, and the dynamic routing and scheduling
of pickup and delivery trucks in a city. But our experience has been that while
these models can be effective, they are always missing something.

8.2 Rolling horizon procedures

We can extend our basic myopic model by including information that is fore-
casted to come in the future. A common procedure is to use a point forecast in
a rolling horizon procedure (Section 8.2.1), but we can also use a distributional
forecast in a stochastic programming algorithm (Section 8.2.2).
314 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

8.2.1 Deterministic models

The most common approach to combining current knowledge with fore-
casted information is to use a point forecast of future events which produces a
classic rolling horizon procedure using point forecasts. Let Rtt  be the forecast
of information that becomes available during time interval t that is actionable
at time t  . We note that these forecasts are known at time 0. Using these fore-
casts, our optimization problem becomes
   
max (c0t x0t ) + ct  x0t  (24)
x0
t∈T ph t  ∈T ph \0
subject to
A00 x00 = R00  (25)

t−1 
t
A0t x0t − Δt  t x0t  = t  t 
R t > 0 (26)
t  =0 t  =0
x0t  0 for t ∈ T ph  (27)
Note that the right-hand side of Equation (26) is the total number of resources
that are actionable at time t  . To obtain this, we have to sum over all the infor-
mation that is forecasted to arrive in future time periods. It is interesting that
many practitioners view this as the “correct” way to solve dynamic problems.
They recognize that point forecasts are not perfect, but argue that you simply
reoptimize as new information comes in.
Rolling horizon procedures have the primary benefit that they use classi-
cal modeling and algorithmic technologies. In addition, point forecasts are the
easiest to understand in the business community. But as with myopic models,
they will over time start to show their weaknesses. For example, rail customers
often exhibit highly variable demands. A forecasting system will tend to fore-
cast the average demand, while a seasoned operations manager will know that
you need to supply the customer more than the average to cover the days when
the demand is high. Waiting until the last minute might not provide enough
time to move cars to satisfy the demand.
Another problem class where point forecasts perform poorly are discrete
routing and scheduling. Here, a customer demand will normally be zero but
will sometimes be one. A point forecast would produce a fraction such as 0.2. It
is not possible to route an entire vehicle to each customer with a nonzero (frac-
tional) demand forecast. Randomly sampling the demands is nothing more
than routing to a random outcome. Allowing a model to send a fraction of
a vehicle to a fraction of a demand will create an extremely difficult integer
program for the first period.

8.2.2 Stochastic programming

If we wish to explicitly model the possibility of multiple outcomes, we find
ourselves in the domain of stochastic programming. Now, we have to allow for
 Ch. 5. Dynamic Models for Freight Transportation 315

the possibility that different decisions will be made in the future depending on
the actual evolution of information.
In this case, our optimization problem becomes
     
max (c0t x0t ) + p̂(ω) ctt  xtt  (ω) (28)
x
t∈T ph 
ω∈Ω t∈T ph \0 t  >t

subject to, for t t  ∈ T ph :

(1) First stage constraints:
A00 x00 = R00  (29)
t  −1

A0t  x0t  − Δt  t  x0t  = R0t   t  > 0 (30)
t  =0

x0t   0 t  t (31)
 t t  ∈ T ph \ 0, t   t:
(2) Later stage constraints, for all ω ∈ Ω,
Att (ω)xtt (ω) = Rtt (ω) (32)
 −1
t
Att  (ω)xtt  (ω) − tt  (ω)
Δt  t  (ω)xtt  (ω) = R (33)
t  =t
xtt  (ω)  0 (34)
This formulation requires that we choose a vector of decisions for the first time
period x0t but allows a different set of decisions for each scenario in the later
time periods. This might not be a bad approximation, but it does allow the
decisions in later time periods to “see” future information. This happens be-
cause we are indexing a decision by an outcome ω which implicitly determines
all future information. In the vocabulary of stochastic programming, we pre-
vent this by adding what are known as nonanticipativity constraints. We can
express these by using some of our notation (and concepts) from Section 4.4.
As before, let ht = (ω1  ω2      ωt ) be the history of the process up through
time t, and let Ht be the set of all possible histories. In addition, define the
subset Ωt (ht ) = {ω ∈ Ω | (ω1  ω2      ωt ) = ht } to be the set of scenarios

ω ∈ Ω which match the history ht up through time t. We would then add the
constraint
xt (ht ) = xt (ω) ∀ω ∈ Ωt (ht ) ht ∈ Ht  (35)
Equation (35) is our nonanticipativity constraint, which has the effect of forc-
 to form a filtration. Literally, we want
ing the set of decisions (x(ω) ω ∈ Ω)
each set of decisions with the same history to be the same.
For most transportation problems, the optimization problem (28)–(35) is
too large to be solved, although special cases exist (see, for example, Morton
et al., 2003; Glockner and Nemhauser, 2000).
316 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

8.3 Dynamic programming

A powerful technique for taking into account uncertain future events is dy-
namic programming, which uses the construct of a value function to capture
the expected value of being in a particular state in the future. If we can let
Rt represent the state of our system, we can calculate the value Vt (Rt ) of being
in state Rt from time t onward using the well-known Bellman equations
  
 
Vt (Rt ) = max Ct (xt ) + P(R |Rt  xt )Vt (R )  (36)
x t ∈Xt
R  ∈R

where P(R |Rt  xt ) is the probability of being in state R if we are in state Rt

and take action xt . A more general way of writing Equation (36) (and more
useful for our purposes here) is using
 
Vt (Rt ) = max Ct (xt ) + E Vt+1 (Rt+1 )|Rt  (37)
x t ∈Xt

where Rt+1 = Rt + Δt xt + R t+1 . It is generally assumed that we are going

to compute Equation (36) for each possible value of Rt . As we computed in
Equation (1), the number of possible values of Rt is extremely large even for
fairly small problems. This is the well-known “curse of dimensionality” that
has been recognized since dynamic programming was first discovered. For our
problem class, there are actually three curses of dimensionality: the state space,
the outcome space, and the action space. Since new information (such as R t ) is
typically a vector, sometimes of high dimensionality, the expectation in Equa-
tion (37) will generally be computationally intractable. The more conventional
form of Bellman’s equation, shown in Equation (36) hides this problem since
the one-step transition matrix is itself an expectation. Finally, if we compute
Vt (Rt ) for each discrete value of Rt , then the only way to find the optimal xt is
to search over each (discrete) element of the set Xt . Again, if xt is a vector (for
our problems, the dimensionality of xt can easily be in the thousands or tens
of thousands) the size of Xt is effectively infinite.
The dynamic programs depicted in Equations (36) and (37) assume a partic-
ular representation of states, information and action. We can write the history
of our process up through time t as
ht = (R0  x0  ω1  R1  x1      ωt  Rt  xt ) (38)
The evolution of the state variable is often depicted using

Rt+1 = RM (Rt  xt  ωt+1 ) (39)

where RM (·) captures the dynamics of the system. We can write the decision
function as
  
Xtπ (Rt ) = arg max ct xt + E Vt+1 (Rt+1 )|Rt  (40)
x t ∈Xt
 Ch. 5. Dynamic Models for Freight Transportation 317

Since the expectation is typically computationally intractable, we can replace

this with an approximation. Let Ω  ⊂ Ω be a small set of potential outcomes.
This would be written
  
 
Xtπ (Rt ) = arg max ct xt + p̂(ω̂)Vt+1 Rt+1 (ω̂)  (41)
x t ∈Xt

ω̂∈Ω
 The problem
where p̂(ω̂) is the probability of the sampled outcome ω̂ ∈ Ω.

we encounter in practice is that even when Ω is relatively small, it can still
greatly complicate solving (41). The problem we encounter in transportation
and logistics is that the myopic problem
Xtπ (Rt ) = arg max ct xt (42)
x t ∈Xt

can be computationally difficult. This could be a vehicle routing problem or

integer multicommodity flow problem. Adding the summation over multiple
outcomes (as we do in Equation (41)) can make a hard problem much larger
and harder. We could avoid this by using a single sample ωt+1 = Wt+1 (ω),
producing
  
Xtπ (Rt  ω) = arg max ct xt + Vt+1 Rt+1 (ωt+1 )  (43)
x t ∈Xt

Now we are choosing xt given Rt+1 (ωt+1 ) which is like making a decision now
given the information that will arrive in the next time period. This will not even
be a good approximation.
We can overcome this problem by introducing the concept of measuring the
state of the system immediately after a decision is made (or equivalently, before
new information has arrived). We refer to our original state variable Rt as the
“pre-decision” state variable, and we introduce a new state variable, denoted
Rxt as the “post-decision” state variable. We can now write the history of our
process as
 
ht = R0  x0  Rx0  ω1  R1  x1  Rx1      ωt  Rt  xt  Rxt  (44)
If we write the evolution of our system in terms of Rxt , we obtain
 x 
Rxt = RM t Rt−1  ωt  xt  (45)
Writing the Bellman equations around Rxt
we obtain
 x      
x
Vt−1 Rt−1 = E max ct xt + Vtx Rxt (xt ) Rxt−1  (46)
x t ∈Xt

The relationship between the pre- and post-decision value functions is best
illustrated using
 
Vt (Rt ) = max ct xt + Vtx Rxt (xt ) 
x t ∈Xt
 x 
Vt−1 Rt−1 = E Vt (Rt )|Rxt−1 
x
318 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

Using Equation (46) we still encounter the problem of computing the expecta-
tion, but now consider what happens when we drop the expectation and solve
it for a single sample realization ω
 
Xtπ (Rt−1  ωt ) = arg max ct xt + Vtx Rxt (xt )  (47)
xt ∈Xt (ω)

Here, we are finding the decision xt given the information ωt = R t (ω) which
would be available when we make a decision. As a result, we are not violating
any informational (measurability or anticipativity) constraints. What is most
important is that Vtx (Rxt (xt )) is a deterministic function of xt , so we do not
have to work with even an approximation of an expectation (the approximation
is implicit in Vtx (Rxt (xt ))).
We have now crossed a major hurdle, but we still have the problem that we
do not know Vtx (Rxt ). We are going to need to use an approximation which will
be estimated iteratively. At iteration n, we would solve (47) using the approxi-
mation from iteration n − 1
   
Xtπ Rt−1  ωnt = arg max n ct xt + Vtn−1 Rxt (xt )  (48)
xt ∈Xt (ω )

where we drop the superscript x when writing the approximation V. We denote
the function produced by solving the right-hand side of (48) by
    
Vtn Rnt ωn = max n ct xt + Vtn−1 Rxt (xt )  (49)
xt ∈Xt (ω )

A brief note on notation is in order here. Vtn (Rt (ωn )) is an approximation of

x (Rx ), which suggests an inconsistency in how we are indexing by time
Vt−1 t−1
(V x is indexed by t, while V is indexed by t − 1). Recall that our time in-
dexing indicates the information content. Vt−1 x is an expectation conditioned

on Rxt−1 , and therefore contains information up through t − 1. Vtn (Rt (ωn )) is

a conditional approximation, computed using the information that arrives in
time t. It does not represent even an approximation of an expectation (by con-
trast Vtn (Rt ) is an approximation of the expected value function conditioned
on Rt ).
We use the information from solving (49) to update the value function
approximation. How this update is performed depends on the type of approxi-
mation used, so for now we represent this updating process simply as
 n−1 n n  n 
Vt−1
n
= U V Vt−1  Vt  Rt ω  (50)
The algorithm works by stepping forward through time using the value func-
tion approximation from the previous iteration. This is the standard approach
used in the field of approximate dynamic programming. Instead of finding the
value function over all possible states, we develop approximations around the
states that are actually visited. The two textbooks that describe these tech-
 Ch. 5. Dynamic Models for Freight Transportation 319

niques (Bertsekas and Tsitsiklis, 1996; Sutton and Barto, 1998) both describe
these procedures in the context of steady state problems. Operational prob-
lems in transportation and logistics require time-dependent formulations.
There are two strategies for updating the value function (especially when
a time-dependent formulation is used). The first uses a single, forward pass,
updating the value function approximations as we proceed. The steps of this
procedure are outlined in Figure 3. The second uses a two-pass procedure,
stepping forward through time determining states and actions, and a backward
pass that updates the value function using information about the states that
were actually visited. The steps of this procedure are given in Figure 4.
In our applications, we will generally be more interested in derivatives than
the actual value function itself. When we solve the approximation Vt (Rt ) in
Equation (49), we will typically be solving these problems subject to flow con-
servation constraints on the resources as in Equation (7). As a general rule our

Step 0. Initialization:
Initialize Vt0 , t ∈ T .
Set n = 1.
Initialize R1 .
Step 1. Do while n  N:
Choose ω = (ωn1  ωn2      ωnT ).
Step 2. Do for t = 1 2     T :
Step 2a. Solve Equation (49) to obtain Vtn (Rnt ) and xnt .
xn xn n n
Step 2b. Compute Rt = RM t (Rt−1  ωt  xt ).
Step 2c. Update the value function approximations using Vt−1 n−1  Vn  R ).
n ← U V (V xn
t−1 t t−1
π  N
Step 3. Return the policy Xt (Rt  Vt ).

Fig. 3. Single pass version of the approximate dynamic programming algorithm.

Step 0. Initialize Vt0 , t ∈ T .

Set n = 1.
Initialize R1 .
Step 1. Do while n  N:
Choose ω = (ωn1  ωn2      ωnT ).
Step 2. Do for t = 0 1     T − 1:
Step 2a. Solve Equation (49) to obtain Vtn (Rt ) and xnt .
xn xn n n
Step 2b. Compute Rt = RM t (Rt−1  ωt  xt ).
Step 3. Do for t = T − 1 T − 2     1:
Step 3a. Recompute Vtn (Rnt ) using Vt+1 n and the decision xn from the forward pass:
t
 n  n 
 n  n
Vt Rt = ct xt + Vt+1 Rt+1 . (51)
Step 3b. Update the value function approximations, Vn ← U V (Vn−1  Vn  R ).
xn
t−1 t−1 t t−1
Step 4. Return policy Xtπ (Rt  VN ).

Fig. 4. Double pass version of the approximate dynamic programming algorithm.

320 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

problems are continuous but nondifferentiable in Rt . If we are solving linear

resource allocation problems, we can obtain stochastic subgradients just by us-
ing the dual variable for the flow conservation constraint. In general, these dual
variables are subgradients of Vtn (Rt ) only, which depends on Vtn−1 (Rt (xt )).
These can be effective, but the use of approximations from previous iterations
will slow the overall convergence.
Ideally, we would prefer to have gradients of all future profits. Let


T
Vtπ (Rt  ω) = ct  Xtπ (Rt   ω) (52)
t  =t

be the total future profits given an initial resource vector Rt , and sample real-
ization ω, under policy π. We can obtain the gradient of Vtπ (Rt  ω) by storing
the basis obtained as we step forward in time. The basis will allow us to iden-
tify the effect of, say, adding one more unit to an element Rta by giving us the
impact on costs in time period t, and the impact on future resources Rt+1 .
By storing this information as we step forward in time, we can then com-
pute the effect of increasing Rta by one by stepping backward through time,
combining the impact of an additional unit of resource of each type at each
time t   t with the value of one more unit of resource at the next time pe-
riod.

8.4 Incorporating low-dimensional patterns

In practice, it is common to carefully develop a math programming-based

optimization model, only to find that a knowledgeable expert will criticize the
behavior of the model. It is possible, although generally unlikely, that the ex-
pert is simply finding a better solution than the algorithm. More often, the
expert simply recognizes behavior that will incur a cost that is not captured by
the model, or expresses a preference for an alternative behavior because of
other unquantified benefits.
One response to such input is to “fix the model” which means improving
the cost function or adding constraints that capture the physical problem more
realistically. For some problems, this is the only solution. However, it can often
take weeks or months to implement these fixes, which might easily require data
that was not originally available. Often, the complaints about the model can be
expressed in fairly simple ways:
• You should avoid sending that type of cargo aircraft into Saudi Arabia
since we do not have the ability to handle certain types of repairs.
• Do not send Cleveland-based drivers into Indianapolis because their
union is in a different conference and it creates a lot of bookkeeping
problems.
• You should put your sleeper teams on longer loads so that you get
higher utilization out of the two drivers.
 Ch. 5. Dynamic Models for Freight Transportation 321

• Do not use the six-axle locomotives on certain trains out of Denver

because the track is too curved to handle the long locomotives.
• Try to send flatcars that can handle trailers into Chicago because that
location moves a lot of trailers from trucking companies.
• Avoid sending drivers domiciled in Florida into Texas because we do
not have very much freight out of that location that terminates in
Florida, making it hard to get drivers home.
• Avoid putting inexperienced drivers on service-sensitive loads.
All of these preferences can be stated as low dimensional patterns, each of
which consists of three elements: the attributes of the resource being acted on
at some level of aggregation, the decision being made (also at some level of ag-
gregation), and the fraction of time that the pattern should occur. Aggregation
on the attributes arises because each rule is typically trying to accomplish a par-
ticular objective, such as getting a driver home or avoiding sending equipment
needing maintenance into areas that do not have the proper repair facilities.
A pattern measures how often a particular decision is applied to a group of
resources, and the decisions will also generally be expressed at an aggregate
level. For example, we are not specifying a rule that a type of driver should (or
should not) be assigned to move a particular load of freight. Instead, we are
expressing preferences about types of loads (high priority, loads into a partic-
ular destination). Finally, these “rules” are generally not hard constraints, but
rather expressions of behaviors to be encouraged or discouraged.
Each pattern can be thought of as a preference about the desirability of a
particular type of state action pair, which can be represented as aggregations
on the attribute vector and decision. For this, we define

P = a collection of different patterns

p
Ga = aggregation function for pattern p ∈ P that is applied
to the attribute vector a
p
Gd = aggregation function for pattern p ∈ P that is applied
to the decision d
These aggregation functions create new attribute spaces and decision sets
p
Ap = Ga (a) | a ∈ A 
p p
Da = Gd (d) | d ∈ Da 
For notational compactness, it is useful to let ā represent a generic aggregated
attribute vector, and d̄ an aggregated decision. We use this notation when the
level of aggregation is either not relevant, or apparent from the context. A pair
(ā d̄) represents an aggregated state/action pair (literally, an attribute vec-
tor/action pair). The last element of a pattern is the fraction of time that we
expect to observe decision d̄ when we are acting on a resource with attribute ā,
322 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

which we represent as follows:

p
ρ = the fraction of time that resources with attributes
ād̄
p p
{a | Ga (a) = ā} are acted on using decisions {d | Gd (d) = d̄}
Given a flow vector xt = (xtad )a∈Ad∈D returned by the model, the number of
times that the flow matches a pattern is given by
p
x̄ = Gp (x)
t ād̄
 
= xtad I{Gpa (a)=ā} I{Gp (d)=d̄}  (53)
d
∀a∈A ∀d∈Da

where IX = 1 if X is true. We also need to express the resource state variable

Rt in an aggregated form

p =
R Rta I{Gpa (a)=ā}
t ā
∀a∈A
= the number of resources with attribute ā at time t
 
Rā (x) = xt ād̄ 
t∈T d̄∈Dp
ā

Finally, we define the model pattern flows as

xt ād̄ p
ρt ād̄ (x) = (∀ā ∈ Ap  ∀d̄ ∈ Dā  ∀t ∈ {1 2     T })
Rt ā
= the fraction of time that resources with attribute ā are
acted on with decision d̄ at time t
T
x p
ρād̄ (x) = t=1 t ād̄ (∀ā ∈ Ap  ∀d̄ ∈ Da )
Rā
= the fraction of time that resources with attribute ā are
acted on with decision d̄ over the entire horizon
The static flow fraction after iteration n can also be written as
 n
n t∈T xt ād̄
ρād̄ =
Rnā
 xn Rntā
t ād̄
=
Rntā Rnā
t∈T
 n 
n Rt ā
= ρt ād̄ n  (54)
Rā
t∈T
 Ch. 5. Dynamic Models for Freight Transportation 323

Our goal is to create a solution x̄p that reasonably matches the exogenous
pattern ρp . We do this by introducing the metric
      p 2
H ρ(x) ρ = Rād̄ ρād̄ (x) − ρ  (55)
ād̄
p∈P ā∈Ap d̄∈Dap

We often refer to H(ρ(x) ρ) as the “happiness function” since the domain ex-
pert evaluating the model tends to be “happier” when the model flows reason-
ably match the exogenous pattern. More formally, H(ρ(x) ρ) is a well-defined
distance matrix used in proximal point algorithms to produce solutions. We
now wish to solve
  
max ct Xtπ − θH ρ(x) ρ  (56)
π∈P
t∈T

where θ is a scaling parameter (these might depend on the specific pattern).

Our decision function at time t is given by
  
Xtπ = arg max ct xt − θH ρ(x) ρ 
x t ∈Xt

We are generally unable to solve this problem, so we use instead sequences of

approximations of the pattern metric, which for the moment we can write as
  
Xtπ = arg max ct xt − θH n−1 ρn−1 (x) ρ 
x t ∈Xt

We encounter an algorithmic challenge because the proximal term is defined

over time but we have to solve the problem one time period at a time. We over-
come this problem by solving a series of separable, quadratic approximations
of the pattern metric. We start by computing the gradient of the pattern metric,
which we denote H n (x ρ), by differentiating (55)


Hn (ρ) = ∂H 
ād̄ ∂ρād̄ ρ =ρn
ād̄ ād̄
  p
= 2Rā ρād̄ − ρād̄  ∀ā ∈ Ap  ∀d̄ ∈ Dā 
n n
(57)

where p is the pattern associated with the level of aggregation in ā and d̄.
A convergent algorithm can be designed using a Gauss–Seidel strategy
which optimizes the pattern metric computed by using flows xnt for t  < t, and
xtn−1
 for t  > t, and then optimizing xt . We then solve a separable, quadratic
programming approximation at each point in time.
To describe the algorithm in detail, we first define


t−1  
T 
Rnā (t) = xnt ād̄ + xn−1

t ād̄
t  =1 d̄∈Dp t  =t d̄∈Dp
ā ā
324 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

and
t−1  xn
  T  xn−1 

 t ād̄  t ād̄
ρ̃nād̄ (t) = + 
Rnā (t  ) Rnā (t  )
t  =1 t  =t
p
∀ā ∈ A ∀d̄ ∈ Dā  ∀t ∈ {1 2     T } (58)
The Gauss–Seidel version of the gradient is now
  p
h̃ntād̄ = 2Rnā (t) ρ̃nād̄ (t) − ρād̄  ∀ā ∈ A ∀d̄ ∈ Dā  ∀t ∈ {1 2     T }
(59)
Our algorithm proceeds by iteratively solving subproblems of the form
 
xnt = arg max ctad xtad
a∈A d∈Da
     xt ād̄  

1 n n−1 
−θ h̃ + 2 u − x du 
Rnā (t) p 0
t ād̄ t ād̄
p∈P ā∈Ap d̄∈Dp d̄∈Dā
ā
(60)
This approach has been found to produce solutions that closely match exoge-
nous patterns within as few as two or three iterations.
The real value of exogenous patterns is that it allows an expert to change the
behavior of the model quickly by editing an external data file. This approach
does not replace the need to ensure the best quality model through careful
engineering, but it does allow us to produce more realistic behaviors quickly
by simply changing an external data file.

8.5 Summary: The four information classes

This section has outlined a series of modeling and algorithmic strategies

that can be largely differentiated based on the information being made avail-
able to the model. We started in Section 8.1 by providing a basic myopic model
that only uses the information available at time t. There was no attempt to in-
corporate any information that had not yet arrived. All the remaining models
use some sort of forecast. Section 8.2 illustrates the use of point and distribu-
tional forecasts of future information, which produces classical rolling horizon
procedures. Section 8.3 introduces a forecast of the value of resources in the
future, represented as a value function. This information brings into play the
theory of dynamic programming. Finally, Section 8.4 uses what might be called
a forecast of a decision, expressed in the form of low-dimensional patterns. As
an alternative, we could have avoided mathematical optimization entirely and
resorted instead to an extensive set of rules; this would be the approach that
the artificial intelligence or simulation communities would have adopted. For
transportation problems, the rules simply become too complex. Instead, we
have found that expert knowledge can often be represented in the form of low
 Ch. 5. Dynamic Models for Freight Transportation 325

dimensional patterns, and incorporated using a proximal point term, bringing

into play the theory of proximal point algorithms.
Up to now, we have represented our decision functions as depending on the
resource vector Rt . In practice, what we know at time t can also include other
data such as estimates of system parameters such as travel times and costs,
and functions such as forecasting models. It is useful to summarize “what we
know now” as consisting of data knowledge (such as the resource vector) and
functional knowledge (such as models for forecasting demand). Let νt be our
estimates at time t of various parameters, such as those that govern the modify
function. Our data knowledge at time t is then given by

Kt = (Rt  νt )
We differentiate between a forecast model (which we know now) from an
actual forecast of an event in the future. A forecast is an actual generation of
information that might arrive in the future. Let

ω[tT ph ] = a vector of potential information events over

 
the planning horizon t t + 1     T ph 
Ω[tT ph ] = the set of all potential information events that we wish to
consider over the planning horizon
We feel that it is important to recognize that Vt (Rt ) is a forecast, using in-
formation available up through time t (just as we would from our demand
forecast) of the value of resources in the future. Furthermore, we feel that
Vt (Rt ) is a sufficiently different type of forecast that it should be recognized as
an information class distinct from Ωt .
Finally, we feel that when an expert in operations expresses opinions about
how a system should behave (represented using the pattern vector ρt ), this is
also a form of forecast that brings into play the experience of our expert. ρt can
be thought of as a forecast of a decision that the model should make. It will not
be a perfect forecast, as we will not in general be able to match these patterns
perfectly, but just as we try to match historical patterns using a demand fore-
casting routine, these patterns are a forecast which reflects patterns of past
decisions.
We have, now, four classes of information:

Kt = the data knowledge at time t

Ω[tT ph ] = a set of potential outcomes of exogenous information
events in the future

Vt (Rt ) = forecasts of the impact of decisions now on the future
ρt = forecasts of future decisions, captured in the form
of low-dimensional patterns
326 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

We can now formulate an information set It which will always include Kt , and
can also include any combination of the remaining information sets. If we are
solving the problem using only a rolling horizon procedure, we would specify
It = (Kt  Ω[tT ph ] ). If we want to use dynamic programming, we would write
It = (Kt  Vt (Ωt )) to express the fact that our decision function depends on
value functions which themselves depend on forecasts of future events. All
four information classes would be written: It = (Kt  Vt (Ωt ) ρt ).
We see, then, that there are four fundamental information classes, and each
brings into play specific algorithmic strategies. All the approaches assume that
we are starting with a basic myopic model, and that we have algorithms for
solving these problems. It is possible in theory to combine rolling horizon pro-
cedures and value functions, but value functions do offer a competing strategy
for using forecasts of exogenous information processes to produce a forecast
of a value of a resource in the future. Proximal point terms which incorporate
expert knowledge can complement any model.
We can put all four information classes into a single decision function. As-
sume that we wish to manage equipment to move freight, including freight that
is in the system now as well as forecasted freight within a planning horizon T ph .
We could formulate a decision function that looks like


T T 
ph

ph

Xtπ (It ) = arg max ct  t  xt  t 

x t ∈Xt
t  =t t  =t 
 
 x   
+ 
Vtt  Rtt  (xt ) − H ρ(x) ρ 
t  >T ph

In this formulation, we are planning the flows of equipment within a plan-

ning horizon, using a point forecast of future information within this horizon.
Equipment that becomes actionable past the end of the planning horizon is
captured by a value function approximation, which we have represented as be-
ing separable over time periods. More commonly, optimizing over a planning
horizon using a point forecast is done specifically to avoid the need for a value
function. If value functions are used, these can be particularly nice mechanisms
for handling stochastic forecasts. For this reason, we prefer to write our general
purpose decision function as
T ph 

Xtπ (Rt ) = arg max ctt  xtt 
x t ∈X
t  =t
 
 x   
+ 
Vtt  Rtt  (xt ) − H ρ(x) ρ  (61)
t  >T ph

The idea of manipulating the information set available to a decision function

is illustrated in Wu et al. (2003) in the context of the military airlift problem.
 Ch. 5. Dynamic Models for Freight Transportation 327

This problem, simply stated, involves assigning cargo aircraft to move “re-
quests” which are loads of freight and passengers. The air mobility command
currently uses a simulation package which captures a host of engineering de-
tails, but which uses very simple rules for choosing which aircraft should move
each request. We use a version of these rules to represent the base policy.
These rules work roughly as follows. The program maintains a list of aircraft
in the order they are available, and requests to be moved, also in the order
that they are available. Starting with the request at the top of the list, the pro-
gram then determines whether the first available aircraft is eligible to move
the request (there are a number of engineering restrictions). If not, it moves
to the next aircraft on the list. As requests are satisfied, they are dropped from
the list, and as aircraft are moved, they become available again in the future.
Since a decision of which aircraft to choose does not compare different aircraft
to find the best one, we can view this as a decision with very low information
content. We refer to this as “rule-based, one aircraft, one request”.
The next step up is to choose a request, and then find the best out of a list
of aircraft using a cost function. The list of aircraft that are considered include
only those which are known now and actionable now (that is, we are limited to
aircraft in the vector Rtt ). This strategy is a cost-based, myopic policy, using a
single aircraft and a list of requests. Next is a strategy that considers a list of
requests and a list of aircraft, which requires solving an assignment problem to
determine the best assignment of each aircraft to each request. This strategy
is also restricted to aircraft and requests that are known now, actionable now.
The fourth information set is the same as the third, but now we are allowed
to consider aircraft and requests that are known now but actionable in the
future. The fifth information set includes value functions, which are used to
capture the cost of putting an aircraft into an airbase where it might break
down, incurring maintenance costs that can vary significantly between airbases.
The results of these simulations are shown in Figure 5, demonstrating steady
improvement in costs as information is added to the decision function. We
note that the value of increasing the information set depends on the problem.
In addition, each increase in the information set adds to the computational re-
quirements of the simulation. The use of value functions, for example, requires
running the simulation iteratively, although it is often possible to estimate the
value functions once and store these.
We did not include expert knowledge in this set of experiments since this
information class exists only because the cost function is incomplete in some
way. Imposing exogenous patterns through a proximal point term will always
increase the costs. The appropriate measure here is the degree to which we are
matching an exogenous pattern. To illustrate this process, we imposed a single
exogenous pattern controlling the fraction of time that a particular aircraft
type moved through a particular set of airbases. Without an exogenous pattern,
this activity occurred approximately 20 percent of the time. Figure 6 shows the
percentage of time this activity happened in the model, as a function of the
weighting parameter θ. As θ is increased, the model produces results that are
328 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

Fig. 5. Total costs, broken down between travel costs, late penalty costs, and repair costs, using
four policies: (1) rule-based, chooses a single aircraft and single request and implements if feasible,
(2) cost-based, considers a single aircraft and assigns to the best of a list of requests based on cost,
(3) minimizes the costs of assigning a set of aircraft to a set of requests, limited to those that are both
known and actionable now, (4) same as (3), but now includes aircraft and requests that are known now
but actionable in the future, and (5) includes value functions to guide the choice of aircraft (from Wu
et al., 2003).

closer to the desired pattern, although there is a limit to this. In practice, if an

expert says that we can do something 10 percent of the time, it is very likely
that a solution that does it 15 or 20 percent of the time would be acceptable.
We are not trying to make the argument that all four information classes
must be used in all models, but the exclusion of any information class should
be recognized as such, and our experience has been that each will add value,
especially when working on complex problems.

8.6 Bibliographic notes

There is a vast literature on myopic models and rolling horizon procedures

that use point forecasts (virtually any deterministic dynamic model falls in this
class). A number of authors have developed algorithms for exploiting the dy-
 Ch. 5. Dynamic Models for Freight Transportation 329

Fig. 6. Actual versus desired flow of a particular aircraft through a particular airbase, as a function of
the pattern weight parameter θ.

namic structure of these models (see, for example, Aronson and Chen, 1986;
Aronson and Thompson, 1984) although this work has been largely replaced by
modern linear programming techniques. Stochastic programming models and
algorithms are reviewed in Infanger (1994), Kall and Wallace (1984), Birge and
Louveaux (1997), Sen and Higle (1999). There are numerous texts on dynamic
programming. Puterman (1994) is the best overview of classical methods of
Markov decision processes, but covers only discrete dynamic programs, with
algorithms that work only for very low dimensional problems. The use of the
post-decision state variable has received very little attention. It is used implic-
itly in an exercise in Bellman’s original text (Bellman, 1957), but not in a way
that identifies its importance as a computational device. Godfrey and Powell
(2002) use the post-decision state variable explicitly in the development of
an approximate dynamic programming algorithm for fleet management, ex-
tending an approximation method used in a series of papers (Powell, 1987;
Cheung and Powell, 1996) where the technique was used implicitly. Van Roy
(2001) introduces pre- and post-decision state variables, taking advantage of
the fact that the post-decision state variable avoids the expectation in the de-
termination of an optimal action. This property is not mentioned in any of the
330 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

standard textbooks on exact or approximate dynamic programming (Bertsekas

and Tsitsiklis, 1996; Sutton and Barto, 1998), but it appears to be a powerful
technique for high-dimensional resource allocation problems. The explicit in-
troduction of four classes of information appears to have been first done in
Powell et al. (2001). The investigation of low-dimensional patterns is investi-
gated in Marar et al. (2006) and Marar and Powell (2004), and is applied to the
military airlift problem in Powell et al. (2004).

9 Approximating value functions in dynamic programming

Section 8.3 introduces the idea of using approximate value functions, but
did not fill in any of the details of how to do this. Approximate dynamic pro-
gramming is an emerging field that must be proven experimentally on different
problem classes, just as different heuristics have to be tested and proven on
different problems. This section, which is easily skipped on a first reading, pro-
vides an introduction to the current state of the art in approximate dynamic
programming, focused on the types of problems that arise in transportation
and logistics.
Section 9.1 focuses on a problem where there is a single resource which
might be characterized by a vector of attributes. This treatment lays the foun-
dation for solving problems with many resources. Section 9.2 describes strate-
gies for estimating continuous value function approximations, which are a
central strategy for avoiding “curse of dimensionality” problems in dynamic
programs.

9.1 Discrete value function approximations for a single resource

Fundamental to the challenge of allocating multiple resources is the prob-

lem of allocating a single resource. This helps us illustrate some basic strate-
gies, but also highlights some issues that arise in more complex problems.
Our discussion is helped by a simple example. Assume we have the case of
a nomadic trucker who moves around the country from state to state (within
the United States). His attribute vector (for the moment) is the “state” that
he is in (such as New York). From a state, he has to choose from a number
of loads (requests to move freight) that he learns about after his arrival. The
load he chooses should reflect the profit he expects from the load, plus the
value of being in the state that the load terminates in (this will maximize our
trucker’s long term profits). If we wanted to make our problem more realistic,
we would include attributes such as the driver’s home domicile, how many days
he has been away from his home, and the maintenance status of his truck (was
there an equipment failure during his last trip that would require a repair at
the destination?).
We let at−1 ∈ A be the attribute of our driver at time t − 1 and Vt−1 (at−1 )
be the expected future value of a driver that has attribute a at time t − 1.
 Ch. 5. Dynamic Models for Freight Transportation 331

We can formulate this problem in two ways: as we have above using Rt as the
state variable and xt as our decision, or using at as the state variable and d
as the decision where the set of potential decisions Dt depends on the infor-
mation available at time t. Recalling that aM (t at−1  d) is the attribute vector
produced by applying the decision d to a resource with attribute at−1 , we can
write Bellman’s equations as
    
Vt−1 (at−1 ) = E max ctat−1 d + Vt aM (t at−1  d) at−1  (62)
d∈Dt

Equation (62) is the more natural way to formulate this specific problem, but it
disguises the relationship to problems with multiple resources.  A natural way
to bring the two formulations  together is as follows. Since a∈A Rta = 1, we
can write Vt (Rt ) = Vt (at ) = a∈A Rta Vt (a), where Rtat = 1. Thus, for the
single resource problem, Vt (Rt ) is linear in Rt .
If the attribute space A is not too large, we can solve Equation (62) using
classical backward dynamic programming techniques. Not surprisingly, these
problems do not generally exist in practice (the problems with resources with
simple attribute vectors, such as containers or freight cars, are exactly the prob-
lems with many resources). The more interesting problems involve managing
people (which tend to exhibit complex attribute vectors) or in multistop vehi-
cle routing problems, where the attribute vector has to capture the stops that a
vehicle has already made. For these problems, we resort to our approximation
strategy, where it is fairly obvious that a linear approximation is the appropri-
ate form. This gives us
  
π
 x n
 n x
Xt Rt−1  ω = max n ct xt + v̄ta Rta (xt )  (63)
xt ∈Xt (ω )
a ∈A
 
where Rxta (xt ) = a∈A d∈Da xtad δa (t a d) is our post-decision resource
vector. Our decision function Xtπ is written as depending on Rxt−1 and ω, but
it actually depends on Rt = Rxt−1 + R t (ω).
n of the value of a resource
The result of solving (63) is a random estimate v̂ta
with attribute at−1 with the information available at time t. If we were using
n−1
the simple one-pass algorithm, we could now update v̄t−1 using
n
  n−1
v̄t−1a = 1 − αn v̄t−1a + αn v̂ta
n
 (64)
Again we note that v̂tan is indexed by t because it contains information from
n
time t, while v̄t−1a represents an approximation of an expectation over all
possible outcomes of Wt , and therefore implicitly contains only information up
to time t − 1.
If we were to use the two-pass version of the algorithm, we would simulate
the trajectory of our nomadic trucker using the policy (Xtπ )t  t . We would
n as the cost of the trajectory over the interval (t T ), and then
then compute v̂ta
update v̄ as we did in Equation (64).
332 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

These procedures are known in the approximate dynamic programming lit-

erature as temporal-differencing methods (Sutton, 1988; Sutton and Barto,
1998; Bertsekas and Tsitsiklis, 1996) where they are normally described in the
context of steady state problems. They can also be viewed as examples of sto-
chastic gradient algorithms. To see this, we can write the problem of estimating
the exact function Vt (a) as one of solving

1  2
min f (v) = E (vta − Vta )  (65)
vta 2
where Vta is a random variable representing a measurement of the function
with noise. We note that Vta is an unbiased estimate only when we use the
two-pass version of the algorithm. Equation (65) can be solved by dropping
the expectation, taking a Monte Carlo sample and then applying a standard
gradient algorithm
n n−1  n−1 n 
vta = vta − αn ∇v f vta ω
 
= vta − α vta − Vtan (ω)
n−1 n n−1
  n−1
= 1 − αn vta + αn Vtan (ω) (66)
The procedure (66) is known as a stochastic approximation algorithm (Robbins
and Monro, 1951; see Kushner and Yin, 1997, for a modern treatment of the
field). In addition to some technical requirements on the properties of the
function being estimated, the proof requires that Vtan (ω) be an unbiased es-
timate of Vta in the limit (biases are allowed initially as long
as theyn go to zero
reasonably quickly), and two conditions on the stepsizes: ∞ n=1 α = ∞ and
∞
n=1 (α ) < ∞. These requirements basically require that the stepsize de-
n 2
crease to zero arithmetically, as in αn = 1/n.
This simple problem highlights an issue that plagues all approximate dy-
namic programming methods. Assume that we start with an initial approxima-
tion v̄ta = 0 for all t and a. In the case of our nomadic trucker, this means that
initially, he will prefer long loads with high initial rewards over shorter loads
with low initial reward. In fact, a shorter load might be more profitable in the
long run if it takes him to a region with high profit loads. If we only update the
values of states that he visits, then he will tend to only visit states that he has
already visited before.
One way to avoid this behavior is to begin by initializing v̄t with an upper
bound. This way, if our trucker turns down a load to a state that he has never
visited, then we can be assured that we are not turning away an option that
might prove to be quite good. In practice, this strategy works poorly, because it
is more likely that he will always choose the load that goes to a state that he has
never visited (for our problem, it is not just a state, but also a state at a point
in time). If the attribute vector includes other dimensions, then our attribute
space A can grow very quickly (a nice example of the curse of dimensionality).
The problem of visiting states in order to estimate the value of being in the
state is called the exploration problem in dynamic programming. There is no
 Ch. 5. Dynamic Models for Freight Transportation 333

cure for this problem, but there are strategies that help overcome it. One is
to use “off-policy iterations” which means to choose actions that are not opti-
mal given the current approximations, but forces the system to visit new states.
In real problems, the number of states and actions tends to make such strate-
gies impractical (at a minimum, they have very slow convergence). A more
effective strategy is to exploit problem structure. For example, assume that one
attribute of our truck driver is the number of days away he has been from home,
and suppose that we are able to establish that the value of the driver declines
monotonically with this quantity. We can then use this structural property to
help estimate the values of states we might never have visited.
The problem of estimating high-dimensional functions has open theoreti-
cal and practical questions. From a theoretical perspective, existing proofs of
convergence for discrete value functions requires visiting all states infinitely of-
ten. This generally requires an algorithm that combines optimizing iterations
(where we choose an action based on our approximate policy) with learning
iterations (where we might choose actions at random). From a practical per-
spective, learning steps are only of value with low-dimensional action spaces.
Practical considerations tend to focus on the rate of convergence, which tend
to be highly problem dependent.

9.2 Continuous value function approximations

A popular strategy for dealing with the curse of dimensionality has been
to replace the value function with a continuous approximation (Bellman and
Dreyfus, 1959). Particularly appealing are the use of low-order polynomials. In
this section, we describe several strategies for estimating continuous approxi-
mations.

9.2.1 Basis functions

A strategy for estimating a continuous approximation of a value function
is to use the concept of basis functions. Assume, for example, that we are
managing a fleet of containers of different types, and we believe that the
number of containers of each type at each location is an important determi-
nant of the√value of a particular allocation. We could devise a function, call it
φa (R) = Rta . φa (R) takes the square root of this quantity because we be-
lieve that there are diminishing returns from additional capacity. In this case,
we would call the square root of the number of containers with attribute a
“a feature” of our problem. Another feature might be the total number of con-
tainers ofall types at a location i. We might then create a new class of functions

φi (R) = a∈Ai Rta , where Ai is the set of attribute vectors that correspond
to resources at a particular location i. Another feature could be the number of
containers minus the forecasted number of orders. It is up to our imagination
and insight to identify what appear to be important features. Let F be the set
of all these functions, known as basis functions or features. We can formulate
334 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

an approximate value function as a linear combination of our basis functions


V(s|θ) = θf φf (s) (67)
f ∈F

Now the challenge is to determine the appropriate weights (θf )f ∈F . We il-

lustrate two strategies for doing this. The first is to simply apply a stochastic
gradient algorithm. We generalize slightly the minimization problem in (65) as
follows

1  n−1 
2
min E V (s|θ) − V (s) 
θ 2
Solving this using a stochastic gradient algorithm produces updates of the form
 
θn = θn−1 − αn Vn−1 (s|θ) − V(s ω) ∇θ V(s ω|θ) (68)
The challenge that this class of update procedures faces is the problem of scal-
ing the stepsize, since the units of θ and the units of the updating term will be
completely different. An alternative strategy that avoids this problem takes a
sample Ω  and then finds θ that minimizes the deviation between the approxi-
mation and the sample
 
1  1  n−1 2
θn = arg min V (s ω|θ) − V(s ω)  (69)
θ |Ω| 2
ω∈Ω 

We can then use the value of θ that solves (69), or smooth this estimate with
the previous estimate.
An introduction to the use of basis functions can be found in Bertsekas and
Tsitsiklis (1996), with an in-depth treatment of the theoretical foundation in
Tsitsiklis and Van Roy (1997). However, there is virtually no experimental work
in the area of transportation and logistics. There are two practical challenges:
identifying effective basis functions that capture the important properties of
the problem, and the second is ensuring that the rate of convergence is suffi-
ciently fast.

9.2.2 Auxiliary functions

Consider the case of a continuous resource allocation problem (where Rt is
continuous). This means that we can find a stochastic gradient of Vtn (Rt (ωn ))
in Equation (49). Let
  
v̂tn = ∇Rt Vtn Rt ωn
be a stochastic gradient of the function. Now let Vt0 (Rt ) be a conveniently
chosen continuous approximation (perhaps a low-order polynomial). We can
improve our approximation using the updating equation
 
Vtn (Rt ) = Vtn−1 (Rt ) + αn v̂tn − ∇Rt Vtn−1 (Rnt ) Rt  (70)
 Ch. 5. Dynamic Models for Freight Transportation 335

This is the primary step of the SHAPE algorithm (Cheung and Powell, 2000).
It starts with an arbitrary (but presumably carefully chosen) initial approxima-
tion, and then tilts it using stochastic gradient information. The second term of
Equation (70) is a linear updating term that adds the difference between the
stochastic gradient of the real function and the exact gradient of the approx-
imate function. Each updated function has the same basic form as the initial
approximation. This can be a nice simplification if it is necessary to precompute
derivatives.
The SHAPE algorithm is optimal for two-stage, continuously differentiable
problems. For nondifferentiable problems, and for multistage problems, it is an
approximation. In principle the initial approximation could be estimated from
a set of basis functions, allowing both methods to be used as a hybrid strategy.
Computational testing of these methods is limited. Their success appears
to be highly dependent on a good choice of initial approximation. Simply
choosing low-order polynomials because they are convenient is not likely to
be successful.

9.2.3 Linear functional approximations

We have already seen in the case of managing a single resource that the
value function can be written as a linear function of the resource vector. Linear
approximations can also be effective when we are managing very large numbers
of resources. For example, shipping companies might manage fleets of hun-
dreds of thousands of containers. Our ability to effectively manage resources
depends primarily on our ability to estimate the slope of the function, rather
than the function itself. When there are large numbers involved, the slopes
tend to be fairly stable.
Throughout this section, we assume that v̂ta is a stochastic gradient of either
Vt (Rt ) with respect to Rta (Equation (49)), or of Vtπ (Rt  ω) (Equation (52),
computed using a backward pass). The choice of which version to use is pri-
marily one of trading off speed of convergence with algorithmic complexity
(backward passes can be hard to implement but can dramatically improve re-
sults). Given an estimate of the slope, we would then smooth on the slope (as
in Equation (64)) to obtain an estimate of the slope, producing the following
approximation
  
 n  n 
 n
Vt Rt ω = max n ct xt + n−1 x
v̄ta Rta (xt )  (71)
xt ∈Xt (ω )
a∈A
Using our δ(·) function, we can write Equation (71) as
  
Vtn Rnt ωn
 
= max n ctad xtad
xt ∈Xt (ω )
a∈A d∈Da
   
n−1
+ v̄ta δta (t a d)xtad
a ∈A a∈A d∈Da
336 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão
   
n−1
= max n ctad xtad + v̄ta M (tad) xtad (72)
xt ∈Xt (ω )
a∈A d∈Da a∈A d∈Da
   
n−1
= max n ctad xtad + v̄ta M (tad) xtad (73)
xt ∈Xt (ω )
a∈A d∈Da a∈A d∈Da
  
n−1
= max n ctad + v̄ta M (tad) xtad  (74)
xt ∈Xt (ω )
a∈A d∈Da

We obtain Equation (72) by using δta (t a d) = 1 for a = aM (t a d) and 0

otherwise, producing the more compact expression in (73). Equation (74) is
obtained by simply rearranging terms, which shows that we are now solving a
problem with the same structural form as the original myopic problem, with a
modified set of costs.
This structure is especially important in transportation problems since the
original myopic problem might be fairly difficult. If we are lucky, it is an as-
signment problem, transportation problem or some linear program. But in
some settings it is a set partitioning problem, vehicle routing problem, or inte-
ger multicommodity flow problem. For these problems, it is especially helpful
when the value function approximation does not complicate the problem fur-
ther.
Not surprisingly, linear value functions do not always work well. It is very
hard to tune a linear function to attract just the right number of resources.
Nonlinear functions are simply more stable, but as we show below, they intro-
duce considerable complexity.

9.2.4 Separable piecewise-linear functional approximations

When we are managing discrete resources (freight cars, trailers, people, air-
craft, locomotives) it is especially important to obtain integer solutions. This
is simplified considerably when we use separable, piecewise-linear approxima-
tions which are linear for noninteger values of the number of resources. For
this strategy, we write our approximation as

Vt (Rt ) = Vta (Rta )
a∈A
where
R
ta −1
  
Vta (Rta ) = v̄ta (r) + v̄ta Rta  Rta − Rta   (75)
r=0
When we introduce this approximation in Equation (49) we obtain
    
Vtn Rnt ωn = max n ctad xtad
xt ∈Xt (ω )
a∈A d∈Da
   
+ Vta (Rta ) δta (t a d)xtad  (76)
a ∈A a∈A d∈Da
 Ch. 5. Dynamic Models for Freight Transportation 337

Assume that the original myopic problem is a pure network. We are interested
in the conditions under which the introduction of value functions destroys this
network structure. The constraint set Xt (ωn ) typically contains upper bounds
representing vehicle capacity constraints, network constraints, or simply con-
straints on the number of tasks to be served. Let utd be the upper bound on
the total number of resources acted on by a particular decision type

xtad  utd  (77)
a∈A

Equation (77) has the effect of creating a bundle constraint. This does not
pose a problem when we use linear value functions, as we demonstrated that
any linear value function is equivalent to solving the original problem with a
modified set of costs. But when we use nonlinear value functions, the network
structure is destroyed.
Of particular interest are problems that satisfy the Markov property:

Definition 9.1. A resource allocation problem satisfies the Markov property if

aM (t a d) = aM (t a  d) ∀a ∈ A

The Markov property means that the attributes of a resource after being
acted on are independent of the attributes of the resource. A special class of
problems that satisfy the Markov property are known as single commodity flow
problems. In this problem class, the attribute vector consists purely of the lo-
cation of a resource. A task consists of an origin and a destination, so after
completing the task, the attribute of the resource equals the destination of the
task (and is therefore independent of the origin). Another example arises when
we have to maintain an attribute of whether a container is clean or dirty. As-
sume there is freight that is classified as clean or dirty. Only a clean container
can move clean freight, but any container can move dirty freight. However, if
a clean car moves dirty freight, then it becomes dirty. Such a problem would
also satisfy the Markov property.
When a problem possesses the Markov property, the attributes of the re-
source after being acted on are completely determined by the decision rather
than the attributes of the resource. As a result, we can create a node for re-
sources that have been acted on by a particular type of decision, and we do not
have to track the identity of the origin of the resource after it completes a task
(or any other action). As a result, it is possible to formulate the problem as a
pure network.
When problems do not possess the Markov property, the resulting problem
will not, in general, be a pure network. But the use of piecewise linear, separa-
ble value function approximations appears to produce integer multicommodity
flow problems with very tight LP relaxations. Experimental evidence is that
when these problems are solved as continuous linear programs, the solution
returned is integer the vast majority of the time (99.9 percent).
338 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

9.3 Bibliographic notes

The field of approximate dynamic programming is relatively young. The first

significant reference is Bertsekas and Tsitsiklis (1996) which gives a general in-
troduction to a variety of methods, but without presenting specific algorithms.
Sutton and Barto (1998) approach the topic from the perspective of reinforce-
ment learning. Temporal-difference learning, a form of stochastic approxima-
tion procedure, was first introduced by Sutton (1988). The vast majority of
the literature on dynamic programming appears to focus on discrete repre-
sentations of the value function, but there has been a steady literature on the
use of continuous approximations, beginning as early as Bellman and Dreyfus
(1959). Tsitsiklis and Van Roy (1997) investigate in depth the convergence
properties of temporal-difference methods using continuous approximations.
The SHAPE algorithm was introduced by Cheung and Powell (2000). The the-
ory behind the estimation of linear value function approximations is based on
the seminal work on stochastic approximation methods by Robbins and Monro
(1951) and Blum (1954) (see Kushner and Yin, 1997, for a modern review of
these techniques). The use of separable, piecewise linear approximations was
first introduced by Cheung and Powell (1996) who used a static approxima-
tion. Godfrey and Powell (2001) proposes an adaptive learning algorithm for
piecewise linear, separable approximations in the context of two-stage resource
allocation problems, which is extended in Godfrey and Powell (2002) for mul-
tistage, single commodity problems. Topaloglu and Powell (2006) shows that
these techniques also work quite well for multicommodity flow problems with
a range of forms of substitution.

10 The organization of information and decisions

Another important issue in the modeling of decisions and information is

how they are organized. Large transportation operations are typically charac-
terized by different decision makers controlling different parts of the problem
with different information. In this section, we provide some basic notation to
model how organizations are controlled.
We adopt the convention that a decision-maker is referred to as an “agent”
which is assumed to have sole control over a set of decisions. We let
Q = set of decision-making agents
Dq = the set of decisions controlled by agent q
Our first challenge is to formalize the decision sets Dq . We assume that the sets
(Dq )q∈Q are mutually exclusive and collectively exhaustive.
There are three dimensions to the set Dq : the types of decisions an agent
can make, the types of resources an agent can act on, and the time periods at
which the decisions might be implemented. For example, a locomotive plan-
ner for a railroad might have the responsibility for planning the movements
 Ch. 5. Dynamic Models for Freight Transportation 339

of locomotives on Monday, Tuesday, and Wednesday. He can only control lo-

comotives that are in service; a local maintenance supervisor has control over
locomotives that are classified as being “in maintenance” (it is common for
a locomotive planner to call the local maintenance supervisor to determine
when a locomotive that is in the shop will become available). Furthermore, he
can make decisions about coupling and uncoupling locomotives to and from
trains, but does not make decisions about buying, selling, leasing, and storing
locomotives. To capture this process, we let

Aq = subset of the attribute space for subproblem q where

Aq = A and Aq1 ∩ Aq2 = φ when q1 = q2 
q∈Q

Implicit in the definition of the attribute space is:

CqD = set of control classes associated with subproblem q As a rule,

a subproblem is formulated for a specific type of decision,
so the set CqD is implicit in the formulation of the problem
Dac = set of decisions in control class c that can be applied to
a resource with attribute vector a
Tqih = the implementation horizon for subproblem q This is the set
of time periods during which subproblem q controls
the decisions. Since time is a dimension of the attribute vector,
we may state that a ∈ Aq ⇒ aactionable ∈ Tqih 
It is important to emphasize that the subsets Aq do not necessarily (or even typ-
ically) imply a geographical decomposition, although this certainly can be the
case. At a railroad, locomotives are managed regionally (with some network-
wide coordination); freight cars are managed network-wide, but are decom-
posed by freight car type.
Our decision set for agent q is now given by

Dq = subset of decisions in subproblem q

= d ∈ Dac  c ∈ CqD  a ∈ Aq  aactionable ∈ Tqih 
The definition of Dq includes the time periods for which the decisions may be
implemented, which produces a more compact notation. In some cases, the
indexing of when a decision is to be implemented needs to be made explicit, so
we can use (t q) to capture the combined information.
Before, the information available to make a decision was represented im-
plicitly either by indexing time t, or explicitly through a state variable such as
St or Rt . When we model multiagent systems, we have to be more explicit. For
340 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

this reason, we define:

Itq = the information content of subproblem q at time t

It is best to think of Itq as a set of functions that return data that is needed to
make a decision: the amount of equipment, people, and freight, available now
and in the future (recall that the index t here refers to information content, not
actionability).
Another dimension of multiagent systems is capturing the property that de-
cisions by agent q can have an impact on another agent q (for example, sending
a piece of equipment from one region to the next, or when the person who buys
and sells equipment changes the amount of equipment that a planner can work
with). We capture these interactions by defining the following:
−
→
Definition 10.1. The forward reachable set I q of subproblem q is the set of
subproblems q with resource states a ∈ Aq that can be reached by imple-
menting a single decision d ∈ D on at least one state a ∈ Aq . More precisely,
the forward reachable set of subproblem q is
−
→
Iq = q ∈ Q \ q | ∃a ∈ Aq  d ∈ Dq 
where M(a d ·) → (a  · ·) a ∈ Aq  (78)

For completeness, we also define:

←
−
Definition 10.2. The backward reachable set I q of subproblem q is the set of
all subproblems for which subproblem q is forward reachable. More precisely,
the backward reachable set of subproblem q is
←
− −
→
Iq = {q ∈ Q \ q | q ∈ I q } (79)

This allows us to define decision vectors for each agent:


xtaa = xtad δa (t a d)
d∈Dq
xtqa = {xtaa  a ∈ Aq }
xtqq = {xtqa  a ∈ Aq }
−
→
xtq = {xtqq  q ∈ I q }
We adapt our notation for making decisions by defining
π
Xtq (Itq ) = a function that determines xtad for a ∈ Aq  d ∈ Dq
using the information available in Itq 
where Itq is the information known by agent q at time t. As we did before, we
can create a general purpose decision function that uses all four information
 Ch. 5. Dynamic Models for Freight Transportation 341

classes
π
Xtq (Itq ) = arg max ctq xtq
xtq
    
+ Vtqq Rtq (xtqq ) − H ρ(xtq ) ρq  (80)
−
→
q ∈ I q

Shapiro and Powell (2006) describe this strategy in depth and give a method
for estimating the value functions Vtqq (Rtq (xtqq )).

11 Illustrative models

We illustrate these ideas in the context of three problem areas: rail car dis-
tribution, load matching for truckload motor carriers, and batch processes that
arise in less-than-truckload motor carriers. The car distribution problem is
characterized by a low-dimensional attribute space (it is roughly a multicom-
modity flow problem) with an interesting set of dynamic information processes.
The load matching problem introduces the challenge of modeling people
(characterized by high-dimensional attribute spaces). Also, we demonstrate
the modeling of a two-layer resource allocation problem (drivers and loads).
Despite their differences, both of these problems have the flavor of time-staged
linear programs. The batch processes that arise in less-than-truckload motor
carrier illustrates how we can handle the types of fixed-charge problems that
arise in network design.

11.1 Dynamic flow problems for rail car distribution

Our car distribution problem provides an opportunity to illustrate the no-

tation that we have introduced. Section 11.1.1 provides a model of an actual
car distribution system developed for a major railroad. Then, Section 11.1.2
illustrates several optimization models that are used in practice. Section 11.1.3
summarizes the results of some experiments run with an actual dataset from a
major railroad. Section 11.1.4 closes with some observations about this appli-
cation.

11.1.1 A dynamic model

The car distribution problem basically involves assigning cars to orders. Cars
can be moved empty or assigned to orders and then moved loaded. Cars or
orders that are not acted on are held. Orders that are not satisfied within a
reasonable time (typically within the same week) are considered lost, but in
practice some orders might simply be held to the next week. When a car be-
comes empty at a shipper, the railroad has roughly three options which are
illustrated in Figure 7: assign the car to a pending order, move the car to a “re-
gional depot” from which it can be used to satisfy orders from other shippers
342 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

Fig. 7. Car distribution through classification yards.

in the region, or send the car to a “classification yard”. A classification yard

is a major sorting facility, from which cars might be sent to anywhere in the
network.
Our problem, then, involves two “resource classes”, namely cars and orders.
We model cars using

a = the vector of attributes characterizing a car

A = the set of attributes of the cars
Rctt  a = the number of cars with attribute a that we know about
at time t that will be available at time t   The attribute vector
includes the location of the car (at time t  ) as well as
its characteristics
 
Rctt  = Rctt  a a∈A 
 
Rct = Rctt  t  ∈T 

For our problem, we model the information process (the arrival of information
and the making of decisions) in 24-hour increments. By contrast, we model the
physical movement of resources in continuous time. As a result, our attribute
 Ch. 5. Dynamic Models for Freight Transportation 343

vector consists of
 
alocation
a = acar_type 
aactionable
The element of aactionable gives the time, down to the minute, of when a car is
able to move. Note that there are some important modeling decisions in how
the attribute is handled. Assume, for example, that a car at origin o has to
pass through classification yards (where cars are sorted and can be rerouted)
at i and j on its path to destination d. If the car is sent to d with expected
arrival time t  , we can update its attribute vector to include alocation = d and
aactionable = t  . This means that we will not consider rerouting the car at i or j.
On the other hand, we might wish to consider rerouting, but we do not want
to ignore that it is already on a path to d. The railroad generates a piece of
paper called a work order that routes the car to d; if we reroute the car to a
different destination, this paperwork has to be changed. If we wish to allow
rerouting, but without forgetting the original destination of a car, we would
simply add a new dimension to a which we might call adestination which would
be the ultimate destination of the car, and we would interpret alocation as the
next location where a car is able to be rerouted (so, as the car departs from
o heading to i, we would set alocation = i and adestination = d).
Orders are modeled in a similar way using
b = the vector of attributes characterizing an order
B = the set of attributes of an order, including the number of days
into the future on which the order should be served (in our
vocabulary, its actionable time)
Rtt  b = the vector of car orders with attribute b ∈ B that we know
o

about at time t which are needed at time t   Ro0bt  is the set

of orders that we know about now
 
Rott  = Rott  b b∈B 
 
Rot = Rott  t  ∈T 
The attributes of orders consist of
⎡ ⎤
borigin
⎢b ⎥
b = ⎣ destination ⎦ 
bcar_type
bactionable
Here bcar_type represents the preferred car type, but substitution opportu-
nities exist. bactionable is the exact time that the order is available to be picked
up. We note that when an order becomes known, it is not necessarily the case
that all the attributes of an order become known at the same time. So we may
344 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

adopt a convention that, say, bdestination = ‘-’ means that the destination is not
yet known.
Our resource, vector, can now be written
 
Rt = Rct  Rot 
One of the challenges of car distribution is dealing with a high level of un-
certainty. Figure 8 illustrates the actual and predicted customer demand, with
10th and 90th percentiles. In addition, there are other forms of uncertainty: the
process of loaded cars becoming empty, the transit times, and the acceptability
of a car to a shipper (is it clean enough?). There are five types of information
processes that railroads have to deal with:
(1) Orders – Customers call in orders for cars, typically the week before
when they are needed. The order does not include the destination of
the order.
(2) Order destination – The destination of the order is not revealed until
after the car is loaded.
(3) Empty cars – Empty cars become available from four potential sources:
cars being emptied by shippers, empty cars coming on-line from other
railroads, cars that have just been cleaned or repaired, and new cars that
have just been purchased or leased.
(4) Transit times – As a car progresses through the network, we learn the
time required for specific steps (after they are completed).

Fig. 8. Actual vs. predicted forecasts of future demands, showing the 10th and 90th percentiles.
 Ch. 5. Dynamic Models for Freight Transportation 345

(5) Updates to the status of a car – Cars break down (“bad order” in the
language of railroads) or are judged (typically by the customer) to be
not clean enough.
Railroads manage this uncertainty using several strategies. First, there is the
fact that customers make their orders partially known in advance. For example,
customers might book their orders the week before, although it is common for
them to do so toward the end of the week. However, there can be discrepancies
between what they book in the computer vs. what they really need (which might
be communicated by phone). For example, if a railroad is running short on
cars, customers have a tendency to overbook. Second, railroads depend on
various forms of substitution when they cannot fill an order. Three forms of
substitution come into play:
(1) Geographic substitution – The railroad may look at different sources of
cars and choose a car that is farther away.
(2) Temporal substitution – The railroad may provide a car that arrives on
a different day.
(3) Car type substitution – The railroad may try to satisfy the order using a
slightly different car type.
It is common to model the decisions of moving cars from one location to an-
other, but in practice car distribution is characterized by a number of decisions,
including:
(1) Move car to a location – An empty car may be moved to a regional depot
or an intermediate classification yard.
(2) Assign to a customer order – Here we make the explicit assignment of
a car to a specific customer order.
(3) Clean or repair a car – This produces a change in the status of the car.
(4) Switch pools – Many cars belong to shipper pools which might be ad-
justed from time to time.
(5) Buy/sell/lease decisions – These decisions affect the overall size of the
fleet.
Our presentation will consider only the first two classes, although classes
3 and 4 represent trivial extensions. To represent these decisions, we define

Dc = the decision class to send cars to specific customers, where

Dc consists of the set of customers (each element of Dc
corresponds to a customer location)
D = the decision to assign a car to a type of order. For d ∈ Do 
o

we let bd ∈ B be the attributes of the order type associated

with decision d
D rd
= the decision to send a car to a regional depot
346 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

Dcl = the decision to send a car to a classification yard (each

element of D cl is a classification yard)
d φ = the decision to hold the car (“do nothing”)
The next step is to model the exogenous information processes. We start by
modeling the arrivals of new resources, which includes both cars and orders.
For example, a railroad might send a car “off line” to another railroad. When
it returns to the railroad, it is as if a new car is arriving to the system. In some
cases, a car sent to a shipper can be modeled as a car that is effectively dis-
appearing from the network; when the car is released and returned from the
shipper, it is a new arrival. New orders, of course, arrive in the form of phone
calls or entries into a computer database (for example, through a web site).
We take advantage of our ability to separate the arrival process of infor-
mation from the physical arrival of cars and orders. For example, we would
represent new orders using:
o  = the vector of changes to the order vector (new customer
Rtt
orders, changes in orders) that arrive during time interval t
that become actionable at time t   t
 
o = R
R o   
t tt t t

We would define similar vectors Rc  and Rct for cars. Let Wt be our general
tt
variable representing the information arriving at time t, where
 o c
Wt = R  R 
t t

11.1.2 Optimization formulation

The most basic optimization model for car problems is the one that matches
known cars to known orders over a predetermined horizon (even the most
basic model captures the fact that some cars and orders are actionable in the
future). This model can be stated as

max c0ad x0ad (81)
x
a∈A d∈D
subject to

x0ad = Rc0a  a ∈ A (82)
d∈Dc

x0ad  Ro0bd  d ∈ Do  (83)
a∈A
x0ad ∈ Z+  a ∈ A d ∈ D c  (84)
Equation (82) captures flow conservation constraints on cars only, while Equa-
tion (83) ensures that we do not assign more cars to orders than we have
orders (effectively enforcing flow conservation on orders). This model follows
 Ch. 5. Dynamic Models for Freight Transportation 347

the normal convention of modeling cars as an active resource (resources that

we actively modify) while orders are a passive resource (they only change their
status if they are assigned to a car). Recall that in Equation (83), for every
element of d ∈ D o there is an order type with attribute bd .
The next step up in sophistication is a model that incorporates point fore-
casts of cars and orders that become known in the future (this model is in use
by at least one railroad). Let Rct and R
ot be point forecasts of cars and orders
of information that will become available during time period t. This model can
be written as

max c0ad x0ad (85)
x
a∈A d∈D
subject to
 
x0ad = Rc0a + c 
R d ∈ D c  a ∈ A (86)
ta
d∈D t∈T ph \0
 
x0ad  Roobd + o 
R d ∈ Do  (87)
tbd
a∈A t∈T ph \0

x0ad ∈ Z+  (88)
Note that this model aggregates forecasts of all future information about cars
and orders. We retain the information about when the order will be actionable
in the vector bd , so that decisions to assign, say, a car available now to an order
that is forecasted to arrive at time t  which will then be actionable at time t  ,
can properly account for the difference between the available time of the car
and the available time of the order.
Another strategy is to use our dynamic programming approximations as we
outlined above. If we use a (possibly nonlinear), separable value function, we
would find ourselves solving
  
n−1  x 
Vtt
πn
Xt (Rt ) = arg max ct xt + a R (x
tt  a t ) (89)
xt
t  t a∈A
subject to

xtad = Rctta  a ∈ A (90)
d∈D

xtad  Rotbd  d ∈ Do  (91)
a∈A
xtad ∈ Z+  (92)
Equation (90) limits us to acting on cars that are actionable now, while Equa-
tion (91) says that we cannot move more cars loaded than the orders we know
about now. This problem is generally quite easy to solve, but does require the
iterative learning logic described in Section 9.
348 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

11.1.3 Some numerical experiments

The approximate dynamic programming algorithm was run on a dataset
from a major railroad using nonlinear, separable value functions. The results
are presented in Figure 9 which shows total profits, along with total revenues,
empty movement costs, and service penalty costs. The figure shows a steady
improvement in total profits as the objective function improves. The total rev-
enue drops in the early iterations but rises back to approximately the same
level as the first iteration, indicating that our improvement is through better
repositioning rather than higher coverage. These experiments indicate that the
adaptive learning logic works, but raises the question: exactly what is it doing
that is smarter than a myopic model?
A common misconception is that we do not need to look into the future be-
cause customer orders are known in advance. For the rail dataset, customers
orders are generally known the week before, with the bulk of customers calling
in their orders on Wednesday and Thursday. Empty movement times are typ-
ically less than five days, which suggests that the prebook time allows the rail-
road to wait until orders are known before moving a car. Figure 10 shows the
percent of total miles that are empty in history (approximately 54 percent for
this dataset), from a myopic optimization model (which uses information about
cars and orders that become available in the future), and from an approximate
dynamic programming model. The myopic model reduces empty miles to ap-
proximately 48 percent, which is a significant reduction. Other railroads have
used estimated savings such as these to justify projects in the millions of dol-

Fig. 9. The contribution of adaptive learning to revenue, costs and overall profits.
 Ch. 5. Dynamic Models for Freight Transportation 349

Fig. 10. Empty miles as a percent of total (a) in history, (b) optimized using a myopic model, and
(c) optimized using an approximate dynamic programming procedure.

lars. The adaptive learning model reduces empties to about 35 percent after
25 iterations, with most reductions after this.
Where are these benefits from approximate dynamic programming coming
from? There appear to be two sources. First, when a car becomes empty and
available at a shipper, it must be moved; it cannot be held at the shipper. Cars
that become empty early in the week often have to be moved before customer
orders have been booked. A common strategy used by railroads is to move
cars to classification yards and regional depots where they can be reassigned to
customers after they arrive (by this time, customers orders will often have ar-
rived). A myopic model will never send a car to these locations because there
are no customer demands at these locations. Using approximate value func-
tions, the model will send cars to locations where they have the highest value,
which could easily be a regional depot or classification yard.
The second, and perhaps more significant reason that approximate dynamic
programming works is that it captures the uncertain presence of cars in the
future. In a snapshot of cars and orders, there is often far more information
about orders in the future than cars. A customer might book his order a week
or more into the future, but we are generally unable to track cars into the
future. We cannot forecast the availability of cars after they fulfill orders that
have not even been called in yet. Furthermore, even orders that have been
called in are incomplete; shippers do not identify the destination for a car until
after the car has been loaded.
350 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

The behavior that we have observed is that a myopic model has a tendency
to assign a car that is available now to an order that is known now, but does
not have to be served until the future. The myopic model is not able to as-
sign a car after it has been used to fulfill an order that is either unknown
or has incomplete information (e.g., the destination). The approximate dy-
namic programming model takes this information implicitly into account when
it solves subproblems in the future. Thus, the adaptive learning technology is
not only learning about future orders that are currently unknown, but also fu-
ture cars.

11.1.4 Notes
The car distribution problem captures some of the richness of dynamic in-
formation processes that arise in real problems. At the same time, it offers
some convenient simplifications. First, we implicitly assume that the attribute
spaces A and B are fairly small and easy to enumerate. This is not too impor-
tant in the myopic and rolling horizon models, but is very important if we use
the dynamic programming approximation. The reason is that if we act on a car
to produce a car with attribute a , we need to have a value function Vtt n−1
 a to
measure the value of the resource. Furthermore, we have to update these value
functions even if Rta is zero, since an early estimate of the value of a car in the
future might have been much lower than what it should be. If Vtt n−1
 a underes-

timates the value of a car of type a , then we might not implement a decision
that produces this car type in the future again. As a result, we have to keep
updating the value function to get a good estimate. This has proven to be very
important in practice.
In addition, although this is an instance of a two-class resource allocation
problem, it is modeled as a one-layer problem. We only model decisions acting
on cars, and we only estimate the value of cars in the future (since an order,
once it has been moved, is assumed to disappear).

11.2 The dynamic assignment problem for truckload trucking

We next consider the problem of assigning drivers to loads that arises in

truckload trucking. We first present the model of the physical process, and
then discuss methods for making decisions.

11.2.1 A model of the dynamic load assignment problem

Unlike our car distribution problem, we are going to model this as a two
layer problem from the perspective of approximating the value of resources
in the future. We also use this problem to illustrate some other modeling de-
vices. For example, in the car distribution problem we represented car types
and demand types. This is more natural since there are often a large number
of cars (and even orders) of exactly the same type. In the dynamic assign-
ment problem, we are modeling the assignment of “resources” to “tasks” which
 Ch. 5. Dynamic Models for Freight Transportation 351

are discrete in nature (and generally highly individualized). It is more natural,

then, to model each resource and task individually. We note that we refer to
these as two “resource layers” but use the more common term “resource” to
refer to the layer that represents the objects (drivers, trucks, locomotives) and
“task” to refer to the customer or requirement to be satisfied.
We begin by defining
C R = set of resource classes
 
= Drivers(D) Loads(L) 
Rc = set of all resource indices for class c ∈ C R that might possibly
enter the system
R = RD ∪ R L 
We use d to index elements of the set RD and l to index elements of the set RL ,
and we use r ∈ R as a generic resource which may be a driver or a load. This
convention allows us to avoid defining everything twice.
To describe the state of our system, we define
1 if resource r ∈ R is known at time t and available
Rtt  r = to be assigned in period t  
0 otherwise
Rt = (Rtt  r )r∈Rt  t 
We use RD L
t and Rt to refer specifically to the resource state vector for drivers
and loads.
Over time, new drivers and loads will arrive to the system. We assume that
there is a master list of identifiers for all drivers and loads (in the set R), but
that there is a specific time at which a driver or load becomes known. We model
this process using
1 if resource r ∈ R which is available to be acted on
tt  r =
R at time t   first becomes known in period t
0 otherwise
t = (R
R tt  r )r∈Rt  t 
In this simple model, the exogenous information arriving in time period t is
given by
 D L
Wt = R R
  (93)
t t
Our decision problem involves the assignment of drivers to loads, which are
represented using
1 if driver d is assigned to load l at time t
xtdl =
0 otherwise
xt = (xtdl )d∈RD l∈RL 
352 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

xLtl = xtdl
d∈RD
1 if any driver is assigned to load l at time t
=
0 otherwise

xD
td = xtdl
l∈RL
1 if driver d is assigned to any load at time t
=
0 otherwise
We use these variables to write our resource dynamics as
RD D D D
t+1 = Rt − xt + Rt+1  (94)
RL = RL − xL + RL  (95)
t+1 t t t+1

11.2.2 Optimization formulation

Our challenge, of course, is one of designing a decision function Xtπ (Rt ).
Our goal is to maximize the total contribution from assigning drivers to loads
over a planning horizon T ph . Let
ctdl = the contribution from assigning driver d to load l at time t
Our one period contribution function is given by
 
Ct (xt ) = ctdl xtdl  (96)
r∈RD
t l∈Rt
L

We assume that we only receive a contribution when a decision is implemented,

which can only occur when both the resource and task are actionable. However,
we can plan decisions in the future, which is what we assume we are doing when
we assign a driver to a load when one or both is actionable in the future. As
discussed in Section 7 we make decisions by solving a problem of the form
Xtπ (Rt ) = arg max Ctπ (x) (97)
x

Keep in mind that the function Xtπ returns a vector xt = (xtt  )t  t , which
means it might include actions to be implemented now, xtt , and plans of what
to do in the future, (xtt  )t  ∈T ph \0 . We may limit our attention to resources that
are knowable now and actionable now, Rtt , or we may include resources that
are actionable in the future, (Rtt  )t  ∈T ph . Both represent forms of deterministic
models.
An interesting (and perhaps unexpected) behavior arises when we consider
resources that are known now but actionable in the future. It would seem
natural that such a model would outperform a model that focuses only on
resources that are actionable now, but looking further into the future can pro-
duce poor results if it is not done well. In most applications, the extent to which
 Ch. 5. Dynamic Models for Freight Transportation 353

we know about internal resources (drivers, trucks, locomotives, freight cars) in

the future can be quite different from how far we know about customer orders.
In rail and trucking, it is typical for the customers to call in orders farther into
the future than we would know about drivers and equipment. For example, it
can be hard to track rail equipment into the future because even when a cus-
tomer order is known, we might not know everything about the order (most
obviously, the destination of the order, but also the time required to complete
the order). This can make it difficult to track a car into the future for more
than a few days. Yet customers will call in orders a week or even two weeks
into the future.
Incorporating point forecasts of the future can be difficult for this problem
class, since all the activities are (0/1). First, while we have made the explicit
choice to model specific drivers and loads, as opposed to driver and load types
(as we did in our car distribution model), when we perform forecasting we
have to forecast by type (but we can treat a forecasted driver or load as if it
were real). Generally, we are not able to forecast as many attributes of a real
resource, which means that forecasted resources have to be somewhat simpli-
fied. Second, we encounter problems if we have to forecast a type of load that
occurs with probability 0.20.
An alternative strategy for handling uncertain, forecasted activities is to re-
sort to value function approximations. Here again, it is more useful to think of
estimating the value of a type of driver or load, rather than a specific driver or
load. To handle this transition, we let

atr = the attributes of resource (driver or load) r at time t

If we decide to hold a driver (not assign the driver to any load) then this
means we have this same driver in the future. This is a “do nothing” decision
(which we have represented using d φ ) which might produce a modified set of
attributes given by aM (t ar  d φ ). We note that
xD  
Rtd = 1 − xD
td
1 if driver d is held at time t
=
0 otherwise
xL
Similarly, let Rtl indicate whether load l is held until the next time period.
xD xL
Rxt = (Rt  Rt ) is our post-decision resource state variable. The decision
function is given by
   
Xtπ (Rt ) = arg max ctdl xtdl + Vt Rxt (98)
x∈Xt
d∈RD
t l∈Rt
L

subject to appropriate flow conservation constraints on drivers and loads. We

then have to determine an appropriate functional form for Vt (Rt ). The most
354 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

obvious to choose is a linear approximation

 
Vt (Rt ) = D D
v̄td Rtd + L L
v̄tl Rtl  (99)
d∈RD
t l∈RL
t

D is the value of driver d if it is held at time t, and v̄L is the value

Here v̄td tl
of load l. We note that this approximation is, for the first time, capturing the
value of holding both resource layers in the future. We can easily estimate
these values directly from the dual variables for the driver and load resource
constraints (94) and (95). One complication is that the dual variables, since
xD
they are simply subgradients, are typically not accurate enough. If Rtd = 1,
xD
then we want the impact of eliminating driver d from the future; if Rtd = 0,
then we want the value of adding driver d to the future. It is possible to find
these left and right gradients (respectively) by solving flow augmenting path
problems into and out of the supersink to all the nodes in the network.
If we introduce our linear approximation into (98), we obtain, after some
manipulation
  
Xtπ (Rt ) = arg max D
ctdl − v̄ta L
M (ta d φ ) − v̄taM (ta d φ ) xtdl 
x∈Xt d l
d∈RD
t l∈Rt
L
(100)
where aM (t a  dφ )
d is the attribute vector of a driver that has been held, and
aM (t al  d φ ) is the attributes of a load. We note that (100) is still an assignment
problem which is an important practical result.
There is still a problem. If xtdl = 1, then we are assigning driver d to load
l, and both are removed from the problem in the future. So, each decision at
time t impacts a driver and a load. If xtdl is increased from 0 to 1, then we
have the effect of decreasing RD L
td and Rtl each by one. This is approximated by
D + v̄L ). Needless to say, this will introduce errors. A better approxima-
−(v̄td tl
tion would be to capture the nonseparability of the value function. We could,
instead, let

v̄tdl = the marginal contribution of holding driver d and load

l at the same time
Of course, we use −v̄tdl to approximate the impact of assigning driver d to
load l and therefore eliminating both of them from the pool of resources in the
future. We now find ourselves solving subproblems of the form
 
Xtπ (Rt ) = arg max (ctdl − v̄tdl )xtdl  (101)
x∈Xt
d∈RD
t l∈Rt
L

As before, (101) is still also an assignment problem, which means that if we

can estimate v̄tdl , then our single-period problem in Equation (98) is no more
difficult than the original myopic problem.
 Ch. 5. Dynamic Models for Freight Transportation 355

Not surprisingly, v̄tdl (see Spivey and Powell, 2004, for details) is con-
siderably harder to compute than v̄td D and v̄L , which require only two flow-
tl
augmenting path calculations. v̄tdl , by contrast, requires a flow augmenting
from each load node l back to each driver node d. This can be accomplished
with a flow-augmenting path for each driver, which is substantially more dif-
ficult than the single flow augmenting path into and out of the supersink we
required for the separable approximation.

11.2.3 Some numerical experiments

The optimization model allows us to provide answers (experimentally) to
two types of questions. First, does an ADP policy (that is, a policy that uses
approximate dynamic programming) provide solutions that are better than
myopic policies, and how does this compare with the availability of advance
information? Second, what is the value of advance information, and how does
the value of advance information change with the sophistication of the policy
used? These two questions illustrate the two perspectives of dynamic models
that we addressed in the Introduction.
These questions were addressed in Spivey and Powell (2004). The issue of
solution quality was addressed by running a series of simulations. For each
simulation, we can compute the optimal solution after all the information is
known (this is simply an assignment problem using all the resources and tasks),
giving us a posterior bound. The first set of simulations used deterministic data,
which is to say that |Ω| = 1. For this case, we can hope to achieve a solution that
is close to the posterior bound. The second set of simulations used a stochastic
data set, with |Ω|  1, where there are a number of different outcomes of new
drivers and loads.
The results of the simulations are summarized in Figure 11, which repre-
sent an average over 10 different datasets. For the deterministic datasets, we
see a steady improvement as we move from a myopic policy (v̄R = v̄L = 0),
to resource gradients (where we use only v̄R ), to resource and task gradients
(where we use v̄R and v̄L ), to the arc gradients (where we use v̄tdl ). It is very
encouraging that the arc gradients give near optimal solutions. For stochastic
datasets, the best results are obtained when we use resource and task gradi-
ents; interestingly, the far more computationally intensive arc gradients do not
seem to add any value. It appears that the added precision that these gradients
bring are lost in the presence of uncertainty.
It is usually the case in transportation that we will know about resources and
tasks in the future, represented by the resource vector Rtt  . Simulations were
run which allowed the decision function to consider resources and tasks that
are known now (at time t) but actionable in the future (at time t  ) as long as
the actionable time is within a planning horizon T ph . We would expect to get
better results as the planning horizon is increased (which is an indication of
the value of advance information) but it is not clear how much this answer is
influenced by the use of value function approximations.
356 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

Fig. 11. The effect of resource gradients alone, resource and task gradients, and arc gradients, on solu-
tion quality for deterministic and stochastic problems.

Figure 12 shows the improvement in total profits (always measured against

the posterior bound) for both myopic and ADP policies. In each case, the poli-
cies were tested not only for different values of the planning horizon T ph (the
time range over which information is used) but also for different decision hori-
zons (the time period over which decisions in the future are locked into place).
The results show that increasing the planning horizon is significant in the case
of a myopic policy, but is surprisingly almost nonexistent when we use an ADP
policy. At a minimum, it appears that ADP policies benefit less from advance
information than a myopic policy. The implication is that measuring the value
of advance information using myopic policies might produce inflated estimates.

11.3 Batch processes for less-than-truckload trucking

Up to now, we have considered problems that can be viewed as “flow prob-

lems”. Although both the car distribution problem and load matching problem
involve discrete resources, these are naturally formed as linear programs with
integrality requirements, and they can either be solved as pure networks or
linear programs which provide near-integer solutions.
A problem class where this is not the case arises in batch processes such as
less-than-truckload trucking, where we have to consolidate smaller packages
onto larger vehicles. The decision is how to route the shipments, and when and
where to send the trucks. The decision to send a truck has to balance the value
 Ch. 5. Dynamic Models for Freight Transportation 357

Fig. 12. The value of advance information as estimated by a myopic policy, and using approximate
dynamic programming (from Spivey and Powell (2004)).

of moving the shipments now against various costs of holding the shipments.
Since sending a truck has the same mathematical structure as building a link in
the network, this has been referred to as the dynamic network design problem
(Crainic, 2000).
As of this writing, we are not aware of any computationally tractable al-
gorithms that have proven successful for realistic instances of the dynamic
network design problem, at least as it arises in less-than-truckload trucking.
Heuristics have been applied with varying degrees of success for static in-
stances (Powell, 1986; Crainic and Roy, 1988; Crainic and Rousseau, 1988),
but the dynamic versions are much harder. Most of the research has focused
on solving the problem with a single link. If we face the problem of dispatch-
ing a truck over a single link with homogeneous customers, then we have a
textbook dynamic program that is easy to solve.
The challenge arises when there is more than one product type. Speranza
and Ukovich (1994, 1996) develop optimal strategies for the deterministic mul-
tiproduct problem. Bertazzi and Speranza (1999a, 1999b) consider the deter-
ministic problem of shipping products from an origin to a destination through
one or several intermediate nodes and compare several classes of heuristics
including decomposition of the sequence of links, an EOQ-type solution and
a dynamic programming-based heuristic. The results assume that demands are
358 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

both deterministic and constant. Bertazzi et al. (2000) consider a stochastic,

multiproduct problem which is solved using approximate dynamic program-
ming techniques, but the method is only tested with a single product type.
In this section, we briefly review the single-link model described in Papadaki
and Powell (2003), which considers multiple product types (the method is
tested on a problem with 100 product types), uncertainty and significant non-
stationarities. For transportation, a “product type” might refer to the des-
tination of the shipment, the time at which the shipment has to arrive at
the destination, and the shipment priority (other attributes include size and
weight).

11.3.1 A single link model of the dynamic dispatch problem

We begin by defining the following:
Problem parameters:
K = set of customer classes
c d = cost to dispatch a vehicle
ckh = holding cost of class k ∈ K per time period per unit product
 
c h = c1h  c2h      c|hK| 
K = service capacity of the vehicle, giving the total number
of customers who can be served in a single dispatch
Activity variables:
Rxtk = number of customers in class k waiting at time t before new
arrivals have been added (the post-decision state variable)
 x
Rxt = Rtk k∈K 
t = vector random variable giving the number of arrivals in time t
R
of each type of customer
t
Rt = Rxt−1 + R
= number of customers waiting just before we make a decision
at time t
Decision variables:
xtk = the number of customers in class k who are served at time t
Xtπ (Rt ) = decision function returning xt 
Our feasible region Xt is defined by:
xtk  Rtk 

 K
k∈K
 Ch. 5. Dynamic Models for Freight Transportation 359

It is useful to define a dispatch function:


1 if k∈K xtk > 0
Zt (xt ) =
0 otherwise
Written this way, xt = Xtπ determines Zt . Later we show that we can actually
determine Zt first, and then compute Xtπ .
Objective function:
The one-period cost function is given by
t  xt ) = c d Zt (xt ) + c h (Rt − xt )
Ct (Rt  R
The objective function over the planning horizon can now be written

T
 
π
F(R0 ) = min E Ct Rt  Xt (Rt ) 
π∈Π
t=0

11.3.2 A solution algorithm for the single-link problem

The most common practical approach for this problem is to use a simple dis-
patch rule that might look like “if we have at least a certain amount of freight,
send the vehicle, otherwise hold it”. Such rules generally have to be time de-
pendent (“if it is Friday night. . . ”) or state dependent (“if the truck has been
held more than 12 hours and it is at least 80 percent full. . . ”) since the arrival
process of freight can be highly nonstationary. Any effort to use optimization
has to recognize that such rules can be fairly effective, and at a minimum are
easy to understand.
The value of solving this problem with optimization is to impose the dis-
cipline of formulating and minimizing a cost function. The problem is that
algorithms for network design have generally not been very effective. We pro-
pose to solve this problem (which is high-dimensional because of the multiple
product types) using approximate dynamic programming.
Using our standard methodology, we can formulate a decision function us-
ing
  
Xtπ (Rt ) = arg min Ct (Rt  xt ) + Vt+1
n
Rt (ωt ) xt  (102)
xt

It is possible to show that the value function increases monotonically in the re-
sources; in effect, the more customers that are waiting, the higher the cost. The
function itself is neither concave or convex, but especially when we introduce
uncertainty, it is approximately linear, suggesting the approximation
Vt (Rt ) = v̄t Rt 
Before, we estimated our slopes using dual variables from solving the subprob-
lems. In the case of batch processes, we do not have access to dual variables,
but we can use finite differences
ṽkt = Vt (Rt + ek  ω) − Vt (Rt  ω)
360 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

where ek is a |K|-dimensional vector with a single 1 in the kth element. Per-

forming these for each product class can be extremely expensive. A short-cut,
suggested in Graves (1981) in the context of multistage lot-sizing, is to assume
that increasing the number of customers (of any type) does not change the dis-
patch decision at time t. This makes it trivial to determine the marginal impact
of extra customers.
Rather than solve (102) to find xt , it actually makes more sense to find
zt first (that is, determine whether the batch will be sent or held), and then
 xt using a simple rule. If zt = 0, then clearly xt = 0. If zt = 1
compute
and k∈K R tk  K, then xtk = Rtk . The only interesting case arises when
zt = 1 and k∈K Rtk > K. Assume that the customer classes are ordered so
that c1h  c2h  · · ·  ckh . Then, we want to make sure customers in class 1
are all added to the batch before trying to add customers of class 2, and so
on, until we reach the capacity of the batch. Assuming we use this rule, the
steps of an approximate dynamic programming algorithm are summarized in
Figure 13.
The algorithm was tested on a single-product problem in Papadaki and
Powell (2003) so that comparisons against an optimal solution could be made
(using classical dynamic programming). The approximate dynamic program-
ming approach was compared to a rule which specified that the vehicle should
be dispatched when full, or if it had been held τ units of time. For each
dataset, τ was chosen so that it produced the best results (the assumption

Step 1. Given R0 : Set v̄t0 = 0 for all t. Set n = 1, t = 0.

Step 2. Set Rn0 = R0 and choose random sample ωn .
Step 3. Calculate
     
ztn = arg min c d zt + c h · Rnt − zt X Rnt + v̄tn Rnt − zt X Rnt
zt ∈{01}

and
   
t+1 ωn 
Rnt+1 = Rnt − zt X Rnt + R

Then define:
       
Vtn Rnt = min c d zt + c h · Rnt − zt X Rnt + v̄tn Rnt − zt X Rnt 
zt ∈{01}

Step 4. Update the approximation as follows. For each k = 1     |K|, let:

   
n Rn + ek − Vn Rn 
n =V
v̂tk t t t t

where ek is a |K|-dimensional vector with 1 in the kth entry and the rest zero. Update the
approximation by smoothing:
 
v̄tn = 1 − αn v̄tn−1 + αn v̂tn

Fig. 13. Approximate dynamic programming algorithm for the batch dispatch problem.
 Ch. 5. Dynamic Models for Freight Transportation 361

being that for any specific problem class, we could choose the best holding
time).
The method was tested on a variety of datasets which included both station-
ary and nonstationary arrival processes. An important determinant of perfor-
mance is the relative size of the holding cost per unit c h , and the dispatch cost
per unit of capacity c d /K. For this reason, the datasets were divided into three
groups: c h > c d /K, c h  c d /K, and c h < c d /K. In the first group, the holding
cost is high enough that we will tend to dispatch vehicles very quickly. In the
last group, the holding cost is low enough that we expect a dispatch-when-full
policy to be best. The middle group is the most interesting.
For each group, we further divided the runs between datasets where the
arrival process followed a periodic (e.g., daily) pattern (which was highly non-
stationary) from datasets where the arrival process was stationary. The results
are reported in Table 3, which shows that the approximate dynamic program-
ming approach works better than the optimized myopic heuristic, even for the
relatively easier case c h < c d /K where a dispatch-when-full strategy tends to
work well (this strategy still struggles when the arrival process of customers is
highly nonstationary).
Table 4 summarizes the results of runs done with 100 product types. In
this case, we cannot solve the problem optimally, but we are able to compare
against our myopic policy, and we have an idea from the single product case of
how well the myopic policy works. In this table, instead of reporting the percent
over optimal, we report the total costs of the approximate dynamic program-
ming policy divided by the total costs of the myopic policy. If we again focus on
the class of problems, where h < c/K (where the myopic policy will work the
best) we find that approximate dynamic programming strategies are approx-
imately three percent better than the myopic policy, which is very consistent
with our results from the single product case.

Table 3.
Fraction of total costs produced by each algorithm over the optimal cost: averages and standard devia-
tions within each group (from Papadaki and Powell, 2003)

Method: Linear Linear Linear Linear DWF-TC

Iterations: (25) (50) (100) (200)

h > c/K Periodic 0082 0070 0058 0062 0856

Stationary 0071 0050 0045 0038 0691
h  c/K Periodic 0040 0031 0024 0024 0270
Stationary 0057 0035 0023 0024 0195
h < c/K Periodic 0029 0025 0019 0019 0067
Stationary 0031 0019 0015 0013 0059
Average 0052 0038 0031 0030 0356
362 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

Table 4.
The (expected) cost of the approximate dynamic programming algorithm as a fraction of the cost of
the DWF-TC myopic heuristic for both scalar (single product) and multiple product problems (from
Papadaki and Powell, 2003)

Method: adp adp adp adp adp adp adp adp

scalar scalar scalar scalar mult. mult. mult. mult.
Iterations: (25) (50) (100) (200) (25) (50) (100) (200)

h > c/K Periodic 0602 0597 0591 0592 0633 0626 0619 0619
Stationary 0655 0642 0639 0635 0668 0660 0654 0650
h  c/K Periodic 0822 0815 0809 0809 0850 0839 0835 0835
Stationary 0891 0873 0863 0863 0909 0893 0883 0881
h < c/K Periodic 0966 0962 0957 0956 0977 0968 0965 0964
Stationary 0976 0964 0960 0959 0985 0976 0971 0969
Average 0819 0809 0803 0802 0837 0827 0821 0820

12 Perspectives on real-time problems

Dynamic models have a number of applications. We can simulate a dynamic

process to better understand how to operate a system under more realistic set-
tings. We might want to investigate the value of better information (through
automated detection systems, better databases, processes that require cus-
tomers to make requests known further into the future). A company might
want to produce short term tactical forecasts (for example, to help identify bot-
tlenecks) one or two days into the future given the state of the system right now.
One application that often arises at many companies is a desire to develop
models to help run their operations in real time. While there are a number of
ways to help a company make better decisions now (for example, with short
term tactical forecasts), there is often an interest in having computers tell dis-
patchers and planners what to do (that is, to automate the decision itself).
There is an important, qualitative difference between a real-time, dynamic
model that is used to look into the future to produce forecasts, and a real-time,
dynamic model that is used to make specific, actionable recommendations. The
first is trying to forecast activities in the future to help a human make better
decisions now. The second, which is much harder, is actually trying to tell a
human what to do right now. True real-time optimization can be viewed as the
information-age version of factory robots. We anticipate that these will steadily
make their way into operations, but adoption will be slow. Sometimes it is only
when we try to replace a human that we fully appreciate the diversity of tasks
that people perform.

References

Aronson, J., Chen, B. (1986). A forward network simplex algorithm for solving multiperiod network
flow problems. Naval Research Logistics Quarterly 33 (3), 445–467.
 Ch. 5. Dynamic Models for Freight Transportation 363

Aronson, J., Thompson, G.L. (1984). A survey on forward methods in mathematical programming.
Large Scale Systems 7, 1–16.
Assad, A. (1978). Multicommodity network flows: A survey. Networks 8 (1), 37–91.
Baker, S., Morton, D., Rosenthal, R., Williams, L. (2002). Optimizing military airlift. Operations Re-
search 50 (4), 582–602.
Barnhart, C., Hane, C.A., Johnson, E.L., Sigismondi, G. (1998). Branch-and-price: Column generation
for solving huge integer programs. Operations Research 46 (3), 316–329.
Bellman, R. (1957). Dynamic Programming. Princeton Univ. Press, Princeton.
Bellman, R., Dreyfus, S. (1959). Functional approximations and dynamic programming. Mathematical
Tables and Other Aids to Computation 13, 247–251.
Bertazzi, L., Speranza, M.G. (1999a). Inventory control on sequences of links with given transportation
frequencies. International Journal of Production Economics 59, 261–270.
Bertazzi, L., Speranza, M.G. (1999b). Minimizing logistic costs in multistage supply chains. Naval Re-
search Logistics 46, 399–417.
Bertazzi, L., Bertsekas, D., Speranza, M.G. (2000). Optimal and neuro-dynamic programming solutions
for a stochastic inventory trasportation problem. Unpublished technical report, Universita Degli
Studi Di Brescia.
Bertsekas, D., Tsitsiklis, J. (1996). Neuro-Dynamic Programming. Athena Scientific, Belmont, MA.
Birge, J., Louveaux, F. (1997). Introduction to Stochastic Programming. Springer-Verlag, New York.
Blum, J. (1954). Multidimensional stochastic approximation methods. Annals of Mathematical Statis-
tics 25, 737–744.
Cheung, R., Powell, W.B. (1996). An algorithm for multistage dynamic networks with random arc ca-
pacities, with an application to dynamic fleet management. Operations Research 44 (6), 951–963.
Cheung, R.K.-M., Powell, W.B. (2000). SHAPE: A stochastic hybrid approximation procedure for two-
stage stochastic programs. Operations Research 48 (1), 73–79.
Cinlar, E. (2003). Private communication.
Crainic, T. (2000). Network design in freight transportation. European Journal of Operational Re-
search 12 (2), 272–288.
Crainic, T., Rousseau, J.-M. (1988). Multicommodity, multimode freight transportation: A general
modeling and algorithmic framework for the service network design problem. Transportation Re-
search B 20, 290–297.
Crainic, T., Roy, J. (1988). OR tools for the tactical planning of freight transportation. European Journal
of Operations Research 33, 290–297.
Desrochers, M., Soumis, F. (1989). A column generation approach to the urban transit crew scheduling
problem. Transportation Science 23, 1–13.
Desrosiers, J., Soumis, F., Desrochers, M. (1984). Routing with time windows by column generation.
Networks 14, 545–565.
Desrosiers, J., Solomon, M., Soumis, F. (1995). Time constrained routing and scheduling. In:
Monma, C., Magnanti, T., Ball, M. (Eds.), Network Routing. Handbooks in Operations Research and
Management Science. North-Holland, Amsterdam, pp. 35–139.
Ford, L., Fulkerson, D. (1962). Flows in Networks. Princeton Univ. Press, Princeton.
Glockner, G.D., Nemhauser, G.L. (2000). A dynamic network flow problem with uncertain arc capaci-
ties: Formulation and problem structure. Operations Research 48, 233–242.
Glover, F., Karney, D., Klingman, D., Napier, A. (1974). A computation study on start procedures,
basis change criteria, and solution algorithms for transportation problems. Management Science 20,
793–813.
Glover, F., Klingman, D., Phillips, N.V. (1992). Network Models in Optimization and Their Application in
Practice. Wiley, New York.
Godfrey, G.A., Powell, W.B. (2001). An adaptive, distribution-free approximation for the newsvendor
problem with censored demands, with applications to inventory and distribution problems. Manage-
ment Science 47 (8), 1101–1112.
Godfrey, G.A., Powell, W.B. (2002). An adaptive, dynamic programming algorithm for stochastic re-
source allocation problems I: Single period travel times. Transportation Science 36 (1), 21–39.
364 W.B. Powell, B. Bouzaïene-Ayari and H.P. Simão

Graves, S.C. (1981). Multi-stage lot sizing: An iterative procedure. In: Schwarz, L.B. (Ed.), Multi-Level
Production/Inventory Control Systems: Theory and Practice. TIMS Studies in the Management Sciences,
vol. 16. North-Holland, New York, pp. 95–110.
Infanger, G. (1994). Planning under Uncertainty: Solving Large-Scale Stochastic Linear Programs. The
Scientific Press Series. Boyd & Fraser, New York.
Kall, P., Wallace, S. (1984). Stochastic Programming. Wiley, New York.
Kennington, J.L. (1978). A survey of linear cost multicommodity network flows. Operations Research 26,
209–236.
Kennington, J., Helgason, R. (1980). Algorithms for Network Programming. Wiley, New York.
Kushner, H.J., Yin, G.G. (1997). Stochastic Approximation Algorithms and Applications. Springer-
Verlag, New York.
Langley, R.W., Kennington, J.L., Shetty, C.M. (1974). Efficient computational devices for the capaci-
tated transportation problem. Naval Research Logistics Quarterly 21, 637–647.
Lasdon, L. (1970). Optimization Theory for Large Systems. MacMillan Co, New York.
Lavoie, S., Minoux, M., Odier, E. (1988). A new approach of crew pairing problems by column genera-
tion and application to air transport. European Journal of Operational Research 35, 45–58.
Marar, A., Powell, W.B. (2004). Using static flow patterns in time-staged resource allocation problems.
Technical report, Princeton University, Department of Operations Research and Financial Engi-
neering.
Marar, A., Powell, W.B., Kulkarni, S. (2006). Combining cost-based and rule-based knowledge in com-
plex resource allocation problems. IIE Transactions 38 (2), 159–172.
Morton, D.P., Rosenthal, R.E., Lim, T.W. (1996). Optimization modeling for airlift mobility. Military
Operations Research 1, 49–67.
Morton, D.P., Salmeron, J., Wood, R.K. (2003). A stochastic program for optimizing military sealift
subject to attack. Stochastic Programming e-print Series. Available at http://www.speps.info.
Papadaki, K., Powell, W.B. (2003). An adaptive dynamic programming algorithm for a stochastic mul-
tiproduct batch dispatch problem. Naval Research Logistics 50 (7), 742–769.
Powell, W.B. (1986). A local improvement heuristic for the design of less-than-truckload motor carrier
networks. Transportation Science 20 (4), 246–257.
Powell, W.B. (1987). An operational planning model for the dynamic vehicle allocation problem with
uncertain demands. Transportation Research 21, 217–232.
Powell, W.B., Shapiro, J.A., Simão, H.P. (2001). A representational paradigm for dynamic resource
transformation problems. In: Coullard, R.F.C., Owens, J.H. (Eds.), Annals of Operations Research.
J.C. Baltzer AG, Basel, pp. 231–279.
Powell, W.B., Shapiro, J.A., Simão, H.P. (2002). An adaptive dynamic programming algorithm for the
heterogeneous resource allocation problem. Transportation Science 36 (2), 231–249.
Powell, W.B., Wu, T.T., Whisman, A. (2004). Using low dimensional patterns in optimizing simulators:
An illustration for the airlift mobility problem. Mathematical and Computer Modelling 39, 657–675.
Puterman, M.L. (1994). Markov Decision Processes. Wiley, New York.
Robbins, H., Monro, S. (1951). A stochastic approximation method. Annals of Mathematical Statis-
tics 22, 400–407.
Sen, S., Higle, J. (1999). An introductory tutorial on stochastic linear programming models. Interfaces 29
(2), 33–61.
Shapiro, J., Powell, W.B. (2006). A metastrategy for dynamic resource management problems based on
informational decomposition. Informs Journal on Computing 18 (1), 43–60.
Speranza, M., Ukovich, W. (1994). Minimizing trasportation and inventory costs for several products
on a single link. Operations Research 42, 879–894.
Speranza, M., Ukovich, W. (1996). An algorithm for optimal shipments with given frequencies. Naval
Research Logistics 43, 655–671.
Spivey, M., Powell, W.B. (2004). The dynamic assignment problem. Transportation Science 38 (4), 399–
419.
Sutton, R. (1988). Learning to predict by the methods of temporal differences. Machine Learning 3,
9–44.
Sutton, R., Barto, A. (1998). Reinforcement Learning. MIT Press, Cambridge, Massachusetts.
 Ch. 5. Dynamic Models for Freight Transportation 365

Topaloglu, H., Powell, W.B. (2006). Dynamic programming approximations for stochastic, time-staged
integer multicommodity flow problems. Informs Journal on Computing 18 (1), 31–42.
Tsitsiklis, J., Van Roy, B. (1997). An analysis of temporal-difference learning with function approxima-
tion. IEEE Transactions on Automatic Control 42, 674–690.
Van Roy, B. (2001). Neuro-dynamic programming: Overview and recent trends. In: Feinberg, E.,
Shwartz, A. (Eds.), Handbook of Markov Decision Processes: Methods and Applications. Kluwer Aca-
demic, Boston, pp. 431–460.
Vance, P.H., Barnhart, C., Johnson, E.L., Nemhauser, G.L. (1997). Airline crew scheduling: A new
formulation and decomposition algorithm. Operations Research 45 (2), 188–200.
White, W. (1972). Dynamic transshipment networks: An algorithm and its application to the distribution
of empty containers. Networks 2 (3), 211–236.
Wu, T.T., Powell, W.B., Whisman, A. (2003). The optimizing simulator: An intelligent analysis tool
for the airlift mobility problem. Technical report, Princeton University, Department of Operations
Research and Financial Engineering.
C. Barnhart and G. Laporte (Eds.), Handbook in OR & MS, Vol. 14
Copyright © 2007 Elsevier B.V. All rights reserved
DOI: 10.1016/S0927-0507(06)14006-2

Chapter 6
Vehicle Routing
Jean-François Cordeau
Canada Research Chair in Logistics and Transportation, HEC Montréal,
3000 chemin de la Côte-Sainte-Catherine, Montréal, H3T 2A7, Canada
E-mail: [email protected]

Gilbert Laporte
Canada Research Chair in Distribution Management, HEC Montréal,
3000 chemin de la Côte-Sainte-Catherine, Montréal, H3T 2A7, Canada
E-mail: [email protected]

Martin W.P. Savelsbergh

School of Industrial and Systems Engineering, Georgia Institute of Technology,
Atlanta, GA 30332-0205, USA
E-mail: [email protected]

Daniele Vigo
Dipartimento di Elettronica, Informatica e Sistemistica, University of Bologna,
Viale Risorgimento 2, 40136 Bologna, Italy
E-mail: [email protected]

1 Introduction

The vehicle routing problem lies at the heart of distribution management. It

is faced each day by thousands of companies and organizations engaged in the
delivery and collection of goods or people. Because conditions vary from one
setting to the next, the objectives and constraints encountered in practice are
highly variable. Most algorithmic research and software development in this
area focus on a limited number of prototype problems. By building enough
flexibility in optimization systems one can adapt these to various practical con-
texts.
Much progress has been made since the publication of the first article on
the “truck dispatching” problem by Dantzig and Ramser (1959). Several vari-
ants of the basic problem have been put forward. Strong formulations have
been proposed, together with polyhedral studies and exact decomposition al-
gorithms. Numerous heuristics have also been developed for vehicle routing
problems. In particular the study of this class of problems has stimulated the
emergence and the growth of several metaheuristics whose performance is
constantly improving.
This chapter focuses on some of the most important vehicle routing prob-
lem types. A number of other variants have been treated in recent articles and

367
368 J.-F. Cordeau et al.

book chapters (see, e.g., Toth and Vigo, 2002a). The pickup and delivery vehi-
cle routing problem, which has also been extensively studied, is covered in the
“Transportation on Demand” chapter.
The remainder of this chapter is organized as follows. Section 2 is devoted to
the classical vehicle routing problem (simply referred to as VRP), defined with
a single depot and only capacity and route length constraints. Problems with
time windows are surveyed in Section 3. Section 4 is devoted to inventory rout-
ing problems which combine routing and customer replenishment decisions.
Finally, Section 5 covers the field of stochastic vehicle routing in which some
of the problem data are random variables.

2 The classical vehicle routing problem

The Classical Vehicle Routing Problem (VRP) is one of the most popular
problems in combinatorial optimization, and its study has given rise to several
exact and heuristic solution techniques of general applicability. It generalizes
the Traveling Salesman Problem (TSP) and is therefore NP-hard. A recent sur-
vey of the VRP can be found in the first six chapters of the book edited by
Toth and Vigo (2002a). The aim of this section is to provide a comprehensive
overview of the available exact and heuristic algorithms for the VRP, most of
which have also been adapted to solve other variants, as will be shown in the
remaining sections.
The VRP is often defined under capacity and route length restrictions.
When only capacity constraints are present the problem is denoted as CVRP.
Most exact algorithms have been developed with capacity constraints in mind
but several apply mutatis mutandis to distance constrained problems. In con-
trast, most heuristics explicitly consider both types of constraint.

2.1 Formulations

The symmetric VRP is defined on a complete undirected graph G = (V  E).

The set V = {0     n} is a vertex set. Each vertex i ∈ V \{0} represents a cus-
tomer having a nonnegative demand qi , while vertex 0 corresponds to a depot.
To each edge e ∈ E = {(i j): i j ∈ V  i < j} is associated a travel cost ce
or cij . A fixed fleet of m identical vehicles, each of capacity Q, is available at
the depot. The symmetric VRP calls for the determination of a set of m routes
whose total travel cost is minimized and such that: (1) each customer is vis-
ited exactly once by one route, (2) each route starts and ends at the depot,
(3) the total demand of the customers served by a route does not exceed the
vehicle capacity Q, and (4) the length of each route does not exceed a pre-
set limit L. (It is common to assume constant speed so that distances, travel
times and travel costs are considered as synonymous.) A solution can be viewed
as a set of m cycles sharing a common vertex at the depot. The asymmetric
VRP is similarly defined on a directed graph G = (V  A), where A = {(i j):
 Ch. 6. Vehicle Routing 369

i j ∈ V  i = j} is an arc set. In this case a circuit (directed cycle) is associated

with a vehicle route. Most results of Sections 2.1 and 2.2 apply to the symmetric
CVRP.
An integer linear programming formulation of the CVRP follows, where
for each edge e ∈ E the integer variable xe indicates the number of times
edge e is traversed in the solution. Let r(S) denote the minimum number of
vehicles needed to serve the customers of a subset S of customers. The value of
r(S) may be determined by solving an associated Bin Packing Problem (BPP)
with item set S and bins of capacity Q. Finally, for S ⊂ V , let δ(S) = {(i j):
i ∈ S j ∈ / S or i ∈ / S j ∈ S}. If S = {i}, then we simply write δ(i) instead
of δ({i}). The CVRP formulation proposed by Laporte et al. (1985) is then:

(CVRP1) minimize ce xe (1)
e∈E
subject to

xe = 2 i ∈ V \ {0} (2)
e∈δ(i)

xe = 2m (3)
e∈δ(0)

xe  2r(S) S ⊆ V \ {0} S = ∅ (4)
e∈δ(S)

xe ∈ {0 1} e∈
/ δ(0) (5)
xe ∈ {0 1 2} e ∈ δ(0) (6)
The degree constraints (2) state that each customer is visited exactly once,
whereas the depot degree constraint (3) means that m routes are created.
Capacity constraints (4) impose both the connectivity of the solution and the
vehicle capacity requirements by forcing a sufficient number of edges to enter
each subset of vertices. We note that since the BPP is NP-hard in the strong
 r(S) may be approximated from below by any BPP lower bound, such as
sense,
 i∈S qi /Q. Finally, constraints (5) and (6) impose that each edge between
two customers is traversed at most once and each edge incident to the depot is
traversed at most twice. In this latter case, the vehicle performs a route visiting
a single customer.
A widely used alternative formulation is based on the set partitioning or
set covering models. The formulation was originally proposed by Balinski and
Quandt (1964) and contains a potentially exponential number of binary vari-
ables. Let R = {R1      Rs } denote the collection of all feasible routes, with
s = |R|. Each route Rj has an associated cost γj , and aij is a binary coefficient
equal to 1 if and only if vertex i is visited (i.e., covered) by route Rj . The binary
variable xj , j = 1     s, is equal to 1 if and only if route Rj is selected in the
370 J.-F. Cordeau et al.

solution. The model is:


s
(CVRP2) minimize γj xj (7)
j=1
subject to
 s
aij xj = 1 i ∈ V \ {0} (8)
j=1

s
xj = m (9)
j=1
xj ∈ {0 1} j = 1     s (10)
Constraints (8) impose that each customer i is covered by exactly one route,
and (9) requires that m routes be selected. Because route feasibility is implic-
itly considered in the definition of R, this is a very general model which may
easily take additional constraints into account. Moreover, when the cost matrix
satisfies the triangle inequality (i.e., cij  cik + ckj for all i j k ∈ V ), the set
partitioning model CVRP2 may be transformed into an equivalent set cover-
ing model CVRP2 by replacing the equality sign with “” in (8). Any feasible
solution to CVRP2 is clearly feasible for CVRP2 , and any feasible solution
to CVRP2 may be transformed into a feasible CVRP2 solution of smaller or
equal cost. Indeed, if the CVRP2 solution is infeasible for CVRP2, then one
or more customers are visited more than once. These customers may there-
fore be removed from their route by applying shortcuts which will not increase
the solution cost because of the triangle inequality. The main advantage of us-
ing CVRP2 is that only inclusion-maximal feasible routes, among those with
the same cost, need be considered in the definition of R. This significantly
reduces the number of variables. In addition, when using CVRP2 the dual so-
lution space is considerably reduced since dual variables are restricted to be
nonnegative. One of the main drawbacks of models CVRP2 and CVRP2 lies
in their very large number of variables, which in loosely constrained medium
size instances may easily run into the billions. Thus, one has to resort to a
column generation algorithm to solve these problems. The linear program-
ming relaxation of these models tends to be very tight, as shown by Bramel and
Simchi-Levi (1997). Further details on these formulations and their extensions,
as well as additional formulations for the symmetric and asymmetric cases, can
be found in Laporte and Nobert (1987) and in Toth and Vigo (2002b, 2002d).

2.2 Exact algorithms for the CVRP

We now review the main exact approaches presented in the last two decades
for the solution of the CVRP. For a thorough review of previous exact meth-
ods, see Laporte and Nobert (1987). We first describe the algorithms based on
 Ch. 6. Vehicle Routing 371

branch-and-bound, including those that make use of the set partitioning for-
mulation and column generation schemes, and we then examine the algorithms
based on branch-and-cut. In practice, the CVRP turns out to be significantly
harder to solve than the TSP. The best CVRP algorithms can rarely tackle in-
stances involving more than 100 vertices, while TSP instances with hundreds
and even thousands of vertices are now routinely solved to optimality.

2.2.1 Branch-and-bound and set partitioning based algorithms

Several branch-and-bound algorithms are available for the solution of the
CVRP. Until the late 1980s, the most effective exact methods were mainly
branch-and-bound algorithms based on elementary combinatorial relaxations.
Recently, more sophisticated bounds have been proposed, namely those based
on Lagrangian relaxations or on the additive bounding procedure, which have
substantially increased the size of the problems that can be solved to optimal-
ity. We now describe some branch-and-bound algorithms with an emphasis on
lower bound computations which constitute the most critical component of
methods of this type. More details on the structure of branch-and-bound al-
gorithm strategies and dominance rules may be found in Toth and Vigo (1998,
2002c, 2002d). We also review in this section exact set partitioning based algo-
rithms for the CVRP.
Many different elementary combinatorial relaxations were used in early
branch-and-bound algorithms, including those based on the Assignment Prob-
lem (AP), on the degree-constrained shortest spanning tree, and on state-space
relaxation. Here we outline the two families of relaxations used as a basis for
the more recent branch-and-bound algorithms for the symmetric and asym-
metric CVRP. A first relaxation is obtained from the integer programming
formulations of these problems by dropping the connectivity and capacity con-
straints. In the symmetric case the resulting problem is a b-Matching Problem
(b-MP), i.e., the problem of determining a minimum cost set of cycles cover-
ing all vertices and such that the degree of each vertex i is equal to bi , where
bi = 2 for all the customer vertices, and b0 = 2m for the depot vertex. It is
easy to see that by adding m − 1 copies of the depot to G the relaxation be-
comes a 2-MP. In the asymmetric case the relaxed problem is the well-known
transportation problem which may be transformed into an AP by introducing
copies of the depot. Also in this case, the AP may be seen as the problem of
determining a set of circuits covering all vertices and such that each vertex has
one entering and one leaving arc. The solution of these relaxed problems may
be infeasible for the CVRP since the demand associated with a cycle or circuit
may exceed the vehicle capacity, and some of these may be disconnected from
the depot. The relaxed problems may then be solved in polynomial time (see,
e.g., Miller and Pekny, 1995, for the b-MP and Dell’Amico and Toth, 2000 for
the AP). However, the quality of the lower bounds obtained with these relax-
ations is generally very poor and not sufficient to solve instances with more
than 15 or 20 customers. Toth and Vigo (2002c) report average gaps in excess
of 20% with respect to the optimal solution value on benchmark CVRP in-
372 J.-F. Cordeau et al.

stances. The situation is slightly better for the AP relaxation of the asymmetric
CVRP that yields average gaps of about 10% or less. Laporte et al. (1986) have
proposed a branch-and-bound algorithm for asymmetric CVRP, based on the
AP relaxation and capable of solving randomly generated problems involving
tens of customers and between two and four vehicles.
The second family of elementary relaxations used in recent branch-and-
bound algorithms is based on degree-constrained spanning trees. These re-
laxations extend the well-known 1-tree relaxation proposed by Held and Karp
(1971) for the TSP. The earliest branch-and-bound algorithm based on this
relaxation, proposed by Christofides et al. (1981a), could only solve relatively
small instances. More recently, Fisher (1994) has presented another tree based
relaxation requiring the determination of a so-called m-tree, defined as a mini-
mum cost set of n + m edges spanning the graph. The approach used by Fisher
is based on CVRP1 with the additional assumption that single-customer routes
are not allowed. Fisher modeled the CVRP as the problem of determining an
m-tree with degree equal to 2m at the depot vertex, with additional constraints
on vehicle capacity and a degree of 2 for each customer vertex. The determi-
nation of an m-tree with degree 2m at the depot requires O(n3 ) time. The
degree-constrained m-tree relaxation is easily obtained from CVRP1 by re-
moving the degree constraints (2) for customer vertices and weakening the
capacity constraints (4) into connectivity constraints, i.e., by replacing their
right-hand side with 1. The m-tree solution is not always feasible for the CVRP
since some vertices may have a degree different from 2 and the demand asso-
ciated with the subtrees incident to the depot may exceed the vehicle capacity.
For the asymmetric CVRP, similar relaxations may be derived from directed
trees, also called arborescences, spanning the graph and having an outdegree
equal to m at the depot vertex. To obtain the final bound a minimum cost set of
m vertex-disjoint arcs entering the depot are added to the constrained arbores-
cence. In this case, the relaxed subproblem may be solved in polynomial time,
but again the quality of the resulting lower bound is very poor. Toth and Vigo
(2002c) report that on benchmark asymmetric instances, the average gap of
these relaxations with respect to the optimal solution value is larger than 25%.
Different improved bounding techniques were later developed to narrow
the gap between the lower bound and the optimal solution value of the
CVRP. These include two bounding procedures based on Lagrangian relax-
ation proposed by Fisher (1994) and Miller (1995). These are strengthenings
of the basic CVRP relaxations obtained by dualizing some of the relaxed
constraints in a Lagrangian fashion. In particular, they both include in the
objective function a suitable subset of the capacity constraints (4), whereas
the Fisher relaxation also incorporates degree constraints (2) which were re-
laxed in the m-tree relaxation. As in related problems, good values for the
Lagrangian multipliers associated with the relaxed constraints are determined
by using a subgradient optimization procedure (see, e.g., Held and Karp, 1971;
Held et al., 1974). The main difficulty associated with these relaxations lies in
the exponential cardinality of the set of relaxed constraints which does not
 Ch. 6. Vehicle Routing 373

allow for their complete inclusion in the objective function. These authors
include a limited family F of capacity constraints and iteratively generate
the constraints violated by the current solution of the Lagrangian problem.
The process terminates when no violated constraint is detected (hence the
Lagrangian solution is feasible) or a preset number of subgradient itera-
tions have been executed. Redundant constraints are periodically removed
from F . The relax-and-cut algorithm of Martinhon et al. (2000) generalizes
these Lagrangian-based approaches by also considering comb and multistar
inequalities, and moderately improves the quality of the Lagrangian bound.
Some exact algorithms for the CVRP are based on the set partitioning
formulation CVRP2. The first of these is due to Agarwal et al. (1989) who con-
sidered a relaxation of model CVRP2 not including constraints (9) and solved
the resulting model through column generation. Agarwal, Mathur, and Salkin
used their algorithm to solve seven Euclidean CVRP instances with up to 25
customers. Hadjiconstantinou et al. (1995) proposed a branch-and-bound al-
gorithm in which the lower bound was obtained by considering the dual of
the linear relaxation of model CVRP2, following the approach introduced
by Mingozzi et al. (1994). By linear programming duality, any feasible solu-
tion to this dual problem yields a valid lower bound. Hadjiconstantinou et al.
(1995) determined the heuristic dual solutions by combining two relaxations of
the original problem: the q-path relaxation of Christofides et al. (1981a) and
the m-shortest path relaxation of Christofides and Mingozzi (1989). The al-
gorithm was able to solve randomly generated Euclidean instances with up to
30 vertices and benchmark instances with up to 50 vertices. Further details on
set partitioning-based algorithms for the CVRP are provided in Bramel and
Simchi-Levi (2002).
Fischetti et al. (1994) have improved the AP relaxation of the asymmet-
ric CVRP by combining into an additive bounding procedure two new lower
bounds based on disjunctions on infeasible arc subsets and on minimum cost
flows. The additive approach was proposed by Fischetti and Toth (1989) and al-
lows for the combination of different lower bounding procedures, each exploit-
ing a different substructure of the problem under consideration. The resulting
branch-and-bound approach was able to solve randomly generated instances
containing up to 300 vertices and four vehicles. Other bounds for the asym-
metric CVRP may be derived by generalizing the methods proposed for the
symmetric case. For example, Fisher (1994) proposed a way of extending to the
asymmetric CVRP the Lagrangian bound based on m-trees. In this extension
the Lagrangian problem calls for the determination of an undirected m-tree
on the undirected graph obtained by replacing each pair of arcs (i j) and (j i)
with a single edge (i j) of cost cij = min{cij  cji }. No computational testing for
this bound was presented by Fisher (1994). Potentially better bounds may be
obtained by explicitly considering the asymmetry of the problem, i.e., by using
m-arborescences rather than m-trees and by strengthening the bound in a La-
grangian fashion as proposed by Toth and Vigo (1995, 1997) for the capacitated
shortest spanning arborescence problem and for the VRP with backhauls.
374 J.-F. Cordeau et al.

2.2.2 Branch-and-cut algorithms

Branch-and-cut algorithms currently constitute the best available exact ap-
proach for the solution of the CVRP. Research in this area has been strongly
motivated by the emergence and the success of polyhedral combinatorics as
a framework for the solution of hard combinatorial problems, particularly the
TSP. However, in a recent survey on branch-and-cut approaches for the CVRP,
Naddef and Rinaldi (2002) state: “. . . the amount of research effort spent to
solve CVRP by this method is not comparable with what has been dedicated to
the TSP [. . . the research in this field] is still quite limited and most of it is not
published yet”. In the following we summarize the main available branch-and-
cut approaches for the CVRP. The reader is referred to Naddef and Rinaldi
(2002) for a more detailed presentation.
The use of branch-and-cut for the CVRP is rooted in the exact algorithm
of Laporte et al. (1985). This algorithm uses the Linear Programming (LP) re-
laxation of model CVRP1 without capacity constraints (4) as a basis for the so-
lution of the VRP with capacity and maximum distance restrictions. This initial
relaxation is iteratively strengthened by adding violated capacity constraints
which are heuristically separated by considering the connected components in-
duced by the set of nonzero variables in the current LP solution. Gomory cuts
are also introduced at the root node of the branch-and-cut tree. The algorithm
was capable of solving randomly generated loosely constrained Euclidean and
non-Euclidean instances with two or three vehicles and up to 60 customers.
The first polyhedral study of the CVRP was presented by Cornuéjols and
Harche (1993). The presence of equalities (2) and (3) makes the CVRP
nonfully-dimensional. Therefore, as in the TSP, Cornuéjols and Harche first
considered the full-dimensional polyhedron, containing the CVRP polyhedron
as a face, associated with the so-called Graphical VRP (GVRP) where cus-
tomers may be visited more than once. The basic properties of the GVRP
polyhedron were also investigated. Conditions under which the nonnegativity,
degree and capacity constraints define facets of the GVRP and CVRP poly-
hedra were also determined. Cornuéjols and Harche have extended to the
GVRP and the CVRP several other families of valid inequalities proposed
for the TSP and the graphical TSP. In particular, comb, path, wheelbarrow,
and bicycle inequalities were extended to the capacitated case and again, suffi-
cient conditions under which these inequalities define facets of the GVRP and
CVRP polyhedra were derived. These inequalities were used by Cornuéjols
and Harche as cutting planes to solve two instances of CVRP with 18 and 50
customers, within a branch-and-cut algorithm. The detection of violated in-
equalities was performed manually, starting from the current optimal LP solu-
tion.
Augerat et al. (1995) have developed the first complete branch-and-cut ap-
proach for the CVRP. They described several heuristic separation procedures
for the classes of valid inequalities proposed by Cornuéjols and Harche, as well
as four new classes of valid inequalities. Separation procedures were further in-
vestigated by Augerat et al. (1999). The resulting approach was able to solve
 Ch. 6. Vehicle Routing 375

several CVRP instances containing up to 134 customers. Ralphs et al. (2003)

have presented a branch-and-cut algorithm for the CVRP in which an exact
separation of valid m-TSP inequalities is used in addition to heuristic separa-
tion of capacity inequalities. The resulting algorithm was implemented within
the SYMPHONY parallel branch-and-cut-and-price framework and was able
to solve several instances involving fewer than 100 vertices. Lysgaard et al.
(2004) have developed new separation procedures for most of the families of
valid inequalities proposed so far (see also Letchford et al., 2002). Their over-
all branch-and-cut approach, which is further enhanced by the use of Gomory
cuts, was able to solve within moderate computing times previously solved in-
stances and three new medium size ones.
Baldacci et al. (2004) have put forward a branch-and-cut algorithm based
on a two-commodity network flow formulation of the CVRP and requiring
a polynomial number of integer variables. It seems to provide an interesting
alternative to other classical formulations (see also Gouveia, 1995, for a single-
commodity formulation). The overall algorithm strengthens the LP relaxation
by adding violated capacity inequalities and implements various variable re-
duction and branching rules. The results obtained with this approach are com-
parable with those of the other branch-and-cut algorithms just described.
Finally, Fukasawa et al. (2006) have proposed a successful branch-and-cut-
and-price algorithm combining branch-and-cut with the q-routes relaxation
of Christofides et al. (1981a), used here in a column generation fashion.
This method produces tighter bounds than other branch-and-cut algorithms
and is capable of solving several previously unsolved instances with up to
75 customers. Baldacci et al. (2006) have used their set partitioning algorithm,
previously developed for a rollon–rolloff VRP, to solve difficult CVRP in-
stances. Their approach yields bounds whose quality is comparable to those
of Fukasawa et al. (2006), but seems much quicker.
Other branch-and-cut algorithms are described in Achuthan et al. (1996,
2003) and Blasum and Hochstättler (2000). We also mention that the poly-
hedral structure of the special case of CVRP where all the customers have a
unit demand was studied by Campos et al. (1991) and by Araque et al. (1990).
Branch-and-cut algorithms for this problem are presented by Araque et al.
(1994) and by Ghiani et al. (2006).

2.3 Heuristics for the VRP

An impressive number of heuristics have been proposed for the VRP. Ini-
tially these were mainly standard route construction algorithms, whereas more
recently powerful metaheuristic approaches have been developed. In the fol-
lowing we separately review these two families of algorithms. Almost all of
these methods were developed, described and tested for the symmetric VRP.
In addition, since finding a feasible solution with exactly m vehicles is itself
an NP-complete problem, almost all methods assume an unlimited number
376 J.-F. Cordeau et al.

of available vehicles. However, it should be observed that many of the pro-

posed methods may be quite easily adapted to take into account additional
practical constraints, although these may affect their overall performance (see,
e.g., Vigo, 1996, for an extension of some classical heuristics to the asymmetric
case).

2.3.1 Classical heuristics

Using the classification proposed by Laporte and Semet (2002), we describe
classical VRP heuristics under these headings: route construction methods,
two-phase methods, and route improvement methods.

Route construction heuristics. Route construction methods were among the

first heuristics for the CVRP and still form the core of many software imple-
mentations for various routing applications. These algorithms typically start
from an empty solution and iteratively build routes by inserting one or more
customers at each iteration, until all customers are routed. Construction algo-
rithms are further subdivided into sequential and parallel, depending on the
number of eligible routes for the insertion of a customer. Sequential methods
expand only one route at a time, whereas parallel methods consider more than
one route simultaneously. Route construction algorithms are fully specified
by their three main ingredients, namely an initialization criterion, a selection
criterion specifying which customers are chosen for insertion at the current
iteration, and an insertion criterion to decide where to locate the chosen cus-
tomers into the current routes.
The first and most famous heuristic of this group was proposed by Clarke
and Wright (1964) and is based on the concept of saving, an estimate of the
cost reduction obtained by serving two customers sequentially in the same
route, rather than in two separate ones. If i is the last customer of a route
and j is the first customer of another route, the associated saving is defined
as sij = ci0 + c0j − cij . If sij is positive, then serving i and j consecutively in
a route is profitable. The Clarke and Wright algorithm considers all customer
pairs and sorts the savings in nonincreasing order. Starting with a solution in
which each customer appears separately in a route, the customer pair list is ex-
amined and two routes are merged whenever this is feasible. Generally, a route
merge is accepted only if the associated saving is nonnegative but, if the num-
ber of vehicles is to be minimized, then negative saving merges may also be
considered. The Clarke and Wright algorithm is inherently parallel since more
than one route is active at any time. However, it may easily be implemented
in a sequential fashion. The resulting algorithm is quite fast but may have a
poor performance (see, e.g., Laporte and Semet, 2002). Golden et al. (1977),
Paessens (1988), and Nelson et al. (1985) have proposed various enhancement
strategies of the savings approach aimed at improving either its effectiveness or
its computational efficiency by means of better data structures. Other attempts
to improve the effectiveness of the savings method were made by Desrochers
and Verhoog (1989), Altinkemer and Gavish (1991), and by Wark and Holt
 Ch. 6. Vehicle Routing 377

(1994) who proposed to implement route merges by using a matching algo-

rithm, together with a more sophisticated estimate of actual merge savings.
The results obtained with these algorithms are in general better than those of
previous savings methods, but matching-based algorithms require much larger
computing times.
Another classical route construction heuristic is the sequential insertion al-
gorithm of Mole and Jameson (1976). The algorithm uses as selection and
insertion criterion the evaluation of the extra distance resulting from the in-
sertion of an unrouted customer k between two consecutive customers i and j
of the current route, namely α(i k j) = cik + ckj − λcij , where λ is a user-
controlled parameter. Variations of this criterion taking into account other
factors, such as the distance of the customer from the depot, were also con-
sidered. After each insertion, the current route is possibly improved by using
a 3-opt procedure. A more general and effective two-step insertion heuristic
was proposed by Christofides et al. (1979). In the first step, a sequential inser-
tion algorithm is used to determine a set of feasible routes. The second step
is a parallel insertion approach. For each route determined in the first step,
a representative customer is selected and a set of single-customer routes is
initialized with these customers. The remaining unrouted customers are then
inserted by using a regret criterion, where the difference between the best and
the second-best insertion cost is taken into account, and partial routes are im-
proved by means of a 3-opt procedure. The resulting algorithm is superior to
that of Mole and Jameson and represents a good compromise between effec-
tiveness and efficiency.

Two-phase heuristics. Two-phase methods are based on the decomposition of

the VRP solution process into the two separate subproblems:
(1) clustering: determine a partition of the customers into subsets, each cor-
responding to a route, and
(2) routing: determine the sequence of customers on each route.
In a cluster-first-route-second method, customers are first grouped into clus-
ters and the routes are then determined by suitably sequencing the customers
within each cluster. Different techniques have been proposed for the clustering
phase, while the routing phase amounts to solving a TSP.
The sweep algorithm, due to Wren (1971), Wren and Holliday (1972), and
Gillett and Miller (1974), is often referred to as the first example of cluster-
first-route-second approach. The algorithm applies to planar VRP instances.
The algorithm starts with an arbitrary customer and then sequentially assigns
the remaining customers to the current vehicle by considering them in order
of increasing polar angle with respect to the depot and the initial customer.
As soon as the current customer cannot be feasibly assigned to the current
vehicle, a new route is initialized with it. Once all customers are assigned to
vehicles, each route is separately defined by solving a TSP. Another early two-
phase method is the truncated branch-and-bound method of Christofides et al.
378 J.-F. Cordeau et al.

(1979) in which the set of routes is determined through an adaptation of an ex-

act branch-and-bound algorithm that uses a branching-on-routes strategy. The
decision tree contains as many levels as the number of available vehicles, and
at each level of the decision tree a given node corresponds to a partial solu-
tion made up of some complete routes. The descendant nodes correspond to
all possible routes including a subset of the unrouted customers. The running
time of the algorithm is controlled by limiting to one the number of routes
generated at each level.
The Fisher and Jaikumar (1981) algorithm solves the clustering step by
means of an appropriately defined Generalized Assignment Problem (GAP)
which calls for the determination of a minimum cost assignment of items to
a given set of bins of capacity Q, and where the items are characterized by a
weight and an assignment cost for each bin. Each vehicle is assigned a repre-
sentative customer, called a seed, and the assignment cost of a customer to a
vehicle is equal to its distance to the seed. The GAP is then solved, either opti-
mally or heuristically, and the final routes are determined by solving a TSP on
each cluster.
Another two-phase method working with a fixed number m of vehicles was
described by Bramel and Simchi-Levi (1995). This algorithm determines route
seeds by solving a capacitated location problem, where m customers are se-
lected by minimizing the total distance between each customer and its closest
seed, and by imposing that the total demand associated with each seed be at
most Q. Once seeds have been determined and the single-customer routes
are initialized, the remaining customers are inserted in the current routes by
minimizing insertion costs. Various ways of approximating the insertion cost
are proposed and analyzed. It is worth noting that all three cluster-first-route-
second approaches just described allow for a direct control of the number of
routes in the final solution, whereas the sweep algorithm does not. The perfor-
mance of these algorithms is generally comparable to that of route construction
algorithms in terms of effectiveness. The location based approach of Bramel
and Simchi-Levi produces better solutions but requires much larger computing
times.
A different family of two-phase methods is the class of so-called petal algo-
rithms. These generate a large set of feasible routes, called petals, and select
the final subset by solving a set partitioning model. Foster and Ryan (1976)
and Ryan et al. (1993) have proposed heuristic rules for determining the set of
routes to be selected, while Renaud et al. (1996b) have described an extension
that considers more involved configurations, called 2-petals, consisting of two
embedded or intersecting routes. The overall performance of these algorithms
is generally superior to that of the sweep algorithm.
Finally, in route-first-cluster-second methods, a giant TSP tour over all cus-
tomers is constructed in a first phase and later subdivided into feasible routes.
Examples of such algorithms are given by Beasley (1983), Haimovich and
Rinnooy Kan (1985), and Bertsimas and Simchi-Levi (1996), but the perfor-
mance of this approach is generally poor.
 Ch. 6. Vehicle Routing 379

Route improvement heuristics. Local search algorithms are often used to im-
prove initial solutions generated by other heuristics. Starting from a given solu-
tion, a local search method applies simple modifications, such as arc exchanges
or customer movements, to obtain neighbor solutions of possibly better cost.
If an improving solution is found, it then becomes the current solution and the
process iterates; otherwise a local minimum has been identified.
A large variety of neighborhoods are available. These may be subdivided
into intra-route neighborhoods, if they operate on a single route at a time, or
inter-route neighborhoods if they consider more than one route simultaneously.
The most common neighborhood type is the λ-opt heuristic of Lin (1965) for
the TSP, where λ edges are removed from the current solution and replaced by
λ others. The computing time required to examine all neighbors of a solution
is proportional to nλ . Thus, only λ = 2 or 3 are used in practice. As an alterna-
tive, one can use restricted neighborhoods characterized by subsets of moves
associated with larger λ values, such as Or-exchanges (Or, 1976) or the 4-opt*
neighborhood of Renaud et al. (1996a) which considers only a subset of all
potential 4-opt exchanges. Laporte and Semet (2002) have conducted a com-
putational comparison of some basic route improvement procedures. More
complex inter-route neighborhoods are analyzed by Thompson and Psaraftis
(1993), Van Breedam (1994), and Kindervater and Savelsbergh (1997).

2.3.2 Metaheuristics
Several metaheuristics have been applied to the VRP. With respect to clas-
sical heuristics, they perform a more thorough search of the solution space and
are less likely to end with a local optimum. These can be broadly divided into
three classes:
(1) local search, including simulated annealing, deterministic annealing,
and tabu search;
(2) population search, including genetic search and adaptive memory pro-
cedures;
(3) learning mechanisms, including neural networks and ant colony opti-
mization.
The best heuristics often combine ideas borrowed from different meta-
heuristic principles. Recent surveys of VRP metaheuristics can be found in
Gendreau et al. (2002), Cordeau and Laporte (2004), and Cordeau et al.
(2005).
Local search algorithms explore the solution space by iteratively moving
from a solution xt at iteration t to a solution xt+1 in the neighborhood N(xt )
of xt until a stopping criterion is satisfied. If f (x) denotes the cost of solution x,
then f (xt+1 ) is not necessarily smaller than f (xt ). As a result, mechanisms
must be implemented to avoid cycling. In simulated annealing, a solution x is
drawn randomly from N(xt ). If f (x)  f (xt ), then xt+1 := x. Otherwise,

x with probability pt 
xt+1 :=
xt with probability 1 − pt 
380 J.-F. Cordeau et al.

where pt is a decreasing function of t and of f (x) − f (xt ). This probability is

often equal to
 
f (x) − f (xt )
pt = exp − 
θt
where θt is the temperature at iteration t, usually defined as a nonincreasing
function of t. Deterministic annealing (Dueck, 1990, 1993) is similar. There
are two main versions of this algorithm: in a threshold-accepting algorithm,
xt+1 := x if f (x) < f (xt ) + θ1 , where θ1 is a user controlled parameter; in
record-to-record travel, a record is the best known solution x∗ , and xt+1 := x
if f (xt+1 ) < θ2 f (x∗ ), where θ2 is also user controlled. In tabu search, in or-
der to avoid cycling, any solution possessing some given attribute of xt+1 is
declared tabu for a number of iterations. At iteration t, the search moves to
the best nontabu solution x in N(xt ). These local search algorithms are rarely
implemented in their basic version, and their success depends on the careful
implementation of several mechanisms. The rule employed to define neighbor-
hoods is critical to most local search heuristics. In simulated annealing several
rules have been proposed to define θt (see Osman, 1993). Tabu search relies
on various strategies to implement tabu tenures (also known as short term
memory), search diversification (also known as long term memory), and search
intensification which accentuates the search in a promising region.
Population search algorithms operate on several generations of solution
populations. In genetic search it is common to repeat the following oper-
ation k times: extract two parent solutions from the populations to create
two offspring using a crossover operation, and apply a mutation operation
to each offspring; then remove the 2k worst elements from the popula-
tion and replace them with the 2k offspring. Several crossover rules are
available for sequencing problems (Bean, 1994; Potvin, 1996; Drezner, 2003;
Prins, 2004). In adaptive memory procedures, an offspring is created by ex-
tracting and recombining elements of several parents. In the initial version
proposed by Rochat and Taillard (1995) for the VRP, nonoverlapping routes
are extracted from several parents to create a partial solution. This solution is
then gradually completed and optimized by tabu search.
Neural networks are models composed of richly interconnected units
through weighted links, like neurons in the brain. They gradually construct a
solution through a feedback mechanism that modifies the link weights to better
match an observed output to a described output. In the field of vehicle routing
neural network models called the elastic net and the self-organizing map are
deformable templates that adjust themselves to the contour of the vertices to
generate a feasible VRP solution. An example is provided by Ghaziri (1993).
Ant colony algorithms (see Dorigo et al., 1999) also use a learning mechanism.
They are derived from an analogy with ants which lay some pheromone on
their trail when foraging for food. With time more pheromone is deposited
on the more frequented trails. When constructing a VRP solution a move can
 Ch. 6. Vehicle Routing 381

be assigned a higher probability of being selected if it has previously led to a

better solution in previous iterations.
In what follows we summarize the most effective metaheuristics for the
CVRP. Initially the best methods were almost exclusively based on tabu search
but in recent years several excellent methods inspired from different paradigms
have been proposed.

Local search heuristics. A limited number of simulated annealing heuristics

for the CVRP were proposed in the early 1990s. Osman’s implementation
(Osman, 1993) is the most involved and also the most successful. It defines
neighborhoods by means of a 2-interchange scheme and applies a different
rule of temperature changes. Instead of using a nonincreasing function, as do
most authors in the field, Osman decreases θt continuously as long as the so-
lution improves, but whenever xt+1 = xt , θt is either halved or replaced by the
temperature at which the incumbent was identified. This algorithm succeeded
in producing good solutions but was not competitive with the best tabu search
implementations available at the same period.
A large number of tabu search algorithms have been produced over the past
fifteen years (a survey is available in Cordeau and Laporte, 2004). In the first
known implementation, due to Willard (1989), a CVRP solution is represented
as a giant tour containing several copies of the depot and inter-depot chains
corresponding to feasible vehicle routes, and neighborhoods are defined by
means of 3-opt exchanges. The method was soon to be superseded by more
powerful algorithms, including those of Osman (1993), Taillard (1993), and
Gendreau et al. (1994).
Taillard’s algorithm remains to this day one of the most successful tabu
search implementations for the CVRP. It is based on the use of an 1-inter-
change mechanism to define neighbor solutions, combined with periodic route
reoptimizations by means of an exact TSP algorithm (Volgenant and Jonker,
1983). The algorithm also uses random tabu durations. A continuous diversifi-
cation mechanism that penalizes frequently performed moves is implemented
in order to provide a more thorough exploration of the search space. Finally,
Taillard’s algorithm employs a decomposition scheme that allows for the use
of parallel computing. In planar problems the customer set is partitioned into
sectors and then concentric rings, while in random instances the regions are
defined by means of shortest spanning arborescences rooted at the depot.
The region boundaries are periodically updated to produce a diversification
effect.
The Taburoute algorithm of Gendreau et al. (1994) moves at each itera-
tion a vertex from its current route to another route containing one of its
closest neighbors. Insertions are performed simultaneously with a local re-
optimization of the route, based on the GENI procedure (Gendreau et al.,
1992). Only a subset of vertices are considered for reinsertion at any given
iteration. No vertex can return to its former route during the next θ itera-
tions, where θ is randomly selected in a closed interval. Taburoute also uses
382 J.-F. Cordeau et al.

a continuous diversification mechanism. During the course of the search in-

feasible solutions are penalized. This mechanism is implemented by replacing
the solution value f (x) associated at solution x with a penalized objective
f  (x) = f (x) + αQ(x) + βL(x), where Q(x) is the total capacity violation
of solution x and L(x) is the total route length violation. The two parameters
α and β self-adjust during the search to produce a mix of feasible and infeasible
solutions: every μ iterations, α (resp. β) is divided by 2 if the past μ solutions
were feasible with respect to capacity (resp. route length), or multiplied by 2
if they were all infeasible with respect to capacity (resp. route length). Other
features of Taburoute include the use of random tabu durations, periodic route
reoptimizations by means of the US procedure of Gendreau et al. (1992), false
starts to initialize the search, and a final intensification phase around the best
known solution.
The Rego and Roucairol (1996) Tabuchain algorithm is based on the use of
ejection chains involving routes to define neighborhoods. This process bumps
a vertex from one route of the chain to another route. The last bumped ver-
tex may be relocated in the position of the first bumped vertex or elsewhere.
The process ensures that no arc or edge is considered more than once in
the solution. As in Taburoute, intermediate infeasible solutions are allowed.
The authors have also implemented a sequential and a parallel version of
their method. Another ejection scheme, called Flower, was later developed
by Rego (1998). It is based on the idea of exploiting the representation of
routes as blossoms and of paths as stems, and of performing ejection moves
by means of edge deletions and creations. This method was not as successful
as Tabuchain. Another method employing ejection chains was developed by
Xu and Kelly (1996). It oscillates between ejection chains and vertex swaps
between two routes. The ejection chains are obtained by solving an auxiliary
network flow problem. On the whole this method succeeded in obtaining sev-
eral good CVRP solutions on benchmark instances but it is rather involved and
time consuming.
More recently, Ergun et al. (2003) have developed a Very Large Neigh-
borhood Search (VLNS) algorithm for the VRP. This algorithm operates
on several routes simultaneously, not unlike what is done in cyclic transfers
(Thompson and Psaraftis, 1993) or in ejection chains. Neighborhoods are de-
fined by a combination of 2-opt moves, vertex swaps between routes, and vertex
insertions in different routes. The best choice of moves and of routes involved
in the moves is determined through the solution of a network flow problem
on an auxiliary graph. One advantage of VLNS is that it allows a broad search
by acting on several routes at once. Its main disadvantage lies in the effort
required at each iteration to perform moves.
A very useful concept put forward by Toth and Vigo (2003) is that of Gran-
ular Tabu Search (GTS). This algorithm a priori removes from the graph long
edges that are unlikely to belong to an optimal solution. To determine these
edges, the problem is first solved by means of a fast heuristic, e.g., the Clarke
and Wright (1964) algorithm, and the average edge cost c̄ in this solution
 Ch. 6. Vehicle Routing 383

is determined. Then only two families of edges are retained: those incident
to the depot, and those whose cost does not exceed βc̄, where β is a user-
defined sparsification parameter. The authors show that on benchmark in-
stances, choosing β in [10 20] yields the elimination of between 80–90% of
all edges. Granular tabu search was implemented in conjunction with some of
the features of Taillard’s algorithm (Taillard, 1993) and Taburoute (Gendreau
et al., 1994), and neighbor solutions were obtained by performing intra-route
and inter-route exchanges.
Deterministic annealing was first applied to the VRP by Golden et al. (1998)
and more recently by Li et al. (2005). The latter algorithm combines the record-
to-record principle of Dueck (1993) with GTS. It works on a sparsified graph
containing only a proportion α of the 40 shortest edges incident to each ver-
tex, where α varies throughout the algorithm. The algorithm is applied several
times from three initial solutions generated by the Clarke and Wright (1964)
algorithm, with savings sij defined as ci0 + c0j − λcij , and λ = 06 14 and 16.
Neighbors are defined by means of intra- and inter-route 2-opt moves, and
nonimproving solutions are accepted as long as their cost does not exceed that
of the incumbent by more than 1%. Whenever the solution has not improved
for a number of iterations, a perturbation is applied to the best known solution
to restart the search. This is achieved by temporarily moving some vertices to
different positions.

Population search heuristics. The Adaptive Memory Procedure (AMP) put

forward by Rochat and Taillard (1995) constitutes a major contribution to the
field of metaheuristics. Initially developed in the context of the VRP, it is of
general applicability and has been used, for example, to solve political district-
ing problems (Bozkaya et al., 2003). An adaptive memory is a pool of good
solutions which is updated by replacing its worst elements with better ones. In
order to generate a new solution, several solutions are selected from the pool
and recombined. In the context of the VRP, vehicle routes are extracted from
these solutions and used as the basis of a new solution. The extraction process
is applied as long as it is possible to identify routes that do not overlap with pre-
viously selected routes. When this is no longer possible, a search process (e.g.,
tabu search) is initiated from a partially constructed solution made up of the
selected routes and some unrouted customers. Any solution constructed in this
fashion replaces the worst solution of the pool if it has a better cost. Tarantilis
and Kiranoudis (2002) have proposed a variant to this scheme. In a first phase
a solution is obtained by means of the Paessens (1988) constructive procedure,
which is an application of the Clarke and Wright savings heuristic followed by
2-opt moves, vertex swaps between routes, and vertex reinsertions. In order to
generate new solutions from the adaptive memory, Tarantilis and Kiranoudis
extract route segments, called bones, as opposed to full vehicle routes as did
Rochat and Taillard.
Prins (2004) has developed an algorithm combining two main features of
evolutionary search, namely crossovers and mutations. Crossovers consist of
384 J.-F. Cordeau et al.

creating offspring solutions from parents, while mutations are obtained here
by applying a local search algorithm to an offspring. This combination of so-
lution recombination and local search is sometimes referred to as a memetic
algorithm (Moscato and Cotta, 2003). In this algorithm, solutions are repre-
sented as a giant tour without trip delimiters. To create an offspring from two
parents, a chain (i     j) is first selected from the first parent and the vertices
of the second parent are scanned from position j + 1 by skipping those of the
chain (i     j). A second offspring is generated in a similar way by reversing
the roles of the two parents. Offspring are improved by applying a combination
of vertex and edge reinsertions, vertex swaps, combined vertex and edge swaps.
Two other memetic algorithms have recently been proposed by Berger and
Barkaoui (2004) and by Mester and Bräysy (2005). The first works on two pop-
ulations whose sizes are kept constant through the replacement of parents by
newly created offspring, and migrations take place between the two popula-
tions. Offspring are obtained by combining routes from two parents as long as
this can be done without overlapping, and by inserting the unrouted customers
according to a proximity criterion. A VLNS heuristic (Shaw, 1998) combin-
ing three insertion mechanisms is then applied to the offspring, followed by
an improvement scheme consisting of removing vertices from the solution and
reinserting them by means of the I1 procedure of Solomon (1987).
The Active Guided Evolution Strategies (AGES) of Mester and Bräysy was
initially developed to solve the VRP with time windows and was later applied to
the classical VRP. It combines local search (Voudouris, 1997) with an evolution
strategy (Rechenberg, 1973) to produce an iterative two-stage procedure. The
evolutionary strategy uses a deterministic rule to select a parent solution and
create a single offspring from a single parent. The offspring replaces the parent
if it improves upon it. Offspring are improved by means of an elaborate search
procedure combining granular tabu search, continuous diversification, vertex
swaps and moves, 2-opt* moves (Potvin and Rousseau, 1995), VLNS (Shaw,
1998), and restarts.

Learning mechanisms. A limited number of heuristics based on learning

mechanisms have been proposed for the VRP. None of the known neural
networks based methods is satisfactory, and the early ant colony based heuris-
tics could not compete with the best available approaches. Recently, how-
ever, Reimann et al. (2004) have proposed a well-performing heuristics called
D-ants. The method repeatedly applies two phases until a stopping criterion
is reached. In the first phase, a first generation of good solutions is gener-
ated through the applications of a savings based heuristic (Clarke and Wright,
1964) and a 2-opt improvement procedure is applied to each solution. New
generations of solutions are then created by benefiting from the knowledge
gained in producing past generations. Thus, instead of using the standard sav-
β
ings sij = ci0 + c0j − cij , an attractiveness value χij = τijα sij is now employed,
where τijα contains information on how good linking i and j turned out to be
in previous generations, and α and β are user-controlled parameters. Vertices
 Ch. 6. Vehicle Routing 385

i and j are linked with probability pij = χij /( (h )∈Ωk χh ), where Ωk is the
set of the feasible (i j) pairs yielding the k best savings. In the second phase
the best solution identified in the first phase is decomposed into subproblems
which are then reoptimized using the procedure used in the first phase.

Computational comparison of metaheuristics. Cordeau et al. (2005) provide

a computational comparison of recent VRP heuristics on the 14 Christofides
et al. (1979) instances (50  n  199) and on the 20 larger Li et al. (2005)
instances (200  n  480). Most metaheuristics used in the comparison con-
sistently yield solutions whose value lies within 1% of the best known value.
On the Christofides et al. (1979) instances, the best solutions are obtained
by Taillard (1993), Rochat and Taillard (1995), and Mester and Bräysy (2005).
If the two instance sets are taken together, the best performers, in terms of ac-
curacy and computing time are probably Mester and Bräysy (2005), Tarantilis
and Kiranoudis (2002), and Prins (2004). It should be noted that these three
methods all combine population search and local search, thus allowing for a
broad and deep exploration of the solution space.
As noted by Cordeau et al. (2002b) heuristics should not be judged solely
on speed and accuracy. Simplicity and flexibility are also important. In this re-
spect the Li et al. (2005) record-to-record algorithm is rather interesting: this
algorithm possesses a simple structure and is capable of generating very high
quality solutions. As far as flexibility is concerned, the granularity principle
(Toth and Vigo, 2003) and the adaptive memory concept (Rochat and Taillard,
1995) are general and useful ideas which can easily be applied to other prob-
lems.

3 The vehicle routing problem with time windows

The Vehicle Routing Problem with Time Windows (VRPTW) is an impor-

tant generalization of the classical VRP in which service at every customer i
must start within a given time window [ai  bi ]. A vehicle is allowed to arrive be-
fore ai and wait until the customer becomes available, but arrivals after bi are
prohibited. The VRPTW has numerous applications in distribution manage-
ment. Common examples are beverage and food delivery, newspaper delivery,
and commercial and industrial waste collection (see, e.g., Golden et al., 2002).
The VRPTW is NP-hard since it generalizes the CVRP which is obtained
when ai = 0 and bi = ∞ for every customer i. In the case of a fixed fleet
size, even finding a feasible solution to the VRPTW is itself an NP-complete
problem (Savelsbergh, 1985). As a result, research on the VRPTW has concen-
trated on heuristics. Nevertheless, when the problem is sufficiently constrained
(i.e., when time windows are sufficiently narrow), realistic size instances can be
solved optimally through mathematical programming techniques. This section
presents a mathematical formulation of the VRPTW followed by a description
of some of the most important available exact and heuristic algorithms. It is
386 J.-F. Cordeau et al.

worth pointing out that while exact methods usually minimize distance, most
heuristics consider a hierarchical objective which first minimizes the number
of vehicles used and then distance.

3.1 Formulation of the VRPTW

The VRPTW can be defined on a directed graph G = (V  A), where

|V | = n + 2, and the depot is represented by the two vertices 0 and n + 1.
Feasible vehicle routes then correspond to paths starting at vertex 0 and end-
ing at vertex n + 1. The set of vehicles is denoted by K, with |K| = m. Let
si denote the service time at i (with s0 = sn+1 = 0) and let tij be the travel time
from i to j. In addition to the time window [ai  bi ] associated with each vertex
i ∈ N = V \ {0 n + 1}, time windows [a0  b0 ] and [an+1  bn+1 ] can also be asso-
ciated with the depot vertex. If no particular restrictions are imposed on vehicle
availability, one may simply set a0 = mini∈N {ai − t0i }, b0 = maxi∈N {bi − t0i },
an+1 = mini∈N {ai + si + tin+1 }, and bn+1 = maxi∈N {bi + si + tin+1 }. As in
the CVRP, let qi denote the demand of customer i, and let Q be the vehicle
capacity.
While several models are available for the VRPTW, this problem is often
formulated as a multicommodity network flow model with time window and ca-
pacity constraints. This model involves two types of variables: binary variables
xkij , (i j) ∈ A, k ∈ K, equal to 1 if and only if arc (i j) is used by vehicle k, and
continuous variables wik , i ∈ N, k ∈ K, indicating the time at which vehicle k
starts servicing vertex i. Let δ+ (i) = {j: (i j) ∈ A} and δ− (j) = {i: (i j) ∈ A}.
The problem can then be stated as follows (see, e.g., Desrochers et al., 1988):
 
minimize cij xkij (11)
k∈K (ij)∈A
subject to
 
xkij = 1 i ∈ N (12)
k∈K j∈δ+ (i)

xk0j = 1 k ∈ K (13)
j∈δ+ (0)
 
xkij − xkji = 0 k ∈ K j ∈ N (14)
i∈δ− (j) i∈δ+ (j)

xkin+1 = 1 k ∈ K (15)
i∈δ− (n+1)
 
xkij wik + si + tij − wjk  0 k ∈ K (i j) ∈ A (16)
ai  wik
 bi  k ∈ K i ∈ V  (17)
 
qi xkij  Q k ∈ K (18)
i∈N j∈δ+ (i)
 Ch. 6. Vehicle Routing 387

xkij ∈ {0 1} k ∈ K (i j) ∈ A (19)

The objective function (11) minimizes the total routing cost. Constraints
(12) state that each customer is visited exactly once, while constraints (13)–(15)
ensure that each vehicle is used exactly once and that flow conservation is
satisfied at each customer vertex. The consistency of the time variables wik
is ensured through constraints (16) while time windows are imposed by (17).
These constraints also eliminate subtours. Finally, constraints (18) enforce the
vehicle capacity restriction.
Formulation (11)–(19) is nonlinear because of constraints (16). These con-
straints can, however, be linearized as follows:
 
wjk  wik + si + tij − Mij 1 − xkij  k ∈ K (i j) ∈ A (20)
where Mij = max{0 bi + si + tij − aj } is a constant. As suggested by Desrochers
and Laporte (1991), the bounds on the time variables bki can also be strength-
ened:

wik  ai + max{0 aj − ai + sj + tji }xkji  k ∈ K i ∈ V  (21)
j∈δ− (i)

wik  bi − max{0 bi − bj + si + tij }xkij  k ∈ K i ∈ V  (22)
j∈δ+ (i)

3.2 Exact algorithms for the VRPTW

As for most other vehicle routing problems, it is difficult to solve the

VRPTW exactly through classical simplex-based branch-and-bound methods,
even for small instances. This is in large part explained by the fact that the LP
relaxation of the problem provides a weak lower bound. The first optimiza-
tion algorithm for the VRPTW can be attributed to Kolen et al. (1987) who
used dynamic programming coupled with state space relaxation (Christofides
et al., 1981b) to compute lower bounds within a branch-and-bound algorithm.
Instances with n  15 were solved using this approach. Most subsequent algo-
rithms rely either on the generation of valid inequalities to strengthen the LP
relaxation or on mathematical decomposition techniques. This section reviews
the three main available approaches: Lagrangian relaxation, column gener-
ation, and branch-and-cut. Additional references on the subject can also be
found in the Cordeau et al. (2002a) review.

3.2.1 Lagrangian relaxation based algorithms

Lagrangian relaxation can be applied to the VRPTW in several ways. It
is well known that when the subproblem obtained by relaxing some of the
constraints possesses the integrality property, the best lower bound obtained
by Lagrangian relaxation (i.e., the value of the Lagrangian dual) is equal to
the value of the linear programming relaxation of the original problem. But
388 J.-F. Cordeau et al.

as mentioned above, the LP relaxation of formulation (11)–(19) provides a

weak lower bound which will usually prevent the problem from being solved
by branch-and-bound. As a result, successful implementations of Lagrangian
relaxation for the VRPTW should retain at least some of the complicating con-
straints in the subproblem.
Fisher (1994) and Fisher et al. (1997) have described Lagrangian relaxation
based on m-trees (see Section 2.2.1). This approach relaxes the flow conserva-
tion constraints as well as the capacity and time window constraints. Violated
capacity constraints are handled by identifying subsets of customers S ⊆ N
that must be visited by at least κ(S) vehicles and imposing the constraint
  
xkij  κ(S) (23)
k∈K i∈V \S j∈S

These constraints are relaxed in a Lagrangian fashion so that the resulting

problem remains an m-tree problem with modified costs. Time windows are
handled similarly by identifying infeasible paths and imposing the constraint
that at least one arc in the path be left out of the solution. This approach has
solved a few of the Solomon (1987) test instances with n = 100. In addition
to the m-tree relaxation method, Fisher et al. (1997) have also experimented
with a variable splitting approach in which additional variables yik , equal to 1
if and only if customer i is visited by vehicle k, are introduced in the formu-

lation, and the constraints j∈V xkij = yik (i ∈ N, k ∈ K) are dualized. The
Lagrangian subproblem decomposes into a semi-assignment problem in the
yik variables which is solvable by inspection, and a set of m elementary shortest
path problems with time windows and capacity constraints.
Another possible Lagrangian relaxation consists of dualizing the demand
constraints. Let λ = (λi ) (i ∈ N) be the vector of multipliers associated with
constraints (12) requiring that each customer be visited exactly once. For given
values of the multipliers, the Lagrangian subproblem L(λ) obtained by relax-
ing these constraints in the objective function is
  
min (cij − λi ) xkij + λi  (24)
k∈K (ij)∈A i∈N

subject to constraints (13)–(19).

This subproblem does not possess the integrality property. It does, however,
decompose into m disjoint elementary shortest-path problems with capacity
and time window constraints. When all vehicles are identical, a single prob-
lem can be solved to compute the lower bound. The Lagrangian dual, i.e.,
the problem of finding optimal multipliers that maximize L(λ), is a concave
nondifferentiable maximization problem. Using subgradient and bundle meth-
ods, Kohl and Madsen (1997) were able to solve some instances with up to
100 customers. They reported optimal solutions to each of the 27 clustered
and short-horizon Solomon instances.
 Ch. 6. Vehicle Routing 389

Kallehauge et al. (2006) have developed a stabilized cutting-plane algorithm

to solve the Lagrangian dual. Cutting planes are generated by solving the La-
grangian subproblem and are introduced in a master problem which imposes
bounds (i.e., a trust region) on the dual variables to ensure the stability of their
values from one iteration to the next. Optimizing the relaxed master problem
(a maximization linear program) provides a lower bound on the value of the
original problem. To obtain feasible integer solutions, the cutting-plane algo-
rithm is embedded within a branch-and-bound algorithm and valid inequalities
are introduced in the master problem. Because the relaxed master problem
is stated on the dual variables, violated subtour elimination constraints and
2-path inequalities (see Section 3.2.2) are added as columns to this problem.
This approach has yielded good results on the Solomon test instances and was
able to solve two large instances with 400 and 1000 customers, respectively.

3.2.2 Column generation algorithms

Column generation is intimately related to constraint generation and can be
seen as a special way of updating the multipliers associated with the relaxed
constraints. Let Ωk denote the set of feasible paths for vehicle k ∈ K. For each
path ω ∈ Ωk , let cωk be the cost of this path and let θk be a binary variable
ω
equal to 1 if and only if vehicle k uses path ω. Let also aiω be the number of
times customer i ∈ N is visited by path ω. As first suggested by Balinski and
Quandt (1964), the VRPTW can be stated as follows:
 
k k
minimize cω θω (25)
k∈K ω∈Ωk
subject to
 
k
aiω θω = 1 i ∈ N (26)
k∈K ω∈Ωk

k
θω = 1 k ∈ K (27)
ω∈Ωk
k
θω ∈ {0 1} k ∈ K ω ∈ Ωk  (28)
Because the sets Ωk are likely to have a very large cardinality, this problem
can be tackled by a branch-and-bound algorithm in which the linear relaxations
are solved by column generation. At each node of the enumeration tree, a re-
stricted column generation master problem is solved over the current set of
columns. New columns of negative reduced cost are generated by solving a re-
source constrained shortest path problem (13)–(19) with modified arc costs
reflecting the current values of the dual variables associated with the con-
straints of the column generation master problem. This process stops when no
negative reduced cost column can be generated. Because the column genera-
tion subproblem is equivalent to the Lagrangian subproblem L(λ), the lower
bound provided by column generation is equal to the value of the Lagrangian
390 J.-F. Cordeau et al.

dual. The dual of the LP relaxation of formulation (25)–(28) is, in fact, equiv-
alent to the Lagrangian dual defined in the previous section. This formulation
can also be obtained by applying the Dantzig–Wolfe decomposition principle
(Dantzig and Wolfe, 1960) to the original formulation (11)–(19).
Branching must be performed at each node of the branch-and-bound tree,
where the optimal solution to the linear relaxation includes fractional path
variables. While it is in principle possible to branch directly on fractional θω
variables, this approach is difficult to implement in practice. Indeed, it is easy
to set such variables equal to 1 but it is much more difficult to impose the op-
posite decision. In the latter case, care must be taken to ensure that the same
path will not be generated more than once by the subproblem. To this purpose,
one could use a modified dynamic programming algorithm to implicitly handle
forbidden paths, or a p-shortest path algorithm where p is equal to the num-
ber of forbidden paths plus one. This would ensure the generation of at least
one valid path of negative reduced cost whenever one exists. A more conve-
nient branching scheme consists of making decisions on the original arc flow
variables xkij or on sums of these variables. For example, binary decisions can
be made on the following sum of variables:
 
xkij 
j∈N  k∈K 

where i ∈ N, N  ⊆ δ+ (i), and K  ⊆ K. Forcing this sum to be equal to 1

requires that some vertex in subset N  be visited immediately after i by some
vehicle. If |N  | = 1, then the corresponding vertex must be visited after i by
some vehicle. If |K  | = 1, then vertex i is implicitly assigned to vehicle k. The
special case |N  | = 1 and |K  | = 1 is equivalent to forcing xkij = 1 for some
given j and k. It is worth pointing out that all such decisions can be handled
directly at the subproblem level through the simple elimination of arcs in the
networks.
Column generation was successfully applied to the VRPTW by Desrochers
et al. (1992) and by Kohl et al. (1999). The latter authors also used valid
inequalities to strengthen the bounds obtained by column generation. More
specifically, let
  
x(S) = xkij
k∈K i∈V \S j∈S

denote the flow into set S ⊆ N and denote by κ(S) the minimum number of
vehicles needed to serve all customers in S. Then the constraint
x(S)  κ(S) (29)
is a valid inequality for the VRPTW and is called a κ-path inequality. Com-
puting κ(S) is a difficult problem which is equivalent to solving the VRPTW
on a subset of vertices with the objective of minimizing the number of vehi-
cles used. Kohl et al. (1999) have, in fact, restricted their attention to the case
 Ch. 6. Vehicle Routing 391

κ = 2. Determining whether κ(S) = 1 for a particular subset S can be achieved

by checking that the capacity of a single vehicle is sufficient and the corre-
sponding TSPTW is feasible. The latter problem is NP-hard but can be solved
relatively quickly by dynamic programming for small instances. The algorithm
of Kohl et al. was capable of solving 70 of the 87 Solomon short-horizon in-
stances to optimality. Cook and Rich (1999) have extended this approach to
the case κ  6 by using parallel computing and replacing the TSPTW feasibility
problem with a VRPTW. They were thus able to solve 80 of the short-horizon
instances. They also solved 30 of the 81 long-horizon instances.
While the constrained elementary shortest path problem is NP-hard, the
relaxation obtained by allowing cycles can be solved by a pseudopolynomial
labeling algorithm (see, e.g., Desrochers and Soumis, 1988). Because of time
windows and capacity constraints, these cycles will nevertheless be of finite
length. This relaxation will of course weaken the value of the lower bound, but
cycle elimination procedures can be used to circumvent this difficulty. A proce-
dure for eliminating 2-cycles (i.e., cycles of the form (i j i)) was first proposed
by Houck et al. (1980). More recently, Irnich and Villeneuve (2003) developed
an efficient approach to forbid cycles of length greater than 2. Experiments
performed by the authors show that k-cycle elimination with k  3 can sub-
stantially improve the lower bounds. Embedding this technique within column
generation enabled the exact solution of 15 previously unsolved instances of
the Solomon benchmark set.
Recently, Chabrier (2006) proposed a modified labeling algorithm to handle
the constrained elementary shortest path problem and thus obtain improved
lower bounds. In this algorithm, both exact and heuristic dominance rules are
considered. Whenever the heuristic approach cannot find a path of negative
reduced cost, the exact but slower implementation is used. This approach has
allowed the author to find the optimal solution to 17 previously unsolved long-
horizon instances from the Solomon benchmark set.
Promising results were also reported by Danna and Le Pape (2003) who de-
veloped a cooperation scheme between column generation and local search
applied to the VRPTW. During the branch-and-price process, local search is
regularly applied from the best known integer solution. This often results in an
improved upper bound that can then be used to prune nodes in the enumer-
ation tree. Furthermore, columns associated with solutions identified during
local search can be fed into the restricted master problem. The branch-and-
price algorithm thus benefits from local search by being provided at an early
stage with high quality upper bounds, resulting in a smaller search tree. In turn,
local search benefits from branch-and-price by working with a variety of differ-
ent initial solutions, resulting in an effective form of diversification.

3.2.3 A branch-and-cut algorithm

A branch-and-cut algorithm for the VRPTW was developed by Bard et al.
(2002). As in most such algorithms for the VRP, the problem is formulated
using two-index variables xij equal to 1 if and only if a vehicle travels directly
392 J.-F. Cordeau et al.

from vertex i to vertex j. The algorithm incorporates five types of inequali-

ties: subtour elimination constraints, capacity constraints, comb inequalities,
incompatible pair inequalities, and incompatible path inequalities. At each
node of the search tree an upper bound is computed by means of the Greedy
Randomized Adaptive Search Procedure (GRASP) described by Kontoravdis
and Bard (1995).
Incompatible pair inequalities rely on the existence of vertex pairs that can-
not belong to the same vehicle route. If i and j denote two incompatible
vertices and P = (i h1      h|P |−2  j) is a path, then the following inequal-
ity is valid:
xi1 h1 + xh1 i + · · · + xh|P |−2 j + xjh|P |−2  |P | − 2 (30)
Incompatible path inequalities are similar to infeasible pair inequalities but
take arc orientations into account. If i and j are two vertices such that i cannot
precede j in a feasible vehicle route then the following inequality is valid for
any path P between i and j:
xih1 + xh1 h2 + · · · + xh|P |−2 j  |P | − 2 (31)
The authors present four separation heuristics to identify violated capacity
constraints. The first is based on the computation of minimum cuts in G. The
second applies a graph shrinking heuristic similar to that proposed by Araque
et al. (1994) for the VRP. The third consists of identifying connected com-
ponents in G that do not contain the depot. Finally, the fourth is a heuristic
proposed by Kohl et al. (1999) to identify violated 2-path inequalities. Heuris-
tic separation algorithms are also described for the identification of violated
comb inequalities, incompatible path inequalities, and incompatible pair in-
equalities. The branch-and-cut algorithm of Bard, Kontoravdis, and Yu has
obtained good results on the Solomon test instances: all 50-customer instances
and several 100-customer instances were solved optimally.

3.3 Heuristics for the VRPTW

Because of the difficulty of the VRPTW and its high practical relevance,
there is a genuine need to develop fast algorithms capable of producing good
quality solutions in short computing times. Heuristics can also be used to pro-
vide upper bounds for the exact algorithms described in the previous section.
This section describes the three main classes of heuristics for the VRPTW:
construction heuristics, improvement heuristics, and metaheuristics.

3.3.1 Construction heuristics

Route construction algorithms work by inserting customers one at a time
into partial routes until a feasible solution is obtained (see Section 2.3.1).
Routes can either be constructed sequentially or in parallel. Construction al-
gorithms are mainly distinguished by the order in which customers are selected
and by the method used to determine where a customer should be inserted.
 Ch. 6. Vehicle Routing 393

Several sequential insertion heuristics for the VRPTW were proposed

by Solomon (1987). Among these heuristics, the most efficient, called I1, con-
sists of first selecting the farthest customer from the depot as a seed customer.
The remaining customers are then inserted one at a time into the current route
by selecting at each iteration the customer that maximizes a saving measure,
taking into account the distance from the depot and the cost of insertion in
the current route. The customer is then inserted in the position minimizing
a weighted combination of extra distance and extra time required to visit the
customer. The process is repeated until all customers have been inserted or it
is no longer possible to insert additional customers without violating either the
capacity or time window constraints. At this point a new route is initialized by
selecting a new seed customer and the process repeats itself until no customers
remain.
A parallel version of this heuristic was later developed by Potvin and
Rousseau (1993) who proposed a generalized regret measure to select the next
customer for insertion. This measure reflects the cost increase likely to result
if a customer is not assigned to the route minimizing the insertion cost. Fur-
ther improvements to the sequential heuristic of Solomon (1987) were also
described by Ioannou et al. (2001) who proposed modifying the criteria for
customer selection and insertion to take into account the impact of the inser-
tion on all routed and unrouted customers.

3.3.2 Improvement heuristics

Improvement heuristics iteratively improve an initial feasible solution by
performing exchanges while maintaining feasibility. The process normally
stops when no further exchange can be made without deteriorating the solu-
tion. Improvement heuristics are mainly characterized by the type of exchanges
considered at each iteration. These define the neighborhood of a solution, i.e.,
the set of solutions reachable from the current solution by performing a single
exchange.
The first improvement heuristics for the VRPTW (see, e.g., Russell, 1977;
Baker and Schaffer, 1986) were adaptations of the 2-opt (Croes, 1958), 3-opt
(Lin, 1965), and Or-opt (Or, 1976) edge exchange mechanisms originally in-
troduced for the TSP. Because of time windows, checking whether a given
exchange maintains feasibility of the solution can be rather time consuming.
Starting with the work of Savelsbergh (1985), several attempts have been made
to develop efficient implementations of neighborhood evaluation procedures
for λ-exchanges (see also Solomon et al., 1988; Savelsbergh, 1990, 1992). A
comparison of 2-opt, 3-opt, and Or-opt exchange heuristics for the VRPTW
was performed by Potvin and Rousseau (1995) who also introduced a new ex-
change, called 2-opt*, a special case of 2-opt that maintains the orientation of
the subroutes involved in the exchange. This is accomplished by removing the
last n1 customers from a route k1 , inserting them after the first n2 customers of
a route k2 , and reconnecting the initial part of route k1 with the terminal part
of route k2 . Another exchange mechanism was described by Thompson and
394 J.-F. Cordeau et al.

Psaraftis (1993) who proposed transferring sets of customers in a cyclic fashion

between routes.
Several attempts have also been made to integrate construction and im-
provement heuristics. Russell (1995) developed a procedure that embeds route
improvement within the solution construction process. More precisely, cus-
tomers can be switched between routes, and routes can be eliminated during
the construction of the solution which is performed by a procedure similar to
that of Potvin and Rousseau (1993). More recently, Cordone and Wolfler Calvo
(2001) have proposed a composite heuristic in which a set of initial solutions is
first constructed by means of Solomon’s I1 insertion heuristic and an improve-
ment procedure is then applied to each of them. This procedure applies 2-opt
and 3-opt exchanges and attempts to reduce the number of routes by relocat-
ing customers. To escape from local optima, the heuristic alternates between
an objective minimizing total distance and an objective minimizing total route
duration (the primary objective being in both cases to minimize the number
of routes). Several deterministic local search heuristics were also proposed by
Bräysy (2002), based on a new three-phase approach. In a first phase, an ini-
tial solution is created with one of two proposed route construction heuristics
(a cheapest insertion-based heuristic with periodic route improvements and a
parallel savings heuristic). The second phase attempts to reduce the number of
routes by applying a local search operator based on ejection chains (see, e.g.,
Glover, 1992). Finally, the third phase applies Or-opt exchanges to reduce the
total length of the routes.

3.3.3 Metaheuristics
Most of the recent research on approximate algorithms for the VRPTW has
concentrated on the development of metaheuristics. Unlike classical improve-
ment methods, metaheuristics usually incorporate mechanisms to continue the
exploration of the search space after a local minimum is encountered.

Tabu search heuristics. Some of the first applications of tabu search to the
VRPTW can be attributed to Semet and Taillard (1993) and to Potvin et
al. (1996) who combined Solomon’s insertion heuristics with improvement
schemes based on vertex and chain exchange procedures.
A more sophisticated algorithm was later developed by Taillard et al. (1997)
for the VRP with soft time windows in which vehicles are allowed to arrive late
at customer locations but time window violations are penalized in the objective
function. This heuristic relies on the concept of adaptive memory introduced
by Rochat and Taillard (1995) and on the decomposition and reconstruction
procedure developed by Taillard (1993) for the classical VRP. An adaptive
memory is a pool of routes extracted from the best solutions found during the
search. This memory is first initialized with routes produced by a randomized
insertion heuristic. At each iteration of the metaheuristic, a solution is con-
structed from the routes belonging to the adaptive memory and is improved
through tabu search. The routes of the resulting solution are then stored in
 Ch. 6. Vehicle Routing 395

the adaptive memory if this solution improves upon the worst solution already
stored. The tabu search heuristic uses an exchange operator, called CROSS ex-
change, which swaps sequences of consecutive customers between two routes.
Individual routes are also optimized by removing two edges from a route and
moving the segment between these two edges to another location within the
route. A parallel computing implementation of this approach is described in
Badeau et al. (1997).
A metaheuristic embedding reactive tabu search (see, e.g., Battiti and
Tecchiolli, 1994) within the parallel construction heuristic of Russell (1995) was
developed by Chiang and Russell (1997). In this implementation, the tabu list
length is increased if identical solutions occur too frequently and is decreased
if no feasible solution can be found. Using a variety of customer ordering
rules and criteria for measuring the best insertion points, the metaheuristic
first constructs six different initial solutions by gradually inserting customers
and repeatedly applying tabu search to the partial solutions. The best solution
obtained after this step is further improved through tabu search. Exchanges
are performed by using some of the λ-interchanges of Osman (1993): switch
a customer from one route to another and swap two customers belonging to
different routes.
More recently, a tabu search heuristic was developed by Cordeau et al.
(2001) for the VRPTW and two of its generalizations: the periodic VRPTW
and the multidepot VRPTW (see also Cordeau et al., 1997). In this heuris-
tic, an initial solution is obtained by means of a modified sweep heuristic.
Infeasible solutions are allowed during the search and violations of capacity,
duration or time window constraints are penalized in the objective function
through dynamically updated penalty factors. At each iteration of the tabu
search, a customer is removed from its current route and inserted into a dif-
ferent route by using a least cost insertion criterion. A continuous diversifica-
tion mechanism that penalizes frequently made exchanges is used to drive the
search process away from local optima. Finally, a post-optimizer based on a
specialized TSPTW heuristic (Gendreau et al., 1998) is applied to individual
routes. An improvement to this heuristic for the handling of route duration
constraints was recently described by Cordeau et al. (2004). The heuristic was
also extended by Cordeau and Laporte (2001) to handle heterogeneous vehi-
cles. Other tabu search algorithms for the VRPTW were proposed by Brandão
(1998), Schulze and Fahle (1999), and Lau et al. (2003).

Genetic algorithms. Homberger and Gehring (1999) have described two evo-
lution strategies for the VRPTW. Both are based on the (μ λ) strategy: start-
ing from a population with μ individuals, subsets of individuals are randomly
selected and recombined to yield a total of λ > μ offspring. Each offspring
is then subjected to a mutation operator, and the μ fittest are selected to
form the new population. In the first method, new individuals are generated
directly through mutations and no recombination takes place. Mutations are
obtained by performing one or several moves from the 2-opt, Or-opt, and
396 J.-F. Cordeau et al.

1-interchange families. In the second method, offspring are generated through

a two-step recombination procedure in which three individuals are involved. In
both methods, the fitness of an individual depends first on the number of vehi-
cles used, and second on the total distance traveled. Gehring and Homberger
(2002) later proposed a two-phase metaheuristic in which the first phase min-
imizes the number of vehicles through an evolution strategy, while the second
one minimizes the total distance through tabu search. A parallelization strat-
egy is also used to run several concurrent searches of the solution space with
differently configured metaheuristics cooperating through the exchange of so-
lutions.
Berger et al. (2003) have developed a genetic algorithm that concurrently
evolves two distinct populations pursuing different objectives under partial
constraint relaxation. The first population aims to minimize the total distance
traveled while the second one focuses on minimizing the violations of the time
window constraints. The maximum number of vehicles imposed in the first
population is equal to kmin whereas the second population is allowed only
kmin − 1 vehicles, where kmin refers to the number of routes in the best known
feasible solution. Whenever a new feasible solution emerges from the second
population, the first population is replaced with the second and the value of
kmin is updated accordingly. Two recombination operators and five mutation
operators are used to evolve the populations. This approach has proved to be
rather efficient in minimizing the number of vehicles used.
More recently, Mester and Bräysy (2005) have developed an iterative meta-
heuristic that combines guided local search and evolution strategies. An initial
solution is first created by an insertion heuristic. This solution is then improved
by the application of a two-stage procedure. The first stage consists of a guided
local search procedure in which 2-opt* and Or-opt exchanges are performed
together with 1-interchanges. This local search is guided by penalizing long
arcs appearing often in local minima. The second stage iteratively removes
a selected set of customers from the current solution and reinserts the re-
moved customers at minimum cost. These two stages are themselves repeated
iteratively until no further improvement can be obtained. Very good results
are reported by the authors on large-scale instances. According to Bräysy and
Gendreau (2005b), the three approaches just described seem to produce the
best results among genetic algorithms. Other such algorithms have also been
proposed by a number of researchers including Potvin and Bengio (1996),
Thangiah and Petrovic (1998), and Tan et al. (2001).

Other metaheuristics. Kontoravdis and Bard (1995) have described a two-

phase GRASP for the VRPTW. A number of routes are first initialized by
selecting seed customers. The remaining customers are then gradually inserted
in the routes by using a randomized least insertion cost procedure. During this
process, periodic attempts are made to improve the routes by local search. In
this phase certain routes may be eliminated by means of a deterministic proce-
dure that attempts to relocate the customers to a different route. To estimate
 Ch. 6. Vehicle Routing 397

the required number of routes, the authors have proposed three lower bounds
for fleet size. Two are based on bin packing structures generated by the capac-
ity or time window constraints. The other is derived from the associated graph
created by pairs of customers having incompatible demands or time windows.
A guided local search algorithm for the VRPTW was introduced by Kilby
et al. (1998). In guided local search, the objective function is augmented with
a penalty term reflecting the proximity of the current solution value to that
of previously encountered local minima. The method is used to drive a local
search heuristic that modifies the current solution by performing one of four
moves: 2-opt exchanges within a route, switching a customer from one route
to another, exchanging customers belonging to two different routes, and swap-
ping the ends of two routes. All customers are first assigned to a virtual vehicle
and the routes for the actual vehicles are left empty. Because a penalty is as-
sociated with not visiting a customer, a feasible solution will be constructed
in the process of minimizing cost. The local search algorithm starts from this
solution and performs a series of exchanges until a local minimum is reached.
The objective function is then modified by adding a term penalizing the pres-
ence of the arcs used in this solution. The search iterates by finding new local
minima and accumulating penalties until a stopping criterion is met. This ap-
proach was later coupled with tabu search and embedded within a constraint
programming framework by De Backer et al. (2000).
Gambardella et al. (1999) have developed an ant colony optimization algo-
rithm for the VRPTW which associates an attractiveness measure to the arcs.
Artificial ants represent parallel processes whole role is to construct feasible
solutions. To deal with the hierarchical objective of first minimizing the num-
ber of vehicles and then minimizing distance, two ant colonies are used, each
dedicated to the optimization of a different objective. These colonies coop-
erate by exchanging information through pheromone updating. Whenever a
feasible solution with a smaller number of vehicles is found, both colonies are
reactivated with the reduced number of vehicles.
Bent and Van Hentenryck (2004) have described a two-stage hybrid algo-
rithm that first minimizes the number of routes by simulated annealing and
then minimizes total distance traveled by using a large neighborhood search
(Shaw, 1998) which may relocate a large number of customers. The first stage
uses a lexicographic evaluation function to minimize the number of routes,
maximize the sum of the squares of the route sizes, and minimize the minimal
delay (a measure of time window tightness) of the solution. The neighborhood
used in this stage consists of 2-opt, Or-opt, relocating, exchange, and crossover
moves. In the second stage, subsets of customers are removed from their cur-
rent route and reinserted in possibly different routes. Customers selected for
removal are randomly chosen but the algorithm favors customers that are
geographically close to each other and belong to different routes. A branch-
and-bound algorithm is then used to reinsert these customers.
A four-phase metaheuristic based on a modification of the variable neigh-
borhood search was described by Bräysy (2003). In the first phase, an initial
398 J.-F. Cordeau et al.

solution is created by using route construction heuristics. During this process,

the partial routes are periodically reoptimized through Or-opt exchanges. In
the second phase, an attempt is made to reduce the number of routes by apply-
ing a route elimination operator based on ejection chains. In the third phase,
four local search procedures embedded within a variable neighborhood search
(see, e.g., Mladenović and Hansen, 1997) are applied to reduce the total dis-
tance traveled. These procedures are based on modifications to the CROSS
exchanges of Taillard et al. (1997) and cheapest insertion heuristics. In the
fourth phase, a modified objective function considering waiting time is used
by the local search operators in the hope of further improving the solution.
More recently, a local search algorithm with restarts was also proposed by
Li and Lim (2003). This algorithm first constructs an initial solution by using
an insertion heuristic. Local search is then performed from this solution us-
ing three exchange operators that move segments of customers either between
routes or within the same route. Whenever a local minimum is reached, mul-
tiple restarts are performed starting from the best known solution, and a tabu
list is used to prevent cycling.
A large number of other metaheuristics based on various paradigms have
been described in recent years. For additional references on approximate al-
gorithms for the VRPTW as well as detailed computational experiments, the
reader is referred to recent surveys by Bräysy and Gendreau (2005a, 2005b).

4 The inventory routing problem

The Inventory Routing Problem (IRP) is an important extension of the VRP

which integrates routing decisions with inventory control. The problem arises
in environments where Vendor Managed Inventory (VMI) resupply policies
are employed. These policies allow a vendor to choose the timing and size of
deliveries. In exchange for this freedom, the vendor agrees to ensure that its
customers do not run out of product. In a more traditional relationship, where
customers call in their orders, large inefficiencies can occur due to the timing of
customers’ orders (resulting in high inventory and distribution costs). Realizing
the cost savings opportunities of vendor managed inventory policies, however,
is not a simple task, particularly with a large number and variety of customers.
The inventory routing problem achieves this goal by determining a distribution
strategy that minimizes long term distribution costs. This description of the in-
ventory routing problem focuses primarily on distribution. Inventory control
is restricted to ensuring that no stockouts occur at the customers. Inventory
control takes a more prominent role when inventory holding costs are consid-
ered. In the inventory control literature, the resulting environment is usually
referred to as a one warehouse multiretailer system.
Inventory routing problems are very different from VRPs. Vehicle routing
problems occur when customers place orders and the vendor, on any given day,
 Ch. 6. Vehicle Routing 399

assigns the orders for that day to routes for vehicles. In IRPs, the delivery com-
pany, not the customer, decides how much to deliver to which customers each
day. There are no customer orders. Instead, the delivery company operates
under the restriction that its customers are not allowed to run out of product.
Another key difference is the planning horizon. Vehicle routing problems typ-
ically deal with a single day, the only requirement being that all orders have to
be delivered by the end of the day. Inventory routing problems are defined on
a longer horizon. Each day the delivery company makes decisions about which
customers to visit and how much to deliver to each of them, while keeping in
mind that decisions made today have an impact on what has to be done in the
future. The objective is to minimize the total cost over the planning horizon
while ensuring that no customer runs out of product.

4.1 Definition of the IRP

The deterministic IRP is concerned with the repeated distribution of a single

product from a single facility, to a set of n customers over a planning horizon
of length T , possibly infinity. Customer i consumes the product at a rate ui
(say volume per day) and can maintain a local inventory of product of up to a
maximum of Ci . The inventory at customer i is Ii0 at time 0. A fleet of m ho-
mogeneous vehicles, with capacity D, is available for the distribution of the
product. If a quantity di is delivered at customer i, the vendor earns a reward
equal to ri di . It takes a vehicle a time tij to traverse arc (i j) of the distri-
bution network and a cost cij is incurred when doing so. The objective is to
maximize the profit (revenue minus cost) over the planning horizon, without
causing stockouts at any of the customers. (Note that because product usage is
assumed to be deterministic and no stockouts are allowed, long run revenues
are fixed and the key is to reduce delivery costs.) A dispatcher has to decide
when to serve a customer, how much to deliver, and which delivery routes to
use to serve customers.
In the stochastic IRP customer demands are defined at discrete time in-
stants t by means of random variables. Let Ut = (U1t      Unt ) denote the
vector of random customer demands at time t. Customer demands on differ-
ent days are independent random vectors with a joint probability distribution F
that does not change with time; that is, U0  U1     is an independent and iden-
tically distributed sequence, and F is the probability distribution of each Ut .
The probability distribution F is known to the decision maker. The vendor
can measure the inventory level Xit of each customer i at any time t. At each
time instant t, the vendor makes a decision that controls the routing of vehicles
and the replenishment of customer inventories. Because demand is uncertain,
there is often a positive probability that a customer will run out of stock, and
thus shortages cannot always be prevented. Shortages result in a penalty pi si
if the unsatisfied demand on day t at customer i is si . Unsatisfied demand is
treated as lost demand. The objective is to construct a distribution policy max-
imizing the expected discounted profit over an infinite time horizon.
400 J.-F. Cordeau et al.

4.2 Motivating example

To illustrate the difficulty of inventory routing problems, we reproduce a

small deterministic example introduced by Fisher et al. (1982) and Bell et al.
(1983). The relevant optimal tour costs can be derived from the network shown
in Figure 1, e.g., the optimal tour costs for visiting customers 1 and 2, denoted
by C12 , is equal to $210. The vehicle capacity is 5000 gallons and customer
tank capacity and usage data, in gallons, are as follows:

Customer i di ui

1 5000 1000
2 3000 3000
3 2000 2000
4 4000 1500

A simple schedule jointly replenishes customers 1 and 2 as well as customers

3 and 4 on a daily basis. This schedule is natural because customers 1 and 2
(3 and 4, respectively) are near each other. Each customer i receives a quantity
equal to its daily consumption ui . The long-run average cost of this schedule
is 420 miles per day. An improved schedule consists of a cycle that repeats
itself every two days. On the first day, one trip replenishes 3000 gallons to cus-
tomer 2 and 2000 gallons to customer 3, at a cost of 340 miles. On the second
day, two trips are made. The first trip replenishes 2000 gallons to customer 1
and 3000 gallons to customer 2. The second trip replenishes 2000 gallons to

Fig. 1. A four-customer example with distances shown on edges.

 Ch. 6. Vehicle Routing 401

customer 3 and 3000 gallons to customer 4. Each trip costs 210 miles. The av-
erage cost of this schedule is 380 miles per day, which is nearly 10% lower than
the first schedule.

4.3 Observations on the IRP

Before describing solution approaches, we present some general obser-

vations concerning inventory routing problems and some common elements
found in most solution approaches.
The IRP is a long-term dynamic control problem which is extremely difficult
to solve. Therefore, most of the available algorithms solve only a short-term
planning problem. In early publications, it was often just a single day but later,
short-term was expanded to a few days. Two key issues need to be resolved
with such approaches: how to model the long-term effect of short-term de-
cisions, and which customers to consider in the short-term planning period.
A short-term approach that only minimizes costs has the tendency to defer
as many deliveries as possible to future planning periods, which may lead to
an undesirable situation in the future. Therefore, a proper incorporation of
the long-term objective into the short-term planning problem is essential. The
long-term effect of short-term decisions needs to capture the costs and benefits
of delivering to a customer earlier than necessary. This usually means deliver-
ing less and may lead to higher future distribution costs, but reduces the risk
of a stockout and may thus reduce future shortage costs. Decisions regard-
ing which customers need to be considered in the short-term planning period
are usually guided by some measure of the urgency to make a delivery to a
customer and the quantity that can be delivered. Usually, it is assumed that
customers considered in the short-term planning period may actually be vis-
ited, but the decision whether or not to actually visit them still has to be made.
When the short-term planning problem consists of a single day, the problem
can be viewed as an extension of the VRP and solution techniques for the
VRP can be adapted. For example, Campbell and Savelsbergh (2004c) have
discussed efficient implementations of insertion heuristics to handle situations
where the delivery amount has to lie between a lower and an upper bound,
as opposed to being fixed. In related work, Campbell and Savelsbergh (2004b)
have studied the problem of determining an optimal delivery schedule for a
route, i.e., given a sequence of customer visits, determine the timing of the
visits so as to maximize the total amount of the product delivered on the route.
Because single day approaches usually base decisions on the latest inventory
measurement and a predicted usage for that day, they avoid the difficulty of
forecasting long-term usage, which makes the problem much simpler.

4.4 Single customer analysis

It is insightful to analyze the “simple” situation in which there is only a single

customer. The results of this type of analysis can be used effectively to guide
402 J.-F. Cordeau et al.

decisions on which customers to consider in a short term planning horizon. The

material presented in this subsection is primarily based on Jaillet et al. (2002),
although much of it dates back to the work of Dror and Ball (1987). We first
consider the deterministic case. For ease of notation, let the usage rate of the
customer be u, the storage capacity of the customer be C, the initial inventory
level be I 0 , the delivery cost to the customer be c, and the vehicle capacity be Q.
It is easy to see that an optimal policy is to fill up the storage space precisely at
the time when it becomes empty. Therefore the cost vT for a planning period
of length T is

T u − I0
vT = max 0 c
min{C Q}
Now consider the stochastic case in which one decides daily whether to make
a delivery to the customer or not. The demand U between consecutive decision
points, i.e., the demand per day, is a random variable with known probabil-
ity distribution and finite mean. Assuming that the storage capacity at each
customer is at least as large as the vehicle capacity and the vendor can only
monitor the inventory in the storage space at the time of a delivery, it can be
shown that for the infinite horizon case, there exists an optimal policy that fills
up the storage space at each delivery and, following any scheduled or stock-
out delivery, plans the next delivery d days after. The optimal replenishment
interval d is a constant chosen to minimize the expected daily cost.
A d-day policy makes a delivery to the customer every d days and de-
livers as much as possible, unless a stockout occurs earlier. In such a case,
the vehicle is sent right away, which generates a cost S. It is assumed that
deliveries are instantaneous, so that no additional stockout penalties are in-
curred. Furthermore, assume that initially the storage space is full. Let pj be
the probability that a stockout first occurs on day j (1  j  d − 1). Then
p = p1 + p2 + · · · + pd−1 is the probability that there is a stockout in period
[1     d − 1]. Furthermore, let vT (d) be the expected total cost of this policy
over a planning period of length T . We now have for d > T
  
vT (d) = pj vT −j (d) + S
1jT

and for d  T
    
vT (d) = pj vT −j (d) + S + (1 − p) vT −d (d) + c 
1jd−1

As a consequence, the expected total cost of filling up a customer’s tank

every d days over a T -day period (T  d) is given by

vT (d) = α(d) + β(d)T + f (T d)

 Ch. 6. Vehicle Routing 403

where α(d) is a constant depending only on d, f (T d) is a function that tends

to zero exponentially fast as T tends to infinity, and
pS + (1 − p)c
β(d) =  
1jd jpj
with pd = 1 − p. The value β(d) is the long-run average cost per day. To
determine the best policy in this class, we need to minimize vT (d) which for
large T means finding a value of d minimizing β(d).

4.5 The two-customer IRP

When more than one customer is served, the problem becomes significantly
harder. Not only is it necessary to decide which customers to visit next, but one
must also determine how to combine them into vehicle tours, and how much to
deliver to each of them. Even if there are only two customers, these decisions
may not be easy. The material in the remainder of this section is primarily
based on Campbell et al. (1998).
If the two customers are visited together, it is intuitively clear that given
the amount delivered at the first customer, it is optimal to deliver as much
as possible at the second one (determined by the remaining amount in the
vehicle, and the remaining capacity at the second customer). Thus the problem
of deciding how much to deliver to each customer involves a single decision.
However, making that decision may not be easy, as the following two-customer
stochastic IRP example shows.
Assume the product is delivered and consumed in discrete units and that
each customer has a storage capacity of 20 units. The daily demands of the cus-
tomers are independent and identically distributed (across customers as well as
across time), with P(U = 0) = 04 and P(U = 10) = 06. The shortage penalty
is s1 = 1000 per unit at customer 1 and s2 = 1005 per unit at customer 2. The
vehicle capacity is 10 units. At the beginning of each day the inventory at the
two customers is measured, and the decision maker determines how much to
deliver to each customer. There are three possible vehicle tours, namely tours
exclusively to customers 1 and 2, of cost 120 each, and a tour to both cus-
tomers 1 and 2, of cost 180. Only one vehicle tour can be completed per day.
This situation can be modeled as an infinite horizon Markov decision process,
with the objective of minimizing the expected total discounted cost. Because of
the small size of the state space, it is possible to compute the optimal expected
value and an optimal policy.
Figure 2 shows the expected value (total discounted cost) as a function of
the amount delivered at customer 1 (and therefore also at customer 2), when
the inventory at each customer is 7, and both customers are to be visited in the
next vehicle tour (which is the optimal decision in the given state). The figure
shows that the objective function is not unimodal, with a local minimum at 3,
and a global minimum at 7. Consequently, deciding just how much to deliver
to each customer may require solving a nonlinear optimization problem with
404 J.-F. Cordeau et al.

Fig. 2. Nonunimodal objective function for determining the optimal delivery quantity.

a nonunimodal objective function. This is a hard problem for which available

search methods may not converge to an optimal solution.

4.6 Literature review on the IRP

Rather than providing a comprehensive review of the IRP literature, we dis-

cuss several research streams representing a variety of solution approaches that
have been proposed and investigated. We encourage the reader to examine the
referenced papers for more elaborate and precise coverage.
A first stream of research uses time-discretized integer programming mod-
els to determine the set of customers to be visited in a short-term planning
horizon as well as the amount of product to deliver to them. In order to accu-
rately reflect costs and time related aspects, the integer linear programs work
with a set of potential delivery routes. Fisher et al. (1982) and Bell et al. (1983)
pioneered this approach when they studied the IRP at Air Products, a pro-
ducer of industrial gases. Their formulation determines the delivery volumes
to customers, the assignment of customers to routes, the assignment of vehi-
cle to routes, and the start times of routes. The core structure of their model
is presented below, where the variable xirtv represents the amount of product
delivered to customer i ∈ N on route r ∈ R starting at time t ∈ T , the variable
yrt is 1 if route r starts at time t and 0 otherwise, Sr the set of customers visited
on route r, pi the value of delivering a unit of product to customer i, Fr the
fixed cost of executing route r, qit a lower bound on the cumulative amount
 Ch. 6. Vehicle Routing 405

delivered to customer i by time t, and qit an upper bound on the cumulative

amount that can be delivered to customer i by time t:
  
maximize pi xirt − Fr yrt
r∈R t∈T i∈Sr
subject to

qit  xirs  qit  i ∈ N t ∈ T
r∈R st

xirt  Qyrt  r ∈ R t ∈ T
i∈Sr
xirt  0 yrt ∈ {0 1}
In the model, the per unit value of a delivery to a customer is used to repre-
sent the effect of decisions on events occurring beyond the planning horizon of
the model. In the short-term planning period considered by the model, there
is considerable discretion in the amount of product to deliver. In the long run
this amount is determined by customer usage. Hence, each unit scheduled for
delivery to a customer within the planning horizon reduces the amount to be
delivered in the future. This is accounted for by setting the unit value to an
estimate of the cost of delivering to a customer at a point in time outside the
planning horizon of the model. Furthermore, rather than explicitly incorpo-
rating customer usage rates into the model, lower and upper bounds on the
cumulative amount to be delivered to each customer in each time period in the
planning horizon are used. It is simple, of course, to convert customer usage
rates into bounds, i.e., qit = max{0 tui −Ii0 } and qit = tui +Ci −Ii0 . Lagrangian
relaxation was a central tool in developing an effective heuristic for solving the
integer program. The size of the integer programs to be solved depends on the
chosen time discretization as well as on the size of set of routes.
Campbell and Savelsbergh (2004a) use an integer linear program with a sim-
ilar structure to determine which customers to visit in the next few days (even
though the integer program covers several weeks) and to suggest quantities and
delivery times to these customers. However, then, in a second phase, they use
modified insertion heuristics to determine the actual delivery routes and quan-
tities. The advantage of such a two-phase approach is that a higher degree of
accuracy (in terms of timing of events) can be provided in the second phase and
other practical details, such as drivers shifts, can be considered. The delivery
quantities and times specified by the solution to the integer program are good
from a long-term perspective; they may need to be modified somewhat to also
be good from a short-term perspective. When constructing the actual delivery
routes, Campbell and Savelsbergh consider delivering more to the customers
than the quantity suggested by the integer program (and slightly altering the
delivery time if needed) since this may result in higher vehicle utilization and
thus higher revenues. As in Bard et al. (1998b) their approach is embedded
into a rolling horizon framework.
406 J.-F. Cordeau et al.

A second stream of research is based on the single customer analysis pre-

sented above. This approach was pioneered by Dror et al. (1985) and Dror and
Ball (1987). The optimal replenishment day ti∗ minimizing the expected total
cost for customer i is used to determine the set of customers considered in a
short-term planning problem for the next t̄ days. If ti∗  t̄, then the customer
will be included and will definitely be visited. A value ct is computed for each
day of the planning period to reflect the expected increase in future cost if the
delivery is made on day t instead of t ∗ . If t ∗  T , i.e., the optimal replenish-
ment day falls outside the short-term planning period, then a future benefit gt
can be computed for making an early delivery to the customer on day t of the
short-term planning period. These computed values reflect the long term ef-
fects of short term decisions. An integer linear program is subsequently solved
to assign customers to a vehicle and a day, or just a day, that minimizes the sum
of these costs plus the transportation costs. (It was shown by Adelman (2004)
that this objective function is in fact equivalent to that used by Fisher et al.
(1982).) The delivery amount to a customer on a specific day is fixed and set to
the quantity needed to fill up the storage tank on that day. This leaves either
TSPs or VRPs to be solved in the second stage. These ideas are extended and
improved in Trudeau and Dror (1992). The most recent work along these lines
is that of Bard et al. (1998a, 1998b) who work with a rolling horizon approach
in which a short term planning problem is defined for a two-week period, but
only the decisions for the first week are implemented. In addition, satellite fa-
cilities are considered, i.e., locations other than the depot where vehicles can
be refilled.
A third stream of research focuses on the asymptotic analysis of delivery
policies. Anily and Federgruen (1990, 1991, 1993) analyze fixed partition poli-
cies for the IRP with an unlimited number of vehicles. Customers within the
same partition are divided into regions so as to make the demand of each re-
gion roughly equal to a vehicle load. A customer may appear in more than one
region, but then a certain percent of his demand is allocated to each region.
When one customer in a region is also visited, all other customers in that re-
gion are also visited. The authors determine lower and upper bounds on the
minimum long-run average cost over all fixed partition policies, and propose a
heuristic, called modified circular regional partitioning, to choose a fixed parti-
tion. Using similar ideas, Gallego and Simchi-Levi (1990) evaluate the long-run
effectiveness of direct deliveries (one customer on each route). Direct ship-
ping is shown to be at least 94% effective over all inventory routing strategies
whenever the minimal economic lot size is at least 71% of vehicle capacity.
This indicates that direct shipping becomes an undesirable and costly policy
when many customers require significantly less than a vehicle load, making
more complicated routing policies the appropriate choice. Another adaptation
of these ideas can be found in Bramel and Simchi-Levi (1995) who consider
a variant of the IRP in which customers can hold an unlimited amount of in-
ventory. To obtain a solution, they transform the problem into a Capacitated
Concentrator Location Problem (CCLP), solve it, transform the solution back
 Ch. 6. Vehicle Routing 407

into a solution to the IRP, and heuristically improve it. The CCLP solution will
partition the customers into disjoint sets, which in the IRP will become the
fixed partitions. Chan et al. (1998) analyze zero-inventory ordering policies, in
which a customer’s inventory is replenished only when it has been depleted,
in combination with fixed partitioning routing policies and derive asymptotic
worst-case bounds on their performance. Gaur and Fisher (2004) consider an
IRP with time varying demand. They propose a randomized heuristic to find a
fixed partition policy with periodic deliveries. Their method was implemented
for a supermarket chain.
The fourth stream of research is based on formulating the stochastic IRP
as a Markov decision process and thus explicitly incorporating demand un-
certainty. This approach was pioneered by Minkoff (1993) who proposed a
decomposition heuristic to overcome the computational difficulties caused by
large state spaces. The heuristic solves a linear program to allocate joint trans-
portation costs to individual customers and then solves individual customer
subproblems. The value functions of the subproblems are added to approxi-
mate the value function of the original problem. The main limitation of the
proposed approach is that it assumes the availability of a set of delivery routes
with fixed delivery quantities for the customers on a route and the dispatcher
only has to decide which of the delivery routes to use at each decision point.
This limitation is removed in the work of Kleywegt et al. (2002, 2004) on ap-
proximate dynamic programming approaches and in that of Adelman (2003a,
2004) on price-directed approaches. Let state x = (x1  x2      xn ) represent
the current inventory at each customer, and let A(x) denote the set of all feasi-
ble decisions when the process is in state x. A decision a ∈ A(x) specifies which
customer inventories to replenish, how much to deliver at each customer loca-
tion, and how to combine customers into vehicle routes. Let Q be the Markov
transition function according to which transitions occur. Let g(x a) denote
the expected single stage net reward if the process is in state x at time t and
decision a ∈ A(x) is implemented. The objective is to maximize the expected
total discounted value over an infinite horizon. Let V ∗ (x) denote the optimal
expected value given that the initial state is x. Then, for any state x,

V ∗ (x) = sup g(x a) + α V ∗ (y)Q[dy|x a]  (32)
a∈A(x)
∗
A policy π∗ is called optimal if V π = V ∗ , where V π represents the value
function of policy π. Solving a Markov decision process involves computing the
optimal value function V ∗ and an optimal policy π ∗ by solving the optimality
equation (32). This requires performing the following major computational
tasks:
(1) The computation of the optimal value function V ∗ . Most algorithms for
computing V ∗ involve the computation of successive approximations to
V ∗ (x) for every state x. These algorithms are practical only if the state
space is small.
408 J.-F. Cordeau et al.

(2) The estimation of the expected value (the integral in (32)). For the sto-
chastic IRP, this is a high dimensional integral. Conventional numerical
integration methods are not practical for the computation of such high-
dimensional integrals.
(3) The maximization problem on the right-hand side of (32) has to be
solved to determine the optimal decision for each state. For the sto-
chastic IRP, this means solving a complex variant of the VRP.
Kleywegt, Nori, and Savelsbergh develop approximation methods to effi-
ciently perform these computational tasks. Furthermore, their approach has
the ability to handle a finite fleet of vehicles, whereas in other Markov decision
process based approaches it is assumed that there exists an infinite fleet of ve-
hicles. The optimal value function V ∗ is approximated by V as follows. First,
the stochastic IRP is decomposed into subproblems defined for specific subsets
of customers. Each subproblem is also a Markov decision process. The subsets
of customers do not necessarily partition the set of customers, but must cover
it. The idea is to define each subproblem so that it provides an accurate rep-
resentation of the overall process as experienced by the subset of customers.
To do so, the parameters of each subproblem are determined by simulating
the overall stochastic IRP process, and by constructing simulation estimates
of subproblem parameters. Next, each subproblem is solved optimally. Finally,
for any given state x, the approximate value V (x) is determined by choosing a
partition of the customers and by setting V (x) equal to the sum of the optimal
value functions of the subproblems corresponding to the partition at states
corresponding to x. The partition is chosen to maximize V (x). Randomized
methods, incorporating variance reduction techniques to limit the required
sample size, are used to estimate the expected value on the right-hand side
of (32). Action determination involves deciding which customers to visit on a
route and how much to deliver to them. This is achieved through a heuristic.
An initial solution consisting of only direct delivery routes is constructed. This
is followed by a local search procedure that examines the benefit of adding
a customer to an existing route and modifying the delivery quantities. Using
their approach Kleywegt, Nori, and Savelsbergh can solve problems involving
up to 50 customers.
More recently, Adelman (2003a, 2004) proposed a price-directed operating
policy based on a simple economic mechanism to determine routing and de-
livery decisions for a given inventory state. Suppose management specifies a
value Vi for replenishing one unit of product at customer i. A dispatcher can
now evaluate a feasible delivery route as follows. If a set S = {s1      sn } of cus-
tomers is visited, quantities d1      dn are
 delivered, and a cost cS is incurred.
Then the net value of the route equals i∈S Vi di − cS . The dispatcher has to
choose delivery routes so as to maximize his total net value without stockouts at
customers. This mechanism motivates the dispatcher to replenish a customer i
whose current inventory level is low, because then di can be set large. When
faced with the option of expanding the set S of customers to visit on a route
 Ch. 6. Vehicle Routing 409

which does not yet use the full vehicle capacity, the dispatcher will consider
the incremental cost cS∪{k} − cS and determine if a quantity dk can be replen-
ished that is large enough to justify it, i.e., whether dk Vk − (cS∪{k} − cS ) > 0 or
dk  (cS∪{k} − CS )/Vk .
The key to success in solving management’s problem is to set the Vi ’s in
such a way that the dispatcher is motivated to (ideally) minimize the long-run
time average replenishment costs. If the dispatcher’s total net value is regularly
positive, then his performance exceeds management’s long range expectations.
Management should decrease the Vi ’s to make them consistent with actual per-
formance. On the other hand, if the dispatcher’s total net value is regularly
negative, then the Vi ’s impose unrealistic expectations on the dispatcher and
management should increase them. Ideally, management should set the Vi ’s
equal to the lowest achievable marginal costs.
Starting from a dynamic control model of the inventory routing problem,
Adelman (2003b) derives the following nonlinear programming relaxation,
which computes a long run “average” solution to the inventory routing prob-
lem. Let zR be a decision variable representing the rate at which a subset R
of customers is visited together. Furthermore, let diR for all i ∈ R be a de-
cision variable representing the average quantity delivered to customer i on a
delivery route visiting subset R. This yields the following formulation:

(NLP) minimize CR zR (33)
R⊆N
subject to

diR zR = ui  i ∈ N (34)
R⊆N

diR  Q R ⊆ N (35)
i∈R
diR  Ci  R ⊆ N i ∈ R (36)
zR  diR  0 R ⊆ N i ∈ R (37)
The objective (33) minimizes the long run average replenishment cost. Con-
straints (34) state that for each customer i the rate at which quantities are
replenished must equal the rate at which they are consumed. Constraints (35)
state that on average vehicle capacity is satisfied, and constraints (36) state
that on average the quantity delivered at customer i is less than the storage
capacity. Consider the following linear program

(D) maximize ui Vi (38)
i∈N
subject to

diR Vi  CR  R ⊆ N (39)
i∈R
410 J.-F. Cordeau et al.

with decision variables Vi . Adelman shows that this semi-infinite linear pro-
gram is dual to the nonlinear program in that there is no duality gap between
them and a version of complementary slackness holds. In (NLP) diR is a de-
cision variable while in (D) it is part of the input. The decision variables Vi at
optimality are the marginal costs associated with satisfying constraints (34) of
(NLP). This means that at optimality ui Vi is the total allocated cost rate for
replenishing customer i in an optimal solution to (NLP). Each Vi can be inter-
preted as the payment management transfers to the dispatcher for replenishing
one unit of product of customer i. Hence, the objective (38) maximizes the to-
tal transfer rate, subject to the constraint (39) that the payments can be no
larger than the cost of any replenishment. NLP can be solved effectively by
means of column generation techniques.
We have opted to focus on only a few research streams with an emphasis on
more recent efforts. However, many other researchers have contributed to the
inventory routing literature, including Federgruen and Zipkin (1984), Golden
et al. (1984), Burns et al. (1985), Larson (1988), Chien et al. (1989), Webb and
Larson (1995), Barnes-Schuster and Bassok (1997), Herer and Roundy (1997),
Viswanathan and Mathur (1997), Christiansen and Nygreen (1998a, 1998b),
Christiansen (1999), Reimann et al. (1999), Waller et al. (1999), Çetinkaya and
Lee (2000), Lau et al. (2002), Bertazzi et al. (2002), Savelsbergh and Song
(2005), and Song and Savelsbergh (2005).

5 Stochastic vehicle routing problems

Stochastic Vehicle Routing Problems (SVRPs) are extensions of the deter-

ministic VRP in which some components are random. The three most common
cases are:
(1) stochastic customers: customer i is present with probability pi and ab-
sent with probability 1 − pi ;
(2) stochastic demands (to be collected, say): the demand ξi of customer i
is a random variable;
(3) stochastic times: the service time si of customer i and the travel time tij
of edge (i j) are random variables.
Because some of the data are random it is no longer required to satisfy the
constraints for all realizations of the random variables, and new feasibility and
optimality concepts are required. With respect to their deterministic counter-
parts, SVRPs are considerably more difficult to solve. Not only is the notion of
a solution different, but some of the properties that were valid in a determin-
istic context no longer hold in the stochastic case (see, e.g., Dror et al., 1989;
Gendreau et al., 1996).
Applications of SVRP arise in a number of settings such as the delivery
of meals on wheels (Bartholdi et al., 1983) or of home heating oil (Dror
et al., 1985), sludge disposal (Larson, 1988), forklift routing in warehouses
 Ch. 6. Vehicle Routing 411

(Bertsimas, 1992), money collection in bank branches (Lambert et al., 1993),

and general pickup and delivery operations (Hvattum et al., 2006).
Stochastic VRPs can be formulated and solved in the context of stochastic
programming: a first stage or a priori solution is computed, the realizations
of the random variables are then disclosed and, in a second stage, a recourse
or corrective action is applied to the first stage solution. The recourse action
usually generates a cost or a saving which may be taken into account when de-
signing the first stage solution. To illustrate, consider a planned vehicle route in
an SVRP with stochastic demands. Because demands are stochastic, the vehi-
cle capacity may be attained or exceeded at some customer j before the route is
completed. In this case several possible recourse policies are possible. For ex-
ample, the vehicle could return to the depot to unload and resume collections
at customer j (if the vehicle capacity was exceeded at j) or at the successor
of j on the route (if the vehicle capacity was attained exactly at j). Another
policy would be to plan preventive return trips to the depot in the hope of
avoiding higher costs at a later stage (see, e.g., Laporte and Louveaux, 1990;
Dror et al., 1993; Yang et al., 2000). A more radical policy would be to re-
optimize the route segment following j upon arrival at the depot (see, e.g.,
Bastian and Rinnooy Kan, 1992; Secomandi, 1998; Haughton, 1998, 2000). The
best choice of a recourse policy depends on the time at which information be-
comes available. For example, information about a customer demand may only
be available upon arriving at that customer or when visiting the previous cus-
tomer, thus allowing for a wider range of recourse actions, such as returning
to the depot in anticipation of failure or postponing the visit of a high demand
customer. An extensive discussion of recourse policies in the context of avail-
ability of information is provided in Dror et al. (1989).
There exist two main solution concepts in stochastic programming. In
Chance Constrained Programming (CCP) the first stage problem is solved un-
der the condition that the constraints are satisfied with some probability. For
example, one could impose a failure threshold α, i.e., planned vehicle routes
should fail with probability at most equal to α. The cost of failure is typically
disregarded in this approach. Stewart and Golden (1983) have proposed the
first CCP formulation for the VRP with stochastic demands. Using a three-
index model they showed that probabilistic constraints could be transformed
into a deterministic equivalent form. Laporte et al. (1989) later proposed a
similar transformation for a two-index model. The interest of such transfor-
mations is that the chance constrained SVRP can then be solved using any of
the algorithms available for the deterministic case. In Stochastic Programming
with Recourse (SPR) two sets of variables are used: first-stage variables char-
acterize the solution generated before the realization of the random variables,
while second-stage variables define the recourse action. The solution cost is de-
fined as the sum of the cost of the first-stage solution and that of the recourse
action. The aim of SPR is to design a first-stage solution of least expected total
cost.
412 J.-F. Cordeau et al.

Stochastic VRPs are usually modeled and solved with the framework of a
priori optimization (Bertsimas et al., 1990) or as Markov decision processes
(Dror et al., 1989). A priori optimization computes a first-stage solution of
least expected cost under a given recourse policy. The most favored a priori
optimization methodology is the integer L-shaped method (Laporte and Lou-
veaux 1993, 1998) which belongs to the same class as Benders decomposition
(Benders, 1962) and the L-shaped method for continuous stochastic program-
ming (Van Slyke and Wets, 1969). While route reoptimization is preferable to
a priori optimization from a solution cost point of view, it is computationally
more cumbersome. In contrast, a priori optimization entails solving only one
instance of an NP-hard problem and produces a more stable and predictable
solution (Bertsimas et al., 1990). It is also superior to solving a deterministic
VRP instance with expected demands (Louveaux, 1998).
The integer L-shaped method is essentially a variant of branch-and-cut. It
operates on a current problem obtained by relaxing integrality requirements
and subtour elimination constraints, and by replacing the cost of recourse Q(x)
of first-stage solution x by a lower bound θ on its value. Integrality and subtour
elimination constraints are gradually satisfied as is commonly done in branch-
and-cut algorithms for the deterministic VRP (see, e.g., Naddef and Rinaldi,
2002) while lower bounding functionals on θ, called optimality cuts, are intro-
duced into the problem at integer or fractional solutions. The method assumes
that a lower bound L on θ is available. In the following description xij is a
binary variable equal to 1 if and only if edge (i j) is used in the first stage
solution.
Step 0. Set the iteration count ν := 0 and introduce the bounding constraint
θ  L into the current problem. Set the value z̄ of the best known solution
equal to ∞. At this stage, the only active node corresponds to the initial
current problem.
Step 1. Select a pendent node from the list. If none exists stop.
Step 2. Set ν := ν+1 and solve the current problem. Let (xν  θν ) be an optimal
solution.
Step 3. Check for any subtour elimination constraint violation. If at least one
violation can be identified, introduce a suitable number of subtour elimina-
tion constraints into the current problem, and return to Step 2. Otherwise,
if cxν + θν  z̄, fathom the current node and return to Step 1.
Step 4. If the solution is not integer, branch on a fractional variable. Append
the corresponding subproblems to the list of pendent nodes and return to
Step 1.
Step 5. Compute Q(xν ) and set z ν := cxν + Q(xν ). If z ν < z̄, set z̄ := z ν .
Step 6. If θν  Q(xν ), then fathom the current node and return to Step 1.
Otherwise, impose the optimality cut
   
  ν  ν
θL+ Q x −L xij − xij + 1 (40)
1<i<jxνij =1 1<i<j
 Ch. 6. Vehicle Routing 413

into the current problem and return to Step 2.

The optimality cut (40) uses the fact that a feasible solution is fully char-
acterized by the xij variables associated with edges nonincident to the depot.
They state that either the current solution must be maintained, in which case
the cut becomes θ  Q(xν ), or a new solution must be identified, in which case
the cut becomes θ  L or less and is thus redundant.
Markov decision models are defined on a state space. The system is ob-
served at various transition times corresponding to moments at which a new
customer is visited, and new decisions are taken at these moments. The state
of the system at a given transition time is described by the set of customers
already visited by the vehicle and by its current load. Because the state space is
typically very large, this approach can only be applied to relatively small scale
instances.
Heuristics for SVRPs are adaptations of methods originally designed for
the deterministic case, which can be rather intricate because of the probability
computations involved. In particular, computing the expected cost of a vehicle
route is itself complicated and it may be advisable to use approximations if such
computations are to be performed repeatedly within a search process (see, e.g.,
Gendreau et al., 1996). In what follows we study some particular classes of
SVRPs.

5.1 The vehicle routing problem with stochastic customers

In vehicle routing problems with stochastic customers each vertex i is

present with probability pi . A first-stage solution consists of a set of vehicle
routes visiting the depot and each customer exactly once. The set of absent
customers is then revealed and the second-stage solution consists of following
the first-stage routes while skipping the absent vertices. Jaillet (1985) laid the
foundations of this line of research in his study of the Traveling Salesman Prob-
lem with Stochastic Customers (TSPSC). He proposed mathematical models
and bounds, and he investigated a number of properties of the problem. For
example, he showed that the solution of a deterministic TSP can be arbitrarily
bad for the TSPSC. Also, even if the TSPSC is defined in a plane with Euclid-
ean distances, an optimal cycle may cross itself, contrary to what happens for
the TSP (Flood, 1956). Jézéquel (1985) and Rossi and Gavioli (1987) have
proposed a number of heuristics for the TSPSC based on adaptations of the
Clarke and Wright (1964) savings principle. Bertsimas (1988) and Bertsimas
and Howell (1993) later investigated further properties of the TSPSC and pro-
posed new heuristics, namely methods based on space filling curves (Bartholdi
and Platzman, 1982) and on a 2-opt edge interchange mechanism. The first
exact algorithm for the TSPSC is an integer L-shaped algorithm developed by
Laporte et al. (1994) and capable of solving instances involving up to 50 cus-
tomers. An extension of the TSPSC, called the Pickup and Delivery Traveling
Salesman Problem with Stochastic Customers (PDTSPSC), was recently in-
vestigated by Beraldi et al. (2005). In this problem there are n requests, each
414 J.-F. Cordeau et al.

consisting of a pickup location and of a delivery location, but request i only ma-
terializes with probability pi . The authors show how to efficiently implement a
low complexity interchange heuristic for this problem.
The Vehicle Routing Problem with Stochastic Customers (VRPSC) has
been mostly studied in the context of unit demand customers. As in the TSPSC,
vehicles follow the first-stage routes while skipping the absent customers and
return to the depot to unload when their capacity is reached. This problem was
first studied by Jézéquel (1985), Jaillet (1987), and Jaillet and Odoni (1988).
The latter reference states two interesting properties of the VRPSC:
(1) even if travel costs are symmetric the overall solution cost is dependent
on the direction of travel;
(2) larger vehicle capacities may yield larger solution costs.
Bertsimas’ PhD thesis (Bertsimas, 1988) is an excellent source of informa-
tion on this problem. It describes several properties, bounds and heuristics.
Waters (1989) has studied the case of general integer demands and has com-
pared three simple heuristics for this problem.

5.2 The vehicle routing problem with stochastic demands

The Vehicle Routing Problem with Stochastic Demands (VRPSD) has been
the most studied stochastic VRP. In this problem customer demands are ran-
dom and usually (but not always) independent. Tillman (1969) was probably
the first to study this problem in a multidepot context. He proposed a savings
based heuristic for its solution. The first, major study of the VRPSD can be
attributed to Golden and Stewart (1978) who presented a chance constrained
model and two recourse models. In the first of these a penalty proportional
to vehicle overcapacity is imposed; in the second, the penalty is proportional
to the expected demand in excess of the vehicle capacity. Several basic heuris-
tics were implemented and tested. Dror and Trudeau (1986) developed further
heuristics and showed that for this problem expected travel cost depends on the
direction of travel even in the symmetric case. Again, Bertsimas’ thesis (1988)
constitutes a major contribution to the study of the VRPSD. It proposes several
bounds, asymptotic results and properties for the case where ξi is equal to 1
with probability pi , and equal to 0 otherwise. In their survey paper, Dror et al.
(1989) have shown that some properties established by Jaillet (1985, 1988) and
Jaillet and Odoni (1988) extend to the VRPSC, namely (1) in an optimal solu-
tion a vehicle route may intersect itself; (2) in a Euclidean problem customers
are not necessarily visited in the order in which they appear on the convex
hull of vertices; (3) segments of an optimal route are not necessarily optimal
when considered separately. The latter property can have a major impact on
the design of a dynamic algorithm for the VRPSD.
Laporte et al. (1989) proposed a two-index chance constrained model for
the VRPSC as well as an associated branch-and-cut algorithm capable of solv-
ing instances with n  30. They also introduced a bounded penalty model in
 Ch. 6. Vehicle Routing 415

which the cost of recourse associated with a given route cannot exceed a pre-
set proportion of the first-stage route cost. The best exact solution approach
for the VRPSD is again the integer L-shaped algorithm. Séguin (1994) and
Gendreau et al. (1995) proposed the first implementation of this method for
the solution of the VRPSD and were able to solve instances of up to 70 vertices.
The most difficult case arises when the expected filling rate f of the vehicles
is large. For example, when f = 03 instances with n = 70 can be solved
optimally, but when f = 10 instances with n = 10 can rarely be solved. Us-
ing a similar approach, Hjorring and Holt (1999) solved one-vehicle instances
(m = 1) with 095  f  105 and n = 90. Laporte et al. (2002) imposed an
additional restriction, namely that the expected demand of a route does not
exceed the vehicle capacity, and they also exploited properties of the demand
under known distributions (Poisson and normal) in the generation of lower
bounding functionals on the cost of recourse. This enabled them to solve larger
instances: for Poisson demands they solved instances with f = 09, m = 4, and
n = 25, or with m = 2 and n = 100; for normal demands they solved instances
with f = 09, m = 3, and n = 50.
Dynamic programming was applied by Secomandi (1998) to the VRP with
stochastic demands. The largest instance solved to optimality with this method
contained only 10 customers. The author also developed Neuro-Dynamic Pro-
gramming (NDP) algorithms (Secomandi, 1998, 2000, 2003) for the same prob-
lem. Neuro-dynamic programming (see, e.g., Bertsekas, 1995) is a heuristic
approach used to solve large-scale dynamic programs. It replaces the “cost-
to-go” computations by proxies based on simulation and parametric function
approximations. Secomandi (2000) compared two NDP implementations for
the VRP with stochastic demands: an optimistic approximate iteration policy
in which a neural network methodology is used to compute the approxima-
tions, and a rollout policy in which the cost-to-go is approximated by means of
a heuristic. Computational results show that the second of these two policies is
consistently and substantially superior to the first.

5.3 The vehicle routing problem with stochastic customers and demands

The VRP with stochastic customers and demands combines two difficult
cases. This problem was first mentioned by Jézéquel (1985), Jaillet (1987),
Jaillet and Odoni (1988), and was later formally defined by Bertsimas (1992).
A first-stage solution visiting all customers is first constructed, the set of
present customers is then revealed and their demand becomes known upon
the arrival of the vehicle at the customer’s location, routes are followed as
planned but absent customers are skipped and the vehicle returns to the de-
pot to unload whenever its capacity becomes attained. Benton and Rossetti
(1992) proposed an algorithm which performs route reoptimizations whenever
demands are revealed. One major difficulty in solving this problem lies in the
computation of the objective function value. Recursions, bounds, asymptotic
results, and a comparison of various reoptimization policies are provided by
416 J.-F. Cordeau et al.

Bertsimas (1992). Séguin (1994) and Gendreau et al. (1995) developed the
first exact algorithm for this problem, based again on the integer L-shaped ap-
proach. They solved instances involving up to 46 customers and concluded that
stochastic customers are a far more complicating factor than are stochastic de-
mands. In a different study, Gendreau et al. (1996) developed a tabu search
algorithm which uses an approximation of the objective function cost in order
to ease computations. On a set of 825 instances with 6  n  46 for which the
optimum was known, an optimal solution was identified in 89.45% of the cases
and the average optimality gap was 0.38%.

5.4 The vehicle routing problem with stochastic travel times

In the Vehicle Routing Problem with Stochastic Travel Times (VRPSTT)

travel times on the edges and service times at the vertices are random vari-
ables. Vehicles follow their planned routes and may incur a penalty if the route
duration exceeds a given deadline. It is natural to make this penalty propor-
tional to the elapsed route duration in excess of the deadline (Laporte et al.,
1992). Another possibility is to define a penalty proportional to the uncollected
demand within the time limit, as is the case in a money collection application
studied by Lambert et al. (1993).
The simplest case of the VRPSTT is the Traveling Salesman Problem with
Stochastic Travel Time (TSPSTT) in which there is only one vehicle. It was
first studied by Leipälä (1978) who computed the expected length of tours with
random travel times. A common version of the TSPSTT is the case where the
objective is to design a tour having the largest probability of being completed
within the deadline. Kao (1978) proposed two heuristics for this problem:
one based on dynamic programming and the other on implicit enumeration.
Sniedovich (1981) has shown that dynamic programming applied to the same
problem can be suboptimal because the monotonicity property required for
this method is not satisfied in the TSPSTT. Carraway et al. (1989) later devel-
oped a so-called generalized dynamic programming algorithm that overcomes
this difficulty. Kenyon and Morton (2003) have shown that an optimal TSPSTT
can be identified by solving a deterministic TSP in which the travel and service
times are replaced by their mean values. Verweij et al. (2003) have developed a
heuristic for the case where a penalty proportional to route duration in excess
of the deadline is incurred. The method uses a sample average approximation
technique in which a sample of instance realizations is drawn and each is solved
optimally by means of a deterministic technique. By repeating the method with
different samples a statistical estimate of the optimality gap can be computed.
Laporte et al. (1992) were probably the first to provide exact algorithms for
the VRPSTT. They formulated the chance constrained version of the problem,
and they modeled a recourse version of the problem in which a penalty propor-
tional to route duration in excess of the deadline is incurred. The problem was
solved optimally by means of an integer L-shaped algorithm for 10  n  20
 Ch. 6. Vehicle Routing 417

and two to five travel time scenarios (each scenario corresponds to a differ-
ent travel speed for the entire network). In a more recent study, Kenyon and
Morton (2003) have investigated properties of VRPSTT solutions and have
developed bounds on the objective function value. They have developed a
heuristic that combines branch-and-cut and Monte Carlo simulation which, if
run to completion, terminates with a solution value within a preset percentage
of the optimum.
Finally, vehicle routing with stochastic travel time is frequently encountered
in pickup and delivery problems such as those arising in truckload operations.
Wang and Regan (2001) have proposed models for this class of problems under
the presence of time windows.

Acknowledgements

This work has been supported by the Canadian Natural Sciences and En-
gineering Research Council under Grants 227837-00 and OGP0039682, by
the Ministero dell’Università e della Ricerca (MIUR), and by the Consiglio
Nazionale delle Ricerche (CNR), Italy. This support is gratefully acknowl-
edged.

References

Achuthan, N.R., Caccetta, L., Hill, S.P. (1996). A new subtour elimination constraint for the vehicle
routing problem. European Journal of Operational Research 91, 573–586.
Achuthan, N.R., Caccetta, L., Hill, S.P. (2003). An improved branch and cut algorithm for the capaci-
tated vehicle routing problem. Transportation Science 37, 153–169.
Adelman, D. (2003a). Price-directed replenishment of subsets: Methodology and its application to in-
ventory routing. Manufacturing & Service Operations Management 5, 348–371.
Adelman, D. (2003b). Internal transfer pricing for a decentralized operation with a shared supplier.
Working paper, Graduate School of Business, The University of Chicago.
Adelman, D. (2004). A price-directed approach to stochastic inventory/routing. Operations Research 52,
499–514.
Agarwal, Y., Mathur, K., Salkin, H.M. (1989). A set-partitioning-based exact algorithm for the vehicle
routing problem. Networks 19, 731–749.
Altinkemer, K., Gavish, B. (1991). Parallel savings based heuristic for the delivery problem. Operations
Research 39, 456–469.
Anily, S., Federgruen, A. (1990). One warehouse multiple retailer systems with vehicle routing costs.
Management Science 36, 92–114.
Anily, S., Federgruen, A. (1991). Rejoinder to “One warehouse multiple retailer systems with vehicle
routing costs”. Management Science 37, 1497–1499.
Anily, S., Federgruen, A. (1993). Two-echelon distribution systems with vehicle routing costs and central
inventories. Operations Research 41, 37–47.
Araque, J.R., Hall, L., Magnanti, T.L. (1990). Capacitated trees, capacitated routing and associated
polyhedra. Discussion Paper 90-61, CORE, University of Louvain-la-Neuve, Belgium.
Araque, J.R., Kudva, G., Morin, T., Pekny, J.F. (1994). A branch-and-cut algorithm for vehicle routing
problems. Annals of Operations Research 50, 37–59.
418 J.-F. Cordeau et al.

Augerat, P., Belenguer, J.M., Benavent, E., Corberán, A., Naddef, D., Rinaldi, G. (1995). Computa-
tional results with a branch and cut code for the capacitated vehicle routing problem. Technical
Report RR 949-M, Université Joseph Fourier, Grenoble.
Augerat, P., Belenguer, J.M., Benavent, E., Corberán, A., Naddef, D. (1999). Separating capacity in-
equalities in the CVRP using tabu search. European Journal of Operational Research 106, 546–557.
Badeau, P., Gendreau, M., Guertin, F., Potvin, J.-Y., Taillard, É.D. (1997). A parallel tabu search heuris-
tic for the vehicle routing problem with time windows. Transportation Research C 5, 109–122.
Baker, E., Schaffer, J. (1986). Computational experience with branch exchange heuristics for vehicle
routing problems with time window constraints. American Journal of Mathematical and Management
Sciences 6, 261–300.
Baldacci, R., Hadjiconstantinou, E., Mingozzi, A. (2004). An exact algorithm for the capacitated vehicle
routing problem based on a two-commodity network flow formulation. Operations Research 52, 723–
738.
Baldacci, R., Bodin, L., Mingozzi, A. (2006). The multiple disposal facilities and multiple inventory
locations rollon–rolloff vehicle routing problem. Computers & Operations Research 33, 2667–2702.
Balinski, M., Quandt, R. (1964). On an integer program for a delivery problem. Operations Research 12,
300–304.
Bard, J.F., Huang, L., Dror, M., Jaillet, P. (1998a). A branch and cut algorithm for the VRP with satellite
facilities. IIE Transactions 30, 831–834.
Bard, J.F., Huang, L., Jaillet, P., Dror, M. (1998b). A decomposition approach to the inventory routing
problem with satellite facilities. Transportation Science 32, 189–203.
Bard, J.F., Kontoravdis, G., Yu, G. (2002). A branch-and-cut procedure for the vehicle routing problem
with time windows. Transportation Science 36, 250–269.
Barnes-Schuster, D., Bassok, Y. (1997). Direct shipping and the dynamic single-depot/multi-retailer
inventory system. European Journal of Operational Research 101, 509–518.
Bartholdi, J.J., Platzman, L.K. (1982). An N log N planar traveling salesman heuristic based on space-
filling curves. Operations Research Letters 1, 121–125.
Bartholdi, J.J., Platzman, L.K., Collins, R.L., Warden, W.H. (1983). A minimal technology routing sys-
tem for meals on wheels. Interfaces 13 (3), 1–8.
Bastian, C., Rinnooy Kan, A.H.G. (1992). The stochastic vehicle routing problem revisited. European
Journal of Operational Research 56, 407–412.
Battiti, R., Tecchiolli, G. (1994). The reactive tabu search. ORSA Journal on Computing 6, 126–140.
Bean, J.C. (1994). Genetic algorithms and random keys for the sequencing and optimization. ORSA
Journal on Computing 6, 154–160.
Beasley, J.E. (1983). Route-first cluster-second methods for vehicle routing. Omega 11, 403–408.
Bell, W., Dalberto, L., Fisher, M.L., Greenfield, A., Jaikumar, R., Kedia, P., Mack, R., Prutzman, P.
(1983). Improving the distribution of industrial gases with an on-line computerized routing and
scheduling optimizer. Interfaces 13 (6), 4–23.
Benders, J.F. (1962). Partitioning procedures for solving mixed variables programming problems. Nu-
merische Mathematik 4, 238–252.
Bent, R., Van Hentenryck, P. (2004). A two-stage hybrid local search for the vehicle routing problem
with time windows. Transportation Science 38, 515–530.
Benton, W.C., Rossetti, M.D. (1992). The vehicle scheduling problem with intermittent customer de-
mands. Computers & Operations Research 19, 521–531.
Beraldi, P., Ghiani, G., Laporte, G., Musmanno, G. (2005). Efficient neighbourhood search for the
probabilistic pickup and delivery travelling salesman problem. Networks 46, 195–198.
Berger, J., Barkaoui, M. (2004). A new hybrid genetic algorithm for the capacitated vehicle routing
problem. Journal of the Operational Research Society 54, 1254–1262.
Berger, J., Barkaoui, M., Bräysy, O. (2003). A route-directed hybrid genetic approach for the vehicle
routing problem with time windows. INFOR 41, 179–194.
Bertazzi, L., Paletta, G., Speranza, M.G. (2002). Deterministic order-up-to level policies in an inventory
routing problem. Transportation Science 36, 119–132.
Bertsekas, D.P. (1995). Dynamic Programming and Optimal Control. Athena Scientific, Belmont, MA.
 Ch. 6. Vehicle Routing 419

Bertsimas, D.J. (1988). Probabilistic combinatorial optimization problems. PhD thesis, Operations Re-
search Center, Massachusetts Institute of Technology, Cambridge, MA.
Bertsimas, D.J. (1992). A vehicle routing problem with stochastic demand. Operations Research 40,
574–585.
Bertsimas, D.J., Howell, L.H. (1993). Further results on the probabilistic traveling salesman problem.
European Journal of Operational Research 65, 68–95.
Bertsimas, D.J., Simchi-Levi, D. (1996). A new generation of vehicle routing research: Robust algo-
rithms addressing uncertainty. Operations Research 44, 286–304.
Bertsimas, D.J., Jaillet, P., Odoni, A.R. (1990). A priori optimisation. Operations Research 38, 1019–
1033.
Blasum, U., Hochstättler, W. (2000). Application of the branch and cut method to the vehi-
cle routing problem. Technical Report ZPR2000-386, ZPR, Universität zu Köln. Available at
http://www.zaik.uni-koeln.de/~paper.
Bozkaya, B., Erkut, E., Laporte, G. (2003). A tabu search algorithm and adaptive memory procedure
for political districting. European Journal of Operational Research 144, 12–26.
Bramel, J., Simchi-Levi, D. (1995). A location based heuristic for general routing problems. Operations
Research 43, 649–660.
Bramel, J., Simchi-Levi, D. (1997). On the effectiveness of set covering formulations for the vehicle
routing problem with time windows. Operations Research 45, 295–301.
Bramel, J., Simchi-Levi, D. (2002). Set-covering-based algorithms for the capacitated VRP. In: Toth,
P., Vigo, D. (Eds.), The Vehicle Routing Problem. SIAM Monographs on Discrete Mathematics and
Applications. SIAM, Philadelphia, pp. 85–108.
Brandão, J. (1998). Metaheuristic for the vehicle routing problem with time windows. In: Voss, S.,
Martello, S., Osman, I.H., Roucairol, C. (Eds.), Meta-Heuristics: Advances and Trends in Local
Search Paradigms for Optimization. Kluwer Academic, Boston, pp. 19–36.
Bräysy, O. (2002). Fast local searches for the vehicle routing problem with time windows. INFOR 40,
319–330.
Bräysy, O. (2003). A reactive variable neighborhood search for the vehicle routing problem with time
windows. INFORMS Journal on Computing 15, 347–368.
Bräysy, O., Gendreau, M. (2005a). Vehicle routing problem with time windows, Part I: Route construc-
tion and local search algorithms. Transportation Science 39, 104–118.
Bräysy, O., Gendreau, M. (2005b). Vehicle routing problem with time windows, Part II: Metaheuristics.
Transportation Science 39, 119–139.
Burns, L.D., Hall, R.W., Blumenfeld, D.E., Daganzo, C.F. (1985). Distribution strategies that minimize
transportation and inventory costs. Operations Research 33, 469–490.
Campbell, A.M., Savelsbergh, M.W.P. (2004a). Delivery volume optimization. Transportation Science 38,
210–223.
Campbell, A.M., Savelsbergh, M.W.P. (2004b). A decomposition approach for the inventory-routing
problem. Transportation Science 38, 488–502.
Campbell, A.M., Savelsbergh, M.W.P. (2004c). Efficient insertion heuristics for vehicle routing and
scheduling problems. Transportation Science 38, 269–378.
Campbell, A.M., Clarke, L., Kleywegt, A.J., Savelsbergh, M.W.P. (1998). The inventory routing prob-
lem. In: Crainic, T.G., Laporte, G. (Eds.), Fleet Management and Logistics. Kluwer Academic,
Boston, pp. 95–112.
Campos, V., Corberán, A., Mota, E. (1991). Polyhedral results for a vehicle routing problem. European
Journal of Operational Research 52, 75–85.
Carraway, R.L., Morin, T.L., Moskowitz, H. (1989). Generalized dynamic programming for stochastic
combinatorial optimization. Operations Research 37, 819–829.
Çetinkaya, S., Lee, C.Y. (2000). Stock replenishment and shipment scheduling for vendor managed
inventory systems. Management Science 46, 217–232.
Chabrier, A. (2006). Vehicle routing problem with elementary shortest path based column generation.
Computers & Operations Research 33, 2972–2990.
Chan, L.M., Federgruen, A., Simchi-Levi, D. (1998). Probabilistic analyses and practical algorithms for
inventory-routing models. Operations Research 46, 96–106.
420 J.-F. Cordeau et al.

Chiang, W.-C., Russell, R.A. (1997). A reactive tabu search metaheuristic for the vehicle routing prob-
lem with time windows. INFORMS Journal on Computing 9, 417–430.
Chien, T., Balakrishnan, A., Wong, R. (1989). An integrated inventory allocation and vehicle routing
problem. Transportation Science 23, 67–76.
Christiansen, M. (1999). Decomposition of a combined inventory and time constrained ship routing
problem. Transportation Science 33, 3–16.
Christiansen, M., Nygreen, B. (1998a). A method for solving ship routing problems with inventory
constraints. Annals of Operations Research 81, 357–378.
Christiansen, M., Nygreen, B. (1998b). Modelling path flows for a combined ship routing and inventory
management problem. Annals of Operations Research 82, 391–412.
Christofides, N., Mingozzi, A. (1989). Vehicle routing: Practical and algorithmic aspects. In: Van Rijn,
C.F.H. (Ed.), Logistics: Where Ends Have to Meet. Pergamon, Oxford, pp. 30–48.
Christofides, N., Mingozzi, A., Toth, P. (1979). The vehicle routing problem. In: Christofides, N.,
Mingozzi, A., Toth, P., Sandi, C. (Eds.), Combinatorial Optimization. Wiley, Chichester, pp. 315–
338.
Christofides, N., Mingozzi, A., Toth, P. (1981a). Exact algorithms for the vehicle routing problem based
on the spanning tree and shortest path relaxations. Mathematical Programming 20, 255–282.
Christofides, N., Mingozzi, A., Toth, P. (1981b). State-space relaxation procedures for the computation
of bounds to routing problems. Networks 11, 145–164.
Clarke, G., Wright, J.W. (1964). Scheduling of vehicles from a central depot to a number of delivery
points. Operations Research 12, 568–581.
Cook, W., Rich, J.L. (1999). A parallel cutting-plane algorithm for the vehicle routing problem with
time windows. Technical Report TR99-04, Computational and Applied Mathematics Department,
Rice University, TX.
Cordeau, J.-F., Laporte, G. (2001). A tabu search algorithm for the site dependent vehicle routing
problem with time windows. INFOR 39, 292–298.
Cordeau, J.-F., Laporte, G. (2004). Tabu search heuristics for the vehicle routing problem. In: Rego, C.,
Alidaee, B. (Eds.), Metaheuristic Optimization via Memory and Evolution: Tabu Search and Scatter
Search. Kluwer Academic, Boston, pp. 145–163.
Cordeau, J.-F., Gendreau, M., Laporte, G. (1997). A tabu search heuristic for periodic and multi-depot
vehicle routing problems. Networks 30, 105–119.
Cordeau, J.-F., Laporte, G., Mercier, A. (2001). A unified tabu search heuristic for vehicle routing
problems with time windows. Journal of the Operational Research Society 52, 928–936.
Cordeau, J.-F., Desaulniers, G., Desrosiers, J., Solomon, M.M., Soumis, F. (2002a). VRP with Time
Windows. In: Toth, P., Vigo, D. (Eds.), The Vehicle Routing Problem. SIAM Monographs on Discrete
Mathematics and Applications. SIAM, Philadelphia, pp. 157–193.
Cordeau, J.-F., Gendreau, M., Laporte, G., Potvin, J.-Y., Semet, F. (2002b). A guide to vehicle routing
heuristics. Journal of the Operational Research Society 53, 512–522.
Cordeau, J.-F., Laporte, G., Mercier, A. (2004). An improved tabu search algorithm for the handling of
route duration constraints in vehicle routing problems with time windows. Journal of the Operational
Research Society 55, 542–546.
Cordeau, J.-F., Gendreau, M., Hertz, A., Laporte, G., Sormany, J.-S. (2005). New heuristics for the vehi-
cle routing problem. In: Langevin, A., Riopel, D. (Eds.), Logistics Systems: Design and Optimization.
Springer-Verlag, New York, pp. 279–297.
Cordone, R., Wolfler Calvo, R. (2001). A heuristic for the vehicle routing problem with time windows.
Journal of Heuristics 7, 107–129.
Cornuéjols, G., Harche, F. (1993). Polyhedral study of the capacitated vehicle routing problem. Mathe-
matical Programming 60, 21–52.
Croes, A. (1958). A method for solving traveling salesman problems. Operations Research 6, 791–812.
Danna, E., Le Pape, C. (2003). Accelerating branch-and-price with local search: A case study on the
vehicle routing problem with time windows. Technical Report 03-006, ILOG.
Dantzig, G.B., Ramser, J.M. (1959). The truck dispatching problem. Management Science 6, 81–91.
Dantzig, G.B., Wolfe, P. (1960). Decomposition principle for linear programming. Operations Re-
search 8, 101–111.
 Ch. 6. Vehicle Routing 421

De Backer, B., Furnon, V., Kilby, P., Prosser, P., Shaw, P. (2000). Solving vehicle routing problems using
constraint programming and metaheuristics. Journal of Heuristics 6, 501–523.
Dell’Amico, M., Toth, P. (2000). Algorithms and codes for dense assignment problems: The state of the
art. Discrete Applied Mathematics 100, 17–48.
Desrochers, M., Laporte, G. (1991). Improvements and extensions to the Miller–Tucker–Zemlin sub-
tour elimination constraints. Operations Research Letters 10, 27–36.
Desrochers, M., Soumis, F. (1988). A generalized permanent labeling algorithm for the shortest path
problem with time windows. INFOR 26, 191–212.
Desrochers, M., Verhoog, T.W. (1989). A matching based savings algorithm for the vehicle routing
problem. Les Cahiers du GERAD G–89–04, HEC Montréal.
Desrochers, M., Lenstra, J.K., Savelsbergh, M.W.P., Soumis, F. (1988). Vehicle routing with time win-
dows: Optimization and approximation. In: Golden, B.L., Assad, A.A. (Eds.), Vehicle Routing:
Methods and Studies. North-Holland, Amsterdam, pp. 65–84.
Desrochers, M., Desrosiers, J., Solomon, M.M. (1992). A new optimization algorithm for the vehicle
routing problem with time windows. Operations Research 40, 342–354.
Dorigo, M., Di Caro, G., Gambardella, L.M. (1999). Ant algorithms for discrete optimization. Artificial
Life 5, 137–172.
Drezner, Z. (2003). A new genetic algorithm for the quadratic assignment problem. INFORMS Journal
on Computing 15, 320–330.
Dror, M., Ball, M.O. (1987). Inventory/routing: Reduction from an annual to a short period problem.
Naval Research Logistics Quarterly 34, 891–905.
Dror, M., Trudeau, P. (1986). Stochastic vehicle routing with modified savings algorithm. European
Journal of Operational Reserach 23, 228–235.
Dror, M., Ball, M.O., Golden, B.L. (1985). A computational comparison of algorithms for the inventory
routing problem. Annals of Operations Research 4, 3–23.
Dror, M., Laporte, G., Trudeau, P. (1989). Vehicle routing with stochastic demands: Properties and
solution frameworks. Transportation Science 23, 166–176.
Dror, M., Laporte, G., Louveaux, F.V. (1993). Vehicle routing with stochastic demands and restricted
failures. Zeitschrift für Operations Research 37, 273–283.
Dueck, G. (1990). New optimization heuristics, the great deluge algorithm and the record-to-record
travel. Technical report, IBM Germany, Heidelberg Scientific Center.
Dueck, G. (1993). New optimization heuristics: The great deluge algorithm and the record-to-record
travel. Journal of Computational Physics 104, 86–92.
Ergun, Ö., Orlin, J.B., Steele-Feldman, A. (2003). Creating very large scale neighborhoods out of
smaller ones by compounding moves: A study on the vehicle routing problem. MIT Sloan work-
ing Paper 4393-02, Massachusetts Institute of Technology, Cambridge, MA.
Federgruen, A., Zipkin, P. (1984). A combined vehicle routing and inventory allocation problem. Oper-
ations Research 32, 1019–1036.
Fischetti, M., Toth, P. (1989). An additive bounding procedure for combinatorial optimization prob-
lems. Operations Research 37, 319–328.
Fischetti, M., Toth, P., Vigo, D. (1994). A branch-and-bound algorithm for the capacitated vehicle rout-
ing problem on directed graphs. Operations Research 42, 846–859.
Fisher, M.L. (1994). Optimal solution of vehicle routing problems using minimum k-trees. Operations
Research 42, 626–642.
Fisher, M.L., Jaikumar, R. (1981). A generalized assignment heuristic for the vehicle routing problem.
Networks 11, 109–124.
Fisher, M.L., Greenfield, A., Jaikumar, R., Kedia, P. (1982). Real-time scheduling of a bulk delivery
fleet: Practical application of Lagrangean relaxation. Technical report, The Wharton School, Uni-
versity of Pennsylvania.
Fisher, M.L., Jörnsten, K.O., Madsen, O.B.G. (1997). Vehicle routing with time windows – two opti-
mization algorithms. Operations Research 45, 488–492.
Flood, M.M. (1956). The travelling salesman problem. Operations Research 4, 61–75.
Foster, B.A., Ryan, D.M. (1976). An integer programming approach to the vehicle scheduling problem.
Operations Research 27, 367–384.
422 J.-F. Cordeau et al.

Fukasawa, R., Longo, H., Lysgaard, J., Poggi de Aragão, M., Reis, M., Uchoa, E., Werneck, R.F. (2006).
Robust branch-and-cut-and-price for the capacitated vehicle routing problem. Mathematical Pro-
gramming A 106, 491–511.
Gallego, G., Simchi-Levi, D. (1990). On the effectiveness of direct shipping strategy for the one-
warehouse multi-retailer r-systems. Management Science 36, 240–243.
Gambardella, L.M., Taillard, É.D., Agazzi, G. (1999). MACS-VRPTW: A multiple ant colony system
for vehicle routing problems with time windows. In: Corne, D., Dorigo, M., Glover, F. (Eds.), New
Ideas in Optimization. McGraw-Hill, London, pp. 63–76.
Gaur, V., Fisher, M.L. (2004). A periodic inventory routing problem at a supermarket chain. Operations
Research 52, 813–822.
Gehring, H., Homberger, J. (2002). Parallelization of a two-phase metaheuristic for routing problems
with time windows. Journal of Heuristics 8, 251–276.
Gendreau, M., Hertz, A., Laporte, G. (1992). New insertion and post-optimization procedures for the
traveling selesman problem. Operations Research 40, 1083–1094.
Gendreau, M., Hertz, A., Laporte, G. (1994). A tabu search heuristic for the vehicle routing problem.
Management Science 40, 1276–1290.
Gendreau, M., Laporte, G., Séguin, R. (1995). An exact algorithm for the vehicle routing problem with
stochastic customers and demands. Transportation Science 29, 143–155.
Gendreau, M., Laporte, G., Séguin, R. (1996). A tabu search algorithm for the vehicle routing problem
with stochastic demands and customers. Operations Research 44, 469–477.
Gendreau, M., Hertz, A., Laporte, G., Stan, M. (1998). A generalized insertion heuristic for the travel-
ing salesman problem with time windows. Operations Research 43, 330–335.
Gendreau, M., Laporte, G., Potvin, J.-Y. (2002). Metaheuristics for the capacitated VRP. In: Toth,
P., Vigo, D. (Eds.), The Vehicle Routing Problem. SIAM Monographs on Discrete Mathematics and
Applications. SIAM, Philadelphia, pp. 129–154.
Ghaziri, H. (1993). Algorithmes connexionistes pour l’optimisation combinatoire. Thèse de doctorat,
École Polytechnique Fédérale de Lausanne, Switzerland.
Ghiani, G., Laporte, G., Semet, F. (2006). The black and white traveling salesman problem. Operations
Research 54, 366–378.
Gillett, B.E., Miller, L.R. (1974). A heuristic algorithm for the vehicle-dispatch problem. Operations
Research 21, 340–349.
Glover, F. (1992). New ejection chain and alternating path methods for traveling salesman problems.
In: Balci, O., Sharda, R., Zenios, S. (Eds.), Computer Science and Operations Research: New Devel-
opments in Their Interfaces. Pergamon, Oxford, pp. 449–509.
Golden, B.L., Stewart, W.R. (1978). Vehicle routing with probabilistic demands. In: Hogben, D., Fife,
D. (Eds.) Computer Science and Statistics: Tenth Annual Symposium on the Interface. NBS Special
Publication, vol. 503, pp. 252–259.
Golden, B.L., Magnanti, T.L., Nguyen, H.Q. (1977). Implementing vehicle routing algorithms. Net-
works 7, 113–148.
Golden, B.L., Assad, A.A., Dahl, R. (1984). Analysis of a large scale vehicle routing problem with an
inventory component. Large Scale Systems 7, 181–190.
Golden, B.L., Wasil, E.A., Kelly, J.P., Chao, I-M. (1998). Metaheuristics in vehicle routing. In: Crainic,
T.G., Laporte, G. (Eds.), Fleet Management and Logistics. Kluwer Academic, Boston, pp. 33–56.
Golden, B.L., Assad, A.A., Wasil, E.A. (2002). Routing vehicles in the real world: Applications in the
solid waste, beverage, food, dairy, and newspaper industries. In: Toth, P., Vigo, D. (Eds.), The Vehicle
Routing Problem. SIAM Monographs on Discrete Mathematics and Applications. SIAM, Philadelphia,
pp. 245–286.
Gouveia, L. (1995). A result on projection for the vehicle routing problem. Journal of Operational Re-
search Society 85, 610–624.
Hadjiconstantinou, E., Christofides, N., Mingozzi, A. (1995). A new exact algorithm for the vehicle rout-
ing problem based on q-paths and k-shortest paths relaxations. Annals of Operations Research 61,
21–43.
Haimovich, M., Rinnooy Kan, A.H.G. (1985). Bounds and heuristics for capacitated routing problems.
Mathematics of Operations Research 10, 527–542.
 Ch. 6. Vehicle Routing 423

Haughton, M.A. (1998). The performance of route modification and demand stabilization strategies in
stochastic vehicle routing. Transportation Research 32, 551–566.
Haughton, M.A. (2000). Quantifying the benefits of route reoptimisation under stochastic customer
demands. Journal of the Operational Research Society 51, 320–332.
Held, M., Karp, R.M. (1971). The traveling salesman problem and minimum spanning trees: Part II.
Mathematical Programming 1, 6–25.
Held, M., Wolfe, P., Crowder, M.P. (1974). Validation of the subgradient optimization. Mathematical
Programming 6, 62–88.
Herer, Y., Roundy, R. (1997). Heuristics for a one-warehouse multiretailer distribution problem with
performance bounds. Operations Research 45, 102–115.
Hjorring, C., Holt, J. (1999). New optimality cuts for a single-vehicle stochastic routing problem. Annals
of Operations Research 86, 569–585.
Homberger, J., Gehring, H. (1999). Two evolutionary metaheuristics for the vehicle routing problem
with time windows. INFOR 37, 297–318.
Houck, D.J., Picard, J.-C., Queyranne, M., Vemuganti, R.R. (1980). The traveling salesman problem as
a constrained shortest path problem: Theory and computational experience. Opsearch 17, 93–109.
Hvattum, L.M., Løkketangen, A., Laporte, G. (2006). Solving a dynamic and stochastic vehicle routing
problem with a sample scenario hedging heuristic. Transportation Science, in press.
Ioannou, G., Kritikos, M., Prastacos, G. (2001). A greedy look-ahead heuristic for the vehicle routing
problem with time windows. Journal of the Operational Research Society 52, 523–537.
Irnich, S., Villeneuve, D. (2003). The shortest path problem with resource constraints and k-cycle
elimination for k  3. Technical report, Rheinisch-Westfälische Technische Hochschule, Aachen,
Germany.
Jaillet, P. (1985). Probabilistic traveling salesman problem. PhD thesis, Operations Research Center,
Massachusetts Institute of Technology, Cambridge, MA.
Jaillet, P. (1987). Stochastic routing problems. In: Andreatta, G., Mason, F., Serafini, P. (Eds.), Stochas-
tics in Combinatorial Optimization. World Scientific, Singapore, pp. 197–213.
Jaillet, P. (1988). A priori solution of a traveling salesman problem in which a random subset of the
customers are visited. Operations Research 36, 929–936.
Jaillet, P., Odoni, A.R. (1988). The probabilistic vehicle routing problem. In: Golden, B.L., Assad, A.A.
(Eds.), Vehicle Routing: Methods and Studies. North-Holland, Amsterdam, pp. 293–318.
Jaillet, P., Bard, J.F., Huang, L., Dror, M. (2002). Delivery cost approximations for inventory routing
problems in a rolling horizon framework. Transportation Science 3, 292–300.
Jézéquel, A. (1985). Probabilistic vehicle routing problems. MSc dissertation, Department of Civil En-
gineering, Massachusetts Institute of Technology, Cambridge, MA.
Kallehauge, B., Larsen, J., Madsen, O.B.G. (2006). Lagrangean duality applied to the vehicle routing
with time windows. Computers & Operations Research 33, 1464–1487.
Kao, E.P.C. (1978). A preference order dynamic program for a stochastic traveling salesman problem.
Operations Research 26, 1033–1045.
Kenyon, A.S., Morton, D.P. (2003). Stochastic vehicle routing with random travel times. Transportation
Science 37, 69–82.
Kilby, P.J., Prosser, P., Shaw, P. (1998). Guided local search for the vehicle routing problem with time
windows. In: Voss, S., Martello, S., Osman, I.H., Roucairol, C. (Eds.), Meta Heuristics: Advances and
Trends in Local Search Paradigms for Optimisation. Kluwer Academic, Boston, pp. 473–486.
Kindervater, G.A.P., Savelsbergh, M.W.P. (1997). Vehicle routing: Handling edge exchanges. In:
Aarts, E.H.L., Lenstra, J.K. (Eds.), Local Search in Combinatorial Optimization. Wiley, Chichester,
pp. 337–360.
Kleywegt, A.J., Nori, V., Savelsbergh, M.W.L. (2002). The stochastic inventory routing problem with
direct deliveries. Transportation Science 36, 94–118.
Kleywegt, A.J., Nori, V., Savelsbergh, M.W.L. (2004). Dynamic programming approximations for a sto-
chastic inventory routing problem. Transportation Science 38, 42–70.
Kohl, N., Madsen, O.B.G. (1997). An optimization algorithm for the vehicle routing problem with time
windows based on Lagrangean relaxation. Operations Research 45, 395–406.
424 J.-F. Cordeau et al.

Kohl, N., Desrosiers, J., Madsen, O.B.G., Solomon, M.M., Soumis, F. (1999). 2-path cuts for the vehicle
routing problem with time windows. Transportation Science 33, 101–116.
Kolen, A.W.J., Rinnooy Kan, A.H.G., Trienekens, H.W.J.M. (1987). Vehicle routing with time windows.
Operations Research 35, 256–273.
Kontoravdis, G., Bard, J.F. (1995). A GRASP for the vehicle routing problem with time windows. ORSA
Journal on Computing 7, 10–23.
Lambert, V., Laporte, G., Louveaux, F.V. (1993). Designing collection routes through bank branches.
Computers & Operations Research 20, 783–791.
Laporte, G., Louveaux, F.V. (1990). Formulations and bounds for the stochastic capacitated vehi-
cle routing problem with uncertain supplies. In: Gabzewicz, J., Richard, J.-F., Wolsey, L. (Eds.),
Economic Decision Making: Games, Econometrics and Optimisation. North-Holland, Amsterdam,
pp. 443–455.
Laporte, G., Louveaux, F.V. (1993). The integer L-shaped method for stochastic integer programs with
complete recourse. Operations Research Letters 13, 133–142.
Laporte, G., Louveaux, F.V. (1998). Solving stochastic routing problems with the integer L-shaped
method. In: Crainic, T.G., Laporte, G. (Eds.), Fleet Management and Logistics. Kluwer Academic,
Boston, pp. 159–167.
Laporte, G., Nobert, Y. (1987). Exact algorithms for the vehicle routing problem. Annals of Discrete
Mathematics 31, 147–184.
Laporte, G., Semet, F. (2002). Classical heuristics for the capacitated VRP. In: Toth, P., Vigo, D. (Eds.),
The Vehicle Routing Problem. SIAM Monographs on Discrete Mathematics and Applications. SIAM,
Philadelphia, pp. 109–128.
Laporte, G., Nobert, Y., Desrochers, M. (1985). Optimal routing under capacity and distance restric-
tions. Operations Research 33, 1050–1073.
Laporte, G., Mercure, H., Nobert, Y. (1986). An exact algorithm for the asymmetrical capacitated
vehicle routing problem. Networks 16, 33–46.
Laporte, G., Louveaux, F.V., Mercure, H. (1989). Models and exact solutions for a class of stochastic
location-routing problems. European Journal of Operational Research 39, 71–78.
Laporte, G., Louveaux, F.V., Mercure, H. (1992). The vehicle routing problem with stochastic travel
times. Transportation Science 26, 161–170.
Laporte, G., Louveaux, F.V., Mercure, H. (1994). A priori optimization of the probabilistic traveling
salesman problem. Operations Research 42, 543–549.
Laporte, G., Louveaux, F.V., Van hamme, L. (2002). An integer L-shaped algorithm for the capacitated
vehicle routing problem with stochastic demands. Operations Research 50, 415–423.
Larson, R.C. (1988). Transportation of sludge to the 106-mile site: An inventory routing problem for
fleet sizing and logistic system design. Transportation Science 22, 186–198.
Lau, H.C., Liu, Q., Ono, H. (2002). Integrating local search and network flow to solve the inventory
routing problem. American Association for Artificial Intelligence 2, 9–14.
Lau, H.C., Sim, M., Teo, K.M. (2003). Vehicle routing problem with time windows and a limited number
of vehicles. European Journal of Operational Research 148, 559–569.
Leipälä, T. (1978). On the solutions of stochastic traveling salesman problems. European Journal of
Operational Research 2, 291–297.
Letchford, A., Eglese, R.W., Lysgaard, J. (2002). Multistars, partial multistars and the capacitated ve-
hicle routing problem. Mathematical Programming 94, 21–40.
Li, F., Golden, B.L., Wasil, E.A. (2005). Very large-scale vehicle routing: New test problems, algorithms
and results. Computers & Operations Research 32, 1165–1179.
Li, H., Lim, A. (2003). Local search with annealing-like restarts to solve the VRPTW. European Journal
of Operational Research 150, 115–127.
Lin, S. (1965). Computer solutions of the travelling salesman problem. Bell System Technical Journal 44,
2245–2269.
Louveaux, F.V. (1998). An introduction to stochastic transportation models. In: Labbé, M., Laporte,
G., Tanczos, K., Toint, P. (Eds.), Operations Research and Decision Aid Methodologies in Traffic
and Transportation Management. NATO ASI Series F: Computer and Systems Sciences, vol. 166.
Springer-Verlag, Berlin/Heidelberg, pp. 244–263.
 Ch. 6. Vehicle Routing 425

Lysgaard, J., Letchford, A., Eglese, R.W. (2004). A new branch-and-cut algorithm for the capacitated
vehicle routing problem. Mathematical Programming 100, 423–445.
Martinhon, C., Lucena, A., Maculan, N. (2000). A relax and cut algorithm for the vehicle routing prob-
lem. Technical Report RT-05/00, Universidade Federal Fluminense, Niterói, Brasil.
Mester, D., Bräysy, O. (2005). Active guided evolution strategies for large scale vehicle routing problem
with time windows. Computers & Operations Research 32, 1593–1614.
Miller, D.L. (1995). A matching based exact algorithm for capacitated vehicle routing problems. ORSA
Journal on Computing 7, 1–9.
Miller, D.L., Pekny, J.F. (1995). A staged primal-dual algorithm for perfect b-matching with edge ca-
pacities. ORSA Journal on Computing 7, 298–320.
Mingozzi, A., Christofides, N., Hadjiconstantinou, E. (1994). An exact algorithm for the vehicle routing
problem based on the set partitioning formulation. Technical report, Department of Mathematics,
University of Bologna, Italy.
Minkoff, A. (1993). A Markov decision model and decomposition heuristic for dynamic vehicle dis-
patching. Operations Research 41, 77–90.
Mladenović, N., Hansen, P. (1997). Variable neighborhood search. Computers & Operations Research 24,
1097–1100.
Mole, R.H., Jameson, S.R. (1976). A sequential route-building algorithm employing a generalized sav-
ings criterion. Operational Research Quarterly 27, 503–511.
Moscato, P., Cotta, C. (2003). A gentle introduction to memetic algorithms. In: Glover, F., Kochen-
berger, G.A. (Eds.), Handbook of Metaheuristics. Kluwer Academic, Boston, pp. 105–144.
Naddef, D., Rinaldi, G. (2002). Branch-and-cut algorithms for the capacitated VRP. In: Toth, P., Vigo,
D. (Eds.), The Vehicle Routing Problem. SIAM Monographs on Discrete Mathematics and Applications.
SIAM, Philadelphia, pp. 53–84.
Nelson, M.D., Nygard, K.E., Griffin, J.H., Shreve, W.E. (1985). Implementation techniques for the
vehicle routing problem. Computers & Operations Research 12, 273–283.
Or, I. (1976). Traveling salesman-type combinatorial problems and their relation to the logistics of
blood banking. PhD thesis, Department of Industrial Engineering and Management Science, North-
western University, Evanston, IL.
Osman, I.H. (1993). Metastrategy simulated annealing and tabu search algorithms for the vehicle rout-
ing problem. Annals of Operations Research 41, 421–451.
Paessens, H. (1988). The savings algorithm for the vehicle routing problem. European Journal of Oper-
ational Research 34, 336–344.
Potvin, J.-Y. (1996). Genetic algorithms for the traveling salesman problem. Annals of Operations Re-
search 63, 339–370.
Potvin, J.-Y., Bengio, S. (1996). The vehicle routing problem with time windows – Part II: Genetic
search. INFORMS Journal on Computing 8, 165–172.
Potvin, J.-Y., Rousseau, J.-M. (1993). A parallel route building algorithm for the vehicle routing and
scheduling problem with time windows. European Journal of Operational Research 66, 331–340.
Potvin, J.-Y., Rousseau, J.-M. (1995). An exchange heuristic for routing problems with time windows.
Journal of the Operational Research Society 46, 1433–1446.
Potvin, J.-Y., Kervahut, T., Garcia, B.L., Rousseau, J.-M. (1996). The vehicle routing problem with time
windows – Part I: Tabu search. INFORMS Journal on Computing 8, 158–164.
Prins, C. (2004). A simple and effective evolutionary algorithm for the vehicle routing problem. Com-
puters & Operations Research 31, 1985–2002.
Ralphs, T.K., Kopman, L., Pulleyblank, W.R., Trotter, Jr., L.E. (2003). On the capacitated vehicle
routing problem. Mathematical Programming B 94, 343–359.
Rechenberg, I. (1973). Evolutionsstrategie. Fromman-Holzboog, Stuttgart, Germany.
Rego, C. (1998). A subpath ejection method for the vehicle routing problem. Management Science 44,
1447–1459.
Rego, C., Roucairol, C. (1996). A parallel tabu search algorithm using ejection chains for the vehi-
cle routing problem. In: Osman, I.H., Kelly, J.P. (Eds.), Meta-Heuristics: Theory and Applications.
Kluwer Academic, Boston, pp. 661–675.
426 J.-F. Cordeau et al.

Reimann, M., Rubio, R., Wein, L.M. (1999). Heavy traffic analysis of the dynamic stochastic inventory-
routing problem. Transportation Science 33, 361–380.
Reimann, M., Doerner, K., Hartl, R.F. (2004). D-ants: Savings based ants divide and conquer for the
vehicle routing problem. Computers & Operations Research 31, 563–591.
Renaud, J., Boctor, F.F., Laporte, G. (1996a). A fast composite heuristic for the symmetric traveling
salesman problem. INFORMS Journal on Computing 8, 134–143.
Renaud, J., Boctor, F.F., Laporte, G. (1996b). An improved petal heuristic for the vehicle routing prob-
lem. Journal of the Operational Research Society 47, 329–336.
Rochat, Y., Taillard, É.D. (1995). Probabilistic diversification and intensification in local search for
vehicle routing. Journal of Heuristics 1, 147–167.
Rossi, F.A., Gavioli, I. (1987). Aspects of heuristic methods in the “Probabilistic traveling salesman
problem”. In: Andreatta, G., Mason, F., Serafini, P. (Eds.), Stochastics in Combinatorial Optimiza-
tion. World Scientific, Singapore, pp. 214–227.
Russell, R.A. (1977). An effective heuristic for the M-tour traveling salesman problem with some side
conditions. Operations Research 25, 517–524.
Russell, R.A. (1995). Hybrid heuristics for the vehicle routing problem with time windows. Transporta-
tion Science 29, 156–166.
Ryan, D.M., Hjorring, C., Glover, F. (1993). Extensions of the petal method for vehicle routing. Journal
of Operational Research Society 44, 289–296.
Savelsbergh, M.W.P. (1985). Local search in routing problems with time windows. Annals of Operations
Research 4, 285–305.
Savelsbergh, M.W.P. (1990). En efficient implementation of local search algorithms for constrained
routing problems. European Journal of Operational Research 47, 75–85.
Savelsbergh, M.W.P. (1992). The vehicle routing problem with time windows: Minimizing route dura-
tion. ORSA Journal on Computing 4, 146–154.
Savelsbergh, M.W.P., Song, J.-H. (2005). Inventory routing with continuous moves. Computers & Oper-
ations Research, in press.
Schulze, J., Fahle, T. (1999). A parallel algorithm for the vehicle routing problem with time window
constraints. Annals of Operations Research 86, 585–607.
Secomandi, N. (1998). Exact and heuristic dynamic programming algorithms for the vehicle routing
problem with stochastic demands. PhD dissertation, Faculty of the College of Business Administra-
tion, University of Houston, TX.
Secomandi, N. (2000). Comparing neuro-dynamic programming algorithms for the vehicle routing
problem with stochastic demands. Computers & Operations Research 27, 1201–1225.
Secomandi, N. (2003). Analysis of a rollout approach to sequencing problems with stochastic routing
applications. Journal of Heuristics 9, 321–352.
Séguin, R. (1994). Problèmes stochastiques de tournées de véhicules. PhD thesis, Département
d’informatique et de recherche opérationnelle, Université de Montréal, Canada.
Semet, F., Taillard, É.D. (1993). Solving real-life vehicle routing problems efficiently using tabu search.
Annals of Operations Research 41, 469–488.
Shaw, P. (1998). Using constraint programming and local search methods to solve vehicle routing
problems. In: Maher, M., Puget, J.-F. (Eds.), Principles and Practice of Constraint Programming.
Springer-Verlag, New York, pp. 417–431.
Sniedovich, M. (1981). Analysis of a preference order traveling salesman problem. Operations Re-
search 29, 1234–1237.
Solomon, M.M. (1987). Algorithms for the vehicle routing and scheduling problems with time window
constraints. Operations Research 35, 254–265.
Solomon, M.M., Baker, E.K., Schaffer, J.R. (1988). Vehicle routing and scheduling problems with time
window constraints: Efficient implementations of solution improvement procedures. In: Golden,
B.L., Assad, A.A. (Eds.), Vehicle Routing: Methods and Studies. North-Holland, Amsterdam, pp. 85–
106.
Song, J.-H., Savelsbergh, M.W.P. (2005). Performance measurement for inventory routing. Transporta-
tion Science, in press.
 Ch. 6. Vehicle Routing 427

Stewart, W.R., Golden, B.L. (1983). Stochastic vehicle routing: A comprehensive approach. European
Journal of Operational Research 14, 371–385.
Taillard, É.D. (1993). Parallel iterative search methods for vehicle routing problems. Networks 23, 661–
673.
Taillard, É.D., Badeau, P., Gendreau, M., Guertin, F., Potvin, J.-Y. (1997). A tabu search heuristic for
the vehicle routing problem with soft time windows. Transportation Science 31, 170–186.
Tan, K.C., Lee, L.H., Ou, K. (2001). Hybrid genetic algorithms in solving vehicle routing problems with
time window constraints. Asia-Pacific Journal of Operational Research 18, 170–186.
Tarantilis, C.-D., Kiranoudis, C.T. (2002). Bone route: Adaptive memory method for effective fleet
management. Annals of Operations Research 115, 227–241.
Thangiah, S.R., Petrovic, P. (1998). Introduction to genetic heuristics and vehicle routing problems
with complex constraints. In: Advances in Computational and Stochastic Optimization, Logic Pro-
gramming, and Heuristic Search. Operations Research/Computer Science Interfaces, vol. 9. Kluwer
Academic, Boston, pp. 253–286.
Thompson, P.M., Psaraftis, H.N. (1993). Cyclic transfer algorithms for multi-vehicle routing and
scheduling problems. Operations Research 41, 935–946.
Tillman, F. (1969). The multiple terminal delivery problem with probabilistic demands. Transportation
Science 3, 192–204.
Toth, P., Vigo, D. (1995). An exact algorithm for the capacitated shortest spanning arborescence. Annals
of Operations Research 61, 121–142.
Toth, P., Vigo, D. (1997). An exact algorithm for the vehicle routing problem with backhauls. Trans-
portation Science 31, 372–385.
Toth, P., Vigo, D. (1998). Exact algorithms for vehicle routing. In: Crainic, T., Laporte, G. (Eds.), Fleet
Management and Logistics. Kluwer Academic, Boston, pp. 1–31.
Toth, P., Vigo, D. (Eds.) (2002a). The Vehicle Routing Problem. SIAM Monographs on Discrete Mathe-
matics and Applications. SIAM, Philadelphia.
Toth, P., Vigo, D. (2002b). An overview of vehicle routing problems. In: Toth, P., Vigo, D. (Eds.),
The Vehicle Routing Problem. SIAM Monographs on Discrete Mathematics and Applications. SIAM,
Philadelphia, pp. 1–26.
Toth, P., Vigo, D. (2002c). Branch-and-bound algorithms for the capacitated VRP. In: Toth, P., Vigo, D.
(Eds.), The Vehicle Routing Problem. SIAM Monographs on Discrete Mathematics and Applications.
SIAM, Philadelphia, pp. 29–51.
Toth, P., Vigo, D. (2002d). Models, relaxations and exact approaches for the capacitated vehicle routing
problem. Discrete Applied Mathematics 123, 487–512.
Toth, P., Vigo, D. (2003). The granular tabu search and its application to the vehicle routing problem.
INFORMS Journal on Computing 15, 333–346.
Trudeau, P., Dror, M. (1992). Stochastic inventory routing: Route design with stockouts and route fail-
ures. Transportation Science 26, 171–184.
Van Breedam, A. (1994). An analysis of the behavior of heuristics for the vehicle routing problem for
a selection of problems with vehicle-related, customer-related, and time-related constraints. PhD
dissertation, University of Antwerp, Belgium.
Van Slyke, R., Wets, R.J.-B. (1969). L-shaped programs with applications to optimal control and sto-
chastic programming. SIAM Journal of Applied Mathematics 17, 638–653.
Verweij, B., Ahmed, S., Kleywegt, A.J., Nemhauser, G.L., Shapiro, A. (2003). The sample average ap-
proximation method applied to stochastic routing problems: A computational study. Computational
Optimization and Applications 24, 289–333.
Vigo, D. (1996). A heuristic algorithm for the asymmetric capacitated vehicle routing problem. Euro-
pean Journal of Operational Research 89, 108–126.
Viswanathan, S., Mathur, K. (1997). Integrating routing and inventory decisions in one-warehouse mul-
tiretailer multiproduct distribution systems. Management Science 43, 294–312.
Volgenant, A., Jonker, R. (1983). The symmetric traveling salesman problem and edge exchange in
minimal 1-trees. European Journal of Operational Research 12, 395–403.
Voudouris, C. (1997). Guided local search for combinatorial problems. Dissertation, University of Es-
sex, United Kingdom.
428 J.-F. Cordeau et al.

Waller, M., Johnson, M.E., Davis, T. (1999). Vendor-managed inventory in the retail supply chain. Jour-
nal of Business Logistics 20, 183–203.
Wang, X., Regan, A.C. (2001). Assignment models for local truckload trucking problems with stochastic
service times and time window constraints. Transportation Research Record 1171, 61–68.
Wark, P., Holt, J. (1994). A repeated matching heuristic for the vehicle routing problem. Journal of
Operational Research Society 45, 1156–1167.
Waters, C.D.J. (1989). Vehicle-scheduling problems with uncertainty and omitted customers. Journal of
the Operational Research Society 40, 1099–1108.
Webb, R., Larson, R. (1995). Period and phase of customer replenishment: A new approach to the
strategic inventory/routing problem. European Journal of Operational Research 85, 132–148.
Willard, J.A.G. (1989). Vehicle routing using r-optimal tabu search. MSc dissertation, The Management
School, Imperial College, London.
Wren, A. (1971). Computers in Transport Planning and Operation. Ian Allan, London.
Wren, A., Holliday, A. (1972). Computer scheduling of vehicles from one or more depots to a number
of delivery points. Operational Research Quarterly 23, 333–344.
Xu, J., Kelly, J.P. (1996). A network flow-based tabu search heuristic for the vehicle routing problem.
Transportation Science 30, 379–393.
Yang, W.H., Mathur, K., Ballou, R.H. (2000). Stochastic vehicle routing with restocking. Transportation
Science 34, 99–112.
C. Barnhart and G. Laporte (Eds.), Handbook in OR & MS, Vol. 14
Copyright © 2007 Elsevier B.V. All rights reserved
DOI: 10.1016/S0927-0507(06)14007-4

Chapter 7
Transportation on Demand
Jean-François Cordeau
Canada Research Chair in Logistics and Transportation, HEC Montréal,
3000, chemin de la Côte-Sainte-Catherine, Montréal, H3T 2A7, Canada
E-mail: [email protected]

Gilbert Laporte
Canada Research Chair in Distribution Management, HEC Montréal,
3000, chemin de la Côte-Sainte-Catherine, Montréal, H3T 2A7, Canada
E-mail: [email protected]

Jean-Yves Potvin
Département d’informatique et de recherche opérationnelle and Centre de recherche sur
les transports, Université de Montréal C.P. 6128, succ. Centre-Ville, Montréal, H3C 3J7,
Canada
E-mail: [email protected]

Martin W.P. Savelsbergh

School of Industrial and Systems Engineering, Georgia Institute of Technology,
Atlanta, GA 30332-0205, USA
E-mail: [email protected]

1 Introduction

Transportation on Demand (TOD) is concerned with the transportation of

passengers or goods between specific origins and destinations at the request of
users. Common examples are dial-a-ride transportation services for the elderly
and the disabled, urban courier services, aircraft sharing, and emergency vehi-
cle dispatching. In all such systems, users formulate requests for transportation
from a pickup point to a delivery (or drop-off) point. These requests are served
by a set of capacitated vehicles that often provide a shared service in the sense
that several passengers or goods may be in a vehicle at the same time.
In recent years, TOD systems have become increasingly popular for a num-
ber of reasons. With the aging of the population and the trend toward the
development of ambulatory health care services, more and more people rely
on door-to-door transportation systems provided by local authorities. Aircraft
sharing has also gained in popularity thanks to cost reduction efforts made
by organizations and to the numerous problems that have recently plagued
the airline industry. Finally, a growing emphasis on electronic commerce,
cycle-time compression, and just-in-time deliveries has increased the need for
demand-responsive freight transportation systems.

429
430 J.-F. Cordeau et al.

TOD systems can be either static or dynamic. In the first case, all requests
are known beforehand while, in the second case, requests are received dy-
namically and vehicle routes must be adjusted in real-time to meet demand.
For instance, courier services are generally highly dynamic whereas dial-a-ride
systems can be regarded as mostly static since they usually require users to
make a reservation at least one day in advance. In practice, dynamic problems
are often treated as sequences of static subproblems. Reoptimization from the
current solution can be performed whenever a new request is formulated, or
requests can be buffered and periodically incorporated in the existing vehicle
routes in batches.
Most TOD problems are characterized by the presence of three often con-
flicting objectives: maximizing the number of requests served, minimizing op-
erating costs, and minimizing user inconvenience. A balance between these
objectives is sometimes obtained by first maximizing the number of requests
that can be accepted given the available capacity and then minimizing the oper-
ating costs while imposing service quality constraints. Service quality is usually
measured in terms of deviations from desired pickup and delivery times and,
in the case of passenger transportation, in terms of excess ride time (i.e., the
difference between the actual ride time of a user and the minimum possible
ride time). Operating costs are mostly related to the number of vehicles used,
to total route duration and to total distance traveled by the vehicles.
Another distinguishing aspect of TOD problems is the importance of the
temporal dimension. Pickups and deliveries are often restricted to take place
within specified time windows. These time windows are sometimes very nar-
row, especially in the case of passenger transportation. In this context, quality
of service is also often controlled by imposing a limit on the ride time of each
user. The latter is particularly important in the case of emergency vehicles.
Waiting while passengers are in a vehicle can also be prohibited. Finally, max-
imum route duration constraints are sometimes imposed to take driver shift
lengths into account.
The day-to-day management of a TOD system involves making decisions
regarding three main aspects: request clustering, vehicle routing, and vehicle
scheduling. Request clustering consists of creating groups of requests to be
served by the same vehicle because of their spatial and temporal proximity.
Given these groups, vehicle routing consists of deciding the order in which the
associated pickup and delivery locations should be visited by each vehicle. Fi-
nally, vehicle scheduling specifies the exact time at which each location should
be visited. These decisions are obviously tightly intertwined and a proper man-
agement of the system calls for their simultaneous optimization.
The Operations Research literature contains numerous studies addressing
both static and dynamic TOD problems. Most variants are, in fact, general-
izations of the Vehicle Routing Problem with Pickup and Delivery (VRPPD).
The aim of this chapter is to present the most important results regarding the
VRPPD and to survey four areas of applications: the dial-a-ride problem, the
urban courier service problem, the dial-a-flight problem, and the emergency
 Ch. 7. Transportation on Demand 431

vehicle dispatch problem. In the latter application, the main emphasis is on

locational issues as opposed to routing choices.
The remainder of the chapter is organized as follows. The next section for-
mally defines the VRPPD, introduces notation that will be used throughout
the chapter, and reviews the related literature. The following four sections
then each focus on a specific application by describing the particularities of
the problem and summarizing the main exact and heuristic solution algorithms
that have been proposed in the literature.

2 The vehicle routing problem with pickup and delivery

The VRPPD is a generalization of the classical VRP which also belongs to

a larger family of Pickup and Delivery Problems (PDPs). One can distinguish
between three well-known types of pickup and delivery problems that have
been studied in the literature. One is the single-commodity PDP in which a
single type of goods is either picked up or delivered at each node (see, e.g.,
Hernández-Pérez and Salazar-González, 2004). This is the case, for exam-
ple, when an armored vehicle transports money between the branch offices
of a bank. Another variant is the two-commodity PDP where two types of
goods are considered and each node may act as both a pickup and a deliv-
ery node (see, e.g., Gendreau et al., 1999; Baldacci et al., 2003). This problem
arises, for instance, in beer or soft drinks delivery where vehicles deliver full
bottles and collect empty ones. A variant of this problem is the VRP with back-
hauls in which all deliveries must be performed before any pickup. Finally, the
n-commodity problem occurs when each commodity is associated with a sin-
gle pickup node and a single delivery node. This is the case when passengers
or goods must be transported from an origin to a destination. This problem is
usually referred to as the VRPPD.
Because most practical applications of the VRPPD include restrictions on
the time at which each location may be visited by a vehicle, it is convenient to
present a slightly more general variant of the problem, called the VRPPD with
Time Windows (VRPPDTW).
Let n denote the number of requests to be satisfied. Assuming that all vehi-
cles are based at a single depot, the VRPPDTW may be defined on a directed
graph G = (N A) where N = P ∪ D ∪ {0 2n + 1}, P = {1     n} and
D = {n + 1     2n} are node sets, and A = {(i j): i j ∈ N} is the arc set.
Subsets P and D contain pickup and delivery nodes, respectively, while nodes
0 and 2n + 1 represent the origin and destination depots. With each request i
are thus associated an origin node i and a destination node n + i. Let K be the
set of vehicles and let m = |K|. Each vehicle k ∈ K has a capacity Qk and the
total duration of its route cannot exceed Tk . With each node i ∈ N are associ-
ated a load qi and a nonnegative service duration di such that q0 = q2n+1 = 0,
qi = −qn+i (i = 1     n), and d0 = d2n+1 = 0. A time window [ei  li ] is also
associated with each node i ∈ N, where ei and li represent the earliest and
432 J.-F. Cordeau et al.

latest time, respectively, at which service may begin at node i. With each arc
(i j) ∈ A are associated a routing cost cij and a travel time tij .
For each arc (i j) ∈ A and each vehicle k ∈ K, let xkij = 1 if and only if
vehicle k travels from node i to node j. For each node i ∈ N and each vehicle
k ∈ K, let Bik be the time at which vehicle k begins service at node i, and Qik be
the load of vehicle k after visiting node i. The VRPPDTW can be formulated
as the following mixed-integer program:

minimize cijk xkij (1)
k∈K i∈N j∈N
subject to

xkij = 1 i ∈ P (2)
k∈K j∈N
 
xkij − xkn+ij = 0 i ∈ P k ∈ K (3)
j∈N j∈N

xk0j = 1 k ∈ K (4)
j∈N
 
xkji − xkij = 0 i ∈ P ∪ D k ∈ K (5)
j∈N j∈N

xki2n+1 = 1 k ∈ K (6)
i∈N
 
Bjk  Bik + di + tij xkij  i ∈ N j ∈ N k ∈ K (7)
 
Qjk  Qik + qj xkij  i ∈ N j ∈ N k ∈ K (8)
Bik + di + tin+i  Bn+i
k
 i ∈ P k ∈ K (9)
k k
B2n+1 − B0  Tk  k ∈ K (10)
ei  Bik  li  i ∈ N k ∈ K (11)
max{0 qi }  Qik  min{Qk  Qk + qi } i ∈ N k ∈ K (12)
xkij ∈ {0 1} i ∈ N j ∈ N k ∈ K (13)
The objective function minimizes the total routing cost. Constraints (2) and (3)
ensure that each request is served exactly once and that the associated pickup
and delivery nodes are visited by the same vehicle. Constraints (4)–(6) guar-
antee that the route of each vehicle k starts at the origin depot and ends at
the destination depot. The consistence of time and load variables is ensured
by constraints (7) and (8). Constraints (9) force the vehicles to visit the pickup
node of a request before its delivery node. Finally, inequalities (10) bound the
duration of each route while (11) and (12) impose time windows and capacity
constraints, respectively.
 Ch. 7. Transportation on Demand 433

The VRPPDTW is NP-hard since it generalizes the Traveling Salesman

Problem (TSP) known to be NP-hard (Garey and Johnson, 1979). In the
presence of time windows, even finding a feasible solution to the problem is
NP-hard since the feasibility problem for the TSP with time windows is itself
NP-complete (Savelsbergh, 1985).
Savelsbergh and Sol (1995) considered a slightly more general formulation
of the pickup and delivery problem and reviewed the relevant literature on the
problem. A more recent survey on pickup and delivery problems was also pre-
pared by Desaulniers et al. (2002). In the remainder of this section, we review
the most important exact and heuristic solution algorithms for the VRPPD
with and without time windows. We first present algorithms for the single-
vehicle case, followed by the multiple-vehicle case.

2.1 The single-vehicle VRPPD

The single-vehicle VRPPD is obtained when m = 1 in formulation (1)–(13).

Although few real-life applications exist for this problem, it may appear as a
subproblem in algorithms for the multiple-vehicle case. It is worth mentioning
that unlike the single-vehicle VRP which necessarily reduces to an (uncapac-
itated) TSP, the single-vehicle VRPPD may incorporate capacity constraints.
Indeed, because both pickup and delivery nodes are considered, any number
of requests may be served by a single vehicle provided that qi  Q for every
request.

2.1.1 Exact algorithms

Kalantari et al. (1985) have presented branch-and-bound algorithms for the
single-vehicle case with finite and infinite vehicle capacity. These algorithms,
which are modifications of the algorithm of Little et al. (1963) for the TSP, work
by eliminating in each branch of the search tree all arcs that would lead to a
violation of a precedence constraint. Fischetti and Toth (1989) have developed
an additive bounding procedure for the more general TSP with precedence
constraints in which some nodes may have one or several predecessors. This
procedure combines the lower bounds obtained from the assignment problem
and shortest spanning 1-arborescence problem relaxations, variable decompo-
sition and disjunctions. A lower bounding procedure and a dynamic program-
ming algorithm were later developed by Bianco et al. (1994) while Balas et al.
(1995) proposed valid inequalities for this problem.
More recently, Ruland and Rodin (1997) introduced a branch-and-cut al-
gorithm for the TSP with Pickup and Delivery (TSPPD). Using the previously
introduced notation, the problem can be formulated on an undirected graph
G = (N E) with binary edge variables xe , e ∈ E.
Given a node set S ⊆ N, denote by E(S) and δ(S), the sets of edges with
both endpoints in S, and with  one endpoint in S and the other  in N \ S, re-
spectively. Let also x(S) = e∈E(S) xe . Similarly, let x(E  ) = e∈E  xe for any
edge set E  ⊆ E. Finally, let Us = {U ⊂ N | 0 ∈ U} and Up = {U ⊂ N |
434 J.-F. Cordeau et al.

0 ∈ U 2n + 1 ∈/ U ∃i ∈ P i ∈
/ U n + i ∈ U}. The formulation can be stated
as follows:

minimize ce xe (14)
e∈E
subject to
 
x {0 2n + 1} = 1 (15)
  
x δ {i} = 2 i ∈ N (16)
 
x δ(U)  2 U ∈ Us  (17)
 
x δ(U)  4 U ∈ Up  (18)
0  xe  1 e ∈ E (19)
|E|
x∈Z  (20)
Constraint (15) simply connects the origin depot to the destination depot and
ensures that the solution is a Hamiltonian cycle. Each node is then required to
have a degree of 2 by constraints (16) while constraints (17) ensure the bicon-
nectedness of the solution. Finally, constraints (18) force the pickup node of
each request to be visited before its delivery node.
As explained by Ruland (1995) a careful analysis of constraints (17) and
(18) reveals that the cardinality of the sets Us and Up can, in fact, be reduced
by exploiting the redundancy of some of the associated constraints. The author
also shows that the resulting subtour elimination constraints and precedence
constraints define faces of the TSPPD polytope.
Two other classes of inequalities were introduced by Ruland (1995). Let
U1      Um ⊂ N be mutually disjoint subsets and let i1      im ∈ P be re-
quests such that 0 2n + 1 ∈ / Ul and il  n + il+1 ∈ Ul for l = 1     m (where
im+1 = i1 ). The following inequality, called a generalized order constraint, de-
fines a proper face of the TSPPD polytope:

m 
m
x(Ul )  |Ul | − m − 1 (21)
l=1 l=1

(Note that similar inequalities were also proposed by Balas et al. (1995) for the
precedence-constrained asymmetric TSP.)
Consider two nodes i j ∈ P and a subset H such that {i j} ⊆ H ⊆ N \ {0
n+i n+j 2n+1}. The following inequality, called an order matching constraint,
also defines a proper face of the TSPPD polytope:
   
x(H) + x {i n + i} + x {j n + j}  |H| (22)
The latter inequality can, in fact, be lifted by considering all requests p for
which p ∈ H and n + p ∈ N \ H. For each type of inequality, Ruland (1995)
described separation algorithms relying on the solution of maximum flow prob-
lems. Computational results were reported on instances with n  15.
 Ch. 7. Transportation on Demand 435

For the single-vehicle VRPPD with time windows and capacity constraints,
Desrosiers et al. (1986) have developed an exact forward dynamic program-
ming algorithm to minimize the total distance traveled. A state (S i) is defined
if there exists a feasible path that starts at the depot 0, visits all nodes in S ⊆ N
and ends at node i ∈ S. For each such state, two-dimensional labels are used to
keep track of the time and distance traveled. A label can be eliminated if there
exists no feasible path that starts at node i and visits all remaining nodes. Com-
putational experiments performed on real-life data with tight time windows
have shown that the algorithm could very quickly solve instances with n  40.

2.1.2 Heuristics
A probabilistic analysis of a simple construction heuristic for the problem
without capacity and time windows was performed by Stein (1978). This heuris-
tic constructs a solution by concatenating two optimal traveling salesman tours:
one through the n origins and one through the n destinations. The author
showed that if the 2n points are drawn independently from the uniform proba-
bility distribution over a subset of the Euclidean plane, then the algorithm has
an asymptotic performance bound of 1.06. Later, Psaraftis (1983a) presented a
worst-case analysis of a two-phase construction heuristic. In the first phase, an
optimal TSP tour is constructed for the 2n points. In the second phase, a solu-
tion to the pickup and delivery problem is obtained by traversing the TSP tour
clockwise until all points are visited. While doing this, points that have already
been visited or that correspond to a destination whose origin has not been vis-
ited should be skipped. Psaraftis showed that if the minimum spanning tree
heuristic of Christofides (1976) is used to construct the TSP tour, the heuristic
has a worst-case performance ratio of 3.0. He also reported computational ex-
periments indicating that on realistic size instances the average performance
of his heuristic was superior to that of Stein’s heuristic. In a related paper,
Psaraftis (1983c) proposed a local search heuristic that extends the TSP inter-
change procedure of Lin (1965) to handle precedence constraints. In addition,
Psaraftis described an approach that identifies the best k-interchange in O(nk )
time. Similar ideas were introduced by Savelsbergh (1990) in the more general
context of constrained routing problems. Healy and Moll (1995) also have de-
scribed a variant of local search for the same problem. Their strategy, called
sacrificing, consists of biasing the search in the direction of solutions with larger
neighborhoods of feasible solutions in the hope of improving the overall qual-
ity of the local optima found.
Renaud et al. (2000) have later described adaptations of some classical TSP
heuristics to handle the pickup and delivery problem. They have also intro-
duced a two-phase method in which the first phase constructs a solution by
means of a double insertion procedure that performs the simultaneous inser-
tion of a pickup node and its associated delivery node, and the second phase
is a deletion and reinsertion improvement procedure based on 4-opt* heuris-
tic of Renaud et al. (1996). Recently, Renaud et al. (2002) have described and
compared several perturbation heuristics whose aim is to help an improve-
436 J.-F. Cordeau et al.

ment algorithm to escape from a local optimum. In particular, they considered

instance perturbation in which the instance data are slightly changed, algorith-
mic perturbation in which the definition of the neighborhood is modified, and
solution perturbation in which a locally optimal solution is perturbed before
restarting the search. Compared with existing heuristics, these perturbation
schemes yielded excellent results on instances with n  220.
For the single-vehicle problem with time windows, Van der Bruggen et al.
(1993) developed a two-phase local search method based on the variable-depth
search of Lin and Kernighan (1973) for the TSP. In the first phase, nodes are
first sorted in increasing order of the middle of their time window. The re-
sulting ordering is then modified so as to ensure that the pickup node of each
request appears before the delivery node and capacity constraints are satisfied.
A solution is then constructed by visiting the nodes in that order. This solution
may violate some of the time windows. Iterative improvements are then per-
formed in the hope of obtaining a feasible solution. This solution is also further
refined by applying the same exchange procedures with a different objective.

2.2 The multiple-vehicle VRPPD

2.2.1 Exact algorithms

Dumas et al. (1991) have proposed a set-partitioning formulation of the
problem and an exact column generation algorithm. For vehicle k ∈ K, let
Ωk be the set of feasible routes and let crk be the cost of route r. In addition
to traditional flow conservation constraints, each route r ∈ Ωk satisfies time
windows, capacity constraints, pairing constraints (i.e., node i ∈ P is visited iff
node n + i ∈ D is also visited by the route), and precedence constraints. For all
i ∈ P and r ∈ Ωk , let akir be a binary constant equal to 1 if request i is served by
route r of vehicle k, and 0 otherwise. Finally, define a binary variable yrk that
takes the value 1 if route r is used for vehicle k, and 0 otherwise. The problem
can be stated as follows:
 
minimize crk yrk (23)
k∈K r∈Ωk
subject to
 
akir yrk = 1 ∀i ∈ P (24)
k∈K r∈Ωk

yrk = 1 ∀k ∈ K (25)
r∈Ωk

yrk ∈ {0 1} ∀k ∈ K r ∈ Ωk  (26)

This formulation is solved by a branch-and-bound method in which the lin-
ear relaxations are solved by column generation. Columns of negative reduced
cost are generated by solving a resource-constrained shortest path problem
 Ch. 7. Transportation on Demand 437

in which the arc costs are modified to reflect the current values of the dual
variables associated with constraints (24) and (25). This problem is solved by
a dynamic programming algorithm which is very similar to the one described
by Desrosiers et al. (1986) for the single-vehicle case. In this case, however,
not all nodes have to be visited by the vehicle. To obtain integer solutions,
branching is performed on additional order variables Oij  i j ∈ P ∪ {0 2n + 1},
indicating the sequence in which pickups are performed. These decisions are
easily transfered to the subproblem and are handled directly by the dynamic
programming algorithm through the introduction of an additional label repre-
senting the last pickup node visited. Several arc elimination rules are proposed
by the authors to reduce the problem size by taking time windows and pairing
constraints into account. For example, arc (i n + j) can be eliminated if the
path j → i → n + j → n + i is infeasible even when setting Bj = ej . The algo-
rithm was successful in solving two real-life instances with 19 and 30 requests,
respectively, as well as randomly generated instances involving up to 55 re-
quests but tight capacity constraints. According to the authors, the algorithm
works well when capacity constraints are restrictive and each route serves a
small number of requests (i.e., five or fewer).
A similar approach was developed by Savelsbergh and Sol (1998). However,
because it was intended to solve large-scale instances, it differs from that of
Dumas et al. (1991) in the following respects:
(i) whenever possible, construction and improvement heuristics are used
to solve the pricing subproblem;
(ii) a sophisticated column management mechanism scheme is used to keep
the column generation master problem as small as possible;
(iii) columns are selected with a bias toward increasing the likelihood of
identifying feasible integer solutions during the solution of the master prob-
lem;
(iv) branching decisions are made on additional assignment variables zik
representing the fraction of request i that is served by vehicle k; and
(v) a primal heuristic is used at each node of the search tree to obtain upper
bounds.
Computational results performed by the authors on instances with n  50 show
that the proposed approach yields high quality solutions in short computing
times even when capacity constraints are not very tight.
Very recently, another column generation method was used by Xu et al.
(2003) to address a complex pickup and delivery problem encountered in
long-haul transportation planning. In their problem, there are multiple car-
riers and multiple-vehicle types available to cover a set of pickup and delivery
requests, each of which has multiple pickup time windows and multiple deliv-
ery time windows. In addition to the classical vehicle capacity, route duration,
pairing and precedence constraints, vehicle routes must satisfy compatibility
constraints between the requests, the carriers, and the vehicle types, as well as
sequencing constraints requiring the goods to be collected and delivered in a
438 J.-F. Cordeau et al.

last-in first-out sequence, i.e., the goods picked-up last must be the first to be
delivered. Constraints regarding maximum driving time and maximum working
time are also taken into account, leading to a complex objective function incor-
porating fixed costs, mileage costs, waiting costs, and layover (driver rest) costs.
This problem is solved by means of a column generation approach in which the
pricing subproblems are solved by fast heuristics. Instead of embedding the col-
umn generation method in a branch-and-bound process, the method reaches
an integer solution by applying an IP solver to the restricted set of columns
generated in solving the linear relaxation of the problem. Comparisons with
lower bounds obtained by solving the LP relaxation of the problem exactly
through the use of dynamic programming for the pricing subproblem show
that the heuristic approaches are capable of generating near-optimal solutions
quickly for randomly generated instances with up to 200 requests. Results are
also reported on larger instances involving 500 requests.

2.2.2 Heuristics
A tabu search heuristic for the pickup and delivery problem with time win-
dows was developed by Nanry and Barnes (2000). Solutions that violate time
window and vehicle capacity constraints are allowed during the search. These
authors have considered three types of move. The first removes a node pair
(i n + i) from its current route and reinserts it in a different route. The sec-
ond swaps two pairs of nodes between two distinct routes. The last consists of
moving a single node within its current route. A hierarchical search mecha-
nism is used to dynamically alternate between these neighborhoods according
to problem difficulty. Computational results are reported on random instances
involving up to 100 requests. A similar tabu search heuristic was also developed
by Lau and Liang (2002).
More recently, Ropke and Pisinger (2004) introduced an adaptive Large
Neighborhood Search (LNS) heuristic for the VRPPD with time windows.
Instead of relying on operators that perform minor changes to the solution
in every iteration, the LNS heuristic uses large moves that can potentially
rearrange up to 30–40% of all requests in a single iteration. This is accom-
plished by using several large neighborhoods in an adaptive way. Three re-
moval heuristics are considered: removing random requests, removing similar
requests likely to be interchangeable in the solution, and removing requests
whose removal yields a large decrease in solution cost. To reinsert these re-
quests, two insertion heuristics are used: a greedy heuristic that inserts at each
iteration the request with the least insertion cost, and a regret heuristic that
takes into account the cost of not being able to insert a request in its least-cost
route. The selection of removal and insertion heuristics is performed randomly
at each iteration according to a probability distribution whose weights are ad-
justed dynamically throughout the search. In experiments performed on test
instances with n  500, the adaptive LNS algorithm outperformed previously
proposed heuristics.
 Ch. 7. Transportation on Demand 439

3 The dial-a-ride problem

The Dial-a-Ride Problem (DARP) is a generalization of the VRPPD arising

in contexts where passengers are transported, either in groups or individually,
between specified origins and destinations. The most common DARP applica-
tion arises in door-to-door transportation services for elderly or handicapped
people. In this context, users often formulate two requests per day: an out-
bound request from home to a destination, and an inbound request for the
return trip. The DARP distinguishes itself from the basic VRPPD by its focus
on controlling user inconvenience. This usually takes the form of constraints
or objective function terms relating to waiting time, ride time (i.e., the time
spent by a user in the vehicle) as well as deviations from desired departure and
arrival times.
In the remainder of this section, we first discuss the scheduling aspect of the
problem and then present the most important exact and heuristic algorithms
for the static single-vehicle and multiple-vehicle cases, respectively. This is fol-
lowed by the dynamic case in the last section. An overview of some of these
methods can also be found in the Cordeau and Laporte (2003a) survey.

3.1 Scheduling

Because of the focus on controlling user inconvenience, the scheduling as-

pect of the problem plays a central role in most applications. As a result, the
problem of finding an optimal schedule for a given vehicle route has been stud-
ied independently in the literature. In general terms, given a sequence of nodes
i1  i2      iq to be visited, the problem of finding an optimal schedule satisfying
time windows can be formulated as

q
minimize gi (Bi ) (27)
i=1
subject to
Bi + di + tii+1  Bi+1  ∀i = 1     q − 1 (28)
ei  Bi  li  ∀i = 1     q (29)
where gi (Bi ) is a convex function defined with respect to the time window
[ei  li ]. Sexton and Bodin (1985a, 1985b) observed that some special cases of
this scheduling problem can be seen as network flow problems and thus solved
very efficiently. Dumas et al. (1989) proposed a dual approach to solve the
general problem by performing q unidimensional minimizations. In the special
cases where the inconvenience functions are quadratic or linear, the complexity
of the algorithm is O(q). The extension of this methodology to handle time-
varying, stochastic travel times was addressed by Fu (2002).
On a related topic, Hunsaker and Savelsbergh (2002) have devised a pro-
cedure for efficiently testing the feasibility of an insertion in construction or
440 J.-F. Cordeau et al.

improvement heuristics. They considered a variant of the DARP with time

windows, waiting time constraints and ride time constraints, and showed how
to check in O(q) time whether the insertion of a given request in a route is
feasible. Cordeau and Laporte (2003b) also proposed a procedure, based on
the forward time slack notion introduced by Savelsbergh (1992), to sequentially
minimize time window constraint violations, route durations, and ride times in
the context of local search.

3.2 The static single-vehicle DARP

Early work on the single-vehicle dial-a-ride problem was carried out by

Psaraftis (1980) who studied the “immediate-request” case in which a list of
requests should be served as soon as possible. His model assumes that no time
windows are specified by the users. Instead the transporter imposes “maxi-
mum position shift” constraints limiting the difference between the position
of a request in the calling list and its position in the vehicle route. The objec-
tive function aims to minimize the sum of route completion time and customer
dissatisfaction. Customer dissatisfaction is itself expressed as a weighted com-
bination of waiting time before pickup and ride time. The problem is solved
using a dynamic programming algorithm in which the state space consists of
vectors (L k1      kn ), where L denotes the node being currently visited and
ki denotes the status of request i. The status of request i is either 3 if user i
has been dropped off, 2 if the user is still in the vehicle or 1 if the user has yet
to be picked up. The complexity of this algorithm is O(n2 3n ) and only small
instances can thus be solved. Psaraftis also explains how to handle the dynamic
case in which new requests occur dynamically in time but no information on
future requests is available. In this context, maximum position shift constraints
become essential to prevent a request from being indefinitely deferred. In a
later paper, Psaraftis (1983b) extended his approach to handle time windows
on departure and arrival times. The new algorithm has the same complexity as
the previous one but uses forward instead of backward recursion.
A heuristic approach based on Benders decomposition was later developed
by Sexton and Bodin (1985a, 1985b) who considered one-sided time windows
on delivery. Their algorithm iterates between a routing master problem and
a scheduling subproblem. The routing problem relaxes the Benders cuts in
the objective function and is solved by a route improvement procedure. The
scheduling subproblem is shown to be the dual of a network flow problem
that can be solved very quickly. These authors minimize a user inconvenience
function made up of the weighted sum of two terms. The first measures the dif-
ference between the actual travel time and the direct travel time of a user. The
second term is the (positive) difference between desired drop-off time and ac-
tual drop-off time, under the assumption that the former is at least as large as
the latter, late drop-offs being disallowed. The approach was tested on real-life
data sets where the number of users varies between 7 and 20.
 Ch. 7. Transportation on Demand 441

3.3 The static multiple-vehicle DARP

One of the first heuristics for the multiple-vehicle DARP was proposed by
Jaw et al. (1986) who impose windows on the pickup times of inbound re-
quests and on the drop-off times of outbound requests. A maximum ride time,
expressed as a linear function of the direct ride time, is given for each user. In
addition, vehicles are not allowed to be idle when carrying passengers. A non-
linear objective function combining several types of disutility is used to assess
the quality of solutions. The authors have developed an insertion heuristic that
selects users in order of earliest feasible pickup time and gradually inserts them
into vehicle routes so as to yield the least possible increase in the objective
function. The algorithm was tested on artificial instances involving 250 users
and on a real data set with 2617 users and 28 vehicles. This approach was also
adapted by Alfa (1986) and applied to a practical case in Winnipeg, Canada.
Several of the heuristics proposed for the multiple-vehicle case are two-
phase algorithms in which the first phase creates and selects clusters of users
that are then combined into vehicle routes in the second phase. An early ap-
proach based on this idea is the interactive optimizer described by Cullen et al.
(1981) for the case of a homogeneous fleet. Another clustering method was
proposed by Bodin and Sexton (1986). Their heuristic creates clusters, ap-
plies the single-vehicle algorithm of Sexton and Bodin (1985a, 1985b) to each
cluster and then moves users between clusters so as to reduce total user incon-
venience. It was applied to real-life instances involving approximately 85 users
each.
Dumas et al. (1989) later improved upon this methodology by creating so-
called “mini-clusters” of users, i.e., groups of users to be served within the
same area at approximately the same time. Users in a mini-cluster should be
transportable by a single vehicle while respecting constraints on time win-
dows, vehicle capacity, pairing, and precedence. The mini-clusters are then
optimally combined to form feasible vehicle routes, using column generation.
In this phase, a time window is imposed on each cluster to ensure feasibility
and columns are generated by solving a constrained shortest-path problem.
Finally, each vehicle route is optimized by means of the single-vehicle algo-
rithm of Desrosiers et al. (1986) and a scheduling step is executed to minimize
user inconvenience (Dumas et al., 1990). Instances with up to 200 users are
easily solved, while larger instances require the use of a spatial and tempo-
ral decomposition technique. The mini-clustering phase was later improved by
Desrosiers et al. (1991) who described a parallel insertion method that relies
on the notion of neighboring requests. Two requests are said to be neighbors if
they satisfy the following conditions:
1. ei  ej  ln+i or ei  lj+n  ln+i or ej  ei  ln+i  ln+j ;
2. tij + tjn+i  αtin+i or tji + tin+j  αtjn+j with α > 1;
3. |θi − θj |  β where θi is the angle between a reference axis and the
direction from i to n + i;
442 J.-F. Cordeau et al.

4. s(i j)  γ where s(i j) are the savings in distance obtained by clustering

requests i and j together.
Clusters are then constructed in parallel by treating the requests in decreas-
ing order of the direct duration tin+i , and considering, at each iteration, the
creation of a new cluster or the insertion of a request in clusters that contain
at least one neighboring request. Finally, Ioachim et al. (1995) showed that
there is an advantage in terms of solution quality to use an optimization tech-
nique for the construction of the clusters. Results were reported on data sets
comprising more than 2500 users.
Another study, by Borndörfer et al. (1999), also uses a two-phase approach
in which clusters of users are first constructed and then grouped together to
form feasible vehicle routes. A cluster is defined as a “maximal subtour such
that the vehicle is never empty”. In the first phase, a large set of good clusters
are constructed and a set partitioning problem is then solved to select a subset
of clusters serving each user exactly once. In the second phase, feasible routes
are enumerated by combining clusters and a second set partitioning problem is
solved to select the best set of routes covering each cluster exactly once. Both
set partitioning problems are solved by a branch-and-cut algorithm. On real-
life instances, the algorithm cannot always be run to completion and terminates
with the best known solution. It was applied to several instances provided by an
operator in Berlin and including between 859 and 1771 transportation requests
per day.
In another real-life application, Toth and Vigo (1996) considered a problem
in which users specify requests with a time window on their origin or destina-
tion. An upper bound proportional to direct distance is imposed on the ride
time. Transportation is supplied by a fleet of capacitated minibuses and by the
occasional use of taxis. The objective is to minimize the total cost of service.
The authors have developed a heuristic consisting of first assigning requests
to routes by means of a parallel insertion procedure, and then performing
intra-route and inter-route exchanges. Tests performed on instances involving
between 276 and 312 requests show significant improvements with respect to
the previous hand-made solutions. Further improvements were later obtained
by Toth and Vigo (1997) through the execution of a tabu thresholding post-
optimization phase after the parallel insertion step.
More recently, Cordeau and Laporte (2003b) have developed a tabu search
heuristic for the problem in which users specify a desired arrival time for their
outbound trip and a desired departure time for their inbound trip, and a max-
imum ride time is associated with each user. Capacity and maximum route
duration constraints are also imposed on the vehicles. The search algorithm is
based on a simple mechanism that iteratively removes a request from its cur-
rent route and reinserts it into another route. As is common in such contexts
(see, e.g., Cordeau et al., 1997), intermediate infeasible solutions are allowed
during the search, but they are discouraged by a penalty term in the objective
function. As explained in the previous section, whenever the cost of an ex-
change is evaluated, the schedules of the two routes involved in the exchange
 Ch. 7. Transportation on Demand 443

must be updated so as to measure the impacts on violations of time windows,

route duration constraints and ride time constraints. The algorithm was tested
on randomly generated instances with up to 144 users and on six data sets with
n = 200 or n = 295 provided by a Danish transporter. Parallel implementa-
tions of this heuristic have also been described by Attanasio et al. (2004) who
studied the case where a fraction of the requests are received dynamically.
Finally, Cordeau (2006) developed a branch-and-cut algorithm for the same
version of the problem. This algorithm relies on several types of valid in-
equalities that are either new inequalities for the problem or adaptations of
known inequalities for the TSP, the TSPPD, or the VRP. The first four types
of inequalities are liftings of subtour elimination constraints for the symmet-
ric and asymmetric TSP. Consider the simple subtour elimination constraint
x(S)  |S| − 1 for S ⊆ P ∪ D. In the case of the DARP, this inequality can
be lifted in two different ways by taking into account the fact that for each
user i, node i must be visited before node n + i. For any set S ⊆ P ∪ D, let
π(S) = {i ∈ P | n + i ∈ S} and σ(S) = {n + i ∈ D | i ∈ S} denote the sets of
predecessors and successors of S, respectively. Balas et al. (1995) have proposed
two families of inequalities for the precedence-constrained asymmetric TSP
that also apply to the DARP by observing that each node i ∈ P ∪ D is either
the predecessor or the successor of exactly one other node. For S ⊆ P ∪ D, the
following inequality, called a successor inequality (or σ-inequality) is valid for
the DARP:
   
x(S) + xij + xij  |S| − 1 (30)
i∈S∩σ(S) j∈S i∈S\σ(S) j∈S∩σ(S)

Similarly, for any set S ⊆ P ∪ D, the following predecessor inequality (or

π-inequality) is valid for the DARP:
   
x(S) + xij + xij  |S| − 1 (31)
i∈S j∈S∩π(S) i∈S∩π(S) j∈S\π(S)

Directed subtour elimination constraints proposed by Grötschel and Pad-

berg (1985) for the asymmetric TSP can also be lifted in a similar fashion,
yielding two other valid inequalities.
Generalized order constraints, introduced by Ruland (1995) for the TSPPD
(see inequalities (21)), can also be lifted in two ways as follows:

m 
m−1 
m 
m
x(Ul ) + xi1 il + xi1 n+il  |Ul | − m − 1 (32)
l=1 l=2 l=3 l=1

m 
m−2 
m−1 
m
x(Ul ) + xn+i1 il + xn+i1 n+il  |Ul | − m − 1 (33)
l=1 l=2 l=2 l=1
Finally, ride time constraints may give rise to paths that are infeasible in
an integer solution but nonetheless feasible in a fractional solution. Forbid-
ding such paths can be accomplished as follows. For any directed path P =
444 J.-F. Cordeau et al.

{i k1  k2      kp  n + i} such that tik1 + dk1 + tk1 k2 + dk2 + · · · + tkp n+i > L
the following inequality is valid for the DARP:
p−1

xik1 + xkh kh+1 + xkp n+i  p − 1 (34)
h=1
Heuristic separation algorithms are proposed for each type of valid inequality.
In addition, several techniques are described to strengthen the formulation
and reduce problem size. In computational experiments, the branch-and-cut
algorithm was able to solve instances with up to four vehicles and 32 users.

3.4 The dynamic DARP

While early studies on the DARP were often motivated by dynamic settings,
the dynamic DARP has received less attention in the literature than its sta-
tic counterpart. An approach inspired by the work of Jaw et al. (1986) was
developed by Madsen et al. (1995) for a real-life problem involving services
to elderly and disabled people in Copenhagen. Users may specify a desired
pickup or drop-off time window, but not both. Vehicles of several types are
used to provide service, not all of which are available at all times. In addi-
tion, some requests arrive dynamically throughout the day. New requests are
inserted in vehicle routes taking into account their difficulty of insertion into
an existing route. The algorithm was tested on a 300-customer, 23-vehicle in-
stance and the authors report that it was capable of generating good quality
solutions within very short computing times.
At about the same time, Dial (1995) introduced the concept of an Au-
tonomous Dial-a-Ride Transit (ADART) service based on fully automated
command-and-control, order-entry, and routing and scheduling systems imple-
mented on computers on-board vehicles. Here, the system is fully automated:
the only human intervention in the process is the customer requesting service.
Furthermore, routing-and-scheduling is not done at some central dispatching
center, but is rather distributed among vehicles through an auction mechanism.
Teodorovic and Radivojevic (2000) have later studied a generic version of
the dynamic dial-a-ride problem using fuzzy logic. Their approach exploits the
fact that passengers, dispatchers and drivers have a fuzzy notion of travel times,
which can thus be expressed with fuzzy sets and numbers. Through fuzzy arith-
metic, calculations about arrival times at customers, waiting times, etc. are
performed and the qualitative results are provided to different approximate
reasoning algorithms to decide about the assignment and insertion of a new
request in a vehicle route.
A software system for demand-responsive passenger services, like variably
routed buses, conventional and maxi-taxis, was proposed by Horn (2002). The
optimization capabilities of the system are based on least-cost insertions of
new requests and periodic reoptimization of the planned routes. The latter is a
steepest descent approach using a neighborhood structure which either moves
 Ch. 7. Transportation on Demand 445

or swaps customers. A so-called “rank-homing” heuristic is also proposed for

governing the relocation of idle vehicles. A set of locations, known as cab-
ranks, are specified in advance and the heuristic chooses the cab-rank where
the idle vehicle should be dispatched. To take a decision, the heuristic exploits
information about future patterns of demand at each cab-rank.
Finally, Coslovich et al. (2003) have addressed a dial-a-ride problem where
people might unexpectedly ask a driver for a trip at a given stop. Clearly, these
requests must immediately be accepted or rejected. In order to accommodate
them, a neighborhood of solutions is generated off-line by considering differ-
ent perturbations to the current planned routes. Using this neighborhood of
solutions, more insertion opportunities are available and can be quickly evalu-
ated when an unexpected customer asks for service. Whenever a new customer
is accepted, both the current solution and the neighborhood must be updated,
but this computationally demanding task can be done while the vehicles are
moving from one stop to the next. In that case, the time pressure is much less
stringent, when compared with the almost immediate response required by un-
expected customers.

4 Urban courier service problems

Every large city has a number of courier companies serving pickup and deliv-
ery requests for the transportation of letters and small parcels. These requests
occur continuously during the day and only a small fraction of these are known
in advance (typically, those requests that have been received the previous day,
but too late for immediate service). The distinctive features of Urban Courier
Service Problems (UCSPs) with regard to the other problems presented in this
chapter are their inherent dynamic nature and the absence of capacity con-
straints, due to the small size of letters and parcels.
In these problems, each request is characterized by a pickup and a deliv-
ery location, plus a time window for service. A usually fixed fleet of vehicles is
available to service the requests. When a new request occurs, it is dispatched
and inserted in a least-cost fashion in the planned route of one vehicle. This
cost typically relates to the total distance traveled by the vehicles plus a penalty
for lateness when the vehicle arrives at a location after its time window’s upper
bound (a vehicle can arrive before the lower bound but, in that case, it must
wait up to the lower bound to start its service). With a fixed size fleet, some
requests received during the day may remain unserviced. This happens, for ex-
ample, when drivers must be back at some location before a given deadline at
the end of the day, when the upper bounds of the time windows are strictly en-
forced, or when the incremental cost for servicing a request exceeds a tolerance
threshold. In practice, these unserviced requests will be serviced the next day
or will be referred to alternative transportation means, including competitors.
This type of problem has not been studied much in the literature. In partic-
ular, we are not aware of any exact methods for solving them and only a few
446 J.-F. Cordeau et al.

heuristics are reported. Furthermore, the only dynamic aspect of the problem
that has been considered is the occurrence of new requests, although other as-
pects are certainly of interest like dynamic travel times. In the following, the
main algorithms for the UCSP are reviewed.
The single-vehicle dynamic pickup and delivery problem (without time win-
dows) was analyzed by Swihart and Papastravou (1999) under different routing
policies and demand intensities. This paper can be seen as an extension of the
work of Bertsimas and van Ryzin (1991) on the Dynamic Traveling Repairman
Problem (DTRP) with single-point customer requests. As the objective is to
minimize the time spent in the system, this work is at the interface of vehi-
cle routing and queuing theory. Dynamic contexts with both unit-capacity and
multiple-capacity vehicles are analyzed.

4.1 A tabu search heuristic for the UCSP

A natural heuristic approach for the UCSP is to use a cheapest insertion

criterion for new requests. Gendreau et al. (1998) go beyond this principle and
reoptimize the planned routes with a tabu search heuristic. The neighborhood
structure is based on ejection chains (Glover, 1996), where the pickup and
delivery locations of a request are taken from one route and moved to another
route, forcing a request from that route to move to yet another route, and
so on. The chain may be of any length and may be cyclic or not. The “best”
ejection chain is obtained by solving a constrained shortest path problem. The
tabu search heuristic also integrates an adaptive memory (Rochat and Taillard,
1995), which combines high quality solutions to produce new ones.
Due to real-time requirements, different computational techniques are pro-
posed to speed up the neighborhood evaluation. In addition, a parallel im-
plementation is developed where a master processor distributes the workload
among the slaves that run the tabu search processes. A two-level parallelization
scheme is used. First, several tabu search threads run in parallel and commu-
nicate through the common adaptive memory: they all feed the memory with
new improved solutions and get starting solutions from it. Second, each solu-
tion (a set of planned routes) within a search thread is partitioned into subsets
of routes, and a different tabu search process is associated with each subset.
This is a form of intensification, as it allows the search to focus on restricted
parts of the solution, while reducing the computational effort as a side effect.
The tabu search heuristic runs between the occurrence of new events. At
each input update, it solves the static problem associated with known requests.
Because it does not account for future requests, it can be characterized as a
“myopic” problem-solving approach. Two different types of event are consid-
ered: the occurrence of new service requests, which are truly dynamic events,
and the completion of service at customer locations (which can be determined
from the current routes, due to deterministic travel times). When a new re-
quest is received, the tabu search processes are stopped and their best solution
is sent to the master for possible inclusion in the adaptive memory. The new
 Ch. 7. Transportation on Demand 447

request is then inserted at least cost in each solution contained in the adaptive
memory. Once updated, the memory feeds the tabu search processes with new
starting solutions. A similar procedure is applied when service is completed at
a given location. In that case, the best solution in the adaptive memory is used
to identify the vehicle’s next destination, and the other solutions in the memory
are updated accordingly.
A simulator was developed to test the algorithm under realistic scenarios
with up to 30 requests per hour. The computational results demonstrated the
superiority of the tabu search heuristic for handling new requests, when com-
pared with more straightforward approaches, like simple insertion heuristics.
It is thus useful to optimize the planned routes with sophisticated procedures,
even when the optimization does not account for future requests. Although this
is not explicitly mentioned by Gendreau et al. (1998), the benefits mostly arise
from the early portion of the planned route (as opposed to the later portion
which is likely to be modified with the arrival of new requests). This observa-
tion naturally leads to the work described in the next subsection.

4.2 A double-horizon strategy for the UCSP

Mitrović-Minić et al. (2004) have proposed a double-horizon strategy for

solving a variant of the UCSP. In this work, the number of vehicles is a free vari-
able, thus allowing all requests to be serviced. Furthermore, each request must
be visited before a deadline, which means that the objective reduces to mini-
mizing the total distance traveled. The authors propose a generalization of the
short-term rolling horizon approach (where only requests with a time window
sufficiently close to the current time are assigned to planned routes (Psaraftis,
1988)). Both a short-term and a long-term planning horizon are considered.
The latter is introduced to alleviate the adverse long-term effects of appar-
ently good short-term decisions. Basically, the idea is to associate a different
objective with each horizon type. The objective associated with the short term
horizon is the true objective (i.e., total distance traveled), while the objective
associated with the long-term horizon is aimed at introducing large waiting
times in the routes to favor the insertion of future requests. Each objective
is optimized with a simplified version of the tabu search heuristic reported in
Section 4.1. The computational results obtained on instances generated from
data collected from two courier companies, operating in Vancouver, Canada,
demonstrate the benefits of the double-horizon approach when compared to a
single horizon approach. Mitrović-Minić and Laporte (2004) have further ana-
lyzed different ways of introducing waiting times when scheduling the planned
routes. They showed that mixed waiting strategies provide better results. The
best approach partitions a planned route into segments made of close loca-
tions. Within a segment, the vehicle always departs as soon as possible from its
current location; but when it is time to cross a boundary between two segments
to travel further, the vehicle waits at its current location for a fraction of the
time available up to the latest possible departure time.
448 J.-F. Cordeau et al.

4.3 Adaptive methods for the UCSP

The approaches reported in this subsection exploit the knowledge accumu-

lated, over the years, by expert human dispatchers to make good dispatching
decisions. They can be considered as learning or adaptive methods. In the work
of Shen et al. (1995), a neural network model learns to assign requests to vehi-
cles by automatically adjusting itself to a sample of decisions previously made
by an expert (see Rumelhart et al., 1986, for an introduction to feedforward
neural networks and the backpropagation learning algorithm). After training
with 140 dispatching scenarios obtained from a small courier company oper-
ating in Montreal, and for which expert decisions were known, the network
was able to take good dispatching decisions on other sets of previously unseen
scenarios.
Leclerc and Potvin (1997) proposed a linear utility function that integrates
the main decision variables considered by expert dispatchers when they make
decisions. The decision variables within the utility function are then weighted
with a genetic algorithm (Holland, 1992). Basically, different sets of weights
“evolve” through genetic mechanisms, with the objective of matching as closely
as possible a sample of decisions previously taken by an expert. Benyahia and
Potvin (1998) extended this work by evolving nonlinear utility functions, using
a genetic programming framework (Koza, 1992).

5 The dial-a-flight problem

Almost 3000 businesses in the United States provide on-demand air char-
ter services (certificated by the FAA as Part 135 on-demand air charter). The
majority of companies in the industry are small businesses regulated by the
FAA with similar oversight to that given to the large scheduled airlines. The
on-demand air charter industry provides a vital transportation link for med-
ical services, important cargo needed to promote commerce, and personal
travel supporting the growth of the economy. These companies use smaller
aircraft to meet the customized needs of the traveling public for greater flexi-
bility in scheduling and access to almost every airport in the country. Flights
are planned according to the customer’s schedule, not the operator’s. On-
demand air charter serves commerce across the country and the world by
providing short notice delivery of parts, important documents, supplies, and
other valuable cargo. On-demand air charter also saves lives, since air ambu-
lances transport critically ill or injured patients to hospitals and trauma centers
that can provide the necessary care, and transport vital organs for those requir-
ing transplants. All of these services are contingent upon the ability to respond
quickly to the needs of customers.
Recently there has been an increased interest in the passenger service sec-
tor of the on-demand air charter industry. This is, in part, due to the changes
taking place in the commercial airline industry. Increased security at airports
 Ch. 7. Transportation on Demand 449

has resulted in longer waiting times, with the associated frustrations, and thus
longer travel times. Furthermore, due to the huge losses suffered by the air-
lines in recent years (airlines worldwide have lost $25 billion and more than
400,000 jobs in 2002 and 2003), airlines have cut back and are operating with
a reduced schedule, affecting the flexibility of the business traveler, especially
when it concerns smaller regional airports. At the same time, technological ad-
vances are paving the way for the development of cheaper jet airplanes. For
example, the Eclipse 500, a six-seat, single-pilot state-of-the art jet will sell for
about $1 million – about one-quarter of the price of the cheapest business jets
made today. As a consequence, the idea of an air taxi service, providing effi-
cient, hassle-free, affordable, on-demand air transportation, is no longer just
fiction; it is rapidly becoming a reality. In fact, an air taxi already exists today.
Since April 2002, SkyTaxi, Inc. (http://www.skytaxi.com) has been providing on-
demand air transportation in the northwestern United States. Some passengers
pay a premium to fly direct and by themselves; others receive a discount by
agreeing to allow stops to pick up or drop off other passengers.
There are obvious advantages to an air taxi system. More than 1.5 million
people board commercial airliners each day. Most fly with hundreds of other
passengers on jumbo jet airplanes to and from a limited number of major
“hub” airports which are heavily congested and often located many miles from
their homes and final destinations. Missed connections and flight delays add
to their frustrations. An air taxi system gives travelers the option of hopping
aboard small jets that fly to and from less congested outlying airports, without
packed parking lots, long lines at security checkpoints, flight delays, and lost
luggage, that are closer to where they live and where they want to go. While
commercial flight service now exists at only about 550 airports in the United
States, air taxis will be able to land at 10,000 of the nation’s 14,000 public and
private runways. And all that at competitive fares.
Even though many characteristics of an air taxi service are similar to those
of the common taxi cab service, there are also some fundamental differences.
Requests for service are typically placed farther in advance, a day to two days
in advance as opposed to minutes to an hour. This gives more time to opti-
mize a flight schedule. Usually, a request for service involves several origin-
destination pairs, because business travelers in the end want to return home.
Furthermore, the set of possible pickup, drop-off, and transit locations is rela-
tively small and known in advance. An air taxi service operates out of a given
set of airports, which implies a flight schedule optimizer can exploit the fixed
transportation network structure. Finally, the FAA imposes strict rules on pi-
lot flying and duty hours as well as on aircraft maintenance, which all affect
scheduling flexibility.
Consequently, the traditional dial-a-ride problem is insufficient as an ab-
stract representation for studying, analyzing, and developing decision technol-
ogy for air taxi services. In this section, we introduce the Dial-a-Flight Problem
(DAFP). As air taxi services are a new phenomenon, there is little or no lit-
450 J.-F. Cordeau et al.

erature on dial-a-flight problems, so we will focus on problem definition and

optimization challenges.

5.1 Problem definition

A key challenge in setting up an air taxi system is the development of a

scheduling engine that takes requests for transportation and schedules planes
and pilots in a cost effective way to satisfy these requests. The static and
dynamic DAFPs defined below capture the essential characteristics of the
scheduling problems encountered by an air taxi service.

5.1.1 The static dial-a-flight problem

The Static Dial-a-Flight Problem (SDAFP) is concerned with the schedul-
ing of a set P of n single passenger requests for air transportation during a
single day. A request i ∈ P specifies an origin airport oi , an earliest acceptable
departure time ei at oi , a destination airport di , and a latest acceptable arrival
time li at di . A request i results in a revenue of ri . A fleet F of m identical
airplanes with capacity Q and operable by a single pilot is available to provide
the requested air transportation. Each airplane j ∈ F has a home base Bj , is
available between Ej and Lj , and returns to its home base at the end of the
day. A set P of pilots, stationed at the home bases of the airplanes, are avail-
able to fly the airplanes. A pilot departs from the home base at the start of his
duty and returns to his home base at the end of his duty. A pilot schedule has
to satisfy FAA regulations governing flying hours and duty period, i.e., a pi-
lot cannot fly more than 8 hours in a day and his duty period cannot be more
than 14 hours. It takes an airplane tuv time to fly from airport u to v (over
a distance duv ) and a cost cuv is incurred when doing so. To ensure acceptable
service a passenger itinerary will involve at most two flights, i.e., a single inter-
mediate stop is allowed. The turnaround time at an airport, i.e., the minimum
time between an arrival at an airport and the next departure, is 30 minutes.
The objective is to maximize the profit, i.e., revenues minus costs, while satis-
fying all requests. (Note that because all requests have to be satisfied, in this
variant of the problem the revenues are fixed and the objective is to minimize
the costs.) A dispatcher has to decide which planes and pilots to use to satisfy
the requests and what the plane and pilots itineraries will be, i.e., the flight legs
and associated departure times.
Several restrictions in the above problem definition represent business deci-
sions rather than physical limitations. For example, the limit on the duration of
a passenger trip from origin to destination is a business rule, set, for example,
to twice the direct flying time. This constraint which is similar to the maximum
ride time in the DARP reflects a trade-off between scheduling flexibility and
customer service. Intuitively, an important factor contributing to profitability
for an air taxi service provider is a high plane utilization. This can be ac-
complished by having an airplane satisfy multiple requests at the same time.
Therefore, an air taxi service provider may “encourage” passengers to accept a
 Ch. 7. Transportation on Demand 451

trip that involves an intermediate stop (to pickup or drop off other passengers).
Note that the turnaround time also affects flexibility, and therefore profitabil-
ity. Another business rule, which has not been made explicit in the problem
definition, is whether or not to allow passengers to change planes at an inter-
mediate stop. Again, allowing this will increase the scheduling flexibility, but
it may have a negative impact on customer service because of potential de-
lays.

5.1.2 The dynamic dial-a-flight problem

In the Dynamic Dial-a-Flight Problem (DDAFP) the set of requests for air
transportation arrives over time and each time a request arrives one must im-
mediately decide whether it is feasible to accept the request given the available
resources and the commitments already made. In addition, if it is feasible to ac-
cept the request, one may also want to decide whether it is desirable to accept
the request, i.e., whether it will increase profit. The latter decision is especially
complex as it depends on the requests that will arrive in the future.
The simplest variant of the DDAFP is to construct a schedule of flights for
a specific day in the not too distant future. Requests for transportation on that
particular day arrive in real-time and are considered up to a certain cut-off
time, which precedes the actual execution of the planned schedule. The rule is
to accept each request if there is available capacity. In the DDAFP, one needs
to accommodate bundles of requests, as customers typically request not one,
but two or more flights, and ultimately want to return to their point of ori-
gin. Clearly, all these flight requests must be accepted as a bundle, rather than
individually. Note that all requests are received prior to the execution of any
flight schedule. A more complex variant incorporates “same day travel” ser-
vice, where requests can arrive during the execution of a flight schedule and
have to be incorporated into the schedule.
The real-time, online nature of the booking process is not specific to the air
taxi service, but a common feature of many transportation problems. The lit-
erature on dynamic and stochastic routing problems is rapidly expanding. For
a discussion of many of the issues related to accept/reject decisions in trans-
portation problems, see Bent and Van Hentenryck (2004) and Campbell and
Savelsbergh (2003, 2005).

5.2 Solution approaches for dial-a-flight problems

As mentioned earlier, little or no literature exists on dial-a-flight problems.

Therefore, we present a natural integer programming formulation for SDAFP
and we discuss some pertinent issues concerning solution approaches for the
DDAFP.

5.2.1 A model for the static dial-a-flight problem

In this subsection, we present a time-discretized multicommodity network
flow model for one variant of the problem. We assume the company providing
452 J.-F. Cordeau et al.

the air taxi service has decided to operate each of its planes with two pilot shifts
per day, a morning shift and an afternoon shift. The plane has to return to its
home base to switch pilots some time in the early afternoon. The shift lengths
are chosen so that the limit on pilot duty hours is automatically satisfied. Fi-
nally, it is assumed that for the expected demand distribution it is unlikely that
the limit on the number of flying hours of the pilots is violated, and therefore,
pilot constraints are not explicitly incorporated in the model.
The time horizon is discretized into time periods of several minutes. Let
T be the set of time periods in the planning horizon and let A be the set of
airports. Define the following decision variables:

1 if passenger i departs from airport u to airport v
xiuvt = at time t
0 otherwise

yuvt = 1 if airplane j departs from airport u to airport v at time t
j
0 otherwise
For each airplane j, let Rj and Sj define the start and end of the period during
which the pilot switch needs to take place.
The mathematical formulation of SDAFP can now be written as follows
(where parameters have been converted to their time period equivalents when-
ever necessary):
 j
minimize cuv yuvt
j u u=v t

subject to

xiuvt = 1 i ∈ P u = oi  t = ei  (35)
v

xivu(t−tvu ) = 1 i ∈ P u = di  t = li  (36)
v
 
xiuu(t−1) + xivu(t−tvu ) − xiuvt − xiuut = 0
v=u v=u
i ∈ P u ∈ A t = ei + 1     li − 1 (37)

xiuvt  2 i ∈ P (38)
u v=u t
  j
xiuvt − Qyuvt  0 u ∈ A v ∈ A u = v t ∈ T (39)
i j
 j
yuvt = 1 j ∈ F u = Bj  t = Ej  (40)
v
 j
yvu(t−tvu ) = 1 j ∈ F u = Bj  t = Lj  (41)
v
 Ch. 7. Transportation on Demand 453

j
 j
 j j
yuu(t−1) + yvu(t−tvu ) − yuvt − yuut = 0
v=u v=u
j ∈ F u ∈ A t = Ej + 1     Lj − 1 (42)
j −tvu
 t=S j
yvut  1 j ∈ F u = Bj  (43)
v t=Rj −tvu

xiuvt ∈ {0 1} i ∈ P u ∈ A v ∈ A t ∈ T (44)

j
yuvt ∈ {0 1} j ∈ F u ∈ A v ∈ A t ∈ T (45)
Constraints (35) and (36) state that each passenger departs from his origin air-
port and arrives at his destination airport. Constraints (37) ensure passenger
flow conservation at airports. Similar constraints are specified for each air-
plane, i.e., (40)–(42). Constraints (38) limit the number of intermediate stops
for each passenger and constraints (39) enforce the airplane capacity, i.e., the
total number of passengers flying on a leg is less than the total capacity of all
airplanes flying on that leg. Finally, constraints (43) force each airplane to re-
turn to its home base to switch pilots.
This time-discretized multicommodity network flow model becomes large
quickly and specialized solution approaches need to be developed to solve even
medium size instances, e.g., involving 15–30 airports and 5–10 airplanes.

5.2.2 Algorithmic issues related to the dynamic dial-a-flight problem

In the DDAFP one has to decide, given a set of already accepted requests,
whether an incoming request can be served or not. The amount of time avail-
able to make this decision depends on the business rules but most likely there
is very little time to do so. Consequently, fast heuristics will have to be part of
the decision technology. It is, however, conceivable that when heuristics fail to
accept a request quickly, a customer may be given the option of receiving final
notification of acceptance or rejection in, say, 30 minutes to allow time for op-
timization based techniques to try and accommodate the request. Research on
the DDAFP is relatively new and, in the current state of knowledge, there are
more questions than answers.
Given a set of already accepted requests and an existing feasible schedule,
it makes sense to ask what is the value of that schedule in deciding whether
it is feasible to accept a new request. What if simple insertion into any of the
existing airplane itineraries is not feasible? Does one try to reoptimize a single
airplane itinerary or try to reoptimize several airplane itineraries? One by one
or simultaneously? How does one select the airplane itineraries to consider?
What if the existing schedule included passenger plane changes? In fact, there
is no need to keep just one feasible schedule for the already accepted requests.
Should one keep several feasible schedules and try to insert new requests in
each of them?
Once an incoming request has been accepted, there may be time available to
optimize the schedule of the accepted requests by solving a static dial-a-flight
454 J.-F. Cordeau et al.

problem. Does that help handle the next request? Should the standard objec-
tive function be used or is it better to use an objective function that focuses
on the remaining flexibility in the schedule? How does one formally define
remaining flexibility?

5.3 The dial-a-flight problem in practice

The DAFP is a complex and challenging optimization problem, but still

ignores many relevant practical aspects. For example, in reality the costs of
a flight leg is a function of the fuel consumption, which depends on the weight
of the plane, the altitude of the flight, and the duration of the flight (shorter
flights burn more fuel as most fuel is consumed during take off and landing).
Furthermore, fuel prices may differ at various locations, so deciding where to
refuel may impact overall costs. Also, taxi service providers will likely offer var-
ious service classes, e.g., for a higher price a direct flight will be guaranteed, for
an even higher price one can charter an entire airplane plus a pilot for certain
period of time.

6 Ambulance fleet management

Transportation on demand problems also arise in the planning of emergency

services associated with fire fighting, police patrols, and ambulance fleet plan-
ning. Several important location, staffing and dispatching issues lie at the heart
of the problems encountered in these three areas but ambulance fleet manage-
ment is certainly the most relevant to this chapter. Readers interested in the
management of fire companies and in police patrols are referred to the work
of Larson (1972), Walker et al. (1979), Swersey (1994), and Adams (1997).
One of the key issues arising in ambulance fleet management is to decide
where to locate ambulances to provide at all times an adequate population
coverage. Over the past thirty-five years, there has been a steady evolution in
the development of ambulance location models, as witnessed by the work of
Marianov and ReVelle (1995) and Brotcorne et al. (2003). The latter study
serves as a basis for this chapter. The first models developed in the 1970s pro-
posed static solutions to a problem that is essentially stochastic and dynamic.
Over the years more realism has gradually been introduced into these basic
models until the emergence, a few years ago, of a truly dynamic solution pro-
cedure.
Most contributions in the field of ambulance location present integer linear
programming formulations, but no solution techniques. Only in recent years
have algorithms been proposed.

6.1 Two early ambulance location models

Ambulance location models are usually defined on a graph G = (V ∪W  A),

where V is a node set representing aggregated demand points, W is a set of
 Ch. 7. Transportation on Demand 455

potential ambulance location sites, and A = {(i j) ∈ V ∪W } is an arc set. With

each arc (i j) is associated a travel time tij . A demand point i ∈ V is covered
by site j ∈ W if and only if tij  r, where r is a preset coverage standard. Let
Wi = {j ∈ W : tij  r} be the set of location sites covering demand point i.
The Location Set Covering Model (LSCM) of Toregas et al. (1971) aims
to minimize the number of ambulances needed to cover all demand points. It
uses binary variables xj equal to 1 if and only if an ambulance is located at j:

(LSCM) minimize xj (46)
j∈W
subject to

xj  1 i ∈ V (47)
j∈Wi
xj ∈ {0 1} j ∈ W (48)
The Maximum Covering Location Problem (MCLP) of Church and ReVelle
(1974) works with a fixed number p of ambulances and attempts to cover the
largest possible demand z(p). Denote by di the demand at node i ∈ V and
let yi be a binary variable equal to 1 if and only if i is covered by at least one
ambulance:

(MCLP) maximize z(p) = di yi (49)
i∈V
subject to

xj  yi  i ∈ V (50)
j∈Wi

xj = p (51)
j∈W
xj ∈ {0 1} j ∈ W (52)
yi ∈ {0 1} i ∈ V (53)
A good way to combine these two models is to repeatedly solve MCLP with in-
creasing values of p and select a solution offering a good compromise between
p and z(p).

6.2 Static models with extra coverage

One major drawback of LSCM and of MCLP is that adequate coverage may
no longer exist once an ambulance has been dispatched. This is the reason why
models with extra coverage have been introduced. The Tandem Equipment
Allocation Model (TEAM) of Schilling et al. (1979) works with two equipment
types A and B, corresponding to advanced life support (ALS) units and basic
life support units (BLS) operating with different time standards (see Mandell,
456 J.-F. Cordeau et al.

1998). Let r A and r B be the respective coverage standards of A and B, and let
WiA = {j ∈ W : tij  r A } and WiB = {j ∈ W : tij  r B }. Let xA B
j (resp. xj )
be a binary variable equal to 1 if and only if a vehicle of type A (resp. B) is
located at j ∈ W , and let yi be a binary variable equal to 1 if and only if i ∈ V
is covered by two types of vehicle.

(TEAM) maximize di yi (54)
i∈V
subject to

xA
j  yi  i ∈ V (55)
j∈WiA

xB
j  yi  i ∈ V (56)
j∈WiB

xA A
j =p  (57)
j∈W

xB B
j =p  (58)
j∈W

xA B
j  xj  j ∈ W (59)
xA B
j  xj ∈ {0 1} j ∈ W (60)
yi ∈ {0 1} i ∈ V (61)
This model is a direct extension of MCLP, with the proviso that constraints (59)
impose a hierarchy between the two vehicle types. In the FLEET model of
Schilling et al. (1979), these constraints are relaxed and the number of poten-
tial location sites is limited to a preset value p. Another variant, proposed by
Daskin and Stern (1981), is to use a hierarchical objective to first maximize the
number of demand points covered more than once.
In the same spirit, Hogan and ReVelle (1986) have proposed the Backup
Coverage Problems, called BACOP1 and BACOP2, in which xj is the number
of ambulances located at j ∈ W , and yi , ui are binary variables equal to 1 if
and only if i ∈ V is covered once or at least twice, respectively:

(BACOP1) maximize di ui (62)
i∈V
subject to

xj  1 + ui  i ∈ V (63)
j∈Wi

xj = p (64)
j∈W
 Ch. 7. Transportation on Demand 457

ui ∈ {0 1} i ∈ V (65)

xj  0 and integer i ∈ V (66)
and
 
(BACOP2) maximize θ di yi + (1 − θ) di ui (67)
i∈V i∈V
subject to

xj  yi + ui  i ∈ V (68)
j∈Wi
ui  yi  i ∈ V  (69)

xj = p (70)
j∈W
ui ∈ {0 1} i ∈ V (71)
yi ∈ {0 1} i ∈ V (72)
xj  0 and integer j ∈ W (73)
where θ is a weight chosen in [0 1].
The Double Standard Model (DSM) of Gendreau et al. (1997) works with
two coverage standards r1 and r2 , with r1 < r2 , as specified by the United States
Emergency Medical Services Act of 1973. A proportion α of the demand must
be covered within r1 while the entire demand must be covered within r2 . In the
DSM, the objective is to maximize the demand covered twice within r1 using
p ambulances, at most pj of which are located at j ∈ W , subject to the double
coverage constraints. Let Wi1 = {j ∈ W : tij  r1 } and Wi2 = {j ∈ W : tij  r2 }.
The integer variable xj denotes the number of ambulances located at j ∈ W
and the binary variable yik is equal to 1 if and only if the demand at node i ∈ V
is covered k times (k = 1 or 2) within r1 . The formulation is then:

(DSM) maximize di yi2 (74)
i∈V
subject to

xj  1 i ∈ V (75)
j∈Wi2
 
di yi1  α di  (76)
i∈V i∈V

xj  yi1 + yi2  i ∈ V (77)
j∈Wi1

yi2  yi1  i ∈ V (78)


xj = p (79)
j∈W
458 J.-F. Cordeau et al.

xj  pj  j ∈ W (80)
yi1  yi2 ∈ {0 1} i ∈ V (81)
xj  0 and integer j ∈ W (82)
Here, the objective function computes the demand covered twice within r1 time
units, and constraints (75) mean that all demand is covered within r2 . The
left-hand side of (77) represents the number of ambulances covering node i
within r1 units, while the right-hand side is equal to 1 if i is covered once within
r1 units, and equal to 2 if it is covered at least twice within r1 units. The combi-
nation of constraints (76) and (77) ensures that a proportion α of the demand
is covered within r1 . Constraints (78) state that node i cannot be covered at
least twice if it is not covered at least once. In constraints (80), pj can be set
equal to 2 since an optimal solution using this upper bound always exists.
Gendreau, Laporte, and Semet solve the DSM by means of a tabu search
procedure in which neighbor solutions are obtained by generating sequences
of ambulance moves to an adjacent location, not unlike what is done in ejection
chain methods (see, e.g., Rego and Roucairol, 1996). Comparisons with the
linear relaxation value of the model indicate that the heuristic typically yields
optimal or near-optimal solutions.

6.3 Probabilistic models with extra coverage

None of the models introduced so far takes into account that ambulances
are not always available to answer a call. A way around this is to assume that
each ambulance has a probability q, called busy fraction, of being unavailable.
This value is obtained by dividing the total time spent by all ambulances on all
calls by the total ambulance time available. If i ∈ V is covered by k ambulances,
then the expected demand covered at that node is Eik = di (1 − qk ) and the
marginal contribution of the kth ambulance is Eik − Eik−1 = di (1 − q)qk−1 .
In the Maximum Expected Covering Location Model (MEXCLP) of Daskin
(1983), up to p ambulances may be located in total, and more than one vehicle
may be located at the same node. Let yik be a binary variable equal to 1 if
and only if node i ∈ V is covered by at least k ambulances. The model can be
written as follows:

p
(MEXCLP) maximize di (1 − q)qk−1 yik (83)
i∈V k=1
subject to
 
p
xj  yik  i ∈ V (84)
j∈Wi k=1

xj  p (85)
j∈W
 Ch. 7. Transportation on Demand 459

xj  0 and integer j ∈ W (86)

yik ∈ {0 1} i ∈ V  k = 1     p (87)
The validity of this model stems from the fact that the objective function is
concave in k. Therefore, if yik = 1, then yih = 1 for h  k. Since the objective
is to be maximized, both (84) and (85) will be satisfied as equalities. It follows
that the two sides of (84) will be equal to the number of ambulances covering
node i ∈ V .
An application of MEXCLP to the City of Bangkok (|V | = 59, |W | = 46,
10  p  30) by Fujiwara et al. (1987) has shown that without reducing ex-
pected coverage, the number of ambulances could be reduced to 15 from the
current 21. Repede and Bernardo (1994) have proposed TIMEXCLP, a dy-
namic implementation of MEXCLP, in which travel speeds are allowed to vary
during the day. Goldberg et al. (1990) have worked with stochastic travel times.
They compute the probability that a given demand point i will be covered
based on three probabilities: the probability that an ambulance located at the
sth preferred site of i can reach i within eight minutes, the probability that
this ambulance is available, and the probability that ambulances located at less
preferred sites are not available. Experiments conducted on data collected in
Tucson, Arizona, have shown that the use of this model could increase the ex-
pected covered demand from 24% to 53.1%.
ReVelle and Hogan (1989) have developed two chance-constrained maximal
covering location models, called the Maximal Availability Location Problem
(MALP I and MALP II), in which the constraint

xj
1−q j∈Wi
 α i ∈ V (88)
ensures that each demand point is covered with probability at least equal to α.
This constraint, which can be linearized as
 
log(1 − α)
xj  = b i ∈ V  (89)
log q
j∈Wi

can be used instead of (47) in LSCM. In MALP I, Hogan and ReVelle maxi-
mize the total demand covered with b ambulances, subject to a global availabil-
ity of p ambulances. Defining binary variables yjk as in MEXCLP, the MALP I
model can then be written as:

(MALP I) maximize di yib (90)
i∈V
subject to

b 
yjk  xj  i ∈ V (91)
k=1 j∈Wi
yik  yik−1  i ∈ V  k = 2     b (92)
460 J.-F. Cordeau et al.

xj = p (93)
j∈W

xj ∈ {0 1} j ∈ W (94)

yik ∈ {0 1} i ∈ V  k = 1     p (95)

Here, constraints (92) are required since the concavity property observed in
MEXCLP no longer holds.
In MALP II, ReVelle and Hogan estimate a busy fraction qi associated with
each i, as the ratio of the total duration of all calls associated to i to the to-
tal ambulance time in Wi . The major difficulty with this is that qi cannot be
computed a priori because it is an output of the model. An iterative procedure
is then required to solve MALP II approximately. Finally we mention the ex-
istence of two studies, by Batta et al. (1989) and by Marianov and ReVelle
(1994), whose aim is to better approximate the busy fraction of the whole
system or associated with a particular location. These are based in part on
Larson’s (1974) hypercube model.
Finally, Ball and Lin (1993) have developed an extension of LSCM, called
Rel-P, which incorporates a linear constraint on the number of vehicles re-
quired to achieve a given reliability level. The model contains binary variables
xjk equal to 1 if and only if k ambulances are located at node j ∈ W , and
constants cjk equal to the cost of locating k vehicles at site j. An upper bound
pj is imposed on the number of ambulances located at site j. Their model is as
follows:
 
(Rel-P) minimize cjk xjk (96)
j∈J 1kpj

subject to

xjk  1 j ∈ W (97)
1kpj
 
ajk xjk  bi  i ∈ V (98)
j∈Wi 1kpj

xjk ∈ {0 1} j ∈ W  1  k  pj  (99)

In constraints (98), the constants ajk and bi are computed to ensure that given
the number of ambulances covering demand point i, the probability of being
unable to answer a call does not exceed a certain value. The computation of
the ajk and bi coefficients are, in fact, carried out by using an upper bound on
that probability.
Ball and Lin incorporate valid inequalities in their model which is then
solved by means of a standard branch-and-bound code for integer linear pro-
gramming.
 Ch. 7. Transportation on Demand 461

6.4 A dynamic model

In practice, ambulances are often relocated over time in order to always en-
sure an adequate coverage. Gendreau et al. (2001) have developed a dynamic
relocation model where a new redeployment can in principle be implemented
at each instant t at which an ambulance is dispatched to a call or returns
from a call. The model is based on the DSM developed by the same authors
(Gendreau et al., 1997). It constitutes, to our knowledge, the only available
dynamic ambulance relocation tool. In addition to the standard coverage and
site capacity constraints, the model takes into account a number of practical
considerations inherent to the dynamic nature of the problem: (1) one should
avoid relocating the same vehicle frequently within a short time interval; (2) re-
peated round trips between the same two location sites must be avoided;
(3) long trips between the initial and final location sites must be avoided.
The dynamic aspect of the redeployment model is captured by time depen-
dent constants Mjt equal to the cost of relocating, at time t, ambulance from
its current site to site j ∈ W . This includes the case where site j coincides with
the current location of the ambulance, i.e., Mjt = 0. The constant Mjt captures
some of the history of ambulance . If it has been relocated frequently prior to
time t, then Mjt will be larger. If relocating ambulance to site j violates any
of the above constraints, then the relocation is simply disallowed. The constant
pt is the number of ambulances available at time t, and ptj is the maximum
number of ambulances that can be located at site j at time t. Binary variables
xj are equal to 1 if and only if ambulance is relocated to site j. The Dynamic
Double Standard Model at Time t (DDSMt ) can now be described:
t
 
p
t
(DDSM ) maximize di yi2 − Mjt xj (100)
i∈V j∈W =1
subject to
t
 
p
xj  1 i ∈ V (101)
j∈Wi2 =1
 
di yi1  α di  (102)
i∈V i∈V
t
 
p
xj  yi1 + yi2  i ∈ V (103)
j∈Wi1 =1

yi2  yi1  i ∈ V  (104)


xj = 1 = 1     pt  (105)
j∈W
462 J.-F. Cordeau et al.

t

p
xj  ptj  j ∈ W (106)
=1
yi1  yi2 ∈ {0 1} i ∈ V (107)
t
xj ∈ {0 1} j ∈ W  = 1     p  (108)
Apart from variables xj , all variables, parameters and constraints of this
model can be interpreted as in the static case. The objective function is the
demand covered twice within r1 time units minus the sum of penalties associ-
ated with ambulance relocations at time t.
Gendreau et al. (2001) solve DDSMt by constructing a redeployment table
in which the first column gives the list of all ambulances that could possibly
be dispatched to the next call, and the second column gives the redeployment
plan associated with the ambulance. Whenever an ambulance is dispatched,
the associated relocation plan is implemented and the table is then recom-
puted from scratch. The authors have used the Gendreau et al. (1997) tabu
search algorithm to compute the relocation strategies and they have also made
use of a simple parallel computing strategy: each of 16 processors was assigned
the computation of a series of rows of the redeployment table but no communi-
cation took place between the processors. The success of this approach rests on
the capability of recomputing the entire table between successive calls. Simu-
lations performed on real data from Montreal have shown that this was indeed
possible in 95% of all cases. Out of all calls, 38% required at least one am-
bulance relocation and in only 0.05% of the situations were more than five
ambulance relocations necessary. Whenever the table cannot be fully com-
puted between two calls, it is deleted at the next call and no relocation takes
place.

Acknowledgements

This work was supported by the Natural Sciences and Engineering Research
Council of Canada under Grants 227837-00, OGP0039862, and 36662-01. This
support is gratefully acknowledged. Most of the material in Section 5 is the re-
sult of discussions among the members of a project team designing and imple-
menting scheduling technology for air taxi services, consisting of Mo Bazaraa,
Marcos Coycoolea, Daniel Espinoza, Yongpei Guan, Renan Garcia, George
Nemhauser, and Martin Savelsbergh from Georgia Tech and Alex Khmelnit-
sky and Eugene Taits from Jetson Systems.

References

Adams, T.F. (1997). Police Field Operations. Prentice Hall, Englewood Cliffs, NJ.
Alfa, A.S. (1986). Scheduling of vehicles for transportation of elderly. Transportation Planning and Tech-
nology 11, 203–212.
 Ch. 7. Transportation on Demand 463

Attanasio, A., Cordeau, J.-F., Ghiani, G., Laporte, G. (2004). Parallel tabu search heuristics for the
dynamic multi-vehicle dial-a-ride problem. Parallel Computing 30, 377–387.
Balas, E., Fischetti, M., Pulleyblank, W.R. (1995). The precedence-constrained asymmetric traveling
salesman polytope. Mathematical Programming 68, 241–265.
Baldacci, R., Hadjiconstantinou, E., Mingozzi, A. (2003). An exact algorithm for the traveling salesman
problem with deliveries and collections. Networks 42, 26–41.
Ball, M.O., Lin, L.F. (1993). A reliability model applied to emergency service vehicle location. Opera-
tions Research 41, 18–36.
Batta, R., Dolan, J.M., Krishnamurty, N.N. (1989). The maximal expected covering location problem:
Revisited. Transportation Science 23, 277–287.
Bent, R., Van Hentenryck, P. (2004). Scenario-based planning for partially dynamic vehicle routing with
stochastic customers. Operations Research 52, 573–586.
Benyahia, I., Potvin, J.-Y. (1998). Decision support for vehicle dispatching using genetic programming.
IEEE Transactions on Systems, Man and Cybernetics 28, 306–314.
Bertsimas, D.J., van Ryzin, G.J. (1991). A stochastic and dynamic vehicle routing problem in the Euclid-
ean plane. Operations Research 39, 601–615.
Bianco, L., Mingozzi, A., Ricciardelli, S., Spadoni, M. (1994). Exact and heuristic procedures for
the traveling salesman problem with precedence constraints, based on dynamic programming.
INFOR 32, 19–31.
Bodin, L.D., Sexton, T. (1986). The multi-vehicle subscriber dial-a-ride problem. TIMS Studies in Man-
agement Science 26, 73–86.
Borndörfer, R., Grötschel, M., Klostermeier, F., Küttner, C. (1999). Telebus Berlin: Vehicle scheduling
in a dial-a-ride system. In: Wilson, N.H.M. (Ed.), Computer-Aided Transit Scheduling. Lecture Notes
in Economics and Mathematical Systems, vol. 471. Springer-Verlag, pp. 391–422.
Brotcorne, L., Laporte, G., Semet, F. (2003). Ambulance location and relocation models. European
Journal of Operational Research 147, 451–463.
Campbell, A., Savelsbergh, M.W.P. (2003). Incentive schemes for attended home delivery services. Tech-
nical Report TLI-03-04, Georgia Institute of Technology, The Logistics Institute.
Campbell, A., Savelsbergh, M.W.P. (2005). Decision support for consumer direct grocery initiatives.
Transportation Science 39, 313–327.
Christofides, N. (1976). Worst case analysis of a new heuristic for the travelling salesman problem.
Technical Report 388, Graduate School of Industrial Administration, Carnegie-Mellon University,
Pittsburgh, PA.
Church, R.L., ReVelle, C.S. (1974). The maximal covering location problem. Papers of the Regional
Science Association 32, 101–118.
Cordeau, J.-F. (2006). A branch-and-cut algorithm for the dial-a-ride problem. Operations Research 54,
573–586.
Cordeau, J.-F., Laporte, G. (2003a). The dial-a-ride problem (DARP): Variants, modeling issues and
algorithms. 4OR – Quarterly Journal of the Belgian, French and Italian Operations Research Societies 1,
89–101.
Cordeau, J.-F., Laporte, G. (2003b). A tabu search heuristic for the static multi-vehicle dial-a-ride prob-
lem. Transportation Research B 37, 579–594.
Cordeau, J.-F., Gendreau, M., Laporte, G. (1997). A tabu search heuristic for periodic and multi-depot
vehicle routing problems. Networks 30, 105–119.
Coslovich, L., Pesenti, R., Ukovich, W. (2003). A two-phase insertion technique of unexpected cus-
tomers for a dynamic dial-a-ride problem. Technical report, Università di Trieste, Italy.
Cullen, F.H., Jarvis, J.J., Ratliff, H.D. (1981). Set partitioning based heuristics for interactive routing.
Networks 11, 125–143.
Daskin, M.S. (1983). A maximum expected location model: Formulation, properties and heuristic so-
lution. Transportation Science 7, 48–70.
Daskin, M.S., Stern, E.H. (1981). A hierarchical objective set covering model for emergency medical
service vehicle deployment. Transportation Science 15, 137–152.
Desaulniers, G., Desrosiers, J., Erdmann, A., Solomon, M.M., Soumis, F. (2002). VRP with pickup and
delivery. In: Toth, P., Vigo, D. (Eds.), The Vehicle Routing Problem. SIAM Monographs on Discrete
Mathematics and Applications. SIAM, Philadelphia, PA, pp. 225–242.
464 J.-F. Cordeau et al.

Desrosiers, J., Dumas, Y., Soumis, F. (1986). A dynamic programming solution of the large-scale single-
vehicle dial-a-ride problem with time windows. American Journal of Mathematical and Management
Sciences 6, 301–325.
Desrosiers, J., Dumas, Y., Soumis, F., Taillefer, S., Villeneuve, D. (1991). An algorithm for mini-
clustering in handicapped transport. Technical Report G-91-22, GERAD, HEC Montréal.
Dial, R.B. (1995). Autonomous dial-a-ride transit introductory overview. Transportation Research C 3,
261–275.
Dumas, Y., Desrosiers, J., Soumis, F. (1989). Large scale multi-vehicle dial-a-ride problems. Technical
Report G-89-30, GERAD, HEC Montréal.
Dumas, Y., Soumis, F., Desrosiers, J. (1990). Optimizing the schedule for a fixed vehicle path with
convex inconvenience costs. Transportation Science 24, 145–152.
Dumas, Y., Desrosiers, J., Soumis, F. (1991). The pickup and delivery problem with time windows.
European Journal of Operational Research 54, 7–22.
Fischetti, M., Toth, P. (1989). An additive bounding procedure for combinatorial optimization prob-
lems. Operations Research 37, 319–328.
Fu, L. (2002). Scheduling dial-a-ride paratransit under time-varying, stochastic congestion. Transporta-
tion Research B 36, 485–506.
Fujiwara, O., Makjamroen, T., Gupta, K.K. (1987). Ambulance deployment analysis: A case study of
Bangkok. European Journal of Operational Research 31, 9–18.
Garey, M.R., Johnson, D.S. (1979). Computers and Intractability: A Guide to the Theory of
NP-Completeness. Freeman, San Francisco.
Gendreau, M., Laporte, G., Semet, F. (1997). Solving an ambulance location model by tabu search.
Location Science 5, 75–88.
Gendreau, M., Guertin, F., Potvin, J.-Y., Séguin, R. (1998). Neighborhood search heuristics for a
dynamic vehicle dispatching problem with pick-ups and deliveries. Technical Report CRT-98-10,
Centre de recherche sur les transports, Montreal.
Gendreau, M., Laporte, G., Vigo, D. (1999). Heuristics for the traveling salesman problem with pickup
and delivery. Computers & Operations Research 26, 699–714.
Gendreau, M., Laporte, G., Semet, F. (2001). A dynamic model and parallel tabu search heuristic for
real-time ambulance relocation. Parallel Computing 27, 1641–1653.
Glover, F. (1996). Ejection chains, reference structures and alternating path methods for traveling sales-
man problems. Discrete Applied Mathematics 65, 223–253.
Goldberg, J., Dietrich, R., Chen, J.M., Mitwasi, M.G. (1990). Validating and applying a model for lo-
cating emergency medical services in Tucson, AZ. European Journal of Operational Research 49,
308–324.
Grötschel, M., Padberg, M.W. (1985). Polyhedral theory. In: Lawler, E.L., Lenstra, J.K., Rinnooy Kan,
A.H.G., Shmoys, D.B. (Eds.), The Traveling Salesman Problem. Wiley, New York, pp. 251–305.
Healy, P., Moll, R. (1995). A new extension of local search applied to the dial-a-ride problem. European
Journal of Operational Research 83, 83–104.
Hernández-Pérez, H., Salazar-González, J.-J. (2004). A branch-and-cut algorithm for a traveling sales-
man problem with pickup and delivery. Discrete Applied Mathematics 145, 126–139.
Hogan, K., ReVelle, C.S. (1986). Concepts and applications of backup coverage. Management Sci-
ence 34, 1434–1444.
Holland, J.H. (1992). Adaptation in Natural and Artificial Systems. MIT Press, Cambridge, MA.
Horn, M.E.T. (2002). Fleet scheduling and dispatching for demand-responsive passenger services.
Transportation Research C 10, 35–63.
Hunsaker, B., Savelsbergh, M. (2002). Efficient testing for dial-a-ride problems. Operations Research
Letters 30, 169–173.
Ioachim, I., Desrosiers, J., Dumas, Y., Solomon, M.M. (1995). A request clustering algorithm for door-
to-door handicapped transportation. Transportation Science 29, 63–78.
Jaw, J., Odoni, A.R., Psaraftis, H.N., Wilson, N.H.M. (1986). A heuristic algorithm for the multi-vehicle
advance-request dial-a-ride problem with time windows. Transportation Research B 20, 243–257.
Kalantari, B., Hill, A.V., Arora, S.R. (1985). An algorithm for the traveling salesman problem with
pickup and delivery customers. European Journal of Operational Research 22, 377–386.
 Ch. 7. Transportation on Demand 465

Koza, J.R. (1992). Genetic Programming. MIT Press, Cambridge, MA.

Larson, R.C. (1972). Urban Police Patrol Analysis. MIT Press, Cambridge, MA.
Larson, R.C. (1974). A hypercube qeueing model for facility location and redistricting in urban emer-
gency services. Computers & Operations Research 1, 67–75.
Lau, H.C., Liang, Z. (2002). Pickup and delivery with time windows: Algorithms and test case genera-
tion. International Journal on Artificial Intelligence Tools 11, 455–472.
Leclerc, F., Potvin, J.-Y. (1997). Genetic algorithms for vehicle dispatching. International Transactions
in Operational Research 4, 391–400.
Lin, S. (1965). Computer solutions of the travelling salesman problem. Bell System Technical Journal 44,
2245–2269.
Lin, S., Kernighan, B.W. (1973). An effective heuristic algorithm for the traveling-salesman problem.
Operations Research 21, 498–516.
Little, J.D.C., Murty, K.G., Sweeney, D.W., Karel, C. (1963). An algorithm for the traveling salesman
problem. Operations Research 11, 972–989.
Madsen, O.B.G., Ravn, H.F., Rygaard, J.M. (1995). A heuristic algorithm for a dial-a-ride problem
with time windows, multiple capacities, and multiple objectives. Annals of Operations Research 60,
193–208.
Mandell, M.B. (1998). Covering models for two-tiered emergency medical services systems. Location
Science 6, 355–368.
Marianov, V., ReVelle, C.S. (1994). The queueing probabilistic location set covering problem and some
extensions. Socio-Economic Planning Sciences 28, 167–178.
Marianov, V., ReVelle, C.S. (1995). Siting emergency services. In: Drezner, Z. (Ed.), Facility Location.
Springer-Verlag, New York, pp. 199–223.
Mitrović-Minić, S., Laporte, G. (2004). Waiting strategies for the dynamic pickup and delivery problem
with time windows. Transportation Research B 38, 635–655.
Mitrović-Minić, S., Krishnamurti, R., Laporte, G. (2004). Double-horizon based heuristics for the dy-
namic pickup and delivery problem with time windows. Transportation Research B 38, 669–685.
Nanry, W.P., Barnes, J.W. (2000). Solving the pickup and delivery problem with time windows using
reactive tabu search. Transportation Research B 34, 107–121.
Psaraftis, H.N. (1980). A dynamic programming approach to the single-vehicle, many-to-many imme-
diate request dial-a-ride problem. Transportation Science 14, 130–154.
Psaraftis, H.N. (1983a). Analysis of an O(N 2 ) heuristic for the single vehicle many-to-many Euclidean
dial-a-ride problem. Transportation Research B 17, 133–145.
Psaraftis, H.N. (1983b). An exact algorithm for the single-vehicle many-to-many dial-a-ride problem
with time windows. Transportation Science 17, 351–357.
Psaraftis, H.N. (1983c). k-interchange procedures for local search in a precedence-constrained routing
problem. European Journal of Operational Research 13, 391–402.
Psaraftis, H.N. (1988). Dynamic vehicle routing problems. In: Golden, B.L., Assad, A.A. (Eds.), Vehicle
Routing: Methods and Studies. North-Holland, Amsterdam, pp. 223–248.
Rego, C., Roucairol, C. (1996). A parallel tabu search algorithm using ejection chains for the vehicle
routing problem. In: Osman, I.H., Kelly, J.P. (Eds.), Meta-Heuristics: Theory & Applications. Kluwer
Academic, Boston, pp. 661–675.
Renaud, J., Boctor, F.F., Laporte, G. (1996). A fast composite heuristic for the symmetric traveling
salesman problem. INFORMS Journal on Computing 8, 134–143.
Renaud, J., Boctor, F.F., Ouenniche, J. (2000). A heuristic for the pickup and delivery traveling salesman
problem. Computers & Operations Research 27, 905–916.
Renaud, J., Boctor, F.F., Laporte, G. (2002). Perturbation heuristics for the pickup and delivery travel-
ing salesman problem. Computers & Operations Research 29, 1129–1141.
Repede, J.F., Bernardo, J.J. (1994). Developing and validating a decision support system for locating
emergency medical vehicles in Louisville, Kentucky. European Journal of Operational Research 75,
567–581.
ReVelle, C.S., Hogan, K. (1989). The maximum availability location problem. Transportation Science 23,
192–200.
466 J.-F. Cordeau et al.

Rochat, Y., Taillard, É.D. (1995). Probabilistic diversification and intensification in local search for
vehicle routing. Journal of Heuristics 1, 147–167.
Ropke, S., Pisinger, D. (2004). An adaptive large neighborhood search heuristic for the pickup and
delivery problem with time windows. Technical Report 2004-13, DIKU, University of Copenhagen.
Ruland, K.S. (1995). Polyhedral solution to the pickup and delivery problem. PhD thesis, Sever Institute
of Technology, Washington University.
Ruland, K.S., Rodin, E.Y. (1997). The pickup and delivery problem: Faces and branch-and-cut algo-
rithm. Computers and Mathematics with Applications 33, 1–13.
Rumelhart, D.E., Hinton, G.E., Williams, R.J. (1986). Learning internal representations by error prop-
agation. In: Rumelhart, D.E., McClelland, J.L. (Eds.), Parallel Distributed Processing, vol. 1. MIT
Press, Cambridge, MA, pp. 318–364.
Savelsbergh, M.W.P. (1985). Local search in routing problems with time windows. Annals of Operations
Research 4, 285–305.
Savelsbergh, M.W.P. (1990). An efficient implementation of local search algorithms for constrained
routing problems. European Journal of Operational Research 47, 75–85.
Savelsbergh, M.W.P. (1992). The vehicle routing problem with time windows: Minimizing route dura-
tion. ORSA Journal on Computing 4, 146–154.
Savelsbergh, M.W.P., Sol, M. (1995). The general pickup and delivery problem. Transportation Sci-
ence 29, 17–29.
Savelsbergh, M.W.P., Sol, M. (1998). DRIVE: Dynamic routing of independent vehicles. Operations
Research 46, 474–490.
Schilling, D.A., Elzinga, D.J., Cohon, J., Church, R.L., ReVelle, C.S. (1979). The TEAM FLEET mod-
els for simultaneous facility and equipment siting. Transportation Science 13, 163–175.
Sexton, T., Bodin, L.D. (1985a). Optimizing single vehicle many-to-many operations with desired deliv-
ery times: I. Scheduling. Transportation Science 19, 378–410.
Sexton, T., Bodin, L.D. (1985b). Optimizing single vehicle many-to-many operations with desired deliv-
ery times: II. Routing. Transportation Science 19, 411–435.
Shen, Y., Potvin, J.-Y., Rousseau, J.-M., Roy, S. (1995). A computer assistant for vehicle dispatching
with learning capabilities. Annals of Operations Research 61, 189–211.
Stein, D.M. (1978). Scheduling dial-a-ride transportation systems. Transportation Science 12, 232–249.
Swersey, A.J. (1994). The deployment of police, fire, and emergency medical units. In: Pollock, S.M.,
Rothkopf, M.H., Barnett, A. (Eds.), Operations Research and the Public Sector. Handbooks in Oper-
ations Research and Management Science, vol. 6. North-Holland, pp. 151–200.
Swihart, M.R., Papastravou, J.D. (1999). A stochastic and dynamic model for the single-vehicle pick-up
and delivery problem. European Journal of Operational Research 114, 447–464.
Teodorovic, D., Radivojevic, G. (2000). A fuzzy logic approach to dynamic dial-a-ride problem. Fuzzy
Sets and Systems 116, 23–33.
Toregas, C.R., Swain, R., ReVelle, C.S., Bergman, L. (1971). The location of emergency service facili-
ties. Operations Research 19, 1363–1373.
Toth, P., Vigo, D. (1996). Fast local search algorithms for the handicapped persons transportation prob-
lem. In: Osman, I.H., Kelly, J.P. (Eds.), Meta-Heuristics: Theory & Applications. Kluwer Academic,
Boston, pp. 677–690.
Toth, P., Vigo, D. (1997). Heuristic algorithms for the handicapped persons transportation problem.
Transportation Science 31, 60–71.
Van der Bruggen, L.J.J., Lenstra, J.K., Schuur, P.C. (1993). Variable-depth search for the single-vehicle
pickup and delivery problem with time windows. Transportation Science 27, 298–311.
Walker, W.E., Chaiken, J.M., Ignall, E.J. (1979). Fire Department Deployment Analysis. North-Holland,
New York.
Xu, H., Chen, Z.-L., Rajagopal, S., Arunapuram, S. (2003). Solving a practical pickup and delivery
problem. Transportation Science 37, 347–364.
C. Barnhart and G. Laporte (Eds.), Handbook in OR & MS, Vol. 14
Copyright © 2007 Elsevier B.V. All rights reserved
DOI: 10.1016/S0927-0507(06)14008-6

Chapter 8
Intermodal Transportation
Teodor Gabriel Crainic
Département management et technologie, École des Sciences de la Gestion,
Université du Québec à Montréal, and Centre de recherche sur les transports – CIRRELT,
Montréal, Canada
E-mail: [email protected]

Kap Hwan Kim

Department of Industrial Engineering, Pusan National University, Korea
E-mail: [email protected]

Keywords: Intermodal transportation, freight transportation, operations re-

search

1 Introduction

Intermodal transportation may be defined as the transportation of a person

or a load from its origin to its destination by a sequence of at least two trans-
portation modes, the transfer from one mode to the next being performed at
an intermodal terminal. The concept is very general and thus, it means many
things to many people: transportation of containerized cargo by a combination
of truck, rail, and ocean shipping, dedicated rail services to move massive quan-
tities of containers and trailers over long distances, main transportation mode
for the international movement of goods, central piece in defining transporta-
tion policy for the European Community, trips undertaken by a combination
of private (e.g., car) and public (e.g., light rail) transport, and so on. One must
therefore start with a few definitions to set the terminology and limit the scope
of this chapter. First, although both people and freight transportation can be
examined from an intermodal perspective, we limit the scope of this chapter to
freight.
In one of its most widely accepted meanings, intermodal freight transporta-
tions refers to a multimodal chain of container-transportation services. This
chain usually links the initial shipper to the final consignee of the container (so-
called door-to-door service) and takes place over long distances. Transportation
is often provided by several carriers. In a classical example of an intercontinen-
tal intermodal chain, loaded containers leave a shipper’s facility by truck either
directly to port or to a rail yard from where a train will deliver them to port.
A ship will move the containers from this initial port to a port on the other
continent, from where they will be delivered to the final destination by a single
or a combination of “land” transportation means: truck, rail, coastal or river

467
468 T.G. Crainic and K.H. Kim

navigation. Several intermodal terminals are part of this chain: the initial and
final seaport container terminals, where containers are transferred between
the ocean navigation and land transportation modes, as well as in-land termi-
nals (rail yards, river ports, etc.) providing transfer facilities between the land
modes.
Container transportation is a major component of intermodal transporta-
tion and international commerce and this importance is reflected in this
chapter. Intermodal transportation is not only about containers and interconti-
nental exchanges, however. On the one hand, a significant part of international
trade that is moved in containers does not involve ocean navigation, land trans-
portation means providing the intermodal chain. On the other hand, other
types of cargo may be moved by a chain of transportation means and re-
quire intermodal transfer facilities, as illustrated by the definition European
Conference of Ministers of Transport (2001) gives for intermodal transporta-
tion: “movement of goods in one and the same loading unit or vehicle, which
uses successively two or more modes of transport without handling the goods
themselves in changing modes”. This last definition is still too restrictive, how-
ever. Thus, for example, the transportation of express and regular mail on a
regional or national scale is strongly intermodal, using various combinations
of road, rail, and air transportation modes, and yet freight is handled (sorted
and grouped) in terminals. More generally, the transportation of less-than-
vehicle-capacity loads by nondedicated services is intermodal, since it involves
pickup (at origin) and delivery (at destination) operations, usually performed
by trucks, at least one long-haul transportation movement by road, rail, river,
or air, as well as transfer activities between these modes in dedicated terminals.
Almost all types of freight carriers and terminal operator may, thus, be
involved in intermodal transportation, either by providing service for part of
the transportation chain or by operating an intermodal transportation system
(network). We limit the scope of the chapter to the latter with a particular fo-
cus on container-based systems, including container terminals in seaports. The
national/regional planning perspective, which considers the flow of multiple
products on multimodal networks, is also addressed.
Compared to several other application areas, Operations Research models
and methods for intermodal freight transportation is still a very young domain.
In many cases, there are not, yet, widely accepted models and methodologies.
Work is indeed proceeding as this chapter is being written. Therefore, the goal
of the chapter is to be informative and provide a starting point for future re-
search. The chapter overviews the evolution of the intermodal transportation
field and presents methodological developments proposed to address a num-
ber of important operations and planning issues: system and service design for
intermodal transportation networks, container fleet management, container
terminal operations and scheduling, national planning. We focus on models.
Algorithmic developments are indicated but not examined in any depth.
To structure the presentation, we follow the somewhat classical approach of
examining issues, models, and methods according to whether they belong to
 Ch. 8. Intermodal Transportation 469

the strategic, tactic, or operational level of planning and management of opera-

tions. The chapter is therefore organized as follows. Section 2 briefly describes
freight transportation and the main actors and issues discussed in the chapter.
Section 3 is dedicated to system and service network design issues for carriers.
Strategic design issues for seaport container terminals are also discussed in this
section. Operational-planning issues are addressed in Section 4 with particular
emphasis on empty container repositioning. Section 5 is dedicated to tactical
and operational planning of operations in container terminals, while Section 6
examines national planning models. Section 7 concludes the chapter with an
enumeration of a number of trends and technological developments that may
influence intermodal transportation in the future, as well as of promising re-
search directions.

2 Freight transportation systems

Demand for freight transportation derives from the interplay between pro-
ducers and consumers and the significant distances that often separate them.
Producers of goods require transportation services to move raw materials and
intermediate products, and to distribute final goods in order to meet customer
demands. Shippers, which may be the producers of goods or some intermedi-
ary firm (e.g., brokers), thus generate the demand for transportation. Carriers
answer this demand by supplying transportation services. Railways, ocean ship-
ping lines, trucking companies, and postal services are examples of carriers.
Considering the type of services they provide, seaports, intermodal platforms,
and other such facilities may be described as carriers as well. Governments con-
tribute the infrastructure: roads and highways, as well as significant portions of
ports, internal navigation, and rail facilities. Governments also regulate (e.g.,
dangerous and toxic goods transportation) and tax the industry.
We do not intend to make a detailed presentation of freight transporta-
tion. Our goal is rather to describe the basic operation and planning issues for
long-haul carriers and terminals that are involved in intermodal transporta-
tion, with a particular focus on container-based systems, including container
terminals in seaports. The first subsection gives a few statistics and trends rela-
tive to container-based transportation. The second and the third are dedicated
to long-haul carriers and seaport container terminals, respectively.
Reviews on these issues may be found in Assad (1980), Cordeau et al.
(1998), and Crainic (1988) for rail transportation, Delorme et al. (1988) and
Powell (1988) for motor carrier transportation, Christiansen et al. (2004, 2007)
for maritime transportation, and Günther and Kim (2005) and Steenken et al.
(2004) for container port terminals. The reviews proposed by Crainic (2000,
2003), Crainic and Laporte (1997), Daganzo (2005), Dejax and Crainic (1987),
Powell (2003), Powell and Topaloglu (2003, 2005), Powell et al. (1995, 2007)
are more general in scope, addressing issues relevant for more than one trans-
portation mode or problem. Issues and methodologies related to the pickup
470 T.G. Crainic and K.H. Kim

and delivery of loads, generally known as vehicle routing and scheduling prob-
lems, are not included in this chapter. Interested readers may refer to Golden
and Assad (1988), Ball et al. (1995), Dror (2000), Toth and Vigo (2002), and
Cordeau et al. (2007).

2.1 Container intermodal transportation

Container-related transportation activities have grown remarkably over the

last 10 years and the trend does not show any sign of slowing down as il-
lustrated by the annual world container traffic figures, in millions of TEUs
(20 feet equivalent container units), displayed in Table 1 (2006 figures are esti-
mated; Koh and Kim, 2001; ISL, 2006). The initial impulse to container-based
transportation came from the safety it offered regarding loss and damage.
Advantages in terms of reduced cargo handling and standardization of trans-
portation and transfer equipment translate in cost economies and efficient,
world-wide door-to-door intermodal service and fuel the growth of the indus-
try. Containerized intermodal transportation supports a significant part of the
international movement of goods.
The performance of container-based transportation in international trade
has had some remarkable consequences. Ports and container terminals have
been built or profoundly modified to accommodate container ships and ef-
ficiently perform the loading, unloading, and transfer operations. Container
terminal equipment and operating procedures are continuously enhanced to
improve productivity and compete, in terms of cost and time, with the other
ports to attract ocean-shipping lines. Models for planning container-terminal
operations are presented in Section 5. Notice that the competitive position of
a container port is also dependent on the capacity and efficiency of the land
transportation system. The models described in Section 6 may be used for the
integrated analysis and planning of port and land transportation systems.

Table 1.
World container traffic

Year Container traffic Growth rate

(Millions TEU) (%)

1993 1132 125

1995 1372 98
1997 1535 42
1999 2032 10
2000 2253 109
2001 2316 28
2002 2406 39
2003 2546 58
2004 2800 106
2005 3105 109
 Ch. 8. Intermodal Transportation 471

With respect to ocean navigation, efficiency reasons have led to the con-
struction of very large container-carrying ships for intercontinental move-
ments; new container ship capacities are of the order of 8000 to 10,000 TEUs.
To operate efficiently, such ships must not stop frequently. Moreover, they are
too large for the vast majority of ports. Consequently, a new link has been
added to the intermodal chain: super-ships stop at a small number of major
seaports and containers are transferred to smaller ships for distribution to var-
ious smaller ports. Notice that these ships cannot navigate through the Panama
Canal. This results in a modification of sea and land shipping routes. The im-
pact of these modifications on particular transportation systems and facilities
may be analyzed by using the models of Section 6.
The impact of the growth in containerized trade has been significant on
land transportation systems as well. Specialized transportation services have
been created, such as the North-American “land-bridges” that provide con-
tainer transportation by long, double-stack trains operated by independent
subsidiaries of the rail companies between the East and the West coasts and
between these ports and the industrial core of the continent. Dedicated rail
services are also being created in Europe to provide shuttle services for con-
tainer transportation. Section 3.2.3 examines some of the issues associated to
planning and operating such dedicated services.

2.2 Customized, consolidation, and intermodal transportation

In an intermodal chain, one may encounter consolidation transportation

systems, where one vehicle or convoy serves to move freight for different
customers with possibly different initial origins and final destinations, and
customized transportation carriers that provide dedicated service to each par-
ticular customer.
Truckload trucking offers a typical example of customized transportation.
When the customer calls, the dispatcher assigns to the task a truck and driver
(or driving team for very long movements). The truck moves to the customer-
designated location, is loaded, and then moves to the specified destination
where it is unloaded. The driver then calls the dispatcher to give its position
and request a new assignment. The dispatcher may indicate a new load, ask
the driver to move empty to a new location where demand should appear in
the near future, or have the driver wait and call later.
The truckload carrier thus operates in a highly dynamic environment, where
little is known with certainty regarding future demands, waiting delays at cus-
tomer locations, precise positions of loaded and empty vehicles at later mo-
ments in time, and so on. Moreover, the time available to the dispatcher to
decide on the next assignment in response to a customer or driver request, is
generally very short (most such decisions are performed in real time). Service
is tailored for each customer and the timely assignment of vehicles to profitable
demands is very important. The development of efficient resource management
and allocation strategies is therefore at the heart of the management process.
472 T.G. Crainic and K.H. Kim

These strategies attempt to maximize the volume of demand satisfied (loads

moved) and the associated profits, while making the best use of the available
resources: drivers, tractor and trailer fleets, etc. We do not address these issues
in this chapter. The interested reader may consult Powell (2003) or Powell et
al. (2007), for example. Notice, however, that most methodology targeting re-
source management and allocation issues for customized carriers may also be
adapted to address operational-level fleet management issues for consolida-
tion carriers (Section 4).
Ocean navigation services provided by for-hire ships share most of these dy-
namic and stochastic characteristics. Variations in travel times are in fact larger
for sea than for road trips. On the other hand, travel and loading/unloading
times are usually much longer, which increases the time available to decide
on the next assignment. Railways also provide customized services by dedi-
cating unit trains to individual shippers. Such decisions respond in most cases
to long-term contracts, which implies that demand is known deterministically.
This contrasts to the stochastic demand that characterizes truckload trucking
and for-hire ocean shipping. The setup of such unit trains is therefore part of
the regular planning process of operations as described later on in this section.
Customized transportation is not always the appropriate answer to shipper
needs. The relations and trade-offs between volume and frequency of ship-
ping, on the one hand, and the cost, frequency, and delivery time of trans-
portation, on the other hand, often dictates the use of consolidation trans-
portation services. In passenger transportation, this choice is equivalent, for
example, between taking a taxi or using the services of a regular public trans-
port line. Freight consolidation transportation is performed by Less-Than-
Truckload (LTL) motor carriers, railways, ocean shipping lines, regular and
express postal services, etc. Freight transportation in some countries where a
central authority more or less controls a large part of the transportation system
also belongs to this category.
Consolidation transportation carriers and fundamentally all intermodal
transportation systems are organized as so-called hub-and-spoke networks, il-
lustrated in Figure 1. In such systems, service is offered between a certain num-
ber of origin–destination points (the local/regional terminals), represented by
nodes 1–9 in Figure 1. This number is significantly larger than the number
of direct, origin to destination services operated by the carrier. Consequently,
and to take advantage of economies of scale, low-volume demands are moved
first to an intermediate point – a consolidation terminal or a hub (nodes A,
B, and C in Figure 1) – such as an airport, seaport container terminal, rail
yard, or intermodal platform. At a hub, traffic is consolidated into larger flows
that are routed to other hubs by high-frequency, high-capacity services (thick
full and dashed lines in Figure 1). As illustrated in the figure, more than one
service, of possibly different modes, may be operated between hubs. Lower fre-
quency services, often operating smaller vehicles, are used between hubs and
origin–destination terminals (the regular lines in the figure). When the level of
demand justifies it, high-frequency, high-capacity services may be run between
 Ch. 8. Intermodal Transportation 473

Fig. 1. Network with consolidation terminals/hubs.

a hub and a regional terminal (link between nodes A and 9) or between two
regional terminals (links between nodes 4 and 5). Notice that cargo is brought
to and distributed from origin–destination terminals by vehicles performing ei-
ther pickup and distribution routes or customized services. LTL motor carriers
and postal services are examples of the first case, rail and ocean shipping of
containers (barges or trucks performing the local transportation activities) of
the second. Yet, because the planning of the long-haul activities of consolida-
tion carriers does not generally include these activities, we do not elaborate
further on the topic.
A hub-and-spoke organization allows a much higher frequency of service
between all origin–destination pairs in the network and a more efficient uti-
lization of resources. The drawback of this type of organization is increased
delays due to longer routes and the time spent in terminals. This explains partly
why there are very few “pure” hub-and-spoke systems in operation; direct
transportation is usually operated for high-demand or high-priority origin–
destination pairs. A second strategy is to separate high- and low-priority traffic
and to dedicate different services and, eventually, infrastructure (e.g., termi-
nals) to each. An example of this trend is the creation of intermodal subdivi-
sions by North-American railways to ensure efficient movements of containers
among the major ports and industrial centers of the continent.
To further mitigate the drawbacks of hub-based operations, consolidation
carriers engage into rather sophisticate planning activities (Section 3). The car-
rier operates a series of services, each characterized by its own route, stops,
frequency, vehicle and convoy type, capacity, speed (travel time), and so on.
Internally, services are often collected in an operational plan (also referred to
as load or transportation plan), generally accompanied by a schedule that indi-
cates departure and arrival times at the terminals of the route. The schedule
474 T.G. Crainic and K.H. Kim

is partially (e.g., latest delivery of cargo at the origin terminal for on-time de-
livery at destination) or totally available to customers. The aim of the load
plan is to ensure that the proposed services are performed as stated (or as
closely as possible), while operating in a rational, efficient, and profitable way.
It also indicates how demand is moved though the system using its terminals
and transportation services.
Freight demand is defined by its specific origin, destination, and commodity-
related physical characteristics (e.g., weight and volume), as well as particular
service requirements in terms of delivery conditions, type of vehicle, and so
on. A profit or cost is also usually associated to a given demand. Once deliv-
ered at the carrier’s terminal, the cargo of several customers is sorted, grouped,
and loaded into the same vehicle or convoy. It is then moved either directly,
when such a service exists, or through a series of services with intermediary op-
erations of transfer and consolidation. More than one consolidation-transfer
operation may occur during a trip. In the case of LTL motor carriers, and often
for postal services, it is individual loads that are consolidated and loaded into
vehicles, trucks, planes, or rail cars. Containerized cargo is not handled before
reaching its destination and, thus, consolidation operations involve only the
containers, which are loaded into ships, airplanes, or rail cars. This constitutes
a first type of consolidation operation: “small” loads into vehicles, containers
into ships, etc. In some cases, vehicles are further sorted and consolidated into
convoys. The most widely-spread and well-known case is that of railway trans-
portation where, first, cars are grouped into blocks (i.e., the cars in a block
travel together, without any re-consolidation operation, until some common,
intermediary or final, terminal) and, then, blocks are put together to make up
trains. Similar make-up operations, albeit on a smaller scale, also occur for
barge trains and multitrailer assemblies.
Terminals are clearly an important component of consolidation and inter-
modal transportation systems, and their efficiency is vital to the performance
of the entire transport chain. The following subsection briefly describes main
terminal operations and planning issues with a particular focus on container
port terminals.

2.3 Consolidation and intermodal terminals

Terminals come in several designs and sizes and may be specialized for par-
ticular transportation modes and the handling of specific products, or may
offer a complete set of services. Major operations performed in consolidation
terminals include vehicle loading and unloading, cargo and vehicle sorting and
consolidation, convoy make up and break down, and vehicle transfer between
services.
When containerized traffic is concerned, only the containers are handled,
not the cargo they contain. Thus, once loaded on a truck, rail car, barge, or
ocean vessel, containers will generally follow the movements and consolidation
activities of the respective vehicle, until reaching either the final destination or
 Ch. 8. Intermodal Transportation 475

an intermodal terminal. There are very few such activities for motor carriers
and barges. We may mention the transfer of a barge or a trailer from a barge
or road train to another. No particular planning model is built for such activ-
ities; rather, these are included in the scope of the models used to build the
transportation plan of the carrier (Section 3.2).
More complex consolidation operations are performed within railway sys-
tems: sorting and consolidation of rail cars into blocks and trains. A rich liter-
ature exists on models targeting these issues and is reviewed in the references
indicated at the beginning of the section. In most cases, however, there are no
differences between intermodal and regular rail traffic with respect to blocking
and train make up planning and operations, even when particular terminals are
dedicated to handling intermodal traffic. Research is under way to develop the
“next generation” intermodal rail services and terminals, but the efforts dedi-
cated to the associated planning issues are still limited (e.g., Bostel and Dejax,
1998; Macharis and Bontekoning, 2004).
The intermodal transfer of containers between truck and rail, taking place at
rail terminals, is specific to intermodal transportation. Containers thus arrive
at the rail terminal by truck and are either directly transferred to a rail car or,
more frequently, are stacked in a waiting area. Then, containers are picked up
from the waiting area and loaded unto rail cars that will be grouped into blocks
and trains. The reverse operations take place when containers arrive by train
to the terminal and are to be transfered to trucks for their next transport leg.
We did not find any paper dedicated to the planning of these operations. The
issues are very similar to those arising in container port terminals, however,
which are probably the most well-known intermodal transfer facilities. Signif-
icant research has been dedicated to the design (Section 3.3) and particularly
to the operations (Section 5) of container port terminals. In the following, we
briefly describe these operations and planning issues.
The main function of a container port terminal is to provide transfer facili-
ties for containers between sea vessels and land transportation modes, trucks
and rail in particular. It is a highly complex system that involves numerous
pieces of equipment, operations, and container handling steps (Steenken et
al., 2004). The assignment of resources to tasks and the scheduling of these
tasks are thus among the major container port terminal planning issues. Three
main areas make up a container terminal. The sea-side area encompasses the
quays where ships berth and the quay cranes that provide the loading and
unloading of containers into and from ships. The land-side area provides the
interface with the land transportation system (the so-called hinterland of the
port) and encompasses the truck and train receiving gates, the areas where
rail cars are loaded and unloaded, and the associated equipment. Trucks are
generally loaded and unloaded directly in the yard area. This third area is dedi-
cated for the most part to stacking loaded and empty containers for import and
export (in some terminals, facilities are also provided for the loading and load-
ing of containers). Various types of yard cranes are associated with this area.
So-called transporters, primarily yard trucks or automated vehicles, move con-
476 T.G. Crainic and K.H. Kim

Fig. 2. Example of a container terminal with an indirect transfer system (Park, 2003).

tainers between the three areas. Figure 2 illustrates part of a container port
terminal. One ship and three quay cranes are displayed in the sea-side area,
while only trucking is shown in the land-side area. Twelve container stacks are
displayed in the yard area, as well as one type of yard crane used to transfer
containers between yard transporters and outside trucks and stacks, as well as
to change the position of containers in the yard as required.
Three main types of handling operations are performed in a container ter-
minal:
(1) ship operations associated with berthing, loading, and unloading con-
tainer ships,
(2) receiving/delivery operations for outside trucks and trains, and
(3) container handling and storage operations in the yard.
When a ship arrives at the container port terminal, it is assigned a berth and a
number of quay cranes. Berth space is a very important resource in a container
terminal (construction costs to increase capacity are very high, even when
space for growth exists) and berth scheduling determines the berthing time and
position of a container ship at a given quay (Section 5.1). Quay-crane alloca-
tion is the process of determining the vessel that each quay crane will serve and
the associate service time (Section 5.2). Stowage sequencing determines the se-
quence of unloading and loading containers, as well as the precise position
each container being loaded into the ship is to be placed in (Section 5.3). Dur-
 Ch. 8. Intermodal Transportation 477

ing the unloading operation, a quay crane transfers a container from a ship to
a transporter. Then, the transporter delivers the import (unloading) container
to a yard crane that picks it up and stacks it into a given position in the yard.
This sequence of operations is called indirect transfer. Some terminals use a
direct transfer system where the equipment used to move containers between
the quay and the yard will also stack them. For export (loading) operation, the
process is carried out in the opposite direction.
On the land-side, the receiving and delivery operations provide the inter-
face between the container terminal activities and the external movements.
A receiving operation starts when containers arrive at the gate of the terminal
carried by one or several outside trucks or a train. Containers are inspected at
the gate to check for damages (to the container not its content) and whether
all documents are in order. Also at the gate, information regarding where the
container is to be stored is provided to the truck driver. When the outside truck
arrives at the indicated transfer point, a yard crane lifts a container from the
truck and stacks it according to the plan. When containers arrive by rail, the
rail cars are brought to the rail area where containers and documents are ex-
amined. Containers are then transferred by a gantry crane to a transporter,
which delivers them to the yard and stacks them. In the case of a delivery op-
eration, the yard equipment delivers a container onto an outside truck, which
leaves the port, or onto a transporter which delivers the container to the rail
area and loads it onto the designated car.
The sea- and land-side operations interact with the yard container handling
and storage operation through the information on where the containers are
or must be stacked within the yard. How containers are stored in the yard is
one of the important factors that affect the turn-around time of ships and land
vehicles. The space-allocation problem is concerned with determining storage
locations for containers either individually or as a group. Yard storage space is
pre-assigned to containers of each ship arriving in the near future to maximize
the productivity of the loading and unloading operations (Section 5.4).
A container yard consists of blocks of containers, which are separated by
aisles for transporters as shown in Figure 2. A block consists of 25–35 yard bays,
and a yard bay has 6–10 stacks of containers. Container handling and storage
operations include the management and handling of containers while they are
in storage in the yard and thus occur between the receiving and delivery op-
erations and the ship operations. Container-handling equipment performs the
placement of containers into storage and their retrieval when needed. Yard
cranes move along blocks of containers to yard bays to perform these opera-
tions. Planning these operations is part of the equipment-assignment process,
which allocates tasks to container-handling equipment. Based on the quay-
crane schedule, one or two yard cranes are assigned to each quay crane for
loading and unloading. The remaining yard cranes are allocated to receiving
and delivery operations. Terminal operators aim to assign and operate yard
cranes in such a way that inefficient moves and interferences among yard
cranes are minimized (Section 5.5).
478 T.G. Crainic and K.H. Kim

3 System and service network design

This section is dedicated to models aimed at strategic and tactical plan-

ning issues for freight carriers and seaport container terminals. At the strategic
level, we focus on decisions that concern the design of the physical infrastruc-
ture network:
where to locate terminals (e.g., consolidation terminals, rail yards, in-
termodal platforms, and so on)?
what type and quantity of equipment (e.g., cranes) to install at each
facility?
what type of lines or capacity to add?
what lines or facilities to abandon?
which customer zones to serve directly and how to serve the others?
and so on.
We group these issues under the label system design. The term service network
design covers tactical planning issues for consolidation transportation firms:
on what routes to provide service?
what type of service (mode) to use?
how often to offer service on each route and according to what sched-
ule?
how to route the loads through the physical and service networks?
how to distribute the work among the terminals of the system?
The output of the process is a transportation (load) plan. The goals are cus-
tomer satisfaction and cost-efficient utilization of resources (assets) leading to
profits.
Most freight carriers face system and service network design issues, irre-
spective of their involvement in intermodal transportation activities. It is not
our intention, however, to present a comprehensive and detailed treatment of
the subject. We focus rather on the main issues and modeling efforts that bear
directly on intermodal transportation.

3.1 Models for system design

The literature on system design models for freight transportation is not

rich. Such issues are often addressed by evaluating alternatives using network
models for tactical or operational planning of transportation activities. When
formal models are proposed, they generally take the form of discrete loca-
tion formulations to address issues related to the location of consolidation
or hub terminals and the routing of demand from its origin to its destination
terminals. The routing of flows determines the direct connections (physical
infrastructure or service links) between origins, destinations, and consolida-
tion terminals. When these connections must be explicitly decided (e.g., the
allocation of “local” terminals to major classification facilities), a combined
 Ch. 8. Intermodal Transportation 479

location-network-design formulation is often used. All formulations aim to

capture the potential economies of scale associated with the consolidation of
freight.
An extensive literature exists on location and network design models and
solution methods: Mirchandani and Francis (1990), Daskin (1995), Drezner
(1995), Labbé et al. (1995), Labbé and Louveaux (1997), Crainic and Laporte
(1997), Drezner and Hamacher (2002), and Daskin and Owen (2003) re-
view location issues and literature, while Magnanti and Wong (1984), Minoux
(1989), Nemhauser and Wolsey (1993), Salkin and Mathur (1989), Ahuja et al.
(1993), Balakrishnan et al. (1997), and Crainic (2000) survey the network de-
sign field.
In the following, we focus on the contributions related to intermodal trans-
portation, the location of consolidation facilities in particular. All the models
presented are static and deterministic location formulations. They assume that
all problem components, particularly the demand as well as the cost and profit
structure, will not vary during the planning period for which their evaluation is
performed, usually from six to twelve months. Moreover, problems where the
design decisions have long-term effects (e.g., location of hubs) also assume
that variances will be negligible for the foreseeable future. These observa-
tions emphasize the need for research into time-dependent stochastic models
for system design problems. Most problems presented in this subsection are
NP-hard and thus heuristics are the solution method of choice in almost all
cases.

3.1.1 Location with balancing requirements

The multicommodity location problem with balancing requirements was
first introduced by Crainic et al. (1989) in the context of the management of a
heterogeneous fleet of containers by an international maritime shipping com-
pany. The land operations of the carrier proceed as follows. Once a ship arrives
at port, the company has to deliver loaded containers, which may come in sev-
eral types and sizes, to designated in-land destinations. Following their unload-
ing by the importing customer, empty containers are moved to a depot. From
there, empty containers may be delivered to customers which request them for
subsequent shipping of their own products. In addition, empty containers are
often moved, or repositioned, to other depots. These interdepot movements
are a consequence of regional imbalances in empty container availabilities and
needs throughout the network: some areas lack containers of certain types,
while others have surpluses. These balancing movements of empty contain-
ers among depots differentiate this problem from classical location–allocation
applications. Interdepot movements are performed by high-density transport
(rail, typically) and their unit transportation cost is lower than for the other
types of movements. The general problem is therefore to locate depots and
allocate customers to depots (for each type of container and direction of move-
ment) in order to collect the supply of empty containers available at customers’
sites and to satisfy the customer requests for empty containers, while minimiz-
480 T.G. Crainic and K.H. Kim

ing the total operating costs: the costs of opening and operating the depots,
and the costs generated by customer-depot and interdepot movements.
The formulation proposed by Crainic et al. (1989) is based on a directed
network G = (N  A), where N is the set of nodes or vertices and A is the set
of arcs or links. The set of nodes is partitioned into three subsets: O, the set
of origin nodes (supply customers); D , the set of destination nodes (demand
customers); and H, the set of hubs or consolidation nodes (depots). The sets
of customers adjacent to each depot j ∈ H are identified as O(j) = {i ∈
O: (i j) ∈ A} and D(j) = {i ∈ D: (j i) ∈ A}. It is assumed that there exists
at least one origin or destination adjacent to each depot j (O(j) ∪ D (j) = ∅).
The sets of depots adjacent to each node i ∈ N in both directions are defined
as H+ (i) = {j ∈ H: (i j) ∈ A} and H− (i) = {j ∈ H: (j i) ∈ A}. The
problem description does not allow direct customer-to-customer movements
of containers. Hence, the set of arcs A may be partitioned into three subsets:
customer-to-depot arcs, AOH = {(i j) ∈ A: i ∈ O j ∈ H}; depot-to-customer
arcs, AHD = {(i j) ∈ A: i ∈ H j ∈ D }; and depot-to-depot arcs, AHH =
{(i j) ∈ A: i ∈ H j ∈ H}. The commodities (types of containers) that move
through the network are represented by the set P .
p
For each supply customer i ∈ O, the supply of commodity p is noted oi  0,
while for each demand customer i ∈ D , the demand for commodity p is noted
p p
dk  0. A nonnegative cost cij is incurred for each unit of flow of commodity
p moving on arc (i j). In addition, for each depot j ∈ H, a nonnegative fixed
cost fj is incurred if the depot is opened.
p
Let xij represent the flow of commodity p moving on arc (i j), and yj the
binary location variable that takes value 1 if depot j is opened, and value 0
otherwise. The problem is then formulated as:
   p p
 p p
minimize fj y j + cij xij + cji xji
j∈H p∈P (ij)∈AOH (ji)∈AHD
 
p p
+ cjk xjk (1)
(jk)∈AHH
subject to
 p p
xij = oi  i ∈ O p ∈ P  (2)
j∈H+ (i)
 p p
xji = dk  i ∈ D p ∈ P  (3)
j∈H− (i)
 p
 p
 p
 p
xji + xjk − xij − xkj = 0
i∈D(j) k∈H+ (j) i∈O(j) k∈H− (j)
j ∈ H p ∈ P  (4)
p p
xij  oi yj  j ∈ H i ∈ O(j) p ∈ P  (5)
 Ch. 8. Intermodal Transportation 481

p p
xji  dk yj  j ∈ H i ∈ D (j) p ∈ P  (6)
p
xij  0 (i j) ∈ A p ∈ P  (7)
yj ∈ {0 1} j ∈ H (8)
Constraints (2) and (3) ensure that supply and demand requirements are
met, relations (4) correspond to flow conservation constraints at depot sites,
while Equations (5) and (6) forbid customer-related movements through
closed depots. The analogous constraints for the interdepot flows are redun-
dant if costs satisfy the triangle inequality and are not included in the for-
mulation. The problem has been addressed by dual-ascent heuristics (Crainic
and Delorme, 1995), sequential and parallel branch-and-bound (Crainic et al.,
1993a; Gendron and Crainic, 1995, 1997; Bourbeau et al., 2000), sequential
and parallel tabu search (Crainic et al., 1993c, 1995a, 1995b, 1997; Gendron et
al., 1999). Gendron et al. (2003a, 2003b) have studied the capacitated version
of the problem and proposed sequential and parallel metaheuristics.

3.1.2 Multicommodity production distribution

The multicommodity production-distribution problem is a simplified ver-
sion of the previous problem where no interhub movements are considered.
In this case, commodities p ∈ P may be shipped from their origins i ∈ O to
their destinations k ∈ D either directly or via a consolidation terminal j ∈ H.
(Note that sets O, D , and H are not necessarily disjoint.) The main decisions
addressed by the production-distribution formulations are the number and lo-
cation of consolidation terminals and the product flow pattern through the
system, either directly from the origin to destination or through a consolida-
tion terminal. Particularly relevant for intermodal transportation system design
is the representation of the economies of scale in transportation costs associ-
ated with concentrated flows (e.g., Croxton et al., 2003, 2006).
The capacity of a consolidation terminal located at site j ∈ H is denoted uj .
The transportation cost per unit of flow of commodity p from origin i to des-
p p
tination k transiting through consolidation terminal j is denoted cijk , while cik
stands for the unit transportation cost of direct movements. The correspond-
p p
ing flow routing decision variables are denoted xijk and xik , respectively. All
other notation is the same as in the previous problem. The model is then:
    p p  p p 
minimize fj y j + cik xik + cijk xijk (9)
j∈H p∈P i∈O j∈H k∈D
subject to
p  p p
xijk  min oi  uj  dk yj  i ∈ O j ∈ H k ∈ D  p ∈ P  (10)
 p
xijk  uj yj  j ∈ H (11)
p∈P i∈O k∈D
482 T.G. Crainic and K.H. Kim

 p
 p

p
xik + xijk  oi  i ∈ O p ∈ P  (12)
k∈D j∈H
 p
 p p
xik + xijk ) = dk  k ∈ D p ∈ P  (13)
i∈O j∈H

yj = 0 or 1 j ∈ H (14)
p p
xijk  0 xik  0 i ∈ O  k ∈ D  j ∈ H p ∈ P  (15)

In this formulation, expression (9) minimizes the total cost of opening the
consolidation centers and distributing the commodities. Constraints (10) and
(11) ensure that only open terminals are used. Relation (11) also enforces the
consolidation-terminal capacity constraint. Commodity supplies and demands
are enforced by relations (12) and (13), respectively. When demand is specified
by origin-to-destination pairs
p
dik : Quantity of commodity p to be transported from origin terminal i
to destination terminal k,
the previous formulation is modified by replacing relations (10) with (16) and
constraints (12) and (13) with (17):
p  p
xijk  min uj  dik yj  i ∈ O j ∈ H k ∈ D  p ∈ P  (16)
p
 p p
xik + xijk = dik  i ∈ O k ∈ D  p ∈ P  (17)
j∈H

Many variants of these formulations may be found in the location literature

dealing with practical considerations such as partial, commodity-specific ca-
pacities at terminals or handling costs at consolidation terminals (e.g., Aikens,
1985). Piecewise linear functions are often used to represent economies of
scale in transportation costs associated with concentrated flows (e.g., Croxton
et al., 2003, 2006). Klincewicz (1990) proposed such a formulation for an
uncapacitated problem with single-sourcing for each commodity and no di-
rect origin-to-destination movements. When linear transportation costs are
assumed either for the origin to consolidation center movements or from the
latter to destination terminals, the problem decomposes into concave-cost un-
capacitated location problems that can be formulated as uncapacitated plant
location problems (with one facility for each piecewise linear segment) for
which efficient solution methods exist (e.g., Erlenkotter, 1978). Heuristics that
solve a series of these formulations are proposed for the general case. Heuris-
tics combining Lagrangian relaxations and ad-hoc rules based on the logis-
tics characteristics of the system (costs and distances, principally) have also
been proposed for the capacitated version of the problem (9)–(15) (Pirkul and
Jayaraman, 1996, 1998).
 Ch. 8. Intermodal Transportation 483

3.1.3 Hub-location models

Consolidation transportation systems are generally structured as hub-and-
spoke networks (Section 2 and Figure 1). Hub-location models may be used
when one must decide simultaneously the location of the hubs (consolida-
tion terminals) and the allocation of regional terminals or customers to the
hubs. The location with balancing requirements and the multicommodity
production-distribution problems presented in the previous subsections are
particular hub-location cases where the allocation of regional terminals or cus-
tomers to potential hubs does not involve any setup (fixed) cost or has already
been decided, respectively. Hub-location formulations appear in many appli-
cation areas (Campbell, 1994b) and several variants have been proposed and
studied (Campbell et al., 2002; Ebery et al., 2000).
The basic model assumes that all traffic passes through two hubs on its route
from its origin to its destination, no hub capacities, no direct transport be-
tween nonhub terminals, and no fixed costs for establishing a link between
a regional and a consolidation terminal. In all hub-based problems, interhub
transportation is assumed to be more efficient due to the concentration of
flows. Consequently, interhub links are assigned a lower unit cost than links
representing the other movements in the system. Most of the notation is sim-
ilar to that of the previous models. Let us add the following decision-variable
definitions:
yj = 1, if a consolidation terminal is located at site j and 0, otherwise;
yij = 1, if terminal i is linked to hub j and 0, otherwise;
p
xijlk : Flow of commodity p with origin i and destination k that passes
through hubs j and l, in that order.
The following formulation assumes that exactly P hubs have to be located
out of the |H| potential sites:
   p   p 
minimize cij yij xijlk
p∈P i∈O j∈H l∈H k∈D
   
p p
+ clk ylk xijlk
l∈H k∈D i∈O j∈H
   
p p
+ cjl yij ylk xijlk (18)
j∈H l∈H i∈O k∈D
subject to

yj = P (19)
j∈H
 p p
xijlk = dik  i ∈ O k ∈ D p ∈ P  (20)
j∈H l∈H
yij  yj  i ∈ O  j ∈ H (21)
484 T.G. Crainic and K.H. Kim

ylk  yl  k ∈ D  l ∈ H (22)
p p
xijlk  dik yj  i ∈ O k ∈ D  j l ∈ H p ∈ P  (23)
p p
xijlk  dik yl  i ∈ O k ∈ D  j l ∈ H p ∈ P  (24)
yj = 0 or 1 j ∈ H (25)
yij = 0 or 1 i ∈ O  j ∈ H (26)
p
xijlk  0 i ∈ O k ∈ D  j l ∈ H p ∈ P  (27)
The objective function (18) minimizes the total transportation cost of the
system. Due to its similarity with the p-median location problem, this class
of formulations is called the p-hub median problem (Campbell, 1994a). Note,
however, that despite the name similarity, several properties of the p-median
problem do not hold for the p-hub problem (e.g., the assignment of demand
nodes to the closest open facility is not optimal for p-hub median problems).
Formulation (18)–(27) was first introduced by O’Kelly (1987). Campbell
(1994a) introduced the first linearization mechanism, later refined by Campbell
(1996), Skorin-Kapov et al. (1996), and O’Kelly et al. (1996). The mechanism
yields a path-based, linear mixed-integer formulation:
  p p
minimize cijlk xijlk (28)
p∈P i∈O j∈H l∈H k∈D

subject to (19), (20), (23)–(25), and (27), where

p
cijlk : Unit cost of transportation for commodity p from origin i to desti-
nation k with consolidation at hubs j and l, in that order.
The two preceding formulations allow multiple allocations of origin and des-
tination terminals to hubs. To enforce single allocation of terminals to hubs,
constraints (23) and (24) are replaced in the first formulation by

yij = 1 i ∈ O (29)
j∈H

ylk = 1 k ∈ D (30)
l∈H
p
xijlk
p = yij  i ∈ O k ∈ D  j l ∈ H p ∈ P  (31)
dik
p
xijlk
p = ylk  i ∈ O k ∈ D  j l ∈ H p ∈ P  (32)
dik
Both these constraints and the yij , ylk variables must be added to the linearized
formulation to enforce single terminal allocation.
Both hub-location formulations are difficult to solve. Heuristics have thus
been mainly proposed (e.g., O’Kelly, 1987; Aykin, 1990; Klincewicz, 1991;
 Ch. 8. Intermodal Transportation 485

Campbell, 1996). Metaheuristics have led to improved results: Neural Net-

works (Smith et al., 1996), Simulated Annealing (Ernst and Krishnamoorthy,
1996), and Tabu Search (Klincewicz, 1992; Skorin-Kapov and Skorin-Kapov,
1994). Lower bounds have also been proposed (e.g., O’Kelly, 1992a; O’Kelly
et al., 1995; Skorin-Kapov et al., 1996), as well as a number of methods that
find the optimal solution to problem instances of limited size (Skorin-Kapov
et al., 1996; O’Kelly et al., 1996; Ernst and Krishnamoorthy, 1996).
A number of important extensions to these models may be formulated.
Fixed cost terms j∈H fj yj added to the objective function (and dropping
constraint (19) on the number of hubs to open) may be used to represent
the costs of opening consolidation centers (O’Kelly, 1992b). Klincewicz (1996)
proposed dual-based heuristics for this formulation, while Addinnour-Helm
and Venkataramanan (1998) developed a genetic search metaheuristic and a
branch-and-bound algorithm. Limited-sized problem instances have been ad-
dressed.
Additional extensions consider fixed costs for allocating origin and destina-
tion terminals to hubs, hub capacities, more complex routing patterns (e.g.,
more than two hubs), minimum levels of traffic in order to allow a terminal-
to-hub connection, etc., and are better able to capture the complexity of trans-
portation systems. Such hub network design formulations are very difficult to
solve, however, and relatively little research had been conducted (e.g., Aykin,
1994, 1995a, 1995b; Kuby and Gray, 1993; Jaillet et al., 1996). Significant re-
search is required in this domain.

3.2 Service network design

Service network design formulations are used to build a transportation (or

load) plan to ensure that the system operates efficiently, serves demand, and
ensures the profitability of the firm. The physical infrastructure (e.g., the termi-
nal locations) and the available resources are fixed for these problems. Service
network design problems address the system-wide planning of operations to
decide the selection, routing, and scheduling of services, the consolidation
activities in terminals, and the routing of freight of each particular demand
through the physical and service network of the company. The goal is cost-
efficient operation together with timely and reliable delivery of demand ac-
cording to customer specifications and the targets of the carrier.
Service network design problems are difficult due to the strong interactions
among system components and decisions and the corresponding tradeoffs be-
tween operating costs and service levels. To illustrate, consider the routing of
a shipment between two terminals of the intermodal transportation system il-
lustrated in Figure 1. (Assume, for simplicity, that each link corresponds to
a service.) A shipment that originates at terminal 4 with destination termi-
nal 8 may be routed according to a number of strategies, including (1) direct
through its designated hub B; (2) indirect through hub A or terminal 5 and
486 T.G. Crainic and K.H. Kim

then through B; (3) indirect through hubs A and C. The number of strate-
gies increases rapidly when one considers additional factors such as the type
of operation (e.g., consolidation or simple cargo transfer from one service to
another) or consolidation policy.
Which alternative is the “best”? Each has its own cost and time measures
that result from the characteristics of each terminal and service, as well as
from the routing of all other shipments. Thus, for example, strategies based
on re-consolidation and routing through intermediate terminals may be more
efficient when direct services are offered rarely due to low levels of traffic de-
mand. Such strategies would probably result in higher equipment utilization
and lower waiting times at the original terminal; hence, in a more rapid service
for the customer. The same decision would also result, however, in additional
unloading, consolidation, and loading operations, creating larger delays and
higher congestion levels at terminals, as well as a decrease in the total reli-
ability of the shipment. Alternatively, offering direct service (introducing the
service or increasing its frequency) would imply faster and more reliable ser-
vice for the corresponding traffic and a decrease in the level of congestion at
some terminals, but at the expense of additional resources, thus increasing the
direct costs of the system.
Service network design thus integrates two types of major decisions. The
first is to determine the service network, that is, to select the routes – origin
and destination terminals, physical route and intermediate stops – on which
services will be offered and the characteristics of each service: mode, frequency
or schedule, etc. The second major type of decision is to determine the rout-
ing of demand, that is, the itineraries used to move the flow of each demand:
services and terminals used, operations performed in these terminals, etc. The
service network specifies the movements through space and time of the vehi-
cles and convoys of the various modes considered. Operating rules indicating,
for example, how cargo and vehicles may be sorted and consolidated are some-
times specified at particular terminals and become part of the service network
(this is the case, in particular, for rail carriers). The itineraries used to move
freight from origins to destinations determine the flows on the services and
through the terminals of the service network.
Minimization of the total operating costs is the primary optimization cri-
terion, reflecting the traditional objectives of freight carriers and expectation
of customers to “get there fast at lowest possible cost”. Increasingly, however,
customers not only expect low rates, but also require a high-quality service,
measured by speed, flexibility, and reliability. Service performance measures
modeled, in most cases, by delays incurred by freight and vehicles or by the
respect of predefined performance targets are then added to the objective
function of the network optimization formulation. The resulting generalized
cost function thus captures the tradeoffs between operating costs and service
quality.
Several efforts have been directed toward the formulation of service net-
work design models. Reviews are presented by Assad (1980), Crainic (1988,
 Ch. 8. Intermodal Transportation 487

2000, 2003), Delorme et al. (1988), and Cordeau et al. (1998). Most of these
contributions aim specific carrier types and modes of transportation (rail, less-
than-truckload trucking, navigation, etc.) and are thus covered in more depth
in other chapters of this book. In the following, we briefly present basic mod-
eling approaches and examine a number of contributions aimed at intermodal
transportation.
One may distinguish between static and time-dependent service network de-
sign formulations. The former assume that demand does not vary during the
planning period that is considered. The time dimension of the service network
is then implicitly considered through the definition of services and interservice
operations at terminals. Time-dependent formulations include an explicit rep-
resentation of movements in time and usually target the planning of schedules
to support decisions related to when services depart. Most service network de-
sign models proposed in the literature take the form of deterministic, fixed cost,
capacitated, multicommodity network design formulations.

3.2.1 Static service network design

Service network design assumes a physical network G = (N  A) on which
the carrier may move its vehicles. On this network, the carrier operates (or
makes use of) a number of terminals, with various characteristics and func-
tionalities (loading/unloading stations, regional terminals with limited sorting
capabilities, consolidation hubs, etc.). Let nodes in H ⊆ N correspond to lo-
cations where the terminals are situated and assume, for simplicity, that all
terminals can perform all operations.
The service network specifies the transportation services that could be of-
fered. Each service s ∈ S is characterized by its (1) mode, which may represent
either a specific transportation mode (e.g., rail and truck services may belong to
the same service network), or a particular combination of traction and service
type (e.g., fast navigation lines by large container ships providing shuttle ser-
vice between South-East Asia and the west coast of North America); (2) route,
defined as a path in A, from its origin terminal to its destination terminal, with
intermediary terminals where the service stops and work may be performed
(e.g., a navigation line that stops at pre-defined ports for unloading and load-
ing containers); (3) capacity, which may be measured in load weight or volume,
number of containers, number of vehicles (when convoys are used to move sev-
eral vehicles simultaneously), or a combination thereof; (4) service class that
indicates characteristics such as preferred traffic or restrictions, speed and pri-
ority, etc. To design the service network thus means to decide what service to
include in the transportation plan such that the demands and the objectives
of the carrier are satisfied. When a service may be operated repeatedly during
the planning period (e.g., several similar ships departing during the month and
visiting the same ports in the same order), the design must also determine the
frequency of each service.
Crainic and Rousseau (1986) proposed a multimodal multicommodity path-
flow service network design formulation. In their model, a commodity p ∈ P is
488 T.G. Crainic and K.H. Kim

defined as a triplet (origin, destination, type of product (or vehicle)) and traffic
moves according to itineraries. An itinerary l ∈ Lp for commodity p specifies
the service path used to move (part of) the corresponding demand: the origin
and destination terminals, the intermediary terminals where operations (e.g.,
consolidation and transfer) are to be performed, and the sequence of services
between each pair of consecutive terminals where work is performed. The de-
mand of product p is denoted dp . Flow routing decisions are then represented
p
by decision variables hl indicating the volume of product p moved by using
its itinerary l ∈ Lp . Service frequency decision variables ys , s ∈ S , define the
level of service offered, i.e., how often each service is run during the planning
period.
p
Let y = {ys } and h = {hl } be the decision-variable vectors. The model
minimizes the total generalized system cost, while satisfying the demand for
transportation and the service standards:
   p
minimize Fs (y) + Cl (y h) + Θ(y h) (33)
s∈S p∈P l∈Lp
 p
subject to hl = d p  p ∈ P (34)
l∈Lp
ys  0 and integer s ∈ S (35)
p
hl  0 l ∈ L p ∈ P  (36)
where
Fs (y): Total cost of operating service s;
p
Cl (y h): Total cost of moving (part of) product p demand by using its
itinerary l;
Θ(y h): Penalty terms capturing various relations and restrictions, such
as limited service capacity.
The objective function defines the total system cost and includes the total
cost of operating a service network at given frequencies, the total cost of mov-
ing freight by using the selected itineraries for each commodity, as well as a
number of terms translating operational and service restrictions, such as facil-
ity and service capacities and on-time delivery targets, into monetary values.
This third term models, for example, service capacity restrictions as utilization
targets, over-assignment of traffic being permitted at the expense of additional
costs and delays. The introduction of the third term aims to enhance the capa-
bility of the model to identify tradeoffs between the cost of increasing the level
of service (service frequency) and the cost of routing freight on less interesting
itineraries.
The objective function computes a generalized cost, in the sense that it may
include various productivity measures related to terminal and transportation
operations. In addition to the actual costs of performing the operations, one
may thus explicitly consider the costs, delays, and other performance measures
 Ch. 8. Intermodal Transportation 489

related to the quality and reliability of the service offered. Nonlinear conges-
tion functions were thus used to reflect the increasingly larger delays that result
when facilities of limited capacity (e.g., consolidation terminals) must serve
a growing volume of traffic brought by vehicles of different services carrying
freight for different products on various itineraries. The notation aims to con-
vey this utilization of more general functional forms.
The model was adapted and applied to rail, long-haul LTL trucking (railway
transportation of trailers was included as a possible mode), and express letter
services (e.g., Crainic et al., 1984; Roy and Delorme, 1989; Roy and Crainic,
1992). The original solution method described by Crainic and Rousseau (1986)
combines a heuristic that iteratively decreases frequencies from initial high
values, and a convex network optimization procedure to distribute the freight.
The latter makes use of column generation to create itineraries and descent
procedures to optimize the flow distribution.
Most service network design models proposed in the literature use arc-flow-
based formulations and focus on the selection decisions. This results into {0 1}
network design formulations (e.g., the liner service design models presented by
Christiansen et al., 2007). Minimum thresholds are imposed in some cases on
the volume of traffic required to allow the selection of a transportation service
(e.g., the load planning model for LTL motor carriers introduced by Powell
(1986), Powell and Sheffi (1983, 1986, 1989), Braklow et al. (1992)).

3.2.2 Design of postal services

The design of postal services, regular and those dedicated to express let-
ter and package delivery, yields particularly large and complex service network
design formulations. These applications involve air and land long-haul trans-
portation, as well as local pick up and delivery operations. Land long-haul
transportation is often performed by using trucks, although rail is also used
in some countries. It should be noted that the expansion of high-speed railway
networks is making rail increasingly attractive even for express services. The
air mode is both expensive and efficient in carrying relatively large loads over
long distances. Due to the high costs, efforts are dedicated to the optimiza-
tion of available capacity and the reduction of the number of aircrafts used.
This translates into a relatively large number of alternatives in terms of air-
craft types, capacities, and costs which adds to the complexity of the problem.
Kuby and Gray (1993) developed an early model for the design of the net-
work of an express package delivery firm. It is a path-based {0 1} network
design model, where multistop aircraft routes must be selected in and out of
a given hub. Paths were generated a priori and the model was solved with a
standard mixed-integer package. Analyses illustrated the cost effectiveness of
a design with multiple stops over a pure hub-and-spoke network.
Kim et al. (1999) and Armacost et al. (2002) propose a more comprehen-
sive model for the design of the multimodal version of the problem where both
air and ground vehicles are considered (Barnhart and Schneur, 1996, address
a simplified, single-hub version of the problem). In this problem, several hubs
490 T.G. Crainic and K.H. Kim

and aircraft types are considered, while trucks may perform pickup and deliv-
ery activities, as well as transportation over limited distances. The focus is on
the design of the overnight air services for next-day delivery.
Two types of terminals are considered in the model: gateways where pack-
ages enter and exit the air network and hubs where packages are unloaded
from in-coming airplanes, sorted, and loaded into an out-going aircraft for
delivery to their destination gateways. Aircrafts may fly nonstop between gate-
ways and hubs or stop at one gateway en-route. Time window restrictions on
pickup and delivery times at gateways as well as on the sorting periods at hubs
limit the number of stops and constrain the design. Aircrafts of different types
are available in limited numbers. Each aircraft type has an operating cost and
a capacity, as well as operating characteristics (type of engine, flying range,
speed, etc.) that determine the routes it can fly. The objective is to design the
minimum cost set of routes, aircraft assignments to these routes, and package
flows to satisfy demand and level-of-service objectives, while complying with
the operating parameters of the terminals (e.g., capacity), aircraft type (e.g.,
capacity, range, speed), and aircraft fleet (number of planes available).
A commodity p ∈ P represents packages to be moved from an origin gate-
way o(p) to a destination gateway d(p). Let d p indicate the corresponding
demand, H the set of hubs, and {o(p) d(p) p ∈ P } the set of origin and
destination gateways. Let F be the set of aircraft types available and Rf the
set of routes (sequence of gateways starting or ending at a hub) that may be
flown by aircrafts of type f ∈ F . The routes become the flight arcs A linking
the hub and gateway nodes N of the system. The cost of flying route r ∈ Rf
f f
with aircrafts of type f ∈ F is denoted cr , while ur gives its capacity. Costs
associated to package flows are not significant compared to those of operat-
ing the air fleets and are thus not included in the formulation. The number of
available aircrafts of type f is nf and the landing capacity of hub h ∈ H is ah .
rf
Three indicators are also defined: (1) δij = 1 when flight arc (i j) is included
in route r flown by aircraft type f and 0, otherwise; (2) βri = 1 (−1) if node i is
the origin (destination) gateway of route r and 0, otherwise; and (3) δhr = 1 if
hub h is included in route r and 0, otherwise.
f
Two types of decision variables are defined: yr indicates the number of times
p
route r ∈ R is flown by aircrafts of type f ∈ F ; continuous variable xij stands
f

for the amount of product p on the air link (i j) ∈ A. The basic frequency
service network design formulation is then:
  f f
minimize cr yr (37)
f ∈F r∈Rf
subject to
  dp if i = o(p)
p p
xij − xji = −d p if i = d(p) i ∈ N  p ∈ P (38)
j∈N j∈N 0 otherwise
 Ch. 8. Intermodal Transportation 491
 p
  rf f f
xij  δij ur yr  (i j) ∈ A (39)
p∈P f ∈F r∈Rf
 f
yr  n f  f ∈ F (40)
r∈Rf
 f
βri yr = 0 i ∈ N  f ∈ F (41)
r∈Rf
 f
δhr yr  ah  h ∈ H (42)
r∈Rf
f
yr  0 and integer r ∈ Rf  f ∈ F  (43)
p
xij  0 (i j) ∈ A p ∈ P  (44)
Constraints (38) enforce flow conservation for each product, while the forc-
ing constraints (39) restrict the flow on each fly arc to the capacity of the
aircrafts flown on that route. Constraints (40) and (42) enforce the number of
available aircrafts of each type and the landing capacities at hubs, respectively.
Constraints (41) are the aircraft balance restrictions: the number of aircrafts of
each given type landing at a location (delivery routes) must equal the number
taking off from that same location (pick up routes).
In Kim et al. (1999), the authors examine arc, path, and tree-based for-
mulations of this model, and select the latter since it yields a problem of
significantly reduced size compared to the two others. The authors solve the
linear relaxation of the resulting formulation by combining heuristics, which
further reduce the size of the problem, cut-set inequalities to strengthen the
relaxation, and column generation to gradually generate a good set of route
variables. Branch-and-bound is then used to obtain an integer solution. A dif-
ferent methodological approach is presented by Armacost et al. (2002). The
authors transform the problem formulation by defining new composite vari-
ables that combine the original air service frequency and the package flow
variables to represent possible air routes from gateway to hub with minimal but
sufficient capacity to transport the required volume. The composite-variable
formulation includes constraints that force the selection of at least one com-
posite variable for each gateway–hub connection. The authors take advantage
of the fact that air routes pass at most through two gateways and thus the num-
ber of composite variables, which allow to capture multiple aircraft routes with
a single variable, is relatively limited. Moreover, the composite route variables
implicitly account for the flow distribution and thus yield a pure design formu-
lation for which stronger bounds and thus more efficient solution methods may
be derived. The methodology has been implemented at a major US express
package delivery company with significant success (Armacost et al., 2004). The
model is used on a continuous basis for the planning of next-day operations as
well as for what-if scenario analyses. Such results emphasize the need to con-
tinue to explore the network design formulations for new insights and more
efficient solution methods.
492 T.G. Crainic and K.H. Kim

The reorganization of the German postal services belongs to the same prob-
lem class, albeit on a more comprehensive scale. Grünert and Sebastian (2000)
(see also Grünert et al., 1999; Buedenbender et al., 2000) decompose the prob-
lem into several subproblems of manageable proportions: the optimization of
the overnight airmail network, the design of the ground-feeding and deliv-
ery transportation system, and the scheduling of operations. Vehicle routing
models and techniques, as well as a discrete, time-dependent network design
formulation (see Section 3.2.3), are proposed for the routing and scheduling
tasks. The air network design formulation is further decomposed into a di-
rect flight problem and a hub system problem; both subproblems are fixed
cost, multicommodity, capacitated network design formulations with side con-
straints. To optimize these formulations, the authors propose combinations
of classical heuristics, tabu search and evolutionary metaheuristics, and exact
mathematical programming methods (branch-and-bound). A decision support
system integrates the models and associated solution methods, as well as the
tools required to handle the data, models, and methods, and to assist in the
decision process.

3.2.3 Time-dependent service network design

When schedules are contemplated, a time dimension must be explicitly in-
troduced into the formulation. This is usually achieved by representing the
operations of the system over a certain number of time periods by using a
space–time network. The representation of the physical network is replicated
in each period. Starting from its origin in a given period, a service arrives (and
leaves, in the case of intermediary stops) at a later period at other terminals.
Services thus generate temporal service links between different terminals at
different time periods. Temporal links that connect two representations of the
same terminal at two different time periods may represent the time required
by terminal activities or the freight waiting for the next departure. The costs
associated with the arcs of these networks are similar to those used in the sta-
tic formulations of the previous subsections. Additional arcs may be used to
capture penalties for arriving too early or too late.
There are again two types of decision variables. Integer design variables are
associated with each service. Restricted to {0 1} values, these variables indicate
whether or not the service leaves at the specified time. When several depar-
tures may take place in the same time period, general (nonnegative) integer
variables must be used. (Note that one can always use {0 1} variables only by
making the time periods appropriately small.) Continuous variables are used
to represent the distribution of the freight flows through this service network.
The resulting formulations are network design models similar to those pre-
sented in the previous subsections but on significantly larger networks due
to the time dimension. The sheer size of the space–time network, as well
as the additional constraints usually required by the time dimension, makes
this class of problems more difficult to solve than static versions. A lim-
ited number of contributions have thus been reported to date. Most address
 Ch. 8. Intermodal Transportation 493

the problem in the context of a single carrier and propose heuristic solu-
tion methods (e.g., Farvolden and Powell, 1991, 1994; Farvolden et al., 1992;
Equi et al., 1997).
Different formulations are being proposed to address the recent trends in
the development and operation policy of rail intermodal networks. One such
trend is the introduction of booking systems that force customers to book in
advance a precise number of container space slots on a given day and train
service. To operate such systems, intermodal rail carriers enforce regular and
cyclically scheduled services. Moreover, in order to decrease operating costs
(by simplifying operations and reducing equipment handling) and to increase
the utilization of their assets, rail cars and engines, mainly, the characteristics
of services in terms of composition, capacity, and so on, are supposed to stay
the same for all the time the schedule is valid. This starts to be known as a
“full-asset-utilization” operating policy.
Not much work has yet been done with respect to these new operation
characteristics, which may be observed both in North America and Europe.
Pedersen et al. (2006) present one of the first contributions. The authors notice
that a “full-asset-utilization” operating policy requires that the asset circula-
tion issue be integrated into the service network design model. They represent
this requirement by enforcing the condition that at each node of the repre-
sentation, the (integer) design flow be balanced. The authors propose two
formulations (a link and a cycle-based one) and a tabu search-based meta-
heuristic that appears to efficiently yield good quality solutions. Much more
work is required in this area, however.

3.3 Port dimensioning

Important strategic planning issues for the design of container port termi-
nals are related to the number of berths, the size of storage space, and the
number of pieces of various equipment to install. Port dimensioning issues re-
quire a trade-off between the amount of investment and the level of customer
service. For example, as the number of berths increases, the turnaround (wait-
ing) time of vessels decreases. Because of the high investment and operating
cost of container ships, delays experienced at a port generate high costs to the
ship’s operators and cause delayed arrivals at successive ports creating serious
downstream operational problems. Thus, availability of empty berths at a port
is a key issue when operators determine container terminals for their ships.
Studies aimed at determining the optimal number of berths usually consider
the costs related to the turnaround time of ships, the berth construction cost,
and the berth operation cost.
Quay cranes are the most expensive (5–10 million dollars per quay crane)
handling equipment in port container terminals. Determining the number of
quay cranes to be installed is thus another major container–port design issue.
In addressing this issue, one must consider that as the number of quay cranes
per berth increases, the throughput rate per berth increases as well, but at a
494 T.G. Crainic and K.H. Kim

declining rate due to interference between adjacent quay cranes. Determining

the number of pieces of other types of equipment, such as transporters, yard
cranes, and so on, are also important port dimensioning issues.
Storage space is another critical resource in port container terminals, espe-
cially in major Asian hub ports such as Singapore, Hong Kong, Busan, Kobe,
and Kaohsiung. As the storage space becomes smaller, the storage stacks be-
come higher, which results in lower productivity of the transfer operations
in the yard. Thus, there is a trade-off relationship between the investment in
storage space and the productivity of the transfer operations in the yard (see
Section 5).
Very few optimization-based models have been proposed for these issues,
however. Descriptive models, mainly based on queuing theory and probabilis-
tic approaches, have been used in most cases. The goal was to evaluate the
performance of particular components of container terminals under various
hypotheses regarding the number, type and, eventually, combination of re-
sources, e.g., Fratar et al. (1960), Miller (1971), Wanhill (1974), and Daganzo
(1990) for the number of berths; Griffiths (1976) and Daganzo (1989a) for the
number of quay cranes; Schonfeld and Sharafeldien (1985) for the number of
quay cranes and berths; van Hee and Wijbrands (1988), Kozan (1997), and
Kim and Kim (2002) for combinations of different types of resources including
yard cranes, storage slots, and yard trailers. Two exceptions to this trend are
nonlinear programming model proposed by Noritake and Kimura (1990) for
berth dimensioning and the maximum flow network model introduced by Vis
et al. (2001) to determine the minimum number of automated guided vehicles
for completing a given set of delivery tasks.
Simulation models have also been proposed to evaluate seaport container
terminal design (e.g., number and layout of various resources), operation poli-
cies (e.g., working hours), operation characteristics (e.g., travel and handling
times of various equipments) and general performance: Ramani (1996), Lai
and Leung (2000), Nam et al. (2002), etc. A different simulation approach
was proposed by Alessandri et al. (2006), which represent containers and their
movements in a terminal as a network of queues. Discrete-time equations are
then used to describe the dynamic evolution of the system, where control vari-
ables represent the utilization of the terminal resources. This model yields an
optimization problem that aims to minimize the transfer delays in the termi-
nal. The problem is stated as an optimal control problem and is addressed by
a receding-horizon solution strategy.

4 Container fleet management

The need for freight carriers to move empty vehicles follows from the differ-
ences in demand and supply for each commodity observed at most locations,
resulting in an accumulation of empty vehicles in regions where they are not
 Ch. 8. Intermodal Transportation 495

needed and in deficits of vehicles in other regions that require them. Vehi-
cles must then be moved empty, or additional loads must be found, in order
to bring them where they are needed to satisfy known and forecast demand
in the following planning periods. This operation is known as repositioning or
empty balancing and is a major component of what is known as fleet manage-
ment. In its most general form, fleet management covers the whole range of
planning and management issues from procurement of power units and vehi-
cles to vehicle dispatch and scheduling of crews and maintenance operations.
Often, however, the term designates a somewhat restricted set of activities: al-
location of vehicles to customer requests and repositioning of empty vehicles.
We follow this definition is this section.
Fleet management belongs to what is usually called the operational level of
planning. Most system elements, demand, travel and handling times, and so
on, vary with time. Most strategic and tactical planning models do not explic-
itly account for these variations. At the operational planning level, however,
the time-dependency of data, decisions, and operations must be explicitly ad-
dressed, including the representation of the outcome of current decisions on
the future state of the system. The time-dependent (dynamic) aspect of opera-
tions is further compounded by the stochasticity inherent to most systems, that
is, by the set of uncertainties that are characteristic of real-life operations and
management activities. If traffic is slower than predicted, vehicles may arrive
late at customers’ locations or at the terminal. Forecast customer requests for
empty containers may not materialize while unexpected demands may have
to be satisfied. The planned supplies of empty vehicles at depots may thus
be unsettled and additional empty movements may have to be performed. In-
creasingly, these characteristics are reflected in the models and methods aimed
at operational planning and management issues, as illustrated by the fleet man-
agement models of this section.
Moving vehicles empty does not directly contribute to the profit of the firm
but it is essential to its continuing operations. Consequently, one attempts to
minimize empty movements within the limits imposed by the demand and ser-
vice requirements. Early models were based on transportation formulations or,
most often, deterministic time-dependent transshipment network models (e.g.,
White, 1972, and Ermol’ev et al., 1976, for container fleet management). De-
tails on the early approaches may be found in Dejax and Crainic (1987). Mod-
els that explicitly consider uncertainties in empty vehicle distribution started
to be proposed in the mid-1980s (e.g., Jordan and Turnquist, 1983). There is
now a significant body of work in this area, mostly dedicated to rail and motor
carrier issues (see the reviews by Crainic (2003), Cordeau et al. (1998), Powell
(2003), Powell and Topaloglu (2003, 2005), Powell et al. (2007)).
Few efforts were dedicated to container fleet management issues. Crainic
et al. (1993b) proposed a series of models for the allocation and management
of a heterogeneous fleet of containers where loaded movements are exoge-
nously accepted. Cheung and Chen (1998) focused on the single-commodity
container allocation problem for liner regular ocean navigation lines operators.
496 T.G. Crainic and K.H. Kim

Powell and Carvalho (1998) addressed the problem of the combined optimiza-
tion of containers and flatcars for rail intermodal operations.
The container fleet management problem addressed by Crainic et al.
(1993b) was cast in the context of the land transport part, as described in Sec-
tion 3.1.1. The land container distribution system is composed of a number
of port terminals, in-land depots, and customer locations. Ships arrive in ports
carrying loaded and empty containers of various types and dimensions. Loaded
containers are delivered to their destinations, using rail and truck under var-
ious types of contracts, while empty containers are available for delivery to
customers in the vicinity of the port or for repositioning. Customers receiv-
ing loaded containers unload them and signal that they may be picked up and
transported to a designated terminal (port or in-land). Similarly, customers
that require empty containers of specific types for future shipments receive
them from an in-land or a port terminal. If the number of containers of a given
type requested by the customer is not available for delivery within the time
window specified by the customer, the company can lease (or borrow from
partner companies) containers or substitute certain other types of contain-
ers. The empty containers available for distribution to customers thus come
from four sources: customers that unloaded their goods, world-wide reposi-
tioning, leasing, and substitution. Containers are repositioned through two
mechanisms: (1) the balancing movements determined by the strategic/tactical
planning of the operations (Section 3.1.1), and (2) the allocation decisions
concerning the depots to which the containers that become empty at customer
locations are to be sent. The problem is dynamic, in the sense that demands
vary in time and transport and customer operations may both require more
than one period (day). It is stochastic due to variations in customer demands,
including the possibility of demands from “unknown” customers, variations in
the time required by customers to unload and return containers, and damage
to containers (travel time variations were not considered in the chapter).
The formulation is defined for a planning horizon discretized in T periods:
t = 1 2     T . Let H be the set of ports and D the set of inland depots. The
sources of randomness considered in the formulation are:
p
di : Demand of containers of type p for demand customer i ∈ I p ; The
containers must be delivered within a time window i ;
pt
dh : Demand of empty containers of type p at port h ∈ H in period t for
export;
pt
os : Supply of empty containers of type p released by supply customer s
at time t; The set of these customers is denoted S pt ;
pt
oh : Supply of empty containers of type p arriving at port h in period t.
The parameters of the problem are:
pt
Ji : Set of depots j that may send a shipment of containers of type p
that would arrive in period t at customer i;
 Ch. 8. Intermodal Transportation 497

pt
Ij : Set of customer requests that may be served (i.e., he containers of
type p will arrive within i ) by a shipment starting from depot j ∈
D ∪ H in period t;
pt
cji : Unit transportation cost for a container of type p from a depot j to
pt
demand customer i ∈ Ij ;
pt
Sj : Set of supply customers from where containers of type p may reach
terminal j in period t;
pt
csj : Unit transportation cost for a container of type p from supply cus-
pt
tomer s ∈ Sj to a depot j;
pt
cjk : Unit transportation cost for a container of type p between depots
j k ∈ D ∪ H;
cpr : 1/(Number of containers of type p needed to replace one container
of type r);
prt
cj : Unit cost at depot j in period t to substitute a container of type p
for a container of type r;
pt
cj : Unit holding cost for a container of type p at depot j in period t;
pt
c̄j : Cost of leasing (or borrowing) a container of type p at depot j in pe-
pt
riod t; It also represents the penalty cost at port h when dh cannot
be satisfied.

With decision variables:

pt
xji : Number of equivalent containers of type p allocated in period t
from depot j to customer i;
pt
xj : Number of containers of type p allocated as type p in period t from
depot j;
prt
xj : Number of containers of type p substituted for containers of type r
in period t at depot j;
pt
xsj : Number of containers of type p picked up at customer s in period t
and delivered to depot j in period t̄  t;
pt
wj : Number of containers of type p in stock at depot j at the end of
period t;
pt
wjk : Number of containers of type p repositioned from depot j to depot
k in period t;
pt
bj : Number of containers of type p rented at depot j in period t;
498 T.G. Crainic and K.H. Kim

the model minimizes the expected total operating cost, including substitutions
and stockouts, over a multiperiod planning horizon
     pt pt  pt pt  pt pt
E csj xsj + cji xji + cjk wjk
t p∈P j∈H∪D s∈S pt pt
i∈Ij k∈H∪D
j
 
prt prt pt pt pt pt
+ cj xj + c j wj + c̄j bj (45)
r∈P

subject to nonnegativity constraints on all variables, bounds on balancing flows,

and constraints representing the dynamics of the system:
  pt p
xji − di = 0 i ∈ I p  (46)
t∈i j∈J pt
i
  pt̄ pt
xsj − os = 0 s ∈ S pt  p ∈ P  t = 1 2     T (47)
t̄t j∈Sj pt¯
pt
 prt
 pt
xj + cpr xj − xji = 0
r∈P i∈Ij
pt

j ∈ D ∪ H p ∈ P  t = 1 2     T (48)
pt pt prt
wj + xj + xj
 pt pt−1
+ wjk − wj
k∈H∪D
 pt   pt̄ pt
− xsj − wkj − bj = 0 (49)
s∈S
pt t̄t k∈H∪D
j

j ∈ D  p ∈ P  t = 1 2     T
pt pt prt
wh + xh + xh
 pt pt pt
+ whk + dh − whk
k∈H∪D
 pt   pt̄ pt pt
− xsh − whk − oh − bh = 0
pt
s∈Sh t̄t k∈H∪D

h ∈ H p ∈ P  t = 1 2     T (50)
where constraints (46) and (47) enforce demand and supply conditions at cus-
tomers, Equations (49) and (50) are the flow conservation conditions at in-land
depots and ports, respectively, while Equations (48) represent the allocation
accounting between true and substitution containers at depots. The authors
presented single and multicommodity deterministic formulations and a two-
stage, restricted-recourse single commodity, stochastic model.
 Ch. 8. Intermodal Transportation 499

The stochastic and dynamic empty container allocation model of Cheung

and Chen (1998) addresses the problem from the point of view of a container
liner company that offers regular service lines between a given number of ports.
Ships transport loaded containers that bring a profit and, space permitting,
empty containers to reposition them in order to be able to satisfy future fore-
cast demand. It is assumed that only one ship travels between two ports at
each time period and that demand for loaded containers between ports does
not exceed the capacity of the ship. Only one product – empty containers – is
considered. The schedule of ships linking the ports in H is assumed fixed for
the planning horizon t = 1 2     T . Similarly to the formulations of Crainic
et al. (1993b), one supposes that the model is to be run into a rolling horizon
mode: the model gives a solution (that accounts for the impact of today’s de-
cisions on the future state of the system), the suggestion solution for the first
period is implemented and then, in the next period, the model is run again
using the up-dated information.
Three sources of randomness are considered in the model:
κtij : residual capacity for empty containers on the ship traveling from
port i ∈ H to port j ∈ H, leaving at time t;
δti : Demand for containers at port i at period t;
σit : Supply of containers at port i at period t (before unloading from
ships).
The parameters of the problem are:
cijt : Unit transportation cost from port i ∈ H to port j ∈ H, leaving at
time t;
lit : Unit cost for loading a container on a ship at port i at period t;
uti : Unit cost for unloading a container from on a ship at port i at pe-
riod t;
cit : Unit holding cost for a container at port i in period t;
c̄it : Unit cost of leasing (or borrowing) at port i in period t;
Rti : Unit revenue from satisfying demand at port i at period t;
τij : Transportation time from port i to port j;
and the decision variables:
lt
xi : Number of containers loaded at port i in period t for any destina-
tion;
ut
xi : Number of containers unloaded at port i in period t from any other
port;
ct
xij : Number of containers to be repositioned from port i to port j de-
parting in period t;
pt
xij : Number of containers currently on the ship, in port i at period t,
with destination port j;
ht
xi : Number of containers stored at port i at period t;
500 T.G. Crainic and K.H. Kim

rt
xi : Number of leased containers at port i and period t used to meet
demand;
dt
xi : Number of owned containers at port i and period t used to meet
demand.
All decision variables are nonnegative. The objective is to minimize the ex-
pected total cost while maximizing the revenue from satisfying demand:
  lt ut ht rt
minimize E lit xi + uti xi + cit xi + c̄it xi
t i∈H
 
ct
+ cijt xij − Rti δti (51)
j∈H

while the following relations define the system dynamics and constraints of the
model:
 ct−τij ut
 pt
xki − xi − xij = 0 i ∈ H t = 1 2     T (52)
k∈H j∈H
ht ht−1 ht
xi + δti − xi − σit − xi = 0 i ∈ H t = 1 2     T (53)
 ct lt
 pt
xij − xi − xij = 0 i ∈ H t = 1 2     T (54)
j∈H j∈H
ct
xij  κij  i ∈ H t = 1 2     T
t
(55)
rt dt
xi = δti − xi  i ∈ H t = 1 2     T (56)
Relations (55) specify that demand must be satisfied. At each port and pe-
riod, Equations (52) compute the volume of containers unloaded from ships
(the difference between the total volumes arriving and departing on all ships),
relations (53) represent the inventory equations that also enforce the require-
ment that demand must be satisfied, and Equations (54) compute the number
of containers being repositioned. Constraints (55) enforce the residual capacity
for repositioning movements and Equations (56) define the number of leased
containers as the difference between the number of owed containers avail-
able and demand. The formulation is represented by a network problem with
random arc capacities and is cast as a two-stage stochastic program. Stochas-
tic quasi-gradient and hybrid approximation solution procedures are proposed
and are compared through a rolling-horizon experiment.
Powell and Carvalho (1998) address fleet management problems within the
context of rail intermodal systems. The goal is to consider simultaneously
the management of the fleet of flat cars the railway uses to provide service, the
fleet of trailers and containers (collectively know as boxes) the railway owns
and rents to its customers, and the complex rules that govern the substitu-
tions among box types and the assignment of boxes to flat cars. The problem
is highly dynamic and characterized by two sources of stochasticity: forecast of
 Ch. 8. Intermodal Transportation 501

customer demand and the return of boxes and flat cars from customers and
other railways.
The authors decompose the problem into a problem addressing the manage-
ment of the boxes the railway owns and a flat car fleet management problem.
Each is formulated as a mixed-integer program over a multicommodity space–
time diagram representing “all” possible movements of vehicles and loads
(including holding activities) over the planning horizon. Decision variables
concern departure times, the type of vehicle (and departure time) for each
load, repositioning movements, flows of vehicles and loads. The objective max-
imizes the expected total profit of the system. Each formulation is then cast as a
recursive dynamic model and, using approximations of the future values of ve-
hicles and boxes at the nodes of the space–time network, it is decomposed into
“easy-to-solve” local subproblems (assignment of boxes to requests or combi-
nations of boxes to flat cars). In a series of forward–backward passes through
the network, the algorithm then refines these approximations and assignments.
Experimentations based on actual data suggest significant improvements over
the planning procedures used in industry. More details on this methodology
may be found in Powell (2003), Powell and Topaloglu (2003, 2005), and Powell
et al. (2007).

5 Models for seaport container terminal operations

This section describes issues and introduces corresponding models for oper-
ational planning and control in port container terminals. The following subsec-
tions introduce models for berth scheduling, quay-crane scheduling, stowage
planning and sequencing, storage space planning, and dispatching of yard
cranes and transporters: yard trucks, straddle carriers and, for automated ter-
minals, automated guided vehicles (AGV). Reviews can be found in Steenken
et al. (2004) and in the book edited by Günther and Kim (2005).

5.1 Berth scheduling

As already mentioned, berths are the most important resource in port con-
tainer terminals because berth construction costs are the highest among all
relevant cost factors. Berth scheduling is the process of determining the time
and position at which each arriving vessel will berth. Quay-crane allocation is
the process of determining the vessel that each quay crane will serve and the
time during which the quay crane will serve the assigned vessel. (The terms
scheduling and allocation are often used interchangeably for both problems.)
The berth scheduling and the quay-crane allocation problems are related be-
cause the number of quay cranes assigned to a vessel impacts the berthing
duration of the ship. Despite this relationship, most studies treat the two is-
sues separately to avoid the complexity of the integrated problem. The study
by Park and Kim (2003) is an exception.
502 T.G. Crainic and K.H. Kim

Ports have long used priority rules to determine the allocation of berths
to incoming ships and the earliest studies focused on this approach (e.g., van
der Heyden and Ottjes, 1985). Simulation was used to evaluate and compare
rules (e.g., Lai and Shih, 1992). Brown et al. (1994, 1997) proposed the first
mathematical models for allocating berths to vessels. Their studies focused on
military vessels and assumed that each vessel required different services (re-
provisioning, maintenance, repair, training, and certification test). Since not
all services were provided at all berths, scheduling vessel shifts between berths
was an important issue. This is different from the case of container ships for
which only loading and unloading operations need to be considered.
Many ports are configured such that berths may be considered as subsec-
tions of a continuous line that ships of finite lengths can share, and several stud-
ies (Lim, 1998; Park and Kim, 2002; Park and Kim, 2003; Kim and Moon, 2003;
Guan and Cheung, 2004) consider the berth allocation problem as a type of
continuous line partitioning problem. Most researchers have treated berth
scheduling as a discrete resource allocation problem, however. This approxi-
mation reduces the problem to that of assigning berths to arriving ships, which
is much easier to address than the continuous berth-scheduling problem. We
start the presentation with this second approach.
Cordeau et al. (2005) proposed an integer programming model for the dis-
crete version of the berth-allocation problem based on a Multidepot Vehicle
Routing Problem with Time Windows (MDVRPTW) formulation, where ships
are seen as customers and berths as depots. One vehicle is located at each de-
pot. Each vehicle starts and ends its tour at its depot. Ships are modeled as
vertices in a multigraph. Every depot is divided into an origin vertex and a des-
tination vertex. Time windows can be imposed on every vertex to represent the
availability period of the corresponding berth. Notation is as follows:
N: Set of ships; n = |N|;
M: Set of berths; m = |M|;
tik : Handling time of ship i at berth k;
ai : Arrival time of ship i;
sk : Start of availability time of berth k;
ek : End of availability time of berth k;
bi : Upper bound of the service time window of ship i;
vi : Value of the service time for ship i;
o(k): Starting operation time of berth k;
d(k): Ending operation time of berth k;
Mijk = max{bi + tik − aj  0}, k ∈ M, i and j ∈ N;
Gk = (V k  Ak ), k ∈ M, where V k = N ∪ {o(k) d(k)} and Ak ⊆
V k × V k.
With decision variables:
xkij : xkij = 1, if and only if ship j is scheduled after ship i at berth k;
0, otherwise;
 Ch. 8. Intermodal Transportation 503

Tik : Berthing time of ship i at berth k, i.e., the time when the ship moors;
k : Starting operation time at berth k, i.e., the time when the first ship
To(k)
moors at the berth;
k : Ending operation time at berth k, i.e., the time when the last ship
Td(k)
departs from the berth;
the MDVRPTW model may be written as:
 
minimize vi Tik − ai + tik xkij (57)
i∈N k∈M j∈N∪{d(k)}
subject to
 
xkij = 1 i ∈ N (58)
k∈M j∈N∪{d(k)}

xko(k)j = 1 k ∈ M (59)
j∈N∪{d(k)}

xkid(k) = 1 k ∈ M (60)
j∈N∪{o(k)}
 
xkij − xkji = 0 k ∈ M i ∈ N (61)
j∈N∪{d(k)} j∈N∪{o(k)}

Tik + tik − Tjk  (1 − xkij )Mijk  k ∈ M (i j) ∈ Ak  (62)

ai  Tik  k ∈ M i ∈ N (63)

Tik + tik xkij  bi  k ∈ M i ∈ N (64)
j∈N∪{d(k)}

sk  To(k)
k
 k ∈ M (65)
k
Td(k)  ek  k ∈ M (66)
xkij ∈ {0 1} k ∈ M i j ∈ Ak  (67)
The objective is to minimize the weighted sum of the service times for all
the vessels. Constraint (58) implies that each vessel must be assigned once to
a berth. Constraints (59) and (60) define the degree of the depots. Constraint
(61) enforces the flow conservation. The consistency of the Tik variables with
the time sequence on a berth is guaranteed by constraints (62). Constraints
(63) and (64) indicate the service time windows for vessels, while constraints
(65) and (66) set the available time windows on the berths. For small instances,
the above formulation can be solved by commercial software for integer linear
programming models. For problem instances of realistic size, Cordeau et al.
(2005) proposed a tabu search-based metaheuristic.
A different approach consists in explicitly representing and penalizing the
difference between the berthing order implied by the ship priority and the
504 T.G. Crainic and K.H. Kim

one proposed. The resulting models take the form of nonlinear integer pro-
gramming formulations (Imai et al., 1997, 2001, 2003). Lagrangian-based and,
for problem instances of realistic size, genetic metaheuristic solution methods
have been proposed. The discrete berth-scheduling problem may also be cast
as a machine-scheduling problem. Li et al. (1998) introduced a formulation
for the scheduling of a single processor (the berth) that can simultaneously
perform multiple jobs (vessels). The authors aimed to minimize the makespan
and, based on the similarity of the problem to the bin-packing problem, sug-
gested various algorithms based on the First-Fit-Decreasing heuristic. Guan
et al. (2002) proposed an m-parallel machine scheduling formulation. In their
model, the machines (quay cranes) are arranged along a straight line and each
job (ship) requires simultaneous processing by multiple consecutive proces-
sors. Ships are characterized by size, processing time, and weight (priority).
A heuristic was proposed to minimize the total weighted completion time of
the jobs (ships).
A drawback of considering a berth as a collection of discrete berthing lo-
cations is that the number of ships that may be served simultaneously is fixed
regardless of ship lengths. The continuous representation does not have this
limitation: for the same length of berth, more vessels can be served simul-
taneously if they are shorter. On the other hand, in contrast to the discrete
variant, the continuous berth-scheduling problem requires determining the ex-
act berthing position of each ship as a real-valued position on a continuous
line. Moreover, the berthing time for each ship must be determined simulta-
neously. The goals of the process include minimizing the ship departure delays
and the container-handling costs that depend on the berthing position of each
ship.
Lim (1998) was the first to consider a berth as a continuous line rather than
a collection of discrete segments and viewed the berth planning problem as
a two-dimensional bin packing problem. He discussed how to locate berthing
positions of vessels so that the throughput of the berth is maximized, but did
not consider the berthing time as a decision variable. Park and Kim (2002)
introduced the first model for the continuous berth-scheduling problem to
determine simultaneously the berthing time and position for each ship. The
objective of the model is to minimize the costs resulting from delayed vessel
departures, plus the additional handling costs resulting from deviations of the
berthing position from the best location on the berth. The following informa-
tion is assumed known:
L: Length of the berth;
l: Number of vessels;
ai : Expected arrival time of vessel i;
bi : The ship operation time required for vessel i;
di : Requested departure time of vessel i;
li : Length of vessel i; this value includes the required gap between ad-
jacent vessels;
 Ch. 8. Intermodal Transportation 505

c1i : Additional travel cost (per unit berth length) of delivering contain-
ers to vessel i resulting from nonoptimal berthing locations;
c2i : Penalty cost (per time unit) of vessel i resulting from a delayed de-
parture after the requested due time.
The problem is set in a two-dimensional space, the berth length and the
planning time defining the horizontal and vertical axes, respectively. The ref-
erence point for the berth-length coordinate is the leftmost boundary of the
berth. Figure 3 illustrates this setting, as well as the relationships between the
input data and the following decision variables:
xi : Berthing position of vessel i;
yi : Berthing time of vessel i;
pi : Best berthing location of vessel i; This location is represented by the
x-coordinate (berth length axis) of the leftmost end of the vessel. It
is determined by considering the distribution of containers to be
loaded into the vessel;
zijx : Equals 1, if vessel i is located to the left of vessel j in the time-berth-
length space; 0, otherwise; Vessel i is located to the left of vessel j
in Figure 3;
y
zij Equals 1, if vessel i is located below vessel j in the time-berth-length
space; 0, otherwise; Vessel i is not considered to be located below
vessel j in Figure 3 because of their partial overlap in the time di-
mension.

Fig. 3. Structure of continuous berth scheduling problem (Park and Kim, 2002).
506 T.G. Crainic and K.H. Kim

An objective function can then be written as:


l
 
minimize c1i |xi − pi | + c2i (yi + bi − di )+  (68)
i=1

where x+ = max{0 x}. The first term of (68) computes the penalty for the
deviation of the berthing positions from the best locations, while the second
computes the penalty corresponding to the departure delays of the vessels leav-
ing after the requested departure time. Let |xi − pi | be α+ i when xi − pi  0
and α−i when xi −p i < 0. Let (y i +bi −d i ) be β+
i when y i +b −
i −di  0 and βi ,
otherwise. Then, the berth scheduling problem can be formulated as follows:


l
 
minimize c1i (α+ − +
i + αi ) + c2i βi (69)
i=1
subject to
xi − pi = α+ −
i − αi  i = 1 2     l (70)
yi + bi − di = β+ −
i − βi  i = 1 2     l (71)
xi + li  L i = 1 2     l (72)

xi + li  xj + M 1 − zijx  j = 1 2     l i = j (73)
y
yi + bi  yj + M 1 − zij  i j = 1 2     l i = j (74)
y y
zijx + zji
x
+ zij + zji  1 i j = 1 2     l i < j (75)
yi  ai  i = 1 2     l (76)
α+ − + −
i  αi  βi  βi  xi  0 i = 1 2     l (77)
y
zijx  zij ∈ {0 1} i j = 1 2     l i = j (78)

Constraints (70) and (71) define α+ − + −

i , αi , βi , and βi . Constraints (72) spec-
ify that the position of the rightmost end of vessel i is restricted by the length of
the berth. A constraint of the type (73) or (74) is effective only when the corre-
y
sponding zijx or zij , respectively, equals one. The constraints define the relative
“left” and “bellow” position of vessel i relative to that of vessel j in the time-
berth-length space. Constraints (75) enforce the condition that two vessels
cannot occupy the same space during the same time. The constraints proceed
by excluding in the time-berth-length diagram the case zijx + zji x + z y + z y = 0,
ij ji
which corresponds to schedule overlapping for vessels i and j. Constraints (76)
imply that vessels cannot berth before they arrive.
Model (69)–(78) is a mixed-integer programming formulation. Park and
Kim proposed a subgradient optimization technique and solved problems with
13–20 ships in a few minutes of computational time on a Pentium II (233 MHz
and 128 Mb RAM) computer. Similar formulations have also been proposed
 Ch. 8. Intermodal Transportation 507

by Kim and Moon (2003) and Guan and Cheung (2004). Kim and Moon exam-
ined a stability property that berthing locations of vessels must satisfy to form
an optimal solution and used it in a simulated annealing procedure. Guan and
Cheung used a different objective function. They minimized the total weighted
flow time instead of the penalties for delayed service and the deviation of
berthing positions from the best locations of Park and Kim (2002). Guan and
Cheung proposed a tabu search metaheuristic based on pair-wise exchanges,
which appears to perform well.

5.2 Quay-crane scheduling

Quay cranes are another important resource in container terminals. In prac-

tice, the quay-crane schedule is usually constructed by operation planners as
part of ship plans. Two types of quay-crane scheduling problems have been
defined according to the detail of the representation. The first problem simul-
taneously schedules the berth and the quay cranes by specifying the starting
and ending times of unloading and loading operations for each quay crane
assigned to a specific ship. Park and Kim (2003) proposed a model and a
two-phase solution procedure. The first phase determines the berthing posi-
tion and time of each vessel as well as the number of quay cranes assigned to
each vessel at each time segment. Subgradient optimization is used to obtain a
near-optimal solution of the first phase. In the second phase, a detailed sched-
ule for each quay crane is constructed by dynamic programming starting from
the solution of the first phase.
The second type of quay-crane scheduling problem determines the detailed
schedule for each quay crane assigned to a vessel for a pre-specified time
window to carry out container unloading or loading operations. Most studies
on quay-crane scheduling address this problem. The modeling methodology
generally used is based on integer programming models. Daganzo (1989b) ad-
dressed the quay-crane scheduling problem in port terminals for general and
containerized cargo. He attempted to determine the number of quay cranes to
assign to each ship bay at each time segment. An integer programming model
was suggested with an objective to minimize the completion time of ship oper-
ations for all the ships.
Peterkofsky and Daganzo (1990) refined the previous model and proposed
a branch-and-bound algorithm. In their model, S ships are moored alongside
the berth of a terminal. Ship i has Hi holds numbered (i 1) (i 2)     (i Hi ).
The time required to complete the loading and unloading operations for hold
(i j) is denoted Wij measured in quay-crane-hours. It is assumed that all ships
are ready for loading/unloading operations at time t = 0. It is also assumed
that sufficient berth space is provided for all the ships to berth at the same
time. M quay cranes are available. Decision variables are yijm (t) = 1 if quay
crane m, m = 1 2     M, is allocated to hold (i j) at time t, and 0 other-
wise. The objective function minimizes the weighted amount of time that ships
spend in port, Si=1 Ci Ti , where Ti and Ci stand for the departure time and
508 T.G. Crainic and K.H. Kim

cost per hour of delay for ship i, respectively. A ship can depart only when the
operations on all its holds have been completed, that is, if
M 
 Ti
yijm (t) dt  Wij  j = 1 2     Hi  (79)
m=1 0
In some cases, there may exist other activities than ship unloading and load-
ing which must be performed (e.g., refill of supplies or repairs) and thus the
earliest possible departure time may be restricted by a limit τi yielding the
constraints
Ti  τi for i = 1 2     S (80)
A quay crane may be allocated to no more than one hold at a time. Thus, for
any value of t, the crane assignments must satisfy:

S 
Hi
yijm (t)  1 m = 1 2     M (81)
i=1 j=1

Finally, when only a maximum of ij quay cranes can be allocated to hold (i j),
the following constraints must be added to the formulation:

M
yijm (t)  ij  i = 1 2     S j = 1 2     Hi  (82)
m=1
Peterkofsky and Daganzo (1990) assumed a situation with no serious in-
terference between cranes and, consequently, did not impose any restrictions
on the movements of cranes. This approach is appropriate for general cargo
handling. However, quay cranes in container terminals travel on the same
rail, which results in various interference possibilities between adjacent cranes.
Consequently, adjacent quay cranes need space of at least two ship bays be-
tween them. At the same time, no crane can pass over adjacent cranes, which
makes the problem more complicated. Kim and Park (2004) addressed the
quay-crane scheduling problem at this level of detail, but for a single vessel
only. The model determines starting and ending times for each quay crane
to serve each ship bay, under various constraints representing detailed move-
ments of quay cranes and interferences among quay cranes. The authors pro-
posed both a branch-and-bound algorithm and a GRASP-based heuristic to
overcome the computational difficulty of solving the mixed-integer program-
ming formulation exactly.
Comparing the work of Peterkofsky and Daganzo (1990) and Kim and Park
(2004), one realizes that the difference comes from the different viewpoints on
how the whole ship planning problem should be addressed. Peterkofsky and
Daganzo (1990) considered the quay-crane scheduling as a part of berth plan-
ning, whereas Kim and Park (2004) separated the detailed crane-scheduling
problem from the berth-scheduling problem. This comparison also provides
 Ch. 8. Intermodal Transportation 509

an illustration of the often-encountered trade off between the integration of

several planning issues into the same model and the level of detail one can
include in the representation of each element of the system being considered.

5.3 Stowage planning and sequencing

A container group is defined as a collection of containers of the same size

and with the same destination port. Stowage planning determines the block
(cluster of adjacent slots) of a ship bay a specific group of containers should be
stacked into. The stowage planning process must consider the burden of ad-
ditional container manipulations when containers bound for succeeding ports
are located in higher tiers than those that must be unloaded at a given port.
The stowage plan must also comply with various measures of ship stability and
strength. Christiansen et al. (2007) review studies on stowage planning.
The stowage-sequencing problem must be addressed next by determining the
sequence of unloading inbound containers and loading outbound containers.
The slot into which each of the outbound containers will be stacked must be de-
termined simultaneously. When an indirect transfer system is used, yard cranes
move containers between yard stacks and yard trucks. The loading sequence of
individual containers thus impacts significantly the total distance traveled by
yard cranes and hence the total handling cost in the yard. This is not the case
when a direct transfer system is used since the yard handling cost is deter-
mined by the cost of yard truck movements only, which is independent of the
container loading sequence.
In the stowage (load/unload) sequencing problem, it is usually assumed that
the stowage plan is already constructed and provided by the vessel carrier. The
problem is then reduced to assigning slots to individual loading containers and
sequencing unloading and loading operations. Most research has focused on
the sequence of loading operations, which influences the handling costs more
significantly than the sequence of unloading operations. Indeed, containers to
be loaded into the slots of a ship must satisfy various constraints on the slots,
which are pre-specified by the stowage planner. Furthermore, locations of out-
bound containers may be scattered over a wide area in a terminal yard. The
time required for loading operations, thus, depends on the cycle time of both
quay and yard cranes. The cycle time of a quay crane depends on the loading
sequence of slots in the vessel, while the cycle time of a yard crane is affected
by the loading sequence of containers in the yard.
Research on load sequencing can be classified according to the scope of
the problem. Several efforts have addressed the pickup-scheduling problem in
which the travel route of each yard crane and the number of containers to
be picked up at each yard bay on the route are determined (Kim and Kim,
1999b, 1999c; Kim and Kim, 1999d, 2003; Narasimhan and Palekar, 2002; Ryu
et al., 2001). Other authors have attempted to determine the loading sequence
of individual containers present in a yard into the slots of a vessel, a process
that requires more detailed scheduling than the previous problem (Cojeen and
510 T.G. Crainic and K.H. Kim

Table 2.
Yard map

Yard-bay 1 2 4 5 7 8 9 11 12 14 15
Container group A B C B A B C B A C A
Number of containers 14 14 13 14 12 7 12 8 13 10 13

Table 3.
Load plan

Subsequence number 1 2 3 4 5 6 7 8
Container group A B A C B A B C
Number of containers 15 20 24 25 11 23 12 10

Dyke, 1976; Kozan and Preston, 1999; Kim et al., 2004). In this subsection, we
focus on the pickup-scheduling problem, with only a brief literature review
dedicated to the individual-container-scheduling problem.
The pickup-scheduling problem for a piece of yard equipment (e.g., a yard
crane) can be described as follows (Kim and Kim, 1999b; Narasimhan and
Palekar, 2002): In a terminal yard, containers are classified into groups, each
yard bay holding a number of containers of a number of particular groups.
Table 2 illustrates a yard map, that is, the distribution of containers by con-
tainer group over the yard bays. It is a simplified case where each yard bay
contains containers of one group only: 14 containers of group A are in yard-
bay 1, 14 containers of group B are in yard-bay 2, and so on. A load plan
specifies a sequence of blocks of pickup operations, each block representing a
subsequence of consecutive pickups of containers belonging to the same group.
Table 3 shows a load plan composed of eight subsequences to pick up sequen-
tially 15 containers of group A, 20 containers of group B, 24 of group A, and so
on. A partial tour denotes a sequence of yard bays the yard crane visits to per-
form the operations for a given subsequence. The pickup-scheduling problem
is to decide (1) the container locations (yard bays) to use for each subsequence;
and (2) the partial tour for each subsequence, i.e., the visiting order of the yard
bays assigned to each subsequence in (1). The objective is to minimize the total
setup and travel time the yard crane requires to pick up all the containers in
the load plan.
The following pickup-scheduling model follows Kim and Kim (1999b). It
addresses the scheduling of a single yard crane. The problem parameters are:
m: Number of subsequences (and partial tours) that constitute a com-
plete tour for a yard crane;
n: Number of yard bays;
l: Number of container groups;
chj : Initial number of containers of group h stacked at yard bay j;
t: Subsequence number, t = 1 2     m;
 Ch. 8. Intermodal Transportation 511

rt : Number of containers in subsequence t;

gt : Container-group number of subsequence t;
S(h): Set of subsequences for the container-group number h;
B(h): Set of the yard bays where containers of group h are stored;
S: Initial location of the yard crane, noted B(g0 ) = {S};
T: Final location of the yard crane, noted B(gm+1 ) = {T };
dij : Travel distance of the yard crane between yard bays i and j;
eij : Equals 1, if yard bays i and j are distinct and 0, otherwise (i.e.,
i = j);
Ts : Setup time of the yard crane for each visit to a yard-bay;
Td : Travel time of the yard crane per the distance of a yard-bay length.
The decision variables are:
yijt = 1 if the yard crane moves from yard bay i to yard bay j after
0 m
completing the partial tour t; 0 otherwise; ySj and yjT denote
the first and the final movements of the yard crane during
the tour, respectively;
zijt = 1 if the yard crane moves from yard bay i to yard bay j during
the partial tour t; 0 otherwise;
xtj = Number of containers picked up at yard bay j during the partial
tour t
The problem may then be formulated as follows:
m 
  
m 
minimize Ts eij yijt + zijt
t=0 i∈B(gt )j∈B(gt+1 ) t=1 ij∈B(gt )
 

m  
m 
+ Td dij yijt + dij zijt
t=0 i∈B(gt )j∈B(gt+1 ) t=1 ij∈B(gt )

m 
= (Ts eij + Td dij )yijt
t=0 i∈B(gt )j∈B(gt+1 )

m 
+ (Ts eij + Td dij )zijt (83)
t=1 ij∈B(gt )
subject to

0
ySj = 1 (84)
j∈B(g1 )

m
− yjT = −1 (85)
j∈B(gm )
512 T.G. Crainic and K.H. Kim
   
yjit−1 − yijt + t
zki t
− zik = 0
j∈B(gt−1 ) j∈B(gt+1 ) k∈B(gt )
i ∈ B(gt ) t = 1 2     m (86)

zijt  |N| − 1 N ⊆ B(gt ) t = 1 2     m (87)
ij∈B(gt )
   
xtj M t
zkj + yijt−1 
k∈B(gt ) i∈B(gt−1 )
j ∈ B(gt ) t = 1 2     m (88)

xtj = rt  t = 1 2     m (89)
j∈B(gt )

xtj  chj  h = 1 2     l j ∈ B(h) (90)
t∈S(h)
yijt ∈ {0 1} i ∈ B(gt ) j ∈ B(gt+1 ) t = 1 2     m (91)
zijt ∈ {0 1} i j ∈ B(gt ) t = 1 2     m (92)
xtj  0 j ∈ B(gt ) t = 1 2     m (93)
where M is a sufficiently large number and |N| denotes the cardinality of set N.
The objective function (83) minimizes the total time required by the yard
crane to perform the load plan, which depends on the total number of setups
and the total distance traveled. Because a setup occurs whenever a yard crane
moves from one yard bay to another, only the inter-yard-bay movements are
considered in the evaluation of the total number of setups. A feasible solution
corresponds to a path from the source node to the terminal node of the net-
work. Constraints (84) and (85) represent conservation of flow at the source
and terminal nodes, respectively, while constraints (86) enforce the flow con-
servation at the other nodes. Relations (87) are subtour elimination constraints
that exclude isolated cycles (i.e., not connected to a path from the source to the
destination node) from a partial tour. Constraints (88) imply that only when a
yard crane visits a yard bay can it pick up containers there. Constraints (89)
enforce the condition that the number of containers picked up in a partial tour
must equal the number of containers requested by the load plan. Constraints
(90) indicate that the total number of containers picked up during the entire
tour must not be larger than the initial number of containers available at each
yard bay.
To build this formulation, Kim and Kim (1999b) took advantage of the lim-
ited motion of yard cranes, which move on a straight rail line, to reduce the
solution space. Kim and Kim (1999d) and Kim and Kim (1999c) generalized
this model for the pickup-scheduling problem for a single straddle carrier and
for multiple straddle carriers, respectively. Kim and Kim (2003) compared ex-
act optimization, a beam search heuristic, and a genetic metaheuristic for solv-
ing the model proposed by Kim and Kim (1999b). Based on objective-function
 Ch. 8. Intermodal Transportation 513

value, the beam search outperformed the genetic metaheuristic on problems

with 30 yard bays for which exact optimization algorithms could not be used
due to excessive computing times. Ryu et al. (2001) suggested an ant system
metaheuristic for the same formulation and compared it to a tabu search meta-
heuristic on problem instances with 39–514 ships and 36–350 yard bays. On
average, the ant colony metaheuristic was 49 times slower than the tabu search
but achieved a solution value 17% better.
Narasimhan and Palekar (2002) analyzed a generalized version of the
scheduling problem proposed by Kim and Kim (1999b). The authors relaxed
the assumption that the order of subsequences is fixed, which may yield a
reduction in the total handling time of transfer cranes. They also proved a
number of properties of the optimal solution to this version of the problem, in
particular that:
(1) the yard bays visited by a transfer crane during a partial tour must be
contiguous in the yard-bay map;
(2) sets of yard bays for partial tours of the same container group coincide
with partitions of the yard-bay map; and
(3) the minimization of the total distance traveled by transfer cranes results
in the minimization of the transfer-crane setup time.
The authors then restated the problem as an odd-node-matching problem
where edges are added at minimum cost such that every node of the graph
has en even degree. They proved that this version of the problem is NP-hard in
the strong sense and proposed both a branch-and-bound and an enumerative
heuristic to address it.
The problem of constructing a containership loading plan at the level of in-
dividual containers appears more complex than the pickup-scheduling one. On
the one hand, the problem is larger because the loading sequence of individual
containers must be determined. On the other hand, the goals and limitations
related to the operation of various pieces of equipment, quay and transfer
cranes notably, and by the constraints imposed by the stowage plan must be ac-
counted for (Kim et al., 2004). The objectives related to the operation of quay
cranes are first to fill slots in the same hold, then stack containers onto the
same tier on deck, and finally stack containers of weights included in the same
weight group as specified in the stowage plan. The objectives related to the
operation of transfer cranes are to minimize the travel time of transfer cranes,
minimize the number of rehandles (the same container that is moved more
than once to, for example, allow access to containers stacked underneath), and
pick up containers in locations nearer to the transfer point earlier than those
located farther from the transfer point. The constraints related to the oper-
ation of quay cranes are to follow precedence relationships among slots that
follow from their own work schedules and the relative positions of slots in a
ship bay, not violate the maximum allowed total weight of the stack on deck,
not violate the maximum allowed height of the stack of a hold, and load the
same type of containers as specified in the stowage plan. A constraint related
514 T.G. Crainic and K.H. Kim

to the operation of transfer cranes is to maintain sufficient distance between

adjacent transfer cranes to avoid interferences.
The early contributions presented software based on heuristic rules and sim-
ple procedures (e.g., Cojeen and Dyke, 1976). Kim et al. (2004) proposed an
integer linear programming model for sequencing individual outbound con-
tainers. The cost-minimization objective accounted for the handling cost of
both quay and transfer cranes, while constraints included restrictions on the
maximum weight allowed on deck, maximum height of stacks in holds, and the
stability of the vessels. The authors proposed a beam search heuristic for this
formulation. Kozan and Preston (1999) focused on the scheduling of transfer
operations where the actual sequencing issues are not considered. The authors
proposed a mixed-integer formulation to determine the storage locations of
containers in the yard, the yard equipment to move the containers, and the
schedules of these movements, such that the total travel cost of the yard equip-
ment is minimized. A genetic metaheuristic was proposed as solution method.

5.4 Storage activities in the yard

The scope of most research dedicated to the operation of yards is to mini-

mize the container handling efforts and utilize the storage space efficiently. In
addressing these issues, one must consider the rather different characteristics
of inbound (usually import) and outbound (export) container flows. Inbound
containers are usually unloaded fast not to delay the ship departure, but are
retrieved over a long period of time compared to the ship unloading time, in a
somewhat random sequence due to uncertainties related to importing formali-
ties and the operations of land transportation modes. The minimization of the
number of container rehandlings is, thus, the most important issue in this case.
The same characteristics of the land and maritime transportation modes make
outbound containers arrive at the yard randomly over a long period of time,
while their loading is performed rapidly. Pre-planning storage spaces for arriv-
ing containers will result in less rehandlings and a more efficient ship loading
operation.
A number of early studies contributed greatly to the understanding of these
problems. They described practical issues related to space allocation, analyzed
basic properties related to container handling in storage yards, and proposed
formulas to evaluate yard productivity measures such as the total handling ef-
fort or the changes in container inventory due to various container-handling
strategies (Taleb-Ibrahimi et al., 1993; Castilho and Daganzo, 1993; Kim, 1997;
Chen, 1999).
The space-allocation problem is concerned with determining storage loca-
tions for containers either individually or as a group. The problem may focus
on inbound or outbound container movements, or it may consider both. Stor-
age activities and requirements in yards may be initiated by several sources,
such as export containers arriving through gates by truck, at quay by feeder
vessels, or by rail service, as well as import containers unloaded from vessels
 Ch. 8. Intermodal Transportation 515

and moved out of the terminal by trucks, feeder vessels, or rail. Each storage
activity is characterized by the amount of storage space required, the source
and destination of the containers involved (e.g., export containers unloaded
from rail have rail as source and the berth as destination), and the starting and
ending times of the storage activity. Information regarding these factors can
be obtained from delivery schedules or forecasts, as well as unloading/loading
schedules for vessels.
We present a formulation that follows that of Kim and Park (2003b) and may
be applied to the planning of storage locations for both inbound and outbound
containers. The formulation is defined for a planning horizon discretized in T
periods, t = 1 2     T , according to the starting and ending times of storage
activities. Let ai and bi represent the starting and the ending times of stor-
age activity i = 1 2     l, respectively, and di the associated storage space
required (i.e., the number of containers in twenty-foot equivalent units). The
space-allocation problem may then be expressed as a minimum cost multicom-
modity network flow problem on a time–space network G = (N  A). Each
storage activity i generates a source node S i and a destination node T i . Interme-
diate nodes jt identify storage locations j = 1 2     n at time periods t. Arcs
from each source node S i to intermediate nodes in period ai represent the pos-
sible movements of containers from their origin toward the storage locations
at the starting time of storage activity i. Symmetrically, arcs from intermediate
nodes to a destination node T i in period bi stand for the possible movements
of containers from the storage locations to their destination at the ending time
of storage activity i. The corresponding transportation costs are associated to
each of these two types of arcs, but no limits are imposed on their capacity.
Holding arcs are defined between intermediate nodes representing the same
storage location at consecutive time periods. The capacity uj of holding arc
(jt  jt+1 ) corresponds to the stowage capacity of the associated stacking area
j = 1 2     n, for t = 1 2     T . Holding costs are not included in the for-
mulation.
A storage activity i can be described as a route Rij in G from node S i to one
of the intermediate nodes (storage location), through several holding arcs, to
node T i . The number of such routes equals the number of available storage
locations:
Rij : S i → jai → jai +1 → jai +2 → · · · → jbi −1 → jbi → T i 
A unit transportation cost cij is associated to each route Rij of storage activity i.
j
Let δRt ij equal 1 if arc (jt  jt+1 ) is included in route Rij of storage activity i and
0, otherwise.
Define the decision variables xij as the number of containers of storage ac-
tivity i that moves following route Rij . The space-allocation problem can then
be formulated as follows:

l 
n
minimize cij fik (94)
i=1 k=1
516 T.G. Crainic and K.H. Kim

subject to

l
j
δRt ij fik  uj  j = 1 2     n t = 1 2     T (95)
i=1
n
fik = di  i = 1 2     l (96)
j=1
fik  0 and integer i = 1 2     l j = 1 2     n (97)
The objective function (94) minimizes the total transportation cost of all
the storage activities. Constraints (95) enforce the storage space limitations
at each storage location and time period. Constraints (96) indicate that the
space requirement of each storage activity must be satisfied. Notice that stor-
age activities have been defined with constant space requirements for all time
periods. A storage activity with a time-varying space requirement may be rep-
resented in this formulation as a combination of multiple storage activities with
different starting and ending times and space requirement, respectively.
Formulation (94)–(97) corresponds to an integer, capacitated, minimum
cost multicommodity flow problem, which is NP-hard (assuming all storage
activities begin and end at the same time, the problem reduces to a knap-
sack problem, which is a well-known NP-hard problem). Moreover, real-world
problems are frequently of very large dimensions. Thus, for example, the for-
mulation for a container terminal with 20 blocks and 4 berths yields a math-
ematical model with more than 100,000 storage activities. Most contributions
that focused on the container terminal storage activities thus propose heuristic
solution methods. Methodology based on mathematical programming meth-
ods for large-scale integer multicommodity network flow problems has been
proposed by Barnhart et al. (1998) and Barnhart et al. (2000).
Cao and Uebe (1995) addressed the static, single-period version of the pre-
vious problem and proposed a transportation problem formulation. Zhang et
al. (2003) generalized the multiperiod problem of Kim and Park (2003b) by
including into the objective function terms representing the balancing of han-
dling activities among different storage areas, as well as the minimization of the
total distance traveled by yard trucks between storage areas and quay cranes.
The authors considered the flows of both inbound and outbound containers.
They decomposed the problem into two stages. The first determined the total
number of unloaded (from ships) and received (from land modes) containers
to balance their distribution among yard blocks. The second stage determined
the number of containers associated to each vessel allocated to each yard block
in order to minimize the total travel cost of yard trucks. The first stage prob-
lem was formulated as a linear programming model, while the second took the
form of a transportation problem.
Several other studies addressed aspects of the storage-space allocation
problem different from Kim and Park (2003b). Kim and Bae (1998) exam-
ined transportation and dynamic programming models to schedule the remar-
 Ch. 8. Intermodal Transportation 517

shaling operations for outbound containers. Kim et al. (2000) addressed the
problem of locating individual outbound containers considering the container
weights. Kim and Park (2003a) focused on the space-allocation problem for
export containers only. They considering various practical constraints and pro-
posed integer programming formulations. Kim and Kim (1999a) addressed the
space-allocation problem for import containers in which the stacking height
and the amount of space (not the storage locations) are simultaneously deter-
mined to accommodate dynamically changing space requirements. The authors
proposed a nonlinear programming model and a solution method based on La-
grangian relaxation. Kozan (2000), Preston and Kozan (2001), and Mattfeld
and Kopfer (2004) proposed integer linear programming models to simul-
taneously determine storage locations and schedules of transfer operations.
Castilho and Daganzo (1991) and Holguin-Veras and Jara-Diaz (1999) pro-
posed methods for determining prices of storage spaces for containers or gen-
eral cargo.

5.5 Allocation and dispatching of yard cranes and transporters

Yard equipment is an important container terminal resource that includes

yard cranes and transporters, i.e., yard trucks and straddle carriers. In auto-
mated container terminals, automated guided vehicles belong to this category.
The main operational decision issue at this level concerns the assignment of
container-handling tasks to different pieces of equipment. Yard equipment is
considered less critical than berths, storage spaces, and quay cranes for which
schedules are constructed several hours, or even days, in advance. Conse-
quently, no such planning is performed for yard equipment. Instead, decisions
on allocation and dispatching yard equipment are made only for the limited
number of activities to be carried out in the near future (e.g., one hour and
less).
The objectives of allocating and dispatching yard cranes and transporters
are usually to minimize the total travel distance, the total waiting time, or
the total delays of equipment beyond prespecified due times. The dispatch-
ing problem for transporters was studied by Bish (2003), Kim and Bae (1999,
2004), Böse et al. (2000), and Grunow et al. (2004). Zhang et al. (2002),
Cheung et al. (2002), Kim et al. (2003), Lai and Lam (1994), and Lai and Leung
(1996) addressed the allocation and dispatching problem for yard cranes.
Consider two ships, sh− and sh+ , that need unloading and loading, respec-
tively, and that are berthed around the same time so that they can be served by
the same set of k vehicles. It is assumed, however, that no ship-to-ship move-
ment is possible. Let N − and N + denote the sets of containers that will be
unloaded (from ship sh− ) and loaded (onto ship sh+ ), respectively. Let ci be
container i in N − . Associated to each container to be loaded onto a ship is its
current storage location in the yard area, which is known. Each such container
will require a loaded vehicle trip from its current storage location to the location
518 T.G. Crainic and K.H. Kim

of ship sh+ . Let L+ denote the set of current storage locations of the contain-
ers in N + . Symmetrically, a set of potential storage locations in the yard area
is reserved for the containers of each unloading ship. Set L− contains the po-
tential storage locations reserved for all containers to be unloaded from sh− .
Each unloaded container will require a loaded vehicle trip from the location of
ship sh− to its selected storage location. We make the simplifying assumption
that sets L+ and L− are disjoint, that is, no container in N − can be stored in
a location currently occupied by a container in N + . Let L = L− ∪ L+ and Wi
be the subset of L− where ci can be stacked.
When the destination of a loaded trip differs from the origin of the next
loaded trip on a vehicle schedule, the vehicle needs to make an empty, repo-
sitioning movement. The total amount of empty vehicle trips thus depends on
the sequence of loaded trips of each vehicle and should be minimized by, for
example, matching each loaded trip for an unloaded container to a loaded
trip for a loading container. The goal therefore is to (1) determine a storage
location for each unloaded container, (2) construct round vehicle trips, and
(3) assign round trips to each vehicle. The first problem consists in assigning
each unloaded container ci ∈ N − to a candidate storage location p ∈ Wi . Con-
structing round vehicle trips, the second problem, corresponds to matching a
trip loaded with an unloaded container with a trip loaded with a loading con-
tainer. Finally, the third problem consists in constructing sequences of round
trips and assigning these to vehicles to minimize the make-span.
The following formulation follows the model proposed by Bish (2003) to si-
multaneously address the first two problems: determine storage locations for
containers being unloaded from ships, schedule unloading and loading oper-
ations, and dispatch AGVs. The model is based on a network where supply
nodes, with unit supply, correspond to unloaded containers ci ∈ N − , demand
nodes, with unit demands, correspond to current storage locations of loading
containers lq , q ∈ L+ , and transshipment nodes lp stand for potential storage
locations p ∈ Wi reserved for unloading containers. The arc set is given by
A = {(ci  lp ) | ci ∈ N −  p ∈ Wi } ∪ {(lp  lq ) | p ∈ L−  q ∈ L+ }, each arc with
unit capacity. Arc (ci  lp ) ∈ A corresponds to the trip of an unloading con-
tainer ci that is to be stored in location lp ; Its cost tlp is thus the corresponding
travel time. Arc (lp  lq ) ∈ A corresponds to the empty trip of a vehicle that
just completed the delivery on an unloaded container to location lq . The vehi-
cle moves to location lq to pick up a loading container. The cost of the arc λpq
is thus the travel time of the empty movement.
Define the decision variables xuv , (u v) ∈ A, that equal 1 if the activity
corresponding to arc (u v) is to be performed (the arc is selected) and 0, oth-
erwise. We can now model the problem as a transshipment formulation:
   
minimize tlp xci lp + λpq xlp lq (98)
ci ∈N − p∈Wi p∈L− q∈L+
subject to
 Ch. 8. Intermodal Transportation 519

xci lp = 1 ci ∈ N −  (99)
p∈Wi
 
xci lp = xlp lq  p ∈ L−  (100)
ci ∈N − q∈L+

xlp lq = 1 q ∈ L+  (101)
p∈L−
xuv ∈ {0 1} for all (u v) ∈ A (102)
The objective function (98) minimizes the total assignment and matching-
related travel time of the yard equipment. Constraints (99) ensure that each
unloaded container is assigned to exactly one loaded trip and thus, to exactly
one unloading location. Constrains (100) enforce the flow-balance require-
ments at storage sites for unloading containers. Relations (101) ensure that
each loaded trip with an unloaded container is matched with a loaded trip with
a loading container. The integrality conditions (102) imposed on the decision
variables are not really required due to the total unimodularity property the
formulation. The author proposed a heuristic for realistically-sized problem
instances. The solution of the model yields a set of round trips for vehicles,
which consist of a loaded movement with an unloaded container to a storage
location, an empty repositioning trip from this location to a pickup location of
a loading container, a loaded movement from this pickup location to a loading
position under a quay crane, and an empty travel from there to the unloading
position of another quay crane. These round trips are then assigned to vehicles
by a list-scheduling heuristic.
Several other studies have been dedicated to issues related to yard equip-
ment allocation and dispatching. Kim and Bae (1999) addressed the dispatch-
ing problem of AGVs in automated container terminals under the assumption
that storage locations of containers as well as the schedules for unloading and
loading operations by quay cranes are given. The scope is thus narrower than
that of Bish (2003). Kim and Bae (1999) considered quay cranes are the most
expensive equipment in container terminals. The model they proposed aimed
therefore to minimize the total idle time of a quay crane resulting from late
arrivals of AGVs as well as the associated total travel time. The authors sug-
gested a network-based mixed-integer linear model and provided a heuristic
algorithm. Nodes in the network represent “events” in time and space that
correspond to moments (the event time) when vehicles pick up containers. An
arc from one node to another indicates that the time lapse between the event
times of the two operations allows a vehicle to drop off a container at the desti-
nation node after completing its pickup at the origin node. The problem is then
to find the optimal routes on the network, each representing a sequence of de-
livery tasks assigned to a specific vehicle. Kim and Bae (2004) extended this
approach to the case with multiple quay cranes in which dual cycle operation
of AGVs is allowed.
520 T.G. Crainic and K.H. Kim

Böse et al. (2000) addressed the dispatching problem of straddle carriers

with the objective of minimizing the delays of quay cranes. They also assumed
that storage locations of containers are already determined. The problem then
becomes to assign delivery tasks of unloading and loading containers to strad-
dle carriers. The problem is basically the same as the one addressed by Kim
and Bae (2004) except that Böse et al. (2000) do not consider the vehicle travel
times. The authors did not provide a formal model but proposed a genetic
metaheuristic to search for solutions. Bish et al. (2005) represented the dis-
patching of vehicles as a machine-scheduling problem and proposed a greedy
heuristic that assigned deliveries (jobs) to the first available vehicle (machine)
which may arrive on time at the designated destination location, i.e., the quay
crane or the location of the loading container. The authors showed that the
greedy heuristic yields a job sequence that minimizes the makespan. Grunow
et al. (2004) generalized the previous studies by considering the dispatch-
ing problem for multiload vehicles in automated container terminals. Indeed,
while previous studies assumed that a vehicle can move only one container
at a time, the authors considered the fact that a vehicle which moves one
forty-foot container can move two twenty-foot containers. They proposed a
mixed-integer linear programming model and suggested priority rules for dis-
patching delivery tasks to two-load AGVs.
Similar studies have been done for yard cranes. The objectives considered in
most cases were the total waiting time and the total delays of trucks. Zhang et
al. (2002) and Cheung et al. (2002) solved static versions of the crane deploy-
ment problem when the total workload at each storage area is known in ad-
vance. Zhang et al. (2002) proposed a mixed-integer programming model and
addressed it by a method based on Lagrangian relaxation. Cheung et al. (2002)
addressed a similar problem but removed the restriction that crane movements
must be completed within a single period. This allows to use a shorter period
length resulting in a more accurate model. A successive piecewise-linear ap-
proximation method was proposed. Kim et al. (2003) addressed the problem
of sequencing transfer tasks of a yard crane for outside trucks in dynamic sit-
uations where new trucks arrive continuously. A dynamic programming model
was suggested and decision rules derived by a reinforcement learning tech-
nique were proposed. Lai and Lam (1994) and Lai and Leung (1996) proposed
various dispatching rules for yard cranes and tested them by simulation.

6 Strategic planning of multimodal systems

The focus of the models and methods presented in this section is broad:
strategic planning issues at the international, national, and regional level,
where the movements of several commodities through the transportation net-
works and services of several carriers are considered simultaneously. The main
questions address the evolution of a given transportation system and its re-
sponse to various modifications in its environment: evolution of the “local” or
 Ch. 8. Intermodal Transportation 521

international socioeconomic environment resulting in modifications to the pat-

terns and volumes of production, consumption, and trade; modifications to ex-
isting policies and legislation and introduction of new ones (e.g., environment-
related taxes); changes to existing infrastructure; variations in energy prices;
modifications to labor conditions; carrier mergers; introduction of new tech-
nologies, and so on and so forth. These questions are often part of cost-benefit
analyzes and comparative studies of policy and investment alternatives.
A strategic planning methodology identifies and represents the fundamental
components of a transportation system – demand, supply, performance mea-
sures, and decision criteria – and their interactions. It models flow volumes
by commodity and transportation mode, as well as associated performance
measures, defined on a network representation of the transportation system.
It aims to achieve a sufficiently good simulation of the global behavior of the
system to offer a correct representation of the current situation and serve as an
adequate analysis tool for planned or forecast scenarios and policies. It must be
tractable and produce results that are easily accessible. No single formulation
may address such a broad scope. Consequently, a strategic planning methodol-
ogy is typically a set of models and procedures. Other than data manipulation
tools (e.g., collection, fusion, updating, validation, etc.) and result analysis ca-
pabilities (e.g., cost-benefit, environmental impacts, energy consumption poli-
cies, etc.), the main components are:
(1) Supply modeling to represent the transportation modes, infrastructure,
carriers, services, and lines; vehicles and convoys; terminals and intermodal
facilities; capacities and congestion; economic, service, and performance mea-
sures and criteria.
(2) Demand modeling to capture the product definitions, identify produc-
ers, shippers, and intermediaries, and represent production, consumption, and
zone-to-zone (region-to-region) distribution volumes, as well as mode choices
for transportation. Relations of demand and mode choice to the performance
of economic policies and transportation system performance are also ad-
dressed here.
(3) Assignment of multicommodity flows (from the demand model) to the
multimode network (the supply representation). This procedure simulates the
behavior of the transportation system and its output forms the basis for the
strategic analyses and planning activities. The assignment methodology must
therefore be both precise in reproducing current situations and sufficiently
general to produce robust analyzes of future scenarios based on forecast data.
The prediction of multicommodity freight flows over a multimodal network
is an important component of transportation science and has generated sig-
nificant interest in recent years. However, perhaps due to the inherent dif-
ficulty and complexity of such problems, the study of freight flows at the
national or regional level has not yet achieved full maturity, in contrast to
passenger transportation where the prediction of car and transit flows over
multimodal networks has been studied extensively and several of the re-
search results have been transferred to practice (Florian and Hearn, 1995;
522 T.G. Crainic and K.H. Kim

Cascetta, 2001). In the following, we review the most frequently used method-
ologies for freight planning and briefly describe associated references.

Demand

The modeling of demand attempts to describe the economic activities of a

region, its production, consumption, import and export of goods. For planning
purposes, the output of demand models is a series of product (or commodity
group) specific demand matrices indicating the volumes to be moved from one
zone to another. The process is often completed by the modeling of mode
choice, which specifies for each product and origin–destination combination
on what set of transportation infrastructure or services the demand may be
moved.
A number of countries have developed input/output models of their econ-
omy that serve to determine the basic production and attraction of goods
(Isard, 1951; Cascetta, 2001 and references within). In order to use an
input/output model as a demand model, it is necessary to disaggregate the
input/output model inputs and outputs by region and zone. This process is
complex, and is usually not integrated with a supply representation and as-
signment procedure. When an input/output model is not available, the initial
determination of origin–destination matrices is carried out by using national
statistics on production, consumption, imports and exports combined with
surveys of particular industrial sectors to complete missing or unreliable in-
formation. This process may be tedious since one has to reconcile data from
several sources that may be collected by using different geographical subdi-
visions or inconsistent product definitions. The results of the disaggregated
input/output model or the ad-hoc estimation procedures serve for the initial
computation of origin-destination matrices for each product but without a sub-
division by mode.
A second class of models that is well studied for the prediction of interre-
gional commodity flows is the spatial price equilibrium model and its variants
(Friesz et al., 1983; Harker and Friesz, 1986a, 1986b; Harker, 1987; see also
Florian and Hearn, 1995; Nagurney, 1993). This class of models determines
simultaneously the flows between producing and consuming regions, as well as
the selling and buying prices that satisfy the spatial equilibrium conditions.
A spatial equilibrium is reached when for all pairs of supply and demand
regions with a positive commodity flow, the unit supply price plus the unit
transportation cost is equal to the unit demand price; the sum is larger than
this price for all pairs of regions with no exchanges. A simple network (bi-
partite graph) is generally used to represent the transportation system. These
models rely to a large extent on the supply and demand functions of producers
and consumers, respectively, which are rarely available and quite difficult to
calibrate. There are relatively few applications of this class of models for the
determination of demand by product and these deal with specific commodities
such as crude oil, coal or milk products.
 Ch. 8. Intermodal Transportation 523

Mode choice

Mode-choice models aim to describe the set of transportation modes or

services that may be used to carry specific products of groups thereof. The
mode-choice definition may be rather general, e.g., petroleum moves by ship
and pipeline, extremely specific indicating a particular set of single or mul-
timodal paths for a given product, shipper, and origin–destination pair, or
anywhere in between. The level of detail of modal specification need not be
the same for all products or interzonal trade flows. The specification of mode
choice may be inferred from historical data and shipper surveys or it may result
from a formal description and modeling effort (Winston, 1983). The output of
this process are either coefficients that indicate how to split the demand of
a given origin–destination pair between the paths of a given set, or origin–
destination demand matrices with particular sets of allowed modes.
Random utility models, developed and largely used for the analysis and plan-
ning of person transportation systems, have been proposed for freight trans-
portation as well but their use in actual applications is scarce (Cascetta, 2001).
The huge number of paths that have to be explicitly generated and stored, cou-
pled to the challenge of performing this task for forecast data, may explain
this phenomenon. At aggregated levels, mode choices have been specified for
particularly important product flows by explicitly surveying the major logistic
chains used between pairs of macro regions.

Supply representation and assignment

Once modal origin–destination matrices have been developed by some

means, the next step is to assign them to the supply network model by us-
ing some route choice mechanism. The results of such an assignment model –
product flows and performance measures – form part of the input to demand
and cost-benefit modeling and analysis.
One class of assignment mechanisms is based again on the application of
random utility models to the choice of paths defined previously by the mode-
choice phase. It is noteworthy that the attributes of pre-defined paths are
determined by the state of the network at generation time and are not re-
sponsive to assignment results. Thus, for example, congestion conditions are
very difficult to represent. Moreover, the utility and choice models have to be
calibrated, and all paths have to be generated, for each scenario, which is quite
difficult to perform when forecast data is used.
Another class, network optimization models enable the prediction of multi-
commodity flows over a multimodal network that represents the transportation
facilities at a level of detail appropriate for a nation or region. The demand and
mode choice are exogenous and intermodal shipments are permitted. Within
the specified mode choice, the optimization (assignment) engine determines
the best (with respect to the specified network performance measures) mul-
timodal paths for each product and origin–destination pair. The emphasis is
524 T.G. Crainic and K.H. Kim

on the proper representation of the network and its different transportation

modes, the corresponding intermodal transfer operations, the various criteria
used to determine the movement of freight, the interactions and competi-
tion for limited resources captured through the representation of congestion
effects, and the associated estimation of the traffic distribution over the trans-
portation system considered to be used for comparative studies or for discrete
time multiperiod analyses.
Studies in the 1970s used rather simple network representations (e.g., Jones
and Sharp, 1977; Sharp, 1979). Several studies also attempted to extend spa-
tial equilibrium models to include more refined network representations and
to consider congestion effects and shipper–carrier interactions. Friesz et al.
(1986) present a sequential model which uses two network representations: de-
tailed separate networks for each carrier, and an aggregate, shipper-perceived
network. On each carrier network commodities are transported at the least to-
tal cost. On the shipper-perceived network, traffic equilibrium principles are
used to determine the carriers that shippers choose to move their traffic. This
approach was quite successful in the study of logistics of products where a very
limited number of shippers and carriers interact and strongly determine the
behavior of the system. A typical example is the coal market between electric
utilities in the United States and their suppliers in exporting countries. Friesz
and Harker (1985), Harker and Friesz (1986a, 1986b), Harker (1987, 1988),
and Hurley and Petersen (1994) present more elaborate formulations. This line
of research has not, however, yielded practical planning models and tools yet,
mainly because the formulations become too large and complex when applied
to realistic situations. For a more detailed review of these efforts see Guélat
et al. (1990) and Crainic et al. (1990b).
Models based on more sophisticated representations of the supply network
were introduced by Guélat et al. (1990) and Jourquin and Beuthe (1996).
We follow the former in the following presentation. The proposed modeling
framework is that of a multimodal network, made up of modes, nodes, links,
and intermodal transfers, on which multiple products are to be moved by spe-
cific vehicles and convoys between given origin and destination points. Here,
a mode is a means of transportation having its own characteristics, such as ve-
hicle type and capacity, as well as specific cost measures. Depending on the
scope and level of detail of the strategic study, a mode may represent a carrier
or part of its network representing a particular transportation service, an ag-
gregation of several carrier networks, or specific transportation infrastructures
such as ports.
The network consists of nodes N , links A, modes M, and transfers T that
represent all possible physical movements on the available infrastructure. To
capture the modal characteristics of transportation, a link a ∈ A is defined as a
triplet (i m j), where i ∈ N is the origin node, j ∈ N is the destination node,
and m ∈ M is the mode. Parallel links are used to represent situations where
more than one mode is available for transporting goods between two adjacent
nodes. This network representation enables easy identification of the flow of
 Ch. 8. Intermodal Transportation 525

goods by mode, as well as various cost functions (e.g., operating cost, time
delay, energy consumption, emissions, noise, risk, etc.) by product and mode.
To model intermodal shipments, one must allow for mode transfers at certain
nodes of the network and compute the associated costs and delays. Intermodal
transfers t ∈ T at a node of the network are modeled as link to link, hence
mode to mode, allowed movements. A path in this network then consists of a
sequence of directed links of a mode, a possible transfer to another mode, a
sequence of directed links of the second mode, and so on.
A product is any commodity (or collection of similar products) – goods or
passengers – that generates a link flow. Each product p ∈ P transported over
the multimodal network is shipped from certain origins o ∈ N to certain desti-
nations d ∈ N within the network. The demand for each product is exogenous
and is specified by a set of O–D matrices. The mode choice for each product
is also exogenous and is indicated by defining for each O–D matrix a subset of
modes allowed for transporting the corresponding demand. Shipper behav-
ior is assumed to be reflected in these O–D matrices and associated mode
choice. Let gm(p) be a demand matrix associated with product p ∈ P , where
m(p) ⊆ M is the subset of modes that may be used to move this particular
part of product p.
The flows of product p ∈ P on the multimodal network are the decision
p
variables of the model. Flows on links a ∈ A are denoted by va and flows
p
on transfers t ∈ T are denoted by vt ; v stands for the vector of all prod-
uct flows. Vehicle and convoy (e.g., train) movements are deduced from these
flows. Cost functions are associated with the links and transfers of the network.
p p
For product p, the respective average cost functions sa (v) and st (v) depend
on the transported volume of goods. Then, the total cost of product p on arc a
p p p p
is sa (v)va , and it is st (v)vt on transfer t. The total cost over the multimodal
network is the function F, which is to be minimized over the set of flow volumes
that satisfy the flow conservation and non negativity constraints:
 p p
 p p

F= sa (v)va + st (v)vt  (103)
p∈P a∈A t∈T
m(p)
Let Lod denote the set of paths that for product p lead from origin o
to destination d using only modes in m(p). The path formulation of the flow
conservation equations are then:
 m(p)
hl = god  o d ∈ N  p ∈ P  m(p) ⊆ M (104)
m(p)
l∈Lod

m(p)
where hl is the flow on path l ∈ Lod . These constraints specify that the total
flow moved over all the paths that may be used to transport product p must be
equal to the demand for that product. The nonnegativity constraints are:
m(p)
hl  0 l ∈ Lod  o d ∈ N  p ∈ P  m(p) ⊆ M (105)
526 T.G. Crainic and K.H. Kim

p
The relation between arc flows and path flows is va = l∈Lp δal hl , a ∈ A,
p ∈ P , where Lp is the set of all paths that may be used by product p, and
δal = 1 if a ∈ l (and 0, otherwise) is the indicator function which identifies the
p
arcs of a particular path. Similarly, the flows on transfers are vt = l∈Lp δtl hl ,
t ∈ T , p ∈ P , where δal = 1 if t ∈ l (and 0, otherwise). Then, the system
optimal multiproduct, multimodal assignment model consists of minimizing
(103), subject to constraints (104) and (105). The optimality principle ensures
that in the final flow distribution, for each product, demand matrix, and origin–
destination pair, all paths with positive flows will have the same marginal cost
(lower than on the other paths). The algorithm developed for this problem
exploits the natural decomposition by product and results in a Gauss–Seidel-
like procedure which allows the solution of large size problems in reasonable
computational times (Guélat et al., 1990).
This model and algorithm are embedded in the STAN interactive-graphic
system where they are complemented by a large number of tools to input, dis-
play, analyze, modify, and output data, as well as implement demand, mode
choice, performance, and analysis models. See Larin et al. (2000) for a detailed
description of the STAN system, components, interfaces, and tools. STAN has
been applied successfully in practice for scenario analysis and planning, and
several agencies and organizations in a number of countries around the world
use it (Crainic et al., 1990a, 1990b, 1994, 1999, 2002; Guélat et al., 1990).

7 Perspectives

Intermodal transportation, particularly container-based, is steadily growing

and will continue to do so in the foreseeable future. This is accompanied by
the evolution of the regulatory, economic, and technological environment of
the industry. Enhanced planning and management procedures and decision
technologies are thus required, offering both great opportunities and signifi-
cant challenges for the Operations Research community. On a general note,
while many significant methodological advances have been achieved and sev-
eral have been successfully transferred to actual practice, many problems have
received scant attention. Moreover, advances in vehicle, infrastructure, and
communication technologies yield new problems and require that problems
already studied be revisited. In this section, we identify some of these trends,
opportunities, and challenges.
Container terminals, mainly located in ports, are a case in point. Most re-
search dedicated to this area is very recent and aimed at operational issues.
This may be explained by the fact that terminals are often seen as bottlenecks
in freight transportation and efforts are therefore dedicated to improving their
efficiency and productivity. The development of comprehensive models for
strategic and tactical planning of container terminals offers significant research
opportunities. Moreover, the trend one observes in container terminal au-
tomation makes this research direction extremely timely. How to represent
 Ch. 8. Intermodal Transportation 527

automatic operations in planning models raises interesting questions, however.

The automation of container terminals also opens up research opportunities
in real-time decision and control of operations. Automated equipment collects
and transmits data in realtime. This data, together with historical information
and the plan of operations, could be used to automate and, in some case nearly
optimize, real-time decisions. New models are required, as well as appropriate
solution methods. Automation of terminals also requires revisiting operational
planning models.
Compared to terminals, more work has been dedicated to carrier strategic
and tactical planning issues. Yet, new problems emerge and many challenging
research opportunities exist. Enhancing the models to better represent opera-
tion characteristics and to better integrate line and terminal activities is such
an opportunity; integrating a representation of resource (vehicles, power, man-
power) circulation and scheduling is another. The planning and operations of
the “new” rail intermodal-service networks, operating regular and fixed ser-
vices on a full-asset-utilization basis and enforcing advance bookings, define
new challenging problems for Operations Research and Transportation Sci-
ence.
The more comprehensive integration of the time-dependency of decisions
and of the stochasticity of data and operations into strategic/tactic models is
a major research challenge and opportunity. Indeed, many models aimed at
strategic and tactical planning issues are static and almost all are deterministic.
The conventional wisdom seems to be that such models plan based on “aver-
age” forecast data, while actual operations provide the “recourse” to adjust
the plan to the day-to-day reality. Are we missing something? Could we build
better plans and operate more efficiently by taking stochasticity into account
at planning level? These and similar questions open up a fascinating research
field.
Many operational problems are represented as time-dependent formula-
tions. Crew scheduling, for example. Some, such as the management of empty
vehicles, are also cast as stochastic formulations. Research opportunities exist
in the development and enhancement of such formulations in many sectors of
the industry. Thus, for example, very few results have been reported relative to
the container fleet management problem. Moreover, each of these efforts was
dedicated to one part of the system only. No comprehensive model of container
fleet management on land and ocean is known to the authors. Moreover, while
the scheduling and operation of various resources – vehicles, traction units,
crews, etc. – are clearly related, most current formulations consider them sep-
arately. Research is required in this area.
The growth in the deployment of Intelligent Transportation Systems (ITS)
and the electronic society will continue to impact the planning and operations
of freight transportation. ITS and e-business technologies and procedures in-
crease the flow of data, improve the timeliness and quality of information,
and offer the possibility to control and coordinate operations in real-time. Re-
search is required to adequately model the various planning and management
528 T.G. Crainic and K.H. Kim

problems under ITS and real-time information and to develop efficient so-
lution methods. These efforts must target carriers, terminals, as well as the
entire intermodal chain. The scheduling, assignment, dispatching, routing, and
re-routing of equipment are obvious and challenging subjects. As important is
the impact on planning. Consider, to illustrate, the uncertainties related to the
exact manifest of arriving container ships, their exact arrival time, and the des-
tination (and, eventually, the carrier contracted for the next leg of the journey)
of each container. ITS technology will deliver precise information earlier thus
significantly reducing the uncertainty for the managers of the terminal and of
the carriers that are next in the intermodal chain. Will uncertainty disappear
completely? Unlikely. But its representation in planning and operations mod-
els will change.
Very few efforts have been dedicated to the intermodal chain. The e-business
environment forces the issue and ITS offers the technological support. This
largely unexplored field offers numerous research opportunities and chal-
lenges. The coordination (synchronization, in some cases) of plans and op-
erations of independently owned or managed carriers and terminals is such a
case. The uncertainties related to the operation of each element of the chain,
the relations among these uncertainties, as well as their propagation within the
intermodal chain are of prime importance in this context and pose consider-
able modeling and algorithmic challenges.
Recent years have brought to the forefront security issues related to trans-
portation, ports, and border crossing. Planning and operations models and
methods must be revisited and new ones must be proposed to address these
issues, for each participant in the intermodal chain, as well as for the entire
chain.
Most problems mentioned in this chapter are NP-hard and the formula-
tions proposed are large-scale, mixed-integer combinatorial models. Stochas-
tic, time-dependent formulations make resolution efforts even more difficult.
And the need to build more comprehensive models is not making them any
easier to solve. Significant research must, thus, be dedicated to the method-
ological aspects, including the study of models to develop stronger formu-
lations and bounds. From an algorithmic point of view, the profession must
continue to aim for powerful exact methods to address continuously larger
problem instances. The dimension and complexity of the problems we face
continuously overcome the capability of exact solution methods, however. Ap-
proximate solution methods, metaheuristics in particular, thus, play an increas-
ingly important role in obtaining good solutions to difficult problems within
reasonable computing times. Much work is still needed to develop more ef-
ficient and robust procedures and to better understand the conditions under
which each method performs best. Solution methods that combine characteris-
tics of two or more metaheuristics in a sequential or parallel computing setting
offer interesting, but challenging perspectives.
Parallel and distributed computation indeed offers interesting perspectives
with potentially great rewards: to solve realistically modeled and dimensioned
 Ch. 8. Intermodal Transportation 529

problem instances within reasonable times. Each class of problems and algo-
rithms presents its own challenges. Promising research areas are the parallel
exploration of branch-and-bound trees, the collaborative search undertaken
by several metaheuristics or by metaheuristics and exact methods, and the
development of hierarchical methods that combine different parallel models
and methods (e.g., a first level of cooperating metaheuristics and branch-and-
bound that each call, at a lower level, parallel procedures to evaluate solutions
or bounds). Advanced decomposition techniques are also required, particu-
larly related to the resolution of time-dependent problems. Parallel computing
offers the possibility to design solution architectures to efficiently address com-
plex requests in real, or quasi-real time. These ideas have just begun to be
considered, but present great potential for the development of intelligent and
efficient decision support tools for real-time intermodal transportation sys-
tems.

Acknowledgements

Funding for this project has been provided by the Natural Sciences and En-
gineering Council of Canada, through its Discovery Grant program. This work
was supported in part by the Regional Research Centers Program (Research
Center for Logistics Information Technology) granted by the Korean Ministry
of Education & Human Resources Development.
While working on this project, Dr. T.G. Crainic was Adjunct Professor at the
Département d’informatique et de recherche opérationnelle of the Université
de Montréal (Canada) and at Molde University College (Norway).

References

Addinnour-Helm, S., Venkataramanan, M.A. (1998). Solution approaches to hub location problems.
Annals of Operations Research 78, 31–50.
Ahuja, R.K., Magnanti, T.L., Orlin, J.B. (1993). Network Flows – Theory, Algorithms, and Applications.
Prentice Hall, Englewood Cliffs, NJ.
Aikens, C.H. (1985). Facility location models for distribution planning. European Journal of Operational
Research 22, 263–279.
Alessandri, A., Sacone, S., Siri, S. (2006). Modelling and optimal receding-horizon control of maritime
container terminals. Journal of Mathematical Modelling and Algorithms, in press.
Armacost, A.P., Barnhart, C., Ware, K.A. (2002). Composite variable formulations for express shipment
service network design. Transportation Science 36 (1), 1–20.
Armacost, A.P., Barnhart, C., Ware, K.A., Wilson, A.M. (2004). UPS optimizes its air network. Inter-
faces 34 (1), 15–25.
Assad, A.A. (1980). Models for rail transportation. Transportation Research Part A: Policy and Prac-
tice 14, 205–220.
Aykin, T. (1990). On a quadratic integer program for the location of interacting hub facilities. European
Journal of Operational Research 46 (3), 409–411.
Aykin, T. (1994). Lagrangian relaxation based approaches to capacitated hub-and-spoke network design
problem. European Journal of Operational Research 79 (3), 501–523.
530 T.G. Crainic and K.H. Kim

Aykin, T. (1995a). The hub location and routing problem. European Journal of Operational Research 83
(1), 200–219.
Aykin, T. (1995b). Networking policies for hub-and-spoke systems with application to the air trans-
portation system. Transportation Science 29 (3), 201–221.
Balakrishnan, A., Magnanti, T.L., Mirchandani, P. (1997). Network design. In: Dell’Amico, M., Maffioli,
F., Martello, S. (Eds.), Annotated Bibliographies in Combinatorial Optimization. Wiley, New York,
pp. 311–334.
Ball, M., Magnanti, T.L., Monma, C.L., Nemhauser, G.L. (Eds.) (1995). Network Routing. Handbooks
in Operations Research and Management Science, vol. 8. North-Holland, Amsterdam.
Barnhart, C., Schneur, R.R. (1996). Network design for express freight service. Operations Research 44
(6), 852–863.
Barnhart, C., Johnson, E.L., Nemhauser, G.L., Savelsbergh, M.W.F., Vance, P.H. (1998). Branch-and-
price: Column generation for solving huge integer programs. Operations Research 46 (3), 316–329.
Barnhart, C., Hane, C.A., Vance, P.H. (2000). Using branch-and-price-and-cut to solve origin–
destination integer multicommodity flow problems. Operations Research 48 (2), 318–326.
Bish, E.K. (2003). A multiple-crane-constrained scheduling problem in a container terminal. European
Journal of Operational Research 144 (1), 83–107.
Bish, E.K., Chen, F.Y., Leong, Y.T., Nelson, B.L., Ng, J.W.C., Simchi-Levi, D. (2005). Dispatching ve-
hicles in a mega container terminal. OR Spectrum 27 (4), 491–506.
Böse, J., Reiners, T., Steenken, D., Voß, S. (2000). Vehicle dispatching at seaport container terminals
using evolutionary algorithms. In: Proceedings of the 33rd Hawaii International Conference on System
Sciences, vol. 2. IEEE Computer Society, p. 2025.
Bostel, N., Dejax, P. (1998). Models and algorithms for container allocation problems on trains in a
rapid transshipment shunting yard. Transportation Science 32 (4), 370–379.
Bourbeau, B., Gendron, B., Crainic, T.G. (2000). Branch-and-bound parallelization strategies applied
to a depot location and container fleet management problem. Parallel Computing 26 (1), 27–46.
Braklow, J.W., Graham, W.W., Hassler, S.M., Peck, K.E., Powell, W.B. (1992). Interactive optimization
improves service and performance for Yellow Freight system. Interfaces 22 (1), 147–172.
Brown, G.G., Lawphongpanich, S., Thurman, K.P. (1994). Optimizing ship berthing. Naval Research
Logistics 41, 1–15.
Brown, G.G., Cormican, K.J., Lawphongpanich, S., Widdis, D.B. (1997). Optimizing submarine
berthing with a persistence incentive. Naval Research Logistics 44, 301–318.
Buedenbender, K., Grünert, T., Sebastian, H.-J. (2000). A hybrid tabu search/branch and bound algo-
rithm for the direct flight network design problem. Transportation Science 34 (4), 364–380.
Campbell, J.F. (1994a). Integer programming formulations of discrete hub location problems. European
Journal of Operational Research 72, 387–405.
Campbell, J.F. (1994b). A survey of network hub location problems. Studies in Locational Analysis 6,
31–49.
Campbell, J.F. (1996). Hub location and the p-hub median problem. Operations Research 44 (6), 923–
935.
Campbell, J.F., Ernst, A.T., Krishnamoorthy, M. (2002). Hub location problems. In: Drezner, Z.,
Hamacher, H. (Eds.), Facility Location: Application and Theory. Springer-Verlag, Berlin, pp. 373–
407.
Cao, B., Uebe, G. (1995). Solving transportation problems with nonlinear side constraints with tabu
search. Computers & Operations Research 22 (6), 593–603.
Cascetta, E. (2001). Transportation Systems Engineering: Theory and Methods. Kluwer Academic, Dor-
drecht, The Netherlands.
Castilho, B.D., Daganzo, C.F. (1991). Optimal pricing policies for temporary storage at ports. Trans-
portation Research Record 1313, 66–74.
Castilho, B.D., Daganzo, C.F. (1993). Handling strategies for import containers at marine terminals.
Transportation Research Part B: Methodological 27 (2), 151–166.
Chen, T. (1999). Yard operations in the container terminal – a study in the unproductive moves. Mar-
itime Policy Management 26 (1), 27–38.
 Ch. 8. Intermodal Transportation 531

Cheung, R.K., Chen, C.-Y. (1998). A two-stage stochastic network model and solution methods for the
dynamic empty container allocation problem. Transportation Science 32 (2), 142–162.
Cheung, R.K., Li, C.L., Lin, W. (2002). Interblock crane deployment in container terminals. Transporta-
tion Science 36 (1), 79–93.
Christiansen, M., Fagerholt, K., Ronen, D. (2004). Ship routing and scheduling: Status and perspectives.
Transportation Science 38 (1), 1–18.
Christiansen, M., Fagerholt, K., Nygreen, B., Ronen, D. (2007). Maritime transportation. In: Barnhart,
C., Laporte, G. (Eds.), Transportation. Handbooks in Operations Research and Management Science.
North-Holland, Amsterdam, pp. 189–284. This volume.
Cojeen, H.P., Dyke, P.V. (1976). The automatic planning and sequencing of containers for containership
loading and unloading. In: Pitkin, M., Roche, J.J., Williams, T.J. (Eds.), Ship Operation Automation,
II: Proceedings of the 2nd IFAC/IFIP Symposium. North-Holland, Amsterdam, pp. 415–423.
Cordeau, J.-F., Toth, P., Vigo, D. (1998). A survey of optimization models for train routing and schedul-
ing. Transportation Science 32 (4), 380–404.
Cordeau, J.-F., Laporte, G., Legato, P., Moccia, L. (2005). Models and tabu search heuristics for the
berth allocation problem. Transportation Science 39 (4), 526–538.
Cordeau, J.-F., Laporte, G., Savelsbergh, M.W.P., Vigo, D. (2007). Vehicle routing. In: Barnhart, C.,
Laporte, G. (Eds.), Transportation. Handbooks in Operations Research and Management Science.
North-Holland, Amsterdam, pp. 367–428. This volume.
Crainic, T.G. (1988). Rail tactical planning: Issues, models and tools. In: Bianco, L., Bella, A.L. (Eds.),
Freight Transport Planning and Logistics. Springer-Verlag, Berlin, pp. 463–509.
Crainic, T.G. (2000). Network design in freight transportation. European Journal of Operational Re-
search 122 (2), 272–288.
Crainic, T.G. (2003). Long-haul freight transportation. In: Hall, R.W. (Ed.), Handbook of Transportation
Science, 2nd edition. Kluwer Academic, Norwell, MA, pp. 451–516.
Crainic, T.G., Delorme, L. (1995). Dual-ascent procedures for multicommodity location–allocation
problems with balancing requirements. Transportation Science 27 (2), 90–101.
Crainic, T.G., Laporte, G. (1997). Planning models for freight transportation. European Journal of Op-
erational Research 97 (3), 409–438.
Crainic, T.G., Rousseau, J.-M. (1986). Multicommodity, multimode freight transportation: A general
modeling and algorithmic framework for the service network design problem. Transportation Re-
search Part B: Methodological 20, 225–242.
Crainic, T.G., Ferland, J.-A., Rousseau, J.-M. (1984). A tactical planning model for rail freight trans-
portation. Transportation Science 18 (2), 165–184.
Crainic, T.G., Dejax, P.J., Delorme, L. (1989). Models for multimode multicommodity location prob-
lems with interdepot balancing requirements. Annals of Operations Research 18, 279–302.
Crainic, T.G., Florian, M., Guélat, J., Spiess, H. (1990a). Strategic planning of freight transportation:
STAN, an interactive-graphic system. Transportation Research Record 1283, 97–124.
Crainic, T.G., Florian, M., Léal, J.-E. (1990b). A model for the strategic planning of national freight
transportation by rail. Transportation Science 24 (1), 1–24.
Crainic, T.G., Delorme, L., Dejax, P.J. (1993a). A branch-and-bound method for multicommodity loca-
tion with balancing requirements. European Journal of Operational Research 65 (3), 368–382.
Crainic, T.G., Gendreau, M., Dejax, P.J. (1993b). Dynamic stochastic models for the allocation of empty
containers. Operations Research 41 (1), 102–302.
Crainic, T.G., Gendreau, M., Soriano, P., Toulouse, M. (1993c). A tabu search procedure for multicom-
modity location/allocation with balancing requirements. Annals of Operations Research 41, 359–383.
Crainic, T.G., Florian, M., Larin, D. (1994). STAN: New developments. In: Khade, Al S., Brown, R.
(Eds.), Proceedings of the 23rd Annual Meeting of the Western Decision Sciences Institute. School of
Business Administration, California State University, Stanislaus, CA, pp. 493–498.
Crainic, T.G., Toulouse, M., Gendreau, M. (1995a). Parallel asynchronous tabu search for multicom-
modity location–allocation with balancing requirements. Annals of Operations Research 63, 277–299.
Crainic, T.G., Toulouse, M., Gendreau, M. (1995b). Synchronous tabu search parallelization strategies
for multicommodity location-allocation with balancing requirements. OR Spectrum 17 (2–3), 113–
123.
532 T.G. Crainic and K.H. Kim

Crainic, T.G., Toulouse, M., Gendreau, M. (1997). Towards a taxonomy of parallel tabu search algo-
rithms. INFORMS Journal on Computing 9 (1), 61–72.
Crainic, T.G., Dufour, G., Florian, M., Larin, D. (1999). Path analysis in STAN. In: Zopounidis, C.,
Despotis, D. (Eds.), Proceedings 5th International Conference of the Decision Sciences Institute New
Technologies Publications, Athens, Greece, pp. 2060–2064.
Crainic, T.G., Dufour, G., Florian, M., Larin, D. (2002). Path recovery/reconstruction and applications
in nonlinear multimodal multicommodity networks. In: Gendreau, M., Marcotte, P. (Eds.), Trans-
portation and Network Analysis: Current Trends: Miscellanea in Honor of Michael Florian. Kluwer
Academic, Norwell, MA, pp. 2060–2064.
Croxton, K.L., Gendron, B., Magnanti, T.L. (2003). Models and methods for merge-in-transit opera-
tions. Transportation Science 37 (1), 1–22.
Croxton, K.L., Gendron, B., Magnanti, T.L. (2006). Variable disaggregation in network flow problems
with picewise linear costs. Operation Research, in press.
Daganzo, C.F. (1989a). Crane productivity and ship delay in ports. Transportation Research Record 1251,
1–9.
Daganzo, C.F. (1989b). The crane scheduling problem. Transportation Research Part B: Methodologi-
cal 23 (3), 159–175.
Daganzo, C.F. (1990). The productivity of multipurpose seaport terminals. Transportation Science 24
(3), 205–216.
Daganzo, C.F. (2005). Logistics Systems Analysis, 4th edition. Springer-Verlag, Berlin.
Daskin, M.S. (1995). Network and Discrete Location: Models, Algorithms, and Applications. Wiley, New
York.
Daskin, M.S., Owen, S.H. (2003). Location models in transportation. In: Hall, R.W. (Ed.), Handbook
of Transportation Science, 2nd edition. Kluwer Academic, Norwell, MA, pp. 321–371.
Dejax, P.J., Crainic, T.G. (1987). A review of empty flows and fleet management models in freight
transportation. Transportation Science 21 (4), 227–247.
Delorme, L., Roy, J., Rousseau, J.-M. (1988). Motor-carrier operation planning models: A state of
the art. In: Bianco, L., Bella, A.L. (Eds.), Freight Transport Planning and Logistics. Springer-Verlag,
Berlin, pp. 510–545.
Drezner, Z. (Ed.) (1995). Facility Location: A Survey of Applications and Methods. Springer-Verlag, New
York.
Drezner, Z., Hamacher, H. (Eds.) (2002). Facility Location: Application and Theory. Springer-Verlag,
Berlin.
Dror, M. (Ed.) (2000). Arc Routing: Theory, Solutions and Applications. Kluwer Academic, Norwell,
MA.
Ebery, J., Krishnamoorthy, M., Ernst, A., Boland, N. (2000). The capacitated multiple allocation hub
location problem: Formulations and algorithms. European Journal of Operational Research 120 (3),
614–631.
Equi, L., Gallo, G., Marziale, S., Weintraub, A. (1997). A combined transportation and scheduling
problem. European Journal of Operational Research 97 (1), 94–104.
Erlenkotter, D. (1978). A dual-based procedure for uncapacitated facility location. Operations Re-
search 26 (6), 992–1009.
Ermol’ev, Y.M., Krivets, T.A., Petukhov, V.S. (1976). Planning of shipping empty seaborne containers.
Cybernetics 12, 644–646.
Ernst, A.T., Krishnamoorthy, M. (1996). Efficient algorithms for the uncapacitated single allocation
p-hub median problem. Location Science 4 (3), 139–154.
European Conference of Ministers of Transport (2001). Terminology for combined transport.
http://www.crmt.org/online/glossaries/index.htm.
Farvolden, J.M., Powell, W.B. (1991). A dynamic network model for less-than-truckload motor carrier
operations. Working Paper 90-05, Department of Industrial Engineering, University of Toronto,
Toronto, ON, Canada.
Farvolden, J.M., Powell, W.B. (1994). Subgradient methods for the service network design problem.
Transportation Science 28 (3), 256–272.
 Ch. 8. Intermodal Transportation 533

Farvolden, J.M., Powell, W.B., Lustig, I.J. (1992). A primal partitioning solution for the arc–chain for-
mulation of a multicommodity network flow problem. Operations Research 41 (4), 669–694.
Florian, M., Hearn, D. (1995). Network equilibrium models and algorithms. In: Ball, M., Magnanti,
T.L., Monma, G.L., Nemhauser, C.L. (Eds.), Network Routing. Handbooks in Operations Research
and Management Science, vol. 8. North-Holland, Amsterdam, pp. 485–550.
Fratar, T.J., Goodman, A.S., Brant Jr., A.E. (1960). Prediction of maximum practical berth occupancy.
Journal of the Waterways and Harbors Division, 69–78.
Friesz, T.L., Harker, P.T. (1985). Freight network equilibrium: A review of the state of the art. In:
Daughety, A.F. (Ed.), Analytical Studies in Transport Economics. Cambridge Univ. Press, Cam-
bridge, Chapter 7.
Friesz, T.L., Tobin, R.L., Harker, P.T. (1983). Predictive intercity freight network models. Transportation
Research Part A: Policy and Practice 17, 409–417.
Friesz, T.L., Gottfried, J.A., Morlok, E.K. (1986). A sequential shipper–carrier network model for pre-
dicting freight flows. Transportation Science 20, 80–91.
Gendron, B., Crainic, T.G. (1995). A branch-and-bound algorithm for depot location and container
fleet management. Location Science 3 (1), 39–53.
Gendron, B., Crainic, T.G. (1997). A parallel branch-and-bound algorithm for multicommodity location
with balancing requirements. Computers & Operations Research 24 (9), 829–847.
Gendron, B., Potvin, J.-Y., Soriano, P. (1999). Tabu search with exact neighbor evaluation for multi-
commodity location with balancing requirements. INFOR 37 (3), 255–270.
Gendron, B., Potvin, J.-Y., Soriano, P. (2003a). A parallel hybrid heuristic for the multicommodity
capacitated location problem with balancing requirements. Parallel Computing 29, 591–606.
Gendron, B., Potvin, J.-Y., Soriano, P. (2003b). A tabu search with slope scaling for the multicommodity
capacitated location problem with balancing requirements. Annals of Operations Research 122, 193–
217.
Golden, B.L., Assad, A.A. (Eds.) (1988). Vehicle Routing: Methods and Studies. North-Holland, Ams-
terdam.
Griffiths, J.D. (1976). Optimal handling capacity at a berth. Maritime Studies and Management 3, 163–
167.
Grünert, T., Sebastian, H.-J. (2000). Planning models for long-haul operations of postal and express
shipment companies. European Journal of Operational Research 122 (2), 289–309.
Grünert, T., Sebastian, H.-J., Thäerigen, M. (1999). The design of a letter-mail transportation network
by intelligent techniques. In: Sprague R. (Ed.), Proceedings Hawaii International Conference on Sys-
tem Sciences 32.
Grunow, M., Günther, H.-O., Lehmann, M. (2004). Dispatching multi-load AGVs in highly automated
seaport container terminals. OR Spectrum 26 (2), 211–235.
Guan, Y., Cheung, R.K. (2004). The berth allocation problem: Models and solution methods. OR Spec-
trum 26 (1), 75–92.
Guan, Y., Xiao, W.Q., Cheung, R.K., Li, C.L. (2002). A multiprocessor task scheduling model for berth
allocation: Heuristic and worst-case analysis. Operations Research Letters 30 (5), 343–350.
Guélat, J., Florian, M., Crainic, T.G. (1990). A multimode multiproduct network assignment model for
strategic planning of freight flows. Transportation Science 24 (1), 25–39.
Günther, H.-O., Kim, K.H. (Eds.) (2005). Container Terminals and Automated Transport Systems.
Springer-Verlag, Berlin.
Harker, P.T. (1987). Predicting Intercity Freight Flows. VNU Science Press, Utrech, The Netherlands.
Harker, P.T. (1988). Issues and models for planning and regulating freight transportation systems. In:
Bianco, L., Bella, A.L. (Eds.), Freight Transport Planning and Logistics. Springer-Verlag, Berlin,
pp. 374–408.
Harker, P.T., Friesz, T.L. (1986a). Prediction of intercity freight flows I: Theory. Transportation Research
Part B: Methodological 20 (2), 139–153.
Harker, P.T., Friesz, T.L. (1986b). Prediction of intercity freight flows II: Mathematical formulations.
Transportation Research Part B: Methodological 20 (2), 155–174.
Holguin-Veras, J., Jara-Diaz, S. (1999). Optimal pricing for priority service and space allocation in
container ports. Transportation Research Part B: Methodological 33, 81–106.
534 T.G. Crainic and K.H. Kim

Hurley, W.J., Petersen, E.R. (1994). Nonlinear tariffs and freight network equilibrium. Transportation
Science 28 (3), 236–245.
Imai, A., Nagaiwa, K., Tat, C.W. (1997). Efficient planning of berth allocation for container terminals
in Asia. Journal of Advanced Transportation 31 (1), 75–94.
Imai, A., Nishimura, E., Papadimitriou, S. (2001). The dynamic berth allocation problem for a container
port. Transportation Research Part B: Methodological 35B (4), 401–417.
Imai, A., Nishimura, E., Papadimitriou, S. (2003). Berth allocation with service priority. Transportation
Research Part B: Methodological 37 (5), 437–457.
Isard, W. (1951). Interregional and regional input–output analysis: A model of a space-economy. The
Review of Economics and Statistics 33, 318–328.
ISL (2006). ISL shipping statistics and market review (SSMR) – short comment (6) 2006. Available at
http://www.isl.org.
Jaillet, P., Song, G., Yu, G. (1996). Airline network design and hub location problems. Location Sci-
ence 4 (3), 195–212.
Jones, P.S., Sharp, G.P. (1977). Multi-mode intercity freight transportation planning for underdeveloped
regions. In: Proceedings of the 18th Meeting of the Transportation Research Forum, pp. 523–531.
Jordan, W.C., Turnquist, M.A. (1983). A stochastic dynamic network model for railroad car distribution.
Transportation Science 17, 123–145.
Jourquin, B., Beuthe, M. (1996). Transportation policy analysis with a geographic information system:
The virtual network of freight transportation in Europe. Transportation Research Part C: Emerging
Technologies 4 (6), 359–371.
Kim, D., Barnhart, C., Ware, K., Reinhardt, G. (1999). Multimodal express package delivery: A service
network design application. Transportation Science 33 (4), 391–407.
Kim, K.H. (1997). Evaluation of the number of rehandles in container yards. Computers & Industrial
Engineering 32 (4), 701–711.
Kim, K.H., Bae, J.W. (1998). Remarshaling export containers in port container terminals. Computers &
Industrial Engineering 35 (3–4), 655–658.
Kim, K.H., Bae, J.W. (1999). A dispatching method for automated guided vehicles to minimize delay of
containership operations. International Journal of Management Science 5 (1), 1–25.
Kim, K.H., Bae, J.W. (2004). A look-ahead dispatching method for automated guided vehicles in auto-
mated port container terminals. Transportation Science 38 (2), 224–234.
Kim, K.H., Kim, H.B. (1999a). Segregating space allocation models for container inventories in port
container terminals. International Journal of Production Economics 59, 415–423.
Kim, K.H., Kim, H.B. (2002). The optimal sizing of the storage space and handling facilities for import
containers. Transportation Research Part B: Methodological 36, 821–835.
Kim, K.H., Kim, K.Y. (1999b). An optimal routing algorithm for a transfer crane in port container
terminals. Transportation Science 33 (1), 17–33.
Kim, K.H., Kim, K.Y. (1999c). Routing straddle carriers for the loading operation of containers using
a beam search algorithm. Computers & Industrial Engineering 36 (1), 109–136.
Kim, K.H., Moon, K.C. (2003). Berth scheduling by simulated annealing. Transportation Research Part
B: Methodological 37 (6), 541–560.
Kim, K.H., Park, K.T. (2003a). A note on a dynamic space–allocation method for outbound containers.
European Journal of Operational Research 148 (1), 92–101.
Kim, K.H., Park, K.T. (2003b). Dynamic space allocation for temporary storage. International Journal
of Systems Science 34 (1), 11–20.
Kim, K.H., Park, Y.M. (2004). A crane scheduling method for port container terminals. European Jour-
nal of Operational Research 156 (3), 752–768.
Kim, K.H., Park, Y.M., Ryu, K.R. (2000). Deriving decision rules to locate export containers in con-
tainer yard. European Journal of Operational Research 124 (1), 89–101.
Kim, K.H., Lee, K.M., Hwang, H. (2003). Sequencing delivery and receiving operations for yard cranes
in port container terminals. International Journal of Production Economics 84 (3), 283–292.
Kim, K.H., Kang, J.S., Ryu, K.R. (2004). A beam search algorithm for the load sequencing of outbound
containers in port container terminals. OR Spectrum 26 (1), 93–116.
 Ch. 8. Intermodal Transportation 535

Kim, K.Y., Kim, K.H. (1999d). A routing algorithm for a single straddle carrier to load export contain-
ers onto a containership. International Journal of Production Economics 59, 425–433.
Kim, K.Y., Kim, K.H. (2003). Heuristic algorithms for routing yard-side equipment for minimizing
loading times in container terminals. Naval Research Logistics 50 (5), 498–514.
Klincewicz, J.G. (1990). Solving a freight transport problem using facility location techniques. Opera-
tions Research 38 (1), 99–109.
Klincewicz, J.G. (1991). Heuristics for the p-hub location problem. European Journal of Operational
Research 53 (1), 25–37.
Klincewicz, J.G. (1992). Avoiding local optima in the p-hub location problem using tabu search and
GRASP. Annals of Operations Research 40, 283–302.
Klincewicz, J.G. (1996). Dual algorithm for the uncapacitated hub location problem. Location Science 4
(3), 173–184.
Koh, Y.K., Kim, S.C. (2001). A study on the impact of extra large-sized containership on the shipping
company & port. Korean International Commerce 16 (2), 165–187.
Kozan, E. (1997). Comparison of analytical and simulation planning models of sea port container ter-
minals. Transportation Planning and Technology 20, 235–248.
Kozan, E. (2000). Optimizing container transfers at multimodal terminals. Mathematical and Computer
Modeling 31, 235–243.
Kozan, E., Preston, P. (1999). Genetic algorithms to schedule container transfers at multimodal termi-
nals. International Transactions in Operational Research 6, 311–329.
Kuby, M.J., Gray, R.G. (1993). The hub network design problem with stopovers and feeders: The case
of Federal Express. Transportation Research Part A: Policy and Practice 27 (1), 1–12.
Labbé, M., Louveaux, F.V. (1997). Location problems. In: Dell’Amico, M., Maffioli, F., Martello, S.
(Eds.), Annotated Bibliographies in Combinatorial Optimization. Wiley, New York, pp. 261–281.
Labbé, M., Peeters, D., Thisse, J.-F. (1995). Location on networks. In: Ball, M., Magnanti, T.L., Monma,
G.L., Nemhauser, C.L. (Eds.), Network Routing. Handbooks in Operations Research and Management
Science, vol. 8. North-Holland, Amsterdam, pp. 551–624.
Lai, K.K., Lam, K. (1994). A study of container yard equipment allocation strategy in Hong Kong.
International Journal of Modeling & Simulation 14 (3), 134–138.
Lai, K.K., Leung, J.W. (1996). Analysis of yard crane deployment strategies in a container terminal. In:
Proceedings of ICC & IE Conference, Kyungju, Korea, pp. 1187–1190.
Lai, K.K., Leung, J. (2000). Analysis of gate house operations in a container terminal. International
Journal of Modelling and Simulation 20 (1), 89–94.
Lai, K.K., Shih, K. (1992). A study of container berth allocation. Journal of Advanced Transportation 26
(1), 45–60.
Larin, D., Crainic, T.G., Simonka, G., James-Lefebvre, L., Dufour, G., Florian, M. (2000). STAN user’s
manual, Release 6. INRO Consultants, Inc., Montréal, QC, Canada.
Li, C.-L., Cai, X., Lee, C.-Y. (1998). Scheduling with multiple-job-on-one-processor pattern. IEE Trans-
actions on Scheduling and Logistics 30 (5), 433–445.
Lim, A. (1998). The berth planning problem. Operations Research Letters 22 (2–3), 105–110.
Macharis, C., Bontekoning, Y.M. (2004). Opportunities for OR in intermodal freight transport re-
search: A review. European Journal of Operational Research 153 (2), 400–416.
Magnanti, T.L., Wong, R.T. (1984). Network design and transportation planning: Models and algo-
rithms. Transportation Science 18 (1), 1–55.
Mattfeld, D.C., Kopfer, H. (2004). Terminal operations management in vehicle transshipment. Trans-
portation Research Part A: Policy and Practice 37 (5), 435–452.
Miller, A.J. (1971). Queuing at single-berth shipping terminal. Journal of Waterways, Harbors and
Coastal Engineering Division 97, 43–56.
Minoux, M. (1989). Network synthesis and optimum network design problems: Models, solution meth-
ods and applications. Networks 19, 313–360.
Mirchandani, P.S., Francis, R.L. (Eds.) (1990). Discrete Location Theory. Wiley, New York.
Nagurney, A. (1993). Network Economics: A Variational Inequality Approach. Kluwer Academic, Nor-
well, MA.
536 T.G. Crainic and K.H. Kim

Nam, K.-C., Kwak, K.-S., Yu, M.-S. (2002). Simulation study of container terminal performance. Journal
of Waterway, Port, Coastal and Ocean Engineering 128 (3), 126–132.
Narasimhan, A., Palekar, U.S. (2002). Analysis and algorithms for the transtainer routing problem in
container port operations. Transportation Science 36 (1), 63–78.
Nemhauser, G.L., Wolsey, L.A. (1993). Integer and Combinatorial Optimization. Wiley, New York.
Noritake, M., Kimura, S. (1990). Optimum allocation and size of seaports. Journal of Waterway, Port,
Coastal, and Ocean Engineering 116 (2), 287–299.
O’Kelly, M.E. (1987). A quadratic integer program for the location of interacting hub facilities. Euro-
pean Journal of Operational Research 32 (3), 393–404.
O’Kelly, M.E. (1992a). A clustering approach to the planar hub location problem. Annals of Operations
Research 40, 339–353.
O’Kelly, M.E. (1992b). Hub facilities location with fixed costs. Papers in Regional Science 71, 292–306.
O’Kelly, M.E., Skorin-Kapov, D., Skorin-Kapov, J. (1995). Lower bounds for the hub location problem.
Management Science 41 (4), 713–721.
O’Kelly, M.E., Bryan, D., Skorin-Kapov, D., Skorin-Kapov, J. (1996). Hub network design with single
and multiple allocation: A computational study. Location Science 4 (3), 125–138.
Park, K.T., Kim, K.H. (2002). Berth scheduling for container terminals by using a subgradient optimiza-
tion technique. Journal of the Operational Research Society 53 (9), 1049–1054.
Park, Y.M. (2003). Berth and crane scheduling of container terminals. Department of Industrial Engi-
neering, Pusan National University, Pusan, Korea.
Park, Y.M., Kim, K.H. (2003). A scheduling method for berth and quay cranes. OR Spectrum 25 (1),
1–23.
Pedersen, M.B., Crainic, T.G., Madsen, O.B.G. (2006). Models and tabu search metaheuristics for
service network design with vehicle balance requirements. Publication CRT-2006-22, Centre de
recherche sur les transports, Université de Montréal, Montréal, QC, Canada.
Peterkofsky, R.I., Daganzo, C.F. (1990). A branch and bound solution method for the crane scheduling
problem. Transportation Research Part B: Methodological 24 (3), 159–172.
Pirkul, H., Jayaraman, V. (1996). Production, transportation, and distribution planning in a multi-
commodity tri-echelon system. Transportation Science 30 (4), 291–302.
Pirkul, H., Jayaraman, V. (1998). A multi-commodity, multi-plant, capacitated facility location problem:
Formulation and efficient heuristic solution. Computers & Operations Research 25 (10), 869–878.
Powell, W.B. (1986). A local improvement heuristic for the design of less-than-truckload motor carrier
networks. Transportation Science 20 (4), 246–357.
Powell, W.B. (1988). A comparative review of alternative algorithms for the dynamic vehicle allocation
problem. In: Golden, B.L., Assad, A.A. (Eds.), Vehicle Routing: Methods and Studies. North-Holland,
Amsterdam, pp. 249–292.
Powell, W.B. (2003). Dynamic models of transportation operations. In: Graves, S., Tok, T.A.G. (Eds.),
Supply Chain Management. Handbooks in Operations Research and Management Science, vol. 11.
North-Holland, Amsterdam, pp. 677–756.
Powell, W.B., Carvalho, T.A. (1998). Real-time optimization of containers and flatcars for intermodal
operations. Transportation Science 32 (2), 110–126.
Powell, W.B., Sheffi, Y. (1983). The load-planning problem of motor carriers: Problem description and
a proposed solution approach. Transportation Research Part A: Policy and Practice 17 (6), 471–480.
Powell, W.B., Sheffi, Y. (1986). Interactive optimization for motor carrier load planning. Journal of
Business Logistics 7 (2), 64–90.
Powell, W.B., Sheffi, Y. (1989). Design and implementation of an interactive optimization system for
the network design in the motor carrier industry. Operations Research 37 (1), 12–29.
Powell, W.B., Topaloglu, H. (2003). Stochastic programming in transportation and logistics. In:
Ruszczynski, A., Shapiro, A. (Eds.), Stochastic Programming. Handbooks in Operations Research and
Management Science, vol. 10. North-Holland, Amsterdam, pp. 555–635.
Powell, W.B., Topaloglu, H. (2005). Fleet management. In: Wallace, S., Ziemba, W. (Eds.), Applica-
tions of Stochastic Programming. Math Programming Society–SIAM Series on Optimization. SIAM,
Philadelphia, PA, pp. 185–216.
 Ch. 8. Intermodal Transportation 537

Powell, W.B., Jaillet, P., Odoni, A. (1995). Stochastic and dynamic networks and routing. In: Ball, M.,
Magnanti, T.L., Monma, G.L., Nemhauser, C.L. (Eds.), Network Routing. Handbooks in Operations
Research and Management Science, vol. 8. North-Holland, Amsterdam, pp. 141–295.
Powell, W.B., Bouzaïene-Ayari, B., Simaõ, H.P. (2007). Dynamic models for freight transportation. In:
Barnhart, C., Laporte, G. (Eds.), Transportation. Handbooks in Operations Research and Manage-
ment Science. North-Holland, Amsterdam, pp. 285–365. This volume.
Preston, P., Kozan, E. (2001). An approach to determine storage locations of containers at seaport
terminals. Computers & Industrial Engineering 28, 983–995.
Ramani, K.V. (1996). An interactive simulation model for the logistics planning of container operations
in seaports. Simulation 66 (5), 291–300.
Roy, J., Crainic, T.G. (1992). Improving intercity freight routing with a tactical planning model. Inter-
faces 22 (3), 31–44.
Roy, J., Delorme, L. (1989). NETPLAN: A network optimization model for tactical planning in the
less-than-truckload motor-carrier industry. INFOR 27 (1), 22–35.
Ryu, K.R., Kim, K.H., Lee, Y.H., Park, Y.M. (2001). Load sequencing algorithms for container ships
by using metaheuristics. In: Proceedings of 16th International Conference on Production Research.
Prague, Czech Republic, 29 July–4 August, CD-ROM.
Salkin, H.M., Mathur, K. (1989). Foundations of Integer Programming. North-Holland, Amsterdam.
Schonfeld, P., Sharafeldien, O. (1985). Optimal berth and crane combinations. Journal of Waterway,
Port, Coastal and Ocean Engineering 111 (6), 1060–1072.
Sharp, G.P. (1979). A multi-commodity, intermodal transportation model. In: Proceedings of the 20th
Meeting of the Transportation Research Forum, pp. 399–407.
Skorin-Kapov, D., Skorin-Kapov, J. (1994). On tabu search for the location of interacting hub facilities.
European Journal of Operational Research 73 (3), 502–509.
Skorin-Kapov, D., Skorin-Kapov, J., O’Kelly, M.E. (1996). Tight linear programming relaxation of un-
capacitated p-hub median problems. European Journal of Operational Research 94 (3), 582–593.
Smith, K., Krishnamoorthy, M., Palaniswami, M. (1996). Neural versus traditional approaches to the
location of interacting hub facilities. Location Science 4 (3), 155–171.
Steenken, D., Voß, S., Stahlbock, R. (2004). Container terminal operation and operations research –
a classification and literature review. OR Spectrum 26 (1), 3–49.
Taleb-Ibrahimi, M., Castilho, B., Daganzo, C.F. (1993). Storage space vs handling work in container
terminals. Transportation Research Part B: Methodological 27 (1), 13–32.
Toth, P., Vigo, D. (Eds.) (2002). The Vehicle Routing Problem. SIAM Monographs on Discrete Mathemat-
ics and Applications. SIAM, Philadelphia, PA.
van der Heyden, W.P.A., Ottjes, J.A. (1985). A decision support system for the planning of the workload
of a grain terminal. Decision Support Systems 1, 293–297.
van Hee, K.M., Wijbrands, R.J. (1988). Decision support system for container terminal planning. Eu-
ropean Journal of Operational Research 34 (3), 262–272.
Vis, I.F.A., de Koster, R., Roodbergen, K.J. (2001). Determination of the number of automated guided
vehicles required at a semi-automated container terminal. Journal of the Operational Research Soci-
ety 52 (4), 409–417.
Wanhill, S.R.C. (1974). Further analysis of optimum size seaport. Journal of the Waterways Harbors and
Coastal Engineering Division 100, 377–383.
White, W. (1972). Dynamic transshipment networks: An algorithm and its application to the distribution
of empty containers. Networks 2 (3), 211–236.
Winston, C. (1983). The demand for freight transportation: Models and applications. Transportation
Research Part A: Policy and Practice 17, 419–427.
Zhang, C., Wan, Y.W., Liu, J., Linn, R.J. (2002). Dynamic crane deployment in container storage yards.
Transportation Research Part B: Methodological 36 (6), 537–555.
Zhang, C., Liu, J., Wan, Y.-W., Murty, K.G., Linn, R. (2003). Storage space allocation in container
terminals. Transportation Research Part B: Methodological 37 (10), 883–903.
C. Barnhart and G. Laporte (Eds.), Handbook in OR & MS, Vol. 14
Copyright © 2007 Elsevier B.V. All rights reserved
DOI: 10.1016/S0927-0507(06)14009-8

Chapter 9
Hazardous Materials Transportation
Erhan Erkut
Faculty of Business Administration, Bilkent University, Ankara, Turkey
E-mail: [email protected]

Stevanus A. Tjandra
University of Alberta School of Business, Edmonton, Canada
E-mail: [email protected]

Vedat Verter
Desautels Faculty of Management, McGill University, Montreal, Canada
E-mail: [email protected]

1 Introduction

The transportation of hazardous materials (or dangerous goods) deserves

to be treated in a separate chapter of this volume, primarily due to the risks as-
sociated with this activity. Although the industry has an excellent safety record,
accidents do happen, and the consequences can be significant, due to the na-
ture of the cargo. Reduction of hazardous material (hazmat) transportation
risks can be achieved in many different ways. Some of these risk reduction
measures, such as driver training and regular vehicle maintenance, have little
connection to operations research (OR), whereas others offer interesting chal-
lenges to OR. This chapter focuses on applications of OR models to hazmat
transportation, providing a relatively comprehensive review of the literature,
and outlining areas of potential impact for operations researchers.
According to the US Department of Transportation (US DOT), a hazardous
material is defined as any substance or material capable of causing harm to
people, property, and the environment. Dependence on hazardous materials
is a fact of life in industrialized societies. There are thousands of different haz-
ardous materials in use today (US DOT, 2004b). The United Nations sorts
hazardous materials into nine classes according to their physical, chemical, and
nuclear properties: explosives and pyrotechnics; gasses; flammable and com-
bustible liquids; flammable, combustible, and dangerous-when-wet solids; ox-
idizers and organic peroxides; poisonous and infectious materials; radioactive
materials; corrosive materials (acidic or basic); and miscellaneous dangerous
goods, such as hazardous wastes (UN, 2001). In almost all instances, hazmats
originate at a location other than their destination. For example, oil is extracted
from oil fields and shipped to a refinery (typically via a pipeline); many oil
products, such as heating oil and gasoline, are refined at the refinery and then

539
540 E. Erkut, S.A. Tjandra and V. Verter

shipped to storage tanks at different locations within a country. As another

example, polychlorinated biphenyls (PCBs) are collected at many industrial
installations, such as old power generation and transfer stations and shipped
to a special waste management facility for safe disposal (usually incineration).
Hence, transportation plays a significant role for hazmats. The magnitude of
this role depends on the size of a country and its level of industrialization.
For example, the Office of Hazardous Materials Safety (OHMS) of the US
DOT estimated that there were 800,000 domestic shipments of hazmats, to-
taling approximately 9 million tons, in the USA each day in 1998 (US DOT,
2000). Transport Canada estimates that nearly 80,000 shipments of dangerous
goods are moved by road, rail, water, and air in Canada (Transport Canada,
2004). Given a conservative estimate of 2% annual growth in the production
of hazmats, it is safe to assume that the total number of shipments in North
America is well over the one million mark in 2005.
In 2002, over 99 percent of hazmat shipments in Canada made it safely to
their destination (Transport Canada, 2004). While the hazmat transport sector
is far safer than other transport sectors (US DOT, 2000), hazmat transport
accidents do happen. Figure 1 shows the distribution of accidents/incidents
by hazmat class in 2003. An accident resulting in a release of the hazmat is
called an incident. The figure shows that flammable–combustible liquids and
corrosive materials accounted for the majority of hazmat accidents/incidents
in the USA (US DOT, 2004a).
The transportation of hazmats can be classified according to the mode of
transport, namely: road, rail, water, air, and pipeline. Some shipments are in-
termodal; they are switched from one mode to another during transit. There

Fig. 1. Accident/incident by hazmat class in 2003 (US DOT, 2004a).

 Ch. 9. Hazardous Materials Transportation 541

are significant differences in the use of these modes. While transportation by

truck accounts for approximately 94% of all individual hazmat shipments in the
USA, this mode carries merely 43% of the hazmat tonnage since the volume
that can be shipped by one truck is limited compared to other modes of trans-
port. In contrast, rail, water, and pipelines carry 57% of the hazmat tonnage
while accounting for less than 1% of all individual shipments. It is possible
to carry huge quantities of hazmats using these modes. While the counting
of individual shipments is less clear with these modes (How do we count the
number of shipments via a pipeline? Does a train consisting of multiple hazmat
tank cars count as a single shipment?), they carry much larger quantities per
shipment than trucks do. The balance of hazmat shipments (5% by count and
0.05% by weight) are made via air (US DOT, 1998).
Hazmat transport incidents can occur at the origin or destination (when
loading and unloading) or en-route. Incidents involving hazmat cargo can lead
to severe consequences characterized by fatalities, injuries, evacuation, prop-
erty damage, environmental degradation, and traffic disruption. In 2003, there
were 488 serious incidents (among a total of 15,178 incidents) resulting in
15 deaths, 17 major and 18 minor injuries, and a total property damage of
$37.75 million (US DOT, 2004c). About 90% of hazmat incidents occur on
highways. As far as causes go, human error seems to be the single greatest
factor (see Figure 2) in all hazardous materials incidents (minor and serious
incidents).
The annual number of nonhazmat transportation accidents in the USA is
estimated to be 126,880, in contrast to the approximately 15,000 hazmat trans-

Fig. 2. Serious incident by mode/cause in 2003 (US DOT, 2004c).

542 E. Erkut, S.A. Tjandra and V. Verter

portation accidents and incidents (FMCSA, 2001). Even though hazmats are
involved in a small minority of all transport accidents, hazmat accidents can
have catastrophic consequences. In 2003, for example, 22 train cars derailed at
Tamoroa, IL, resulting in the release of various types and quantities of haz-
ardous materials from seven tank cars. The evacuation of over a thousand
residents within a three-mile radius and the closing of Highway 51 followed
the derailment.
Table 1 contrasts the average costs (per event) of hazmat and nonhazmat
motor carrier accidents and incidents for one year. Although the cost of an
average hazmat incident is not significantly higher than the cost of a non-
hazmat incident, the cost of a hazmat incident resulting in fire or explosion
is significantly higher. Hazmat transportation accidents are perceived as low
probability–high consequence (LPHC) events and data seem to support this
perception. For example, chlorine leaking from damaged tank cars due to a de-
railment in Mississauga, Ontario in 1979, forced the evacuation of 200,000 peo-
ple. In 1982, a gasoline truck explosion in a tunnel in Afghanistan caused 2700
fatalities. Most transport accidents that impact a large number of people and
result in significant economic loss involve a hazmat cargo.
Hazmat transportation involves multiple players such as shippers, carri-
ers, packaging manufacturers, freight forwarders, consignees, insurers, govern-
ments, and emergency responders; each has a different role in safely moving
hazardous materials from their origins to their destinations. There are often
multiple handoffs of material from one party to another during transport. The
various parties, ranging from individuals or small firms to large multinational
organizations, may have overlapping and unclear responsibilities for managing
the risks (ICF Consulting, 2000). Furthermore, each party may have differ-
ent priorities and viewpoints. Although the transportation department or local
government is responsible for designating allowable routes that reduce risk,
a carrier company would, in general, try to identify the route that minimizes
the fuel costs and travel times, between the origin and destination for each
shipment. Some routes have short lengths but move through heavily populated
areas; some routes avoid heavily populated areas but are longer, resulting in
higher transport costs and accident probabilities; while other routes use major

Table 1.
Comparative costs of hazmat and nonhazmat motor carrier accident/incident events (FMCSA, 2001)

Type of accident/incident event Average cost Average traffic

(in US$) delay (in hours)

Nonhazmat events 340000 2

All hazmat events 414000 –
Hazmat events with spill/release 536000 5
Hazmat events with fire 1200000 8
Hazmat events with explosion 2100000 12
 Ch. 9. Hazardous Materials Transportation 543

freeways and thus minimize travel time but may be associated with higher ac-
cident rates. Thus, hazmat transportation is a typical multiobjective problem
with multiple stakeholders.
Multiobjective/multistakeholder problems are complicated to solve. Haz-
mat transport problems are further complicated by public sensitivity surround-
ing these problems. The concept of social amplification of risk (see Kasperson
et al., 1988; Renn et al., 1992) indicates that public assessment of a risk de-
pends not only on its magnitude but also on subjective perceptions. The in-
dividual and social perceptions of risk can be heightened or attenuated by
many factors such as extensive media coverage of the hazard event (see, e.g.,
Horlick-Jones, 1995), involvement of social groups (see, e.g., Moore, 1989),
inaccuracies and inconsistencies in the communication process that lead to ru-
mors and speculations on risk magnitude (see, e.g., Mileti and O’Brien, 1992;
Barnes, 2001). The amplification of the risk of a relatively minor hazmat acci-
dent may imply much stronger public reaction and results in a call for action,
such as tighter transport regulations or even the banning of hazmat shipments
via a certain mode of transport, in some extreme cases.
Public sensitivity to hazmat transport is rooted not only in public risk per-
ceptions, but also in equity concerns. Those individuals benefiting from hazmat
shipments are usually those who live near the production facility or the delivery
points. Yet the population living along a major highway connecting the hazmat
origin and destinations is exposed to the transport risks regardless of whether
or not they benefit from the hazmat shipments. This lack of burden-benefit
concordance is another source of public opposition to hazmat shipments. The
shipment of spent nuclear fuel rods from nuclear power plants to the proposed
repository at Yucca Mountain in Nevada, USA, offers a good example of eq-
uity-based public opposition. The shipping reduces the risk at the power plants.
Yet some risk is imposed on the population living along the major east–west
highways or railways, who are asked to assume the risk with no clear bene-
fits to them. Furthermore, if the same main route segment were selected for
shipments from multiple origins, the objection of people living along this route
would increase considerably. These people are likely to prefer alternate rout-
ings that would spread the risks.
Public opposition to hazmat shipments has increased in recent years, due
to fears of terrorist attacks on hazmat vehicles. The Research and Special
Projects Administration (RSPA) of US DOT accepts that hazmats could pose
a significant threat during transportation, when they are particularly vulner-
able to sabotage or misuse as weapons of mass destruction or as weapons of
convenience by terrorists – particularly given how easy it is to identify a haz-
mat vehicle (as well as the specifics of their cargo) given the current system of
hazmat placards. As a result some jurisdictions are trying to force a rerouting
of hazmat vehicles away from populated areas by implementing local laws.
Much of the discussion to this point also applies to the location of hazardous
facilities. If anything, the risks and the public opposition are higher for fixed
facilities than for transport. Operations researchers have dealt with both types
544 E. Erkut, S.A. Tjandra and V. Verter

of problems, and we will include references to facility location as well as trans-

portation problems in this chapter, particularly for facility location models that
treat the transportation component explicitly.
The rest of this chapter is organized in the following way. In Section 2, we of-
fer a high-level view of hazmat logistics literature where we summarize special
journal issues, reference books, reports, and web sites that are potentially use-
ful to an operations researcher who wishes to conduct research in this area. We
also offer a classification of journal papers, which provides the organizational
structure for the rest of the chapter. Section 3 contains a treatment of risk, the
main ingredient of hazmat logistics problems that separate them from other
logistics problems. We review different models of risk for hazmat transport
and discuss how one can go from point risk to edge risk and then to route risk.
Section 4 deals with hazmat routing and scheduling problems. In Section 5,
we turn our attention to models that combine undesirable facility location and
hazmat transportation. In the final section we offer a critique of the existing
literature and suggest directions for future research.

2 A high-level view of hazmat logistics research

2.1 Special issues of journals

Hazmat logistics has been a very active research area during the last twenty
years. In 1984 Management Science published a special issue on Risk Analysis
(Vol. 30, No. 4) where five papers dealt with hazmats and hazardous facilities.
This was followed by a number of special issues of refereed academic journals
that focus on hazmat transportation or location problems.
• Transportation Research Record published two special issues on hazmat
transportation in 1988 (No. 1193) that included four papers and 1989
(No. 1245) that included six papers.
• Transportation Science devoted an issue to hazmat logistics in 1991
(Vol. 25, No. 2) that contained six papers.
• There was a special section on hazmat transportation in the March/
April 1993 issue of the Journal of Transportation Engineering that in-
cluded four papers.
• A special double-issue of INFOR on hazardous materials logistics was
published in 1995 (Vol. 33, No. 1 and 2) with nine papers.
• Four papers were included in a special issue of Location Science dealing
with hazmats in 1995 (Vol. 3, No. 3).
• Transportation Science produced a second special issue on hazmat lo-
gistics in 1997 (Vol. 31, No. 3) with seven papers.
• Studies in Locational Analysis published a special issue on undesirable
facility location in April 1999 (Issue 12) that contained seven papers.
 Ch. 9. Hazardous Materials Transportation 545

• Computers & Operations Research will publish a hazmat logistics special

issue in 2007 which will contain results of the most recent research in
the area in 13 papers.
These special issues contain many useful papers and they offer a good start-
ing point for research in this area. Likewise, the book chapter by Erkut and
Verter (1995a) offers a relatively comprehensive survey of the literature up to
1994, and the annotated bibliography by Verter and Erkut (1995) offers a good
list of pre-1994 references in risk assessment, location, and routing.

2.2 Books

Perhaps an even better starting point for those who wish to familiarize them-
selves with the terminology and the problem context are the following books.
• Transportation of Hazardous Materials: Issues in Law, Social Science,
and Engineering (1993), edited by L.N. Moses and D. Lindstrom,
Kluwer Academic Publishers. This book contains 18 articles presented
at Hazmat Transport ‘91, a national conference held at Northwestern
University on all aspects of hazmat transport. While only a few of
the articles use OR models and techniques, the book offers a multi-
dimensional treatment of the subject and it is good reading for new
researchers in the area.
• Three books were produced by Institute for Risk Research, University
of Waterloo, as a result of the First International Consensus Confer-
ence on the Risks of Transporting Dangerous Goods, held in Toronto,
Canada in April, 1992:
• Transportation of Dangerous Goods: Assessing the Risks (1993), edited
by F.F. Saccomanno and K. Cassidy. This book contains 30 articles
which are organized into five main chapters: Application of QRA
models to the transport of Dangerous Goods; Analysis of Dangerous
Goods Accident and Releases; Application of Simple Risk Assess-
ment Methodology; Uncertainty in Risk Estimation; Risk Tolerance,
Communication and Policy Implications.
• Comparative Assessment of Risk Model Estimates for the Transport
of Dangerous Goods by Road and Rail (1993), edited by F.F. Sac-
comanno, D. Leming, and A. Stewart. This book documents the
assessment of a corridor exercise involving the application of sev-
eral risk models to a common transport problem involving the bulk
shipment of chlorine, LPG, and gasoline by road and rail along pre-
defined routes. The purpose of the corridor exercise was to provide
a well-defined transportation problem for analysis in order to exam-
ine the sources of variability in the risk estimates. Seven agencies in
six countries participated in this exercise.
• What is the Risk (1993), edited by F.F. Saccomanno, D. Leming, and
A. Stewart. This book documents the small group discussions and
546 E. Erkut, S.A. Tjandra and V. Verter

consensus testing process from the corridor exercise conducted as

part of the international consensus conference.
• Hazardous Materials Transportation Risk Analysis: Quantitative Ap-
proaches for Truck and Train (1994), Rhyne WR, Van Norstrand Rein-
hold. This book explains the quantitative risk analysis (QRA) method-
ologies and their application to hazmat transportation. It also provides
an extended example of a QRA for bulk transport of chlorine by truck
and train. This detailed example explores every step of the QRA from
preliminary hazards analysis to risk reduction alternatives. This book is
a valuable reference for hazmat transportation risks, and it is intended
for practitioners. It is not an OR book, but it provides useful informa-
tion for OR research in hazmat transportation modeling and analysis.
• Guidelines for Chemical Transportation Risk Analysis (1995), American
Institute of Chemical Engineers, Center for Chemical Process Safety
(CCPS), New York. This book completes two other books in the se-
ries of process safety guidelines books produced by CCPS: Guide-
lines for Chemical Process Quantitative Risk Analysis (CPQRA, 1989)
and Guidelines for Hazard Evaluation Procedures (HEP, 2nd edition,
1992). It is intended to be used as a companion volume to the CPQRA
and HEP Guidelines when dealing with a quantitative transportation
risk analysis (TRA) methodology. This book offers a basic approach to
TRA for different transport modes (pipelines, rail, road, barge, water,
and intermodal containers). It can be useful to an engineer or man-
ager in identifying cost effective ways to manage and reduce the risk of
a hazmat transportation operation.
• Quantitative Risk Assessment of Hazardous Materials Transport Systems
(1996), M. Nicolet-Monnier and A.V. Gheorge, Kluwer Academic Pub-
lishers. This book contains a comprehensive treatment of the analysis
and assessment of transport risks due to hazmat transport on roads,
rail, by ship, and pipeline. It contains European case studies as well
as a discussion of computer-based DSS (Decision Support System) for
hazmat transport problems. It is a useful reference book in the area.

2.3 Reports

In addition to these books, there are also a number of recent government

reports that contain a wealth of useful information for researchers in OR as
well as other relevant fields:
• AND-DIS-01-1 A National Risk Assessment for Selected Hazardous Ma-
terials in Transportation (2000), Argonne National Laboratory, Depart-
ment of Energy.
• ANL-DIS-00-1: Development of the Table of Initial Isolation and Pro-
tective Action Distances for the 2000 Emergency Response Guidebook
(2000), Argonne National Laboratory, Department of Energy.
 Ch. 9. Hazardous Materials Transportation 547

• Comparative Risks of Hazardous Materials and Non-Hazardous Materi-

als Truck Shipment Accidents/Incident (2001), Batelle.
• A Resource Handbook on DOE Transportation Risk Assessment (2002),
DOE Transportation Risk Assessment Working Group Technical Sub-
committee.
(Note: All URLs in this chapter were functional as of May 2005.)

2.4 Web sites

The following web sites contain useful information for practitioners as well
as researchers on hazmat transport:
• The Office of Hazardous Materials Safety (US DOT Research and
Special Programs Administration): http://hazmat.dot.gov/.
• The Hazmat 101 Web: http://www.hazmat101.com/.
• Hazmat Magazine: http://www.hazmatmag.com/.
• On-line hazmat school: http://www.hazmatschool.com/.
• National Hazardous Materials Route Registry: http://hazmat.fmcsa.dot.
gov/.
• United Nations Economic Commission for Europe (UNECE) – Dan-
gerous Goods and Special Cargo: http://www.unece.org/trans/danger/
danger.htm.
• The Canadian Transport Emergency Centre (CANUTEC) of the De-
partment of Transport: http://www.tc.gc.ca/canutec/.
• A mailing list for those interested in hazmat transport: http://groups.
yahoo.com/group/DangerousGoods/.

2.5 Software

There exists some software which has been developed to aid the analysts
or decision makers in dealing with hazmat logistics. For example, ALOHA
(Areal Locations of Hazardous Atmospheres) predicts how a hazardous gas
cloud might disperse in the atmosphere after an accidental chemical release.
This software (see US EPA, 2004) has been developed jointly by the Na-
tional Oceanic and Atmospheric Administration’s (NOAA) Hazardous Mate-
rials Response and Assessment Division and the US Environmental Protection
Agency’s (EPA) Chemical Emergency Preparedness and Prevention Office.
ALOHA can be useful for transport risk assessment. However, this software
is more useful for fixed facility risk assessment than for route selection.
In contrast to the availability of many software packages for regular truck
routing, we know of only one off-the-shelf hazmat routing package that is
currently available: PC*Miler|HazMat (ALK Associates, 1994). It has fea-
tures that allow transportation and logistics companies to determine routes
and mileages for hauling hazardous materials while ensuring compliance with
548 E. Erkut, S.A. Tjandra and V. Verter

government regulations. Routes can be generated for general, explosive, in-

halant and radioactive hazmats. This software contains all of the features and
functionality of PC*Miler, a routing, mileage and mapping software, which is
also developed by ALK. Here we note that HazTrans (Abkowitz et al., 1992)
and PC*HazRoute (ALK Associates, 1994) were marketed in the last decade,
but both are off the market as of 2005.

2.6 Classification

While we offer references to books, reports, and web sites in this section,
the rest of this chapter deals mainly with the academic literature consisting of
refereed journal articles. Figure 3 displays the number of papers published in
this area between 1982 and 2004. It seems that this area of research has peaked
in mid-1990s and has declined somewhat since.
Given the large number of papers in this area, we believe a simple classifi-
cation can be useful in providing some structure to the rest of the chapter. The
articles in this area deal with different aspects of the problem. One possible
classification is the following (in no particular order):
(1) risk assessment,
(2) routing,
(3) combined facility location and routing,
(4) network design.
Although we have offered this simple classification, it is fair to say that nu-
merous papers deal with problems that lie at the intersection of the above areas
and such problems are receiving increasingly more attention in the literature.

Fig. 3. Number of hazmat-transportation related papers published in refereed journals between 1982
and 2004.
 Ch. 9. Hazardous Materials Transportation 549

Tables 2(a–d) provides a classification of papers using the above four prob-
lem classes as well as other important attributes such as transport mode, par-
adigm (deterministic vs. stochastic) and number of objectives, and whether or
not the paper uses GIS (Geographic Information System) or proposes a DSS.
The rest of this chapter provides a comprehensive literature survey follow-
ing the problem classification presented above, and points out directions for
future research.

Table 2a.
A classification of hazmat transportation models – risk assesment

Road Jonkman et al., 2003; Nardini et al., 2003; Martinez-Alegria et al., 2003G ;
Rosmuller and Van Gelder, 2003; Abkowitz, 2002C ; Fabiano et al., 2002;
Kimberly and Killmer, 2002; Saccomanno and Haastrup, 2002N ; Hollister,
2002; Hwang et al., 2001; Abkowitz et al., 2001; Verter and Kara, 2001G ;
Efroymson and Murphy, 2000; ICF Consulting, 2000; Leonelli et al.,
2000; Zhang et al., 2000G ; Pet-Armacost et al., 1999; Cassini, 1998; Mills
and Neuhauser, 1998; Cutter and Ji, 1997; Groothuis and Miller, 1997;
Lovett et al., 1997G ; Pine and Marx, 1997; Alp and Zelensky, 1996;
Ertugrul, 1995; Sissell, 1995; Chakraborty and Armstrong, 1995; Erkut and
Verter, 1995aU ; Erkut and Verter, 1995b; Moore et al., 1995G ; Spadoni
et al., 1995; Verter and Erkut, 1995U ; Gregory and Lichtenstein, 1994;
Macgregor et al., 1994; Hobeika and Kim, 1993; Sandquist et al., 1993;
Harwood et al., 1993; Abkowitz et al., 1992; Glickman, 1991; Grenney
et al., 1990DSS ; Kunreuther and Easterling, 1990; Chow et al., 1990;
Abkowitz and Cheng, 1989; Ang and Briscoe, 1989; Harwood et al., 1989;
Abkowitz and Cheng, 1988; Hillsman, 1988; Horman, 1987; Keeney and
Winkler, 1985; Scanlon and Cantilli, 1985; Pijawka et al., 1985; Kunreuther
et al., 1984; Philipson et al., 1983; Wilmot, 1983; Keeney, 1980; Shappert
et al., 1973
Rail Anderson and Barkan, 2004; Barkan et al., 2003; Fronczak, 2001; Orr
et al., 2001; Dennis, 1996; Larson, 1996; Glickman and Golding, 1991;
McNeil and Oh, 1991; Saccomanno and Elhage, 1991; Glickman and
Rosenfield, 1984; Glickman, 1983; Saccomanno and El-Hage, 1989
Marine Douligeris et al., 1997; Roeleven et al., 1995; Romer et al., 1995
Air LaFrance-Linden et al., 2001
Road + rail Brown and Dunn, 2007; Milazzo et al., 2002; Bubbico et al., 2000; Neill
and Neill, 2000; Deng et al., 1996; Leeming and Saccomanno, 1994; Purdy,
1993; Saccomanno and Shortreed, 1993; Saccomanno and El-Hage, 1989;
Vanaerde et al., 1989; Glickman, 1988; Swoveland, 1987
Road + rail + marine Andersson, 1994
Road + rail + marine + air Kloeber et al., 1979

C with security consideration;

DSS decision support system model;
G using GIS;
N through road tunnels;
U survey/annotated bibliography.
550 E. Erkut, S.A. Tjandra and V. Verter

Table 2b.
A classification of hazmat transportation models – routing

Local routing Road Akgün et al., 2007; Duque, 2007; Erkut and Ingolfsson, 2005;
Huang and Cheu, 2004CG ; Huang et al., 2003CM ; Kara et
al., 2003; Luedtke and White, 2002CU ; Marianov et al., 2002;
Frank et al., 2000; Erkut and Ingolfsson, 2000; Leonelli et
al., 2000; Zografos et al., 2000DSS ; Erkut and Verter, 1998;
Tayi et al., 1999M ; Bonvicini et al., 1998; Marianov and ReV-
elle, 1998M ; Verter and Erkut, 1997; Sherali et al., 1997M ;
Nembhard and White, 1997M ; Erkut and Glickman, 1997;
Jin and Batta, 1997; Verter and Erkut, 1997; Erkut, 1996;
Jin et al., 1996; Ashtakala and Eno, 1996S ; Beroggi and Wal-
lace, 1995; Boffey and Karkazis, 1995; Erkut, 1995; Moore et
al., 1995G ; Karkazis and Boffey, 1995; Glickman and Son-
tag, 1995M ; McCord and Leu, 1995M ; Sivakumar et al.,
1995; Beroggi, 1994; Beroggi and Wallace, 1994; Ferrada and
Michelhaugh, 1994; Patel and Horowitz, 1994G ; Sivakumar
and Batta, 1994; Lassarre et al., 1993G ; Sivakumar et al.,
1993; Turnquist, 1993MS ; Wijeratne et al., 1993M ; Lepofsky
et al., 1993G ; Beroggi and Wallace, 1991; Miaou and Chin,
1991; Gopalan et al., 1990a; Chin, 1989M ; Zografos and Davis,
1989M ; Abkowitz and Cheng, 1988M ; Batta and Chiu, 1988;
Vansteen, 1987; Cox and Turnquist, 1986; Belardo et al., 1985;
Saccomanno and Chan, 1985; Urbanek and Barber, 1980;
Kalelkar and Brinks, 1978M
Rail Verma and Verter, 2007; McClure et al., 1988; Coleman, 1984;
Glickman, 1983
Marine Iakovou, 2001; Li et al., 1996; Haas and Kichner, 1987
Road + rail Glickman, 1988
Road + rail + marine Weigkricht and Fedra, 1995DSS
Local routing Erkut and Alp, 2006; Chang et al., 2005MST ; Zografos
and scheduling and Androutsopoulos, 2004M ; Zografos and Androutsopou-
(on road) los, 2002M ; Miller-Hooks and Mahmassani, 2000ST ; Bowler
and Mahmassani, 1998T ; (Miller-Hooks and Mahmassani,
1998)ST ; Sulijoadikusumo and Nozick, 1998MT ; (Nozick et
al., 1997)MT ; Smith, 1987M ; Cox and Turnquist, 1986
Global routing Road Carotenuto, et al. (2007a, 2007b); Dell’Olmo et al., 2005;
Akgün et al., 2000; Marianov and ReVelle, 1998; Lindner-
Dutton et al., 1991; Gopalan et al. (1990a, 1990b); Zografos
and Davis, 1989
Marine Iakovou et al., 1999

C with security consideration;

DSS decision support system model;
G using GIS;
M multiobjective;
S stochastic;
T time-varying;
U survey/annotated bibliography.
 Ch. 9. Hazardous Materials Transportation 551

Table 2c.
A classification of hazmat transportation models – combined facility location and routing

Alumur and Kara, 2007; Cappanera et al., 2004; Berman et al., 2000; Giannikos, 1998M ; Helander and
Melachrinoudis, 1997; List and Turnquist, 1998; Current and Ratick, 1995M ; Jacobs and Warmerdam,
1994; Boffey and Karkazis, 1993; Stowers and Palekar, 1993; List and Mirchandani, 1991M ; List et al.,
1991U ; ReVelle et al., 1991; Zografos and Samara, 1989; Peirce and Davidson, 1982; Shobrys, 1981

M multiobjective;
U survey/annotated bibliography.

Table 2d.
A classification of hazmat transportation models – network design

Berman et al., 2007; Erkut and Alp, 2006; Erkut and Gzara, 2005; Erkut and Ingolfsson, 2005; Verter
and Kara, 2005; Kara and Verter, 2004

3 Risk assessment

Risk is the primary ingredient that separates hazmat transportation prob-

lems from other transportation problems. In this section we will provide a de-
tailed treatment of how risk is incorporated into hazmat transport models,
starting with the basic building blocks and moving our way into risk assess-
ment along a route. In the context of hazmat transport, risk is a measure of
the probability and severity of harm to an exposed receptor due to potential
undesired events involving a hazmat (Alp, 1995). The exposed receptor can be
a person, the environment, or properties in the vicinity. The undesired event in
this context is the release of a hazmat due to a transport accident. The conse-
quence of a hazmat release can be a health effect (death, injury, or long-term
effects due to exposure), property loss, an environmental effect (such as soil
contamination or health impacts on flora and fauna), an evacuation of nearby
population in anticipation of imminent danger, or stoppage of traffic along the
impacted route.
Risk assessment can be qualitative or quantitative. Qualitative risk assess-
ment deals with the identification of possible accident scenarios and attempts
to estimate the undesirable consequences. It is usually necessitated by a lack
of reliable data to estimate accident probabilities and consequence measures.
The goal is to identify events that appear to be most likely and those with the
most severe consequences, and focus on them for further analysis. It may be
the only option in the absence of data – for example, assessing the risks due
to the location of a permanent nuclear waste repository. While hazmat trans-
port analysts are known to complain about the quality of their data (we will
return to this topic later in this section), they do have access to considerable
historical information on accident frequencies and fairly accurate consequence
models for hazmat releases in case of accidents in many developed countries.
552 E. Erkut, S.A. Tjandra and V. Verter

Due to this, and the necessity of quantitative information for OR models, in

this section we focus on quantitative risk assessment.
Quantitative risk assessment (QRA) involves the following key steps:
(1) hazard and exposed receptor identification;
(2) frequency analysis; and
(3) consequence modeling and risk calculation.
Identification of hazard refers to identifying the potential sources of release
of contaminants into the environment, the types (e.g., thermal radiation due to
jet and pool fires and fireballs, explosions, flying pieces of metals or other ob-
jects due to blast waves, toxic clouds, and flame) and quantities of compounds
that are emitted or released, and the potential health and safety effects as-
sociated with each substance. In some cases (for example, when a release of
carcinogenic substances is involved), we also need to investigate the long-term
health risks of a hazmat accident. Examination of risks on different types of
exposed receptor is also essential to cover different response characteristics in
the risk assessment.
The language of QRA is one of frequencies and consequences, and unlike
in qualitative risk analysis, QRA results in a numerical assessment of risks in-
volved, for example, an expected number of individuals impacted per year. In
the next two sections we discuss frequency analysis and consequence modeling
along with risk calculation.

3.1 Frequency analysis

The frequency analysis involves (a) determining the probability of an unde-

sirable event; (b) determining the level of potential receptor exposure, given
the nature of the event; and (c) estimating the degree of severity, given the level
of exposure (Ang and Briscoe, 1989). Each stage of this assessment requires
the calculation of a probability distribution, with stage (b) and (c) involving
conditional distributions. Consider a unit road segment. Suppose that there is
only one type of accident, release, incident, and consequence. Let A be the
accident event that involves a hazmat transporter, M the release event, and
I the incident event. Suppose that the consequence of the hazmat release is
expressed in terms of the number of injuries. We denote the event of an injury
to an individual as D. Using Bayes’ theorem, we obtain the probability of an
injury resulting from an accident related to the hazmat as
p(A M I D) = p(D|A M I)p(A M I)
= p(D|A M I)p(I|A M)p(A M)
= p(D|A M I)p(I|A M)p(M|A)p(A) (3.1)
where p(E) denotes the probability of the event E occurring on the road seg-
ment and p(E|F) the associated conditional probability. Despite its simplicity,
the above model already contains many of the necessary elements for hazmat
 Ch. 9. Hazardous Materials Transportation 553

risk assessment. For example, Chow et al. (1990) used a Bayesian model that
includes multiple levels of event severity to predict severe nuclear accidents
and to estimate the associate risks. Glickman (1991) used a Bayesian model
in the assessment of the risks of highway transportation of flammable liquid
chemicals in bulk.
Furthermore, let slm denote the number of shipments of hazmat m on
road segment l per year. Note that a highway transport route from the ori-
gin to the destination consists of finitely many road segments. The product
slm pl (A Mm  I D) determines the frequency of the occurrence of the haz-
ardous release event that measures the individual risk for a person in the
neighborhood of road segment l. Usually, the individual risk is defined as the
yearly death frequency for an average individual at a certain distance from
the impact area (see, e.g., Mumpower, 1986; Leonelli et al., 2000). Although
no universally accepted individual risk criteria exist, one tends to compare the
risk of death to de minimis of 10−6 to 10−5 deaths per year (Mumpower, 1986).
Hazmat incidents usually impact a number of individuals. Hence, we need to
move from individual risk toward societal risk. The societal risk is a character-
istic of the hazardous activity in combination with its populated surroundings.
There are several ways to express societal risk. Perhaps the simplest method
is to compute the expected number of impacted individuals by multiplying the
probability of impact per person with the number of persons present in the im-
pact zone. Hence, the societal risk (or just risk for short) on road segment l of
hazmat m, Rlm , can be expressed as

Rlm := slm pl (Dxy |A Mm  I)pl (I|A Mm )
L
× pl (Mm |A)pl (A)POPl (x y) dx dy (3.2)
where pl (Dxy |A Mm  I) is the probability that individuals on location (x y)
in the impact area L will be dead due to the incident on a route segment l and
POPl (x y) is the population density on location (x y) in the neighborhood of
road segment l. By assuming that each individual in the affected population
will incur the same risk, Rlm can be simply expressed as
Rlm := slm pl (D|A Mm  I)pl (I|A Mm )pl (Mm |A)pl (A)POPl  (3.3)
Thus, if few people are present around the hazardous activity, the societal risk
may be close to zero, whereas the individual risk may be quite high.
While this expected consequence is a convenient measure for OR models,
the risk assessment literature prefers a richer measure, namely the FN-curve
which expands the point estimate of the expectation to the entire distribution.
To produce an FN-curve, one has to compute the probability that a group
of more than N persons would be impacted due to a hazmat accident, for
all levels of N. The risk level is communicated by the FN-curve, a graph
with the ordinate representing the cumulative frequency distribution F of the
hazardous release events which result in at least N number of impacts (e.g.,
554 E. Erkut, S.A. Tjandra and V. Verter

Fig. 4. A conditional FN-curve (given an evacuation incident).

number of fatalities or number of people evacuated) and abscissa representing

the consequence (N impacts). Furthermore, if F is a conditional cumulative
frequency distribution, then the associated FN-curve is called the conditional
FN-curve. Figure 4 shows a conditional FN-curve for PCB transport through
Edmonton, Canada (Erkut and Verter, 1995b). The ordinate F is the annual
cumulative frequency of incidents with at least N evacuations conditioned on
the occurrence of an evacuation incident in the city. This figure shows that
if an evacuation incident occurs, then the probability of evacuating more than
500 people is 0.8. Some countries (such as Denmark, Netherlands and the UK)
use decision rules for hazmat installations based FN-curves (Jonkman et al.,
2003).
Clearly, more than one type of accident, release, incident, and consequence
can occur during the hazmat transport activity. For example, a release of flam-
mable liquid can lead to a variety of incidents such as a spill, a fire, or an
explosion. To accommodate this, let A, M, I , and C denote the set of pos-
sible accidents, releases, incidents, and consequences that may occur on road
segment l. Suppose that all consequences (injuries and fatalities, property
damage, and environmental damage) can be expressed in monetary terms (see
Section 3.2.3). Then, the hazmat transport risk associated with road segment l
can be expressed as
  
Rl := slm pl (Aa  Mm  Ii  Cc )CONSc  (3.4)
a∈A m∈M i∈I c∈C
where CONSc is the possible c-type consequence.
 Ch. 9. Hazardous Materials Transportation 555

To summarize, we started with individual risk due to a single incident, then

we moved on to risk due to multiple shipments, and on to societal risk, and
finally to societal risk with multiple incidents.
However, in practice, researchers frequently neglect these conditional prob-
abilities and simplify the analysis by considering the expected loss (or the
worst-case loss) as the measure of risk. The expected value is calculated as the
product of the probability of a release accident and the consequence of the
incident (List et al., 1991). Hence the hazmat transport risk associated with
a road segment l can be expressed as

Rl := slm p(Mm )clm  (3.5)
m∈M

where clm is the undesirable consequence due to the release of hazmat m on

road segment l. This risk model is sometimes referred to as the technical risk
(Erkut and Verter, 1998). The US DOT use this expected loss definition in
their guidelines for transporting hazmats (US DOT, 1994). These simplifica-
tions are mainly due to the lack and inaccuracy of accident and exposure data.
As it is clear from the discussion above, QRA depends heavily on an esti-
mation of probabilities. There are two primary means to estimate the accident,
release, and incident probabilities: historical frequencies and logical diagrams
(fault tree and event tree analysis).

Historical frequencies
We can use the number of hazmat transport accidents in a given time pe-
riod and the total distance traveled by hazmat trucks in the same time period
to calculate the accident rate on a unit road segment (i.e., accidents per km).
The hazmat accident probability on road segment l, pl (A), can be obtained
by multiplying the accident rate by the length of road segment l. To esti-
mate pl (Mm |A), we need to calculate the percentage of hazmat accidents
that result in a release of hazmat m. Similarly, we can use historical data to
estimate pl (I|A Mm ) and pl (D|A Mm  I). However, the occurrence of an
accident may be influenced by intrinsic factors such as tunnels, rail bridges,
road geometry, weather conditions, and human factors, as well as other fac-
tors correlated to traffic conditions, such as traffic volume and frequency of
hazmat shipment. Consequently, some locations are more vulnerable to ac-
cidents than others. Therefore, a careful analysis should be done prior to
the use of historical data. The rarity of hazmat accidents may result in in-
sufficient information to determine whether historical figures are relevant to
the circumstances of concern, particularly regarding rare catastrophic acci-
dents. Moreover, in estimating the associate probabilities on road segment l
of a hazmat transportation route, the dependency to the impedances of pre-
ceding road segments should also be taken into account (Kara et al., 2003;
Verter and Kara, 2001). We will discuss this dependency issue in more detail in
Section 3.3.1.
556 E. Erkut, S.A. Tjandra and V. Verter

Logical diagram-based techniques

An alternative way to estimate the frequency (and possibly consequences)
of hazmat release incidents is the use of logical diagram-based techniques,
namely fault tree and event tree analysis. Fault Tree Analysis (FTA) is a top-
down analysis tool to identify the causes of events and to quantify various
accident scenarios that would cause the system fail. It starts with an identified
hazard (e.g., chlorine release due to a transport accident) as the root of a tree
(or top event) and works backward to determine its possible causes (e.g., col-
lision accident, derailment, and relief valve poorly sealed) using two logical
functions: OR and AND. The causative events are laid out in a tree with the
branches connected by gates comprising one of these logical functions. The
OR gate represents the union of events attached to the gate. An OR gate can
have any number of inputs (branches). The event above the gate is realized if
any one or more of the inputs occur. The AND gate represents the intersec-
tion of events attached to the gate. An AND gate can also have any number
of inputs, but the event above the gate is only realized if all the inputs occur.
Moreover, several fault trees can be combined into a single complex fault tree.
FTA enables us to determine the probability of the top event on the basis of
the probabilities of the basic events (e.g., p(D|A M I) in (3.1), where death
of an individual in hazmat transport accident is the top event) for which suffi-
cient historical data exist or expert judgments are reliable.
Event Tree Analysis (ETA), on the other hand, is a bottom-up analysis tool.
It takes at its starting point the event that can affect the system (e.g., an initial
release of hazmat) and tracks them forward through sequences of interfacing
system components to determine their possible consequences. It examines all
possible responses to the initiating event, such as the functioning, failure, or
partial failure of subsystems or different systems, in a tree structure with the
branches developing from left to right. Each outcome of the branches is usually
binary (i.e., the outcome occurs or does not occur). By assigning a probability
to each branch, the probabilities of every possible outcome following the ini-
tiating event can be determined. ETA can be used in conjunction with FTA,
called FETA, to identify and quantify the possible consequences of the top
event in fault tree. Figure 5 shows a fault tree in conjunction with an event
tree. For additional details and examples of fault and event tree construction,
we refer to Henley and Kumamoto (1981), Vesely et al. (1981), and Rhyne
(1994).
Boykin et al. (1984) applied FETA to analyze the risks associate with the se-
lection of technology alternatives in the chemical storage system. Pet-Armacost
et al. (1999) used FETA in conjunction with two Monte Carlo simulations
(one uses spreadsheet add-in @RISK and the other uses discrete event simula-
tion software ARENA) to conduct a transportation risk analysis of Hydrazine
in order to determine whether or not a relief valve should be used. FETA
was used to decompose the transport process into its basic components and
to identify the major sources of uncertainty. The event probabilities in the
event trees were derived as functions of the parameters whose effects were
 Ch. 9. Hazardous Materials Transportation
Fig. 5. A fault tree in conjunction with an event tree for hazmat release (adapted from Alp, 1995).

557
558 E. Erkut, S.A. Tjandra and V. Verter

not known. The impact of these parameters on the risks of toxic exposure, fire,
and explosion was analyzed through Monte Carlo analysis and analysis of vari-
ance. Rosmuller and Van Gelder (2003) used FETA to conduct a QRA for
the hazmat transportation in the Netherlands. The results were used to for-
mulate appropriate risk and rescue policies. They suggested that emergency
response teams could use the release data for determining impact circles for
road accidents and subsequently decide on detour routes. Moreover, expected
distributions of release quantities could be used to facilitate the training of
hazmat response personnel.

3.2 Consequence modeling and risk calculation

3.2.1 Modeling the impact area

There are many undesirable consequences of a hazmat transportation ac-
cident, such as economic losses, injuries, environmental pollution, damage
to wildlife, and fatalities. These consequences are a function of the impact
area (or exposure zone) and population, property, and environmental assets
within the impact area. The shape and size of an impact area depends not only
on the substance being transported but also on other factors, such as topol-
ogy, weather, and wind speed and direction. Estimating, a priori, the impact
area of a potential accident is difficult. Researchers used different geomet-
ric shapes to model the impact area such as a band of fixed width around
each route segment (e.g., Batta and Chiu, 1988; ReVelle et al., 1991), a cir-
cle (called danger circle), with a substance-dependent radius centered at the
incident location (e.g., Erkut and Verter, 1998; Kara et al., 2003), rectan-
gle around the route segment (e.g., ALK Associates, 1994), and an ellipse
shape based on the Gaussian plume model (e.g., Patel and Horowitz, 1994;
Chakraborty and Armstrong, 1995; Zhang et al., 2000). Figure 6 shows these
four shapes of the impact area that have been used in the literature.
Perhaps the most common approximation of the impact zone is the dan-
ger circle. By moving the danger circle along a route segment between two
nodes (see Kara et al., 2003), we get the fixed-bandwidth approximation and by
cutting off the circular segments at the two ends we get the rectangle approx-
imation. The bandwidth or radius is substance-dependent but it is assumed to
be constant for a given shipment, which means that this approximation does
not consider effect of the distance on the level of impact. One can determine
the radius by considering the evacuation distance (i.e., the initial isolation zone)
when a hazmat incident occurs, for example, 0.8 km for flammable hazmats and
1.6 km for flammable and explosive hazmats (CANUTEC, 2004). The central
assumption in these models is that each individual within the danger zone will
be impacted equally and no one outside of this area will be impacted.
The modeling of an impact area can also be considered from the point
of view of the affected population center. For example, a population center
is commonly modeled as a point on the plane, where all inhabitants of the
 Ch. 9. Hazardous Materials Transportation 559

Fig. 6. Possible shapes of impact area around the route segment.

population center are considered to experience the same impact from a haz-
mat incident on a road segment nearby. The impact on this aggregation point
depends on the distance between the point and the incident location. For ex-
ample, the impact can be inversely proportional to the square of the Euclidean
distance between the two points. However, a GIS enables researchers to rep-
resent the spatial distribution of population density more accurately (see, for
example, Figure 8) rather than using aggregation points. Erkut and Verter
(1995b) proposed a model of the spatial distribution of population by using
a polygon. Verter and Kara (2001) incorporated this in a GIS, and developed
a large-scale risk assessment model for the provinces of Ontario and Quebec.

3.2.2 Gaussian plume model

In an airborne hazmat (e.g., chlorine, propane, and ammonia) accident,
the concentration of the airborne contaminant varies with distance from the
source of accident. It will be lower as the gas disperses with distance and wind.
Therefore, the three approaches discussed above can result in poor approxima-
tions of the impact area. In this case, researchers have usually resorted to the
Gaussian plume model (GPM) (Hanna et al., 1993; Patel and Horowitz, 1994;
Chang et al., 1997; Zhang et al., 2000; Puliafito et al., 2003). The Gaussian
plume model is based on several limiting assumptions:
(1) the gas does not change its chemical properties during dispersion;
(2) the terrain is unobstructed and flat;
(3) the ground surface does not absorb the gas;
(4) the wind speed and direction is stable during the dispersion period; and
(5) the emission rate is constant.
560 E. Erkut, S.A. Tjandra and V. Verter

These assumptions certainly limit the application of GPM, for example, as-
sumption (1) restricts the applicability of the GPM to stable chemicals and to
accidents which do not result in an explosion (Zhang et al., 2000). The GPM is
formulated as
   
Q 1 y 2
C(x y z he ) = exp −
2πμσy σz 2 σy
        
1 z − he 2 1 z + he 2
× exp − + exp − 
2 σz 2 σz
where C is the concentration level (mass per unit volume – μg/m3 or parts
per million – ppm), x is the distance downwind from the source (m), y is the
distance crosswind (perpendicular) from the source (m), z is the elevation of
the destination point (m), he is the elevation of the source (m), Q is the release
rate of pollutant (mass emission rate – g/s or volumetric volume rate – m3 /s),
μ is the average wind speed (m/s), σy and σz are the dispersion parameters in
the y and z directions (m).
In hazmat dispersion from traffic accidents, it is usually assumed that the
source is on the ground (i.e., he = 0) and we are interested in the ground
concentration level (i.e., z = 0). Therefore, we obtain
   
Q 1 y 2
C(x y z he ) = C(x y) = exp − 
πμσy σz 2 σy
Figure 7 shows bell-shaped curves of concentration levels C(x y) for two dif-
ferent downwind distances: (a) the concentration of the pollutant is high at the
source of the spill (x = 0) and the Gaussian distribution has a pronounced
peak; (b) as the pollutant drifts farther downwind (x  0), it spreads out and
the bell-shape becomes wider and flatter.
The release rate, Q, depends on container volume, hazmat type, and rupture
diameter. To calculate Q, one can use ALOHA (see Section 2.5). ALOHA can
also be used for estimating the concentration level, C(x y), but its results are
only reliable within one hour of the release event, and 10 kilometers from the
release source. The dispersion parameters, σy and σz , can be determined as a
function of downwind distance x (Pasquill and Smith, 1983; Arya, 1999).
The individual risk, that is the probability that an individual at location
j with coordinate (j x  j y ) will experience an undesirable consequence (such
as evacuation, or injury, or death) as a result of a release at i, pij , can be
represented as a function of the concentration of airborne contaminant at j,
Cij := C(|j x − ix | |j y − iy |). The American Institute of Chemical Engineers
(2000) suggests a probit function to model pij (Cij ). Consequently, the social
risk can be obtained by multiplying pij (Cij ) with the population size at loca-
tion j.
A simpler alternative way to estimate the consequence of airborne hazmat
accident is to use the standard concentration level to determine the threshold
 Ch. 9. Hazardous Materials Transportation 561

(a)

(b)
Fig. 7. The bell-shape of concentration level C(x y): (a) Gaussian distribution at x = 0 and
(b) Gaussian distribution at x  0 (Chakraborty and Armstrong, 1995).

distances for different consequences (e.g., fatalities and injuries), such as Im-
mediately Dangerous to Life and Health (IDLH) (NIOSH, 1994) developed by
the National Institute for Occupational Safety and Health (NIOSH) and the
Occupational Safety and Health Administration (OSHA). The IDLH-values
represent the maximum concentration from which one could escape without
injury or irreversible health effects (e.g., severe eye or respiratory irritation,
disorientation, or lack of coordination) within 30 minutes of exposure. For ex-
ample, the IDLH-values for carbon dioxide and propane are 40,000 ppm and
2100 ppm respectively. These numbers hold for enclosed spaces (and not open-
air). To be used in an open environment, for example, Verma and Verter (2007)
considered a propane dispersion of 2100 ppm per second and assumed that
562 E. Erkut, S.A. Tjandra and V. Verter

Fig. 8. Population densities within different concentration levels (Zhang et al., 2000).

roughly a 4–5 minute propane exposure at this IDLH level can cause minor
injury while a 30–35 minute exposure can cause major injury or fatality. Us-
ing these assumptions, they defined a fatality zone (if the concentration level
C  4,200,000), an injury zone (if 600,000  C < 4,200,000) and a nonexpo-
sure zone (if C < 600,000) where C is given in ppm. Hence, the threshold
distance is determined by the level curve of the associated hazmat IDLH-value
and the associated consequence can be represented as the function of the pop-
ulation size within the level curve. Figure 8 shows the population densities
within different concentration levels of a single source release.
The following conceptual example demonstrates how an improper assess-
ment of the impact area may lead to a high-risk routing decision. Consider
two east–west routes around a city that may be used for propane shipments:
South (P1 ) and North (P2 ) routes (see Figure 9(a)). Assume each route seg-
 Ch. 9. Hazardous Materials Transportation 563

(a)

(b)
Fig. 9. (a) Gaussian plume model vs. (b) danger circle.

ment in both routes has the same incident probability. Suppose these routes
divide the city into three regions A, B, C, where each region has uniform popu-
lation density. Among these regions, suppose that region B is the most densely
populated one and region C is the least densely populated one. Moreover,
suppose that the prevalent wind direction is south-east. Figure 9(b) shows con-
centration contours of route segments in P1 and P2 , according to the IDLH
value. Since the population density in the impact area of route P1 is less than
that of P2 , one might send propane via route P1 . In contrast, if one were to use
a danger circle instead of the Gaussian plume model, neglecting the type of
hazmat and the wind direction, one may select route P2 instead of P1 . This de-
cision would expose more people in case of an incident as propane would drift
south-eastwardly into region B. As this simple example demonstrates a careful
analysis is necessary prior to defining the impact area.

3.2.3 Risk cost

To estimate the cost of a hazmat release incident, various consequences
must be considered. The consequences can be categorized into (Abkowitz
et al., 2001; FMCSA, 2001): injuries and fatalities (or often referred to as pop-
ulation exposure), cleanup costs, property damage, evacuation, product loss,
traffic incident delay, and environmental damage. All impacts must be con-
verted to the same unit (for example dollars) to permit comparison and com-
plication of the total impact cost. The discussion of risk costs presented here
deals primarily with hazmat incidents on highways.
564 E. Erkut, S.A. Tjandra and V. Verter

Injuries and fatalities. Finding a dollar value of human life and safety is per-
haps the most difficult and controversial issue. Some find it offensive; others
argue that any dollar value assigned to human life would be too low. Yet it
is possible to estimate the value indirectly. Insurance payments offer a simple
estimate. Perhaps more relevant is the figure used by government agencies to
prioritize their projects that reduce fatalities and injuries. Clearly if an agency
is making a choice between Project A which will save X lives and cost P dol-
lars per year and Project B which will save Y lives and cost Q dollar per year,
they are implicitly using a trade-off value that converts fatalities to dollars –
regardless of whether or not the trade-off is made explicit.
The value of an injury or fatality in a hazmat incident can be estimated from
different perspectives (FMCSA, 2001). For example, one can value an injury
or fatality in terms of lost income and economic productivity to society. The
National Highway Transportation Safety Administration (NHTSA) estimates
the cost of fatalities and injuries by considering both direct and indirect costs
to individuals and to society (NHTSA, 1996). Direct costs include emergency
treatment, initial medical costs, rehabilitation costs, long-term care and treat-
ment, insurance administrative expenses, legal costs, and employer/workplace
costs. Indirect costs are productivity losses in the workplace due to temporary
and permanent disability and decreases in productivity at home resulting from
these disabilities. In 1996 dollars, a fatality costs about $913,000 and a critical
injury costs about $780,000.
In addition to the economic cost components discussed above, The National
Safety Council (NSC) also includes the value of a person’s natural desire to live
longer or to protect the quality of one’s life (NSC, 2003). This value indicates
what people are willing to pay to reduce their safety and health risks. Hence,
the cost estimates include wage and productivity losses, such as wages and
fringe benefits, replacement cost and travel delays caused by the accident; med-
ical expenses, such as doctor fees, hospital charges, cost of medication, future
medical costs, and other emergency medical services; administrative expenses,
such as insurance premiums and paid claims, police and legal costs; motor ve-
hicle damage, such as property damage to vehicles; and employer costs, such
as time lost by uninjured workers, investigation and reporting time, produc-
tion slowdowns, training of replacement workers, and extra costs of overtime
for uninsured workers (FMCSA, 2001). The 2003 estimates of incapacitating
injury and fatality costs are $181,000 and $3,610,000, respectively.
Finally, US DOT values injuries and deaths at the amount they would spend
to avoid an injury or fatality (FMCSA, 2001). This averages out to be $400,000
to avoid an injury requiring hospitalization and $2,800,000 to avoid a fatality.

Cleanup costs. Cleanup costs are assumed to encompass the costs of both
stopping the spread of a spill and removing spilled materials (Abkowitz et al.,
2001; FMCSA, 2001). Such costs vary widely depending on the size, type of ma-
terials, and location of the spill. Some national database systems, such as the
Hazardous Materials Information System (HMIS) of US DOT and The Work-
 Ch. 9. Hazardous Materials Transportation 565

place Hazardous Materials Information System (WHMIS) of Health Canada,

can be used as references for the cleanup costs. For the period 1990–1999,
cleanup costs averaged about $24,000 per en-route accident, $1300 per cleanup
for an en-route incident spill, and $260 for an unloading/loading accident and
incident spill cleanup (HMIS database).

Property damage. Property damage encompasses damage to other vehicles,

which may have been involved in the incident, as well as damage to both public
and private property (e.g., private buildings, public utilities, public roadways).
For example, from HMIS database of the period 1990–1999, the average prop-
erty damage for flammable and combustible liquids en-route accidents was
$16,041, while the average property damage for en-route incident spills was
$274. Average property damage for leaks occurring during loading and unload-
ing incidents and accidents was $68. Average property damage for flammable
gases en-route accidents, en-route spills, and loading/unloading incidents were
$3147, $173, and $2315, respectively. For corrosive materials, the average val-
ues for en-route accidents, en-route spill incidents, and loading/unloading in-
cidents were $3104, $67, and $17, respectively (FMCSA, 2001).

Evacuation. There are numerous variables which complicate the estimation

of the cost of evacuation. These include the expense for temporary lodging and
food, losses due to lost wages and business disruptions, inconvenience to the
public, and the cost of agencies assisting in evacuation. A reasonable estimate
would be $1000 per person evacuated (TRB, 1993). This $1000 estimate is also
used by the Federal Railroad Administration (FRA) to estimate impacts from
railroad evacuations.

Product loss. Product loss refers to the quantity and value of the haz-
mats lost during a spill. For example, from the HMIS database for period
1990–1999, the average cost of product lost of flammable and combustible
liquids en-route accident related spills was $3208 per spill. Similarly, for flam-
mable gases accidents, the average cost of product lost per en-route accident
related spill was $1140 per spill. Corrosive material spill accidents averaged
$4910 per spill in product loss.

Traffic incident delay. Hazmat spills typically require an emergency response

that causes a significant traffic delay. This type of traffic delay is called in-
cident delay. If traffic volume and incident situation (e.g., the traffic arrival
rate, road capacity reduction, and incident duration) is known, a deterministic
model can be used to estimate the incident delay. For example, Morales (1989)
used a deterministic queueing model and Wirasinghe (1978) and Alp (1995)
used models based on shock-wave theory. Due to its simplicity, the Morales
model is often used by practitioners (see, for example, Abkowitz et al., 2001;
FMCSA, 2001). However, these deterministic models are inappropriate for
prediction of incident delay in real-time situation where the incident duration
566 E. Erkut, S.A. Tjandra and V. Verter

is unknown. In this case, incident delay is best modeled by a random variable

that represents the stochastic characteristics associated with the incident (as in
Fu and Rilett, 1997).
To obtain the associated costs of incident delays, information on the oc-
currence of an incident or the split between trucks and other vehicles on the
various highway systems are required. Earlier studies (Grenzeback et al., 1990)
assumed the hourly cost of incident delay to be about $20 for trucks and $10
for other vehicles, which accounts for the value of a driver’s time and fuel
consumption costs. The total cost traffic incident delay is then obtained by
multiplying this dollar value of incident delay with the total number of person-
hours of delay given by the model discussed above.

Environmental damage. Environmental damage consists of damage to the en-

vironment that remains after the cleanup. This damage can be calculated in
terms of loss of economic productivity, such as agricultural production lost
and/or in loss of habitat or ecosystem deterioration (FMCSA, 2001). The loss
of agricultural productivity can be estimated, for example, using the quantity
of crops that are not grown during a 20-year period due to contamination. Us-
ing wheat as an example, a contaminated field that can produce 35 bushels per
acre/year would result in an (undiscounted) gross income loss of $3500/acre
over a 20-year period assuming a fixed value of $5/bushel. To calculate the
natural resource environmental damage from a hazmat incident is more com-
plicated. We need to know how much material was spilled, where the spill
occurred, and what sort of surface it covered. Using, for example, HMIS data,
one can estimate the dollar cost of this damage.
As the discussion in this section points out, while there are different types of
costs associated with a hazmat transport incident, in most cases all other costs
are dwarfed by the cost of fatalities and injuries and the cost of evacuations
in cases of major spills. Perhaps this is a reason why many OR analysts focus
exclusively on populations inside a danger circle.

3.2.4 Perceived risks

All consequences we discussed so far assume that society is risk-neutral;
i.e., we are indifferent between two consequence distributions, as long as their
expected values are equal. For example, risk neutrality assumes that a sin-
gle incident causing 100 fatalities is equivalent (or equally undesirable to the
society) to 100 incidents causing one fatality each, since in both cases the to-
tal number of fatalities is the same. However, most individuals would judge
a low probability–high consequence (LPHC) event as more undesirable than
a high probability–low consequence (HPLC) event even if the expected con-
sequences of the two events are equal (Erkut and Verter, 1998). Consequently
when dealing with LPHC events, most human decision makers tend to exhibit
risk aversion; a single incident causing 100 fatalities is perceived as much more
undesirable than 100 incidents each causing a single fatality.
 Ch. 9. Hazardous Materials Transportation 567

A simple way to incorporate risk attitude to risk models is to add a risk

preference (or tolerance) factor α as an exponent to the consequence values.
For example, if the risk assessment deals with the population exposure, then
the societal risk on road segment l (see (3.2), dropping the hazmat index m)
can be expressed as (see, e.g., Slovic et al., 1984; Abkowitz et al., 1992)

Rl := sl pl (Dxy |A M I)pl (I|A M)pl (M|A)pl (A)
L
 α
× POPl (x y) dx dy
By considering only one shipment (or one trip) and one type of hazmat spill,
the traditional expected loss model of risk (3.5) can thus be modified as (see,
e.g., Slovic et al., 1984; Abkowitz et al., 1992; Erkut and Verter, 1998; Erkut
and Ingolfsson, 2000):
Rl := pl (POPl )α 
Figure 10 shows three different values of α associated with three different
risk preferences: α = 1 models risk neutrality; α > 1 models risk aversion;
and α < 1 models risk-taking behavior. The greater the value of α, the higher
the aversion to the risk of a hazmat incident. The risk-aversion model assumes
that the (i + 1)st life lost is more costly than the ith life lost, for all possible
values of i. Of course as α is increased, any route selection model that operates
with an objective of minimizing total risk is eventually reduced to a model
that minimizes the maximum risk, as shown by the following small example.
Consider a hazmat shipment from an origin O to a destination D. There are
two routes (north and south) between O and D, P1 and P2 , each consisting of
two route segments. Suppose that the incident probability and the population
density in the impact area of the two segments of route P1 are (10−4 ; 25) and
(10−4 ; 75), and those of P2 are (10−5 ; 100) and (10−5 ; 400). The total risks
associated with P1 and P2 are 10−2 and 5×10−3 , respectively, and the maximum
risks are 75 × 10−4 and 4 × 10−3 , respectively. For α = 1, we select P2 , the
route with lower total risk. As α approaches infinity, the problem turns into
one of minimizing the maximum risk, and we select P1 . Figure 11 shows how

Fig. 10. Three different risk preferences: α = 1 risk neutral; α > 1 risk aversion; α < 1 risk proneness.
568 E. Erkut, S.A. Tjandra and V. Verter

Fig. 11. Routing decision for different values of α, based on perceived risk.

the optimal routing decision changes from P2 to P1 as the risk-aversion factor α

increases.
The perceived risk model can be thought of as a simple (dis)utility model.
It is possible to model risk disutility in other ways. For example, Kalelkar and
Brinks (1978) proposed an empirical disutility function that was constructed
by using a series of questions posed to decision makers. Erkut and Ingolfsson
(2000) proposed an exponential disutility function to model risk aversion.

3.3 Risk on a hazmat transportation route

Up to this point, we discussed hazmat transport risk in general. Now we dis-

cuss the modeling of risk along an edge, and then along a route, of a transport
network. In other words, we now move from point risk (risk due to accident
at a given point) to linear risk (risk along an edge and route). Consider a road
network G = (N E) with node set N and edge set E. The nodes correspond
to the origin, the destination, road intersections, and population centers and
the edges correspond to road segments connecting two nodes. (We note that
one does not have to model population centers as nodes if one uses a GIS as
discussed earlier.) We first focus our discussion on road transportation, and
then move to hazmat transportation on rail.
Note that in the context of hazmat routing it is desirable that each point on
an edge has the same incident probability and level of consequence (e.g., popu-
lation density). Therefore, a long stretch of a highway that goes through a series
 Ch. 9. Hazardous Materials Transportation 569

of population centers and farmland should not be represented as a single edge,

but as a series of edges. Thus, a network to be used for hazmat routing is usu-
ally different from a network to be used for other transport planning purposes.
This difference is quite important since it limits the portability of network data-
bases between different transport applications (Erkut and Verter, 1998). We
first discuss the modeling of risk along an edge.

3.3.1 Edge risk

Erkut and Verter (1998) proposed a risk model that takes into account the
dependency to the impedances of preceding road segments (see also Jin et al.,
1996; Jin and Batta, 1997). Suppose that an edge is a collection of n unit road
segments each with the same incident probability p and consequence c. The
probability p is obtained from (3.1) and the consequence c is determined by
taking a proper impact area of a unit road segment. If, for example, the impact
area of a unit road segment is modeled as a danger circle, then the impact area
of an edge is a semicircular shape with the same radius as the danger circle, as
shown in Figure 12. The vehicle will either have an incident on the first road
segment, or it will make it safely to the second segment. If it makes it safely
to the second segment, it will either have an incident in the second segment,
or it will not, and so on. They assumed that the trip ends if an incident occurs.
Hence, the expected risk associated with this edge would be
pc + (1 − p)pc + (1 − p)2 pc + · · · + (1 − p)n−1 pc (3.6)
Since the incident probability p is at most on the order of per trip 10−6
per kilometer (based on North American data, Harwood et al., 1993), we can
approximate
ps ∼
= 0 for s > 1 (3.7)
Consequently, the risk of hazmat transport on this edge becomes pnc. For
edge i, we can, thus, define the risk as
ri = pi ci  (3.8)
where the probability of an incident on edge i is pi := np, and the associated
consequence is ci := c.
Note that this simple risk model works under an assumption of uniform
incident probability and uniform consequence along an edge. If these two

Fig. 12. Semicircular impact area around link (i j).

570 E. Erkut, S.A. Tjandra and V. Verter

attributes are not uniform, the risk computation on either an edge or an origin–
destination route will be more complicated. In practice, however, this assump-
tion will be valid if we define long stretches of a highway as a series of edges.
(In other words, it is not only the network topology, but also the value of the
edge attributes that define an edge. The edges must be short enough that the
accident probability and the consequence are constant along the entire edge.)
This edge risk definition can be considered as a generalization of the classical
(or traditional) risk definition, which considers risk as an expected loss (see
Section 3.1). The expected loss can be obtained from (3.6) by defining n = 1,
i.e., each unit road segment is considered as an edge of the road network. Next
we will discuss in detail some ways to model and calculate the edge risk.
Recall that according to Equation (3.2), the risk of a hazmat accident on
road segment l can be calculated by considering the probability that individu-
als in neighborhood L (of road segment l) will be affected due to the incident
and the population density in L. A hazmat vehicle at point v on edge (i j)
poses a threat to the population at each point v in the impact area L. The
hazmat incident probability pij (v), can be obtained from (3.1) and it is mea-
sured in probability of accident per-unit length of movement. Moreover, let us
assume that the consequence is determined by assuming that the impact area
is a danger circle with radius λ.
The edge-risk formulation can be derived as follows. Let lij denote the
length of edge (i j) and wv denote the population density at a point v . The
risk at point v , rv ij , due to the hazmat transport on an edge (i j) is deter-
mined by
 lij
rv ij := wv δ(v v )pij (v) dv (3.9)
v=0

where

1 d(v v )  λ
δ(v v ) :=
0 otherwise
with d(v v ) the Euclidean distance of two points v and v . To calculate the
lij
integral v=0 δ(v v )pij (v) dv, one can move the origin to node i and rotate
the axes so that edge (i j) lies on the positive abscissa. Denote this integral by
Fij (vv ). The semicircular area around an edge (i j) consists of four regions
with different expressions to calculate Fij (v ), as shown in Figure 13. We note
that region II is empty when lij > 2λ. If the coordinate of v is (xv  yv ) and
x+ (v ) and x− (v ) are the intersections of the abscissa with a circle of radius λ
centered at v , then

x+ (v ) = xv + λ2 − yv2 and x− (v ) = xv − λ2 − yv2 

if λ > |yv | (3.10)
 Ch. 9. Hazardous Materials Transportation 571

Fig. 13. Regions inside the semicircular impact area with radius λ around edge (i j) (Batta and Chiu,
1988).

for every point v in the road network. In this case, Batta and Chiu (1988)
showed that
⎧ + 
⎪
⎪
x (v )
pij (v) dv v is in region I
⎪
⎪ 0
⎪
⎪ lij
⎪
⎨ 0 pij (v) dv v is in region II

Fij (v ) = lij
p (v) dv v is in region III (3.11)
⎪
⎪ x− (v ) ij
⎪
⎪ + 
⎪
⎪
x (vv )
pij (v) dv v is in region IV
⎪ − 
⎩ x (v )
0 v is outside the semicircular area
Hence, the total risk of a hazmat vehicle travels on edge (i j) is

rij = rv ij dv 
v ∈L
Batta and Chiu (1988) assumed that population centers are located at nodes
and along the edges of the road network. Thus, a hazmat vehicle at point v
on edge (i j) poses a threat to the population at node v and/or at point v
on edge (i  j  ). Let wv denote the population density at node v , and fkl (v )
denote the population density function associated with edge (i  j  ). Moreover,
assume that the function fi j  (v ) has been normalized so that its integral from
zero to li j  equals one. The nodal risk at node v , rv ij , is determined by (3.9)
and the edge risk on edge (i  j  ), ri j  ij , due to the hazmat transport on edge
(i j) is determined by
 l   lij
ij

ri j  ij := fi j  (v ) δ(v v )pij (v) dv dv 
v =0 v=0
To calculate the edge risk ri j  ij , we need to partition edge (i  j  ) into regions
as discussed earlier (see Figure 14). Let consider a point v on (i  j  ), which is
v units from node i . By definition, the coordinates of this point are
v v
xv = xi + (xj  − xi ) and yv = yi + (yj  − yi ) 
li  j  li  j 
572 E. Erkut, S.A. Tjandra and V. Verter

Fig. 14. Partition of edge (i  j  ) into regions inside the semicircular impact area with radius λ around
edge (i j).

Using this coordinate and (3.10) and (3.11), one can calculate x+ (v ), x− (v ),
Fij (v ), and finally the edge risk ri j  ij . Hence, the total risk of a hazmat vehicle
travels on edge (i j) is
 
rij = ri j  ij + rv ij 
(i j  ) v

3.3.2 Path risk

An origin–destination route P for a hazmat shipment is a collection of edges,
where travel on this path can be viewed as a probabilistic experiment as shown
in Figure 15 (see, e.g., Jin et al., 1996; Jin and Batta, 1997; Erkut and Verter,
1998). Similar to the argument used above, a hazmat vehicle will travel along
the ith edge of P only if there is no incident on the first (i − 1) edges of P
(i.e., an incident terminates a trip along P). Suppose that the path P has n
edges. Note that n may represents the length of the path if each edge ei ∈ E,
i = 1     n, has length of one unit. The expected path risk associated with this
trip can be expressed as

n 
i−1
R(P) = (1 − pj )pi ci  (3.12)
i=1 j=1

where (1 − p1 )(1 − p2 ) · · · (1 − pi−1 )pi ci is the impedance of the ith edge of P.

By this definition, edge impedances are path-dependent.
Using an approximation similar to (3.7), that is, assuming
p i pj ∼
= 0 for all edges i j (3.13)
we obtain a very simple linear form of path risk

n
R (P) = pi ci  (3.14)
i=1
 Ch. 9. Hazardous Materials Transportation 573

Fig. 15. Partial probability path displaying possible outcomes of a hazmat transport along a path, where
pi is the incident probability and ci is the associated consequence along the ith edge of the path (Erkut
and Verter, 1998).

This model is often referred to as the traditional risk model, since it explic-
itly uses the expected consequence definition of risk. Note that this model is
simple to explain and justify, and it is not data intensive; it requires only one
accident probability and one consequence figure per edge. Furthermore, it is
rather easy to work with in optimization models. In fact, minimizing (3.14) for
a given OD pair in a hazmat transport network is a shortest path problem which
is solved easily for even large networks. For these reasons, most papers on haz-
mat transportation use this traditional risk model (Erkut and Verter, 1998).
The US DOT also uses this approach in their guidelines (US DOT, 1994).
This simple risk model makes a tacit assumption that the hazmat vehicle will
travel along every edge on the path, regardless of what happened on earlier
edges. Consequently, a single hazmat trip can result in several incidents (with
a very small probability). In some cases, though very unlikely, this assumption
is practically reasonable. After a minor incident, the cargo may still be trans-
ported to the destination, perhaps on a different vehicle and/or on different
route. Nevertheless, since incident rates for hazardous materials are very low,
the probability of the conditioning event that an incident has not yet occurred
when an edge i is reached will always be very close to 1. Therefore, the two
assumptions (an incident terminates a trip and an incident does not terminate
a trip) and (3.12) and (3.14), consequently, will differ insignificantly. Erkut
and Verter (1998) point out that this approximation is likely to result in a very
small error (less than 0.25% in most cases) in measuring the incidence proba-
bility along a hazmat transport route. Erkut and Ingolfsson (2000) provide an
upper bound of exp(npmax ) − 1 on the percent of error introduced by (3.14)
574 E. Erkut, S.A. Tjandra and V. Verter

relative to (3.12), where pmax is an upper bound on the incident probability on

any edge. Formally
R (P) − R(P)
< exp(npmax ) − 1
R(P)
For example, for a path P with length 4800 km, using an incident probability
of 10−6 per trip-km, we can compute an upper bound of exp(npmax ) − 1 =
exp(0004800) − 1, which is 048%. This upper bound is obtained by assuming
that accidents along any edge i with length li occur according to a spatial Pois-
son process with rate λi per distance unit. Under this assumption, the risk on
path P can be obtained as

n 
i−1

R (P) = exp(−pj ) 1 − exp(−pi ) ci  (3.15)
i=1 j=1

where pi = λi li . By defining
qi := 1 − exp(−pi ) (3.16)
for all edges i, (3.15) reduces to (3.12) with qi replacing pi .
Although the traditional risk model has been the most popular one, many
other hazmat transport risk models have been proposed in the literature. Ta-
ble 3 summarizes nine models and cites studies that have used each model.
Each of the seven models that use probabilities are based on approxima-
tion (3.13), even though this approximation is usually not mentioned explicitly.
We will refer to the alternate expressions of these seven models without using
approximation (3.13) as “exact.” Most of the models use population exposure
as the consequence measure. In the population exposure model, ci denotes
the total population in the rectangle shape impact area that stretches along
edge i. Other models use the circle-shaped impact area. Based on the empiri-
cal analysis on the US road network, Erkut and Verter (1998) suggest that the
choice of risk model is important because it effects the path selection decision
and the optimal path for a certain criterion can perform very poorly under an-
other. Therefore, researchers as well as practitioners must pay considerable
attention to the risk modeling in hazmat transport.
In addition to the path risk models summarized in Table 3, Jin and Batta
(1997) proposed six exact risk models, which relate the number of shipments
or trips S that need to be made and the threshold number of accidents T . The
shipments cease after T accidents occur or S trips have been made, whichever
come first. The hazmat shipments are considered as a sequence of independent
Bernoulli trials. Moreover, it is assumed that a trip is over if an accident occurs
on that trip or the destination is reached. Here, we provide a summary of these
risk models and refer the reader to Jin and Batta (1997) for more detail.
• Expected consequence of each trip given that shipment will continue
no matter how many accidents occur (i.e., when S = T = ∞).
 Ch. 9. Hazardous Materials Transportation 575

Table 3.
Alternative models of path risk (adapted from Erkut and Ingolfsson, 2005)

Model Approximation approach Satisfy axioms? Sample references

(approximation/
exact model)
Y = yes; N = no;
NA = not applicable
Axiom 1 Axiom 2 Axiom 3
n
Traditional i=1 pi ci Y/N Y/N Y/N Batta and Chiu, 1988;
risk US DOT, 1994;
Alp, 1995;
Zhang et al., 2000
n
Population i=1 ci NA/Y NA/Y NA/Y Batta and Chiu, 1988;
exposure ReVelle et al., 1991
n
Incident i=1 pi Y/Y Y/Y Y/Y Saccomanno and
probability Chan, 1985;
Abkowitz et al., 1992
n α
Perceived i=1 pi ci , α > 0 Y/N Y/N Y/N Abkowitz et al., 1992
risk n n
Conditional i=1 pi ci / i=1 pi N/N N/N N/N Sivakumar et al. 1993,
risk 1995;
Sherali et al., 1997
Maximum maxei ∈P ci NA/Y NA/Y NA/Y Erkut and Ingolfsson,
population 2000
exposure n
Expected i=1 pi (exp(αci ) − 1), α > 0 Y/N Y/N Y/N Erkut and Ingolfsson,
disutility 2000
n 2
Mean– i=1 (pi ci + βpi ci )4, β > 0 Y/N Y/N Y/N Sivakumar and Batta,
variance 1994;
Erkut and Ingolfsson,
2000
n n
Demand i=1 (1 − exp(−pi ))ci j=i exp(pj ) NA/Y NA/N NA/Y Erkut and Ingolfsson,
satisfaction 2005

Note: The three axioms tabulated here are discussed in the next subsection.

• Expected total consequence given that shipments will cease either

when T accidents occur or S shipments are finished (i.e., when T <
S < ∞).
• Expected total consequence given that shipments will cease when T ac-
cidents have occurred (i.e., when T < ∞ and S = ∞).
• Expected total consequence given that shipments will cease when T ac-
cidents have occurred (i.e., when T < ∞ and S = ∞) and parameters
change after an accident.
• Expected consequence per trip given that shipments will cease when
T accidents have occurred (i.e., when T < ∞ and S = ∞).
576 E. Erkut, S.A. Tjandra and V. Verter

• Expected number of trips between two successive accidents.

Most of the exact expressions in Jin and Batta (1997) are too complicated
for optimization purposes, and hence only the associated approximate models
are of interest for practical purposes. Yet, approximations for the fourth and
fifth models above are still not available. Further research on situation-specific
models, such as the six listed above, is warranted.
We now discuss briefly the last three rows in Table 3, which are the most
recently proposed hazmat transport risk models.

Expected disutility model. The disutility model incorporates the risk aversion
of the society toward hazmat incidents, especially the catastrophic incidents
(incidents with very large consequences). Erkut and Ingolfsson (2000) assumed
that hazmat incidents occur according to a spatial, nonhomogeneous Pois-
son process defined over the edges of the network. Let Ni and Xi denote
the number of hazmat incidents that occur on the ith edge and the num-
ber of people affected by a hazmat incident on the ith edge, respectively, of
path P, where Ni has a Poisson distribution with a parameter pi , the incident
probability on ith edge of path P. We can thus define Xi = ci Ni , where ci de-
notes the associate population exposure. The disutility function is defined as
u(X) := exp(αX), where the constant α > 0 is a measure of catastrophe aver-
sion. The higher the values of α, the more extreme the catastrophe aversion.
By assuming that a single trip can result in several incidents,
 the expected disu-
tility for a path P can be obtained as E(u(X)) = exp[ ni=1 pi (exp(αci ) − 1)].
Minimizing E(u(X))
 is then equivalent to minimizing the summation in the ex-
ponent, i.e., ni=1 pi (exp(αci ) − 1). Hence, finding a minimum disutility path
is equivalent to finding a shortest path with edge attribute pi (exp(αci )−1). The
magnitude of the edge attributes can become very large. For example, suppose
the population exposure is 10,000, the incident probability is 10−6 , and the risk
aversion constant is 0.01. Then, the edge attribute is 10−6 (exp(100)−1) ≈ 1036 .
As the risk aversion constant α increases, the edge attribute will approach infin-
ity. Consequently, this will ban the associated edge from consideration during
a route selection process that seeks a finite solution. Under an assumption that
an incident terminates the trip, the expected utility for a path P (i.e., the exact
model) can be obtained as
 n 
  
E u(X) = exp ri exp(αci ) − 1 
i=1
where

i−1
 
ri := exp(−pj ) 1 − exp(−pi )  ei ∈ P (3.17)
j=1

denotes the incident probability on edge i conditioned that no incident oc-

curred on the first (i − 1) edges. By definition, ri are path dependent.
 Ch. 9. Hazardous Materials Transportation 577

Mean–variance model. Many available risk models are based solely on the ex-
pected value of the risk and ignore how risk may deviate from the mean value.
Sivakumar and Batta (1994) proposed a risk model that identifies the least
expected length path subject to the constraint that the variance of the path
length is within a pre-specified threshold. The model is formalized as an inte-
ger programming problem with linear objective function and both linear and
nonlinear constraints. The nonlinear constraints contain quadratic terms which
account for the covariance of length between two edges. Since the covariance
terms can be negative, subtour elimination constraints are added to ensure
a simple-path solution. The authors developed an efficient solution procedure,
based on the Lagrange multipliers, to solve the equivalent linear integer pro-
gramming problem, which is obtained by linearizing the quadratic terms.
Under the same Poisson distribution for the incident rates as in the disutil-
ity model, Erkut and Ingolfsson (2000) proposed a risk model that takes into
account both the expected value and variance of the number of people af-
fected by an incident. Using the same definition of Xi , Ni , and ci , and assuming
that a single trip can result in several incidents (i.e., the approximate model),
the expected value and the variance of X(P), the total numberof people af-
fected by incidents caused by travel along P, are E[X(P)] = ni=1 ci pi and
 
Var[X(P)] = ni=1 ci2 pi . The associate exact models are E[X(P)] = ni=1 ci ri
 
and Var[X(P)] = ni=1 ci2 ri − ( ni=1 ci ri )2 , where ri are defined as in (3.17).
One can consider these two measures E[X(P)] and Var[X(P)] simultane-
ously in a multiobjective model and search for paths that are Pareto-optimal
with respect to both E[X(P)] and Var[X(P)]. To deal with the multiobjective
model, one can use the weighted sum technique and obtain a disutility model
E[X(P)] + β Var[X(P)] for a given constant β. By minimizing this for several
values of β, several Pareto-optimal paths can be found.

Demand satisfaction model. When a hazmat is transported to satisfy a de-

mand (e.g., a shipment of chlorine from a producer to a chemical processing
plant), an incident will result in a subsequent shipment. Hence, we must con-
sider the possibility of multiple trips to fulfill the demand. Erkut and Ingolfsson
(2005) proposed a simple demand satisfaction model by assuming that an in-
cident will terminate a trip (i.e., referring to exact model) and a new shipment
must be arranged to fulfill the demand. The exact probability that transport
along a path P results in at least one incident is


n
p̄(P) = 1 − (1 − qi )
i=1

where qi are defined as in (3.16). By assuming that this probability is indepen-

dent of any previous trips that were terminated by an incident, then one can
consider each trip as a Bernoulli trial, with probability 1− p̄(P) = ni=1 (1−qi )
of success in any given trial. The number of trips required (on the same path)
578 E. Erkut, S.A. Tjandra and V. Verter

before the first success (i.e., trip arrives at the destination

 safely) will then fol-
lows a geometric distribution with expected value 1/ ni=1 (1 − qi ). By taking
the expected consequence per trip as in (3.15), the expected total consequence
from all trips required to fulfill demand is
n i−1
i=1 j=1 (1 − qj )qi ci n n

R (P) = n = qi ci (1 − qj )−1  (3.18)
j=1 (1 − qj ) i=1 j=i

The expression in (3.18) has the following intuitive interpretation: the term qi ci
is
nthe expected risk associated with traversing edge i once and the term
−1
j=i (1 − qj ) is the expected number of times that edge i and the subse-
quent edges on the path must be traversed before the shipment reaches the
destination.

3.3.3 Path risk axioms

Now, we will discuss three important axioms which can be used to assess the
merits of the different models listed in Table 3. Define v(P) to be an evaluation
function that operates on path P (such as distance, cost, or risk). Let P1 denote
the set of all paths between an origin O1 and a destination D1 , and P2 denote
the set of all paths between an origin O2 and a destination D2 . Let assume that
for any P1 ∈ P1 there is P2 ∈ P2 such that P1 ⊂ P2 .

Axiom 1 (Monotonicity axiom for path evaluation models (Erkut, 1995)). If

a path P1 is contained in a path P2 , then v(P1 )  v(P2 ).

Axiom 2 (Optimality principle for path selection models (Erkut and Verter,
1998)).
v(P2 ) = min v(P) ⇒ v(P1 ) = min v(P)
P∈P2 P∈P1

For the third axiom, we assume that v(P) is a function of K edge vec-
tor attributes uk (P) of size n, the number of edges in P, i.e., v(P) =
f (u1 (P)     uK (P)). For example, the attributes of any edge in P can be
the incident probability and its associated consequence. In this case, we have
K = 2.

Axiom 3 (Attribute monotonicity axiom (Erkut and Verter, 1998)). If hk , k =

1     K, are nonnegative vector of reals of size n, then
 
f u1 (P)     uK (P)  f u1 (P) + h1      uK (P) + hK 

The first axiom implies that the evaluation value of a path will not decrease
as edges are added to the path. Clearly additive value functions (e.g., distance,
cost, travel time) satisfy this monotonicity axiom. The second axiom is merely
a restatement of Bellman’s optimality principle that implies a concatenating
 Ch. 9. Hazardous Materials Transportation 579

property of the shortest path. That is, all subpaths of an optimal path should
themselves be optimal. Evaluation functions that satisfy Axiom 2 are called
order-preserving functions. The third axiom states that the path evaluation func-
tion is a nondecreasing function of edge attributes. Consequently, path risk
is a nondecreasing function of edge incident probabilities and edge conse-
quences, i.e., increased probability or consequence on an edge cannot result
in reduced path risk.
One of the nine models in Table 3, namely the conditional risk model, vio-
lates all three axioms. Erkut (1995) and Erkut and Verter (1998) argued that
this model has some undesirable properties which make the model inappro-
priate for planning of hazmat shipments. For example, increasing the accident
probability on a link may reduce the conditional risk of a route that includes
that link.
We now consider the remaining eight models in Table 3. Most of the
approximate versions of the models listed in Table 3 satisfy all three of
these axioms. However, without assumption (3.7) or (3.13), most of the “ex-
act” models containing probabilities do not satisfy the axioms. For example,
consider the exact version of the traditional risk model defined in (3.12).
One can easily construct a simple example to demonstrate that looping re-
duces the risk (see, e.g., Boffey and Karkazis, 1995; Erkut and Verter, 1998;
Erkut and Ingolfsson, 2005). A loop in hazmat route is clearly undesirable
(and indefensible). Therefore when using this exact model one must restrict
the feasible set to loopless paths (as in Boffey and Karkazis, 1995). However,
if one makes assumption (3.7), looping will not occur. Hence, the approximate
version of the traditional risk model does not have the undesirable property of
the exact version.
The simple example in Figure 16 demonstrates how the exact traditional
risk model may result in an indefensible route selection. Node 1 is the origin
and node 4 is the destination. The incident probabilities and consequences are
given along the edges.
The exact risks associated with the two paths are as follows:
Path(1 2 4) :

10−4 × 10 + 1 − 10−4 × 10−4 × 110000 = 109999
Path(1 3 4) : 1 × 10 + 0 = 10
Hence, the exact version of the traditional risk model would select
Path(1 3 4), and this selection is guaranteed to result in an incident. The
downstream consequences on edges (2 4) and (3 4) are so high that the model
chooses the path which guarantees the truck will not reach the downstream
edges. Such a model is not suitable for decision making.
In general, in spite of their more realistic assumption (i.e., an incident will
terminate the trip) most of the exact versions of risk models have some puz-
zling properties and they may be unsuitable for hazmat transportation plan-
ning. We suggest that researchers and practitioners consider the properties of
the risk models carefully before selecting one.
580 E. Erkut, S.A. Tjandra and V. Verter

Fig. 16. Numerical example to demonstrate an undesirable property of the exact traditional risk model.

4 Routing and scheduling

Routing hazmat shipments involves a selection among the alternative paths

between origin–destination pairs. From a carrier’s perspective, shipment con-
tracts can be considered independently and a routing decision needs to be
made for each shipment, which we call the local route planning problem. A ship-
ment typically involves multiple vehicles that have to be scheduled. Since the
risk factors pertaining to each alternative route (such as accident probability
and population exposure) can vary with time, the vehicle routing and schedul-
ing decisions are intertwined, which we call the local routing and scheduling
problem. At the macro level, hazmat routing is a “many to many” routing prob-
lem with multiple origins and an even greater number of destinations (List and
Abkowitz, 1986). In the sequel, we refer to this problem as global route plan-
ning.
The local routing problem is to select the route(s) between a given origin–
destination pair for a given hazmat, transport mode, and vehicle type. Thus, for
each shipment order, this problem focuses on a single commodity and a single
origin–destination route plan. Since these plans are often made without taking
into consideration the big picture, certain links of the transport network tend to
be overloaded with hazmat traffic. This could result in a considerable increase
of accident probabilities on some road links as well as leading to inequity in
the spatial distribution of risk. Although large-scale hazmat carriers are known
to consider transport risk in their routing and scheduling decisions (Verter and
Erkut, 1997), transport costs remain as the carriers’ main focus.
In contrast, the government (municipal, state/provincial, or federal) has to
consider the global problem by taking into account all shipments in its juris-
diction. This leads to a harder class of problems that involve multicommodity
and multiple origin–destination routing decisions. In addition to the total risk
imposed on the public and environment, a government agency may need to
consider promoting equity in the spatial distribution of risk. This becomes cru-
cial in the event that certain population zones are exposed to intolerable levels
of risk as a result of the carriers’ routing and scheduling decisions. The gov-
ernments’ task is further complicated by the need to keep the transport sector
 Ch. 9. Hazardous Materials Transportation 581

economically viable – despite the regulations to ensure public safety – since

dangerous goods shipments are an integral part of our industrial lifestyle.
Hazmat local route planning has attracted the attention of many OR re-
searchers. The existing local route planning models cover a wide area that
includes different transport modes: road (e.g., Akgün et al., 2000; Kara et al.,
2003), rail (e.g., Glickman, 1983; Verma and Verter, 2007), water (e.g., Iakovou
et al., 1999; Iakovou, 2001); deterministic (e.g., Batta and Chiu, 1988; ReVelle
et al., 1991) or stochastic models (e.g., Miller-Hooks and Mahmassani, 1998;
Erkut and Ingolfsson, 2000); and single objective (Erkut and Verter, 1998;
Erkut and Ingolfsson, 2005) or multiple objective models (e.g., Sherali et al.,
1997; Marianov and ReVelle, 1998). Tables 2(a–d) provides a more complete
list of references.
The local routing models fail to capture the dynamic nature of transport
risk factors at the tactical level (e.g., traffic conditions, population density, and
weather conditions). Moreover, most of these risk factors cannot be known a
priori with certainty. They are both time-dependent and stochastic in nature;
i.e., they are random variables with probability distribution functions that vary
with time. Therefore, the local routing and scheduling problem is best mod-
eled as a path selection problem in a stochastic time-varying network (see, for
example, Hall, 1986; Fu and Rilett, 1998; Miller-Hooks and Mahmassani, 1998;
Miller-Hooks, 2001).
The global route planning problem has attained relatively little attention in
the literature, much less compared to the local route planning problem. The re-
sults in this area include the works of Gopalan et al. (1990b), Lindner-Dutton
et al. (1991), Marianov and ReVelle (1998), and Iakovou et al. (1999). The
works of Akgün et al. (2000) and Dell’Olmo et al. (2005) on the problem of
finding a number of spatially dissimilar paths between an origin and a destina-
tion can also be considered in this area.
The rest of this section provides a discussion on the known models and so-
lution algorithms pertaining to the three problem categories discussed above.

4.1 Local routing problems

As we have discussed in Section 3, almost all approximate versions of the

path evaluation functions listed in Table 3 are additive and satisfy the optimal-
ity principle (i.e., Axiom 2). Therefore, the static, deterministic and single ob-
jective local routing problems that minimize those evaluation functions reduce
to the classical shortest path problem. Consequently, a label-setting algorithm
(e.g., Djikstra’s algorithm) can simply be applied to find an optimal route.
Most of the exact versions of these path evaluation functions, on the other
hand, do not satisfy Axiom 2. Therefore, Djikstra’s algorithm cannot be ap-
plied directly to find an optimal route. Kara et al. (2003) proposed a simple
modification of Djikstra’s algorithm to find a route that minimizes the ex-
act version of the path incident probability. The modification relies on the
adjustment of the link attribute that is used to update the node label and
582 E. Erkut, S.A. Tjandra and V. Verter

the scanning process. The algorithm is called the impedance-adjusting node

labeling shortest path algorithm and is explained briefly as follows. Let P =
{(i1  i2 ) (i2  i3 )     (in−1  in )} with i1 the origin node and in the destination
node, and let q(ik ) denote the probability of safely arriving at node ik of P.
From (3.12), we obtain q(ik ) = q(ik−1 )(1 − pik−1 ik ) for k = 2     n, where
pik−1 ik denotes the incident probability of (ik−1  ik ) and q(i1 ) := 1. The at-
tribute aij of each link (i j) is defined as aij := q(i)pij . During the scanning
process of node i, aij for each (i j) is recomputed. This new value is used
to update the node label θj of node j. If the current value of θj is greater
than θi + aij , then we set θj := θi + aij , q(j) := q(i)(1 − pij ) and update
the predecessor of node j. This modified algorithm has the same computa-
tional complexity as that of Djikstra’s. Kara et al. (2003) also proposed the
impedance-adjusting link labeling algorithm to minimize the path population
exposure. This algorithm eliminates the errors resulting from double-counting
of population exposure, which is caused by the network topology. Using a sim-
ilar modification technique to the impedance-adjusting node labeling shortest
path algorithm, Djikstra’s algorithm can be used to solve the local routing prob-
lem with the exact version of perceived risk, the expected disutility, and the
mean-variance path evaluation functions.

4.1.1 Rail transport

A significant majority of the literature on hazmat routing focus on road ship-
ments. This is not surprising, since trucks account for the largest percentage of
hazmat shipments, as discussed in Section 1. Although train shipments can
reach comparable levels to truck shipments from a total tonnage perspective
(particularly in Europe and Canada), they received considerably less atten-
tion from researchers. Remarkably, the literature on marine, air, and pipeline
transport of dangerous goods is in its infancy. McClure et al. (1988) pointed
out a number of differences between rail and highway routing of hazmat trans-
portation. Rail infrastructure is typically owned and maintained by private rail
companies. Consequently, railroad networks are sparse and do not contain
as many potential alternative routes as highway networks. More importantly,
railroads do not have tracks circumventing major population centers that are
comparable to interstate beltways around metropolitan areas. A given ship-
ment is likely to be handled by more than one railroad carrier, whereas truck
shipments are usually limited to a single company. The rail carriers are mo-
tivated to maximize their portion of the movement. In a recent paper, Verma
and Verter (2007) highlighted additional differences between the two transport
modes. A train usually carries nonhazardous and hazardous cargo together,
whereas these two types of cargo are almost never mixed in a truck shipment.
Furthermore, a rail tank car has roughly three times the capacity of a truck-
tanker (80 tons and 25–30 tons respectively) and the number of hazmat railcars
varies significantly among different trains. Another important characteristic of
trains is the possibility of incidents that involve multiple railcars. Verma and
Verter (2007) noted that there is an average of about one major railroad ac-
 Ch. 9. Hazardous Materials Transportation 583

cident per year during the 1990–2003 period in the United States. Thus, there
is a need for the development of risk assessment and routing procedures that
incorporate the differentiating features of railroad hazmat shipments.
The academic literature has mostly focused on the comparison of rail and
road from the viewpoint of hazmat transport risk. For example, Glickman
(1988) observed that the accident rate for significant spills (when release
quantities exceed 5 gallons or 40 pounds) is higher for truck tankers than
for rail tank cars and that rail tank cars are more prone to small spills.
Saccomanno et al. (1989) showed that the safer mode varies with the haz-
mat being shipped and differing volumes complicate comparison between the
two transport modes. Leeming and Saccomanno (1994) reported that although
hazmat railway shipments pose more risk to residents in the vicinity of railroad
tracks, the total risk of these two transport modes does not differ significantly.
Their conclusion is based on a single case study in England. In summary, there
is no consensus among researchers with regards to the dominant transport
mode in terms of public and environmental safety.
Over the past three decades, railroad industry has focused on reducing the
frequency of tank car accidents as well as the likelihood of releases in the event
of an accident – rather than routing and scheduling of trains with potentially
hazardous cargo. The industry’s most recent initiatives have aimed at improv-
ing tank car safety at the design stage. By studying the risks associated with
nonpressurized materials, Raj and Pritchard (2000) report that the DOT-105
tank car design constitutes a safer option than DOT-111. Barkan et al. (2000)
showed that tank cars equipped with surge pressure reduction devices expe-
rienced lower release rates than those without this technology. Barkan et al.
(2003) undertook a study to identify proxy variables that can be used to predict
circumstances most likely to lead to a hazmat release accident. They concluded
that the speed of derailment and the number of derailed cars are highly corre-
lated with hazmat release.

4.2 Multiobjective approaches to local routing

As discussed in Section 1, hazmat transportation is multiobjective in nature

with multiple stakeholders. In general, there is no solution that simultaneously
optimizes all the conflicting objective functions in a multiobjective problem.
Instead, a set of nondominated solutions (or Pareto-optimal solutions) can be
determined. A Pareto-optimal solution is one where we cannot improve on an
objective without worsening at least one other objective. Local route planning
often involves finding the set of Pareto-optimal routes between a given origin–
destination pair. In the event that the decision maker’s preferences among the
conflicting objectives are available in advance, the problem can be reduced to
a single objective optimization problem (via utility theory). The most preferred
solution can then be identified among the Pareto-optimal solutions so as to max-
imize the preference function of the decision maker.
584 E. Erkut, S.A. Tjandra and V. Verter

Nembhard and White (1997) considered the problem of determining the

most preferred path that maximizes a multiattribute, nonorder-preserving
value function both with and without intermediate stops. For the no-stop case,
the problem is solved approximately by applying the dynamic programming al-
gorithm as if a subpath of an optimal path were always optimal (i.e., by using
an approximate method on the exact problem). The intermediate-stop case
is solved approximately by approximating the nonorder preserving criterion
with the linear order-preserving criterion and by properly applying the dynamic
programming algorithm (i.e., by using an exact method on the approximated
problem). Marianov and ReVelle (1998) proposed a linear optimization model
to find the routes that minimize both the cost and the exact version of the prob-
ability of accident. The weighted sum technique is used to solve the biobjective
problem and to approximate the set of Pareto-optimal routes. The associated
weighted, single objective problem can thus be solved by simply applying the
classical shortest path algorithm. Tayi et al. (1999) dealt with the cost equity
and feasibility problem in hazmat routing, where each edge of the network is
associated with a vector of costs incurred by different zones due to an accident
along that edge. The zones represent the community clusters, and each com-
ponent of the cost vector represents the impact of an accident on a zone. The
notion of cost equity is represented by six objective functions, including mini-
mization of the average cost path, the maximum cost path, and the imbalanced
cost path.
As discussed earlier, many transport risk factors involve considerable un-
certainty, which increases the difficulty of routing decisions. Two methods
that are frequently used in incorporating uncertainty are mean-risk (e.g.,
Markowitz, 1987; Ogryczak and Ruszczynski, 2002) and stochastic dominance
(e.g., Yitzhaki, 1982; the survey by Levy, 1992). The mean-risk criterion is
based on comparing only two values: the mean, representing the expected
outcome; and the risk, a scalar measure of the variability of outcomes (e.g.,
variance and semivariance). Mean–variance (MV) criterion is probably the
most well-known mean-risk criterion. It states that if E(v(P1 ))  E(v(P2 ))
and Var(v(P1 ))  Var(v(P2 )) with at least one strict inequality, then v(P1 )
is MV-strictly smaller than v(P2 ), where v(P) is an evaluation function that
operates on path P.
Stochastic dominance (SD) criterion, on the other hand, considers the entire
probability distribution rather than just the two moments. It uses the cumula-
tive distribution function (CDF) as a basis for comparison. Let FP1 and FP2
be the CDFs of two random variables v(P1 ) and v(P2 ). The first- and second-
order stochastic dominance (FSD and SSD) are defined as follows. A random
variable v(P1 ) is strictly smaller, with respect to FSD, than a random variable
v(P2 ), if FP1 (t)  FP2 (t) for all values of t, and at least one of the inequalities
holds strictly. If two CDFs do not intersect, then one of them should stochasti-
cally dominate the other, regardless of their variances. Furthermore, a random
variable v(P1 ) is strictly smaller, with respect to SSD, than a random variable
t
v(P2 ), if −∞ (FP1 (ω) − FP2 (ω)) dω  0 for all values of t, and at least one the
 Ch. 9. Hazardous Materials Transportation 585

inequalities holds strictly. For SSD, the two CDFs may intersect, but the total
accumulated area between FP1 and FP2 must stay nonnegative up to any t. FSD
implies SSD but not vice versa.
Figure 17(a) shows that the distribution FP1 is above distribution FP2 every-
where, and therefore, the probability of “t or less” is higher under FP1 than FP2 .
In Figure 17(b), if the two distributions cross within the range of t, then the
FSD does not hold, but SSD holds. Figure 17(c) shows that v(P1 ) is nei-
ther FSD nor SSD smaller than v(P2 ) and vice versa. Mean–variance crite-
rion offers a much simpler computational tool than SD criterion. However,
a Pareto-optimal solution with respect to the MV criterion may be stochasti-
cally dominated by other feasible solutions if the normality of distributions is
not guaranteed (see, e.g., Yitzhaki, 1982; Ogryczak and Ruszczynski, 2002). On
the other hand, as the CDFs of v(P1 ) and v(P2 ) (or their integration) have to
be compared for every t, the stochastic dominance itself is actually a multiob-
jective model with a continuum of criteria. The stochastic dominance criterion
usually leads to large efficient sets, and it does not provide us with a simple
computational tool.
The problem with the efficient set becomes worse in the multiobjective rout-
ing problem, as the number of Pareto-optimal solutions can be exponential in
the number of nodes (Hansen, 1980). To reduce the size of this efficient set,

Fig. 17. (a) P1 is FSD P2 ; (b) P1 is not FSD P2 but P1 is SSD P2 ; (c) P1 is not SSD P2 and vice versa;
(d) P1 dominates P2 for q0  01382.
586 E. Erkut, S.A. Tjandra and V. Verter

Wijeratne et al. (1993) proposed a two-stage evaluation procedure for normally

distributed path evaluation functions. This procedure includes a probability pa-
rameter that allows the analyst or the decision maker to control the degree to
which a comparison deviates from the FSD criterion. That is, a path P1 domi-
nates P2 if either of the following occurs:
• Primary comparison rule: both the mean and the variance of v(P1 ) are
smaller than those of v(P2 ) (i.e., MV criterion).
• Secondary comparison rule: the mean of v(P1 ) is smaller, the variance
of v(P2 ) is smaller, and the CDF of v(P1 ) exceeds the CDF of v(P2 )
for probability values greater than (1 − q0 ).
The higher the value of q0 , the smaller the size of the efficient set. However,
a small set may exclude some interesting Pareto-optimal paths. Consider the
following small example. Suppose there are two paths P1 and P2 from an origin
to a destination, where the mean and standard deviation of v(P1 ) and v(P2 ) are
(176; 64) and (213; 30), respectively. Figure 17(d) shows that MV and FSD
criteria do not hold. By applying the criterion of Wijeratne et al. (1993), path
P1 will dominate P2 for q0 > 01382.
Although the example involves a single evaluation function v(P), observe
that the incorporation of uncertainty results in a multiobjective problem.
Wijeratne et al. (1993) proposed a simple procedure to deal with a stochastic
multiobjective routing problem (or with a mixture of deterministic and stochas-
tic path evaluation functions). We illustrate this procedure by a small example.
Suppose that there are two stochastic path evaluation functions v1 (P), v2 (P),
one deterministic path evaluation function v3 (P) (all of these functions are to
be minimized) and 4 paths to be compared. Hence, the set of feasible paths is
P = {P1  P2  P3  P4 }. The comparison is done separately for each evaluation
function, where the user-controlled probability parameter q0 may be different
for each stochastic evaluation function. Suppose we find (after applying the
two-stage evaluation procedure) that with respect to v1 (P), the set P can be
partitioned into a ranked set P 1 = {(P1 ) (P2  P3 ) P4 }, which means P1 domi-
nates all other paths, P2 is indifferent to P3 , and both P2 and P3 dominate P4 .
Furthermore, with respect to v2 (P), the set P may be partitioned into a ranked
set P 2 = {(P1  P3 ) P4  P2 }. Suppose that v3 (P1 ) = 100, v3 (P2 ) = 150,
v3 (P3 ) = 50, and v3 (P4 ) = 100, resulting in P 3 = {P3  (P1  P4 ) P2 }. We
can thus combine the relative ranking for each path to create a ranking vec-
tor of evaluation functions for each path: Path P1 : (1 1 2); Path P2 : (2 3 3);
Path P3 : (2 1 1); Path P4 : (3 2 2). (For the deterministic evaluation function,
one may put its value, instead of the relative ranking directly in the ranking
vector.) The final step is to examine this set of ranking vectors to eliminate
the dominated paths. If we require strict dominance across all evaluation func-
tions, we obtain two Pareto-optimal routes: P1 and P3 .
Turnquist (1993) assumed that both accident probability and population
exposure are stochastic. He studied the problem of identifying a set of Pareto-
optimal routes with the following objectives: minimize the incident rate; min-
 Ch. 9. Hazardous Materials Transportation 587

imize the population exposed within a certain distance of the roadway; and
minimize the travel distance. Turnquist used the distribution functions of each
Pareto-optimal path on each criterion to highlight the trade-offs among the
Pareto-optimal solutions.
There are very few static and stochastic routing models (either single or
multiobjective) in the literature for hazmat transportation. In addition to
Wijeratne et al. (1993) and Turnquist (1993), the mean–variance models pro-
posed by Sivakumar and Batta (1994) and Erkut and Ingolfsson (2000) are
noteworthy (see the discussion on these papers in subsection “Mean–variance
model” of Section 3.3.2). There are static, stochastic path finding models
that are designed for other transportation applications (e.g., Frank, 1969;
Mirchandani, 1976; Kulkarni, 1986; Corea and Kulkarni, 1993), which the
reader may find useful. Nonetheless, the dynamic, stochastic routing is more
relevant to hazmat transportation, which we discuss in the next section.

4.3 Local routing and scheduling problems

The traffic conditions and other risk factors in hazmat transportation net-
works (e.g., incident probabilities and population exposure) often vary with
time and can at best be known a priori with uncertainty. For example, for
a hazmat truck, the travel time and the accident probability on certain road
segments can be uncertain and depend on traffic congestion, weather condi-
tions, and road conditions during the vehicle’s trip across those links. Hence,
the transport risk and arrival time at the destination can vary with the dispatch
schedule from the origin. Also, allowing the vehicle to stop during its trip in
order to avoid peak risk periods on certain road segments can be an effective
strategy to reduce the total transport risk (Erkut and Alp, 2006). To represent
this phenomenon appropriately, the transport network should be modeled as
a stochastic, time-varying (STV) network.
In an STV network, the link attributes (such as travel times, incident prob-
abilities, and population exposure) are represented as random variables with
a priori probability distributions that vary with time. STV network-based mod-
eling has been an important and well-researched topic since the late 1980s (see,
e.g., Hall, 1986; Fu and Rilett, 1998; Miller-Hooks and Mahmassani, 1998;
Miller-Hooks, 2001). Most of the existing results are devoted to the Intelligent
Transportation System (ITS), and only some of them are designed specifically
for the hazmat transportation problem (e.g., Bowler and Mahmassani, 1998;
Miller-Hooks and Mahmassani, 1998). The prevailing studies can be classified
into three different groups:
1. A priori optimization: the optimal routes are chosen before the travel be-
gins. Hence, an update on the routing decision en-route is not allowed.
2. Adaptive route selection: the routing decision is subject to change en-route
based on the realization of the estimated data.
588 E. Erkut, S.A. Tjandra and V. Verter

3. Adaptive route selection with real-time updates: the routing decision is

subject to change en-route due to real-time updates of the traffic data
followed by re-optimization procedures.
In the following, we will discuss some of the results in each class that can be
applied to hazmat transportation.

4.3.1 A priori optimization

This class of problems assumes that the optimal route is chosen before trip
begins. Hence, an update on the routing decision en-route is not allowed. All
routing decisions in static (time-invariant) networks fall into this category.
Hall (1986) showed that in STV networks, one cannot simply set the random
arc travel times to their expected values and identify the shortest (expected)
travel time by applying standard shortest path algorithms based on Bellman’s
equation (Bellman, 1958), such as Djikstra’s algorithm. The expected travel
time on an arc in STV networks depends on the arrival time of the vehicle at
the beginning of that arc. A partial route with a higher expected travel time
might be selected, if this choice results in a preferable outcome in the rest of
the route. This is demonstrated by the numerical example in Figure 18.
The objectives are to minimize the expected total travel time and to mini-
mize the expected total risk as defined by the expected total number of exposed
individuals. Suppose the hazmat truck must leave node O at 15:00. On the way
to node T , arc e1 has both, the lowest expected travel time, and the lowest ex-
pected population exposure (120 minutes and 100 individuals as opposed to
125 minutes and 120 individuals on arc e2 ). Hence, Bellman’s principle would
include arc e1 in the optimal path. However, note that a vehicle traveling on

Fig. 18. An STV network for the fastest and least risk path problem.
 Ch. 9. Hazardous Materials Transportation 589

arc e2 has a higher probability of arriving at node T before 16:45 (0.3 probabil-
ity as opposed to zero probability on arc e1 ). Hence, the total expected travel
time and the total expected population exposure via e2 are lower (03(90 +
60) + 07(140 + 120) = 227 minutes and 03(120 + 50) + 07(120 + 200) = 275
individuals as opposed to 240 minutes and 300 individuals).
Hall (1986) proposed an exact, nonpolynomial algorithm that combines a
branch-and-bound technique with a k-shortest paths algorithm to find the
fastest path in STV networks. This algorithm, however, applies only to acyclic
networks or to cyclic networks with First-In First-Out (FIFO) travel times. (We
say that travel times are FIFO if they are nondecreasing functions of time;
i.e., if Vehicle A leaves before Vehicle B, Vehicle A will arrive no later than
Vehicle B.) Miller-Hooks and Mahmassani (2000) extended Hall’s model to
allow cycles or non-FIFO travel times. They proposed a time-dependent label-
correcting algorithm to solve this fastest path problem. Under the assumption
that travel times are continuous functions of time, Fu and Rilett (1998) pro-
posed a heuristic algorithm based on the k-shortest path algorithm to solve
the fastest path problem without the FIFO assumption. The differentiating
feature of their model is the propagation of mean and variance of travel time
along a path in the process of determining the fastest path.
Chang et al. (2005) adapted the continuous-time mean and variance propa-
gation method of Fu and Rilett (1998) to discrete-time intervals and minimized
the total cost as well as the total travel time. The path evaluation functions
(except the total travel time) of two paths in STV networks are comparable
at a node only if the arrival times of those paths at this node are the same.
This condition, however, implies a large efficient set, as it may be unlikely that
two paths arrive at a node at precisely the same time. To tackle this prob-
lem, Sulijoadikusumo and Nozick (1998) and Chang et al. (2005) suggested
a time-window criterion: two paths are comparable only if their arrival times
Pj
are “close enough” as defined by the analyst/decision maker. Suppose Yi , the
arrival time at node i along a path Pj , is normally distributed. The probability
that the difference of two path travel times is less than or equal to a predefined
time window Δ can be approximated as
 P P 
 P1
 P2  Δ − (E[Yi 1 ] − E[Yi 2 ])
p Yi − Yi  Δ =
P P
Var[Yi 1 ] + Var[Yi 2 ]
 P P 
−Δ − (E[Yi 1 ] − E[Yi 2 ])
− 
P P
Var[Yi 1 ] + Var[Yi 2 ]
where (z) denotes the cumulative distribution function of a standard normal
P P
random variable. If p(|Yi 1 − Yi 2 |  Δ)  δ, where δ is the pre-specified
threshold, then these two paths are comparable at node i. If the two paths are
comparable, then the stochastic comparison methods discussed in the previous
subsection can be used to choose the preferred path.
590 E. Erkut, S.A. Tjandra and V. Verter

4.3.2 Adaptive route selection

When traveling along a network, the motorist gathers new information that
can be useful in making better routing decisions. For example, the arrival time
at a node can be used in making a choice among the partial emanating from
that node. This is called adaptive route selection. The optimal route depends
on intermediate information concerning past travel times, road and weather
conditions and hence, a single (and simple) path is not adequate.
Hall (1986) showed that the optimal adaptive route in STV networks that
minimizes the expected travel time is not a simple path but an acyclic subnet-
work (called a hyperpath) that represents a set of routing strategies (see, e.g.,
Nguyen and Pallottino, 1986). The adaptive route specifies the road link to be
chosen at each intermediate node, as a function of the arrival time at the node.
As an illustration, consider the example depicted in Figure 19.
The hazmat truck is to leave node O at 15:00. The a priori expected travel
times of two paths P1 := {e1  e2 } and P2 := {e1  e3 } are 03(90+60)+07(140+
120) = 227 minutes and 03(90 + 100) + 07(140 + 30) = 176 minutes,
respectively. The associated a priori expected total risks for P1 and P2 are
03(100 + 100) + 07(150 + 80) = 221 and 03(100 + 200) + 07(150 + 50) = 230
individuals at risk, respectively. Hence, the a priori fastest path is path P2 , and
the a priori least risk path is P1 . However, if the motorist is permitted to select
the rest of the path upon arrival at node T , we will obtain the following routing
strategy:
• Travel time: If the arrival time at node T is 16:30, take arc e2 with
a travel time of 60 minutes. If the arrival time at node T is 17:20, take
arc e3 with a travel time of 30 minutes. The expected travel time for the
adaptive fastest path from O to D is 03(90 + 60) + 07(140 + 30) = 164
minutes. The associated total risk is 03(100 + 100) + 07(150 + 50) =
200 individuals at risk.
• Total risk: The routing strategy is the same as for that of the adaptive
fastest path.
The resulting hyperpath of the optimal adaptive routing strategy, depicted as
a decision tree, is shown in Figure 20.
It is, in general, quite unlikely that the optimal adaptive routing strategies of
different objectives coincide. In this case, the multiobjective version of the la-
bel correcting and Stochastic Decreasing Order of Time (SDOT) algorithms
from Miller-Hooks (2001) can be used to generate a set of Pareto-optimal
adaptive routing strategies.

4.3.3 Adaptive route selection with real-time updates

The recent advances in information and communication technologies, such
as satellite-based Automatic Vehicle Location (AVL) and mobile phones, en-
able the driver and dispatch center to obtain and exchange real-time infor-
mation. Satellite-based AVL is a computer-based vehicle tracking system that
uses signals from satellite systems, such as Navstar Global Positioning System
 Ch. 9. Hazardous Materials Transportation 591

Fig. 19. An STV network for the adaptive routing strategy.

Fig. 20. The resulting hyperpath of the adaptive routing strategy, depicted as a decision tree.

(GPS), to identify a vehicle’s location. Mobile communication systems such as

cellular phones, paging systems, and mobile satellite communication systems,
provide two-way communication between the driver and the dispatch center
or among drivers. These AVL and mobile communication systems enable the
driver and the dispatch center to monitor and/or change the route of vehicles
based on real-time information.
These technological advances are challenging OR researchers to develop
routing models and robust optimization procedures that are able to respond
quickly to changes in the data. In this real-time environment, the quality of the
decision depends not only on the appropriateness of the decision, but also on
its timeliness (Seguin et al., 1997). Another main issue in this area, besides
592 E. Erkut, S.A. Tjandra and V. Verter

route planning, is the data updating procedure. New real-time information

obtained by the dispatch center is used to update the estimation of future val-
ues of some network attributes (e.g., travel times, incident probabilities, and
population in the impact area). However, this information is of limited use
if the information is about parts of the transport network that are far away
from the current location of the vehicle (either spatially or temporally). There-
fore, either a spatial or a temporal discounting procedure must be applied
before this real-time information is used to update the estimates of network
attributes (see, e.g., Hoffman and Janko, 1990; Koutsopoulos and Xu, 1993;
Yang, 2001).
We observe a lack of papers in this area that consider both adaptive routing
decisions and data updates based on real-time information. Moreover, none
of the prevailing studies are designed specifically for hazmat transportation
problems. Koutsopoulos and Xu (1993) proposed an information discounting
procedure for travel times in finding the shortest path in STV networks with
an FIFO assumption. For temporal discounting, they used the results from
Hoffman and Janko (1990), where the ratio of the historical mean over its
current travel time is used to estimate the future travel times on the same arc.
Suppose that the route planning is defined in discrete time T := {tk : k =
0     K} with tk+1 := tk + Δ, k = 0     K − 1. If we denote the travel time
ratio on arc (i j) at time t ∈ T by δijt , then
λ̄ijt
δijt := 
λijt
where λ̄ijt is the historical average travel time on (i j) at time t and λijt is
the associated actual travel time. This ratio is set to 1.0 when real-time infor-
mation for an arc is not available. To incorporate changes in neighboring arcs,
a smoothed mean ratio is computed as
1 
δijt := δklt 
|Aij |
(kl)∈Aij

where Aij is a set of all adjacent arcs of (i j). The new estimated travel time
λijt  on arc (i j) at a future time period t  = t + t     tK is then given by

λ̄ijt 
λijt  := 
δijt
Koutsopoulos and Xu (1993) claimed that actual information obtained on arc
(i j) will be less useful, as either the distance between the origin node and
node i increases or the variability of the historical travel time on (i j) in-
creases. The new estimation of travel time (after being temporally and spatially
discounted) on arc (i j) is

λ∗ijt0 +Psi (t0 ) = λ̄ijt  + e−θσijt  Psi (t0 ) (λijt  − λ̄ijt  )

 Ch. 9. Hazardous Materials Transportation 593

where Psi (t0 ) is the shortest travel time from the origin node s to node i de-
parting from the origin at time t0 , θ is a positive constant scalar that can be
adjusted to produce a good fit between the estimated and actual travel times,
t  −  t0 +Pij (t0 )  t  , and σijt  is the standard deviation of historical travel
time λ̄ijt  . The larger the value of Psi (t0 ) and σijt  , the larger the discounting
of the actual information. This travel time updating procedure is incorporated
in the label setting algorithm to find the shortest routes from an origin s. For
each arc (i j) out of the last permanently labeled node i, calculate (if node j is
not yet permanently labeled):

Psj (t0 ) = min Psj (t0 ) Psi (t0 ) + λ̄ijt 

+ e−θσijt  Psi (t0 ) (λijt  − λ̄ijt  ) 
Set the label of a node with the smallest Psj (t0 ) to permanent and update its
predecessor node, which is needed to construct a path from the origin. The
process is repeated until all nodes are labeled permanently.
Yang (2001) discussed an adaptive route selection with real-time updates in
discrete STV networks, which is applied to ITS. To update the travel times,
Yang considered both spatial and temporal information discounting, which
are determined by spatial and temporal depth. The spatial depth determines
the maximum reachable distance, with respect to the number of arcs, from
the current node. The temporal depth is defined as the maximum number
of time periods in which the information is still considered valuable. Fur-
thermore, Yang also proposed two re-optimization algorithms to find the new
adaptive route strategy that incorporates the new estimated travel times. The
re-optimization algorithms are based on the ELB (Expected Lower Bound) al-
gorithm of Miller-Hooks and Mahmassani (2000) and the SDOT algorithm of
Miller-Hooks (2001). These re-optimization algorithms, called “adapted ELB”
and “adapted SDOT,” assume that the realization of the travel time must coin-
cide with one of the possible values known a priori. Hence, it is assumed that
the analysts are able to predict all possible values of future travel times, which
is not realistic in many cases.

4.4 Global routing problems

The global route planning problem typically belongs to a government agency

charged with the management of hazmat shipments within and through its ju-
risdiction. Although the transportation industry has been deregulated in many
countries, hazmat transportation usually remains as part of the governments’
mandate mainly due to the associated public and environmental risks. The two
main concerns for a government agency are the total risk and the spatial distri-
bution of risk in its jurisdiction. A number of policy tools are available to the
government in mitigating public risk. These include proactive measures such
as the establishment of inspection stations (Gendreau et al., 2000), insurance
requirements (Verter and Erkut, 1997), and container specifications (Barkan
594 E. Erkut, S.A. Tjandra and V. Verter

et al., 2000) as well as reactive measures such as the establishment of haz-

mat emergency response networks (Berman et al., 2007). Another common
tool for governments is to ban the use of certain road segments by potentially
hazardous vehicles. For an example of such regulation, we refer the reader
to the local authority bylaws section of the Alberta Dangerous Goods Trans-
portation and Handling Act (Government of Alberta, 2002). In the context of
global route planning, the road segments to be closed by the government can
be identified by solving a hazmat network design problem, which we discuss in
Section 4.4.2.
Equity in the spatial distribution of risk can be important for a government
agency for two reasons: (i) the perception of risk inequity frequently results
in public opposition to the routing of vehicles carrying hazmats through the
nearby passageways; and (ii) the overloading of certain road segments with
hazmat flows (i.e., risk inequity) may lead to an increase in the incident proba-
bilities as well as the severity of consequences. The concept of equity has been
studied in the OR literature primarily within the context of undesirable facility
location. Marsh and Schilling (1994) provided a comprehensive review of eq-
uity measures for location problems. Erkut (1993) offered two equity axioms
for location problems and showed that the Gini coefficient and the coefficient
of variation are the only measures that satisfy both of these axioms. Defining
n = number of zones, ti = individual risk at population zone i, and t̄ = average
individual risk, these two equity measures can be represented as follows:

i (ti − t̄)
2
Coefficient of variation = 
  nt̄
j i |ti − tj |
Gini coefficient = 
2n2 t̄
Coefficient of variation evaluates equity in terms of the deviations of the in-
dividual risks from the average. In contrast, Gini coefficient focuses on the
differences between individual risks. Clearly, smaller values of these equity
measures correspond to higher levels of fairness in risk distribution. A value
of zero represents perfect equity, whereas a value of one represents absolute
inequality. Using GIS, Verter and Kara (2001) estimated these two equity
measures for gasoline shipments in Ontario and Quebec under four routing
criterion: minimum length, minimum expected risk, minimum population ex-
posure, and minimum incident probability.

4.4.1 Equity considerations in global route planning

The multiobjective model proposed by Zografos and Davis (1989) was per-
haps the first attempt to explicitly incorporate equity considerations in global
route planning for dangerous goods shipments. Their objectives were to min-
imize the total risk, the risk imposed on special population categories, travel
time, and property damage. Equity is achieved by constraining the capacity
of the road links. Zografos and Davis used pre-emptive goal programming in
 Ch. 9. Hazardous Materials Transportation 595

solving the problem, and demonstrated (using hypothetical data) that forcing
equity could increase the total risk up to 35%.
Gopalan et al. (1990a) proposed an equity constrained shortest path model
that minimizes the total risk of travel between and origin–destination pair,
while maintaining a desired level of equity among disjoint zones of a trans-
portation network. Each zone constitutes a jurisdiction of a government agency
that regulates hazmat transportation. The travel risk associated with road
link (i j) is the sum of risks imposed on the zones in the vicinity of the link. An
origin–destination path is considered equitable if the difference between the
risks imposed on any two arbitrary zones is under a given threshold. This eq-
uitable path definition can be incorporated in the shortest path model through
additional constraints. Gopalan et al. (1990b) developed a subgradient algo-
rithm to solve the Lagrangian dual, which is obtained by relaxing the equity
constraints. They proposed a labeling shortest path procedure to close any
remaining duality gap. The model was applied to a 50-node network from Al-
bany, New York.
Gopalan et al. (1990b) extended their earlier work so as to identify a set of
routes to be utilized for T trips between a single and origin–destination pair. In
this case the equity threshold for a zone pair is the sum of the risk differences
over T trips. Note that the T routes do not need to be distinct in their model.
Gopalan et al. (1990b) proposed a heuristic procedure that repeatedly solves
single trip problems using a Lagrangian dual approach with the gap-closing
procedure, as in Gopalan et al. (1990a). To avoid having T identical routes,
the link risks are modified using information from the previous t routes during
iteration (t + 1). This iterative procedure can easily be adapted to multiple
origin–destination pairs.
In extending Gopalan et al. (1990b), Lindner-Dutton et al. (1991) focused
on finding an equitable sequence of T trips, where the cumulative risk in-
curred by any zone after t < T trips is equitable to that incurred by the other
zones in the previous t trips. Both integer programming and dynamic program-
ming (DP) formulations of this problem were presented. Lindner-Dutton et al.
(1991) showed that a DP approach combined with the relaxation and fathom-
ing methods of the Branch and Bound algorithm (as described in Morin and
Marsten, 1976) could not solve moderate size problems to optimality within
reasonable time. Therefore, they developed five upper bound heuristics to
tackle large problems.
Marianov and ReVelle (1998) proposed a linear optimization model to solve
the global route planning problem that minimizes both total cost and (the exact
version of) accident probability. To introduce equity, they used an upper bound
on the total risk associated with each arc. Similarly, Iakovou et al. (1999) in-
corporated equity through the use of a capacitated transport network model.
Their multicommodity network flow model has two objectives: minimize trans-
port cost and minimize expected risk cost. They used a weighted sum of these
costs in conducting a trade-off analysis. A two-phase solution procedure, simi-
596 E. Erkut, S.A. Tjandra and V. Verter

lar to that of Gopalan et al. (1990a), was proposed. The model was applied to
marine transportation of oil products in the Gulf of Mexico.
The studies on generation of a set of spatially dissimilar (not necessarily
disjoint) paths are also relevant to equity considerations in global route plan-
ning (e.g., Akgün et al., 2000; Dell’Olmo et al., 2005). Iterative penalty method
(IPM), gateway shortest paths method, and minimax method are among the
procedures that can be used to generate such a set of paths set between an
origin–destination pair. However, Akgün et al. (2000) showed that the gate-
way shortest path method may not be suitable for generating dissimilar paths.
They posed the dissimilarity problem as a p-dispersion problem (Erkut, 1990).
In the p-dispersion context, p of m candidate paths are selected so that the
minimum spatial dissimilarity between any pair of selected paths is maximized.
The m candidate paths can be constructed using k-shortest path method or
IPM.
Erkut and Verter (1998) proposed four indexes to measure the dissimilarity
among paths P1 and P2 :
• Arithmetic average of two ratios:
L(P1 ∩ P2 ) L(P1 ∩ P2 )
1− + ;
2L(P1 ) 2L(P2 )
• Geometric average of two ratios:

L(P1 ∩ P2 )2
1− ;
L(P1 )L(P2 )
• Ratio of the intersection length and the length of the longest path:
L(P1 ∩ P2 )
1− ;
max{L(P1 ) L(P2 )}
• Ratio of the intersection length and the length of the union of the two
paths:
L(P1 ∩ P2 )
1− ;
L(P1 ∪ P2 )
where L(P) denotes the length of path P.
Dell’Olmo et al. (2005) provided a multicriteria formulation of the dis-
similar path problem. They used travel distance and transport risk as their
criteria. After finding the Pareto-optimal set of paths, a buffer zone is con-
structed for each path in this set. This buffer zone approximates the impact
area of a hazmat incident. Based on the buffer zones, a dissimilarity index can
be calculated for each pair of paths by replacing L(P) in the above defini-
tions with A(P) that represents the area of the buffer zone around path P.
For example, the average arithmetic dissimilarity index can be defined as
 Ch. 9. Hazardous Materials Transportation 597

1 − A(P1 ∩ P2 )/(2A(P1 )) + A(P1 ∩ P2 )/(2A(P2 )). A subset of maximally dis-

similar paths (spatially speaking) can thus be found, for example, by applying
the p-dispersion method.
The above models can be useful in identifying a global routing plan for a ma-
jor hazmat producer/carrier that takes into account the equitable distribution
of transport risk in a region. However, these models are of little use in the im-
plementation of a comprehensive global transportation plan in a jurisdiction
with multiple carriers since governments have no authority to impose routes
on individual carriers. Yet many governments have the authority to close cer-
tain road segments to hazmat shipments (permanently or during certain hours
of the day), and equity concerns can be incorporated into a hazmat network
design problem. This is an interesting and challenging OR problem that has
not been studied in the past. In the next section we review a closely related
problem: the hazmat network design problem with a risk minimization objec-
tive.

4.4.2 Hazmat transportation network design

Network design problems have wide applications in both transporta-
tion and telecommunication planning (see, e.g., Magnanti and Wong, 1984;
Balakrishnan et al., 1997). It is important to recognize the differentiating char-
acteristics of this problem in the context of dangerous goods shipments. The
transportation infrastructure is built mainly to connect heavily populated areas
and not to avoid them. Therefore, the question becomes which road segments
to close in an existing network rather than identifying the most appropriate
ways to expand the infrastructure. Kara and Verter (2004) provide the follow-
ing definition: given an existing road network, the hazardous network design
problem involves selecting the road segments that should be closed to haz-
mat transport so as to minimize total risk. The carriers will select minimum
cost routes on the designated hazmat network, and they are likely to incur
higher costs due to reduced availability of routes. Hence, this can be consid-
ered a two-level decision problem where the government designates a subset
of the transport network for hazmat transport and carriers select routes on this
subset.
Note that these two levels cannot be considered in isolation. If one were to
select minimum risk routes and offer the union of such routes to the carriers,
the carriers would select minimum cost routes on this network which could
result in much higher risk levels than the government had intended. This can
be illustrated using the example depicted in Figure 21(a) (Erkut and Gzara,
2005). Suppose that hazmat type 1 is to be sent from node 1 to node 8, and
hazmat type 2 is to be sent from node 2 to node 8. Assume that the transport
cost for each commodity is the same.
If the carrier is allowed to route freely, it will select the minimum cost
routes {(1 3) (3 8)} and {(2 5) (5 6) (6 8)} with a total cost of 3 + 3 = 6
units and total risk of 8 + 8 = 16 units. In contrast the minimum risk routes
are {(1 3) (3 6) (6 8)} and {(2 5) (5 6) (6 7) (7 8)} with a total risk of
598 E. Erkut, S.A. Tjandra and V. Verter

(a) (b)
Fig. 21. (a) Multicommodity hazmat transport network design problem; (b) The feasible roads for
routing cost minimization.

7 + 4 = 11 units. Figure 21(b) shows the union of the two minimum risk
paths. If the government designates this network as the hazmat transport net-
work, but allows the carrier to choose its routes, it will select the minimum
cost routes {(1 3) (3 6) (6 8)} and {(2 5) (5 6) (6 8)} with a total cost of
8 units and total risk of 13 units. This risk is higher than what the government
anticipates. As this example demonstrates, the design problem cannot be sim-
plified to a one-level risk minimization problem, and the government must take
into account the cost-minimizing behavior of carriers in designing the network.
The hazmat transportation network design problem has received the atten-
tion of researchers only recently. Kara and Verter (2004) proposed a bi-level
integer linear programming formulation for this design problem that involves
multiple types of hazmats. Their aim is to design a transport network so that
the total risk resulting from the carriers’ route choices is minimized. At the
outer-level, risk is measured as the total number of people exposed to hazmat
transport incidents. The inner-level problem represents the carriers’ routing
decisions on the available transport network so as to minimize their cost. This
problem is represented by the linearized Karush–Kuhn–Tucker (KKT) condi-
tions of its LP relaxation. As a result, the bi-level integer programming (IP)
problem is transformed into a single-level mixed integer programming prob-
lem. The proposed model is solved by using CPLEX and applied to the hazmat
transport network in Western Ontario, Canada. Kara and Verter demonstrate
that carriers can benefit from the government’s efforts and involvement in the
regulation of dangerous goods shipments.
 Ch. 9. Hazardous Materials Transportation 599

Erkut and Gzara (2005) considered a bi-level bi-objective (cost and risk
minimization) network design problem similar to that discussed by Kara and
Verter (2004). They proposed a heuristic algorithm that exploits the network
flow structure at both levels, instead of transforming the bi-level IP problem to
a single-level formulation. As a result, they achieved a significant increase in
the computational performance.
Erkut and Alp (2007) posed the minimum risk hazmat network design prob-
lem as a Steiner tree selection problem. This topology takes away the carriers’
freedom in route selection and simplifies the bi-level problem to a single level.
However, it also results in circuitous (and expensive) routes. To avoid an eco-
nomically infeasible solution, they suggested adding edges to the Steiner tree.
They proposed a greedy heuristic that adds shortest paths to the tree so as to
keep the risk increase to a minimum. They also posed a bi-objective version of
the problem to minimize cost and risk, and solved it using a weighted additive
objective. Their approach allows the decision maker to determine the density
of the hazmat network where the options range from a tree to a completely
connected network.
Verter and Kara (2005) provided a path-based formulation for the haz-
ardous network design problem. Their main modeling construct is a set of al-
ternative paths for each shipment. This facilitates the incorporation of carriers’
cost concerns in regulator’s risk reduction decisions. Paths not economically
viable for carriers can be left out of the model. Alternative solutions to the
network design problem can be generated by varying the number of routing
options included in the model. To this end, Verter and Kara use pre-specified
thresholds, e.g., for the maximum acceptable additional travel distance com-
pared to the shortest path. Therefore, each solution corresponds to a certain
compromise between the regulator and the carriers in terms of the associated
transport risks and costs. Information about the nature of the cost-risk trade
off can facilitate negotiation between the two parties. By using a GIS-based
model of Quebec and Ontario, the authors demonstrate that their path-based
formulation can be used for identifying road closure decisions that are mutu-
ally acceptable.

5 Facility location and transportation

Hazmat shipments often originate from facilities that themselves are poten-
tially harmful to public and environmental safety, such as petroleum refineries
or nuclear power plants. Also, the destinations of hazmat shipments can be
noxious facilities such as gas stations and hazardous waste treatment centers.
The location decisions pertaining to such facilities have a considerable effect
on the routing of hazmat shipments. Therefore, integration of facility loca-
tion and routing decisions can be an effective means to mitigate the total risk
in a region where hazmats are processed and transported. It is interesting to
600 E. Erkut, S.A. Tjandra and V. Verter

note that, in general, location decisions are considered strategic, whereas rout-
ing decisions are dealt with at the tactical level. However, the risk constitutes
a coupling factor for these decisions in the context of dangerous goods. We
refer the reader to Erkut and Neuman (1989) and Cappanera (1999) for exten-
sive surveys of the location-only literature dealing with undesirable facilities.
In this section, we provide a review of the prevailing studies on integrated lo-
cation and routing models for hazmats.
The location–routing problem (LRP) involves determining the optimal num-
ber, capacity, and location of facilities as well as the associated optimal set of
routes (and shipping schedules) to be used in serving customers. The distribu-
tion of goods from the facilities to the customers can be on a full-truck load or
less than full-truck load basis. In the latter case, routes involving multiple cus-
tomers are commonly used. From the solution method perspective, the LRP is
NP-hard and offers a variety of challenges to OR researchers. The literature
addressing LRP with different real-world applications has evolved since the
late 1960s. Christofides and Elon (1969) were among the first to consider LRP
with multiple customers on each route. The literature surveys on LRP include
Madsen (1983), Balakrishnan et al. (1997), and Min et al. (1998).
Two types of risk need to be taken into account in integrating location and
routing decisions pertaining to hazmat shipments: transport risk, RT , and facil-
ity risk, RF . Figure 22 illustrates these two types of risk. An individual at point x
is exposed to (i) a transport incident on a nearby route segment l of a path P
that involves a vehicle carrying volume vP and (ii) an incident at the hazmat
treatment center at site j with capacity uj . The transport risk, RT Pl (vP  x), can
be determined as a function of the undesirable consequence at point x, taking
into account the impact zone of a hazmat incident on segment l (see Section 3),
and the estimated incident probability. The facility risk, RFj (uj  x), can be de-
termined in a similar way, with site j replacing the route segment l. Let O and D
denote sets of origins and destinations, respectively, POD denote the set of all
utilized paths for each O–D pair (O ∈ O and D ∈ D), and L denote the set
of hazmat facility locations. Assuming additivity of risk, the individual risk at
point x can be determined as
   
R(x) := RTPl (vP  x)+ RFj (uj  x)
O∈OD∈D P∈POD l∈P j∈L

Let A denote the region of interest and POP(x) denote the population density
at point x ∈ A. The total risk in A is

R(A) = R(x)POP(x) dx
x∈A
Now consider a location–routing problem where L = D (e.g., storage loca-
tions for spent nuclear fuel shipments). Let VO denote the hazmat volume at
O ∈ O (e.g., a nuclear power plant) that needs to be transported, and let uD
denote the capacity of a hazmat treatment facility at site D ∈ D. Note that
 Ch. 9. Hazardous Materials Transportation 601

Fig. 22. Individual risk at point x due to transportation and processing of dangerous goods (adapted
from List and Mirchandani, 1991).

D and POD now represent the sets of candidate locations for hazmat treatment
facilities and the set of potential paths for each origin–destination pair, respec-
tively. The set POD may represent the set of available routes on the hazmat
road network designated by the government (see Section 4.4.2). We define two
types of variables:
• binary location variables yD , where
⎧
⎨ 0 if a new hazmat treatment facility is located
yD = in site D
⎩
1 otherwise
• nonnegative continuous flow variables vP representing the quantity of
hazmat shipped along path P.
Thus, the total risk in region A is
    
R(A) := RT
Pl (vP  x)
x∈A O∈OD∈D P∈POD l∈P
 
+ RFD (uD  x)yD POP(x) dx
D∈D
602 E. Erkut, S.A. Tjandra and V. Verter

In addition to the total risk, the costs (i.e., transportation, operation, and
fixed costs) should be also minimized. Let cPT denote the transportation cost per
unit volume of hazmat along path P, cD F denote the (annualized) installation

cost and cDO denote the unit operation cost of a hazmat treatment facility at
site D. The total cost, TC, is determined as
     
T F O
TC := cP vP + cD y D + c D vP 
O∈OD∈D P∈POD D∈D O∈O P∈POD

Also, equity in the spatial distribution of risk due to the location and routing
decisions can be a relevant objective. Risk equity can be enforced, for example,
by minimizing the maximum individual risk in the region, i.e.,

R(A) := max R(x)
x∈A
Hence, a mathematical programming formulation of the capacitated LRP
to minimize the total risk and total cost and to force the risk equity can be
constructed as follows:
min R(A) (5.1)
TC (5.2)

R(A) (5.3)
subject to:
 
vP = VO  for all O ∈ O (5.4)
D∈D P∈POD
 
vP  uD yD  for all D ∈ D (5.5)
O∈O P∈POD
  

R(A)  RT
Pl (vP  x)
O∈OD∈D P∈POD l∈P

+ RFD (uD  x)yD  for all x ∈ A (5.6)
D∈D
yD ∈ {0 1} for all D ∈ D (5.7)
vP  0 for all P ∈ POD and O–D pairs
O ∈ O D ∈ D (5.8)
Constraints (5.4) ensure that all hazmat generated must be shipped out of the
origins, whereas constraints (5.5) stipulate that if a facility at location D is open
(i.e., yD = 1), then total quantity of hazmat to be treated at D cannot exceed
the pre-specified capacity of the facility. Constraints (5.6) are used to incorpo-
rate the risk equity. It is evident from the above model that the hazmat LRP is
multiobjective by nature. The surveys by List et al. (1991), Boffey and Karkazis
(1993), and Cappanera et al. (2004) observed that literature on hazmat LRP is
 Ch. 9. Hazardous Materials Transportation 603

sparse. In this section, rather than duplicating these surveys, we highlight the
important results.
Shobrys (1981) is the first study on hazmat LRP with a focus on selecting
routes and storage locations for spent nuclear fuel shipments. A decomposition
approach is used to separate the routing problem from the location problem.
Two routing objectives are minimized; ton-miles and population exposure-
tons. The associated bi-objective shortest path model identifies a set of Pareto-
optimal paths between each waste source (origin) and each candidate storage
site (destination). The weighted costs associated with each Pareto-optimal path
determine the cost coefficients of the p-median problem that is used to select
the storage site.
Zografos and Samara (1989) considered an LRP with three objectives,
namely minimization of transport risk, minimization of travel times, and min-
imization of disposal risk, to establish locations of a given number of waste
treatment facilities and determine the associated shipment routes. Their model
requires that the hazardous waste at each population center must be disposed
of entirely. Each population center is assigned to its nearest disposal facility.
Moreover, links of the transportation network are capacitated. Pre-emptive
goal programming is used to generate solutions under a few different scenar-
ios.
List and Mirchandani (1991) proposed a hazmat LRP model that simulta-
neously considers total transportation and treatment risk, total transportation
cost, and risk equity. Risk equity is enforced by minimizing the maximum con-
sequence per unit population for all mutually disjoint zones of the transporta-
tion network. Their formulation served as a basis for the model in (5.1)–(5.8).
However, the List and Mirchandani model is more general since it allows for
different types of hazardous materials and treatment technologies. This model
assumes that the impact to point x in a zone Z from a vehicle incident is
inversely proportional to the square of the Euclidean distance between the
vehicle and point x, and the impact is directly proportional to the volume vP
being shipped regardless of material. Hence, the transport risk faced by an
individual at point x is determined as

T
RPl (vP  x) := αvP l − x−2 c(x)π(l) dl
l∈P
where α is a constant of proportionality, c(x) is a likelihood of impact at
point x, and π(l) is the probability of an incident at road segment l. The fa-
cility risk from an incident at a hazardous waste treatment facility at site j of
waste type w with treatment technology t and volume ujwt , RFjwt (ujwt  x), is
determined in a similar way. However, their facilities have unlimited capacity
and the total cost of establishing treatment facilities is bounded by a budget
constraint. Uncertainty is considered in constructing the risk formulations, but
it is not incorporated in solving the example case. Instead, the expected num-
ber of fatalities is used to calculate the risk. The LRP problem is solved using
LINDO. The weighted sum technique is used to study the tradeoffs among
604 E. Erkut, S.A. Tjandra and V. Verter

the objectives in identifying the transportation routes, locating the hazardous

waste treatment facilities, and choosing the treatment technologies.
ReVelle et al. (1991) developed a combined discrete location–routing model
for shipments of spent nuclear fuel that minimizes both transportation cost
and perceived risk. As in Shobrys (1981), the transportation cost is measured
in ton-miles, and the perceived risk is measured using population exposure as
people-tons. The total people-ton of an arc is the product of the number of
people within a certain bandwidth on the arc and the tons of hazardous waste
shipped on that arc. The problem is solved in two stages. In the first stage,
a weighted sum of the arc distance and the number of people in the impact
area around that arc (called hybrid distance) is calculated for every arc in the
network. Floyd’s shortest path algorithm is used to generate (hybrid) shortest
paths for all origin–destination pairs. In the second stage, the location prob-
lem is modeled as a p-median problem, where the coefficients of the objective
function are calculated by taking the product of the tons of spent fuel at the
origin and the hybrid shortest distance from the origin to the destination.
Stowers and Palekar (1993) proposed a bi-objective network LRP with a sin-
gle facility and a single commodity. In a network LRP, the waste facility can
be located anywhere on the network. Two objectives are considered, namely
minimizing the total exposure (minisum) and minimizing the maximum expo-
sure (minimax). The total exposure to a node or to an arc of the network is
represented as a convex combination of location exposure and travel expo-
sure, where the impact area is modeled as a danger circle. Stowers and Palekar
showed that an optimal solution to the minisum and minimax problems with
only travel exposure occurs at a node. The nodal optimality is still valid for
any positive linear combination of travel cost and travel exposure as long as
the travel cost is an increasing function of distance, as in ReVelle et al. (1991).
Moreover, when population is concentrated at nodes only, a finite dominat-
ing set of facility locations can be identified which is guaranteed to contain an
optimal solution.
Giannikos (1998) proposed a multiobjective model for a discrete hazardous
waste LRP that minimizes the following four objectives:
(1) total transportation cost and fixed cost of opening the treatment facili-
ties;
(2) total perceived risk due to the shipment of hazardous waste;
(3) maximum individual risk (to force the risk equity); and
(4) maximum individual disutility due to the treatment facilities.
The disutility imposed on a population center i by the establishment of a treat-
ment facility at site j is a function of the capacity of facility j and the distance
between i and j. The total disutility at population center i is obtained by adding
the disutilities imposed upon i by all treatment facilities. A weighted goal pro-
gramming technique is used to solve the problem.
Cappanera et al. (2004) presented a single objective LRP model that mini-
mizes the total transportation and facility establishment costs. In their model,
 Ch. 9. Hazardous Materials Transportation 605

an arc formulation is given instead of a path formulation as in (5.1)–(5.8). Their

model includes constraints that require both routing and population exposures
for each affected site to remain within given threshold values. Arcs of the
network are incapacitated, but the facilities are capacitated. Cappanera et al.
(2004) consider only a single commodity and seek to find the optimal number
of facilities. By dualizing the capacity constraints, the LRP is decomposed into
location and routing subproblems to obtain a lower bound. To find the upper
bounds, two Lagrangian heuristics, called the Location–Routing heuristic and
Routing–Location heuristic, are proposed.
In closing this section, we note that almost all existing models for hazmat
LRP are static and deterministic. Only the model of List and Mirchandani
(1991) considers different types of hazmats and technology selection for haz-
mat treatment facilities as well as uncertainty in problem parameters. The lack
of multiple hazmat models that consider stochasticity in a time-dependent en-
vironment constitutes an area for further LRP research.

6 Synthesis and future research directions

To summarize the material we have reviewed, Tables 2(a–d) groups the mod-
els into classes distinguished by
• the main aspects of the problem (risk assessment, routing, combined
facility location and routing, and network design),
• transport mode,
• single vs. multiple objectives,
• whether or not stochastic elements are included,
• whether or not time-variant elements are included,
• whether or not GIS is used.
Tables 2(a–d) suggests that the hazmat transportation problems on highways
received the most attention from the operations researchers. In contrast, haz-
mat transportation via air or pipeline, as well as intermodal hazmat transporta-
tion has received almost no attention. From the methodological perspectives,
we observe that:
• global routing problems on stochastic time-varying networks received
no attention despite their relevance and application potential,
• hazmat transportation network design problem which considers all in-
volved parties (government and the carriers) is a relatively young re-
search topic. The most obvious extension of the existing models in this
area is to incorporate uncertainty and consider multiple objectives as
the hazmat transportation problems are highly stochastic in nature and
involve multiple criteria (and players),
• there is an increase on utilizing a GIS either for data input or combined
with optimization models to conduct more realistic risk assessment.
606 E. Erkut, S.A. Tjandra and V. Verter

Erkut and Verter (1995a) reflected on the state-of-the-art as of 1995, and

pointed out a number of directions for future research. In the following ten
years, some of the problem areas proposed in Erkut and Verter (1995a) were
investigated by researchers, whereas many others remained relatively unex-
plored. We discuss some of the underexploited areas discussed in Erkut and
Verter (1995a), as well as other potential problem areas, that can lead to fruit-
ful research.

Risk calculation – probabilities

QRA relies heavily on empirical accident/incident probabilities. However

past data is not very reliable. Using general truck accident data for hazmat
trucks overestimates the accident probabilities. What makes matters worse is
that there is no agreement on general truck accident probabilities and con-
flicting numbers are reported by different researchers. Furthermore, applying
national data uniformly on all road segments of similar type is quite problem-
atic since it ignores hot spots such as road intersections, highway ramps, and
bridges. Researchers need to have access to high quality accident probability
data and empirical or theoretical research that leads to improvements in the
quality of such data would be welcome.

Risk calculation – perceived risks

Given the limitation of QRA, and the fact that public opposition is a func-
tion of perceived risks, perhaps more attention should be paid to quantifying
and modeling of perceived risks. We believe more work is needed to improve
our understanding of how perceived risks change as a function of the haz-
ardous substance, the distance to a hazardous activity, and the volume of the
activity.

Risk calculation – consequences

The second important input in QRA is the population exposed as a result of

an incident. Many past studies used uniform population density along transport
links which is a very blunt approach. A GIS makes it possible to use more pre-
cise population information. However, using census-based population data for
daytime hazmat movements makes little sense since census data is residence-
based and most residents are not at home during the day. Researchers need to
take the next step and incorporate day versus night population distributions,
as well as high-density population installations such as schools and hospitals.
While this is done relatively easily for QRA of a single route, it is more com-
plicated to generate the necessary data for an entire transportation network.
 Ch. 9. Hazardous Materials Transportation 607

Risk calculation – time dependence

There are very significant differences in risks between day and night (due
to differences in accident probabilities, population distributions, and weather
conditions). Yet most of the OR literature pays little attention to this. Risk
radii (or safe distances) strongly depend on transport mode and weather condi-
tions. Hence, it is impossible to speak of a single “minimum risk” route; hazmat
routing problems must be solved with real-time information. Solving problems
with static parameter values can result in poor solutions and decisions.

Risk calculation – model

We emphasized the importance of using the proper risk model throughout

the chapter. It is important to use as accurate a model as technically and com-
putationally feasible. For example, it is not only possible, but also necessary
to combine GIS data, plume dispersion modeling, and real-time weather in-
formation to determine bypass routes for chlorine shipments. In fact, analysis
that does not use such level of detail is of little use in the case of hazmats that
can generate plumes.

Risk calculation – nonhuman risks

The vast portion of the hazmat risk literature is concerned with fatalities,
and to some extent injuries and property damage. Little if any attention is paid
to environmental damage. Environmental risks are usually only included in
multiattribute utility models. We believe that hazmat risk models should take
into account all risks to humans and environment for broader acceptance by
the public.

Multicriteria approach to risk minimization

It is well known that different routes can emerge as minimum risk routes
depending on the definition of risk used. Hence, it is crucial to use multiple
measures and provide decision-makers with a set of efficient solutions instead
of a single “risk minimizing” route. Development of methodology that would
allow for the decision-makers to effectively search the efficient solution set and
select a route would be of great practical use.

Risk equity

The academic literature suggests that equity in the spatial distribution of risk
is a critical concern in designing hazmat management strategies acceptable to
the public. Yet, risk equity is not a great concern to the hazmat industry. If
equity is a valid concern then it must be imposed by a regulatory agency.
608 E. Erkut, S.A. Tjandra and V. Verter

Local vs. global route planning

Most hazmat transport models deal with only one commodity. While it may
make sense for carriers to decompose a transport planning problem into mul-
tiple single commodity problems, if one is concerned about concentration and
distribution of risks, one has to pose a multicommodity problem where risk
and equity concerns couple the different materials. For example, hazmat facil-
ity location models should include the hazmat distribution network for proper
risk assessment. Likewise, the hazmat network design problem requires con-
sideration of all hazmats.

Multidisciplinary nature of the problem

It is rather unfortunate that research in this highly multidisciplinary area

continues to be compartmentalized. Chemical and civil engineers tend to pub-
lish in their own journals, decision analysts and quantitative risk assessment
researchers limit their focus to their paradigms, and operations researchers
seldom wonder outside their safe zone. For fruitful research and applications
researchers from different disciplines have to reach out to one another.

Cost consequences of risks

One of the reasons why hazmat carriers are not too interested in hazmat
routing research is that there are no consequences to not using a decision-
support system before making routing decisions. If carriers are faced with
lawsuits as a result of poor routing decisions, or if their insurance companies
(or creditors) required the use of QRA in route planning to avoid such law-
suits, or if a government agency required the use of QRA and OR tools in
route planning, we believe that research in this area would accelerate consid-
erably.

Implementation

It is inconceivable to imagine a hazmat transport DSS that does not take

advantage of a GIS while most academic researchers solve small problems on
made-up (realistic) networks. In fact the ideal hazmat transport DSS would
combine GIS, QRA, OR, and MCDA. We suggest that research in this area
follow the same recipe. This increases adoption probability by the industry. We
note that clever use of GIS can enable one to incorporate nonhuman risks into
the analysis. Another necessary condition for successful implementation of OR
research in this area is cooperation between the researchers, the government
agencies, and the carriers – something we cannot claim has happened with
regularity in the past.
 Ch. 9. Hazardous Materials Transportation 609

Recent concern: security

In addition to the concerns discussed above there is a new concern in

hazmat transport planning, namely potential for a terrorist attack on a haz-
mat vehicle. The terrorist attacks in the USA in 2001 have focused atten-
tion on what other targets terrorists may choose. It was quickly recognized
that hazmat vehicles could be desirable targets for terrorists, and certain haz-
mat vehicles were designated as “weapons of mass destruction” (TRB, 2002;
Abkowitz, 2002). Such concerns changed the way the hazmat transport indus-
try operates. For example, the US Federal Government now requires hazmat
truckers to submit to fingerprinting and criminal background checks (Glaze,
2003).
This security issue, however, has not yet received much attention from op-
erations researchers. Clearly, the problem is complex and there are many
solutions that involve little or no OR. However, there is potential for OR con-
tributions and we list three here:
• Rerouting around major cities: the risk of terrorist attacks made it
very undesirable to route hazmat vehicles (particularly trains) through
major population centers. Traditional OR algorithms can be used to
find alternate routes for shipments. Erkut and Glickman (1997) show
that significant risk reductions are possible through rerouting, and
Erkut and Ingolfsson (2000) develop new methodology for routing with
a catastrophe-avoidance objective.
• Changes in the modeling of incidence risks: The traditional risk as-
sessment for hazmat transport assumes incidents are caused by traffic
accidents or human error. Yet we now know that there is a nonzero
probability of a terrorist attack or a hijack. Not only does this increase
the incident probabilities, but it also requires a new way of modeling
consequences since the impact may no longer be limited to the planned
route. Furthermore, attack probabilities are unlikely to be uniform.
For example, a location in a tunnel, on a bridge, or near a “trophy
building” is likely to have a higher attack probability than a location in
a remote and unpopulated area. In contrast, sparsely populated areas
may be associated with a higher hijack probability. A hijacked vehi-
cle’s future route is unpredictable and special precautions may have to
be taken to prevent it from having an incident in a highly populated
area. As a result, traditional risk assessment-based route planning is
no longer adequate. There are very few papers in this new area. (See
Paté-Cornell, 2002, for probabilistic modeling of terrorist threats, and
Huang and Cheu, 2004 and Huang et al., 2003, for incorporation of
security concerns in route planning.)
• Changes in route planning methodology: Past hazmat routing liter-
ature focuses on finding a minimum risk route. The problem with
determining quantitative measures and selecting routes accordingly is
610 E. Erkut, S.A. Tjandra and V. Verter

that terrorists could predict such routes by using similar methods. To

minimize the probability of a successful terrorist attack or hijacking,
shippers could alternate routes – game theory can be applied to the
problem of alternating among routes to minimize predictability – or
change them en-route in real time in ways that would be difficult to
predict. Video surveillance or Global Positioning Systems (GPS) and
communication equipment installed on all hazmat vehicles would not
only allow for precise tracking of vehicles, but also allow the imple-
mentation of such real-time decision making (see, e.g., Glaze, 2003;
Zografos and Androutsopoulos, 2001).
We believe that there are still many important OR problems in hazmat
transportation. However, we think the focus will shift from a priori optimiza-
tion toward real-time adaptive decision making for several reasons, such as the
availability of the necessary technology and data, as well as security concerns.
While it is rather unfortunate that terrorist attacks can and do happen, their
possibility opens up a new frontier for operations researchers in general, and
hazmat transport researchers in particular. We expect that hazmat transport
research will intensify in the near future and we hope that this chapter will be
useful to future researchers in this area.
We finish with an attempt to explain why we find research in hazmat lo-
gistics particularly interesting and challenging, in addition to the potential for
practical applications. The realm of OR can be crudely divided into two major
paradigms: deterministic and stochastic. Optimization is the major tool in the
deterministic area while the stochastic domain requires probabilistic model-
ing. Much of the research in OR can be classified in one of these two regions.
Hazmat logistics research lies in the cross-section of these two domains, and
it requires a good knowledge of probabilistic modeling as well as optimization
techniques. Hazmat transport can be modeled as a probabilistic phenomenon,
but one needs to add optimization of appropriate objectives to realize the pos-
sible policy benefits. The fact that we are modeling an inherently probabilistic
process results in the natural consequence that there are many appropriate ob-
jectives. The exact probabilistic expressions are usually too complicated, which
results in the use of approximations for optimization. Hence, the researchers
must understand probabilistic modeling well enough to capture the essence of
the activity, but they must also be sufficiently proficient in optimization tech-
niques to decide which approximations are necessary and what tools to use.
The multicriteria/multistakeholder nature of the problems adds to the com-
plexity as well as the attraction of this area. We found research in hazmat
logistics quite rewarding and we encourage others to explore this area further.

Acknowledgements

This research has been supported in part by two Discovery Grants from
NSERC (OGP 25481 and 183631). The authors acknowledge the input pro-
 Ch. 9. Hazardous Materials Transportation 611

vided by Manish Verma on rail and intermodal transportation of hazardous

materials.

References

Abkowitz, M. (2002). Transportation risk management: A new paradigm. Security papers, Southeastern
Transportation Center, University of Tennessee, pp. 93–103.
Abkowitz, M., Cheng, P.D.M. (1988). Developing a risk cost framework for routing truck movements
of hazardous materials. Accident Analysis and Prevention 20 (1), 39–51.
Abkowitz, M., Cheng, P.D.M. (1989). Hazardous materials transport risk estimation under conditions
of limited data availability. Transportation Research Record 1245, 14–22.
Abkowitz, M., Lepofsky, M., Cheng, P. (1992). Selecting criteria for designating hazardous materials
highway routes. Transportation Research Record 1333, 30–35.
Abkowitz, M.D., DeLorenzo, J.P., Duych, R., Greenberg, A., McSweeney, T. (2001). Assessing the eco-
nomic effect of incidents involving truck transport of hazardous materials. Transportation Research
Record 1763, 125–129.
Akgün, V., Erkut, E., Batta, R. (2000). On finding dissimilar paths. European Journal of Operational
Research 121 (2), 232–246.
Akgün, V., Parekh, A., Batta, R., Rump, C.M. (2007). Routing of a hazmat truck in the presence of
weather systems. Computers & Operations Research 34 (5), 1351–1373.
ALK Associates (1994). ALK’s PC*HazRoute (Version 2.0). ALK Associates, Inc., 1000 Herrontown
Road, Princeton, NJ, USA.
Alp, E. (1995). Risk-based transportation planning practice overall methodology and a case example.
INFOR 33 (1), 4–19.
Alp, E., Zelensky, M.J. (1996). Risk quantification for meteorology- and direction-dependent hazards
due to point and linear risk sources. Journal of Loss Prevention in the Process Industries 9 (2), 135–
145.
Alumur, S., Kara, B.Y. (2007). A new model for the hazardous waste location–routing problem. Com-
puters & Operations Research 34 (5), 1406–1423.
American Institute of Chemical Engineers (2000). Guidelines for Chemical Process Quantitative Risk
Analysis, 2nd edition. Center for Chemical Process Safety, New York.
Anderson, R.T., Barkan, C.P.L. (2004). Railroad accident rates for use in transportation risk analysis.
Transportation Research Record 1863, 88–98.
Andersson, S.E. (1994). Safe transport of dangerous goods – road, rail or sea – a screening of technical
and administrative factors. European Journal of Operational Research 75 (3), 499–507.
Ang, A., Briscoe, J. (1989). Development of a systems risk methodology for single and multimodal
transportation systems. Final report, Office of University Research, US DOT, Washington, DC.
Arya, S.P. (1999). Air Pollution Meteorology and Dispersion. Oxford Univ. Press, New York.
Ashtakala, B., Eno, L.A. (1996). Minimum risk route model for hazardous materials. Journal of Trans-
portation Engineering – ASCE 122 (5), 350–357.
Balakrishnan, A., Magnanti, T.L., Mirchandani, P. (1997). Network design. In: Dell’Amico, M., Maffioli,
F., Martello, S. (Eds.), Annotated Bibliographies in Combinatorial Optimization. Wiley, Chichester.
Barkan, C.P.L., Treichel, T.T., Widell, G.W. (2000). Reducing hazardous materials releases from railroad
tank car safety vents. Transportation Research Record 1707, 27–34.
Barkan, C.P.L., Dick, C.T., Anderson, R. (2003). Railroad derailment factors affecting hazardous ma-
terials transportation risk. Transportation Research Record 1825, 64–74.
Barnes, P. (2001). Regulating safety in an unsafe world (risk reduction for and with communities).
Journal Hazardous Materials 86, 25–37.
Batta, R., Chiu, S.S. (1988). Optimal obnoxious paths on a network: Transportation of hazardous ma-
terials. Operations Research 36 (1), 84–92.
Belardo, S., Pipkin, J., Seagle, J.P. (1985). Information support for hazardous materials movement.
Journal Hazardous Materials 10 (1), 13–32.
612 E. Erkut, S.A. Tjandra and V. Verter

Bellman, R.E. (1958). On a routing problem. Quarterly of Applied Mathematics 16, 87–90.
Berman, O., Drezner, Z., Wesolowsky, G.O. (2000). Routing and location on a network with hazardous
threats. Journal of the Operational Research Society 51 (9), 1093–1099.
Berman, O., Verter, V., Kara, B. (2007). Designing emergency response networks for hazardous mate-
rials transportation. Computers & Operations Research 34 (5), 1374–1388.
Beroggi, G.E.G. (1994). A real-time routing model for hazardous materials. European Journal of Oper-
ational Research 75 (3), 508–520.
Beroggi, G.E.G., Wallace, W.A. (1991). Closing the gap-transit control for hazardous material flow.
Journal of Hazardous Materials 27 (1), 61–75.
Beroggi, G.E.G., Wallace, W.A. (1994). Operational risk management – a new paradigm for decision-
making. IEEE Transactions on Systems Man and Cybernetics 24 (10), 1450–1457.
Beroggi, G.E.G., Wallace, W.A. (1995). Operational control of the transportation of hazardous materi-
als: An assessment of alternative decision models. Management Science 41 (12), 1962–1977.
Boffey, T.B., Karkazis, J. (1993). Models and methods for location and routing decisions relating to
hazardous materials. Studies in Locational Analysis 4, 149–166.
Boffey, T.B., Karkazis, J. (1995). Linear versus nonlinear models for hazardous materials routing.
INFOR 33, 114–117.
Bonvicini, S., Leonelli, P., Spadoni, G. (1998). Risk analysis of hazardous materials transportation:
Evaluating uncertainty by means of fuzzy logic. Journal of Hazardous Materials 62 (1), 59–74.
Bowler, L.A., Mahmassani, H.S. (1998). Routing of radioactive shipments in networks with time-varying
costs and curfews. Technical Report ANRCP-1998-11, Amarillo National Resource Center for Plu-
tonium, TX (United States).
Boykin, R.F., Freeman, R.A., Levary, R.R. (1984). Risk assessment in a chemical storage facility. Man-
agement Science 30 (4), 512–517.
Brown, D.F., Dunn, W.E. (2007). Application of a quantitative risk assessment method to emergency
response planning. Computers & Operations Research 34 (5), 1243–1265.
Bubbico, R., Ferrari, C., Mazzarotta, B. (2000). Risk analysis of LPG transport by road and rail. Journal
of Loss Prevention in the Process Industries 13 (1), 27–31.
CANUTEC (2004). Emergency Response Guidebook.
Cappanera, P. (1999). A survey on obnoxious facility location problems. Technical Report TR-99-11,
University of Pisa.
Cappanera, P., Gallo, G., Maffioli, F. (2004). Discrete facility location and routing of obnoxious activi-
ties. Discrete Applied Mathematics 133, 3–28.
Carotenuto, P., Giordani, S., Ricciardelli, S. (2007a). Finding minimum and equitable risk routes for
hazmat shipments. Computers & Operations Research 34 (5), 1304–1327.
Carotenuto, P., Giordani, S., Ricciardelli, S., Rismondo, S. (2007b). A tabu search approach for schedul-
ing hazmat shipments. Computers & Operations Research 34 (5), 1328–1350.
Cassini, P. (1998). Road transportation of dangerous goods: Quantitive risk assessment and route com-
parison. Journal of Hazardous Meterials 61 (1–3), 133–138.
Chakraborty, J., Armstrong, M.P. (1995). Using geographic plume analysis to assess community vulner-
ability to hazardous accidents. Computers, Environment, and Urban Systems 19 (5/6), 341–356.
Chang, N.B., Wei, Y.L., Tseng, C.C., Kao, C.Y.J. (1997). The design of a GIS-based decision support
system for chemical emergency preparedness and response in an urban environment. Computers,
Environment, and Urban Systems 21 (1), 67–94.
Chang, T.S., Nozick, L.K., Turnquist, M.A. (2005). Multi-objective path finding in stochastic dynamic
networks, with application to routing hazardous materials shipments. Transportation Science 39 (3),
383–399.
Chin, S.-M., Cheng, P.D.-M. (1989). Bicriterion routing scheme for nuclear spent fuel transportation.
Transportation Research Record 1245, 60–64.
Chow, T.C., Oliver, R.M., Vignaux, G.A. (1990). A Bayesian escalation model to predict nuclear acci-
dents and risk. Operations Research 38 (2), 265–277.
Christofides, N., Eilon, S. (1969). Expected distances in distribution problems. Operational Research
Quarterly 20, 437–443.
 Ch. 9. Hazardous Materials Transportation 613

Coleman, J.A. (1984). Railroad–highway crossings and route selection for transporting hazardous ma-
terials. Public Roads 48 (2), 63–71.
Corea, G., Kulkarni, V.G. (1993). Shortest paths in stochastic networks with arc lengths having discrete
distributions. Networks 23, 175–183.
Cox, R., Turnquist, M. (1986). Scheduling truck shipments of hazardous materials in the presence of
curfews. Transportation Research Record 1063, 21–26.
Current, J.R., Ratick, S. (1995). A model to assess risk, equity, and efficiency in facility location and
transportation of hazardous materials. Location Science 3, 187–202.
Cutter, S.L., Ji, M.H. (1997). Trends in US hazardous materials transportation spills. Professional Ge-
ographer 49 (3), 318–331.
Dell’Olmo, P., Gentili, M., Scozzari, A. (2005). On finding dissimilar Pareto-optimal paths. European
Journal of Operational Research 162, 70–82.
Deng, C., Oberg, S.G., Downs, J.L. (1996). Risk assessment for transportation of radioactive material
within the state of Idaho. Health Physics 70 (6), 41. Annual Meeting of the Health Physics Society,
Seattle, WA (United States).
Dennis, S.M. (1996). Estimating risk costs per unit of exposure for hazardous materials transported by
rail. Logistics and Transportation Review 32 (4), 351–375.
Douligeris, C., Iakovou, E., Yudhbir, L. (1997). Maritime route risk analysis for hazardous materials
transportation. Proceedings of the 8th IFAC/IFIP/IFORS Symposium on Transportation Systems, Cha-
nia, Crete, Greece, pp. 574–579.
Duque, J., Barbosa-Póvoa, A.P.F.D. (2007). Synthesis and optimization of the recovery route for resid-
ual products under uncertain product demand. Computers & Operations Research 34 (5), 1463–1490.
Efroymson, R.A., Murphy, D.L. (2000). Ecological risk assessment of multimedia hazardous air pollu-
tants: Estimating exposure and effects. Science of the Total Environment 274, 219–230.
Erkut, E. (1990). The discrete p-dispersion problem. European Journal of Operational Research 46, 48–
60.
Erkut, E. (1993). Inequality measures for location problems. Location Science 1 (3), 199–217.
Erkut, E. (1995). On the credibility of the conditional risk model for routing hazardous materials.
Operations Research Letters 18, 49–52.
Erkut, E. (1996). The road not taken. OR/MS Today 23, 22–28.
Erkut, E., Alp, O. (2006). Integrated routing and scheduling of hazmat trucks with stops en-route.
Transportation Science, in press.
Erkut, E., Alp, O. (2007). Designing a road network for dangerous goods shipments. Computers &
Operations Research 34 (5), 1389–1405.
Erkut, E., Glickman, T. (1997). Minimax population exposure in routing highway shipments of haz-
ardous materials. Transportation Research Record 1602, 93–100.
Erkut, E., Gzara, F. (2005). A bi-level programming application to hazardous material transportation
network design. Research report, Department of Finance and Management Science, University of
Alberta School of Business, Edmonton, Alberta, Canada.
Erkut, E., Ingolfsson, A. (2000). Catastrophe avoidance models for hazardous materials route planning.
Transportation Science 34 (2), 165–179.
Erkut, E., Ingolfsson, A. (2005). Transport risk models for hazardous materials: Revisited. Operations
Research Letters 33 (1), 81–89.
Erkut, E., Neuman, S. (1989). Analytical models for locating undesirable facilities. European Journal of
Operational Research 40, 275–291.
Erkut, E., Verter, V. (1995a). Hazardous materials logistics. In: Drezner, Z. (Ed.), Facility Location:
A Survey of Applications and Methods. Springer-Verlag, New York, pp. 467–506. Chapter 20.
Erkut, E., Verter, V. (1995b). A framework for hazardous materials transport risk assessment. Risk
Analysis 15 (5), 589–601.
Erkut, E., Verter, V. (1998). Modeling of transport risk for hazardous materials. Operations Research 46
(5), 625–642.
Ertugrul, A. (1995). Risk-based transportation – planning practice – overall methodology and a case
example. INFOR 33 (1), 4–19.
614 E. Erkut, S.A. Tjandra and V. Verter

Fabiano, B., Curro, F., Palazzi, E., Pastorino, R. (2002). A framework for risk assessment and decision-
making strategies in dangerous good transportation. Journal of Hazardous Materials 93 (1), 1–15.
Ferrada, J.J., Michelhaugh, R.D. (1994). Development of an expert system for transportation of haz-
ardous and radioactive materials. In: International Topical Meeting on Nuclear and Hazardous Waste
Management, Spectrum ’94, Atlanta. American Nuclear Society, Inc., l.2, pp. 997–1002.
FMCSA (2001). Comparative risks of hazardous materials and non-hazardous materials truck shipment
accidents/incidents. Federal Motor Carrier Safety Administration, Washington, DC.
Frank, H. (1969). Shortest paths in probabilistic graphs. Operations Research 17, 583–599.
Frank, W.C., Thill, J.C., Batta, R. (2000). Spatial decision support system for hazardous material truck
routing. Transportation Research Part C – Emerging Technologies 8 (1–6), 337–359.
Fronczak, R.E. (2001). Public health risks of railroad hazardous substance emergency events. Letters
to the Editor. Journal of Occupational and Environmental Medicine 43 (9), 738–739.
Fu, L., Rilett, L.R. (1997). Real-time estimation of incident delay in dynamic and stochastic networks.
Transportation Research Record 1603, 99–105.
Fu, L., Rilett, L.R. (1998). Expected shortest paths in dynamic and stochastic traffic networks. Trans-
portation Research Part B 32 (7), 499–516.
Gendreau, M., Laporte, G., Parent, I. (2000). Heuristics for the location of inspection stations on a
network. Naval Research Logistics 47 (4), 287–303.
Giannikos, I. (1998). A multiobjective programming model for locating treatment sites and routing
hazardous wastes. European Journal of Operational Research 104 (2), 333–342.
Glaze, M. (2003). New security requirements for hazmat transportation. Occupational Health &
Safety 72 (9), 182–185.
Glickman, T.S. (1983). Rerouting railroad shipments of hazardous materials to avoid populated areas.
Accident Analysis and Prevention 15 (5), 329–335.
Glickman, T.S. (1988). Benchmark estimates of release accident rates in hazardous materials trans-
portation of rail and truck. Transportation Research Record 1193, 22–28.
Glickman, T.S. (1991). An expeditious risk assessment of the highway transportation of flammable liq-
uids in bulk. Transportation Science 25 (2), 115–123.
Glickman, T.S., Golding, D. (1991). For a few dollars more: Public trust and the case for transporting
nuclear waste in dedicated trains. Policy Studies Review 10, 4.
Glickman, T.S., Rosenfield, D.B. (1984). Risks of catastrophic derailments involving the release of haz-
ardous materials. Management Science 30 (4), 503–511.
Glickman, T.S., Sontag, M.A. (1995). The tradeoffs associated with rerouting highway shipments of
hazardous materials to minimize risk. Risk Analysis 15 (1), 61–67.
Gopalan, R., Kolluri, K.S., Batta, R., Karwan, M.H. (1990a). Modeling equity of risk in the transporta-
tion of hazardous materials. Operations Research 38 (6), 961–975.
Gopalan, R., Batta, R., Karwan, M.H. (1990b). The equity constrained shortest path problem. Comput-
ers & Operations Research 17 (3), 297–307.
Government of Alberta (2002). Dangerous goods transportation and handling act. Available at
http://www.qp.gov.ab.ca/documents/acts/D04.cfm.
Gregory, R., Lichtenstein, S. (1994). A hint of risk-tradeoffs between quantitative and qualitative risk-
factors. Risk Analysis 14 (2), 199–206.
Grenney, W.J., Stevens, D.K., Doucette, W.J. (1990). Decision support model for hazardous-waste ap-
plication rates at land treatment systems. Civil Engineering Systems 7 (4), 219–228.
Grenzeback, L.R., Reilly, W.R., Roberts, P.O., Stowers, J.R. (1990). Urban freeway gridlock study:
Decreasing the effects of large trucks on peak-period urban freeway congestion. Transportation Re-
search Record 1256, 16–26.
Groothuis, P.A., Miller, G. (1997). The role of social distrust in risk-benefit analysis: A study of the
siting of a hazardous waste disposal facility. Journal of Risk and Uncertainty 15 (3), 241–257.
Haas, T.J., Kichner, J.J. (1987). Hazardous materials in marine transportation – a practical course.
Journal of Chemical Education 64 (1), 34–35.
Hall, R.W. (1986). The fastest path through a network with random time-dependent travel times. Trans-
portation Science 20 (3), 182–188.
 Ch. 9. Hazardous Materials Transportation 615

Hanna, S.R., Chang, J.C., Strimaitis, D.G. (1993). Hazardous gas model evaluation with field observa-
tions. Atmospheric Environment 27A (15), 2265–2285.
Hansen, P. (1980). Bicriterion path problems. In: Fandel, G., Gal, T. (Eds.), Multiple Criteria Decision
Making: Theory and Applications. Lecture Notes in Economics and Mathematical Systems, vol. 177.
Springer-Verlag, Berlin, pp. 109–127.
Harwood, D.W., Russell, E.R., Viner, J.G. (1989). Characteristics of accidents and incidents in highway
transportation of hazardous materials. Transportation Research Record 1245, 23–33.
Harwood, D.W., Viner, J.G., Russell, E.R. (1993). Procedure for developing truck accident and release
rates for hazmat routing. Journal of Transportation Engineering – ASCE 119 (2), 189–199.
Helander, M.E., Melachrinoudis, E. (1997). Facility location and reliable route planning in hazardous
material transportation. Transportation Science 31 (3), 216–226.
Henley, E.J., Kumamoto, H. (1981). Reliability Engineering and Risk Assessment. Prentice Hall, Engle-
wood Cliffs, NJ.
Hillsman, E.L. (1988). Estimating population at risk from release of hazardous materials. In: Seminar
on Multipurpose Land Information Systems: Application of Information Technology for Natural Re-
source Planning, Management, and Monitoring, Univ. of Wisconsin, October 24, 1986. Report 32, Inst.
for Environmental Studies, pp. 79–96
Hobeika, A.G., Kim, S. (1993). Databases and needs for risk assessment of hazardous materials ship-
ments by trucks. In: Moses, L.N. (Ed.), Transportation of Hazardous Materials. Kluwer Academic,
Boston, MA, p. 146.
Hoffman, G., Janko, J. (1990). Travel times as a basic part of the LISB guidance strategy. Paper pre-
sented at the IEEE Road Traffic Control Conference, London.
Hollister, K.A.K. (2002). A risk/cost framework for logistics policy evaluation: Hazardous waste man-
agement. Journal of Business & Economic Studies 8 (1), 51–65.
Horlick-Jones, T. (1995). Modern disasters as outrage and betrayal. International Journal of Mass Emer-
gencies and Disasters 13 (3), 305–315.
Horman, R.L. (1987). Time/loss analysis. Technical report, EG and G Idaho, Inc., Idaho Falls, ID
(United States), System Safety Development Center.
Huang, B., Cheu, R.L. (2004). GIS and genetic algorithms for HAZMAT route planning with security
considerations. International Journal of Geographical Information Science 18 (8), 769–787.
Huang, B., Long, C.R., Liew, Y.S. (2003). GIS – AHP model for HAZMAT routing with security con-
siderations. In: IEEE 6th Int’l Conf. on ITS (ITSC2003).
Hwang, S.T., Brown, D.F., O’Steen, J.K., Policastro, A.J., Dunn, W.E. (2001). Risk assessment for
national transportation of selected hazardous materials. Transportation Research Record 1763, 114–
124.
Iakovou, E.T. (2001). An interactive multiobjective model for the strategic maritime transportation of
petroleum products: Risk analysis and routing. Safety Science 39 (1–2), 19–29.
Iakovou, E., Douligeris, C., Li, H., Ip, C., Yudhbir, L. (1999). A maritime global route planning model
for hazardous materials transportation. Transportation Science 33 (1), 34–48.
ICF Consulting (2000). Risk management framework for hazardous materials transportation. Submit-
ted to Research and Special Programs Administration, US Department of Transportation.
Jacobs, T.L., Warmerdam, J.M. (1994). Simultaneous routing and siting for hazardous-waste operations.
Journal of Urban Planning and Development – ASCE 120 (3), 115–131.
Jin, H.H., Batta, R. (1997). Objectives derived from viewing hazmat shipments as a sequence of inde-
pendent Bernoulli trials. Transportation Science 31 (3), 252–261.
Jin, H.H., Batta, R.J., Karwan, M.H. (1996). On the analysis of two new models for transporting haz-
ardous materials. Operations Research 44 (5), 710–723.
Jonkman, S.N., van Gelder, P.H.A.J.M, Vrijling, J.K. (2003). An overview of quantitative risk measures
for loss of life and economic damage. Journal of Hazardous Materials A 99, 1–30.
Kalelkar, A.S., Brinks, R.E. (1978). Use of multidimensional utility functions in hazardous shipment
decisions. Accident Analysis and Prevention 10, 251–265.
Kara, B.Y., Verter, V. (2004). Designing a road network for hazardous materials transportation. Trans-
portation Science 38 (2), 188–196.
616 E. Erkut, S.A. Tjandra and V. Verter

Kara, B.Y., Erkut, E., Verter, V. (2003). Accurate calculation of hazardous materials transport risks.
Operations Research Letters 31 (4), 285–292.
Karkazis, J., Boffey, T.B. (1995). Optimal location of routes for vehicles transporting hazardous mate-
rials. European Journal of Operational Research 86 (2), 201–215.
Kasperson, R.E., Renn, O., Slovic, P., Brown, H.S., Emel, J., Goble, R., Kasperson, J.S., Ratick, S.
(1988). The social amplification of risk: A conceptual framework. Risk Analysis 8 (2), 177–187.
Keeney, R. (1980). Equity and public risk. Operations Research 28, 527–534.
Keeney, R.L., Winkler, L.R. (1985). Evaluating decision strategies for equity of public risks. Operations
Research 33 (5), 955–970.
Kimberly, A., Killmer, H. (2002). A risk/cost framework for logistics policy evaluation: Hazardous waste
management. Journal of Business & Economic Studies 8 (1), 51–65.
Kloeber, G., Cornell, M., McNamara, T., Moscati, A. (1979). Risk assessment of air versus other trans-
portation modes for explosives and flammable cryogenic liquids. Volume I: Risk assessment method
and results. Report DOT/RSPA/MTB-79/13, US Department of Transportation.
Koutsopoulos, H.N., Xu, H. (1993). An information discounting routing strategy for advanced traveler
information systems. Transportation Research Part C 1 (3), 249–264.
Kulkarni, V.G. (1986). Shortest paths in networks with exponentially distributed arc lengths. Net-
works 16, 255–274.
Kunreuther, H., Easterling, D. (1990). Are risk-benefit tradeoffs possible in siting hazardous facilities.
American Economic Review 80 (2), 252–256.
Kunreuther, H., Linnerooth, J., Vaupel, J.W. (1984). A decision-process perspective on risk and policy
analysis. Management Science 30 (4), 475–485.
LaFrance-Linden, D., Watson, S., Haines, M.J. (2001). Threat assessment of hazardous materials trans-
portation in aircraft cargo compartments. Transportation Research Record 1763, 130–137.
Larson, K.M. (1996). Hazmat on the rails: A closer look. Environmental Solutions 9 (9).
Lassarre, S., Fedra, K., Weigkricht, E. (1993). Computer-assisted routing of dangerous goods for Haute-
Normandie. Journal of Transportation Engineering 119 (2), 200–210.
Leeming, D.G., Saccomanno, F.F. (1994). Use of quantified risk assessment in evaluating the risks of
transporting chlorine by road and rail. Transportation Research Record 1430, 27–35.
Leonelli, P., Bonvicini, S., Spadoni, G. (2000). Hazardous materials transportation: A risk-analysis-
Based routing methodology. Journal of Hazardous Materials 71, 283–300.
Lepofsky, M., Abkowitz, M., Cheng, P. (1993). Transportation hazard analysis in integrated GIS envi-
ronment. Journal of Transportation Engineering – ASCE 119 (2), 239–254.
Levy, H. (1992). Stochastic dominance and expected utility: Survey and analysis. Management Science 38
(4), 555–593.
Li, H., Iakovou, E., Douliger, C. (1996). A strategic planning model for marine oil transportation in the
Gulf of Mexico. Transportation Research Record 1522, 108–115.
Lindner-Dutton, L., Batta, R., Karwan, M. (1991). Equitable sequencing of a given set of hazardous
materials shipments. Transportation Science 25, 124–137.
List, G., Abkowitz, M. (1986). Estimates of current hazardous material flow patterns. Transportation
Quarterly 40, 483–502.
List, G., Mirchandani, P. (1991). An integrated network planar multiobjective model for routing and
siting for hazardous materials and wastes. Transportation Science 25 (2), 146–156.
List, G.F., Turnquist, M.A. (1998). Routing and emergency response team siting for high-level radioac-
tive waste shipments. IEEE Transactions on Engineering Management 45 (2), 141–152.
List, G.F., Mirchandani, P.B., Turnquist, M.A., Zografos, K.G. (1991). Modeling and analysis for haz-
ardous materials transportation-risk analysis, routing scheduling and facility location. Transportation
Science 25 (2), 100–114.
Lovett, A.A., Parfitt, J.P., Brainard, J.S. (1997). Using GIS in risk analysis: A case study of hazardous
waste transport. Risk Analysis 17 (5), 625–633.
Luedtke, J., White, C.C. (2002). Hazmat transportation and security: Survey and directions for future
research. Department of Ind. & Sys. Engineering, Georgia Institute of Technology.
Macgregor, D., Slovic, P., Mason, R.G., Detweiler, J., Binney, S.E., Dodd, B. (1994). Perceived risks
of radioactive-waste transport through Oregon: Results of a statewide survey. Risk Analysis 14 (1),
5–14.
 Ch. 9. Hazardous Materials Transportation 617

Madsen, O.B.G. (1983). Methods for solving combined two level location–routing problems of realistic
dimensions. European Journal of Operational Research 12 (3), 295–301.
Magnanti, T.L., Wong, R.T. (1984). Network design and transportation planning: models and algo-
rithms. Transportation Science 18, 1–55.
Marianov, V., ReVelle, C. (1998). Linear, non-approximated models for optimal routing in hazardous
environments. Journal of the Operational Research Society 49 (2), 157–164.
Marianov, V., ReVelle, C., Shih, S. (2002). Anticoverage models for obnoxious material transportation.
Environment and Planning B – Planning & Design 29 (1), 141–150.
Markowitz, H.M. (1987). Mean – Variance Analysis in Portfolio Choice and Capital Markets. Blackwell
Sci., Oxford.
Marsh, M.T., Schilling, D.A. (1994). Equity measures in facility location analysis: A review and frame-
work. European Journal of Operations Research 74 (1), 1–17.
Martinez-Alegria, R., Ordonez, C., Taboada, J. (2003). A conceptual model for analyzing the risks
involved in the transportation of hazardous goods: Implementation in a geographic information
system. Human And Ecological Risk Assessment 9 (3), 857–873.
McClure, T.A., Brentlinger, L.A., Drago, V.J., Kerr, D.C. (1988). Considerations in rail routing of ra-
dioactive materials. Technical report, Office of Transport Systems and Planning, Battelle Memorial
Institute, Columbus, OH.
McCord, M.R., Leu, A.Y.C. (1995). Sensitivity of optimal hazmat routes to limited preference specifi-
cation. INFOR 33 (2), 68–83.
McNeil, S., Oh, S.C. (1991). A note on the influence of rail defects on the risk associated with shipping
hazardous materials by rail. Risk Analysis 11 (2), 333–338.
Miaou, S.P., Chin, S.M. (1991). Computing k-shortest path for nuclear spent fuel highway transporta-
tion. European Journal of Operational Research 53 (1), 64–80.
Milazzo, M.F., Lisi, R., Maschio, G., Antonioni, G., Bonvicini, S., Spadoni, G. (2002). HazMat transport
through, Messina town: From risk analysis suggestions for improving territorial safety. Journal of
Loss Prevention in the Process Industries 15 (5), 347–356.
Mileti, D., O’Brien, P. (1992). Warning during disasters: Normalizing communicated risk. Social Prob-
lems 39, 40–44.
Miller-Hooks, E.D. (2001). Adaptive least-expected time paths in stochastic, time-varying transporta-
tion and data networks. Networks 37 (1), 35–52.
Miller-Hooks, E.D., Mahmassani, H.S. (1998). Optimal routing of hazardous materials in stochastic,
time-varying transportation networks. Transportation Research Record 1645, 143–151.
Miller-Hooks, E.D., Mahmassani, H.S. (2000). Least expected time paths in stochastic, time-varying
transportation networks. Transportation Science 34, 198–215.
Mills, G.S., Neuhauser, K.S. (1998). Urban risks of truck transport of radioactive material. Risk Analy-
sis 18 (6), 781–785.
Min, H., Jayaraman, V., Srivastava, R. (1998). Combined location–routing problems: A synthesis and
future research directions. European Journal of Operational Research 108 (1), 1–15.
Mirchandani, P.B. (1976). Shortest distance and reliability of probabilistic networks. Computers & Op-
erations Research 3, 347–355.
Moore, J.A. (1989). Speaking of the data: The alar controversy. EPA Journal 15, 5–9.
Moore, J.E., Sandquist, G.M., Slaughter, D.M. (1995). A route-specific system for risk assessment of
radioactive materials transportation accidents. Nuclear Technology 112 (1), 63–78.
Morales, J.M. (1989). Analytical procedures for estimating freeway traffic congestion. TRB Research
Circular 344, 38–46.
Morin, T.L., Marsten, R.E. (1976). Branch-and-bound strategies for dynamic programming. Operations
Research 24 (4), 611–627.
Mumpower, J.L. (1986). An analysis of the de minimis strategy for risk management. Risk Analysis 6,
437–446.
Nardini, L., Aparicio, L., Bandoni, A., Tonelli, S.M. (2003). Regional risk associated with the transport
of hazardous materials. Latin American Applied Research 33 (3), 213–218.
Neill, H.R., Neill, R.H. (2000). Transportation of transuranic nuclear waste to WIPP: A reconsideration
of truck versus rail for two sites. Natural Resources Journal 40 (1), 93–124.
618 E. Erkut, S.A. Tjandra and V. Verter

Nembhard, D.A., White, C.C. (1997). Applications of non-order-preserving path selection to hazmat
routing. Transportation Science 31 (3), 262–271.
Nguyen, S., Pallottino, S. (1986). Hyperpaths and shortest hyperpaths. In: Combinatorial Optimization.
Lecture Notes in Mathematics, vol. 1403. Springer-Verlag, Berlin, pp. 258–271.
NHTSA (1996). The economic cost of motor vehicle crashes. National Highway Traffic Safety Admin-
istration, US Department of Transportation.
NIOSH (1994). NTIS publication PB-94-195047. Available at http://www.cdc.gov/niosh/idlh/idlh-1.html.
Nozick, L.K., List, G.F., Turnquist, M.A. (1997). Integrated routing and scheduling in hazardous mate-
rials transportation. Transportation Science 31 (3), 200–215.
NSC (2003). Estimating the costs of unintentional injuries, 2003. Available at http://www.nsc.org/
lrs/statinfo/estcost.htm.
Ogryczak, W., Ruszczynski, A. (2002). Dual stochastic dominance and related mean-risk models. SIAM
Journal on Optimization 13 (1), 60–78.
Orr, M.F., Kaye, W.E., Zeitz, P., Powers, M.E., Rosenthal, L. (2001). Public health risks of railroad
hazardous substance emergency events. Journal of Occupational and Environmental Medicine 43
(2), 94–100.
Pasquill, F., Smith, F.B. (1983). Atmospheric Diffusion, 3rd edition. Ellis Horwood Series in Environmental
Science. Prentice Hall, New York.
Paté-Cornell, E. (2002). Probabilistic modeling of terrorist threats: A system analysis approach to set-
ting priorities among counter-measures. Military Operations Research 7 (4), 5–23.
Patel, M.H., Horowitz, A.J. (1994). Optimal routing of hazardous materials considering risk of spill.
Transportation Research Part A – Policy and Practice 28 (2), 119–132.
Peirce, J.J., Davidson, G.M. (1982). Linear programming in hazardous waste management. Journal
Environmental Engineering 108 (5), 1014–1026.
Pet-Armacost, J.J., Sepulveda, J., Sakude, M. (1999). Monte Carlo sensitivity analysis of unknown pa-
rameters in hazardous materials transportation risk assessment. Risk Analysis 19 (6), 1173–1184.
Philipson, L.L., Napadensky, H.S., Maxey, M.N. (1983). Hazardous materials transportation risk assess-
ment. Transportation of Hazardous Materials: Toward a National Strategy, vol. 2, TRB Special Report
197, pp. 43–57.
Pijawka, K.D., Foote, S., Soesilo, A., Scanlon, R.D., Cantilli, E.J. (1985). Improving transportation of
hazardous materials through risk assessment and routing. Technical Report PB-86-112471/XAB;
TRB/TRR-1020, National Research Council, Washington, DC (USA), Transportation Research
Board.
Pine, J.C., Marx, B.D. (1997). Utilizing state hazardous materials transportation data in hazardous
analysis. Journal of Hazardous Materials 54 (1–2), 113–122.
Puliafito, E., Guevara, M., Puliafito, C. (2003). Characterization of urban air quality using GIS as a
management system. Environmental Pollution 122, 105–117.
Purdy, G. (1993). Risk analysis of the transportation of dangerous goods by road and rail. Journal of
Hazardous Materials 33 (2), 229–259.
Raj, P.K., Pritchard, E.W. (2000). Hazardous materials transportation on US railroads: Application of
risk analysis methods to decision making in development of regulations. Transportation Research
Record 1707, 22–26.
Renn, O., Burns, W.J., Kasperson, J.X., Kasperson, R.E., Slovic, P. (1992). The social amplification of
risk: Theoretical foundations and empirical applications. Journal of Social Issues 48, 137–160.
ReVelle, C., Cohon, J., Shobrys, D. (1991). Simultaneous siting and routing in the disposal of hazardous
wastes. Transportation Science 25 (2), 138–145.
Rhyne, W.R. (1994). Hazardous Materials Transportation Risk Analysis. Van Nostrand–Reinhold, New
York.
Roeleven, D., Kok, M., Stipdonk, H.L., Devries, W.A. (1995). Inland waterway transport – modeling
the probability of accidents. Safety Science 19 (2–3), 191–202.
Romer, H., Haastrup, P., Petersen, H.J.S. (1995). Accidents during marine transport of dangerous goods
– distribution of fatalities. Journal of Loss Prevention in the Process Industries 8 (1), 29–34.
Rosmuller, N., Van Gelder, P.H.A.J.M. (2003). Hazardous materials release analysis: Probabilistic input
for emergency response organizations. In: Bedford, T., Van Gelder, P.H.A.J.M. (Eds.), Safety and
Reliability. A.A. Balkema, pp. 1337–1344.
 Ch. 9. Hazardous Materials Transportation 619

Saccomanno, F., Chan, A. (1985). Economic evaluation of routing strategies for hazardous road ship-
ments. Transportation Research Record 1020, 12–18.
Saccomanno, F.F., El-Hage, S. (1989). Minimizing derailments of railcars carrying dangerous commodi-
ties through effective marshaling strategies. Transportation Research Record 1245, 34–51.
Saccomanno, F.F., Elhage, S.M. (1991). Establishing derailment profiles by position for corridor ship-
ments of dangerous goods. Canadian Journal of Civil Engineering 18 (1), 67–75.
Saccomanno, F., Haastrup, P. (2002). Influence of safety measures on the risks of transporting danger-
ous goods through road tunnels. Risk Analysis 22 (6), 1059–1069.
Saccomanno, F.F., Shortreed, J.H. (1993). Hazmat transport risks – societal and individual perspectives.
Journal of Transportation Engineering – ASCE 119 (2), 177–188.
Saccomanno, F.F., Shortreed, J.H., Van Aerde, M., Higgs, J. (1989). Comparison of risk measures for
the transport of dangerous commodities by truck and rail. Transportation Research Record 1245,
1–13.
Sandquist, G.M., Bennion, J.S., Moore, J.E., Slaughter, D.M. (1993). A route-specific transportation
risk assessment model. Transactions of the American Nuclear Society 68 (9), 151–153.
Scanlon, R.D., Cantilli, E.J. (1985). Assessing the risk and safety in the transportation of hazardous
materials. Transportation Research Record 1020.
Seguin, R., Potvin, J.-Y., Gendreau, M., Crainic, T.G., Marcotte, P. (1997). Real-time decision problems:
An operational research perspective. Journal of the Operational Research Society 48 (2), 162–174.
Shappert, L.B., Brobst, W.A., Langhaar, J.W., Sisler, J.A. (1973). Probability and consequences of
transportation accidents involving radioactive-material shipments in the nuclear fuel cycle. Nuclear
Safety 14 (6), 597–604.
Sherali, H.D., Brizendine, L.D., Glickman, T.S., Subramanian, S. (1997). Low-probability high-
consequence considerations in routing hazardous materials shipments. Transportation Science 31,
237–251.
Shobrys, D. (1981). A model for the selection of shipping routes and storage locations for a hazardous
substance. PhD thesis, Johns Hopkins University, Baltimore.
Sissell, K. (1995). Defining risk assessment. Chemical Week (September 27), S16.
Sivakumar, R.A., Batta, R. (1994). The variance-constrained shortest path problem. Transportation Sci-
ence 28 (4), 309–316.
Sivakumar, R.A., Batta, R., Karwan, M.H. (1993). A network-based model for transporting extremely
hazardous materials. Operations Research Letters 13 (2), 85–93.
Sivakumar, R.A., Batta, R., Karwan, M.H. (1995). A multiple route conditional risk model for trans-
porting hazardous materials. INFOR 33 (1), 20–33.
Slovic, P., Lichtenstein, S., Fischhoff, B. (1984). Modeling the societal impact of fatal accidents. Man-
agement Science 30, 464–474.
Smith, R.D. (1987). Routing and scheduling of radioactive material shipments. PhD thesis, Texas Uni-
versity, Austin, USA.
Spadoni, G., Leonelli, P., Verlicchi, P., Fiore, R. (1995). A numerical procedure for assessing risks from
road transport of dangerous substances. Journal of Loss Prevention in the Process Industries 8 (4),
245–252.
Stowers, C.L., Palekar, U.S. (1993). Location models with routing considerations for a single obnoxious
facility. Transportation Science 27 (4), 350–362.
Sulijoadikusumo, G.S., Nozick, L.K. (1998). Multiobjective routing and scheduling of hazardous mate-
rials shipments. Transportation Research Record 1613, 96–104.
Swoveland, C. (1987). Risk analysis of regulatory options for the transport of dangerous commodities
by rail. Interfaces 17 (4), 90–107.
Tayi, G.K., Rosenkrantz, D.J., Ravi, S.S. (1999). Path problems in networks with vector-valued edge
weights. Networks 34 (1), 19–35.
Transport Canada (2004). On the move – keeping Canadians safe. Available at http://www.tc.gc.ca/
Publications/TP14217e/onthemove-e.htm.
TRB (1993). Hazardous materials shipment information for emergency response. Special Report 239,
Transportation Research Board.
620 E. Erkut, S.A. Tjandra and V. Verter

TRB (2002). Deterrence, protection, and preparation: The new transportation security imperative. Spe-
cial Report 270, Transportation Research Board.
Turnquist, M. (1993). Multiple objectives, uncertainty and routing decisions for hazardous materials
shipments. In: Cohn, L.F. (Ed.), Computing in Civil and Building Engineering: Proceedings of the
5th International Conference on Computing in Civil and Building Engineering. ASCE, New York,
pp. 357–364.
UN (2001). UN recommendation on the transport of dangerous goods, model regulations. United Na-
tions Economic and Social Council’s Committee of Experts on the Transport of Dangerous Goods.
Urbanek, G.I., Barber, E.J. (1980). Development of criteria to designate routes for transporting haz-
ardous materials. Final Report FHWA–RD–80-105, Federal Highway Administration, Washington,
DC.
US DOT (1994). Guidelines for applying criteria to designate routes for transporting hazardous mate-
rials. Report FHWA-SA-94-083, US Department of Transportation, Federal Highway Administra-
tion, Washington, DC.
US DOT (1998). Hazardous materials shipments. Office of Hazardous Materials Safety, Research and
Special Programs Administration, US Department of Transportation, Washington, DC.
US DOT (2000). Department wide program evaluation of the hazardous materials transportation pro-
grams. The Office of Hazardous Materials Safety, US Department of Transportation, Washington,
DC.
US DOT (2004a). 2002 and 2003 summary of hazardous materials transportation incidents.
Research and Special Programs Administration, Office of Hazardous Materials Safety,
US Department of Transportation, Washington, DC. Available at http://hazmat.dot.gov/
files/summary/2003/2003sum.pdf .
US DOT (2004b). List of hazardous materials. The Office of Hazardous Materials Safety, US Depart-
ment of Transportation, Washington, DC.
US DOT (2004c). Hazmat summary by mode/cause: Calendar year 2003. Serious incidents, printed
on January 4, 2005. The Office of Hazardous Materials Safety, US Department of Transportation,
Washington, DC. Available at http://hazmat.dot.gov/pubs/inc/data/2003/2003scause.pdf .
US EPA (2004). Computer-aided management of emergency operations software with ALOHA.
US Environmental Protection Agency. Complete Kit available at http://response.restoration.
noaa.gov/cameo/aloha.html.
Vanaerde, M., Shortreed, J., Stewart, A.M., Matthews, M. (1989). Assessing the risks associated with
the transport of dangerous goods by truck and rail using the riskmod model. Canadian Journal of
Civil Engineering 16 (3), 326–334.
Vansteen, J.F.J. (1987). A methodology for aiding hazardous materials transportation decisions. Euro-
pean Journal of Operational Research 32 (2), 231–244.
Verma, M., Verter, V. (2007). Rail transportation of hazardous materials: Population exposure to air-
borne toxins. Computers & Operations Research 34 (5), 1287–1303.
Verter, V., Erkut, E. (1995). Hazardous materials logistics: An annotated bibliography. In: Haurie,
A., Carraro, C. (Eds.), Operations Research and Environmental Management. Kluwer Academic,
pp. 221–267.
Verter, V., Erkut, E. (1997). Incorporating insurance costs in hazardous materials routing models. Trans-
portation Science 31 (3), 227–236.
Verter, V., Kara, B.Y. (2001). A GIS-based framework for hazardous materials transport risk assess-
ment. Risk Analysis 21 (6), 1109–1120.
Verter, V., Kara, B.Y. (2005). A path-based approach for the hazardous network design problem. Work-
ing paper, Faculty of Management, McGill University.
Vesely, W.E., Goldberg, F.F., Roberts, N.H., Haasl, D.F. (1981). Fault tree handbook. NUREG - 0492.
US Nuclear Regulatory Commission.
Weigkricht, E., Fedra, K. (1995). Decision-support systems for dangerous goods transportation.
INFOR 33 (2), 84–99.
Wijeratne, A.B., Turnquist, M.A., Mirchandani, P.B. (1993). Multiobjective routing of hazardous mate-
rials in stochastic networks. European Journal of Operational Research 65, 33–43.
 Ch. 9. Hazardous Materials Transportation 621

Wilmot, E.L., Cashwell, J.W., Joy, D.S. (1983). Analysis of population densities along transportation
routes. The 7th international symposium on packing and transportation of radioactive materials, New
Orleans, LA, USA, p. 10.
Wirasinghe, S. (1978). Determination of traffic delays from shock-wave analysis. Transportation Re-
search 12, 343–348.
Yang, B. (2001). Robust on-line routing in intelligent transportation systems. PhD dissertation, Depart-
ment of Civil and Environmental Engineering, The Pennsylvania State University.
Yitzhaki, S. (1982). Stochastic dominance, mean variance, and Gini’s mean difference. The American
Economic Review 72, 178–185.
Zhang, J.J., Hodgson, J., Erkut, E. (2000). Using GIS to assess the risks of hazardous materials trans-
port in networks. European Journal of Operational Research 121 (2), 316–329.
Zografos, K.G., Androutsopoulos, K.N. (2001). Assessing impacts from introduction of advanced trans-
port telematics technologies in hazardous materials fleet management. In: Proceedings of the 80th
Annual Meeting of Transportation Research Board, Washington DC, USA.
Zografos, K.G., Androutsopoulos, K.N. (2002). Heuristic algorithms for solving hazardous materials
logistical problems. Transportation Research Record 1783, 158–166.
Zografos, K.G., Androutsopoulos, K.N. (2004). A heuristic algorithm for solving hazardous materials
distribution problems. European Journal of Operational Research 152 (2), 507–519.
Zografos, K.G., Davis, C. (1989). Multi-objective programming approach for routing hazardous mate-
rials. Journal of Transportation Engineering 115, 661–673.
Zografos, K.G., Samara, S. (1989). Combined location–routing model for hazardous waste transporta-
tion and disposal. Transportation Research Record 1245, 52–59.
Zografos, K.G., Vasilakis, G.M., Giannouli, I.M. (2000). A methodological framework for developing
a DSS for hazardous material emergency response operations. Journal of Hazardous Materials 71,
503–521.
C. Barnhart and G. Laporte (Eds.), Handbook in OR & MS, Vol. 14
Copyright © 2007 Elsevier B.V. All rights reserved
DOI: 10.1016/S0927-0507(06)14010-4

Chapter 10
Traffic Equilibrium
Patrice Marcotte
Department of Computer Science and Operations Research, University of Montreal,
Montreal, QC, Canada H3C 3J7
E-mail: [email protected]

Michael Patriksson
Department of Mathematics, Chalmers University of Technology,
SE-412 96 Gothenburg, Sweden
E-mail: [email protected]

1 Background

The subject of traffic equilibrium is the description, through analytical tools,

of the stationary distribution of vehicles in a transportation network. Assum-
ing that travelers seek to minimize their individual travel cost, an equilibrium is
reached when no traveler has an incentive to modify its travel decision. Histor-
ically, the term traffic assignment was used to describe the same phenomenon,
reflecting the fact that the practice was not so much of estimating the traffic
distribution through analytical models – including design or pricing aspects –
than performing an assignment of travelers onto the network, typically in order
to assess the performance of traffic control policies.
Traffic equilibrium is the cornerstone, or the ‘inner loop’, of any modern
network analysis. Its definition dates back to the 1950s and hundreds of pa-
pers have been devoted to its study. In the last two decades, it has served as
the quintessential benchmark by which the performance of equilibrium algo-
rithms could be measured, and became a field of research of its own. In fact,
much of the early development of algorithms for finite-dimensional variational
inequality problems defined over convex sets took place within this field, and
at one point or another several of the foremost researchers in mathematical
programming and operations research have contributed to the theory of traf-
fic equilibrium. Presenting a comprehensive account of the traffic equilibrium
problem (or TEP, for short) would however require an entire book and, since
such books already exist, is beyond the scope of the current chapter.
In the 21st century, we understand rather well how to solve the basic TEP;
consequently, the challenges and interests of the transportation community
have shifted toward enhanced models that generalize the basic equilibrium
model. In this chapter, the authors briefly present the main theoretical and al-
gorithmical results pertaining to the TEP, along the way improving theoretical
results that were established some 20 years ago, and then focus their attention

623
624 P. Marcotte and M. Patriksson

on topics that have been overlooked, such as the relationship between the Nash
and Wardrop concepts, as well as on new paradigms built around the basic
TEP. In that respect, the presentation style and choice of topics are highly per-
sonal, and strongly reflect the authors’ inclinations. Each section is completed
by a ‘Bibliographical notes’ section, where we outline the relevant literature
and briefly mention topics not covered in the main text. Finally, the appen-
dices provide a list of notation, as well as a primer on variational inequalities,
which constitute the adequate framework for modeling network equilibrium
problems, as was recognized independently by Stella Dafermos and Michael
J. Smith more than 25 years ago.

1.1 Bibliographical notes

Most books on quantitative transportation analysis discuss the issue of traf-

fic assignment. A book entirely devoted to the topic has been written by
Patriksson (1994b). Among several surveys in handbooks, let us mention that
of Florian and Hearn (1995). These works all assume that costs increase with
volumes, although this does not hold in reality, where flow rates may decrease
as the network becomes congested. Although such effects are difficult to assess
in a static model, let us mention the proposal by Nesterov (2000) for addressing
this issue in situations where congestion levels are not critical.
Much of the analysis in this chapter can be directly transposed from traf-
fic networks to computer communication networks, where origin–destination
pairs are connected computer communication centers, and where vehicles cor-
respond to message packets being routed between them. In the fixed demand
case – the most natural framework in computer communication networks – the
goal is to minimize total (that is, average) delay, which usually is modeled as
that of minimizing the function
 vl
φ(v) := + pl vl 
c l − vl
l∈L
where vl denotes the arrival rate of packets at link l, cl is the transmission
capacity of the link, and pl is the processing and propagation delay on the link.
Provided that the capacities cl are sufficiently large, i.e., capacity constraints
are not tight at equilibrium, most of the results presented in this chapter go
through with few alterations.

2 The basic theme

This section is concerned with the basic traffic equilibrium model, includ-
ing the elastic demand case, together with its many mathematical statements.
Indeed, Wardrop’s conditions, which express a variational principle, can be for-
mulated in terms of a complementarity, variational inequality or optimization
program, either in terms of route or link flows. Throughout the section, key
 Ch. 10. Traffic Equilibrium 625

notations and definitions are introduced, while a small numerical example will
help in absorbing them. A comprehensive list of the notation is provided in Ap-
pendix B.

2.1 The Wardrop conditions

The basic elements of the standard traffic equilibrium model are: (i) a trans-
portation network, (ii) travel requirements, and (iii) cost functions,1 from
which traffic volumes (or flows), expressed as vehicular rates, must be deduced.
More precisely, let us consider a directed graph G := (N  L), consisting of a
set of nodes i ∈ N and directed links l ∈ L, sometimes also denoted l = (i j)
by the tail and head nodes i and j, respectively. Consider also a set C ⊂ N × N
of origin–destination (OD) pairs (p q), defining starting and ending nodes of
network trips in G . Throughout the section we consider an example based on
the graph illustrated in Figure 1, that involves five nodes, eight links and two
OD pairs: the first has node 1 as starting node and node 5 as ending node, and
the second OD pair has starting and ending nodes 2 and 4, respectively.
Any well-founded traffic model recognizes the individual network user’s
right to decide when, where, and how to travel. A traffic equilibrium model,
in which one aims at providing a macroscopic description or prediction of the
traffic volume resulting from route choices made in the traffic network, must
therefore be based on a sound route-choice behavioral principle.
The equilibrium condition refers to the concept initially introduced by the
statistician J.G. Wardrop of the British Road Research Laboratory. Ever since
his seminal paper, user equilibrium conditions have been known as Wardrop’s
first principle. The precise definition is as follows:

Fig. 1. An example traffic network.

1 Throughout the chapter, the term ‘cost’ refers to the disutility experienced by a user of the network.
In the basic models, it is a synonym for ‘travel time’ or ‘travel delay’.
626 P. Marcotte and M. Patriksson

The journey times on all the routes actually used are equal, and not greater than
those which would be experienced by a single vehicle on any unused route.

Wardrop also introduced a second principle, whereby users choose their routes
such that the average (or total) travel cost is minimal. This concept will be used
in Section 2.9.
To translate the concept of user equilibrium into mathematical terms, we
denote by Rpq the (finite) set of simple (cycle-free) routes for OD pair (p q),
by hr the volume of traffic on route r ∈ Rpq , and by cr the travel cost on the
route as experienced by an individual user, given the current volume of traffic.
Fixing the travel costs to these values, the user equilibrium conditions can be
written as follows:
hr > 0 ⇒ cr = πpq  r ∈ Rpq  p q ∈ C  (1a)
hr = 0 ⇒ cr  πpq  r ∈ Rpq  p q ∈ C  (1b)
where πpq denotes the minimal (that is, equilibrium) route cost for OD pair
(p q).
These conditions express the optimality conditions of a shortest route prob-
lem for each OD pair, where route costs are given by cr . What makes the
problem more complex than a shortest route problem is that the supply and
demand characteristics are not fixed: normally, the route costs cr depend on
the volume of traffic on the routes, and the total traffic volume in an OD pair
(that is, the OD demand) may depend on the least cost πpq of travel, perhaps
even upon other OD pairs’ costs as well. In order for the above system to de-
scribe an equilibrium state, we must further incorporate the cost perception of
the users given the volume of traffic, and the mechanism by which flow demand
is generated.
|R|
Let therefore cr : + →  be a real-valued function, describing the cost
of utilizing route r ∈ R, and whose argument is the vector h ∈ |R| of route
flows hr . The vector of these route costs forms the aggregated vector-valued
|R|
function c : + → |R| . Further, we assume that the demand on OD pair
(p q) ∈ C is given by the real-valued, nonnegative function gpq : |C | → + ,
whose argument is the vector π ∈ |C | of OD pair least travel costs, πpq ,
introduced in (1). The vector-valued demand function of the least travel costs
|C |
is denoted by g : |C | → + . We also introduce the route–OD pair incidence
matrix  ∈ |R|×|C | whose element γrk is set to 1 if route r joins OD pair
k = (p q) ∈ C , and 0 otherwise.
Based on the above notation, one may express the conditions (1) and the
demand constraints as the system:
 
0|R|  h ⊥ c(h) −  π  0|R|  (2a)
 h = d (2b)
d = g(π) (2c)
 Ch. 10. Traffic Equilibrium 627

where, for two arbitrary vectors a b ∈ n , the notation a ⊥ b means that

a b = 0. An interesting special case of the above conditions occurs when de-
|C |
mand is fixed, that is, when g ≡ d̄ ∈ + . In the sequel, we denote by Hd the set
of couples (h d) that satisfy equalities (2b)–(2c), and by H the corresponding
set when demand is fixed.
Formulation (2) is a special instance of a mixed complementarity problem; it
is the basic traffic equilibrium model from which we will derive several equiv-
alent ones, in the forms of variational inequality, nonlinear complementarity,
and optimization problems. This framework involves the pair (h π) and the
demand vector d ∈ |C | , although one could substitute its value g(π) to the
latter. If demand is fixed, condition (2c) vanishes.
The notion of equilibrium, as described in (2), should be thought of as a
steady-state evolving after a transient (disequilibrium) phase in which travelers
successively adjust their route choices, seeking to minimize travel costs under
prevailing traffic conditions, until a situation with stable route travel costs and
route flows has been reached.

2.2 Link flow representations

As the number of routes in a practical application can be enormous, we are

interested in developing representations of the Wardrop conditions that are
defined in terms of link flows only.
Suppose that for every route r ∈ R, the route cost cr (h) is additive, i.e., cr (h)
is the sum of the costs of using all links defining r. Let  ∈ {0 1}|L|×|R| be the
link–route incidence matrix, whose element λlr equals one if route  r ∈ R uti-
lizes link l ∈ L, and zero otherwise. We can then write cr (h) = l∈L λlr tl (v),
or, in compact form,
c(h) =  t(v) (3)
where v ∈ |L| is the vector of total volumes of traffic on the links. The cost
|L|
tl (v) of traversing link l ∈ L at the flow v is given by a function tl : + → ;
|L|
the vector of these link costs define the function t : + → |L| .
Implicit in the relationship (3) is the assumption that the pair (h v) is con-
sistent, in the sense that v equals the sum of the route flows:
 
vl = λlr hr  l ∈ L
(pq)∈C r∈Rpq

or, in vector-matrix form,

v = h (4)
From the above we derive the formula c(h) =  t(h) for evaluating a set
of route costs consistent with given route flows. Adding the two equations (3)
and (4) to (2) yields a formulation where route flow variables are relegated to
the constraints.
628 P. Marcotte and M. Patriksson

In our example, the link flow vector v is v = (v13  v14  v21  v23  v25  v34  v35 
v45 ) . The graph in Figure 1 then corresponds to the following link–route in-
cidence matrix: in the first OD pair we identify three loop-free routes, namely
the node ordering 1 → 3 → 5, 1 → 4 → 5, and 1 → 3 → 4 → 5, while
the second OD pair has the three routes 2 → 1 → 4, 2 → 3 → 4, and
2 → 1 → 3 → 4. The corresponding link–route matrix  ∈ 8×6 and vector t
of link cost functions are defined as
⎛1 0 1 0 0 1⎞
⎜0 1 0 1 0 0⎟
⎜0 0 0 1 0 1⎟
⎜ ⎟
⎜0 0 0 0 1 0⎟
=⎜
⎜0
⎟ and
⎜ 0 0 0 0 0⎟ ⎟
⎜0 0 1 0 1 1⎟
⎝ ⎠
1 0 0 0 0 0
0 1 1 0 0 0
⎛ ⎞ ⎛ ⎞
t13 (v13 ) 1 + 12 v13 
⎜ t14 (v14 ) ⎟ ⎜ 1 + v14  ⎟
⎜ ⎟ ⎜ ⎟
⎜ t21 (v21 ) ⎟ ⎜ 1 + 5v21  ⎟
⎜ ⎟ ⎜ ⎟
⎜ t23 (v23 ) ⎟ ⎜ 2 + 3v23  ⎟
⎜ ⎟ ⎜ ⎟
t(v) = ⎜ ⎟ := ⎜ 4 + v25  ⎟ 
⎜ t25 (v25 ) ⎟ ⎜ ⎟
⎜ t34 (v34 ) ⎟ ⎜ 3 + v34  ⎟
⎜ ⎟ ⎜ ⎟
⎝ t35 (v35 ) ⎠ ⎝ 2 + 2v35  ⎠
t45 (v45 ) 1 + 4v45
In this network we also define the (separable) OD demand functions
g15 (π15 ) := 11−π15 and g24 (π24 ) := 11−π24 , with inverses π15 = g−1 (d15 ) =:
ξ15 (d15 ) = 11 − d15 and π24 = g−1 (d24 ) =: ξ24 (d24 ) = 11 − d24 , respectively.
The reader is invited to check that π ∗ = (8 9) , with g(π ∗ ) = (3 2) .
Further, a vector of equilibrium route flows is h∗ = (h∗11  h∗12  h∗13  h∗21  h∗22 
h∗23 ) = (2 1 0 1 1 0) , and a vector of equilibrium link flows is v∗ =
(2 2 1 1 0 1 2 1) . The corresponding link costs are t(v∗ ) = (2 3 6 5 4 4
6 5) . Regarding the equilibrium conditions, we observe, for the first OD pair,
that the route costs are 8, 8, and 11, respectively. According to the equilibrium
principle, the more costly third route will not be used. Similarly, in the second
OD pair, the route costs are 9, 9, and 12, respectively. The reader may check
that the equations v∗ = h∗ and c(h∗ ) =  t(v∗ ) are satisfied.
Another formulation involves the origin–destination flows, also referred to
as commodity flows,2 and dispenses with route flow variables. Towards this aim,
we introduce the link–node incidence matrix E ∈ {−1 0 1}|N |×|L| , whose ele-

2 The term commodity stems from the economics literature, where an OD flow is associated with a
specific good.
 Ch. 10. Traffic Equilibrium 629

ment eil equals −1 if node i is the origin node of link l, 1 if node i is the
destination node of link l, and 0 otherwise. The link–node version of Wardrop’s
equilibrium conditions states that at an equilibrium link flow v, equal to the ag-
gregate of commodity (OD pair) volumes wk ∈ |L| , k := (p q) ∈ C , there
exist vectors π k ∈ |N | , k ∈ C , of node prices (alternatively node potentials or
dual variables) such that for a given link (i j) ∈ L,
 
0  wijk ⊥ tij (v) − [πjk − πik ]  0 k ∈ C  (5)
or
 
0|L|  wk ⊥ t(v) − E π k  0|L|  k ∈ C (6a)
To confirm the agreement with (2a), we select any OD pair k = (p q) ∈ C and
a route r ∈ Rpq , and consider a consistent set of volumes w, v, and h. Summing
the above conditions (5) over route r, we obtain that
  
λlr tl (v) − [πjk − πik ] = cr (h) − [πqk − πpk ]
l=(ij)∈L

and since l=(ij)∈L λlr wkl = hr can be made to hold by the Flow Decom-
position Theorem 2 (see below), we conclude that (2) and (6) are equivalent,
provided we identify πk with πqk − πpk .
Indeed, let us introduce the indicator vector ik ∈ {−1 0 1}|N | , which is
zero in all positions but two where, by the sign convention introduced earlier
for the incidence matrix E, the element with value 1 (−1) corresponds to the
sink (source) node. The demand-feasibility relation (2b) is then expressed as
 
Ewk = ik gk iκ π κ (κ∈C )  k ∈ C  (6b)
In the case of fixed demands, the latter right-hand sides are simply replaced
by ik d̄k , where d̄k ∈ + is the fixed demand for the OD pair k. We will later
establish under which additional conditions the representations (2) and (6) of
Wardrop’s user equilibrium conditions are equivalent.
Some comments are in order. First, the vector ik relates the OD node price
vectors π k ∈ |N | appearing in the link–node flow representation (6) and the
OD least cost vector π ∈ |C | appearing in the link–route version (2), through
the relation
 
π = (πpq )(pq)∈C = ik π k (k∈C ) ;
this is the same as stating that, for each k ∈ C , πk = πqk − πpk , as discussed
above.
This relationship allows for a simplification of the right-hand side of (6b).
Further, the vectors ik clarify the relationships between the respective de-
mand vectors. In link-flow based transportation problems, the (fixed) demand
in some commodity k is denoted by an |N |-vector dk , wherein a positive
630 P. Marcotte and M. Patriksson

(negative) element represents a sink (source) node and a null element a trans-
shipment node. In the traffic equilibrium problem, commodities are normally
associated with OD pairs, whence the vector dk just mentioned equals the
|N |-vector ik d̄k , which has precisely two nonzero entries.
Returning to our numerical example, the node–link incidence matrix
E ∈ 5×8 is
⎛ ⎞
−1 −1 1 0 0 0 0 0
⎜ 0 0 −1 −1 −1 0 0 0 ⎟
⎜ ⎟
E=⎜ 1 0 0 1 0 −1 −1 0 ⎟ 
⎝ 0 1 0 0 0 1 0 −1 ⎠
0 0 0 0 1 0 1 1
A commodity link flow in equilibrium is given by w∗1 = (0 1 1 1 0 1 0 0)
and w∗2 = (2 1 0 0 0 0 2 1) ; notice that w∗1 + w∗2 = v∗ . Further, we can
relate the equilibrium OD travel costs π ∗ = (9 8) to the node prices associ-
ated with the above network at equilibrium. Under the natural ordering of the
nodes, the node price vector for the respective OD pair is
⎛ ⎞ ⎛ ⎞
0 6
⎜ +∞ ⎟ ⎜0⎟
⎜ ⎟ ⎜ ⎟
π ∗1 = ⎜ 2 ⎟ and π ∗2 = ⎜ 5 ⎟ 
⎝ 3 ⎠ ⎝9⎠
8 5
as a simple calculation, (in principle) using Dijkstra’s algorithm, would show.
With the vectors i1 = (−1 0 0 0 1) and i2 = (0 −1 0 1 0) , we see that
the difference in node prices between the ending and starting nodes of the
two OD pairs yield precisely the elements of π ∗ = (9 8) , which confirms
the above connection. This also shows the familiar behavior of node prices in
network flow models: their values are only given up to an arbitrary common
additive constant; we have resolved this problem by setting the node price for
the starting node in each OD pair to zero.
In Section 2.7 we show that suitable properties of the demand and travel
cost functions imply that the entities π ∗ , d∗ = g(π ∗ ), and v∗ are unique at
equilibrium, while h∗ , and w∗ are not guaranteed to be unique. The reader is
asked to verify that in our case they are unique regardless.

2.3 Variational inequality representations

Variational inequalities provide a convenient framework for the modeling

of equilibrium problems, and are briefly reviewed in Appendix A. From now
on we often use the short-hand ‘VIP’ to refer to a variational inequality prob-
lem. The notation VIP(f X) refers to a variational inequality problem of the
following form: find a vector x∗ ∈ X such that
f(x∗ ) (x − x∗ )  0 x ∈ X
 Ch. 10. Traffic Equilibrium 631

where X ⊆ n is nonempty and closed and f : n → n is continuous.

An equivalent statement of the variational inequality problem that will also
often be used is the following:
−f(x∗ ) ∈ NX (x∗ )
where NX (x) is the normal cone to the set X at x ∈ X. See also (94).
Link-based variational inequality models are readily developed from
model (2) by incorporating the definitional constraints (4), and from the
model (6) by adding the compatibility constraint

v= wk  (7)
k∈C

Starting from the link–route representation, the fixed demand model is

 where
VIP(t F),
 
 := v ∈ |L| | ∃h ∈ H with v = h 
F
In the case of invertible demand functions, we can write the elastic model as
−t(v) ξ(d) ∈ NFd (v d) (8)
where
 
d := (v d) ∈ |L| × |C | | ∃(h d) ∈ Hd with v = h 
F
The corresponding link–node representations are as follows. The fixed demand
model is VIP(t F), with
 
 |L|
F := v ∈ |L|  ∃wk ∈ + with Ewk = ik d̄k  k ∈ C 
 
and vl = wlk  l ∈ L
k∈C

and the elastic demand model is

−t(v) ξ(d) ∈ NFd (v d) (9)
where
 
 |L|
Fd := (v d) ∈ |L| × |C |  ∃wk ∈ + with Ewk = ik dk  k ∈ C 
 
and vl = wlk  l ∈ L 
k∈C
 and F
Note that both sets F d are implicitly defined, and therefore not avail-
able in closed form (since the routes in R normally are not enumerated and
therefore unavailable), whereas the sets F and Fd are explicit.
632 P. Marcotte and M. Patriksson

2.4 A fixed demand reformulation of the elastic demand model

Suppose that the demand function g is upper bounded by a positive vec-

tor ḡ ∈ |C | , that is, g(π) < ḡ holds for every π ∈ |C | . Then, it is possible
to construct a fixed demand model which is equivalent to the elastic demand
Wardrop conditions (2). The model is constructed as follows. For each OD
pair (p q) ∈ C , we introduce an additional link from node p to node q which
carries a nonnegative flow epq , and which has as its link cost ξpq (ḡ − e),
e ∈ |C | . This link carriesa flow which is the difference between the upper
bound ḡpq and the flow r∈Rpq hr on that OD pair in the original network;
hence,  h + e = ḡ holds. We call this new link the excess demand link for OD
pair (p q). See Figure 2 for an illustration.
The form of this fixed demand model is such that, at equilibrium, the excess
demand link carries some flow. Therefore, the costs on all used routes on that
OD pair must match the value of the inverse demand function for that OD
pair, and the flow in the original network equals that specified by the demand
function on a least-cost route. We conclude that the fixed demand equilibrium
in the modified network is equivalent to an elastic demand in the original net-
work. The complete fixed demand model consists in finding flows (h e) such
that for some vector π ∈ |C | ,
 
0|R|  h ⊥ c(h) −  π  0|R|  (10a)
|C |
  |C |
0  e ⊥ ξ(ḡ − e) − π  0  (10b)
 h + e = ḡ (10c)
We remark that this model is equivalent to a fixed demand VIP of the form
 
c(h)
− ∈ NHḡ (h e) (11a)
ξ(ḡ − e)
where
 |R| |C |  
Hḡ := (h e) ∈ R+ × R+  h + e = ḡ  (11b)
We further note that the excess demand cost

κ(e) := ξ(ḡ − e) (12)

whose domain is the open box {e ∈ |C |
| 0|C |
< e < ḡ}, is strictly monotone if
and only if −g is. For this model, we have the following result.

Proposition 1 (Equivalent fixed demand model). Suppose that the function

g : |C | → |C | is invertible, with inverse ξ : |C | → |C | . Suppose further that g
is upper bounded by a positive demand vector ḡ ∈ |C | . Then, the fixed demand
model (10) is equivalent to the elastic demand model (2).
 Ch. 10. Traffic Equilibrium 633

Fig. 2. The excess demand link for OD pair (p q).

2.5 The equivalence between equilibria in different representations

The models discussed in Section 2.3 are distinct. The difference between the
models VIP(t F) (respectively, VIP([t −ξ] F d )) and VIP(t F) (respectively,
VIP([t −ξ] Fd )) lies in possible occurrence of cycles in the latter’s admissible
sets. (Note that the definition of the sets Rpq specifies that routes are cycle-
free.) We therefore have that F  ⊆ F.3 The equivalence between the link–route
and link–node representations in terms of equilibria thus hinges on whether
an equilibrium flow in the latter can contain a cycle; we provide such a result
below. The interest in this topic is not only theoretical, as a poor modeling
clearly can lead to spurious results; normally, in traffic networks, no one would
travel in a cycle, but the improper modeling of a traffic equilibrium problem
may lead to such a result in theory, when using the F (or Fd ) representation of
flows. We note that no cycles are present in the numerical example.
Since the demand part of the model plays no role in the analysis, we can
focus on the fixed demand case. The result below carries over trivially to elastic
demand models involving bounded demand functions. The proof rests on the
following flow decomposition result.

Theorem 2 (Flow Decomposition Theorem). Every route and cycle flow has
a unique representation as nonnegative link flows. Conversely, every nonnegative
link flow may be represented as a route and cycle flow (though not necessarily
uniquely), which utilizes at most |L| + |N | routes and cycles, and of which at most
|N | are cycles.

We say that a link flow v ∈ F is weakly acyclic if there exists a decomposition

into commodity link flows wk that does not use all links of a directed cycle.
The flow is said to be strongly acyclic if there exists no decomposition into com-
modity link flows wk that use all the links of a directed cycle. The following

3 The “  is bounded, while F may be unbounded.

· ” notation reflects that the set F
634 P. Marcotte and M. Patriksson

relationships are immediate from the Flow Decomposition Theorem:

 ⊆ F
{v ∈ F | v is strongly acyclic} ⊆ {v ∈ F | v is weakly acyclic} ⊆ F

The decomposition of a link flow solution can of course be performed with

respect to other entities than routes and cycles, as was done in the Flow De-
composition Theorem 2. For example, for a fixed link flow vector v ∈ F,
consider the linear program to

minimize 1 v̂ (13a)

(v̂δ)

subject to v̂ ∈ F (13b)
v̂ + δ = v (13c)
|L|
δ0  (13d)

The result of this linear program is a decomposition of the link flow v into a
sum v̂ + δ of a strongly acyclic link flow v̂, and a circulation link flow δ, that
is, a flow which satisfies Eδ = 0|N | . Moreover, v is strongly acyclic if and only
if v = v̂. This linear program, thus, also provides a decomposition of a link
flow in terms of extreme points and directions of the individual flow sets Fpq ,
although not disaggregated into routes and cycles in the individual commodity
spaces. It is therefore different from the representation described in the Flow
Decomposition Theorem, as there may be a circulation in v̂, although such a
flow must necessarily include more than one OD pair.
It follows from the above analysis that one has to be careful when imple-
menting algorithms based on link (vs. route) flows. Indeed, convex combina-
tions of cycle-free flows may fail to be cycle-free, unless restrictive assumptions
hold.
Finally, we say that the link cost vector t is cycle-wise nonnegative (respec-
tively, cycle-wise positive) on F if the sum of the link costs of every cycle is
nonnegative (respectively, positive) on F.

Proposition 3 (Equilibria for two representations).

 is
(a) If t is cycle-wise nonnegative on F, then every solution to VIP(t F)
also a solution to VIP(t F).

(b) Every weakly acyclic solution to VIP(t F) solves VIP(t F).
(c) If t is cycle-wise positive on F, then every solution to VIP(t F) is
strongly acyclic.

Clearly, then, if link costs are everywhere positive the issue of possible cyclic
equilibrium flows is avoided. One must however note that when considering
general link toll models, zero and even negative link costs may occur, as they
may appeal to certain travelers.
 Ch. 10. Traffic Equilibrium 635

2.6 Reduction to an optimization problem

 defines the
In the case where t is a gradient mapping,4 the model VIP(t F)
first-order optimality conditions for an optimization problem of the form
 v
minimize φ(v) := t(s) ds (14)

v∈F 0| L |

where denotes a line integral. The equivalence between equilibria, that is,
solutions to the model VIP(t F),  and stationary points to the problem (14)
follows immediately from the fact that ∇φ ≡ t. In particular, every stationary
point of problem (14) is an equilibrium link flow.
Further, if t is separable, so that tl is a function only of vl , l ∈ L, the opti-
mization problem assumes the more familiar form
  vl
minimize φ(v) := tl (s) ds (15)

v∈F 0
l∈L
−1 = ξ , we
When the demand function is separable and invertible, that is, gpq pq
similarly have the equivalent optimization problem
  vl   dpq
minimize φ(v d) := tl (s) ds − ξpq (s) ds (16)
d
(vd)∈F 0
l∈L (pq)∈C 0

It follows as above from the identity ∇φ ≡ (t −ξ) that elastic demand equi-
libria are stationary points for (16).
Since the cost and demand mappings are identical irrespective of the flow
representation, the traffic equilibrium models for the link–node flow represen-
tation correspond to optimization models that have the same form as above,
except that F  and F
d are replaced, respectively, by F and Fd . The reader may
provide the corresponding optimization formulation to the above numerical
example, and show that the equilibrium solution satisfies the first-order opti-
mality conditions associated with the mathematical program (16).

2.7 Properties of equilibria

In this section we derive the basic properties of traffic equilibria, i.e., the
existence and uniqueness of the following entities: flows, demands, and equi-
librium travel costs.5

4 If t is continuously differentiable relative to an open convex set containing S, then t is a gradient

mapping over S if and only if its Jacobian matrix ∇t(v) is symmetric over S. The gradient property
is a more general property than this symmetry property, since t need not always be differentiable;
nevertheless, the term ‘symmetry’ is frequently used in traffic planning when it should really refer to
the gradient property, also called the ‘integrability’ property.
5 For ease of presentation, we work with additive link costs throughout although, with regards to the
existence result in Theorem 4(a), this condition is not necessary.
636 P. Marcotte and M. Patriksson

We state the properties of the models that hold throughout this section:
(i) The network G is strongly connected with respect to the OD pairs6 ;
|L| |L|
(ii) the function t : + → |L| is single-valued and continuous on + ;
(iii) the function g : |C | → |C | is single-valued, continuous, nonnegative
and upper bounded on |C | .

2.7.1 Existence of equilibria

Theorem 4 (Existence of an equilibrium).
(a) Link–route representation. There exists a compact set of vectors
(h d π) of flows, demands, and least travel costs that solve the traffic equilib-
rium model (2).
(b) Link–node representation. Suppose further that the function c is addi-
tive, and that the function t is cycle-wise nonnegative. Then, there exists a compact
set of vectors ((wk )k∈C  v d (π k )k∈C ) of flows, demands, and node prices, solv-
ing the traffic equilibrium model (6). Moreover, the link flows ((wk )k∈C  v) can be
taken to be strongly acyclic.

Proof.
(a) The proof hinges on a general existence result for variational inequal-
ities, which we will utilize by converting the traffic equilibrium model
into an equivalent VIP involving a continuous operator defined over a
nonempty, convex and compact set. Since g is nonnegative and upper
bounded, the variable h resides in a bounded subset of the nonnegative
orthant; since c is continuous, the variable π resides in a compact subset
of |C | , although not necessarily in the nonnegative orthant. Hence, the
variables of the problem, (π h), lie in a compact set. We can therefore,
with no loss of generality, introduce additional, redundant, bounds for
each variable vector, such that

π  π  π̄
h  h̄
where π  π̄ ∈ |C | are small and large enough, respectively, and
h̄ ∈ |R| is large enough. Incorporating these constraints yields a VIP
model for (2) that involves a continuous operator defined over a non-
empty, convex and compact set. Hence, its solution set is nonempty and
compact.7 Furthermore, the bounds can clearly be selected such that
they all are fulfilled strictly at any equilibrium solution.

6 This condition means that there exists at least one route between each OD pair.
7 The idea is to construct, by means of the projection operator, a fixed point problem equivalent to the
original VIP, and then to apply Brouwer’s fixed point theorem.
 Ch. 10. Traffic Equilibrium 637

(b) Suppose that the vectors (h π) and d := g(π) solve (2), and set
v := h; the existence of a solution follows from the result in (a) above.
Then, the vector v solves the fixed demand traffic equilibrium model
 We now make use of Proposition 3(a), to conclude that v
VIP(t F).
also solves VIP(t F). We next need to establish that this flow is con-
sistent with the elastic demand model (6). To this end, we define the
|N |-vectors dk and π k , k ∈ C , by the use of the incidence vectors ik ,
k ∈ C , discussed in Section 2.2, and note that the vectors π k , k ∈ C are
consistent with the vector π. The first result then follows.
Last, we establish that the link flow v can be taken to be strongly
acyclic. Consider the perturbed link cost operator t + ετ · 1|L| , ετ > 0.
This mapping is cycle-wise positive, so Proposition 3(c) states that every
equilibrium link flow is strongly acyclic. Reasoning as above, and letting
{ετ } tend to zero as τ tends to infinity, we obtain the desired results,
since the set of strongly acyclic flows is compact. 

We note that since a fixed demand vector is a special case of the function g
stated previously, the above result provides an existence result in both the elas-
tic and fixed demand cases.

2.7.2 Uniqueness of equilibria

To establish the uniqueness of the equilibrium solution is advantageous from
several perspectives, but most of all from that of practice: using an equilibrium
model to make predictions about entities that are not unique at equilibrium
is a hazardous at best. Secondly, uniqueness is required for the convergence
of some algorithms that seek an equilibrium. Thirdly, when traffic equilibrium
is but a submodel in a more general model, uniqueness is crucial in order for
the overall model to be solvable, or even consistent. For example, a sensitivity
analysis is impossible to perform in entities that are not properly defined.
While the existence of an equilibrium was ensured under conditions that
are almost identical for all of the models that we have formulated, this is not
the case with the uniqueness issue. For example, there is a clear distinction
between uniqueness results for variational inequality models and optimiza-
tion models, when the cost mapping is not integrable. Further, uniqueness is
a property that depends on the level of aggregation. For example, route flows
(or, commodity link flows) are almost never unique at equilibrium, at least
when travel costs are additive. The reason is of course that a link flow v al-
most always can be distributed across the routes (or, disaggregated over the
commodities on the links) in infinitely many ways into a route flow h while
the relation v = h is satisfied (or, similarly, into an infinite variety of OD
link flows wk satisfying k∈C wk = v). In the results to follow, we will see that
the uniqueness of the entities c (and t if present), π, and d are ensured under
the mildest conditions. In contrast, stronger conditions are required to ensure
the uniqueness of even the link flow vector v.
638 P. Marcotte and M. Patriksson

The results below are provided for the link–route representation only. In-
deed, provided that the conditions of Theorem 4(b) hold, the assumptions that
imply uniqueness are the same for both representations. The theorem is stated
under fairly weak conditions that involve technical concepts that are detailed
in Appendix A.

Theorem 5 (Uniqueness of the equilibrium: The general case).

(a) Convexity. Suppose that (c −g) is monotone. Then, the sets of equilibrium
route and link flows (h v), least travel costs π, and travel demands d, are
convex. Suppose further that the travel demand is fixed. Then, the same
conclusion holds if the monotonicity requirement on c is replaced by the
weaker pseudo-monotonicity condition.
(b) Travel costs and demands. Suppose that (c −g) is monotone+ . Then both
the functions c and g are constant over the set of equilibria. This implies
that both the route costs c(h), least travel costs π, and travel demands d,
are unique. If c is additive, and t is monotone+ , then t(v) is unique on the
set of equilibria as well.
If further, the travel demand function is constant, the same conclusion
holds for the entities c(h) and π under the weaker requirement that c is
pseudomonotone+ +
∗ . If c is additive and t is pseudomonotone∗ , then t(v) is
unique on the set of equilibria as well.
(c) Link flows. Suppose that (c −g) is monotone, that c is additive, and that
t is strictly monotone. Then, the equilibrium link flow v is unique, as well
as c, π, and d, and the set of equilibrium route flows is a polytope.

Proof (Sketch). (a) That the solution set is nonempty and compact follows
from the existence Theorem 4. The convexity of the solution set and its entities
is a consequence of pseudomonotonicity and the properties of projections onto
convex sets. The reason why pseudomonotonicity is not enough for convexity
to hold in the elastic demand case is that pseudomonotonicity is not preserved
under addition.
(b) This result follows from a general one for the variational inequality prob-
lem VIP(f X): if f is pseudomonotone+ ∗ on X, then the value f(x) is constant
on SOL(f X). Again, the property pseudomonotone+ ∗ must be replaced by
monotone+ for the elastic demand case, because the former property is not
preserved under addition.
(c) Follows easily by arguing through contradiction. 

The result for the separable case follows as a corollary, because the prop-
erty of monotonicity+ of a mapping is identical to that of monotonicity when
it is integrable, and pseudomonotonicity+ ∗ reduces to pseudomonotonicity for
functions defined over .

Corollary 6 (Uniqueness of the equilibrium: The separable case).

 Ch. 10. Traffic Equilibrium 639

(a) Travel costs and demands. Suppose that each function gpq , (p q) ∈ C ,
is decreasing, and that each function tl , l ∈ L, is increasing. Then, the
route and link costs c and t, least travel costs π, and travel demands d, are
unique. In particular, if each function gpq , (p q) ∈ C , is invertible, then
it is strictly decreasing, and on the set of optimal solutions for the convex
optimization formulation (16), the equilibrium link travel costs t(v), least
costs π, and demands d are unique.
Suppose further that the travel demand is fixed. Then, the same conclu-
sion holds for the entities c, t, and π on the set of optimal solutions to the
convex optimization formulation (15) if the monotonicity requirement on
tl is replaced by pseudomonotonicity.
(b) Link flows. Suppose that each function tl , l ∈ L, is strictly increasing.
Then, in the solutions to the optimization formulations (15) and (16), the
equilibrium link flow v is unique.

To summarize in a more convenient and compact form, we have for the

elastic demand case:
t monotone −g strictly monotone ⇒ π d unique
+
(t −g) monotone ⇒ c(h) t(v) π d unique
t strictly monotone −g monotone ⇒ v c(h) t(v) π d unique
The first, and slightly weaker, of these results is not established above, but
follows easily by arguing through contradiction. For fixed demands, the corre-
sponding results are:
t pseudomonotone+
∗ ⇒ c(h) t(v) π unique
t strictly monotone ⇒ v c(h) t(v) π unique

2.8 Alternative notions of traffic equilibria and their relations

In this section, we present a few equilibrium conditions that have emerged

since the original one defined by Wardrop, and relate them to his. Some of the
motivation for this development is that when modeling the effect of congestion
pricing instruments, the cost term associated with the pricing scheme might
take the form of a (discontinuous) step function; in order to be able to study the
existence and computations of equilibria for such discontinuous travel costs, we
require that the traditional equilibrium conditions be generalized to allow for
certain types of discontinuities.

Alternative definitions
We consider (without any loss of generality) a fixed demand traffic network
G = (N  L). In this network, we assume that the traffic volume h ∈ H can be
observed. We further focus on an arbitrary OD pair (p q) ∈ C . More precisely,
640 P. Marcotte and M. Patriksson

we consider a route r ∈ Rpq for which hr > 0. We also consider a cost function
|R| |R|
c : + → |R| , each component of which cr : + →  is assumed to be
single-valued. Last, we consider a vector p in |R| describing the move of one
unit of traffic volume from route r ∈ Rpq to route s across the same OD pair,
that is,

−1 if i = r
pi = 1 if i = s
0 otherwise
The following definitions summarize the equilibrium conditions considered
in this section. Notice that we here distinguish the terms ‘Wardrop equilibrium’
and ‘user equilibrium’, although they are frequently used interchangeably in
this text. These notions coincide whenever the travel cost is continuous, as es-
tablished below, whence there is no confusion in most situations.

Definition 7 (Equilibrium conditions).

(a) Wardrop equilibrium. The traffic volume h is a Wardrop equilibrium if
cr (h)  cs (h) (17)
The traffic volume is a Wardrop equilibrium if for each driver, the
present travel cost on any alternative route is at least as high as the
present one on its present route.
(b) User optimized. The traffic volume h is user optimized if for some
α > 0,
cr (h)  cs (h + εp) ε ∈ 0 minimum{α hr }  (18)
The traffic volume is user optimized if any driver who switches routes
experiences a travel cost that is at least as high as the present one on its
present route.
(c) Equilibrated. The traffic volume h is equilibrated if
cr (h + εp)  cs (h + εp) ε ∈ [0 hr ] (19)
The traffic volume is equilibrated if any driver who switches routes
experiences a travel cost that is at least as high as the new one on its
present route.
(d) User equilibrium. The traffic volume h is a user equilibrium if
cr (h)  lim inf cs (h + εp) (20)
ε↓0

The traffic volume is a user equilibrium if no arbitrarily small packet

of flow can reroute itself and lower its travel cost.
(e) Normal equilibrium. The traffic volume h is a normal equilibrium if it
satisfies
c(y) (y − h)  0 y ∈ H (21)
 Ch. 10. Traffic Equilibrium 641

The definition states that h ∈ H solves the Minty variational inequal-

ity associated with the traffic system.

Relationships
In the following, we show how the above definitions are related to each
other.
 We assume here that the travel costs are additive, that is, cr (h) =
l∈L lr tl (v) for every consistent pair (h v) of route and link volumes. Note
λ
that the l.s.c. and u.s.c. properties8 are additive, so if link travel costs are l.s.c.
(respectively, u.s.c.) then the route costs are l.s.c. (respectively, u.s.c.) as well.

Proposition 8 (Relationships between equilibrium definitions). Let the route

travel cost c be additive.
(a) Wardrop equilibria vs. user optimized flows. Suppose that the route cost
function c is continuous. Then, user optimized flows are Wardrop equilib-
ria. If the route cost function c is in C 1 on H and, for every h ∈ H and
routes r s ∈ Rpq , (p q) ∈ C , the inequality
∂cs (h) ∂cs (h)
 (22)
∂hs ∂hr
holds, Wardrop equilibria are user optimized flows. In particular, if the link
cost function t is separable and increasing, Wardrop equilibria are user op-
timized.
(b) Wardrop equilibria vs. user equilibrated flows. Suppose that the route
cost function c is continuous. Then, equilibrated flows are Wardrop equi-
libria. If the route cost function c is pseudomonotone on H, Wardrop
equilibria are equilibrated flows.
(c) Wardrop equilibria vs. user equilibria. Suppose that for each link l ∈ L,
tl is l.s.c. Then, Wardrop equilibria are user equilibria. If, for each link
l ∈ L, tl is u.s.c., then user equilibria are Wardrop equilibria. Whenever
the link cost is continuous, the concepts of Wardrop and user equilibria
coincide.
(d) Wardrop equilibria vs. normal equilibria. Suppose that the link cost is
l.s.c. Then, normal equilibria are Wardrop equilibria. If the link cost oper-
ator t is pseudomonotone, Wardrop equilibria are normal equilibria.
(e) User equilibria vs. normal equilibria. Suppose that the link cost is l.s.c.
Then, normal equilibria are user equilibria.

Proof. In the case of (b), that Wardrop equilibria are equilibrated flows, sup-
pose that c is pseudomonotone on H, and let h ∈ H be a Wardrop equilibrium.

8 A function φ : n →  ∪ {−∞ +∞} is lower semicontinuous (l.s.c.) at x ∈ n if φ(x) =

lim infy→x φ(y) x ∈ n , and upper semicontinuous (u.s.c.) at x if φ(x) = lim supy→x φ(y) x ∈ n .
A function φ is l.s.c. (respectively, u.s.c.) on the set S ⊆ n if it is l.s.c. (respectively, u.s.c.) at every
point x ∈ S. The function φ is continuous at x if it is both l.s.c. and u.s.c. at x.
642 P. Marcotte and M. Patriksson

Then, by definition and for all ε > 0 with h + εp ∈ H, there holds

 
0  c(h + εp) [h + εp] − h = εc(h + εp) p
= ε cs (h + εp) − cr (h + εp) 
where the inequality follows since h ∈ H is a Wardrop equilibrium and c is
pseudomonotone on H. As ε > 0, we are done.
In the case of (d), the result is established by referring to the equivalence
(under the assumption of pseudomonotonicity of t) between the variational
inequality problem VIP(c H) and the Minty formulation (21); see also (104).
In the proof of the first part (that solutions to the Minty variational inequality
are solutions to the standard variational inequality problem), continuity can be
replaced by lower semicontinuity. The reverse implication utilizes the notion
of pseudomonotonicity. For the other results, we refer to the Notes section. 

Remark 9 (On the relationships between the equilibrium conditions). The ma-
trix condition (22) appearing in (a) describes a restriction on the dependence
between costs on distinct routes. In broad terms, if routes that are alternatives
to each other do not interact too strongly, then Wardrop equilibria are user
optimized. Differentiable, separable and increasing link travel cost functions
satisfy the matrix condition (22).
Under weak conditions, viz. continuity and monotonicity, four of the con-
cepts collapse to Wardrop’s definition of user equilibrium. The outlier is the
user optimized concept, akin to a nonatomic Nash equilibrium solution.9 The
following counter-example shows that a Wardrop equilibrium may not be user
optimized when the travel cost is asymmetric.
Consider a two-node network with parallel routes (links) and with respective
costs
c1 (h1  h2 ) = 2h1 + 3h2 
c2 (h1  h2 ) = 2h2 + 25
Note that this mapping is monotone. If demand is equal to 15, the flow vector
h = (10 5) is an equilibrium, with common path cost 34. Now, let us shift
one flow unit from the second to the first path (δ = 1). The cost vector of
the new flow pattern h = (11 4) is c(h1  h2 ) = (34 33). Since the cost to
users that have switched path has decreased,10 the initial flow h is not a Nash
equilibrium. Actually, there exist neither user optimized (nor Nash) equilibria
for this example.

The relationships between the various concepts are illustrated in Figure 3.

9 A Nash equilibrium between finitely many players is achieved when no player, acting on its own, can
improve its payoff. A nonatomic game involves a continuum of players.
10 In fact, all user costs have decreased.
 Ch. 10. Traffic Equilibrium 643

Fig. 3. Relations among definitions of equilibria.

2.9 User versus system optimality

In his seminal paper, Wardrop introduced the concepts of user and system
equilibrium, the second depicting flows that minimize the total system cost
S(v) := t(v) v.
A question that naturally arises is: how different are equilibrium and system-
optimal flows? How much would be gained, in terms of efficiency, from a
centralized control of all traffic flow? Looking at the question from a worst-
case perspective, we want to determine
t(vUO ) vUO
r(Δ) = maximum  (23)
Δ t(vSO ) vSO
where vSO (respectively vUO ) denotes a system-optimal (respectively user-
optimal) flow pattern, and Δ regroups the data of the equilibrium problem,
that is, Δ = G ∪ g ∪ C ∪ t in the fixed demand case. This ratio r(Δ) is the price
of anarchy and, in order that it be well defined, we need that the system cost
associated with user equilibria be unique. This condition will be satisfied under
the monotonicity requirements outlined in Theorem 5.
In certain circumstances, for instance if the cost function is link-separable
and assumes the monomial form
μ
tl (vl ) = κl vl  κl  μ  0 l ∈ L
user and system-optimal flows coincide. This is not the case in general, as can
be observed in the example of Figure 4, where vUO = (vAB  vAC  vBC  vBD 
vCD ) = (4 2 2 2 4)11 and vSO = (3 3 0 3 3). In this situation, every user
is worse off at (user) equilibrium, and the price of anarchy is equal to 92/83,

11 This corresponds to assigning flow in equal proportions on each of the three routes of the network.
644 P. Marcotte and M. Patriksson

Fig. 4. The Braess paradox network.

as can readily be computed. The following theorem provides an exact value of

r(Δ) for the important class of polynomial cost functions.

Theorem 10 (The price of anarchy). If the mapping t is link-separable, increasing

and polynomial of degree p, the price of anarchy is equal to
 
(p + 1)1+1/p p
∈ Θ 
(p + 1)1+1/p − p ln p
If, further, t is affine and monotone (p = 1), then the price of anarchy is equal to
4/3.

Let us now define the steepness of a cost mapping t as

t(v) v
σ(t) = sup  (24)
l∈Lv∈F t(v) v + γl (v)tl (γl (v)) − γl (v)tl (v)
where β is a function such that ∇S(β(v)) = t(v).

Theorem 11 (The price of anarchy revisited). Let S denote the class of cost
mappings that are monotone gradient mappings and for which the system cost S is
convex. Then
r(Δ)  sup σ(t)
0=t∈S
holds.

Note that the above result is mainly of theoretical interest, as the compu-
tation of the steepness appears intractable for general asymmetric mappings.
Moreover, these bounds are unlikely to be tight in practical situations.

2.10 Bibliographical notes

While Wardrop was first in providing a complete mathematical description

of the user equilibrium conditions, the first to use the term equilibrium to de-
scribe the traffic pattern is perhaps the economist Knight (1924). The early
 Ch. 10. Traffic Equilibrium 645

historical development of the equilibrium concept within the transportation

field has been traced in Boyce (2004) and Patriksson (2006).
The presentation in Sections 2.1–2.8 follows that of the forthcoming book
by Patriksson (2006). The modeling development differs from the traditional
one (e.g., Aashtiani and Magnanti, 1981; Ferris and Pang, 1997; Facchinei and
Pang, 2003a) in which the elastic demand Wardrop conditions are transformed
into a nonlinear complementarity problem (NCP). The reason why we adopted
the mixed complementarity approach is that it is the most natural, and that it
possesses stronger properties: in the case of the NCP model, the equivalence
with the Wardrop conditions requires that travel cost function be nonnegative
everywhere – this is associated with having to impose the requirement that
π  0|C | holds. In the MCP model, such a condition is not needed. For a
general introduction to NCP and MCP problems, see Ferris and Pang (1997)
and Facchinei and Pang (2003a).
The fixed demand reformulation and Proposition 1, presented in Sec-
tion 2.4, stem from Patriksson (2006), and extend the work of Gartner (1980)
on problems with a separable costs structure. Sheffi (1985, Section 6.3) re-
marks that the selection of a good upper bound ḡ on the demand function is
crucial to the performance of the fixed-demand algorithm, and proposes means
to iteratively update it during the computational process, when a linearization
(Frank–Wolfe) strategy is adopted.
The Flow Decomposition Theorem 2 can be found, for example, in Bu-
sacker and Saati (1965, Theorems 3.7 and 7.2), Rockafellar (1984, Sections
4.B and 4.J), or Ahuja et al. (1993, Theorem 3.5). The linear program (13)
which can be used to decompose a link flow into the sum of a strongly acyclic
flow and a circulation flow is due to Gallager (1977). Proposition 3 in Sec-
tion 2.5 stems from Hagstrom and Tseng (1998); its extension to the case of
elastic demands is immediate and is found in Patriksson (2006).
The existence results of Section 2.7 (taken from Patriksson (2006)) are
new. This is surprising, given that variational inequality models of traffic equi-
librium problems have been around since the late 1970s (e.g., Smith, 1979;
Dafermos, 1980) and the optimization models in Section 2.6, corresponding
to problems with separable costs and demands, have been around since the
1950s (Beckmann et al., 1956). The improvements have been made possible
for two reasons; for the case of the link–route representation, we have already
remarked that the cost function need not be nonnegative. For the link–node
representation, we have been able to utilize the recent results on the acyclic
nature of equilibria, stated in Proposition 3. Cost additivity was assumed in
our presentation for simplicity only.
Some of the uniqueness results (taken from Patriksson (2006)) are like-
wise new, in particular regarding costs and demands. The classical result is
that t monotone, −g strictly monotone ⇒ π, d unique (e.g., Florian, 1979;
Aashtiani and Magnanti, 1981). That the cost mapping of a variational in-
equality is constant over the solution set whenever the mapping is monotone+
646 P. Marcotte and M. Patriksson

(see Crouzeix et al., 2000) is used to good effect, as well as the result that
“f monotone+ ” coincides with “f monotone” whenever f is a gradient.
The content of Section 2.8 is taken from Patriksson (2006). It is clear that
under continuity and some form of monotonicity of the travel cost function,
all definitions collapse to the traditional one. Proposition 8 is collected partly
from the following sources:
(a) For the first result, see Smith (1984, Theorem 1). See also Dafermos
(1971, Theorem 2.6; 1982) (the latter for the elastic demand case). For the
second result, see Heydecker (1986, Theorem 7). For the third result, see
Dafermos (1971, Theorem 2.6) and Heydecker (1986).
(b) That equilibrated flows are Wardrop equilibria is established in Hey-
decker (1986, Theorem 5). Our result improves upon it by reducing the re-
quirements to pseudomonotonicity.
(c) See Bernstein and Smith (1994, Theorem 2.1).
(d) See de Palma and Nesterov (1998, Theorem 4(i)); however, their result
is weaker in that monotonicity is assumed.
In Section 2.9, the worst-case behavior of Wardrop equilibria, with respect
to a system-optimal flow pattern, has been analyzed in Roughgarden and Tar-
dos (2002), Chau and Sim (2003), Correa et al. (2004). We remark that the
result that user and system optimal solutions coincide for the exponent μ = 0
was established first by Jorgensen (1963), while the result for μ  1 is due to
Dafermos and Sparrow (1969, Equation (1.34)); see also Bennett (1993).

3 Variations

3.1 Multi-mode and multi-attribute models

A model of traffic equilibrium where user behavior is captured by a sin-

gle function, either route-based or link-based, is a mathematical idealization
of real life. It fails to account for travel time (mis-)perceptions and for the
fact that delay is but one attribute of travel, together with cost, driving stress,
landscape, number of stops, vehicle type, etc. The first issue, mis-perception,
is best dealt with by using stochastic models (logit, mixed logit, probit) while
the second issue yields multi-mode or multi-attribute12 models. In multi-mode
models, distinct cost functions tm are assigned to distinct modes m ∈ M. One
possible functional form is
|M| 
 1  
tm v      v|M| = tm αi vi  (25)
i=1

12 We refrain from using the term ‘multi-criterion’ that usually refers to vector optimization or vector
equilibrium problems.
 Ch. 10. Traffic Equilibrium 647

where vm denotes the flow of mode m and αi may be interpreted as the car-
equivalent associated with mode m. An equilibrium is then characterized as a
solution to the variational inequality
   
−tm v1      v|M| ∈ NFm v1      v|M|  m ∈ M (26)
where Fm denotes the set of demand-feasible flows for mode m. The theo-
retical drawbacks of such a functional form have been pointed out by some
researchers, who suggested more complex but better behaved functional forms.
Multi-mode models can easily be converted to the standard model by assigning
to each mode its own copy of the transportation network, with a cost function
that becomes nonseparable, and most likely asymmetric.
Closely related to multi-mode models are multi-attribute (or multiclass)
models, where the disutility (generalized cost) of each user is a function of
route attributes, and the perception of disutility associated with each attribute
varies across the population. To fix ideas, let us consider a model with two
attributes, namely travel time t and travel cost f, the latter including both
out-of-pocket cost (tolls, for example) and variable costs such as petrol or
maintenance. Let us adopt monetary cost as numéraire, define αg as the value
of one time unit (VOT) associated with population group g ∈ G , and make the
hypothesis that the disutility of each class g assumes the linear form

ug (v) = αg t(v) + f(v) (27)

g
where v = (v1      v|G | ). Let us denote by dk
the set of demand-feasible
flow for population group g and origin–destination pair k, and by Fg the cor-
responding set of demand-feasible link flows. Due to the separability of the
feasible sets, the equilibrium is then characterized by the variational inequali-
ties
   
−ug v1      v|G | ∈ NFg vg  g ∈ G  (28)
Although the above model is subsumed, from the mathematical point of view,
by the variational inequality (26), its analysis yields insight into models involv-
ing distinct classes of vehicles or customers, and is well worth investigating for
its own sake. As an introductory example, let us consider a two-group, one-link
example, with cost mapping
 
  α1 t(v1 + v2 ) + f (v1 + v2 )
u v1  v2 = (29)
α2 t(v1 + v2 ) + f (v1 + v2 )
and with Jacobian matrix
 
∇u v1  v2
 
α1 t  (v1 + v2 ) + f  (v1 + v2 ) α1 t  (v1 + v2 ) + f  (v1 + v2 )
= (30)
α2 t  (v1 + v2 ) + f  (v1 + v2 ) α2 t  (v1 + v2 ) + f  (v1 + v2 )
648 P. Marcotte and M. Patriksson

It is instructive to check the monotonicity of the mapping u. To this aim, we

compute the determinant of the symmetrized Jacobian matrix:
 
det ∇u + ∇u
 
 2α1 t  (v1 +v2 )+2f  (v1 +v2 ) (α1 +α2 )t  (v1 +v2 )+2f  (v1 +v2 ) 
=  
(α1 +α2 )t  (v1 +v2 )+2f  (v1 +v2 ) 2α2 t  (v1 +v2 )+2f  (v1 +v2 ) 
  2   2
= 4α1 α2 − (α1 + α2 )2 t  v1 + v2 = −(α1 − α2 )2 t  v1 + v2 
This calculation shows that u can only be monotone if either α1 = α2 or the
cost function t is constant. In the first case, the problem reduces to the single-
group case while, if t is constant, congestion is absent. Now, since the feasible
domain is separable by group, one may scale each group cost mapping by an
arbitrary positive number, to derive an equivalent formulation of the equilib-
rium conditions. For instance, if the parameters α1 and α2 are not zero, we
can divide the cost functions by their respective VOT parameter, yielding the
equivalent cost mapping, which we still denote by u:
 1 
 1 2 t(v + v2 ) + α1 f (v1 + v2 )
u v v =  (31)
t(v1 + v2 ) + α2 f (v1 + v2 )
and where αg represents the inverse VOT of group g. Although both for-
mulations are equivalent from the equilibrium viewpoint, their mathematical
properties are quite different. Indeed, repeating the previous analysis, we ob-
tain that u is monotone if f (rather than t previously) is constant. This will be
satisfied if tolls and variable travel costs are flow-independent, a reasonable
first-approximation assumption. For this reason we will, from now on, switch
to travel time as numéraire. Note that the monotonicity results extend to the
general multi-attribute case, i.e., a sufficient condition that the cost mapping u
be monotone is that all cost functions, with the possible exception of travel
time, be flow independent. We shall see, in a more general setting, that this
condition is necessary as well.
A natural question that arises concerning the variational inequality involv-
ing the mapping u is the existence of a ‘natural’ optimization problem whose
set of stationary (Karush–Kuhn–Tucker) points coincides with the set of multi-
attribute equilibria. In other words: under what conditions is the mapping  ua
gradient mapping? Let C denote the index set of attributes and v̄ = i∈C vi
the total link flow vector. The answer to our question is that f be constant and
t be a gradient mapping, and the relevant optimization program is simply to
 v̄ 
minimize t(s) ds + αi f vi  (32)
v 0 i∈C

where C denotes the index set of criteria. Indeed, it is easy to check that the
optimality conditions of (32) fulfill the multi-attribute equilibrium conditions.
Furthermore, if the mapping t is monotone, the program (32) is convex. If t is
 Ch. 10. Traffic Equilibrium 649

strictly monotone, the total flow equilibrium is unique, although group and
origin–destination flows will not in general be unique.
We now focus our attention on a two-attribute model where the VOT (or
inverse VOT) parameter is continuously distributed across the population, and
described by a continuous probability density function h defined over the non-
negative axis. For ease of notation, we assume that the density h is identical for
all origin–destination pairs. For each origin–destination pair k, we then have
that the demand density for the α-group13 is equal to h(α)dk , and Fα = h(α)F
is the feasible set for the α-group. The variables of the equilibrium problem
are regrouped into a link flow density vector v(α) = {vl (α)}l∈L and, similar to
the finite-dimensional case, we denote by v̄ the total link flow vector which, in
this case, is defined by the integral
 ∞
v̄ = v(α) dα (33)
0
Replacing the index g by α yields the equilibrium conditions:
   
− t(v̄) + αf(v̄) ∈ NFα v(α)  α ∈ [0 ∞) (34)
The above formulation involves an infinite collection of variational inequali-
ties, one for each α-group. Upon introduction of the vectorial disutility func-
tion u(v̄) defined as
u(v̄) (α) = t(v̄) + αf(v̄) (35)

the vector function v = v(α)α0 and the set F = α Fα , the equilibrium condi-
tions (34) can be aggregated into the infinite-dimensional variational inequality
−u(v̄) ∈ NF (v) (36)
While this may be considered elegant from the mathematical point of view, the
approach has drawbacks. First, the proper setting for such formulation is the
set of square-integrable functions. Since two functions that differ over a set of
measure zero are equivalent, standard existence results cannot be invoked to
prove the existence of a solution to the system (35) for every value of the para-
meter α. Second, this framework hides the network structure of the problem,
which is essentially discrete.
Both these drawbacks can be remedied by considering finite-dimensional
formulations. The remainder of the section is devoted to this topic, while al-
gorithms will be discussed in Section 5.1. To ease the presentation,
 we assume
from now on that F denotes the unit simplex F = {v̄i  0 | ni=1 v̄i = 1} and
that h(α) = 0 for all α larger than some finite threshold ᾱ. Since the origi-
nal set F is a bounded polyhedron, the simplicial form of F can be achieved
by expressing each of the points of F as a convex combination of its vertices.

13 Although the terminology α-group is convenient, the notion actually relates to an infinitesimal user.
650 P. Marcotte and M. Patriksson

Alternatively, one may consider a traffic equilibrium problem involving one

origin, one destination and a unit demand. For this reason, we will use, with-
out any loss of generality, the term ‘path’ to denote an extreme point of the
unit simplex.
Our finite-dimensional approach is based on the fixed point formulation of
a variational inequality. In our setting, this takes the form:
 
v(α) ∈ h(α)F ∩ arg minimum t(v̄) + αf(v̄) w(α) α ∈ [0 ∞) (37)
w(α)∈Fα

A key observation is that the right-hand side of (37) is equivalent to the set of
linear programs
 
minimize t(v̄) + αf(v̄) w(α) α ∈ [0 ∞) (38)
w∈F
This is nothing else than a parametric linear program, where the parame-
ter α scans the nonnegative real axis. Because the number of extreme points
of F, that is, combinations of shortest paths, is finite, this infinite-dimensional
program can actually be solved in finite time by the parametric simplex algo-
rithm. To this aim, let us sort the paths in decreasing order of their slopes.
As the valuation of time decreases with α, the cost-conscious users will use
high-index paths, while the time-conscious users will travel on low-index paths.
The solution to the parametric linear program is characterized by a vector
α = (α0 = 0 α1      αn = ᾱ) of critical values, where users having an in-
verse VOT α belonging to the interval (αi−1  αi ) are assigned to path i (see
Figure 5). Several comments are in order:
(i) The flow on path i is given by the integral
 αi
w̄(v̄) = h(α) dα (39)
αi−1

If the total flow vector T (v̄) = (Ti (v̄))ni=1 agrees with v̄, then the solu-
tion to the parametric LP is an equilibrium for the bi-attribute problem.
Since this solution only depends on v̄, it can be recovered from the so-
lution to the finite-dimensional fixed-point problem w̄(v̄) ∈ v̄. Alterna-
tively, a finite-dimensional fixed point formulation can be built around
the vector of critical points α. This amounts to initializing with α the
sequence of evaluations
total flow v̄ → critical vector α → (through integration) w̄
(ii) The assignment of users whose inverse VOT is equal to one of the
critical points (α1      αn−1 ) is ambiguous. Generically, the number
of such points is finite, i.e., of zero Lebesgue measure.
(iii) In the degenerate situation where two or more paths have identical
t and f values over an interval (αi−1  αi ), users may be assigned to any
of these paths, a situation similar to that occurring in the standard,
single-attribute model.
 Ch. 10. Traffic Equilibrium 651

Fig. 5. Parametric path costs. Path 3 is dominated and therefore carries no flow.

(iv) If all slopes are distinct, the path assignment is unambiguous, with the
exception of at most n−1 values of the parameter α, and the solution of
the parametric LP is unique, almost everywhere. This regularity condi-
tion, which is easily achieved by perturbing the f -value of each path (or
link), allows for a finite-dimensional variational inequality formulation
and has beneficial algorithmic implications.
We now consider four important issues related to any fixed point or varia-
tional formulation, namely: existence and uniqueness of solutions, monotonic-
ity and the gradient property. In the infinite-dimensional setting, existence of
a solution can be proved using measure-theoretic arguments. In either finite-
dimensional fixed point formulation, one can actually show that the fixed point
mapping w̄ is upper semicontinuous whenever the ordering of path slopes
remain constant, and existence follows from Kakutani’s theorem. As in the
discrete case, monotonicity holds if and only if t is monotone and f is constant.
Similarly, the gradient property is satisfied if and only if t is a gradient mapping
(not necessarily monotone) and, again, f is constant. The equivalent convex op-
timization problem is to
 v̄   i 

n 
minimize t(s) ds + (fi − fi+1 )ν v̄j  (40)
v̄∈F 0 i=1 j=1

where ν = φ−1 and φ denotes the cumulative distribution function φ of the
density h.
The case for uniqueness deserves more attention. If t is constant and f is
strictly monotone, the (total flow) solution v̄ to is unique. Moreover, if path
slopes are distinct, the solution to the parametric LP (37) is unique, except for
the critical points αi . It follows that path flows are unique, a surprising result.
652 P. Marcotte and M. Patriksson

Indeed, the Jacobian matrix ∇u(v1  v2      v|G | ) of the disutility function u has
rank one and is therefore ‘less and less’ definite as the number of user groups
increases, making futile the standard argument for establishing uniqueness.
Although the situation might be expected to worsen in the infinite-dimensional
case, exactly the opposite occurs. This is reminiscent of stochastic assignment
models, for which path uniqueness can be established under weak regularity
assumptions.

3.2 Side constrained traffic equilibrium models

3.2.1 Introduction and equilibrium characterizations

Traffic flows are frequently subject to restrictions, physical or otherwise, that
have not been taken into account in the basic model. For example, in the pres-
ence of a stationary traffic control system, flows might have to obey constraints
of the form

sk (v)  0 k ∈ K (41)
where, for mathematical convenience, we assume that the functions sk :
|L|
+ → , k ∈ K, are convex and differentiable. Simple flow capacities on
some links, describing the stationary effect of the traffic control, correspond to
 ⊆ L, and
letting K := L

sl (v) := vl − cl  
l∈L
where cl > 0 is the (stationary) link volume capacity. Side constraints can also
be used to force equilibrium flows to comply with traffic management goals.
For instance, system optimal link flow patterns can be enforced by setting cl to
the elements of a system optimal link flow vector.
Unfortunately, under the presence of joint constraints, the Cartesian prod-
uct structure of the feasible sets of, for example, (15) and (16), are not satisfied
any more. This has rather far reaching consequences on the equilibrium char-
acterization of optimal flows. An equilibrium, in the sense of Wardrop, may
no longer exist in terms of the original cost structure. If, however, one incor-
porates into the cost the Lagrange multipliers associated with the side con-
straints (41), then one can describe an equilibrium in terms of a generalized
cost which, under some circumstances, can be given a natural interpretation.
We introduce multipliers βk  0 for each constraint in (41) and append to (2)
the additional condition

0|K|  β ⊥ −s(v)  0|K|  (42)

Letting S := {v ∈ |L| | sk (v)  0 k ∈ K} the side constrained inelastic
demand model becomes

−t(v) ∈ NF∩S (v) = NF (v) + NS (v) = NF (v) + ν ν ∈ NS (v);

 Ch. 10. Traffic Equilibrium 653

if S satisfies a constraint qualification, for instance if each function sk is affine,

then for some vector β satisfying (42), we have that, at equilibrium,
 !
− t(v) + βk ∇sk (v) ∈ NF (v)
k∈K

holds. Hence, for any optimal vector of Lagrange multipliers, v is a Wardrop

equilibrium
 volume in terms of the generalized link costs t̄(v) := t(v) +
k∈K kβ ∇s k (v). In the case of link capacities, this reduces to t̄(v) := t(v) + β,
where the elements of β for l ∈ / L are fixed to 0. This expression may be
interpreted as a generalized delay, where β is a vector of queueing delays at in-
tersections. While this interpretation may be valid, it is important to note that
the vector β of Lagrange multipliers is seldom unique, even in this simple case.
This implies that link queueing delays are not unique, even if link flows v are.
The following example illustrates how this situation can be utilized to improve
traffic conditions.
We consider the directed graph illustrated in Figure 6, that involves two
OD pairs, (1 5) and (2 5), with a demand of 2 and 3, respectively. Travel
cost functions and link capacities are given in Table 1. The solution to the
traffic equilibrium problem (that is, in the absence of upper bounds) is
v = (200 209 091 127 282 218) , while the solution to the capacitated
model is v∗ = (2 2 1 2 2 3) .
Applying an augmented Lagrangian algorithm (see further Section 5.2) to
this problem produces the multiplier vector β := (0 0219 0219 0 80 0) .
This is, however, far from the only multiplier vector in this problem. We first
note that the set of multiplier vectors β for this problem satisfies

β1 + β4 + β6 − π15  −18
β1 + β5 − π15  −10
β2 + β4 + β6 − π25  −24
β2 + β5 − π25  −16
β3 + β6 − π25  −24
2β1 + 2β2 + β3 + 2β4 + 2β5 + 3β6 − 2π15 − 3π25 = −92
β1  β2  β3  β4  β5  β6  0
β6 = 0
Clearly, this set is not a singleton, and is even unbounded. We see that link 6
is overcapacitated at v∗ , whence its multiplier must be zero by complementar-
ity, while the remaining link multipliers are nonnegative, by definition. If the
model is valid, it is then interesting to note that we may interpret the multi-
pliers βl as link tolls. These ensure that the uncapacitated traffic equilibrium
problem with link cost t(·) + β has the same solution v∗ as the capacitated
model with the original link travel costs. Actually, the set of multipliers shown
654 P. Marcotte and M. Patriksson

Fig. 6. A capacitated traffic network.

Table 1.
Network data

Link tl (vl ) cl

1: (1 3) t1 (v1 ) = v1 2
2: (2 3) t2 (v2 ) = 4v2 2
3: (2 4) t3 (v3 ) = 12v3 1
4: (3 4) t4 (v4 ) = 2v4 2
5: (3 5) t5 (v5 ) = 4v5 2
6: (4 5) t6 (v6 ) = 4v6 4

above is only a subset of interesting choices. It may even turn out to be ben-
eficial to impose negative tolls, or to set tolls on links that are not saturated,
in order to achieve some preset management target. This possibility brings
forward an alternative perspective on how to induce a favored equilibrium so-
lution.
Suppose that we wish to impose upon the users link tolls (be they positive or
negative) such that a link flow v∗ ∈ F  is in equilibrium. Then, if one interprets
the side constraints as an enforcement of the equality v = v∗ (that is, sl (v) =
vl − vl∗ , l ∈ L, and s|L|+l (v) = vl∗ − vl , l ∈ L), one observes that the set of link
tolls corresponds to the polyhedron defined by the solutions in β of the linear
system
 
 t(v∗ ) + β   π (43a)
 ∗  ∗
t(v ) + β v = d π (43b)
In the above example, this set corresponds to the removal of the last two con-
ditions, i.e., nonnegativity and complementarity.
Using the previous polyhedron as the feasible set of a mathematical pro-
gram, we may, for example, devise a minimum-revenue toll by minimizing the
 Ch. 10. Traffic Equilibrium 655

revenue function β → β v∗ , whose minimal value 16 is achieved for the toll

vector β∗ = (0 0 0 0 8 0) (compare with the value 16657 obtained from
the augmented Lagrangian solution mentioned above) and π = (18 24) .
Obviously, there does not exist a maximum-revenue toll schedule, since the
revenue can be made arbitrarily large along the line β = (0 γ γ 0 8 0) ,
γ  0. (In the case of the set (43), we can let γ → −∞ to prove that no
minimum-revenue toll exists.) Such toll problem will be investigated in more
detail in Section 6.2.
In the elastic demand case, the toll polyhedron that induces given link flow
and demand patterns (d∗  v∗ ) is
 
 t(v∗ ) + β   π (44a)
 ∗  ∗
t(v ) + β v = ξ(d∗ ) d∗  (44b)
that is, the immediate extension to the system (43). We note that the value of
β v∗ is constant over this set, i.e., the toll revenue is constant! This apparent
paradox can be explained by the fact that network users might not travel when
the travel cost becomes too large.

3.2.2 Wardrop-like principles

According to the above, side constrained equilibria satisfy the Wardrop con-
ditions in terms of generalized travel costs, but they will, in general, not satisfy
any similar conditions in terms of actual travel costs. One can therefore, in gen-
eral, not relate the actual travel costs of the unused routes to those of the used
ones; for example, the least costly route in an OD pair may be unused because
its generalized cost is too high. This deficiency is due to the fact that the side
constrained problem does, in contrast to the standard model, lack a Cartesian
product structure.14
Wardrop-like principles in terms of actual travel costs may however be es-
tablished if the following assumption is fulfilled at the solution to the side
constrained model.


Assumption 12 (Nondecreasing side constraint functions). At the flow v ∈ F,
∂sk (v)
 0 l ∈ L k ∈ K
∂vl
holds.

 then a flow increase along one

If this assumption holds for any flow v ∈ F,
or more links of the network can never result in any side constraint becom-
ing ‘looser’. Conversely, this assumption holds whenever the side constraints
correspond to general capacity restrictions, that is, when they represent upper

14 For capacitated networks, an alternative approach is analyzed in Section 3.3.

656 P. Marcotte and M. Patriksson

bounds on traffic volumes on certain links or routes, or within an area of the

traffic system. In the sequel, we shall use the notions of links and routes that
are unsaturated with respect to the side constraints.

Definition 13 (Unsaturated link and route). A link l ∈ L is unsaturated at the

 if for all k ∈ K,
flow v ∈ F
∂sk (v)
>0 ⇒ sk (v) < 0
∂vl
 if all links l ∈ L on route r are
A route r ∈ R is unsaturated at the flow v ∈ F
unsaturated.

 if ∂sk (v)/∂vl > 0 holds for

A route is clearly saturated at the flow v ∈ F
some link l on the route and some k ∈ K such that sk (v) = 0.

Theorem 14 (Wardrop-type principles). Suppose that (h∗  v∗ ) solves the side

constrained model and that Assumption 12 holds at v∗ . Then, the following con-
clusions hold for any OD pair (p q) ∈ C .
(a) The routes utilized in the OD pair have equal and minimal generalized
route costs.
(b) Assume, without loss of generality, that the first  routes are utilized in
the OD pair and that m of these are unsaturated. Then, the routes may be ordered
so that

c1 = c2 = · · · = cm  cm+1  cm+2  · · ·  c 
(c) For any pair of routes r s ∈ Rpq ,

route r is unsaturated
⇒ h∗s = 0
cs > c r
(d) For any pair of routes r s ∈ Rpq ,

route r is utilized
⇒ route s is saturated
cs < cr

The (simple) proof of the above results is based upon the equilibrium char-
acterization of a side constrained equilibrium.
If the implication in either of the results (c) and (d) were not fulfilled for
some pair of routes, then some traveler might shift to a less costly and un-
saturated alternative route; hence, these results are quite natural. As touched
upon above, the OD routes that are unused in a solution to the side constrained
model are not necessarily more costly (in actual travel cost) than those used in
the OD pair; this is implied by the result (d) since a route may be saturated at
zero flow.
 Ch. 10. Traffic Equilibrium 657

3.3 Strategic models

The standard traffic equilibrium models assume that path selection, per-
formed at the origin nodes, does not change en-route. This assumption is
reasonable as a first approximation, but is not entirely compatible with user
behavior. In this section, we outline an approach that allows users to take on-
line decisions, based on current traffic conditions. While such an approach is
common place in transit systems, where users may need to transfer one or more
times to reach their destination, its application to congested traffic assignment
is recent.
Let us first consider a transit system where, at each stop served by sev-
eral transit lines, waiting customers must make strategic choices of the form:
Should one board the incoming vehicle, or wait for another, more attractive
(quicker) transit line? Such a decision must balance the travel time of the in-
coming vehicle against travel times plus waiting times of vehicles yet to reach
the boarding station. Let us assume that the transit stop is served by two lines
with respective frequencies φ1 , φ2 , and travel times to destination t1 < t2 . As-
suming that the arrival process is random and memoryless (Poisson process),
and that users are rational and risk-neutral, they will naturally board a vehicle
of line 1 if it shows up first, and a vehicle of line 2 if the expected waiting time
of line 1, i.e., 1/φ1 , exceeds the difference t2 − t1 , i.e.,
1
t2  
φ 1 + t1
This reasoning can be generalized to n common bus lines indexed in decreasing
order of their travel times to destination. The optimal strategy is characterized
by a threshold index l̄, below which all transit lines are attractive, that satisfies
the relation
 
1  l
l̄ ∈ arg minimum l 1+ φ i ti  (45)
1ln
i=1 φi i=1

The key feature of this approach is that, while strategies are deterministic ob-
jects, the route traveled from day to day by users is stochastic. The applicability
of the method relies on the existence of efficient algorithms for computing op-
timal strategies. Such algorithms, akin to shortest path methods, are able to
address large-scale problems.
In private transportation, the situation is different, since randomness is asso-
ciated with demand (users) rather than supply (vehicles). Defining equilibrium
meaningfully in this context represents a nontrivial task. To gain some insight
into the situation, let us consider the network illustrated in Figure 7, where
each link is endowed with a cost (shown next to the corresponding link) and,
possibly, a capacity (bracketed number). Paths from origin node 1 to destina-
tion node 5 are listed, together with their features, in Table 2.
658 P. Marcotte and M. Patriksson

Fig. 7. A capacitated network.

Table 2.
Network paths for a small example

Path index Path Cost Capacity

1 1–3–5 100 2
2 1–2–3–5 175 5
3 1–2–5 200 ∞
4 1–3–4–5 250 2
5 1–2–3–4–5 325 ∞

Clearly, no Wardrop equilibrium is compatible with this data, since the

shortest path 1–3–5 cannot accommodate all demand of 10 units. Quite natu-
rally, one could settle for an equilibrium satisfying the following Wardrop-like
principle:
At equilibrium, the cost of a path with a positive residual capacity is larger or
equal to the cost of any path carrying positive flow.

In our small example, any path flow vector of the form

x = (5 − δ δ 5 0 0)
with δ ∈ [3 5] satisfies the above principle. However, whenever δ is strictly
greater than 3 and the users of the second path are endowed with the gift of
prescience, it is tempting for them to switch to the less costly first path on which
some capacity has been set free by themselves. In this sense, the equilibrium
associated with the value δ = 3 is more natural. Indeed this solution, associated
with a queueing delay of 75 on link (1, 3) and 25 on link (3, 5), corresponds to
the equilibrium notion of Section 3.2. Notwithstanding the queueing delay, it
is a system-optimal solution as well.
 Ch. 10. Traffic Equilibrium 659

Table 3.
A set of strategies for the small example

Node 1 2 3 4 5

s1 [3, 2] [3] [5, 4] [5] []

s2 [3, 2] [5] [5, 4] [5] []
s3 [2] [3] [5, 4] [5] []
s4 [2] [5] [] [] []
s5 [3, 2] [5, 3] [5, 4] [5] []
s6 [3] [] [5, 4] [5] []
s7 [3, 2] [3] [4, 5] [5] []

Let us now consider the strategic approach where a subset of strategies,

represented as vectors whose elements consist of an ordered list of outgoing
nodes, is provided in Table 3. For instance, when leaving its origin node 1,
a user adopting strategy s1 will prefer going to node 3 (first node in the ordered
list) over the alternative choice, that is, going to node 2, provided that link
(1, 3) is not saturated yet. A similar situation occurs at node 3, where the out-
going link (3, 5) is preferred over (3, 4). At nodes 2 and 4, however, the strategic
choice shrinks to a singleton, due to the fact that the respective preferred out-
going links are uncapacitated. It follows that a user adopting strategy s1 could
end up traveling on path 1, 2, 4, or 5, depending on the availability of the capac-
itated links (1, 3) and (3, 5). The strategies that, such as s4 , avoid capacitated
links, correspond to ordinary paths in the network.
Whenever the number va of users that wish to access link l exceeds its ca-
pacity ul , the probability p of accessing the link is set to p = ul /vl . This is
equivalent to assuming that users are independently and uniformly distributed
at the tail node of link l, in a dimensionless queue. These access probabilities
(or access proportions) allow us to compute the expected cost of strategies. For
instance, if all 10 users adopt strategy s1 , the probability of accessing link (1, 3)
is equal to 2/10. The users, whether they access link (1, 3) or not, clash again at
node 3, where the access probability of link (3, 5) is 5/10. These numbers, which
are required to compute the access probabilities associated with the paths of
the network, are shown in Table 4.
The expected value of each user’s expected delay is then equal to the sum of
the path costs, weighted by the respective path access probabilities, i.e.,
     
1 4 0
× 100 + × 175 + × 200
10 10 10
   
1 4
+ × 250 + × 325 = 235
10 10
A strategic equilibrium is reached when all users are assigned to strategies
of minimal expected delays. This condition is clearly violated if all commuters
are assigned to strategy s1 , since strategy s4 , which corresponds to the unca-
660 P. Marcotte and M. Patriksson

Table 4.
Path access probabilities for the small example

Path Access probability Cost

1–3–5 2 × 5 = 1 100
10 10 10
1–2–3–5 8 × 5 = 4 175
10 10 10
1–2–5 0 200
1–3–4–5 2 × 5 = 1 250
10 10 10
1–2–3–4–5 8 × 5 = 4 325
10 10 10

Table 5.
Equilibrium strategic path flows

Path Flow from s1 Flow from s2 Cost

1–3–5 5 5 100
6 6
1–2–3–5 20 0 175
6
1–2–5 0 4 200
1–3–4–5 1 1 250
6 6
1–2–3–4–5 4 0 325
6

pacitated path 1–2–5, is available at cost 200 < 235. It is not too difficult to
verify that the unique equilibrium corresponds to the assignment of 5 users to
strategy s1 and of the remaining 5 to strategy s2 ; the resulting strategic flows
and path flows are shown in Table 5. One checks that the expected delay of
each strategy is equal to 185, which is less than the expected delay of unused
strategies, thus fulfilling the equilibrium conditions.
In general, a vector x = {xk }k∈C of strategic flows (one for each commodity
or OD pair k ∈ C ) is a strategic equilibrium if and only if it is demand-feasible
(xk belongs to the set of feasible strategic flows Xk for every commodity k ∈ C )
and satisfies the variational inequality
−c(x) ∈ NX (x) (46)
where X denotes the Cartesian product of the sets Xk . For a given strat-
egy s, the component cs (x) of c(x) represents the expected delay associated
with strategy s and the strategic vector x. Not that the evaluation of c requires
the knowledge of total link flows. Unfortunately, in contrast with the standard
model where this information is readily available from path flows (via the link–
route incidence matrix), the situation is different in the strategic model, as
the dependence of the link-hyperpath matrix on strategies complicates mat-
ters significantly. In particular, the task of determining strategic flows that are
compatible with a given link flow vector is an algorithmic challenge by itself,
and most likely ‘intractable’.
 Ch. 10. Traffic Equilibrium 661

In our introductory example, the link access probabilities were easy to com-
pute. However, the situation becomes complex when the forward star of a node
involves more than two links, and when strategies active at that node have dif-
ferent priority orders. In this situation, should one assume that users strictly
obey a FIFO (First-In-First-Out) rule, or does one allow a user to jump the
queue if its preferred choice is available, as in banks with tellers dedicated to
specific services? The distinction is important, since it impacts the assignment
of users to outgoing links. In the first case, users are assigned to their pre-
ferred node according to their position in the queue. In the second case, users
are assigned simultaneously to their preferred node, until some residual capac-
ity becomes zero. Throughout the process, users that are denied their current
preferred link keep a priority compatible with their arrival instant. Both situa-
tions are illustrated in Figure 8 where each square, for ease of understanding,
represents an atomic user whose preferred outgoing node is either A or B. The
queue is ‘virtual’ (sometimes denoted ‘vertical’) in the sense that it occupies no
physical space.
At node j, the loading of the flow onto the outgoing links is an iterative
 of first
process. In the single queue case, it is initiated by constructing the set K
choices. Next, one computes15
 
ūjk
η = minimum 1 

k∈K dk
where dk is the total demand for outgoing node k and ūjk is the residual ca-
pacity of link (j k). If η = 1, all users access their preferred node k and the
loading terminates trivially. Otherwise, let
 
ūjk
k̄ ∈ arg minimum
k∈K  dk
denote the head node of the link (j k) that gets saturated first. One then loads
a fraction η of each demand to node k, updates demands, removes node k
from the preference lists and repeats the process until demand is exhausted.
Let us consider a nontrivial example involving three outgoings links and two
user groups of size 10 and 20 having respective preference lists [k1  k2  k3 ]
and [k2  k1  k3 ]. See Figure 9. We set the capacities of the three links to 8,
10, and 20. At the first iteration, link (j k2 ) gets saturated first; 5 flow units
from strategy s1 are assigned to k1 , and 10 units from s2 to k2 . Next, one
s s
deletes node k2 from the preference orders Ej 1 and Ej 2 . At the second iter-
ation, 15 units (5 from s1 and 10 from s2 ) select link (j k1 ), whose residual
capacity is 3. A third iteration is required to terminate the process, as the 10
residual units are assigned to link (j k3 ). Note that a by-product of the loading
process is the link access probabilities πjk that allow to compute the strategic

15 Here, the index k refers to a node, not an OD pair.

662 P. Marcotte and M. Patriksson

Fig. 8. Flow dispersion at a node. In the FIFO case, A1 is assigned to A, A2 is assigned to B and all
other users are assigned to C. In the second case, A1 is assigned to A while, in parallel, B1 is assigned
to B; the four remaining users are assigned to C.

Fig. 9. Flow assignment at a node.

cost function c. The process is summarized in Table 6 while, based on the same
data, the outcome of the ‘parallel’ assignment is summarized in Table 7.
Although there are considerable differences in the assignments resulting
from the two rules, these agree in important situations, namely when at most
two choices are available at each decision node. Other theoretical points worth
mentioning are:
(i) The cost function c is continuous which, together with the compact-
ness of the feasible set, implies that the set of equilibrium solutions is
nonempty.
(ii) The loading procedure only makes sense if the network is acyclic, and
nodes are processed in the relevant topological order. If this condition
is not fulfilled, it is yet possible to perform the loading operation by
iterating, à la Gauss–Seidel, with respect to the OD pairs. At a given
 Ch. 10. Traffic Equilibrium 663

Table 6.
Outcome of single file loading at node j

Link: (j k1 ) (j k2 ) (j k3 )

Iteration 1 Residual capacity: 8 10 20

Flow: 5(s1 ) 10(s2 ) 0
Iteration 2 Residual capacity: 3 0 20
Flow: 5(s1 ) + 1(s1 ) + 2(s2 ) 10(s2 ) 0
Iteration 3 Residual capacity: 0 0 20
Flow: 6(s1 ) + 2(s2 ) 10(s2 ) 4(s1 ) + 8(s2 )
s s s s
6 , π 1 = 4 , π 2 = 2 , π 2 = 10 , π 2 = 8 s
Probabilities: πjk1 = 10 jk3 10 jk1 20 jk2 20 jk3 20
1

Table 7.
Outcome of the parallel loading at node j

Link: (j k1 ) (j k2 ) (j k3 )

Iteration 1 Residual capacity: 8 10 20

Flow: 8(s1 ) 8(s2 ) 0
Iteration 2 Residual capacity: 0 2 20
Flow: 8(s1 ) 8(s2 ) + 27 (s1 ) + 12
7 (s2 ) 0
Iteration 3 Residual capacity: 0 0 20
Flow: 8(s1 ) 68 (s ) + 2 (s ) 12 (s ) + 72 (s )
7 2 7 1 7 1 7 2
s s s s
8 , π 1 = 2 , π 1 = 12 , π 2 = 68 , π 2 = 72 s
Probabilities: πjk1 = 10 jk2 7·10 jk3 7·10 jk2 7·20 jk3 7·20
1

iteration, the access probabilities for all OD pairs, with the exception
of the current one, are frozen at their previous values.
(iii) The cost mapping c may fail to be monotone under both queue disci-
plines.
(iv) In the single queue case, the issue of convexity of the equilibrium set is
open.

From the practical side, static strategic models are of limited scope. How-
ever, they can be adapted to time-dependent networks and help in developing
dynamic models that take into account the rational reaction of users to online
information. In a dynamic context, models must take into account the natural
temporal priorities; this leads to a more complex loading procedure involving
a cost mapping c that may fail to be continuous. On the positive side, the map-
ping’s upper semi-continuity ensures that the set of equilibria is nonempty, and
the underlying time–space network is trivially acyclic.
664 P. Marcotte and M. Patriksson

3.4 Nonadditive route costs

In our previousdiscussions, we have assumed that route costs were additive,

that is, cr (h) = l∈L λlr tl (v) holds for all r ∈ R and consistent flows (h v).
This is not true, however, in several situations involving route-specific tolls or
route-specific vehicle emissions. Cost additivity also fails in loss networks aris-
ing, among others, in telecommunication systems.
Let us reconsider the model with two attributes, time and money, intro-
duced in Section 3.1. Empirical studies show that, in contrast with our previous
assumption, the VOT parameter can be nonlinear. Consider, for example, that
 
cr (h) = fr + βr λlr tl (v)  (47)
l∈L
where fr is the monetary outlay and βr is now a function of travel time, for
route r ∈ R. In this model, time is converted into money through the func-
tion β. It is also possible to work with time as the numéraire, thus considering
instead a cost model of the form

cr (h) = λlr tl (v) + αr (fr ) (48)
l∈L
where the scalar αr converts money into time along route r. Interestingly, even
if we set αr to β−1
r , the two models are not equivalent. From the analytic point
of the view, the latter model has the advantage that there exists an equivalent
optimization formulation in the case of separable link costs: provided that fr is
flow-independent, and that the value of νr does not vary within OD pair r, the
partial derivative of the objective function
  vl  
tl (s) + hr αk (fr )
0
l∈L k∈C r∈Rk

with respect to hr is exactly cr (h) given in (48), and Wardrop’s conditions re-
duce to the optimality conditions of this optimization problem. We also see that
the objective function is convex under the same conditions as for the additive
model. Of course, these results hold because monetary costs are flow indepen-
dent; observe the similarity between this result and those of Section 3.3.
It is clear that an algorithm in link-flow space based on first-order approxi-
mation cannot be immediately extended to this model; indeed the ‘best’ route
cannot be determined through a standard shortest route calculation; see Sec-
tion 5.4.

3.5 Bibliographical notes

References to the basic multi-mode model of Section 3.1 can be found

in Florian (1977, 1979), Toint and Wynter (1996), Florian et al. (2002),
Marcotte and Wynter (2004). The introduction of the bi-criterion model in the
 Ch. 10. Traffic Equilibrium 665

transportation literature is due to Dial (1979), who also proposed a solution al-
gorithm (Dial, 1996a, 1996b). Early on, Dafermos (1981) gave a necessary con-
dition in order that the variational formulation of the problem be monotone.
Similar results appear in Nagurney and Dong (2002). Marcotte (1998) proved
that these conditions were necessary for both the monotonicity and integra-
bility properties to hold. Equivalent finite-dimensional formulations of the
problem have been proposed in Leurent (1993) and Marcotte (1998).
Section 3.2 introduces a modeling paradigm from the viewpoint of a con-
trol policy. This way of interpreting the model leads to interesting interpre-
tations of the Lagrange multipliers, and the use of the ‘side constraint’ that
v = v∗ must hold for some a priori set feasible flow v∗ is the direct route
to the use of link tolls for achieving a system optimum or any other desired
flow, as presented, for example, in Section 6.2. Side constraints can emerge
from other sources than controls, however, and Larsson and Patriksson (1994b,
1997, 1999) describe several such circumstances, and show how the Lagrange
multiplier terms can be associated with link queues. That the Lagrange multi-
pliers are not unique was established in Larsson and Patriksson (1998), whence
interpretations must be used with great care, for example, in the context of
equilibrium queues. Previously, equilibrium characterizations of link capaci-
tated models have been provided in Jorgensen (1963) and Hearn (1980), and
the interpretation of the multipliers βl as stationary link queues can be found
in Miller et al. (1975) and Payne and Thompson (1975). See also Ferrari (1995).
The numerical example stems from Larsson and Patriksson (1998), as does the
description of the toll polyhedron development. Related, and in many ways
parallel, work exists for the special case where one wishes to achieve a system-
optimal solution. In this field, the master’s thesis by Bergendorff (1995) was
followed by a series of work by Don Hearn et al. (e.g., Bergendorff et al., 1997;
Hearn and Ramana, 1998).
The strategic approach outlined in Section 3.3 was motivated by the transit
equilibrium model described in Nguyen et al. (2001), and is based on Marcotte
et al. (2004). Additional information, including technical details, is available
in Marcotte and Nguyen (1998), Hamdouch et al. (2004a, 2004b). Mathemat-
ical programming formulations of transit equilibrium problems, based on the
notion of strategy (or hyperpath), were independently proposed in Nguyen and
Pallottino (1988) and Spiess and Florian (1989).
The nonadditive models described in Section 3.4 are due to Gabriel and
Bernstein (1997) (see also Chen and Bernstein, 2003), for the case (47) of
converting money into time, and Larsson et al. (2002), for the case (48) of con-
verting time into money. That the two models are not equivalent was observed
in Larsson et al. (2002), and later established in more generality theoretically
in Bernstein and Wynter (2000). The logit-based stochastic user equilibrium
model is due to Fisk (1980). For further reading on stochastic traffic equi-
librium models, see Sheffi (1985), Akamatsu (1996), Watling (1999), Cascetta
(2001). The effects of modeling emission effects are discussed in Larsson et al.
(2002). Recent models of loss networks in telecommunication can be found in
666 P. Marcotte and M. Patriksson

Altman et al. (2002). As is stated in the section, the model considered con-
verts money into time, and models based on this transformation have better
properties.

4 Solution algorithms: The basic problem

4.1 General guidelines

Finding a flow pattern satisfying Wardrop’s user equilibrium conditions (2)

amounts to solving a large-scale, network structured variational inequality
problem. The computational challenge is threefold: How to exploit the un-
derlying network structure of the problem efficiently? How to obtain fast local
convergence? How to ensure global convergence?
Since the constraint set F, in the link–node representation, is the feasible set
of an uncapacitated multicommodity flow problem, minimizing a linear func-
tion over F is a task that can be efficiently performed by finding the shortest
route trees rooted at the origin nodes. At this level, the distinction between
link, route or commodity flows becomes irrelevant, as the shortest route pro-
cedure provides information at all three levels of (dis)aggregation.
As far as convergence speed is concerned, algorithms combining the above-
mentioned linear subproblems with simple line searches only achieve sublinear
convergence. To achieve linear or super-linear convergence in such a frame-
work, one may have to solve a quadratic problem over the multicommodity
network; unfortunately, such a task cannot be performed on large scale in-
stances. Suppose, however, that we consider the link–route representation,
and consider the set of routes carrying positive flow to be fixed. Then, solv-
ing the corresponding restricted problem by fast algorithms is indeed possible.
Since, in practical applications, a relatively small set of routes is required to
describe an equilibrium flow, this fact suggests that route (or, column) gener-
ation strategies are prime suspects for solving traffic equilibrium problems. It
must be understood that the maximum size of the working set of routes will
influence the behavior of the algorithm: the larger the maximum cardinality
of this set, the better the convergence. However, a trade-off must be achieved
between a good convergence rate on the one hand and, on the other hand, the
size and the computational difficulty associated with the restricted problems.
As a limiting case, the slow converging Frank–Wolfe method is obtained by
limiting the size of the working set to two feasible points, one corresponding
to the current iterate, and the second to the solution to the linear subproblem.
If t is not a gradient mapping, then one must be careful that an aggressive
column dropping scheme will not result in a sequence of iterates that cycles
between identical working sets. The third issueis more or less irrelevant if t
is a gradient mapping, since the potential φ = t then may be used to mon-
itor descent toward a stationary (Karush–Kuhn–Tucker) point, that is, a user
equilibrium. If t is not a gradient, then some monotonicity property (such as
 Ch. 10. Traffic Equilibrium 667

pseudomonotonicity) is required in order to be able to implement a provably

convergent method. In the following, we elaborate more on these issues for the
basic traffic equilibrium model.

4.2 Model structures and the verification of equilibria

Some favorable problem structures are observed when looking at the mod-
els (15) and (16):
• OD pair separability. The feasibility of the demand and traffic volume in
one OD pair does not affect that of another. The problems indeed are
multicommodity flow problems, where there are no side constraints
acting on more than one commodity, such as link flow capacity con-
straints. This property holds for more general cost and demand models
as well.
• Cost separability. The objective function, as well as the link cost and
demand functions, is separable in the link volume and demand vari-
ables. This property does not extend to more general cost and demand
models.
• Primal–dual relations. In both the link–route and link–node represen-
tations, the vector π is a Lagrange multiplier vector, which further
measures the least cost. This property holds for more general cost and
demand models as well.
The link–route representation has a simpler constraint structure than the
link–node one; the constraints in the former describe a simplex, while those of
the latter describe flow conservation. The former, on the other hand, has an
exponential number of variables, which must be enumerated iteratively, while
the latter has a polynomial number of variables, and a richer constraint struc-
ture.
The very definition of a variational inequality provides a convenient crite-
rion for verifying whether a vector is in equilibrium. Indeed, the inequality
f(x) (y − x)  0 x ∈ X
is satisfied if and only if x is optimal for the linear program
minimize f(x) y
y∈X
If one keeps demand fixed, this amounts to checking whether the gap function
(see also Appendix A)
gap(x) = maximum f(x) (x − y) (49)
y∈X

is zero.16 This LP problem defines a set of shortest route problems; an equilib-

rium is reached if and only if the total cost of transportation along the shortest

16 The index k in the second equality relates to feasibility sets and subvectors for individual OD pairs.
668 P. Marcotte and M. Patriksson

routes in each OD pair equals the total travel cost of the current OD flow.
The value of gap(h) is always nonnegative, and a flow vector h is an equilib-
rium if and only if gap(h) = 0. If gap(h) is positive, one has identified a set of
attractive routes that are not currently used (at least not enough).
In the elastic demand case, it seems natural to consider the iteration
d → (g ◦ π ◦ t ◦ v)(d) (50)
that takes the current demand and calculates a corresponding inelastic demand
solution, from which the least route cost is calculated. The value of the de-
mand function then yields the new demand estimate. A sufficient condition
that this fixed-point iteration converge is that it be contractive, that is, for an
equilibrium demand d∗ and any feasible demand vector d, there exists a scalar
θ ∈ (0 1) such that g(π(t(v(d)))) − d∗   θd − d∗  holds. Given the com-
plicated form of this mapping, whether the contraction condition is satisfied is
difficult to analyze from the original data, and updating rules for flows and de-
mands are instead based on approximations of one or more of these mappings,
examples of which will be provided below.

4.3 Decomposition–coordination

The separability of the OD pairs (or, user classes, etc.) in the constraints
suggests the use of a decomposition strategy, wherein flow of a given type is
updated upon while keeping the others fixed; the remaining problem is then
a single-commodity network flow problem. Among the advantages of such a
scheme we mention first that the link cost function has stronger monotonicity
properties when viewed on this smaller subspace than on the entire space of
flows; for example, if vl → tl (vl ) is strictly monotonically increasing, then it
is (usually) nonstrictly monotone in the vector of OD link flows wk , k ∈ C ,
but it is strictly monotonically increasing when keeping flows in all but one
OD pair fixed. Second, several single-commodity network flow algorithms are
available, among which are (potentially) superlinearly convergent ones. Third,
algorithms of this type can be viewed as block versions of Gauss–Seidel (or
Jacobi) algorithms, which have a sound convergence theory. They can how-
ever be slow when the number of blocks is large, since too much of the cost
interaction among the commodities is lost. One can then take advantage of the
problem structure by creating a small number of large blocks of variables such
that, for example, all flow from one given origin is updated at any one time.
Contrary to the use of decomposition over OD pairs or origins, a decomposi-
tion over links is less convenient, as costs are not always defined by separable
functions.
A natural means to decompose the problem over OD pairs is to temporarily
fix link costs, thereby removing the cost dependence between OD pairs. In the
fixed demand problem, what is left is a number of shortest route problems,
on which demand is assigned. In the elastic demand case, the demand to be
assigned onto the shortest routes is given by the demand function evaluated
 Ch. 10. Traffic Equilibrium 669

at the shortest route costs. In this way, we obtain the algorithms of Frank and
Wolfe (1956) and Evans (1976), respectively.

4.3.1 Decomposition by Lagrangian relaxation

In order to take advantage of the link cost separability in the models
(15) and (16), we must eliminate their dependence in the constraints by in-
troducing some form of relaxation. One possibility
 is to introduce explicit
multipliers, ν ∈ |L| , for the constraints ‘v = k∈C wk ’. With the (redundant)
|L|
constraints v ∈ + appended to the Lagrangian relaxed problem, we obtain
a problem, equivalent to VIP(t Fd ), which amounts to finding (v w d π ν)
such that
 
0|L|  v ⊥ t(v) − ν  0|L|  (51a)
Ewk = ik dk  k ∈ C  (51b)
 
0|L|  wk ⊥ ν − E π k  0|L|  k ∈ C (51c)
g(d) = π (51d)

v− wk = 0|L| (51e)
k∈C
holds. Fixing ν̄  t(0), and noting that (51e) is the dual optimality condi-
tion, the problem separates into one over v and wk , k ∈ C . In the separable
case, (51a) vfurther decomposes into a set of single-variable problems with the
objective 0 l (tl (s) − ν̄l ) ds, to be minimized over + , and with a solution given
by the inverse of tl evaluated at ν̄l , whenever the latter is nonnegative (other-
wise it is set to zero). In the general case, this is a nonlinear complementarity
problem (NCP) in |L| . The second problem, (51b)–(51d), amounts to find-
ing, for each k ∈ C , the shortest route based on the vector ν̄ of link costs.
The shortest route costs are used to calculate the demand through the de-
mand function Demand is then distributed onto the shortest routes to form
the vectors wk . A step is then taken along the direction ν, in order to produce
a result that better fulfills the last constraint. Although being a primal–dual al-
gorithm, the aggregate of the wk is a feasible link flow; this is a consequence of
the fact that the dualization is made with respect to the definitional constraint
(51e).
Another way to decompose the problem is to relax the linear equality con-
straints defining the network, that is, (51b). Introducing explicit multipliers
π k ∈ |N | for each commodity k ∈ C , we obtain a problem, equivalent to
VIP(t Fd ), which amounts to finding (v w d π) such that
 
0|L|  wk ⊥ t(v) − E π k  0|L|  k ∈ C  (52a)
Ewk = ik dk  k ∈ C (52b)
g(d) = π (52c)

v− wk = 0|L| (52d)
k∈C
670 P. Marcotte and M. Patriksson

holds. Fixing π̄ k (and letting π̄pq := ik π̄ k ), and noting that (52b) is the
dual optimality condition, the remaining problem is solved as follows. The
demand d is given, through (52c), by the demand function at π̄. The com-
plementarity system (52a) is solved, in the separable case, by first defining,
for each link l = (i j) ∈ L, p̄ij (π) := maximumk∈C {πjk − πik } to be the
largest node potential difference among the commodities. Let Cij (π) ⊆ C
denote the set of commodities for which πjk − πik = p̄ij (π) holds. Then,
vij (π) := maximum{0 tij−1 (p̄ij (π))} is the total link flow in link (i j). The
commodity link flows wijk , k ∈ C , can be taken as any distribution of the total
flow vij (π) onto the commodities in Cij (π). It remains to define an updating
rule for the vectors π̄ k such that we approach also a solution to (52b). If this
Lagrangian relaxation is utilized within a cyclic decomposition algorithm over
OD pairs, where the subproblems are single-commodity network flow prob-
lems, then more efficient dual algorithms are available, since the duals then
are differentiable.

4.3.2 Decomposition by linearization

A decomposition (relaxation) can also be achieved by manipulating the
cost structure. Suppose, for example, that the link costs are fixed at their cur-
rent values, and consider the resulting relaxation. Taking VIP([t g] F d ) as the
τ τ τ
model for discussion, and the iterate (v  d  (π k )k∈C ) at iteration τ, fixing the
link costs to t(vτ ) yields the following procedure for solving this relaxation of
VIP([t g] Fd ): first, for each k = (p q) ∈ C , calculate a least-cost route,
cf. (6a), whence the vectors π̄ τk , k ∈ C , (π̄ τ )(pq)∈C = (ik π̄ τk )k∈C are given;
second, calculate the demand through d̄τ := g(π̄ τ ); third, assign this demand
to each
 OD pair’s shortest route, giving rise to the link flows w̄τk , k ∈ C , and
v̄ = k∈C w̄k . If link costs and demands are gradient mappings, we can per-
τ τ

form a line search in the direction of (v̄τ  d̄τ  (π̄ k )τk∈C ) − (vτ  dτ  (π k )τk∈C ),
whence we have defined a partial linearization algorithm known as Evans’
algorithm; whenever demand is inelastic, this is precisely the Frank–Wolfe al-
gorithm.
The route information generated during the process of calculating shortest
routes is badly utilized in line search algorithms, considering the effort in gen-
erating it. If, instead, we were to store the routes that carry flow or that are new
and promising, and solve the equilibrium problem over those, we would in fact
be addressing a restriction to the link–route formulation of the original model
(a restricted master problem). As we have remarked earlier, this model form
has an advantage in that the constraint structure can be effectively utilized;
if we keep down the size of the route set, we also have a chance of avoiding
the drawback of this formulation. Consider then the model VIP(t F),  where

we have available a restriction R ⊂ R of the route set. We recognize that
the simplex structure can be put to good use, for instance by devising a de-
composition over OD pairs or origins. This is achieved in a cyclic Newton-like
algorithm, wherein we keep the flow from all but one group of users fixed, and
 Ch. 10. Traffic Equilibrium 671

then utilize a diagonalized second-order approximation of the remaining ob-

jective function. At a given iteration where the flow is hτ , we would, for one
OD pair k, say, solve the following problem:
   bτr 2

τ
minimize cr h hr + h  (53a)
hk 2γτ r

r∈Rk

subject to hr d̄k  (53b)
k
r∈R

hk  0|Rk |  (53c)
where 0 < bτr ≈ ∂cr (hτ )/∂hr , and the value of γτ > 0 is chosen such that
the improvement made is the best possible; the resulting solution is taken to
be hτ+1
k . The above problem has a low complexity: solving for the optimal La-
grange multiplier for the equation can be done in O(|R k |) time, that is, linear
in the number of routes. Achieving low complexity is essential in the subprob-
lems of an iterative algorithm, as the size of traffic problems can be very large.
A final remark on algorithms for the basic models is that it seems difficult to
devise truly superlinearly convergent algorithms that utilize problem structure
efficiently. One main reason is the presence of multiple commodities; for sepa-
rable cost nonlinear single-commodity flow problems, superlinearly convergent
algorithms are available, which also utilize the network property in an efficient
way. However, the Gauss–Seidel algorithm – which is needed to coordinate the
search in the multicommodity setting – is linearly convergent, at best. There is
therefore still plenty of room for algorithmic improvements, in particular for
new and clever ways of decomposing and coordinating the problem.

4.4 Merit functions and convergence for asymmetric models

When solving models involving cost and demand functions that are gradient
mappings,17 we have access to natural merit functions (cf. (15) or (16)), with
which we can measure the progress of an algorithm and perform line searches;
within this framework, the whole area of nonlinear optimization is open to ex-
ploration. When turning to nonintegrable models, these merit functions are no
longer well defined, and progress toward an equilibrium must be measured by
some artificial construct. One such function is the primal gap function given in
(49) or (106), which can be calculated entirely in link flow space. As with most
other merit functions used for variational inequality problems, its evaluation
requires the solution of an approximation to the original problem. The com-
putation of equilibria for nonseparable models are in fact much more complex
than for separable ones:

17 This is referred to as the ‘integrable case’.

672 P. Marcotte and M. Patriksson

• Decomposition. In the integrable case, some of the most natural decom-

positions lead to low-complexity subproblems. In the nonintegrable
case, however, they either become unavailable altogether, or they be-
come more complex. For example, a decomposition over links will no
longer be natural or even possible. Further, if the form of the cost
and demand is complex, it also means that computing their values and
derivatives will be more complex, and this will also affect the computa-
tional complexity of an algorithm.
• Search direction → merit function. In the integrable case, virtually every
algorithm proposed is a feasible direction-based descent method which
utilizes the natural objective in (15) or (16). Establishing convergence
of such algorithms is relatively straightforward, as it is easy to satisfy
the descent condition that the gradient of the merit function makes
an obtuse angle with the search direction, and it does not rely on
the monotonicity properties of the cost and demand functions. In the
nonseparable case, however, the angle condition becomes much more
difficult to establish for a suitable merit function. The merit function is
intimately associated with the search direction, and in some cases it is
even defined by the solution to the search direction-finding problem.
That also means that, in contrast to the separable case, the merit func-
tion and its gradient depends on the search direction chosen. In order
to derive convergence results, more stringent requirements on both the
search direction and the step length must be enforced.
As an example of the complications that arise in the general, we mention
that an algorithm that would mimic the Frank–Wolfe strategy by treating the
cost function as if it were a gradient mapping, need not converge for asymmet-
ric cost models.18 However, it is straightforward to show that, whenever the
‘Frank–Wolfe’ direction is unique and the cost mapping is monotone, it con-
stitutes a feasible descent direction for the primal gap function. If the set of
Frank–Wolfe directions is not a singleton, it is yet possible to prove that there
exists a member of that set that is gap-decreasing. Based upon this result, an
implementable and convergent algorithm that operates in link flow space can
be designed.
There exist several variational inequality algorithms that so far have not
caught the attention of the traffic research community, and that converge un-
der rather mild assumptions on the travel cost and demand mappings. For one,
the projection algorithm, which takes an iterate vτ and maps it onto the next
one through the formula
  
vτ+1 := ProjF vτ − γτ t vτ  γτ > 0

18 In such an algorithm, line searches are performed based on first-order information, not the objective
function.
 Ch. 10. Traffic Equilibrium 673

where ProjF denotes the Euclidean projection onto F,  can be supplied with
special line searches that make it convergent for monotone cost mappings. (In
its original statement it requires strong monotonicity or co-coercivity, and re-
lies on the estimate of the cost mapping’s strong monotonicity modulus and
Lipschitz constant.) The class of extra-gradient algorithms involves two pro-
jections of the above type, and is convergent under even milder monotonicity
requirements: pseudomonotonicity.
The most natural environment for these algorithms is the singly constrained
restricted master problem in a route (column) generation algorithm, that is,
a restriction of the feasible sets defined in (15) or (16) to a subset R  of the
routes. The reason why we favor this environment is that projections are easy
to perform onto such sets, cf. (53).
An especially interesting avenue for constructing algorithms for nonstrictly
monotone problems is to combine an algorithm of one’s choice (and which
perhaps requires strong monotonicity, such as the Jacobi algorithm) and the
proximal point algorithm. It is defined thus: for the given cost function t in the
variational inequality VIP(t F), one would add a multiple (γτ > 0) times an
affine cost of the form v → v − vτ , if we work in link space. The resulting map-
ping v → t(v) + γτ (v − vτ ) is strongly monotone with a modulus of at least γτ ,
provided that the original cost function t is monotone. Therefore, if we have
an algorithm which requires strong monotonicity, we apply it instead to this
perturbed problem. At the (approximate) solution to the perturbed problem,
we again perturb the original function at the new iterate, vτ+1 , and proceed,
perhaps with a different value of γ.
Diagonalization is a favorite among heuristics for nonseparable problems.
It amounts to, at an iteration, temporarily removing the dependency of a link’s
cost on the flow in other links. This is related to the Jacobi algorithm, and may
suffer from quite poor convergence characteristics: if the correlation in reality
is quite large, one can expect the algorithm to diverge, or at least converge
slowly; on the other hand, if the cost dependence between links is mild, or
local to a few links only, then convergence can be expected to be fast.
We close this section by mentioning that minimizing the dual gap function

maximum f(y) (x − y)
x∈X

can be achieved by solving the semiinfinite linear program to

minimize z (54)
x∈Xz∈

subject to z  f(y) (x − y) y ∈ X

This structure lends itself naturally to decomposition algorithms similar to the
linear programming method of Dantzig and Wolfe, and to more recent interior
point cutting plane methods.
674 P. Marcotte and M. Patriksson

4.5 Implementations and computational experience

Commercial implementations of equilibrium solvers are frequently based on

the Frank–Wolfe method, or variants thereof. Although it has been known for
quite some time that the Frank–Wolfe algorithm has the worst (sublinear) con-
vergence rate among equilibrium algorithms, its high initial efficiency, small
memory requirements, efficient handling of the underlying network structure
and the natural interpretation behind its workings – it is reminiscent of heuris-
tics that were devised already in the mid-1960s – have made it popular. Its
counterpart for elastic demand models is Evans’ algorithm, which has similar
convergence characteristics: demand quickly converges, and once demand sta-
bilizes, the algorithm’s behavior is similar to Frank–Wolfe’s.
Among the favored extensions of the Frank–Wolfe algorithm is the class of
simplicial decomposition/column generation algorithms (see the discussion on
the DSD algorithm above) which also are based on shortest route subprob-
lems. While the simplicity of the subproblem is preserved, the results of the
computations are better utilized, as the shortest routes are kept in memory
and demand is optimally assigned onto them. This restricted master problem,
which has the same form as that of the original problem except that the set R
is replaced by the known subset R , is then solved using a specialized, perhaps
second-order, method. In later implementations of the DSD algorithm a di-
agonalized Newton algorithm is applied, which also efficiently detects routes
with zero flow; these routes may be removed from the set R  in order to save
space. Other implementations use different versions of reduced gradient and
gradient projection methods. As projection methods are available also for non-
separable models, it is anticipated that the best versions can be transferred to
solve also such more general models.
For a numerical illustration of two of the above discussions, let us consider
the ‘famous’ Sioux Falls network, composed of 24 nodes, 76 links, and 528 OD
pairs. For a MATLAB19 implementation of the DSD and FW algorithms, Fig-
ure 10 illustrates how fast (in terms of CPU time) convergence occurs toward
an equilibrium solution. The implementation of the DSD algorithm comes with
several methods for solving the restricted master problems: two versions of the
Goldstein–Levitin–Polyak gradient projection method (marked ‘Gradient P.’
in the figure), and a diagonalized Newton method. For the small-scale network
representing Sioux Falls, the gradient projection method is faster, but for larger
networks, Newton’s method becomes relatively more favorable. Clearly, the
Frank–Wolfe method is very slow for this problem, and its relative efficiency is
even further reduced for larger problems.
A recent contribution to the field is the quasi-Newton algorithm of Bar–
Gera. It utilizes the fact that the equilibrium flow from a given origin is acyclic

19 The absolute CPU times should not be taken at face value; a corresponding C implementation can
be two to three orders of magnitude faster.
 Ch. 10. Traffic Equilibrium 675

Fig. 10. The performance of DSD vs. FW on the Sioux Falls network.

(cf. Theorem 4(b)), and tries in a combinatorial fashion to determine the cor-
rect acyclic networks. For each origin and acyclic network, a quasi-Newton
algorithm is implemented for finding the best single-commodity flow, based
not only on link flows, but also on the flow through the nodes. The numerical
experience so far is impressive, and it is believed that it will, if not replace, then
at least become an alternative to the algorithms currently used in practice. The
main drawback of this algorithm is that it has not yet been shown how to make
it work for nonseparable models; see Section 5.4 for alternatives.
Somewhat surprisingly, the field of computational testing, in particular com-
parative ones, remains sparse. Comparisons are still quite often performed
against the Frank–Wolfe algorithm; such results are not very enlightening,
given the bad convergence behavior of the latter (see Figure 10).
Several interesting alternatives to the classical algorithms have never been
tested, and the protocols of those that are tested on a regular basis would not
pass the scrutiny of a numerical analysis specialist. The latter is not so surpris-
ing, considering that most of the algorithmic development relates to specific
applications. The former is not so surprising either, given that most researchers
in the field come from an area different from mathematical programming.
The low-complexity subproblems, like the shortest route problem and the
quadratic knapsack problem (53), are also mathematical objects of interest.
676 P. Marcotte and M. Patriksson

Although there exist efficient algorithms for these problems, the fact that they
have to be solved repeatedly leaves room for algorithms that are less efficient
according to a worst-case criterion, but lend themselves better to reoptimiza-
tion. As is the case for sorting, the worst-case behavior does not tell the whole
story.

4.6 Bibliographical notes

The gap function defined by (49) is a measure of the distance to an equilib-

rium. Its negative is the directional derivative of the traffic equilibrium objec-
tive function φ (see (15)) along the direction toward a shortest route solution
– which explains its role in the Frank–Wolfe algorithm. The gap function was
originally considered by Zuhovickiı̌ et al. (1969), and introduced into the field
of transportation science by Murchland (1970); see also Auslender (1976).
Fixed-point type results concerning the algorithmic map (50) are based on
contraction arguments (which are difficult to verify in practice), and can be
found in Dunn (1973), Bruck (1975), and Baillon (1975). One reason we do
not cover the fixed-point literature is that such formulations, through the use of
problem mappings similar to (50), hide the salient structures of the underlying
model. We prefer, departing somewhat from orthodoxy, to view such fixed-
point problems not as traffic equilibrium models in their own right, but rather
as algorithmic constructs.
The uncapacitated and link capacitated single-commodity flow problems
discussed in Section 4.3, and algorithms for their solution, are reviewed in
Kennington and Helgason (1980), Ahuja et al. (1993), Patriksson (1994b),
Bertsekas (1998), Patriksson (2006).
The procedure following the statement (51) is taken from Larsson et al.
(1997, 1999), and the procedure following the statement (52) stems from Pa-
triksson (1994b, Section 4.3.2); both are covered in Patriksson (2006). The
Evans and Frank–Wolfe methods were first implemented for traffic equilib-
rium models by Evans (1976), and Nguyen (1974) and LeBlanc et al. (1975),
respectively.
The first successful attempt at building an efficient route-based column gen-
eration algorithm was made by Larsson and Patriksson (1992), who devised
the disaggregate simplicial decomposition (DSD) framework. Although simi-
lar approaches had been proposed earlier, for example, in Leventhal et al.
(1973) and Bertsekas and Gafni (1982), the restricted master problems were
then deemed too difficult to solve, and route-based methods were therefore
for many years considered inefficient. A few years later, Hearn et al. (1987)
designed a similar algorithm where routes however are aggregated into all-or-
nothing solutions, thereby restricting the number of variables in the restricted
problems and at the same time reducing the potential speed of convergence
for the respective route flows. The quadratic subproblem shown in (53) is the
one preferred for the restricted problems, and is also used in the DSD algo-
rithm. The convergence of this scaled gradient projection method has been
 Ch. 10. Traffic Equilibrium 677

known for many years (e.g., Bertsekas, 1976). What makes it attractive is
that the quadratic term can incorporate derivative information on route costs
in order to enhance convergence, and also that the solution to (53) can be
computed in linear time (Brucker, 1984). Although route-based formulations
yield variational formulations that are not strongly monotone, convergence can
yet be proved under the assumption that the link cost functions are strongly
monotone (see Bertsekas and Gafni, 1982 and Zhu and Marcotte, 1996). (In
the separable case, the problem’s convexity is always enough to ensure conver-
gence provided that a line search is performed in each iteration.)
For separable cost single-commodity problems, superlinearly convergent
Newton-like methods that exploit network structure have been proposed by
Best and Griffin (1975), Klincewicz (1983), Escudero (1986), Gafni and Bert-
sekas (1984); they are also overviewed in Patriksson (2006). In the multicom-
modity, nonseparable, case, a modified Newton method has been implemented
by Marcotte and Guélat (1988) within the framework of simplicial decompo-
sition. The linear convergence of the Gauss–Seidel algorithm is established in
Bertsekas and Tsitsiklis (1989); the convergence rate depends of course on the
extent to which the different variable components (that is, flow variables) in-
teract in the objective function. Choosing the right decomposition – by OD
pair, by origin, or even by grouping together some origins or OD pairs – may
therefore be crucial for the practical convergence rate of the algorithm.
In the nonseparable case, the Frank–Wolfe strategy may lead to cycling
(see, e.g., Hammond, 1984). It is interesting to note that column gener-
ation/simplicial decomposition algorithms are guaranteed to converge, as
long as all routes are kept in memory, or if column dropping is done with
care. Contributions to the theory and practice of such methods are found in
Lawphongpanich and Hearn (1984) and Patriksson (1998). The existence of
gap-decreasing Frank–Wolfe directions, and the design of a convergent frame-
work built around such directions, is due to Marcotte (1986a).
The remainder of Section 4.4 discusses several ways in which convergence
can be achieved by utilizing recent contributions to the field of variational in-
equality algorithms. General references for the development of line search
methods for variational inequality problems until the late 1990s can be found
in the monographs of Patriksson (1998) and Facchinei and Pang (2003b). Spe-
cial line searches incorporated into the projection method are found in Bruck
(1977) and Patriksson (1998). The class of extra-gradient algorithms has devel-
oped quite far since the original paper by Korpelevich (1977); recent contribu-
tions can be found in Iusem (1998), Konnov (2001), Wang et al. (2001), Solodov
(2003). An implementation of such an algorithm in the context of network
equilibrium can be found in Marcotte (1991). The proximal point algorithm
was developed largely by Rockafellar (1976), following previous developments
in the 1960s for more general problems. In that paper, an inexact solution of
the regularized problems is already shown to be valid for monotone problems;
more recent contributions, where the proximal point method is combined with
projection methods, and where previous convergence rate analyses are im-
678 P. Marcotte and M. Patriksson

proved, are found in Solodov and Svaiter (1999a), Solodov and Svaiter (1999b),
Solodov (2003). ‘Diagonalization’ is the favorite term in the transportation
research literature (e.g., Sheffi, 1985) for the method known as ‘Jacobi’ to re-
searchers in numerical analysis and mathematical programming. Cutting-plane
methods for solving the Minty formulation or, equivalently, the semi-infinite
linear program (54), have been proposed by Zuhovickiı̌ et al. (1969) in a
game-theoretical framework. It has been modified and adapted to the traf-
fic assignment problem, using a column generation framework, by Nguyen and
Dupuis (1984). The analytic center cutting plane method described in Goffin
et al. (1997) may provide a viable alternative.
The contributions by Bar-Gera to the practice of solving separable, fixed
demand problems are found in Bar-Gera (2002a, 2002b). Algorithms based on
the DSD framework that have emerged since 1992 are found in, for example,
Jayakrishnan et al. (1994) and Chen et al. (2002). The web site:
http://www.bgu.ac.il/~bargera/tntp/
provides test problems and computational results. The reader may also consult
http://www-rocq.inria.fr/metalau/ciudadsim/.
Finally, the reoptimization issue has been addressed in Nguyen et al. (2002)
and Pallottino and Scutellà (2003).

5 Solution algorithms: Variations

5.1 Multi-mode and multi-attribute models

In a multimodal context, decomposition is the strategy of choice for comput-

ing an equilibrium. If all modal flows are fixed, save one, the model reduces to
a standard traffic equilibrium model which frequently involves a gradient map-
ping. At subiteration j of a major cycle τ, one solves the single-mode problem
of finding the solution wjτ to the variational inequality
   
tm v1τ  vj−1τ      wjτ  vj+1τ−1      v|M|τ−1 ∈ NFj wjτ  (55)
in the Gauss–Seidel variant and to the variational inequality
   
tm v1τ−1  vj−1τ−1      wjτ  vj+1τ−1      v|M|τ−1 ∈ NFj wjτ  (56)
in the Jacobi (diagonalization) implementation. After the completion of an
entire cycle, one updates the multi-attribute flow vector to

vτ+1 = λvτ+1 + (1 − λ)wτ  (57)

where λ ∈ (0 1) is a relaxation parameter. The convergence analysis of the
above two schemes is available in classical numerical analysis textbooks and is
reproduced in the transportation literature. It relies on the strong monotonic-
ity and Lipschitz moduli of the cost mapping, two constants for which reliable
 Ch. 10. Traffic Equilibrium 679

estimates may however be difficult to obtain; a trade-off has to be achieved be-

tween large values for λ that may yield to cycling, and small values that induce
slow convergence.
The two-attribute structure is more interesting and lends itself to three lines
of attacks. The first consists in discretizing the density function h and solv-
ing the resulting finite-dimensional problem by the decomposition approaches
outlined above. A second strategy consists in solving a fixed point problem,
either with respect to the total flow vector v̄ or the critical vector α, with the
drawback that the choice of a suitable relaxation parameter may be difficult to
estimate. We favor a third approach, viz. addressing the infinite-dimensional
formulation (34) by a cost approximation method akin to the Frank–Wolfe lin-
earization scheme. At iteration τ = 1, one sets
 
vτ+1 = vτ + λτ wτ − vτ  (58)
where wτ is the solution to the infinite-dimensional parametric shortest paths
problem. Since the information provided by the total flow vector v̄ is sufficient
to recover the flow densities v(α), the algorithm can also be defined in the
finite-dimensional total flow space. If the conditions ensuring the validity of the
optimization formulation (40) are fulfilled, λ can be determined by performing
a line search along the direction dτ = wτ − vτ . Otherwise, one may minimize,
along dτ , the primal gap function
 
gap(v) = maximum t(v̄) + αf(v̄) (v − w)
w∈F
 ᾱ " #
= maximum t(v̄) + αf(v̄) v(α) − w(α) dα
w∈F 0

#" ᾱ " #
= maximum t(v̄) v̄ − w̄ + α f(v̄) v(α) dα
w∈F 0
 ᾱ " #
− α f(v̄) w(α) dα
0
The first term of this expression is the term found in the standard model, while
the last two are the average costs associated with flow densities v and w, respec-
tively. If f is constant then the update of the latter quantity can be performed
efficiently.
Without further assumptions on the problem’s structure, the above de-
scribed algorithm is a heuristic. However, the algorithm can be made globally
convergent by imposing two simple conditions on the cost functions:
• Invariance condition: The mapping f is constant over its domain F.
• Slope condition: For any two distinct routes r1 and r2 there holds
 
fl (v̄) = fl (v̄) v̄ ∈ F (59)
l∈r1 l∈r2
680 P. Marcotte and M. Patriksson

The slope condition derives its name from the graph of the parametric LP,
illustrated in Figure 5, whose solution yields the search direction dτ =
wτ − vτ . It states that the monetary cost associated with distinct routes are dis-
tinct. This condition can be enforced through a perturbation of the link costs,
whenever the invariance condition holds.

Theorem 15. Let the invariance and slope conditions hold. If t is monotone
over F, then the link-flow solution v∗ of the multi-attribute problem is unique,
almost everywhere. Moreover, the path-flow solution is unique as well.

The above results are surprising, as both conclusions may fail to hold
for the standard model. Indeed, the first conclusion usually requires a strict
monotonicity condition while the second fails if there exist more than one
path flow vector compatible with a given link-flow solution. The situation is
altogether different in the multi-attribute model, where the slope condition
suffices to insure the uniqueness of the solution to a parametric LP. At equi-
librium, every α-group is assigned to a unique path. While such a solution is
extremal in the infinite-dimensional setting, the corresponding total flow vec-
tor will in all likelihood not coincide with an extreme point of the polyhedron
F. Another consequence of the slope condition is that the search direction dτ
induces a feasible but nonextremal total flow direction, relative to the polyhe-
dron F. It is superior to an ordinary Frank–Wolfe direction in the sense that it
is a descent direction for the gap function and induces convergence at a linear
rate!

Theorem 16. Let the invariance and slope conditions hold. Let t be strongly
monotone (with modulus b) and Lipschitz continuous (with modulus L) over F.
If the step size if fixed and satisfies 0 < λτ ≡ λ < minimum{1 2b/L}, then the
cost approximation algorithm converges geometrically to the equilibrium solution.
More precisely:
   
g vτ+1  (1 − λ)g vτ 
$ τ $  
$v̄ − v̄∗ $ = O (1 − λ)τ/2 
$ τ $  
$w̄ − w̄∗ $ = O (1 − λ)τ/2
hold.

The last equality in the theorem states that the total flow vector w̄τ corre-
sponding to the solution wτ of the parametric shortest paths problem converges
to the optimal total-flow equilibrium. Although the asymptotic rate of conver-
gence is similar to that of the sequence v̄τ , the actual convergence is much
slower, as observed empirically on large scale problems.
One drawback of the fixed step size scheme is that its validity rests on para-
meters b and L that might be difficult to estimate. This can be fixed by adopting
a step size rule that obeys the Armijo condition with respect to the primal gap
 Ch. 10. Traffic Equilibrium 681

function. Alternatively, if this solution is deemed to expensive computationally,

the step size can be made to obey one of the following two rules:
∞

RULE I: 0 < λτ < 1 lim λτ = 0 λτ = +∞;
τ→∞
τ=0
1
RULE II: λ0 = 1 λτ+1 = λτ − λ2τ 
2
In either case, global convergence is preserved, but asymptotic convergence
may be sublinear.
Implementation. The main challenge in implementing the cost approxima-
tion algorithm lies in the numerical resolution of the parametric shortest paths
problem. One may of course solve this parametric LP for its optimal solution by
a parametric network simplex procedure, but this would prove costly on large
networks involving thousands of paths, many of them irrelevant, and many
others only active within a critical interval (αi−1  αi ) of insignificant width.
An alternative is to replace the density function h by a suitable discrete ap-
proximation. In practice, a 20-bar histogram yields results that are virtually
indistinguishable from the exact solution. Might we then argue that the run-
ning time of a multi-attribute algorithm is roughly 20 times that of a standard
model of similar size? No. We claim that the actual ratio is much smaller, as
the convergence rate partially offsets the number of discrete problems (20) to
be solved. Finally, one might ask what is the advantage of this approach over
discretizing h before linearizing the functions t and f, a dilemma reminiscent
of the Jacobi–Newton vs. Newton–Jacobi paradigm. Our answer is twofold.
First, it is easier to monitor the convergence process through the control of a
well-defined combinatorial subproblem. Second, let us not be shy about it: the
analysis has some mathematical elegance.

5.2 Side constraints

The derivation of the side constrained traffic equilibrium model and its equi-
librium characterization suggests the following algorithmic construct: guess a
value of the multiplier vector β, solve a standard traffic equilibrium model in
terms of generalized link costs, and update the value of β based on a conver-
gent algorithm for solving the Lagrangian dual problem. If the number of side
constraints is relatively small, then this approach can be extremely efficient,
since the subproblems are standard traffic equilibrium problems and the dual
space has a small dimension.
Provided that the link travel cost functions are strictly increasing, the
Lagrangian is strictly convex. Numerical experiments show however that,
whenever an augmented Lagrangian algorithm is used for a link capacitated
problem, convergence to a near-optimal, near-feasible solution is reached even
more quickly. Starting from a near-optimal and near-feasible solution, it is then
682 P. Marcotte and M. Patriksson

relatively straightforward to generate a primal feasible and near-optimal solu-

tion by means of specialized graph search techniques working in a residual
graph.

5.3 Strategic model

In this section, we develop a strategy for solving the basic model. These ideas
can be extended to the dynamic and priority models. The challenge in finding
an equilibrium solution is threefold:
(i) The cost mapping c is not available in closed form.
(ii) c is neither differentiable nor monotone.
(iii) The entire information pertaining to the strategic flow x must be pre-
served at every iteration.
The second issue implies that only heuristic methods based on first-order in-
formation are implementable. One such proposal is based on the solution of a
linear program; at iteration τ + 1, one sets
 
xτ+1 = xτ + λτ yτ − xτ 
where λτ ∈ [0 1] and
 
yτ ∈ arg minimum c xτ y
y∈X

the vector of best strategic responses to xτ , is obtained by solving Bellman’s

dynamic programming equations, for every destination node q:
⎧
⎨ ∞ if j > q
s ∗
ωj =  0 if j = q
∗ ∗ (60)
⎩ s ∗ π s tjk + ω s  if j < q
k∈E jkj k

and assigning all strategic flow yτ to strategy s∗ . In the above expression,

∗ s∗ is the probability of ac-
Ejs denotes the preference order of strategy s∗ , πjk
cessing node k from node j, using s∗ , tjk is the traversal time of link (j k) and
∗
ωsk is the cost-to-go function. Since the vector π is a by-product of the load-
ing process, 20 the recursion will be well-defined once the preference orders
are set. At a given node, one may of course consider all possible such orders.
However, under the FIFO queue discipline, one can prove the optimality of
the following greedy approach:
∗   ∗
Ejs ∈ arg minimum
s
s
πjk tjk + ωsk  (61)
Ej
k∈Ejs

20 This result also holds for strategies that have not been considered yet, i.e., strategies that carry no
flow.
 Ch. 10. Traffic Equilibrium 683

The time required to compute c(x) favors the use of a preselected sequence
of step sizes, such as the harmonic sequence λτ = 1/τ. An alternative which
proves efficient in practice is to weight the harmonic step size 1/τ by the non-
negative vector c(xτ )−c(yτ ), which is a measure of departure from equilibrium
or, in layperson terms, a measure of user dissatisfaction. The third point is best
addressed by resorting to restriction (simplicial decomposition) techniques.

5.4 Nonadditive route costs

Algorithms for nonadditive traffic equilibrium models need explicit route

information; a simplicial decomposition/route (column) generation algorithm
à la DSD therefore seems natural. In the case of the above cost model, the only
place where some complications arise is in the generation of new, profitable
routes. The Wardrop conditions suggest the following extension of the shortest
route problem, for an OD pair k ∈ C , at some fixed link volume v̄:

minimize λlr tl (v̄) + αr (fr )
r∈Rk
l∈L

If the current solution is not in equilibrium, it can be improved through

the reallocation of flow onto shortest routes, exactly as in Frank–Wolfe’s lin-
earization method. Unfortunately, the above shortest route problem cannot
be solved using Dijkstra’s algorithm, since each link possesses two attributes:
travel time and money outlay; in particular, the dynamic programming prin-
ciple – that subroutes are optimal if the entire routes are – does not apply.
Instead, more complexmultilabel shortest route methods must be used.
Suppose that fr = l∈L λlr flr , so that the monetary outlay on the route is
the sum of link tolls. Then we can devise an algorithm in which node labels in-
corporate both accumulated travel times and monetary outlays, from a specific
node on to the destination nodes. Domination tests (based on the principle
of Pareto optimality) are used to keep down the number of labels stored, and
those remaining at the terminal node are compared with respect to the func-
tion νk to determine the best route. The corresponding extension of the DSD
algorithm works well for this problem.

5.5 Bibliographical notes

Algorithms for the bi-attribute model have been analyzed in Marcotte and
Zhu (1997). An implementation based on the parametric network simplex
method is described in Marcotte et al. (1996) in the congestion-free case.
The side constrained model discussed in Section 5.2 had previously been
considered algorithmically mainly for the link capacitated case, typically
through penalty methods. The algorithm in Larsson and Patriksson (1995)
(whose literature section traces the history of such methods within the context
of transportation science) combines an augmented Lagrangian method with
684 P. Marcotte and M. Patriksson

the DSD algorithm used for the subproblem. One of the few algorithms for
the general side constrained model is found in Larsson et al. (2004), which
combines Lagrangian relaxation and column generation.
The implementation of the strategic model is discussed in Marcotte et al.
(2004), while numerical results on the nonadditive model are found in Larsson
et al. (2002).

6 Optimization in a user equilibrium context

Equilibrium models are by construction descriptive, passive mathematical

objects that can be used to assess the impact of managerial policies. Ideally,
a design model should be able to integrate equilibrium equations within the
overall control process. In this section, we address the design problem both
from the local and global perspectives. In this respect, the first subsection pro-
vides sensitivity results for the traffic equilibrium problem, while the next two
subsections are concerned with the problem of imposing tolls either for con-
trolling flow patterns through the network, or for raising revenues.

6.1 Stability and sensitivity of traffic equilibria

Although Braess’ paradox, which was discovered in the late 1960s, prompted
a study into the qualitative behavior of a traffic equilibrium under varying
conditions, a thorough quantitative study, in the form of a variational and
differential analysis, is much more recent. This section traces the main vari-
ational characteristics of a traffic equilibrium, and shows when the equilibrium
is (directionally) differentiable as a function of input data; the results cited
improve quite substantially on previous analyses that relied heavily on the Im-
plicit Function Theorem and are therefore only applicable in very restrictive
settings.

6.1.1 Introduction
The basis of our sensitivity analysis is a result which is stated for a gen-
eral variational inequality problem over a polyhedron and with a differentiable
mapping, f : d × n → n in the parameters ρ ∈ d and variables x ∈ n :
find x∗ ∈ X such that

f(ρ x∗ ) (x − x∗ )  0 x ∈ X (62)
where X ⊆ n is a polyhedral set.21 We let S : d ⇒ 2d denote the mapping
that assigns to each vector ρ ∈ d the set S(ρ) of solutions to this problem.

21 That the set X is polyhedral is crucial. It is however possible to take into account nonlinear con-
straints by reformulating the problem such that Lagrange multipliers are explicitly used.
 Ch. 10. Traffic Equilibrium 685

Letting ρ = ρ∗ be the current value of the parameter vector, we are interested

in the direction of change of the solution x∗ as ρ∗ is perturbed along a direc-
tion ρ . This directional derivative of S is the solution to an auxiliary variational
inequality, which has the following form: find x ∈ K such that

r(ρ  x ) (x − x )  0 x ∈ K (63a)
where

K := TX (x∗ ) ∩ f(ρ∗  x∗ )⊥ (63b)

and

r(ρ  x ) := ∇ρ f(ρ∗  x∗ )ρ + ∇x f(ρ∗  x∗ )x  (63c)

We let DS(ρ∗ |x∗ ) : d ⇒ d denote the mapping that assigns to each pertur-
bation ρ ∈ d the set DS(ρ∗ |x∗ )(ρ ) of solutions to this problem. The set
K above is referred to as the critical cone, and denotes the set of variations
around x∗ that, roughly speaking, retains feasibility and optimality to the first
order. TX denotes the tangent cone to X, which means that if X is defined by
linear constraints, we have that
 
X = x ∈ n | Ax  b; Bx = d
 
⇒ TX (x∗ ) = z ∈ n |  Az  0; Bz = 0 

where A consists of the rows Ai of A corresponding to the binding inequality

constraints at x∗ , that is, the indices i with Ai x∗ = bi . Further, for any vector
z ∈ n , z⊥ := {y ∈ n | z y = 0} is the orthogonal subspace associated with
the vector z. The mapping r is a linearization of f around (ρ∗  x∗ ); it is an affine
mapping in x .
Suppose now that f(ρ ·) is monotone on X around ρ = ρ∗ , and that the
parameterization is such that the rank of the matrix ∇ρ f(ρ∗  x∗ ) is equal to n.
(The latter result can always be fulfilled by including dummy parameters.) We
say that the mapping S is strongly regular at ρ∗ if S is single valued and Lipschitz
continuous on some neighborhood of ρ∗ . Then,

S is strongly regular at ρ∗ (64a)

∗ ∗
⇐⇒ DS(ρ |x ) is single valued (64b)
Moreover, the unique solution x ∈ n to (63) is the directional derivative of
the solution x∗ to (62) at ρ∗ , in the direction of ρ . A sufficient condition for
the property (64b) to hold is that

∇x f(ρ∗  x∗ ) is positive definite on (K − K) (65)

We refer to this as a sufficient second-order condition. A result stronger than
directional differentiability can also be obtained under additional assumptions.
686 P. Marcotte and M. Patriksson

Indeed, under strong regularity,

S is differentiable at ρ∗ (66a)
∗ ∗   d
⇐⇒ DS(ρ |x )(ρ ) ∈ −K ρ ∈  (66b)
Moreover, if the critical cone K is a subspace, that is, if K = K ∩ (−K), then
the gradient can be represented as
−1
∇ρ x(ρ∗ ) = −Z Z ∇x f(ρ∗  x∗ )Z Z ∇ρ f(ρ∗  x∗ ) (67)
for any n ×  matrix Z such that Z Z is nonsingular and z ∈ K ∩ (−K) if and
only if z = Zy for some y ∈  , where  is the dimension of K ∩ (−K). This
differentiability result is a kind of implicit function theorem; the relationships
in (64) and (66) show how the implicit function theorem naturally extends to
more general cases.
The latter property has been invoked in many sensitivity analysis frame-
works. That is, they hinge on being able to invoke the implicit function the-
orem. Unfortunately, not only does the property DS(ρ∗ |x∗ )(ρ ) ∈ −K fail to
hold in many cases (cf. below), but also there may not exist a nonsingular ma-
trix of the kind referred to above.

6.1.2 Sensitivity analysis of separable traffic equilibrium

For simplicity of presentation, we focus on the case of separable cost and
demand functions. First, we cast the VIP (8) in the sensitivity framework. Let
   0|R| 
h
|R|
x= v ; f(ρ x) = t(ρ v) ; X = + × |L| × |C | 
d −ξ(ρ d)
Then, we can identify the sensitivity problem through the following identifica-
tions:
 h   

K= v ∈ |R| × |L| × |C |  h = d ; v = h ; h ∈ H  
d
where
⎧   ⎫
⎪ hr free if h∗r > 0 ⎪
⎪
⎨   ⎪
 

|R| hr  0 if h∗r = 0 and cr (ρ∗  h∗ ) = πk∗  ⎬
H = h ∈   
⎪
⎪  hr = 0 if h∗r = 0 and cr (ρ∗  h∗ ) > πk∗  ⎪⎪
⎩ [r ∈ R  k ∈ C ] ⎭
k
and
 0|R| 
r(ρ  x ) = ∇ρ t(ρ∗  v∗ )ρ + ∇v t(ρ∗  v∗ )v 
−[∇ρ ξ(ρ∗  d∗ )ρ + ∇d ξ(ρ∗  d∗ )d ]
By the monotonicity and separability of t and −ξ, the resulting ‘sensitivity vari-
ational inequality’ (63) can be equivalently written as the convex quadratic
 Ch. 10. Traffic Equilibrium 687

optimization problem to

minimize φ (v  d ) := ∇ρ t(ρ∗  v∗ )ρ v

(v d h )
1  ∂tl (ρ∗  vl∗ )  2
+ (vl )
2 ∂vl
l∈L

− ∇ρ ξ(ρ∗  d∗ )ρ d
1  ∂ξk (ρ∗  dk∗ )  2
− (dk )  (68a)
2 ∂dk
k∈C

subject to  h = d  (68b)
 
v = h  (68c)
 
h ∈H (68d)
The sensitivity problem is closely related to the original model, with two
notable differences: the link cost and demand functions are replaced by their
linearizations, and the sign restrictions on h are replaced by individual restric-
tions on the route flow perturbations hr that depend on whether the route in
question was used at equilibrium or not, cf. the set H  . Although the appear-
ance of H  depends on the choice of route flow solution h∗ , it is an interesting
fact that the possible choices of v in K do not; this is a general consequence of
aggregation, and amounts to the possibility of (essentially) eliminating route
flow variables from explicit consideration; this is only possible because of the
special connection between route and link flow variables. Finally, we see that
the resemblance to the original problem implies that the sensitivity variational
inequality problem can be solved using software similar to those for the orig-
inal traffic equilibrium model, provided of course that route flow information
can be extracted.

Assumption 17 (Properties of the network model).

(a) For each l ∈ L, the link travel cost function tl (· ·) is continuously dif-
ferentiable, and strictly increasing in its second argument.
(b) For each k ∈ C , the demand function gk (· ·) is continuously differen-
tiable, nonnegative, upper bounded, and strictly decreasing in its second
argument. The function gk (ρ ·) is therefore invertible, and has a single-
valued inverse, ξk (ρ ·), which also is continuously differentiable and
strictly decreasing.

Theorem 18 (Sensitivity of separable traffic equilibrium problems). Let As-

sumption 17 hold, and consider an arbitrary vector ρ∗ ∈ d . Then, the so-
lution (v∗  d∗ ) to (16) is unique, and so are the (negative) travel cost entities
(s∗  π ∗− ) = −(t(ρ∗  v∗ ) ξ(ρ∗  d∗ )).
688 P. Marcotte and M. Patriksson

Assume that the link travel cost function t(ρ∗  ·) is such that
∂tl (ρ∗  vl∗ )
> 0 l ∈ L (69)
∂vl
Assume further that the demand function g(ρ∗  ·) is such that22
∂gk (ρ∗  πk∗ )
< 0 k ∈ C (70)
∂πk
Then, in the solution to (68), the values of the link flow and demand perturba-
tion v and d are unique; therefore, the value v (respectively, d ) is the directional
derivative of the equilibrium link flow (respectively, demand), at ρ∗ , in the direc-
tion ρ .

6.1.3 Limitations of the classical sensitivity framework

Sensitivity analyses that extend standard techniques for nonlinear opti-
mization and variational inequality problems are widespread, and have been
adopted by several researchers over the past fifteen years, particularly as a sub-
routine in the solution of more complex problems of a bilevel nature (see, for
example, Section 6.4). These analyses have however a drawback: they often are
inapplicable because of the strong assumptions underlying their validity. Part
of the problem lies in the utilization of the implicit function theorem, which
requires the problem to be expressed as a system of equations; this leads to a
strict complementarity type of condition, which may fail to be in force at dif-
ferentiable points. Additional requirements pose, for example, conditions on
the topology of the network itself, and are not necessary either. It appears that
some researchers are not aware of these limitations, and of the existence of
more widely applicable analyses. This justifies devoting some space to exam-
ples where this ‘classical’ sensitivity framework fails, while the techniques of
the previous section still apply.

Example 19 (A nonstrictly complementary example). The classical analysis re-

quires that the solution be strictly complementary. In our context this means
that there must exist some equilibrium route flow solution h∗ which has the
property that for every route in R, h∗r > 0 holds if and only if route r is a
shortest route given the equilibrium travel costs. The definition in the classical
work is however based on the total link flows and an aggregated ‘node price’,
common to all OD pairs, which may not exist (see, for example, the numerical
example in Section 2).
To show that strict complementarity is not necessary for differentiability (as
we have already seen from the above analysis), we consider the network de-
picted in Figure 11. It involves two OD pairs, (1 2) and (4 2), with a fixed

22 The two derivative conditions (69) and (70) imply that the functions t (ρ∗  ·) and −g (ρ∗  ·) are
l k
strictly increasing.
 Ch. 10. Traffic Equilibrium 689

Fig. 11. Network for the first example.

and unperturbed demand of 2 and 1 units of flow, respectively. The link cost
functions are given by
t1 (v1  ρ) := 2v1 + ρ; t2 (v2 ) := v2 ; t3 (v3 ) := 1;
and
t4 (v4 ) := v4 + 2; t5 (v5 ) = v5 
We have four routes: {1}, {2 3}, {4}, and {5 3}, two for each OD pair.
If ρ∗ = 0, the unperturbed traffic equilibrium solution is v∗ = (1 1 1 1 1)
and the route flow h∗ = (1 1 0 1) is unique. We observe that the travel cost
on route 3 is equal to 2, as is the case for route 4, so this equilibrium solution
is not strictly complementary.
In order to check whether the solution v∗ is nevertheless differentiable at
ρ∗ = 0, we solve the sensitivity problem for both ρ := 1 and ρ := −1. For
ρ = 1, we obtain the following unique solution to the sensitivity problem, thus,
being the directional derivative of v∗ with respect to the direction ρ = 1 at
ρ∗ = 0: v = (− 13  13  13  0 0) . The effect of perturbing link 1’s cost, such that
it becomes more expensive, is that of sending flow along the cycle {−1 2 3},
where the minus sign reflects that flow is sent backward on link 1. When solving
the sensitivity problem for ρ := −1, we obtain the directional derivative v =
( 13  − 13  − 13  0 0) , that is, the negative of the directional derivative of v∗ in
the direction of ρ := 1. This proves that the directional derivative mapping
is linear, and thus that the derivative of v∗ with respect to ρ at ρ∗ = 0 equals
dv∗ /dρ = (− 13  13  13  0 0) .

Example 20 (Second example). A key assumption of the classical analysis is

that the travel cost function t(ρ ·) be strongly monotone in a neighborhood of
the equilibrium solution. This is clearly not needed, as we have shown above.
More serious, however, is the presence of a condition on the topology of
the graph representing the traffic network. Suppose that we were to limit the
discussion to a subgraph of G in which only the links carrying positive flow are
included. In this subnetwork, the analysis must be performed from an equilib-
rium route flow solution that is an extreme point of H (the routes using the
690 P. Marcotte and M. Patriksson

Fig. 12. Network for the second example.

above-stricken zero-flow links also having been stricken) which has exactly as
many routes with positive flow as the rank of the matrix [+ | + ] (the + sign
indicates that we have eliminated the zero-flow routes, as discussed). The rank
of this matrix is never larger than the number of links with positive flow at v∗
plus |C |. Although the choice of the route flow solution is immaterial, it must
be extremal, as explained above. The resulting formula then takes the form
   −1  
∇ρ h+ ∇h c+ (ρ∗  h∗+ ) −+ −∇ρ c+ (ρ∗  h∗+ )
=  (71)
∇ρ π + 0 ∇ρ g(ρ∗ )
Unfortunately, it is not difficult to construct examples where the solution is
differentiable but where the matrix condition fails, so that the formula (71)
breaks down. Consider the network shown in Figure 12, involving a single OD
pair, (1 3), and a fixed demand of 2 flow units. Set the link cost functions to
t1 (v1  ρ) = v1 + ρ; t2 (v2 ) = v2 ; t3 (v3 ) = v3 ; t4 (v4 ) = v4 
In this example, we have four routes: {1 3}, {1 4}, {2 3}, and {2 4}.
If ρ∗ = 0, the unperturbed traffic equilibrium solution is v∗ = (1 1 1 1) .
We can easily see that the solution is differentiable, and even strictly comple-
mentary. The derivative with respect to ρ at ρ∗ is (− 12  0 12  0) . This result is
intuitive: if the value of ρ increases, then the flow on link 1 should decrease,
whence the flow on link 2 must increase by the same amount. If, on the other
hand, the value of ρ decreases, the reverse should happen.
Consider then the workings of the classical formula (71). We obviously fulfill
the strong monotonicity conditions on the travel cost function. Since all links
carry flow at equilibrium, we need not remove any links or routes when con-
sidering the sensitivity analysis problem. We last try to comply with the linear
independence condition, by choosing the right equilibrium route flow solution.
Note then that
⎛ ⎞
1 0 1 0 1
⎜1 0 0 1 1⎟
 | = ⎝
0 1 1 0 1⎠

0 1 0 1 1
which has rank 3. So, we should find a route flow solution, h∗ , in which exactly 3
routes have a positive flow. This is however impossible; the only alternatives
 Ch. 10. Traffic Equilibrium 691

are 2 or 4. To see why, let us suppose that the flow on the first route, {1 3},
is α ∈ [0 1]. Then, the flows on routes {1 4} and {2 3} must both be 1 − α,
in order to comply with the total flow on the links. This implies that the flow
on route {2 4} is α and that, for any value of α ∈ [0 1], the number of routes
carrying nonzero flow is either 2 or 4. We can therefore not comply with the
matrix condition stated, and the classical formula (71) fails, even though the
gradient does exist.

Note that the technique of Section 6.1.2 carries over to the asymmetric case.
This approach yields the recommended source for sensitivity information, not
only because it is widely applicable, but also because the sensitivity analysis
problem, being closely related to that of the original equilibrium problem, can
be solved using similar computational tools. This is especially important when
relying on sensitivity information for an extended traffic model, for example
of the bilevel type, where many such analyses must be performed (cf. Sec-
tion 6.4). The classical formula may then fail, either because of the topology
being ‘wrong’ or for lack of strict complementarity (and the latter will typi-
cally be the case at a stationary point of a bilevel program); in either case,
the formula breaks down, regardless of the existence, or not, of a gradient. It is
common to claim that the possibility of the nondifferentiability of a given point
in the space of ρ can be ignored because of Rademacher’s theorem, which
essentially states that a locally Lipschitz continuous function is differentiable
everywhere except possibly on a set with a zero (Lebesgue) measure. This type
of argument ignores that several of the most interesting points to look at are
points of the latter category; for example, optimal solutions to bilevel programs
and, more generally, MPECs (Mathematical Program with Equilibrium Con-
straints) problems, are indeed extremal.23

6.2 Inducing system-optimal flows through link tolls

Tolls, either physical or virtual, may be used not only to raise money but also
to alter user behavior in order to improve the performance of a transportation
system with respect to indicators such as travel time or pollution. In this regard,
let vSO be a globally optimal solution to the mathematical program

minimize t(v) v (72)

v∈F

and let us first consider the problem of setting tolls that induce system-optimal
flows vSO , while being compatible with the selfish behavior of users.24 Then

23 The reader may draw a piece-wise linear convex function in  and see for herself that if there exists
a global minimum, then there exists also a ‘nondifferentiable’ global minimum.
24 This is a particular instance of (43) with v∗ = v . For notational simplicity, we have adopted a link
SO
flow formulation of the equilibrium problem.
692 P. Marcotte and M. Patriksson

any toll vector ρ that satisfies

− t(vSO ) + ρ ∈ NF (vSO ) (73)
clearly satisfies our requirements. We say that Equation (73) describes an in-
verse optimization problem where, given an optimal solution to a mathematical
program, one looks for a vector of parameters (ρ in the above) that reproduces
the given solution. To pursue the analysis, we assume that the polyhedron F
takes the form F = {v | Bv = b v  0}, so that one can rewrite the toll
Equation (73) in the Karush–Kuhn–Tucker format
t(vSO ) + ρ − B π  0 (74a)
vSO t(vSO ) + ρ − B π = 0 (74b)
vSO ∈ F (74c)
Because vSO is known and feasible, we obtain that the solution set of (74) as-
sociated with the system-optimal flow vSO is the polyhedron P defined as
ρ  B π − t(vSO )
ρi = B π − t(vSO ) i  if (vSO )i > 0
Since vSO is solution to the variational inequality
− t(vSO ) + t (vSO )vSO ∈ NF (v) (75)
the marginal cost solution ρ = t (vSO )vSO is an element of P. If the system-
optimal solution is not unique in link-flow space, then the set P must be
enlarged to include all polyhedra associated with system-optimal flows. The
resulting union of polyhedra is, in the general situation, not convex.
Under suitable assumptions, vSO is uniquely defined. However, the set P
contains several elements, which suggests optimizing over a secondary crite-
rion. For technical or political reasons, one may wish to minimize the total
revenue levied, and insist that link tolls be nonnegative. This yields the mathe-
matical program
minimize ρ vSO  (76a)
ρ0π

subject to t(vSO ) + ρ − B π  0 (76b)

 
vSO t(vSO ) + ρ − B π = 0 (76c)
Since both terms of the complementarity equation are nonnegative, this con-
straint can be replaced by the constraint (76c). Writing its dual linear program
as that to
maximize t(vSO ) (αvSO − v)
α0v
subject to v − αvSO  vSO 
Bv = αb
 Ch. 10. Traffic Equilibrium 693

and performing the change of variable v = αw yields the parametric multicom-

modity flow problem to

maximize t(vSO ) (vSO − w) (77)

α0w∈F
 
1
subject to w  1 + vSO 
α
which is equivalent to maximizeα0 φ(α), where

φ(α) := maximum t(vSO ) (vSO − w) (78)

w∈F
 
1
subject to w  1 + vSO 
α
The scalar function φ has interesting properties: φ(0) = 0, φ(α) is nonnega-
tive (vSO is a feasible solution), φ is piecewise linear, φ is bounded from above
(the primal problem (76) is feasible and bounded), φ is concave, and φ is in-
creasing.
The properties of the function φ imply that there exists a threshold value α∗
such that φ(α) = φ(α∗ ) whenever α  α∗ . In other words, whenever α is
sufficiently large, a single multicommodity flow problem needs to be solved.
The optimal tolls are then retrieved from the dual vector corresponding to the
flow constraints Bw = b.
In a multi-attribute context, marginal cost pricing may not be applicable any
more, as the resulting tolls will be different for distinct classes. It is therefore
surprising that ‘system-optimal’ tolls that achieve the same goal do actually
exist. Let us first analyze a two-attribute (time and money) problem where we
look for a link toll vector that induces a system-optimal solution with respect
to time.25 As in Section 3.1, we denote by αg the VOT parameter associated
with the link-flow vector vg of group g ∈ G . Let

vSO ∈ arg minimum t(v̄) v̄ + αg f(v̄) vg
v∈F
g∈G

denote the system-optimal solution, and v̄SO the associated total link-flow vec-
tor. Consider the linear program

minimize t(v̄SO ) v̄ + αg f(v̄SO ) vg 
v∈F
g∈G

g
subject to v = v̄SO 
g∈G

25 There also exist system-optimal solutions with respect to cost. One must be aware that they need not
coincide.
694 P. Marcotte and M. Patriksson

and the optimal dual vector α∗ associated with its sole explicit constraint. Since
its primal solution is also a solution to the Lagrangian problem
 
minimize t(v̄SO ) v̄ + αg f(v̄SO ) vg + α∗g vg
v∈F
g∈G g∈G

= t(v̄SO ) + αg f(v̄SO ) + α∗ vg 
g∈G

it satisfies the Wardrop equilibrium principle for every individual class g ∈ G .

It follows that the dual vector α∗ induces system-optimal flows and is a suitable
choice for a toll vector. If oneinsists that tolls be positive, then
one can replace
the compatibility constraint g∈G vg = v̄SO by the inequality g∈G vg  v̄SO ,
which must be tight at the optimal solution to the linear program.
The analysis can be repeated in the infinite-dimensional case, setting tolls to
the dual vector of the infinite-dimensional LP to
 ∞
minimize t(v̄SO ) v̄ + αf (v̄SO ) v(α) dα
v∈F 0
 ∞
subject to v(α) dα = v̄SO 
0
By formulating the multi-attribute problem with respect to total flows (see Sec-
tion 3.1), it is also possible to derive a finite-dimensional formulation variant,
and obtain the same result via arguments invoked in the discrete case.
Note that, in the previous analysis, no monotonicity condition was required
from the mappings t and f, and that cycle-free flow patterns other than system-
optimal ones could be induced by the above technique. Note also that, if tolls
cannot be set on all links of the transportation network, it might not be pos-
sible to achieve system optimality through a toll schedule. The resulting prob-
lem, which involves equilibrium constraints, can be formulated as an MPEC.
This class of hard nonconvex and nondifferentiable problems is usually not
amenable to theoretically efficient algorithms, although much attention has
been paid, recently, to the design of locally convergent algorithms. Instances
of such problems are presented in the next two sections.

6.3 Tolls for raising revenues

Let us reconsider the Braess paradox network illustrated in Figure 4. By

imposing a toll ρ23 larger than 13 on the diagonal link (2, 3), one can induce
the link-flow system-optimal solution
v = (v12  v13  v23  v24  v34 ) = (3 3 0 3 3)
with equilibrium cost 83. Alternatively, one may wish to maximize the revenue
generated by the toll link (2, 3). By solving the equilibrium problem for ρ23
fixed, one can express the diagonal flow v23 as a function of ρ23 and obtain the
 Ch. 10. Traffic Equilibrium 695

revenue

4ρ23   if ρ23  −26
ρ23 · v23 (ρ23 ) = ρ23 2 − 2
13 ρ23  if ρ23 ∈ [−26 13]
0 if ρ23  13
The optimal revenue of 13/2 is achieved when ρ23 = 13/2. This toll not only
maximizes revenue but, surprisingly, also reduces each user perceived travel
cost from 92 to 87.5.
In the case of a general network, let L1 denote the set of toll links and
L2 = L \ L1 ; let v = (v1  v2 ), ρ = (ρ1  ρ2 ) = (ρ1  0), B = (B1  B2 ), and
t = (t1  t2 ) be the respective partitions of the link flow vector, toll vector,
constraint matrix and cost mapping. The revenue maximizing problem is to

maximize ρ1 v1 
v∈Fρ1
 
subject to t(v) + ρ (v − w)  0 w ∈ F (79)
ρ2 = 0
If the mapping t is monotone, one can replace the lower level variational
inequality by its Karush–Kuhn–Tucker conditions. This yields the MPEC for-
mulation

maximize ρ1 v1 
v∈Fρ1 π

subject to t1 (v) + ρ1 − B1 π  0
t2 (v) + ρ2 − B1 π  0 (80)
v1 t1 (v) + ρ1 − B2 π = 0
v2 t2 (v) − B2 π = 0
or, upon the introduction of a vector z of binary variables and a suitably large
constant M, the mixed discrete-continuous program

maximize ρ1 v1 
v∈Fρ1 πz

subject to t(v) + ρ − B π  0 (81)

v  Mz  t(v) + ρ − B π
zl ∈ {0 1} l ∈ L
If the mapping t is monotone and separable, (81) can be approximated by a
piecewise linear function, hence the importance of investigating the linear case,
where t(v) = t v. First, due the complementarity conditions, ρ1l = (B1 π)l −
t1l whenever the flow v1l is positive. On the other hand, ρ1l may be set to any
value larger than (B2 π)l − t1l when v1l = 0. Therefore, one may set a priori
ρ1 = B1 π in the optimal solution. Replacing ρ1 by its value in the objective
696 P. Marcotte and M. Patriksson

of (80) one obtains the simplified formulation

 
maximize B1 π − t1 v1 = π (b − B2 v2 ) − t1 v1 = b π − t v
v∈Fπ

subject to t2 − B2 π  0 (82)
 
v2 t2 − B2 π = 0
If one solves a mixed integer reformulation of (82) by an implicit enumeration
method, an upper bound on the revenue must be available at each node of the
branch-and-bound tree. At such a node, let L0 ⊆ L2 denote the index set of
toll-free links with null flow, and let us relax the complementarity constraint
corresponding to flows vl , l ∈ L2 \ L0 , yielding the upper bound
maximum b π − minimum t v
π v∈F
subject to B2 π  t 2 
 
B2 π l = t2l  l ∈ L0 
Replacing the left-hand side LP by its dual, the upper bound can be expressed
as
minimum t v − minimum t v (83)
v∈Fv1 =0 v∈F
subject to v2l free if l ∈ L0 
and can be interpreted as the difference between two shortest distances, one in
a network with tolls set at 0 and some link flows unrestricted in sign, the other
in a network with tolls set at ∞. Once the subset of positive flows has been
selected, a set of optimal tolls compatible with this choice can be retrieved
from the dual solution associated with the LP that appears on the left-hand side
of (83). This operation, which amounts to solving an inverse linear program, can
be performed very efficiently, by computing shortest paths in a network from
which toll links have been removed, and where backward copies of toll-free
links carrying positive flow are introduced.

6.4 A continuous network design problem

In this section, we address the problem that consists in designing a network

subject to equilibrium constraints, through the specification of continuous con-
trol parameters. More specifically, let ϕ(ρ) denote the cost of implementing a
traffic control ρ belonging to some set Z, assumed to be convex. Based on
this notation, the toll problem considered in Section 6.2 constitutes special in-
stances of the general problem to
minimize v t(v ρ) + ϕ(ρ) (84)
ρ∈Zv∈F(ρ)
subject to −t(v ρ) ∈ NF (ρ)(v)
 Ch. 10. Traffic Equilibrium 697

where the vector of design parameters ρ may impact both the cost mapping t
and the feasible set F(ρ). For instance, one may consider improving a road
network through capacity enhancement, balancing long term investment costs
against recurrent travel delays. In this context, let ϕl represent the capacity
of link l, ϕl (zl ) the cost of achieving capacity zl , where we assume that the
functions t and ϕ are link-separable. If, furthermore, we assume that tl is a
function of the flow-capacity ratio, that the set F is independent of ρ, and that
the design vector ρ is unconstrained, then a continuous variant of the network
design problem (84) takes the form
  
vl
minimize vl tl + ϕl (ρl ) (85)
ρ∈Zv∈F ρ l
l∈L
  wl  u 
subject to v ∈ arg minimum tl du
w∈F 0 ρl
l∈L
While the resulting problem is theoretically difficult, it is amenable to effi-
cient heuristic procedures. Maybe the simplest one consists in first solving the
system-optimal problem to
  
vl
H1: minimize vl tl + ϕl (ρl ) (86)
ρ∈Zv∈F ρl
l∈L
for the design vector ρSO and then finding the equilibrium flows compatible
with ρSO . This is Heuristic H1. Another natural procedure consists in iterat-
ing, à la Gauss–Seidel, between flow and capacity assignment subproblems: for
fixed ρ, one solves a traffic equilibrium problem while, for fixed flow pattern v,
each ρl = ρl (vl ) satisfies the single-variable equation
   2  
vl vl vl
H2: tl + tl + ϕl (ρl ) = 0 (87)
ρl ρl ρl
and is frequently available in closed form. If the procedure converges, it does
so to an equilibrium flow that is compatible with the ‘greedy’ optimality condi-
tion (87). This is Heuristic H2, whose solution can also be computed by solving
the single variational inequality
 
−t v ρ(v) ∈ NF (v) (88)
A general class of heuristic procedures consists in solving a problem where
the flow part of the solution automatically obeys Wardrop’s equilibrium prin-
ciple. This can be achieved by replacing the total cost by its integral in the
objective and, for generality, scaling the investment cost term. This yields the
mathematical program
  vl  u  
H3: minimize t(v ρ) = tl du + ξl ϕl (ρl ) (89)
ρ∈Zv∈F 0 ρl
l∈L l∈L
698 P. Marcotte and M. Patriksson

Next, one might look for the set of parameters that yields the best solution
to the original program (85). This problem, however, is of the same compu-
tational complexity as solving (85) directly, and is thus ‘intractable’. In this
respect, it is natural to restrict our attention to a subfamily H3, for instance
that where all ξl are set to a common value ξ. To pursue the analysis, we sup-
pose that ϕ is the monomial ϕl (ρl ) = dl ρml and that the cost function t assumes
the BPR form tl (x) = αl +βl xp . Under these assumptions, the solution to (87)
is given by
 
pβl 1/(p+m) (p+1)/(p+m)
ρl (vl ) = vl  (90)
mdl
If one introduces the function
 1/p  
1 pβl 1/(p+m) (p+1)/(p+m)
ζl (vl ) = vl 
p+1 mdl
one may check that the solution to the convex program
  vl  p !
u
H4: minimize αl + βl du (91)
v∈F 0 ρl (u)
l∈L

coincides with the system-optimal flow of program (86).

Note that, for the above functional forms, the class H3 defined by (89)
subsumes the Gauss–Seidel solution, the latter corresponding to the choice
ξl = 1/(p + 1) for every l ∈ L, and that all heuristics presented find a de-
sign vector z by solving a convex optimization problem, and are thus easy to
implement. To assess the quality of their respective solutions, we adopt the
worst-case analysis point of view, and aim to determine (or estimate) the ratio
p cost of heuristic solution
Rm (H) = sup
αβd cost of (unknown) optimal solution
F(v(ρH ) ρH )
= sup (92)
αβΔ F(v∗  ρ∗ )
over all possible network configurations represented by the symbol Δ. This
analysis is reminiscent of the worst-case analysis presented in Section 2.9. We
have the following approximation results.
p
Theorem 21. For heuristics H1 through H4, the following bounds on Rm hold:
p
(a) limp→∞ R1 (H1)  2;
p
(b) R1 (H2) = p + 1;
p
(c) Rm (H4) = m(p + 1)/(m + p) + p(m + p)(p + 1)−m/p ;
p p
(d) limm↓0 Rm (H4) = 1 and limp→∞ R1 (H4) = 2;
p p ξp/(p+1) p
(e) 1 + ξ(p+1)  R1 (H3)  (p+1)1/(p+1)
[1 + ξ(p+1) ]2 ;
 Ch. 10. Traffic Equilibrium 699

p
(f) 2  limp→∞ R1 (H4) = 4.

Remark 22. (i) For linear investment functions (m = 1), the value that yields
p
the tightest upper bound on R1 (H3) is ξ = 2 − p/(p + 1).
(ii) The bounds provided by the theorem are very pessimistic. Indeed,
numerical experiments show that the error is of the order of one percent. This
is not surprising if one observes that the objective of both ‘players’ are not
antagonistic: the leader actually wishes to minimize the users’ travel times.
(iii) If the investment function ϕ is linear, Equation (87) fixes a value for
the ratio vl /ρl . It follows that t(v ρ(v)) is constant in the variational inequal-
ity (88) and that its solution is an extremal flow pattern that can be efficiently
computed by shortest routes methods.

Once a heuristic solution is obtained, it can be improved using the sensitivity

results of Section 6.1 with respect to the design (toll) vector ρ. At differentiable
points, the procedure reduces to gradient projection. At points of nondifferen-
tiability, a descent direction can be tentatively built from the coordinate-wise
directional derivatives. In the numerical example below, these were always suf-
ficient to provide a descent direction at every iteration.
We consider the special case of problem (85) where Z is a box, declaring
that 0  ρl  ul must hold for every l. An implementation of the DSD algo-
rithm, the directional derivative problem, and a gradient projection algorithm
with a rather simple line search,26 form a complete descent-based framework
for this network design problem. We have applied the procedure to a prob-
lem defined for the Sioux Falls network (with 10 potential links to improve),
and compared it to some heuristic schemes that have been proposed for the
continuous network design problem. The results are shown in Table 8.
In the table, we use the following short-hand notation: SBD stands for ‘sen-
sitivity-based descent’, that is, the algorithm presented in this section; H–J
stands for ‘Hooke–Jeeves’, a direct search method based only on objective
values; EDO stands for ‘equilibrium decomposed optimization’, an algorithm
which utilizes the ‘classical’ sensitivity analysis formula (71); SA stands for ‘sim-
ulated annealing’; and PIPA stands for ‘penalty interior point algorithm’. The
Notes section provides the sources for the results in the table.
The values in the last row of the table corresponds to a much more accu-
rately calculated upper-level objective value for the terminal designs; in the
case of our algorithm, the difference between the nominal values are small,
since accurate traffic equilibrium computations are used throughout. For the
other algorithms – which typically use 50–100 Frank–Wolfe iterations on each
lower-level problem – the differences are much larger. Clearly, their final solu-
tions suffer from a poor estimation of traffic equilibrium solutions, in addition

26 We have used the Armijo step length rule; the main motivation is that the objective value calculations
are quite expensive.
700 P. Marcotte and M. Patriksson

Table 8.
Results for network design on the Sioux Falls network

SBD SBD SBD SBD H–J EDO SA H–J EDO SA PIPA

Initial value ρ0 : 0 2 3 6.25 2 – 6.25 3 0 6.25 –
Upper bound u: 25 25 25 25 – 25 25 – 25 25 –

ρ16 53027 51492 53457 52773 48 459 538 4507 4276 5322 54680
ρ17 20560 20214 19786 20533 12 152 226 4509 2288 2596 20039
ρ19 53430 51679 53741 53002 48 545 550 4520 4080 5664 54471
ρ20 19901 20012 19460 20369 08 233 201 4052 1618 1309 19395
ρ25 25216 24945 27856 27670 20 127 264 4299 1654 2498 29448
ρ26 25548 25447 28245 28222 26 233 247 2949 1130 2732 28191
ρ29 29883 29535 29257 30124 48 041 454 3000 3219 4123 34039
ρ39 48559 48330 47528 47348 44 459 445 3601 3326 4508 48061
ρ48 30026 29798 29732 29746 48 271 421 3006 1981 3736 32364
ρ74 48496 48212 47347 47511 44 271 467 3200 3190 3903 47779
Obj, reported 799969 799968 799987 800026 8078 8308 8087 83316 83703 81983 808669
Obj, recomputed 799961 799971 799990 800043 8067 8234 8042 81185 81345 80304 800528

to being further from a local optimum (or, a stationary point) than in the pro-
posed algorithm.
The reason why traffic equilibrium computations need to be accurate be-
comes obvious when one realizes that they serve as input to the sensitivity
analysis procedure, which in turn provides the gradient values or, at least,
coordinate-wise directional derivatives. Unless the equilibrium solution is ac-
curately computed, the sensitivity analysis may provide erroneous results, lead-
ing to bad search directions, or to the algorithm halting prematurely. However,
this is not a serious problem with the DSD algorithm, which can solve the traf-
fic equilibrium problem to a sufficiently good accuracy quickly. This must be
taken as an indication that the Frank–Wolfe method, or any other slowly con-
verging method for the traffic equilibrium problem, should definitely not be
used.
We remark finally that the necessity to provide accurate equilibrium solu-
tions arises in several other applications involving traffic equilibrium models.
We mention here only two, in addition to all models of a hierarchical nature,
such as traffic management or toll optimization type models: (i) assessing the
environmental impact of the transportation network (exhaust fumes, for in-
stance) relies on accurate link usage data, which in turn rely on accurate traffic
equilibrium solutions; (ii) travel forecasting, subject to changes in the network
configuration, also relies on accurate flow estimates, both for the ‘before’ and
‘after’ simulations.

6.5 Bibliographical notes

The sensitivity analysis of Section 6.1 was first developed in Patriksson and
Rockafellar (2002, 2003) and later refined in Patriksson (2004). (Regarding
 Ch. 10. Traffic Equilibrium 701

the background material, (64) stems from Dontchev and Rockafellar, 2001,
and (66) from Kyparisis, 1990; see also Robinson 1980, 1985, 1991.) The mate-
rial presented here is a condensed version that for the main part discusses the
standard, separable model. That the sensitivity analysis does not depend on
the equilibrium route flow chosen was first noticed in Tobin and Friesz (1988),
albeit in a framework that has limitations, as we remark and show; this inde-
pendence was observed and utilized later in Patriksson and Rockafellar (2002,
2003), Patriksson (2004), Patriksson (2006). The examples showing the limita-
tions of the Tobin/Friesz formula (71) are taken from Patriksson and Josefsson
(2003), Josefsson and Patriksson (2005), Patriksson (2006).
While it is a classical result from economic theory that marginal cost pricing
can induce a system-optimal flow pattern, Section 6.2 focuses on toll sched-
ules that achieve that aim at minimal user cost. Since the set of system-optimal
tolls is polyhedral (see Bergendorff et al., 1997; Hearn and Ramana, 1998;
Larsson and Patriksson, 1998), such schedules are solutions to a linear pro-
gram. Dial (1999, 2000), either in the single or multicommodity case, proposed
an algorithm that exploits the underlying network structure. A similar result
was obtained by Marcotte and Savard (2002), who made the link, in the single-
commodity case, with the inverse shortest route problem (see Ahuja and Or-
lin, 2001). In a multi-attribute context, where marginal taxation may not be
feasible, Yang and Huang (2004) gave a constructive proof of the existence
of system-optimal tolls. The extension of their method from the discrete to
the continuous case is new to this book. A more complex existence proof, of
a highly nonconstructive nature (even for the discrete case), was given by Cole
et al. (2003).
Whenever tolls cannot be set on every link of a network, the nature of the
problem becomes more complex, and falls within the realm of bilevel program-
ming. Although bilevel programs, together with the closely related MPECs
(Mathematical Programs with Equilibrium Constraints), have been topics of
several studies in the past years, much work remains to be done before robust
procedures may address real-life instances. The reader is referred to the books
by Luo et al. (1996), Bard (1998), and Shimizu et al. (1997) for introductions
to the subject.
Section 6.3 addresses one such instance, where one aims at maximizing toll
revenues. References for theory, algorithms, and applications relevant to un-
congested networks are Labbé et al. (1998, 1999), Brotcorne et al. (2001),
Marcotte and Savard (2001, 2002), Bouhtou et al. (2003), Côté et al. (2003).
A linearization algorithm applicable to the congested case can be found in
Julsain (1998).
The continuous variant of the network design problem considered in Sec-
tion 6.4 was first proposed in Abdulaal and Leblanc (1979). Heuristics were
analyzed from the theoretical and computational perspectives in Marcotte
(1986b) and Marcotte and Marquis (1992). The numerical results displayed
in Figure 10 and Table 8 demonstrate that Frank–Wolfe type methods are
to recommend for use neither in the solution of the basic traffic equilibrium
702 P. Marcotte and M. Patriksson

model nor for more complex models where accurate equilibrium solutions
are needed as inputs. The numerical results are taken from Patriksson and
Josefsson (2003), Patriksson (2006). In Table 8, columns five and six appeared
in Suwansirikul et al. (1987), column seven in Friesz et al. (1992), columns
eight–ten in Huang and Bell (1998), and column eleven in Lim (2002). The
remarks made at the end of the section refer to application work performed
in Larsson et al. (2001), Boyce et al. (2002).

Appendix A: A primer on variational inequalities

A.1 Definition and formulations

Equilibrium problems naturally fit the variational inequality framework,

which constitutes an extension of equation systems, in contrast with optimiza-
tion problems. A variational inequality problem VIP(f X) in n is characterized
by a nonempty, closed and convex set X ⊆ n and a mapping f from n
into n . A vector x ∈ X is a solution of VIP(f X) if and only if it satisfies
f(x) (y − x)  0 y ∈ X (93)
If one denotes by NX (x) the normal cone to the set X at the point x ∈ X,
defined as
  
NX (x) = p ∈ n p (y − x)  0 y ∈ X  (94)
then x is a solution to VIP(f X) if and only if it satisfies the generalized equation
(or, normal cone inclusion)
−f(x) ∈ NX (x) (95)
If f is a gradient mapping, that is, there exists a real function φ such that
f ≡ ∇φ, then solving VIP(f X) reduces to finding a point that satisfies the
first-order optimality condition of the mathematical program
minimize φ(x) (96)
x∈X
If X = {x ∈ n | gi (x)  0 i = 1     m} and the functions gi , i = 1     m,
are continuously differentiable, then, under suitable constraint qualifications
(CQs), any solution x to VIP(f X), together with a vector of multipliers λ =
(λ1      λm ), satisfies the Karush–Kuhn–Tucker system
gi (x)  0m  (97a)

m
f(x) + λi ∇gi (x) = 0n  (97b)
i=1
λi gi (x) = 0 i = 1     m (97c)
λi  0 i = 1     m (97d)
 Ch. 10. Traffic Equilibrium 703

If X = n+ , the Karush–Kuhn–Tucker system reduces to the nonlinear comple-

mentarity problem

x  0n  f(x)  0n  x f(x) = 0 (98)

or, using a more ‘geometric’ shorthand notation:

0n  x ⊥ f(x)  0n  (99)
If X = n+ , the KKT system (97) can still be formulated as a nonlinear com-
plementarity problem, through a suitable redefinition of the primal and dual
variables. It follows that both formulations are equivalent.
Variational inequalities can also be formulated as fixed point problems. Two
such examples, involving either the projection operator ProjX over the set X,
or a linear program defined over X, are
 
x = ProjX x − αf(x)  α > 0; (100)
x ∈ arg minimum f(x) (x − y) (101)
y∈X

A.2 Monotonicity properties

A mapping f is monotone over X if

f(x) − f(y) (x − y)  0 x y ∈ X (102)

It is strictly monotone over X if the above inequality is strict whenever x = y
and strongly monotone if there exists a positive number β such that

f(x) − f(y) (x − y)  βx − y2  x y ∈ X (103)

If f is continuously differentiable, it is monotone over X if and only if its Jaco-
bian ∇f(x) is positive semidefinite over X, that is, (y − x) ∇f(x)(y − x)  0
holds for every pair x y ∈ X.27 Monotonicity is to variational inequality prob-
lems what convexity is to optimization. Indeed, if f is the gradient of φ, then f is
(strictly, strongly) monotone if and only if φ is (strictly, strongly) convex. Other
related concepts of interest used in this chapter are:
pseudomonotonicity:

f(x)T (y − x)  0 ⇒ f(y)T (y − x)  0 x y ∈ X;

monotonicity+ :

f(x) − f(y) (x − y) = 0 ⇒ f(x) = f(y)

27 Beware: this is not equivalent to ∇f(x) being positive semidefinite for every x ∈ X.
704 P. Marcotte and M. Patriksson

A mapping that is monotone (respectively pseudomonotone) and satisfies the

additional condition that

f(x) (x − y) = f(y) (x − y) = 0 ⇒ f(x) = f(y)

is monotone+ + +
∗ (respectively pseudomonotone∗ ). Pseudomonotonicity∗ ensures
that the value of f is constant over the solution set of VIP(f X). Note that the
stronger monotone+ condition is satisfied by convex gradient mappings.
If f is pseudomonotone and continuous, an equivalent formulation of the
variational inequality (93) is the Minty (or dual) variational inequality

f(y) (y − x)  0 x ∈ X (104)

A.3 Existence and uniqueness

Assume that the mapping f is continuous. If the set X is compact, the so-
lution set X ∗ of VIP(f X) is nonempty (and compact). If X is closed but
unbounded, coercivity of f, that is, the existence of a point x0 ∈ X such that
f(x)
lim = ∞
x→∞x∈X x − x0 
yields the same conclusion. Uniqueness of the solution usually requires that
the mapping f be strictly monotone over X.

A.4 Algorithms

The fixed point formulation (100) suggests the iteration

  
xτ+1 = ProjX xτ − αf xτ  τ = 1 2     (105)
which converges if f is strongly monotone on X and α is sufficiently small,
although it is possible to modify this basic scheme and achieve convergence
under weaker assumptions. Although similar strategy is not directly applicable
to the fixed point formulation (101), it is yet possible to construct an algorithm
that finds a solution of VIP(f X) by minimizing the gap function

gap(x) = maximum f(x) (x − y) (106)

y∈X

While the above function is in general neither convex nor differentiable, one
can show that all of its stationary points are solutions to VIP(f X), provided
that f is monotone on X. The descent direction, built around the solutions of
the LP problem to

minimize f(x) y
y∈X
 Ch. 10. Traffic Equilibrium 705

can, for traffic equilibrium models, be obtained simply by computing shortest

route trees rooted at each origin (or destination) node of the network. Alter-
natively, under pseudomonotonicity, one can minimize the dual gap function

dgap(x) = maximum f(y) (x − y) (107)

y∈X

The resulting convex program can be expressed as a semi-infinite linear pro-

gram and solved using cutting-plane methods.
A general framework for solving variational inequality problems consists
in iteratively solving an approximation of the original problem (93), of a sim-
pler form. One sets xτ+1 to the solution of VIP(f̃(xτ  ·) X), where f̃ satisfies
the condition that f̃(x x) = f(x) for all x in X. Under relevant choices of
the approximation mapping f̃, one recovers the Jacobi, Gauss–Seidel, New-
ton, quasi-Newton and projection algorithms, and many others. In particular,
if ψ is a strongly convex function, the approximation

f̃(x y) = f(x) + ∇ψ(y) − ∇ψ(x) x y ∈ n  (108)

suggests the continuously differentiable merit function

φ(x) := maximum f(x) − ∇ψ(x) (x − y) − ψ(y) x ∈ X (109)

y∈X

Under suitable monotonicity conditions on f, the solution y(x) to (109) yields

a feasible descent direction p := y(x) − x for φ, around which a convergent
algorithm can be designed. In particular, if one sets ψ(x) = α2 x2 , the solution
to (109) is precisely the projection ProjX (x − αf(x)).

A.5 Bibliographical notes

A classical reference on variational inequality problems, mostly in the

infinite-dimensional setting, is Glowinski et al. (1981). Variational inequal-
ities were introduced by Hartman and Stampacchia (1966). Several results
pertaining to finite-dimensional variational inequalities appear in the mono-
graph Auslender (1976). A more recent and comprehensive treatment is the
twin volume Facchinei and Pang (2003a, 2003b). Algorithms based on cost ap-
proximations, of which the projection method is the quintessential example,
have been analyzed in Cohen (1988), Larsson and Patriksson (1994a), Patriks-
son (1994a, 1994b, 1998) and several others.
706 P. Marcotte and M. Patriksson

Appendix B: A summary of key notations

h = (hr ): vector of flows on routes indexed by r

c(h) = (cr (h)): vector of route travel costs
d = (dk ): vector of demands for commodity (OD pair) k
π = (πpq ): vector of travel costs for origin–destination couples (p q)
g(π): demand function
: route-OD incidence matrix
v = (vl ): vector of flows along links indexed by l
t(v) = (tl (v)): vector of traversal costs on links indexed by l
: link–route incidence matrix
E: link–node incidence matrix
wk : vector of OD flows
ξ(d): inverse demand function: ξ = g−1
Hd : set of route flows and demand vectors (variable demand
case)
H: set of route flows and demand vectors (fixed demand case)
d :
H set of link flows and demand vectors (link–route
representation)

F: set of link flows (link–route representation, fixed demand)
F: set of link flows (link–node representation, fixed demand)
d :
F set of link flows (link–route representation, variable
demand)
Fd : set of link flows (link–node representation, variable
demand)
f(v): vector of out-of-pocket travel costs along links
NX (x): normal cone to a convex set X at a point x
φl : frequency on transit line l
ϕ(ρ): cost vector associated with a link-improvement vector ρ
c(x): vector of link costs associated with strategic flow vector x
Ejs : preference order of strategy s at node j
s :
πjk probability of accessing node k from node j, when using
strategy s

References

Aashtiani, H.Z., Magnanti, T.L. (1981). Equilibria on a congested transportation network. SIAM Jour-
nal on Algebraic and Discrete Methods 2, 213–226.
Abdulaal, M., Leblanc, L.J. (1979). Continuous equilibrium network design models. Transportation Re-
search B 13, 19–32.
Ahuja, R.K., Orlin, J.B. (2001). Inverse optimization. Operations Research 49, 771–783.
Ahuja, R.K., Magnanti, T.L., Orlin, J.B. (1993). Network Flows: Theory, Algorithms, and Applications.
Prentice Hall, Englewood Cliffs, NJ.
Akamatsu, T. (1996). Cyclic flows, Markov process and stochastic traffic assignment. Transportation
Research B 30, 369–386.
 Ch. 10. Traffic Equilibrium 707

Altman, E., El Azouzi, R., Abramov, V. (2002). Non-cooperative routing in loss networks. Performance
Evaluation 49, 43–55.
Auslender, A. (1976). Optimisation: Méthodes Numériques. Masson, Paris.
Baillon, J.B. (1975). Un théorème de type ergodique pour les contractions non linéaires dans un
espace de Hilbert. Comptes Rendus Hebdomadaires des Séances de l’Académie des Sciences (Paris),
Série A 280, A1511–A1514.
Bard, J.F. (1998). Practical Bilevel Optimization: Algorithms and Applications. Kluwer Academic, Boston,
MA.
Bar-Gera, H. (2002a). Origin-based algorithm for the traffic assignment problem. Transportation Sci-
ence 36, 398–417.
Bar-Gera, H. (2002b). Origin-based network assignment. In: Patriksson, M., Labbé, M. (Eds.), Trans-
portation Planning: State of the Art. Kluwer Academic, Dordrecht, The Netherlands, pp. 1–17.
Beckmann, M., McGuire, C.B., Winsten, C.B. (1956). Studies in the Economics of Transportation. Yale
Univ. Press, New Haven, CT.
Bennett, L.D. (1993). The existence of equivalent mathematical programs for certain mixed equilibrium
traffic assignment problems. European Journal of Operational Research 71, 177–187.
Bergendorff, P. (1995). The bounded flow approach to congestion pricing. Division of Optimization and
Systems Theory, Department of Mathematics, Royal Institute of Technology, Stockholm, Sweden.
Bergendorff, P., Hearn, D.W., Ramana, M.V. (1997). Congestion toll pricing of traffic networks. In:
Pardalos, P.M., Hearn, D.W., Hager, W.W. (Eds.), Network Optimization. Proceedings of the Network
Optimization Conference, University of Florida, Gainesville, FL, February 12–14, 1996. Lecture
Notes in Economics and Mathematical Systems, vol. 450. Springer-Verlag, Berlin, pp. 51–71.
Bernstein, D., Smith, T.E. (1994). Equilibria for networks with lower semicontinuous costs: With an
application to congestion pricing. Transportation Science 28, 221–235.
Bernstein, D., Wynter, L. (2000). Issues of uniqueness and convexity in non-additive bi-criteria equi-
librium models. Conference paper, 8th Meeting of the EURO Working Group on Transportation,
Rome Jubilee 2000 Conference, La Sapienza, Rome, Italy, September 11–14.
Bertsekas, D.P. (1976). On the Goldstein–Levitin–Polyak gradient projection method. IEEE Transac-
tions on Automatic Control 21, 174–184.
Bertsekas, D.P. (1998). Network Optimization: Continuous and Discrete Models. Athena Scientific, Bel-
mont, MA.
Bertsekas, D.P., Gafni, E.M. (1982). Projection methods for variational inequalities with application to
the traffic assignment problem. Mathematical Programming Study 17, 139–159.
Bertsekas, D.P., Tsitsiklis, J.N. (1989). Parallel and Distributed Computation: Numerical Methods. Pren-
tice Hall, London.
Best, M.J., Griffin V.J. (1975). Optimisation of nonlinear performance criteria subject to flow con-
straints. Conference paper presented at the ORSA/TIMS National Meeting, Chicago, IL, USA, 30
April–2 May.
Bouhtou, M., van Hoesel, S., van der Kraaij, A.F., Lutton, J.-L. (2003). Tariff optimization in networks.
Report RM03011, Maastricht Economic Research School on Technology and Organisation.
Boyce, D. (2004). Forecasting travel on congested urban transportation networks: Review and prospects
for network equilibrium models. Conference paper presented at TRISTAN V, The Fifth Triennial
Symposium on Transportation Analysis, Le Gosier, Guadeloupe, June 13–18.
Boyce, D.E., Ralevic-Dekic, B., Bar-Gera, H. (2002). Convergence of traffic assignments: How much is
enough? The Delaware Valley Region Case Study. Paper presented at the 16th Annual International
EMME/2 Users’ Group Conference, Albuquerque, NM, March.
Brotcorne, L., Labbé, M., Marcotte, P., Savard, G. (2001). A bilevel model for toll optimization on a
multicommodity transportation network. Transportation Science 35, 1–14.
Bruck, R.E. Jr (1975). Asymptotic convergence of nonlinear contraction semigroups in Hilbert space.
Journal of Functional Analysis 18, 15–26.
Bruck, R.E. Jr (1977). On the weak convergence of an ergodic iteration for the solution of variational
inequalities for monotone operators in Hilbert space. Journal of Mathematical Analysis and Appli-
cations 61, 159–164.
708 P. Marcotte and M. Patriksson

Brucker, P. (1984). An O(n) algorithm for quadratic knapsack problems. Operations Research Letters 3,
163–166.
Busacker, R.G., Saaty, T.L. (1965). Finite Graphs and Networks: An Introduction with Applications.
McGraw-Hill, New York.
Cascetta, E. (2001). Transportation Systems Engineering: Theory and Methods. Applied Optimization,
vol. 49. Kluwer Academic, Dordrecht, The Netherlands.
Chau, C.K., Sim, K.M. (2003). The price of anarchy for non-atomic congestion games with symmetric
cost maps and elastic demands. Operations Research Letters 31, 327–334.
Chen, A., Lee, D.-H., Jayakrishnan, R. (2002). Computational study of state-of-the-art path-based traf-
fic assignment algorithms. Mathematics and Computers in Simulation 2060, 1–10.
Chen, M., Bernstein, D. (2003). Solving the toll design problem with multiple user groups. Transporta-
tion Research B 37, 1–19.
Cohen, G. (1988). Auxiliary problem principle extended to variational inequalities. Journal of Optimiza-
tion Theory and Applications 59, 325–333.
Cole, R., Dodis, Y., Roughgarden, T. (2003). Pricing network edges for heterogeneous selfish users. In:
Proceedings of the 35th Annual ACM Symposium on the Theory of Computing, 2003, pp. 521–530.
Correa, J.R., Schulz, A.S., Stier-Moses, N.E. (2004). Selfish routing in capacitated networks. Mathemat-
ics of Operations Research 29, 961–976.
Côté, J.-P., Marcotte, P., Savard, G. (2003). A bilevel modeling approach to pricing and fare optimiza-
tion in the airline industry. Journal of Revenue and Pricing Management 2, 23–36.
Crouzeix, J.-P., Marcotte, P., Zhu, D. (2000). Conditions ensuring the applicability of cutting-plane
methods for solving variational inequalities. Mathematical Programming 88, 521–539.
Dafermos, S.C. (1971). An extended traffic assignment model with applications to two-way traffic.
Transportation Science 5, 366–389.
Dafermos, S.C. (1980). Traffic equilibrium and variational inequalities. Transportation Science 14, 42–
54.
Dafermos, S.C. (1981). A multicriteria route-mode choice traffic equilibrium model. Unpublished man-
uscript, Lefschetz Center for Dynamical Systems, June.
Dafermos, S.C. (1982). The general multimodal network equilibrium problem with elastic demand.
Networks 12, 57–72.
Dafermos, S.C., Sparrow, F.T. (1969). The traffic assignment problem for a general network. Journal of
Research of the National Bureau of Standards 73B, 91–118.
de Palma, A., Nesterov, Y. (1998). Optimization formulations and static equilibrium in congested
transportation networks. CORE Discussion Paper 9861, CORE, Université Catholique de Louvain,
Louvain-la-Neuve, Belgium.
Dial, R.B. (1979). A model and algorithm for multicriteria route-mode choice. Transportation Re-
search B 13, 311–316.
Dial, R.B. (1996a). Bicriterion traffic assignment: Basic theory and elementary algorithms. Transporta-
tion Science 30, 93–111.
Dial, R.B. (1996b). Bicriterion traffic assignment: Efficient algorithms plus examples. Transportation
Research B 31, 357–359.
Dial, R.B. (1999). Minimal-revenue congestion pricing part I: A fast algorithm for the single-origin
case. Transportation Research B 33, 189–202.
Dial, R.B. (2000). Minimal-revenue congestion pricing part II: An efficient algorithm for the general
case. Transportation Research B 34, 645–665.
Dontchev, A.L., Rockafellar, R.T. (2001). Ample parameterization of variational inclusions. SIAM Jour-
nal on Optimization 12, 170–187.
Dunn, J.C. (1973). On recursive averaging processes and Hilbert space extensions of the contraction
mapping principle. Journal of the Franklin Institute 295, 117–133.
Escudero, L.F. (1986). A motivation for using the truncated Newton approach in a very large scale
nonlinear network problem. Mathematical Programming Study 26, 240–244.
Evans, S.P. (1976). Derivation and analysis of some models for combining trip distribution and assign-
ment. Transportation Research 10, 37–57.
 Ch. 10. Traffic Equilibrium 709

Facchinei, F., Pang, J.-S. (2003a). Finite-Dimensional Variational Inequalities and Complementarity Prob-
lems. Springer Series in Operations Research and Financial Engineering, vol. I. Springer-Verlag, Berlin.
Facchinei, F., Pang, J.-S. (2003b). Finite-Dimensional Variational Inequalities and Complementarity Prob-
lems. Springer Series in Operations Research and Financial Engineering, vol. II. Springer-Verlag,
Berlin.
Ferrari, P. (1995). Road pricing and network equilibrium. Transportation Research B 29, 357–372.
Ferris, M.C., Pang, J.-S. (1997). Engineering and economic applications of complementarity problems.
SIAM Review 39, 669–713.
Fisk, C.S. (1980). Some developments in equilibrium traffic assignment. Transportation Research B 14,
243–255.
Florian, M. (1977). A traffic equilibrium model of travel by car and public transit modes. Transportation
Science 11, 166–179.
Florian, M. (1979). Asymmetrical variable demand multi-mode traffic equilibrium problems: Existence
and uniqueness of solutions and a solution algorithm. Publication 347, Département d’Informatique
et de Recherche Opérationnelle, Université de Montréal, Montréal, Canada.
Florian, M., Hearn, D.W. (1995). Network equilibrium models and algorithms. In: Ball, M.O., Mag-
nanti, T.L., Monma, C.L., Nemhauser, G.L. (Eds.), Network Routing. Handbooks in Operations
Research and Management Science, vol. 8. North-Holland, Amsterdam, pp. 485–550, Chapter 6.
Florian, M., Wu, J.H., He, S. (2002). Multi-mode variable demand network equilibrium model with
hierarchical logit structure. In: Gendreau, M., Marcotte, P. (Eds.), Transportation and Network
Analysis: Current Trends. Kluwer Academic, Dordrecht, The Netherlands, pp. 119–133.
Frank, M., Wolfe, P. (1956). An algorithm for quadratic programming. Naval Research Logistics Quar-
terly 3, 95–110.
Friesz, T.L., Cho, H.-J., Metha, N.J., Tobin, R.L., Anandalingam, G. (1992). A simulated annealing
approach to the network design problem with variational inequality constraints. Transportation Sci-
ence 26, 18–26.
Gabriel, S.A., Bernstein, D. (1997). The traffic equilibrium problem with non-additive path costs. Trans-
portation Science 31, 337–348.
Gafni, E.M., Bertsekas, D.P. (1984). Two-metric projection methods for constrained optimization.
SIAM Journal on Control and Optimization 22, 936–964.
Gallager, R.G. (1977). Loops in multicommodity flows. In: Proceedings of the 10th IEEE Conference on
Decision and Control, New Orleans, TX, pp. 819–825.
Gartner, N.H. (1980). Optimal traffic assignment with elastic demands: A review. Part II: Algorithmic
approaches. Transportation Science 14, 192–208.
Glowinski, R., Lions, J.-L., Trémolières, R. (1981). Numerical Analysis of Variational Inequalities. North-
Holland, Amsterdam.
Goffin, J.-L., Marcotte, P., Zhu, D.L. (1997). An analytic center cutting-plane method for
pseudomonotone variational inequalities. Operations Research Letters 20, 1–6.
Hagstrom, J.N., Tseng, P. (1998). Traffic equilibrium: Link flows, path flows and weakly/strongly acyclic
solutions. Technical Report M/C 294, Department of Information and Decision Sciences, University
of Illinois, Chicago, IL. Available at http://www.math.washington.edu/~tseng/papers.html.
Hamdouch, Y., Marcotte, P., Nguyen, S. (2004a). Capacitated transit assignment with loading priorities.
Mathematical Programming B 101, 205–230.
Hamdouch, Y., Marcotte, P., Nguyen, S. (2004b). A strategic model for dynamic traffic assignment.
Networks and Spatial Economics 4, 291–315.
Hammond, J.H. (1984). Solving asymmetric variational inequality problems and systems of equations
with generalized nonlinear programming algorithms. PhD thesis, Department of Mathematics,
Massachusetts Institute of Technology, Cambridge, MA.
Hartman, P., Stampacchia, G. (1966). On some non-linear elliptic differential-functional equations.
Acta Mathematica 115, 271–310.
Hearn, D.W. (1980). Bounding flows in traffic assignment models. Research Report 80-4, Department
of Industrial and Systems Engineering, University of Florida, Gainesville, FL.
Hearn, D.W., Ramana, M.V. (1998). Solving congestion toll pricing models. In: Marcotte, P., Nguyen,
S. (Eds.), Equilibrium and Advanced Transportation Modelling. Kluwer Academic, Boston, MA,
pp. 109–124.
710 P. Marcotte and M. Patriksson

Hearn, D.W., Lawphongpanich, S., Ventura, J.A. (1987). Restricted simplicial decomposition: Compu-
tation and extensions. Mathematical Programming Study 31, 99–118.
Heydecker, B.G. (1986). On the definition of traffic equilibrium. Transportation Research B 20, 435–440.
Huang, H.-J., Bell, M.G.H. (1998). Continuous equilibrium network design problem with elastic de-
mand: Derivative-free solution methods. In: Bell, M.G.H. (Ed.), Transportation Networks: Recent
Methodological Advances. Pergamon Press, Amsterdam, pp. 175–183.
Iusem, A.N. (1998). On some properties of paramonotone operators. Journal of Convex Analysis 5,
269–278.
Jayakrishnan, R., Tsai, W.K., Prashker, J.N., Rajadhyaksha, S. (1994). Faster path-based algorithm for
traffic assignment. Transportation Research Record 1443, 75–83.
Jorgensen, N.O. (1963). Some aspects of the urban traffic assignment problem. Master’s thesis, Institute
of Transportation and Traffic Engineering, University of California, Berkeley, CA.
Josefsson, M., Patriksson, M. (2005). On the applicability of sensitivity analysis formulas for traffic
equilibrium models. In: Hearn, D., Lawphongpanich, S., Smith, M. (Eds.), Mathematical and Com-
putational Methods for Congestion Charging. Proceedings of the Theory and Practice of Congestion
Charging Symposium, Imperial College, London, August 18–20, 2003. The Applied Optimization
Series, vol. 148. Springer-Verlag, Berlin, pp. 117–141.
Julsain, H. (1998). Tarification dans les réseaux de télécommunications: une approche par programma-
tion mathématique à deux niveaux. Master’s thesis, École Polytechnique, Montréal.
Kennington, J.L., Helgason, R.V. (1980). Algorithms for Network Programming. Wiley, New York.
Klincewicz, J.G. (1983). A Newton method for convex separable network flow problems. Networks 13,
427–442.
Knight, F.H. (1924). Some fallacies in the interpretation of social cost. Quarterly Journal of Eco-
nomics 38, 582–606.
Konnov, I.V. (2001). Combined Relaxation Methods for Variational Inequalities. Lecture Notes in Eco-
nomics and Mathematical Systems, vol. 495. Springer-Verlag, Berlin.
Korpelevich, G.M. (1977). The extragradient method for finding saddle points and other problems.
Matecon 13, 35–49.
Kyparisis, J. (1990). Solution differentiability for variational inequalities. Mathematical Programming 48,
285–301.
Labbé, M., Marcotte, P., Savard, G. (1998). A bilevel model of taxation and its application to optimal
highway pricing. Management Science 44, 1595–1607.
Labbé, M., Marcotte, P., Savard, G. (1999). On a class of bilevel programs. In: Di Pillo, G., Giannessi, F.
(Eds.), Nonlinear Optimization and Related Topics. Kluwer Academic, Dordrecht, The Netherlands,
pp. 183–206.
Larsson, T., Patriksson, M. (1992). Simplicial decomposition with disaggregated representation for the
traffic assignment problem. Transportation Science 26, 4–17.
Larsson, T., Patriksson, M. (1994a). A class of gap functions for variational inequalities. Mathematical
Programming 64, 53–79.
Larsson, T., Patriksson, M. (1994b). Equilibrium characterizations of solutions to side constrained
asymmetric traffic assignment models. Le Matematiche 49, 249–280.
Larsson, T., Patriksson, M. (1995). An Augmented Lagrangean dual algorithm for link capacity side
constrained traffic assignment problems. Transportation Research B 29, 433–455.
Larsson, T., Patriksson, M. (1997). Price-directive traffic management – an approach utilizing side
constrained traffic equilibrium models. Rendiconti del Circolo Matematico di Palermo, Serie II 48,
147–170.
Larsson, T., Patriksson, M. (1998). Side constrained traffic equilibrium models – traffic management
through link tolls. In: Marcotte, P., Nguyen, S. (Eds.), Equilibrium and Advanced Transportation
Modelling. Kluwer Academic, New York, pp. 125–151.
Larsson, T., Patriksson, M. (1999). Side constrained traffic equilibrium models: Analysis, computation
and applications. Transportation Research B 33, 233–264.
Larsson, T., Liu, Z.-W., Patriksson, M. (1997). A dual scheme for traffic assignment problems. Optimiza-
tion 42, 323–358.
 Ch. 10. Traffic Equilibrium 711

Larsson, T., Patriksson, M., Strömberg, A.-B. (1999). Ergodic, primal convergence in dual subgradient
schemes for convex programming. Mathematical Programming 86, 283–312.
Larsson, T., Lundgren, J., Patriksson, M., Rydergren, C. (2001). Most likely traffic equilibrium route
flows: Analysis and computation. In: Giannessi, F., Maugeri, A., Pardalos, P.M. (Eds.), Equilibrium
Problems: Nonsmooth Optimization and Variational Inequality Models. Proceedings of an Interna-
tional Workshop in Memory of Marino De Luca, Hotel Villa Diodoro, Taormina, Italy, Decem-
ber 3–5, 1998. Kluwer Academic, Dordrecht, The Netherlands, pp. 129–159.
Larsson, T., Lindberg, P.-O., Lundgren, J., Patriksson, M., Rydergren, C. (2002). On traffic equilibrium
models with a nonlinear time/money relation. In: Patriksson, M., Labbé, M. (Eds.), Transportation
Planning: State of the Art. Kluwer Academic, Dordrecht, The Netherlands, pp. 19–31.
Larsson, T., Patriksson, M., Rydergren, C. (2004). A column generation procedure for the side con-
strained traffic equilibrium problem. Transportation Research B 38, 17–38.
Lawphongpanich, S., Hearn, D.W. (1984). Simplicial decomposition of the asymmetric traffic assign-
ment problem. Transportation Research B 18, 123–133.
LeBlanc, L.J., Morlok, E.K., Pierskalla, W.P. (1975). An efficient approach to solving the road network
equilibrium traffic assignment problem. Transportation Research 9, 308–318.
Leurent, F. (1993). Cost versus time equilibrium over a network. European Journal of Operational Re-
search 71, 205–221.
Leventhal, T., Nemhauser, G., Trotter, L. Jr. (1973). A column generation algorithm for optimal traffic
assignment. Transportation Science 7, 168–176.
Lim, A.C. (2002). Transportation network design problems: An MPEC approach. PhD thesis, Depart-
ment of Mathematical Sciences, The Johns Hopkins University Baltimore, MD.
Luo, Z.-Q., Pang, J.-S., Ralph, D. (1996). Mathematical Programs with Equilibrium Constraints. Cam-
bridge Univ. Press, Cambridge, UK.
Marcotte, P. (1986a). A new algorithm for solving variational inequalities, with application to the traffic
assignment problem. Mathematical Programming 33, 339–351.
Marcotte, P. (1986b). Network design problem with congestion effects: A case of bilevel programming.
Mathematical Programming 34, 142–162.
Marcotte, P. (1991). Application of Khobotov’s algorithm to variational inequalities and network equi-
librium problems. INFOR 29, 258–270.
Marcotte, P. (1998). Reformulations of a bicriterion equilibrium problem. In: Fukushima, M., Qi, L.
(Eds.), Reformulation: Nonsmooth, Piecewise Smooth, Semismooth and Smoothing Methods. Kluwer
Academic, Dordrecht, The Netherlands, pp. 269–292.
Marcotte, P., Guélat, J. (1988). Adaptation of a modified Newton method for solving the asymmetric
traffic equilibrium problem. Transportation Science 22, 112–124.
Marcotte, P., Marquis, G. (1992). Efficient implementation of heuristics for the continuous network
design problem. Annals of Operations Research 34, 163–176.
Marcotte, P., Nguyen, S. (1998). Hyperpath formulations of traffic assignment problems. In: Marcotte,
P., Nguyen, S. (Eds.), Equilibrium and Advanced Transportation Modelling. Kluwer Academic, Dor-
drecht, The Netherlands, pp. 175–200.
Marcotte, P., Savard, G. (2001). Bilevel programming: Formulation, applications, algorithms. In:
Floudas, C.A., Pardalos, P.M. (Eds.), The Encyclopedia of Optimization. Kluwer Academic, Dor-
drecht, The Netherlands, pp. 155–159.
Marcotte, P., Savard, G. (2002). A bilevel programming approach to price setting. In: Zaccour, G. (Ed.),
Decision and Control in Management Science. Essays in Honor of Alain Haurie. Kluwer Academic,
Dordrecht, pp. 97–117.
Marcotte, P., Wynter, L. (2004). A new look at the multi-class network equilibrium problem. Transporta-
tion Science 38, 282–292.
Marcotte, P., Zhu, D.L. (1997). Equilibria with infinitely many differentiated classes of customers. In:
Pang, J.-S., Ferris, M. (Eds.), Complementarity and Variational Problems: State of the Art. SIAM,
Philadelphia, pp. 234–258.
Marcotte, P., Nguyen, S., Tanguay, K. (1996). Implementation of an efficient algorithm for the multiclass
traffic assignment problem. In: Lesort, J.-B. (Ed.), Transportation and Traffic Theory. Proceedings of
the 13th International Symposium on Transportation and Traffic Theory, Lyon, France, 24–26 July,
1996. Elsevier, Tarrytown, NY, pp. 217–236.
712 P. Marcotte and M. Patriksson

Marcotte, P., Nguyen, S., Schoeb, A. (2004). A strategic model of traffic assignment in static capacitated
networks. Operations Research 52, 191–212.
Miller, S.D., Payne, H.J., Thompson, W.A. (1975). An algorithm for traffic assignment on capacity con-
strained transportation networks with queues. Paper presented at the Johns Hopkins Conference
on Information Sciences and Systems, The Johns Hopkins University, Baltimore, MD, April 2–4.
Murchland, J.D. (1970). Road network traffic distribution in equilibrium. In: Henn, R., Künzi, H.P.,
Schubert, H. (Eds.), Mathematical Models in the Social Sciences, vol. 8. II Oberwolfach-Tagungüber
Operations Research, Mathematisches Forschungsinstitut, Oberwolfach, 20–25 October, 1969. An-
ton Hain Verlag, Meisenheim am Glan, pp. 145–183 (in German). Translation by H.A. Paul.
Nagurney, A., Dong, J. (2002). Supernetworks: Decision-Making for the Information Age. Edward Elgar,
Cheltenham, UK.
Nesterov, Y. (2000). Stable traffic equilibria, properties and applications. Optimization and Engineer-
ing 3, 29–50.
Nguyen, S. (1974). An algorithm for the traffic assignment problem. Transportation Science 8, 203–216.
Nguyen, S., Dupuis, C. (1984). An efficient method for computing traffic equilibria in networks with
asymmetric transportation costs. Transportation Science 18, 185–202.
Nguyen, S., Pallottino, S. (1988). Equilibrium traffic assignment for large scale transit networks. Euro-
pean Journal of Operational Research 37, 176–186.
Nguyen, S., Pallottino, S., Malucelli, F. (2001). A modeling framework for the passenger assignment on
a transport network with time-tables. Transportation Science 35, 238–249.
Nguyen, S., Pallottino, S., Scutellá, M.G. (2002). A new dual algorithm for shortest path reoptimization.
In: Gendreau, M., Marcotte, P. (Eds.), Transportation and Network Analysis: Current Trends. Kluwer
Academic, New York, pp. 221–235.
Pallottino, S., Scutellà, M.G. (2003). A new algorithm for reoptimizing shortest paths when the arc costs
change. Operations Research Letters 31, 149–160.
Patriksson, M. (1994a). On the convergence of descent methods for monotone variational inequalities.
Operations Research Letters 16, 265–269.
Patriksson, M. (1994b). The Traffic Assignment Problem: Models and Methods. Topics in Transportation.
VSP BV, Utrecht, The Netherlands.
Patriksson, M. (1998). Nonlinear Programming and Variational Inequalities: A Unified Approach. Kluwer
Academic, Dordrecht, The Netherlands.
Patriksson, M. (2004). Sensitivity analysis of traffic equilibria. Transportation Science 38, 258–281.
Patriksson, M. (2006). Traffic Equilibrium Models: Analysis, Applications, and Optimization Algorithms.
Springer-Verlag, in preparation.
Patriksson, M., Josefsson, M. (2003). Sensitivity analysis of separable traffic equilibrium solutions,
with application to bilevel optimization in network design. Report, Department of Mathematics,
Chalmers University of Technology, Gothenburg, Sweden. Transportation Research B, in press.
Patriksson, M., Rockafellar, R.T. (2002). A mathematical model and descent algorithm for bilevel traffic
management. Transportation Science 36, 271–291.
Patriksson, M., Rockafellar, R.T. (2003). Sensitivity analysis of variational inequalities over aggregated
polyhedra, with application to traffic equilibria. Transportation Science 37, 56–68.
Payne, H.J., Thompson, W.A. (1975). Traffic assignment on transportation networks with capacity
constraints and queueing. Paper presented at the 47th National ORSA Meeting/TIMS 1975 North-
American Meeting, Chicago, IL, April 30–May 2.
Robinson, S.M. (1980). Strongly regular generalized equations. Mathematics of Operations Research 5,
43–62.
Robinson, S.M. (1985). Implicit B-differentiability in generalized equations. Technical Summary Re-
port 2854, Mathematics Research Center, University of Wisconsin at Madison, Madison, WI.
Robinson, S.M. (1991). An implicit-function theorem for a class of nonsmooth functions. Mathematics
of Operations Research 16, 292–309.
Rockafellar, R.T. (1976). Monotone operators and the proximal point algorithm. SIAM Journal on Con-
trol and Optimization 14, 877–898.
Rockafellar, R.T. (1984). Network Flows and Monotropic Optimization. Wiley, New York. Also published
by Athena Scientific, Belmont, MA, 1998.
 Ch. 10. Traffic Equilibrium 713

Roughgarden, T., Tardos, É. (2002). How bad is selfish routing? Journal of the ACM 49, 236–259.
Sheffi, Y. (1985). Urban Transportation Networks: Equilibrium Analysis with Mathematical Programming
Methods. Prentice Hall, Englewood Cliffs, NJ.
Shimizu, K., Ishizuka, Y., Bard, J.F. (1997). Nondifferentiable and Two-Level Programming. Kluwer Aca-
demic, Boston, MA.
Smith, M.J. (1979). The existence, uniqueness and stability of traffic equilibria. Transportation Re-
search B 13, 295–304.
Smith, M.J. (1984). Two alternative definitions of traffic equilibrium. Transportation Research B 18, 63–
65.
Solodov, M.V. (2003). Convergence rate analysis of iterative algorithms for solving variational inequal-
ity problems. Mathematical Programming 96, 513–528.
Solodov, M.V., Svaiter, B.F. (1999a). A hybrid approximate extragradient–proximal point algorithm
using the enlargement of a maximal monotone operator. Set-Valued Analysis 7, 323–345.
Solodov, M.V., Svaiter, B.F. (1999b). A hybrid projection-proximal point algorithm. Journal of Convex
Analysis 6, 59–70.
Spiess, H., Florian, M. (1989). Optimal strategies: A new assignment model for transit networks. Trans-
portation Research B 23, 83–102.
Suwansirikul, C., Friesz, T.L., Tobin, R.L. (1987). Equilibrium decomposed optimization: A heuristic
for the continuous equilibrium network design problem. Transportation Science 21, 254–260.
Tobin, R.L., Friesz, T.L. (1988). Sensitivity analysis for equilibrium network flow. Transportation Sci-
ence 22, 242–250.
Toint, Ph., Wynter, L. (1996). Asymmetric multiclass traffic assignment: A coherent formulation. In:
Lesort, J.-B. (Ed.), Transportation and Traffic Theory. Proceedings of the 13th International Sym-
posium on Transportation and Traffic Theory, Lyon, France, 24–26 July, 1996. Pergamon Press,
Oxford, pp. 237–260.
Wang, Y.J., Xiu, N.H., Wang, C.Y. (2001). Unified framework of extragradient-type methods for
pseudomonotone variational inequalities. Journal of Optimization Theory and Applications 111, 641–
656.
Watling, D. (1999). Stability of the stochastic equilibrium assignment problem: a dynamical systems
approach. Transportation Research B 33, 281–312.
Yang, H., Huang, H.-J. (2004). The multi-class, multi-criteria traffic network equilibrium and systems
optimum problem. Transportation Research B 38, 1–15.
Zhu, D.L., Marcotte, P. (1996). Co-coercivity and its role in the convergence of iterative schemes for
solving variational inequalities. SIAM Journal on Optimization 6, 714–726.
Zuhovickiı̌, S.I., Polyak, R.A., Primak, M.E. (1969). Two methods of search for equilibrium points of
n-person concave games. Soviet Mathematics Doklady 10, 279–282.
C. Barnhart and G. Laporte (Eds.), Handbook in OR & MS, Vol. 14
Copyright © 2007 Elsevier B.V. All rights reserved
DOI: 10.1016/S0927-0507(06)14011-6

Chapter 11
ITS and Traffic Management
M. Papageorgiou
Dynamic Systems and Simulation Laboratory, Technical University of Crete,
731 00, Chania, Crete, Greece
E-mail: [email protected]

M. Ben-Akiva
Department of Civil and Environmental Engineering, Massachusetts Institute of Technology,
77 Massachusetts Ave., Cambridge, MA 02139, USA
E-mail: [email protected]

J. Bottom
Charles River Associates, Inc., Boston, MA, USA
E-mail: [email protected]

P.H.L. Bovy
Transportation and Traffic Planning Section, Faculty of Civil Engineering and Geosciences,
Delft University of Technology, Delft, The Netherlands
E-mail: [email protected]

S.P. Hoogendoorn
Transportation and Traffic Planning Section, Faculty of Civil Engineering and Geosciences,
Delft University of Technology, Delft, The Netherlands
E-mail: [email protected]

N.B. Hounsell
Transportation Research Group, School of Civil Engineering and the Environment,
University of Southampton, Hants, SO17 1BJ, UK
E-mail: [email protected]

A. Kotsialos
School of Engineering, University of Durham, South Road, Durham, DH1 3LE, UK
E-mail: [email protected]

M. McDonald
Transportation Research Group, School of Civil Engineering and the Environment,
University of Southampton, Hants, SO17 1BJ, UK
E-mail: [email protected]

1 Introduction

The observed traffic conditions on road and highway networks result from
a quite complex-to-describe confrontation of supply and demand. Supply is

715
716 M. Papageorgiou et al.

mainly determined from the available road and highway infrastructure, most
notably its capacity. Demand is the collective outcome of individual driver de-
cisions regarding the effectuation (or not) of a trip, the choice of transportation
mode, of departure time, of the route to be followed, etc. Traffic congestion is
observed in increasing levels on road and highway networks around the world,
with detrimental consequences for traffic efficiency, safety as well as for the
environment. Traffic congestion affects the nominal capacity of the available
infrastructure leading to a vicious cycle of further congestion increase, fur-
ther infrastructure degradation, and so forth. In fact, the traffic throughput
measured in congested road or highway networks is usually well below the
nominal network capacity. Traffic control measures and strategies described
in this chapter aim at maintaining the available infrastructure capacity close
to nominal levels, protecting the traffic networks from the detrimental effects
of oversaturation and even gridlock. In this sense, traffic control is deemed to
mainly act on the supply side of the basic traffic equation. Other operational
measures have been employed in an attempt to reduce congestion by influenc-
ing the manifest traffic demand; this includes various forms of administrative
restrictions or of demand management (road pricing being the most promi-
nent), which, however, are not addressed in this chapter.
The chapter consists of 4 main overview sections, Section 2 presenting an
overview of traffic flow modeling advancements, Section 3 addressing the issue
of route guidance and information systems, while Sections 4 and 5 are con-
cerned with specific road and motorway network control systems, respectively.

2 Traffic flow modeling

2.1 Traffic flow modeling approaches

Understanding traffic flow characteristics (e.g., headway distributions, rela-

tion between density and speed, capacity distributions) and knowledge of the
associated analytical tools (e.g., queuing models, shockwave theory, simulation
models) to predict the dynamics of these characteristics under given demand,
supply and control conditions, is an essential requirement for the planning,
design and operation of a transportation system. For the analysis of a simple
arterial, on-ramp, or merge area, as well as for studying traffic flow operations
in urban or motorway networks, being able to predict the traffic performance
is an essential factor in the analysis of the system.
Traffic flow modeling research started when Lighthill and Whitham (1955)
presented their seminal paper on the wave dynamics of traffic flow. Their work
was based on the analogy of vehicles in traffic flow and particles in a fluid. Since
then, the mathematical description of traffic flow has been a lively subject of
research and debate for traffic engineers. This has resulted in a broad scope of
low-end and high-end models.
 Ch. 11. ITS and Traffic Management 717

Before giving an overview of these models, we need to emphasize that

it is highly unlikely that traffic science will ever produce a complete theory
on the motion of individual cars. Despite of this, the last five decades have
provided tools to construct a framework of useful – albeit incomplete – theo-
ries from traffic observations and experimentation. These incomplete theories
are neither deductive (i.e., stemming from excellent theories), nor inductive
(black box), but rather intermediate; basic mathematical model structures are
adopted, after which specific flow properties are determined from empirical
or experimental data. Since the only accurate physical law in traffic flow the-
ory is the conservation of vehicle’s equation, the main challenge of traffic flow
researchers is to look for intuitive and useful theories of traffic flow.

2.1.1 Microscopic and macroscopic characteristics and models

A key distinction made in the study of traffic systems is that between mi-
croscopic and macroscopic variables. Microscopic characteristics (e.g., time
headways, individual speeds and distance headways) pertain to the individ-
ual driver–vehicle unit in relation to the other drivers in the flow. Microscopic
models describe the behavior of individual vehicles in relation to the infrastruc-
ture and other vehicles in the flow. On the contrary, macroscopic characteris-
tics pertain to the properties of the traffic flow as a whole (for instance at a
cross-section, or at a time instant). Examples of macroscopic characteristics are
flow, time-mean speed, density, and space-mean speed. Macroscopic models
describe traffic flow in terms of its macroscopic characteristics. The intermedi-
ate mesoscopic level is used to indicate the behavior of groups of drivers.

2.1.2 Properties of traffic flow

Microscopic and macroscopic characteristics of traffic flow have been stud-
ied for many years. Studies concern a large variety of aspects such as the dis-
tribution of headways, statistical relations between speed and density, capacity
of the infrastructure, propagation of shock-waves, etc. A lot of consideration
has been put into average behavior of drivers under the assumption of station-
ary flow conditions. Under these conditions, it is reasonable to assume that the
average behavior of drivers is the same for the same average conditions. That
is, drivers having a certain speed u, will on average maintain the same distance
headway s = 1/k (where k denotes the vehicular density, i.e., the mean num-
ber of vehicles per unit roadway length) with respect to the preceding vehicles.
This in turn implies that if we may assume that there exists some statistical (but
not necessarily causal!) relation among the density k (or equivalently, the mean
distance headway s), the (space) mean speed u and flow q = ku, then it holds:

q = Q(k) = kU(k) (1)

Figure 1 shows an example of the three forms of this fundamental rela-
tion, showing some of its important properties (dU/dk < 0, U(kj ) = 0,
Q(kj ) = 0, qc = maxk Q(k), etc.). The fundamental relation will depend on
718 M. Papageorgiou et al.

Fig. 1. Examples of the fundamental relations between flow, density, and speed.

the different properties of the road (width of the lanes, grade), flow composi-
tion (percentage of trucks, fraction of commuters, experienced drivers, etc.),
external conditions (weather and ambient conditions), traffic regulations, etc.
Traffic flow observations however show that many data are not on the fun-
damental diagram. While some of these points can be explained by stochastic
fluctuations (e.g., vehicles have different sizes, drivers have different desired
speeds and following distances), a number of researchers have suggested that
these differences are structural and stem from the dynamic properties of traffic
flow. That is, they reflect so-called transient states (i.e., changes from conges-
tion to free flow (acceleration phase) or from free flow to congestion (deceler-
ation phase)) of traffic flow.
Several authors have studied the nonlinear or even chaotic-like behavior of
the traffic system (cf. Bovy and Hoogendoorn, 2000; Pozybill, 1998). Among
these behaviors are hysteresis and metastable or unstable behavior of traf-
fic flow. The latter implies that in heavy traffic a critical disturbance may be
amplified and develop into a traffic jam (spontaneous phase-transitions). In il-
lustration, empirical experiments performed in Forbes et al. (1958), and Edie
and Foote (1958, 1960) have shown that a disturbance at the foot of an upgrade
propagates from one vehicle to the next, while being amplified until at some
point a vehicle came to a complete stop. This instability effect implies that once
the density crosses some critical value, traffic flow becomes rapidly more con-
gested without any obvious reasons. More empirical evidence of this instability
and start–stop wave formation can be found in among others (Verweij, 1985;
Ferrari, 1989; Leutzbach, 1991). In Kerner and Rehborn (1997) and Kerner
(1999) it is empirically shown that local jams can persist for several hours, while
maintaining their form and characteristic properties. In other words, the sta-
ble complex structure of a traffic jam can and does exist on motorways.1 These
findings show that traffic flow has some chaotic-like properties, implying that

1 Apart from the formation of stop-and-go waves and localized structures, a hysteric phase-transition
from free-traffic to synchronized flow that mostly appears near on-ramps is described in Kerner and
 Ch. 11. ITS and Traffic Management 719

microscopic disturbances in the flow can result in the on-set of local traffic jams
persisting for several hours.
Having said this, it should be clear that traffic flow shows some interesting
phenomena, which must be reflected correctly by the different models that
have been proposed. The remainder of this section focuses on these different
models, while discussing their most important properties.

2.1.3 Approaches to traffic flow modeling

Traffic flow models may be categorized using various dimensions (determin-
istic or stochastic, continuous or discrete, analytical or simulation, etc.), sto-
chastic. The most common classification is the distinction between microscopic
and macroscopic traffic flow modeling approaches. However, this distinction is
not unambiguous, due to the existence of hybrid models. This is why below,
models are categorized based on the following aspects:
1. Representation of the traffic flow in terms of flows (macroscopic), groups
of drivers (mesoscopic), or individual drivers (microscopic).
2. Underlying behavioral theory, which can be based on characteristics of the
flow (macroscopic), or individual drivers (microscopic behavior).
The remainder of this section uses this classification to discuss some important
flow models. Table 1 presents an overview of these models and the relevant
sections.

2.1.4 Microscopic traffic flow models

A microscopic model provides a description of the movements of individ-
ual vehicles that are considered to be a result of the characteristics of drivers
and vehicles, the interactions between driver–vehicle units, the interactions be-
tween driver–vehicle units and the road characteristics, external conditions,
and the traffic regulations and control. In general, two types of driver tasks
are distinguished: longitudinal tasks (acceleration, maintaining speed, distance
keeping relative to leading vehicle) and lateral tasks (lane changing, overtak-
ing). With respect to the longitudinal movement, most microscopic simulation

Table 1.
Overview of traffic flow model classification

Representation Behavioral rules

Microscopic Macroscopic

Vehicle-based Microscopic flow models (2.1.4) Particle models (2.1.5)

Flow-based Gas-kinetic models (2.1.7) Macroscopic models (2.1.6)

Rehborn (1997). In addition, transitions from synchronized flow to the jammed traffic state occur in
congestion, upstream of the bottleneck.
720 M. Papageorgiou et al.

models assume that a driver will only respond to the one vehicle (the leader)
that is driving in the same lane, directly in front of her.
When the number of driver–vehicle units on the road is very small, the driver
can freely choose her speed given her preferences and abilities, the roadway
conditions, curvature, prevailing speed-limits, etc. In any case, there will be
little reason for the driver to adapt her speed to the other road-users. The
target-speed of the driver is the so-called free speed. In real life, the free
speed will vary from one driver to another, but also the free speed of a sin-
gle driver may change over time. Most microscopic models assume however
that the free speeds have a constant value that is driver-specific. When traf-
fic becomes denser, drivers will no longer be able to choose the speed freely,
since they will not always be able to overtake a slower vehicle. The driver will
need to adapt her speed to the prevailing traffic conditions, i.e., the driver is
following. In the remainder, we will discuss some of these car-following mod-
els. Models for the lateral tasks, such as deciding to perform a lane-change
and gap-acceptance, will not be discussed in this section in detail; a concise
framework of lane changing modeling is provided by Ahmed et al. (1996).

Safe-distance models. The first car-following models were developed in Pipes

(1953) and were based on the assumption that drivers maintain a safe distance:
a good rule for following vehicle i − 1 at a safe distance si is to allow at least the
length S0 of a car between vehicle i and the vehicle ahead for every ten miles
per hour of speed vi at which i is traveling:
si = S(vi ) = S0 + Tr · vi  (2)
where S0 is the effective length of a stopped vehicle (including additional
distance in front), and Tr denotes a parameter (comparable to the reaction
time). A similar approach was proposed in Forbes et al. (1958). Both theo-
ries were compared to field measurements. It was concluded that according to
Pipes’ theory, the minimum headways are slightly less at low and high velocities
than observed in empirical data. However, considering the models’ simplicity,
agreement with real-life observations was amazing (cf. Pignataro, 1973).
In Leutzbach (1988) a more refined model describing the spacing of con-
strained vehicles in the traffic flow was proposed. Considering an overall re-
action time Tr , the distance needed to come to a full stop given the initial
speed vi , the friction coefficient μ, and gravity g, equals
vi2
S(vi ) = S0 + Tr · vi +  (3)
2μg

Stimulus-response models. Stimulus-response models are dynamic models

that describe the reaction of drivers as a function of changes in distance,
speeds, etc., relative to the vehicle in front. These models are applicable to
relatively busy traffic flows, where the overtaking possibilities are small and
drivers are obliged to follow the vehicle in front of them. Drivers do not want
 Ch. 11. ITS and Traffic Management 721

the gap in front of them to become too large, so that other drivers might en-
ter it. At the same time, the drivers will generally be inclined to keep a safe
distance.
Stimulus-response models assume that drivers control their acceleration.
The well-known model proposed in Chandler et al. (1958) is based on the in-
tuitive hypothesis that a driver’s acceleration is proportional to the relative
speed vi−1 − vi :
 
ai (t + Tr ) = v̇i (t) = α vi−1 (t) − vi (t)  (4)
where Tr again denotes the overall reaction time, and α denotes the sensitivity.
Based on field experiments, conducted to quantify the parameter values for
the reaction time Tr and the sensitivity α, it was concluded that α depended on
the distance between the vehicles: when the vehicles were close together, the
sensitivity was high, and vice versa. The following specification was proposed
by
α0
α=  (5)
xi−1 (t) − xi (t)
One of the main aspects of a dynamic model is its stability, i.e., whether
small disturbances will damp out or be amplified. For the stimulus-response
model (4), two types of stability can be distinguished, namely local stability
(stability of response of a driver on the leading vehicle i − 1), and asymptotic
stability (propagation of disturbances along a platoon). Asymptotic stability is
of more practical importance than local stability. If a platoon of vehicles is as-
ymptotically unstable, a small disturbance in the movement of the first vehicle
is amplified as it is passed over to the next vehicle, which in turn can lead to
dangerous situations. Let us briefly consider both kinds of stability. The local
and asymptotic stability of the model depends on the sensitivity α and the re-
action time Tr , i.e., the model is locally stable if C = αTr < π/2. Asymptotic
stability requires C = αTr < 1/2. Note that local stability is less critical than as-
ymptotic stability because the stimulus-response model becomes unstable only
for (unrealistically) large response times or large sensitivity values.
This simple model has several undesirable and unrealistic properties. For
one, vehicles tend to get dragged along when the vehicle in front is moving at
a higher speed. Furthermore, when the distance si (t) is very large, the speeds
can become unrealistically high. To remedy this deficiency, sensitivity α can be
defined as a decreasing function of the distance. In more general terms, the
sensitivity thus can be defined as follows
α0 (vi (t + Tr ))m
α=  (6)
(xi−1 (t) − xi (t))l
Equation (6) implies that the following vehicle adjusts its speed vi (t) propor-
tionally to both distances and speed differences with delay Tr . The extent to
which this occurs depends on the values of α, l, and m. Combining Equations
(4) and (6), and integrating the result, relations between the speed vi (t + Tr )
722 M. Papageorgiou et al.

and the distance headway xi−1 (t)−xi (t) can be determined. Assuming station-
ary traffic conditions, the following relation between the equilibrium speed U
and the density k results
  (l−1) 1/(1−m)
0 k
U(k) = U 1 − (7)
kj
for m = 1 and l = 1; k = 1/(xi−1 − xi ) denotes the density (average number
of vehicles per unit roadway length); kj is the so-called jam-density (density at
which U = 0); U 0 is the mean free speed (at k = 0). We refer to Leutzbach
(1988) for a more general expression of (7).
An alternative approach was proposed in Helly (1959), which includes an
additional term describing the tendency of drivers to maintain a certain desired
following distance Si (t):
   
ai (t + Tr ) = α1 vi−1 (t) − vi (t) + α2 xi−1 (t) − xi (t) − Si (t)  (8)
where
Si (t) = β0 + β1 vi (t) + β2 ai (t) where βj  0 (9)
Car-following models have been mainly applied to single lane traffic (e.g.,
tunnels, cf. Newell (1961)) and traffic stability analysis (Herman et al., 1959;
May, 1990). The parameters of the model (7) have been estimated using
macroscopic and microscopic data by a large number of researchers. It should
be noted that no generally applicable set of parameter estimates has been
found so far, i.e., estimates are site-specific. An overview of parameter esti-
mates can be found in Brackstone and McDonald (1999).

Optimal speed models. So far, the models considered mainly describe the car-
following task where the follower (in time) will aim to drive at the speed of the
leader, at a certain distance gap. Of course, there can be choices of the desired
speed other than the speed of the leader. In Bando et al. (1995) it is assumed
that the desired speed is a function of the distance between the vehicles under
consideration, i.e.,
Udes (xi−1 (t) − xi (t)) − vi (t)
ai (t) =  (10)
Tr
where Udes (xi−1 − xi ) = U0 tanh(xi−1 − xi ).

Psycho-spacing models. The car-following models discussed so far have

a mechanistic character. The only human element is the presence of a finite
reaction time Tr . However, in reality a driver is not able to:
• observe a stimulus lower than a given value (perception threshold);
• evaluate a situation and determine the required response precisely, for
instance due to observation errors resulting from radial motion obser-
vation;
 Ch. 11. ITS and Traffic Management 723

• manipulate the gas and brake pedal precisely.

Furthermore, due to the need to distribute her attention to different tasks,
a driver will generally not be permanently occupied with the car-following task.
This type of considerations has inspired a different class of car-following mod-
els, namely the psycho-spacing models. The first psycho-spacing models were
based on theories borrowed from perceptual psychology provided in Michaels
(1963); cf. Leutzbach and Wiedemann (1986). In these models, car-following
behavior is described using a plane with relative speed and headway distance
as axes. The model is illustrated in Figure 2.
It is assumed that the vehicle in front has a constant speed and that the
potential car-following driver catches up with a constant relative speed vr =
vi−1 −vi . As long as the headway distance xi−1 −xi is larger than sg , there is no
response. If the absolute value of the relative speed is smaller than a boundary
value vrg , then there is also no response because the driver cannot perceive the
relative speed. The threshold value is not a constant but depends on the speed
difference. If the vehicle crosses the boundary, it responds with a constant pos-
itive or negative acceleration. This happens in Figure 2 first at point A, then at
point C, then point B, etc. The term pendeling (the pendulum of a clock) for
the fact that the distance headway varies around a constant value, even if the
vehicle in front has a constant speed, has been introduced in Leutzbach (1988).
In this action-point model the size of the acceleration is arbitrary in the first
instance, whereas it was the main point of the earlier discussed car-following
models.
Action point models form the basis for a large number of contemporary
microscopic traffic flow models. In Brackstone and McDonald (1999) it is con-
cluded that it is hard to come to a definitive conclusion on the validity of these
models, mainly because the calibration of its elements has not been successful
so far.

Fig. 2. Basic action-point car-following model.

724 M. Papageorgiou et al.

Submicroscopic simulation models. In addition to describing the time–space

behavior of the individual entities in the traffic system, submicroscopic simula-
tion models describe the functioning of specific parts and processes of vehicles
and driving tasks. For instance, a submicroscopic simulation model describes
the way in which a driver applies the brakes, considering among other things
the driver’s reaction time, the time needed to apply the brake, etc. These sub-
microscopic simulation models are very suited to predict the impacts of driver
support systems on the vehicle dynamics and driving behavior. Examples of
submicroscopic models are SIMONE (Minderhoud, 1995), MIXIC (van Arem
and Hogema, 1995), and PELOPS (Ludmann, 1998). For a review on micro-
scopic and submicroscopic simulation models, we refer to Ludmann (1998) and
Minderhoud (1995).

Cellular Automaton (CA) or particle hopping models. CA-models aim to com-

bine advantages of micro-simulation models, while remaining computationally
efficient by use of efficient storage and computation algorithms. The car-
following rules generally lack intuitive appeal and their exact mechanisms are
not easily interpretable from the driving-task perspective. These models de-
scribe the traffic system as a lattice of cells of equal size (typically 7.5 m).
A CA-model describes in a discrete way the movements of vehicles from cell
to cell (cf. Nagel, 1996, 1998). The size of the cells are chosen such that a ve-
hicle driving with a speed equal to one, moves to the next downstream cell
during one time step. The vehicle’s speed can only assume a limited number of
discrete values ranging from zero to vmax .
The process can be split-up into three steps:
• Acceleration. Each vehicle with speed lower than its maximum speed
vmax , accelerates to a higher speed, i.e., v ← min(vmax  v + 1).
• Deceleration. If the speed is greater than the distance gap d to the pre-
ceding vehicle, then the vehicle will decelerate: v ← min(v d).
• Dawdling (“Trödeln”). With given probability pbrake , the speed of a ve-
hicle decreases spontaneously: v ← max(v − 1 0).
Using this minimal set of driving rules, and the ability to apply parallel com-
puting,2 the CA-model is very fast, and can consequently be used both to sim-
ulate traffic operations on large-scale motorway networks, as well as for traffic
assignment and traffic forecasting purposes. The initial single-lane model of
Nagel (1996) has been generalized to multilane multiclass traffic flow. In Wu
et al. (1999) time-oriented car-following rules have been proposed, instead of
the traditional space-oriented heuristic rules. It is argued that the resulting
model describes drivers’ behavior more realistically.

2 When one relaxes the parallel update requirement, we generally do not speak of Cellular Automata
models. However, the term particle hopping model still applies (cf. Nagel, 1998).
 Ch. 11. ITS and Traffic Management 725

Verification of CA-models for car traffic on German and American motor-

ways and urban traffic networks (Wu et al., 1999; Esser et al., 1999), shows
fairly realistic results on a macroscopic scale, especially in the case of urban
networks in terms of reproduction of empirical speed-density curves.

Fuzzy logic-based models. The first application of fuzzy logic systems is due
to Kikuchi and Chakroborty (1992), aiming at fuzzifying the stimulus-response
model. The model was used to illustrate how a fuzzy logic system can be used
to describe car-following and local instability. More recent developments are
reported in Rekersbrink (1995) and Henn (1995).

2.1.5 Particle models

Particle models can be considered as a specific type of numerical solution
approach (so-called particle discretization methods; cf. Hockney and Eastwood,
1988), applied to mesoscopic or macroscopic continuum traffic flow models.
These models distinguish individual vehicles, but their behavior is described by
aggregate equations of motion, for instance a macroscopic traffic flow model.
An example of a particle model is INTEGRATION (van Aerde, 1994).

2.1.6 Continuum traffic flow models

Continuum traffic flow models deal with traffic flow in terms of aggregate
variables, such as flow, densities and mean speeds. Usually, the models are
derived from the analogy between vehicular flow and flow of continuous me-
dia (e.g., fluids or gasses), complemented by specific relations describing the
average macroscopic properties of traffic flow (e.g., the relation between den-
sity and speed). Continuum flow models generally have a limited number of
equations that are relatively easy to handle.
Most continuum models describe the dynamics of density k = k(x t), space
mean speed u = u(x t), and flow q = q(x t). The density k(x t) describes
the expected number of vehicles per unit length at instant t. The flow q(x t)
equals the expected number of vehicles flowing past cross-section x during per
time unit. The speed u(x t) equals the mean speed of vehicles defined ac-
cording to q = ku. Some macroscopic traffic flow models also contain partial
differential equations of the speed variance θ = θ(x t) or the traffic pressure
P = P(x t) = kθ.

Conservation of vehicles and the kinematic wave model. Assuming that the de-
pendent traffic flow variables (density, flow, speed) are differentiable functions
of time t and space x, the following partial differential equation represents the
fact that on a roadway, vehicles cannot be lost or created:
∂k ∂q
+ = r(x t) − s(x t) (11)
∂t ∂x
describing that the number of vehicles on a small part of the roadway of length
dx increases according to the balance of inflow and outflow at the boundaries
726 M. Papageorgiou et al.

(interfaces) x and x + dx, respectively, and the inflow r(x t) and outflow
s(x t) at on-ramps and off-ramps, respectively (source and sink terms). To-
gether with the fundamental relation q = ku, Equation (11) constitutes a
system of two independent equations and three unknown variables. Conse-
quently, to get a complete description of traffic dynamics, a third independent
model equation is needed.
In combining the fundamental relation Equation (1) with Equation (11),
a nonlinear first-order partial differential equation results: the kinematic wave
model (Lighthill and Whitham, 1955):
∂k ∂q ∂k ∂k dQ
+ = + c(k) = 0 where c(k) =  (12)
∂t ∂x ∂t ∂x dk
Here c(k) denotes the so-called kinematic wave speed, describing the speed at
which small disturbances propagate through the traffic flow.
Generalized solutions to the kinematic wave model can be determined by
the method of characteristics, see, e.g., Logghe (2003). For the kinematic wave
model, it can be shown that characteristic curves are straight lines in the
(x t)-plane with slope c(k) that emanate from the boundary (i.e., at x = x0 or
t = t0 ) of the considered time–space region. Along the characteristics, den-
sities are conserved and are thus equal to the density at the point on the
boundary from which the characteristic emanates. When on this boundary
∂c(k)/∂x < 0, the characteristic curves will in time intersect ( focusing) and
a shockwave will result. The shock wave speed ω can be determined from the
shock wave equation (May, 1990):
q2 − q1
ω=  (13)
k2 − k1
where (k1  q1 ) and (k2  q2 ) respectively denote the traffic conditions down-
stream and upstream of the shock S. Besides shockwaves, acceleration fans are
formed in case of discontinuities in the density, characterized by k(x t) >
k(y t) for x < y. These acceleration fans describe the way vehicles drive
away from a high-density region into a low-density region. A typical situation
in which this occurs, is a traffic light turning to green, where the acceleration
fan describes the way vehicles drive away from the formed vehicle queue.
The kinematic wave model can be solved efficiently either analytically or nu-
merically, and its properties and limitations are well understood. Amongst the
drawbacks of the model is the formation of shocks irrespective of the smooth-
ness of the initial conditions. Moreover, the kinematic wave model assumes
that the traffic speeds adapt to the stationary speed U(k) immediately (no
fluctuations around the equilibrium speed), and thus does not respect the fi-
nite reaction times and bounded acceleration possibilities of its constituent
elements. The latter drawback has been remedied in Lebacque (2002), by im-
posing additional constraints on the solutions of the kinematic wave model
prescribing the maximum acceleration of the cars. The kinematic wave model
is not able to predict stop-and-go waves with amplitude-dependent oscillation
 Ch. 11. ITS and Traffic Management 727

times, which are quite common in real-life traffic flow (Verweij, 1985), nor is
traffic hysteresis (average headways of vehicles approaching a jam are smaller
than vehicles driving out of a jam, see Treiterer and Myers, 1974) described.
Traffic instability is also not captured by the kinematic wave model.
Recent improvements in the model are reported in Daganzo (1997, 2002a,
2002b), considering multiple lanes, as well as dividing the driver population
into different user-classes showing different driving characteristics. The con-
cept of motivation, indicating that passing drivers will temporarily accept
smaller headways, is also introduced.

Payne-type models. To relax the assumptions that the speeds cannot differ
from the stationary speed u = U(k), the latter expression has been replaced by
a dynamic equation for the speeds alongside the conservation of vehicle equa-
tion (Payne, 1971). Payne-like models can be derived from car-following laws.
Considering a driver at location xi (t), looking ahead to location
xεi = (1 − ε)(xi−1 − xi ) + ε(xi−2 − xi−1 ) where 0  ε  1 (14)
In Zhang (2003) the following expression for the speed vi of vehicle i is used
    
vi (t + Tr ) = U k xεi (t) t + β viε (t) − vi (t) 
 
where viε (t) := u xεi (t) t  (15)
U denotes the equilibrium speed as a function of the density, Tr denotes the
reaction time, and β is a dimensionless parameter. Using Taylor series expan-
sions (Zhang, 2003), the following dynamic expression for the mean speed
v(x t) is derived
∂u   ∂u c 2 (k) ∂k U(k) − u ∂2 u
+ u + 2βc∗ (k) + = + μ(k) 2  (16)
∂t ∂x k ∂x Tr ∂x
where
dU
c∗ (k) = k and μ(k) = 2βTr c 2 (k) (17)
dk
denote the sound speed and the traffic viscosity, respectively.
From Equation (16), different factors can be identified that can be inter-
preted from driver behavior. The term (c 2 (k)/k)(∂k/∂x) denotes the effect
of driver anticipation, showing how drivers anticipate on downstream condi-
tions: in regions of increasing density, drivers will anticipate and reduce their
speeds accordingly. The relaxation term (U(k) − u)/Tr describes the smooth
adaptation of the speed u to an equilibrium state U(k), given the relaxation
time Tr ; under stationary conditions, we have u = U(k). The viscosity term
μ(k) ∂2 u/∂x2 reflects the influence of higher-order anticipation, i.e., the way
drivers react to changes in relative speeds viε (t) − vi (t).
For specific parameter choices, the general expression (16) can be reduced
to other models: the original model of Payne (1971) can be derived by choosing
728 M. Papageorgiou et al.

β = 0 (no higher-order anticipation). For a constant sound speed c∗ (k) = c0 ,

the viscous model of Kühne (1991) results.
It can be shown that the model is unstable in a certain density range (small
perturbations in the density grow into traffic jams), and that the model is able
to (qualitatively) describe stop-and-go traffic. Moreover, small perturbations
in the stable flow will dissipate. In general, solutions to the model are smooth.
As a result, the model does backward smoothing to a sharp concentration/speed
profile, thus possibly predicting negative driving speeds. We can therefore con-
clude that the model may violate the anisotropic character (traffic mainly reacts
to downstream traffic conditions) of traffic flow. This nonanisotropic nature
manifests itself prominently in the workings of shock and expansion waves: in
contrast to the kinematic model, which has only one family of kinematic waves,
Payne-type models have two, associated with the characteristic curves ξ12 de-
fined by
  
dξ12
= λ12 = u + β ± 1 + β2 c(k) (18)
dt
The waves in the first characteristic ξ1 field travel with a speed less than the
speed u, and are qualitatively identical to the kinematic waves in the kinematic
wave model. The waves in the second characteristic field travel faster than the
average traffic flow, implying that in this field, information reaches vehicles
from behind. In Zhang (2003) it is argued that for β < 1, this will never occur,
since the speed of the second characteristic ξ2 approaches u.
Extensions are reported in Hoogendoorn and Bovy (1999), Hoogendoorn
et al. (2002), and Helbing et al. (2001), which pertain to the modeling of mul-
ticlass and multilane traffic flow in networks, including nonlocal, forwardly
directed interactions, effects of vehicle space requirements. It is important to
note that these models are based on gas-kinetic models (see Section 2.1.7),
rather than on car-following models like equation (15).

2.1.7 Gas-kinetic flow models

Starting point of the gas-kinetic models, is the so-called phase-space density
(PSD),
κ(x t v) = k(x t) · f (v|x t) (19)
where κ describes the mean number of vehicles k(x t) per unit roadway length
and f (v) the speed distribution at that location and instant. Prigogine and
Herman (1971) were the first to use the notion of the PSD to derive a model
describing the behavior of traffic flow. They achieved this by assuming that the
PSD changes according to the following processes:
1. Convection ∂(vk)/∂x. Vehicles with a speed v flow into and out of the
roadway segment [x x + dx), causing a change in the PSD κ(x t v).
2. Acceleration towards the desired speed (V 0 (v) − v)/Tr , where V 0 (v)
denotes the expected desired speed of vehicles driving with speed v;
Tr denotes the acceleration time.
 Ch. 11. ITS and Traffic Management 729

3. Deceleration when catching up with a slower vehicle,

 while not being able
to immediately overtake (1 − p(k))κ(x t v) (w − v)κ(x t w) dw =
(1 − p(k))κ(x t v)(u(x t) − v).
Their deliberations yielded the following partial differential equation (PH-mo-
del):
∂κ ∂(vk)
+
∂t ∂x
 
∂ V 0 (v) − v    
= + 1 − p(k) κ(x t v) u(x t) − v  (20)
∂v Tr
where the density k and the mean speed u are defined according to

k(x t) = κ(x t v) dv and

u(x t) = vf (v|x t) dv (21)

The most complex process here is probably the interaction process. Let us
briefly discuss how this term is determined from the following, simple behav-
ioral assumptions:
1. The “slow-down event” is instantaneous and occurs with a probability of
(1 − p(k)), where p denotes the so-called immediate overtaking prob-
ability, reflecting the event that a fast car catching up with a slow car
can immediately overtake to another lane, without needing to reduce its
speed.
2. The speed of the slow car is not affected by the encounter with the fast
car, whether the latter is able to overtake or not.
3. The lengths of the vehicles can be neglected.
4. Only two vehicle encounters are to be considered, multivehicle encoun-
ters are excluded.
The model of Prigogine and Herman has been criticized and improved by
Paveri-Fontana (1975). He considers a hypothetical scenario where a free-
flowing vehicle catches up with a slow moving queue, and considers two ex-
treme cases:
1. The incoming vehicle passes the whole queue as if it were one vehicle.
2. It consecutively passes each single car in the queue independently.
In Paveri-Fontana (1975) it is shown that the Prigogine and Herman formalism
reflects the second case, while the real-life situation falls between these two ex-
tremes. He also shows that the term reflecting the acceleration process yields
a desired speed distribution that is dependent on the local number of vehicles.
This is in contradiction to the well-accepted hypothesis that driver’s personality
is indifferent with respect to changing traffic conditions (the so-called person-
ality condition; cf. Daganzo, 1995). To remedy this deficiency, Paveri-Fontana
730 M. Papageorgiou et al.

generalized the PSD κ(x t v) by also including the distribution of the desired
speeds

κ̃(x t v v0 ) = k(x t)f˜(v v0 |x t) (22)

where f˜(v v0 |x t) denotes the joint probability density function of speed v
and free speed v0 .
Other researchers have objected to the validity of the vehicular chaos as-
sumption underlying the expression for the effects of vehicle interactions. In
Munjal and Pahl (1969) it is argued that the interaction term “corresponds to
an approximation in which correlation between nearby drivers is neglected”,
being only valid in situations where no vehicles are platooning. This issue has
been remedied explicitly in Hoogendoorn and Bovy (1999) by distinguishing
between platooning and nonplatooning vehicles.
In Nelson et al. (1997) it is argued that plausible speed-density relations can
only be determined from the Prigogine–Herman model, based on the nontriv-
ial assumption that the underlying distribution of desired speeds is nonzero for
very small speeds. The situation when this assumption does not hold is investi-
gated in Nelson and Sopasakis (1998). It is found that at concentrations above
some critical value, there is a two-parameter family of solutions, and hence
a continuum of mean velocities for each concentration. This result holds for
both constant values of the passing probability and the relaxation time, and for
values that depend on concentration in the manner assumed by Prigogine and
Herman. It is hypothesized that this result reflects the well-known tendency
toward substantial scatter in observational data of traffic flow at high concen-
trations.
Paveri-Fontana model generalizations are reported in Hoogendoorn and
Bovy (1999), Hoogendoorn et al. (2002), and Helbing et al. (2001), where
gas-kinetic models for multiclass and multilane traffic flow including nonlo-
cal, forwardly directed interactions, effects of vehicle space requirements are
presented. These gas-kinetic models serve as the starting point to derive con-
tinuum models by application of the so-called method-of-moments. Another
multilane gas-kinetic model was proposed in Klar and Wegener (1999a, 1999b).
In Tampére et al. (2002) adaptive driver behavior is introduced into the gas-
kinetic modeling approach.

2.1.8 Model application

Traffic flow and microsimulation models designed to characterize the behav-
ior of the complex traffic flow system have become an essential tool in traffic
flow analysis and experimentation. The application areas of these tools are very
broad, e.g.:
• Evaluation of alternative treatments in (dynamic) traffic management.
• Design and testing of new transportation facilities (e.g., geometric de-
signs).
 Ch. 11. ITS and Traffic Management 731

• Operational flow models serving as a submodule in other tools (e.g.,

traffic state estimation, model-based traffic control and optimization,
and dynamic traffic assignment).
• Training of traffic operators.
Which or even if a model should be used, depends largely on the type of
problem at hand. Important issues are the purpose of the study, the required
level-of-detail (is the individual driver behavior and changes therein impor-
tant), what kind of data is available for model calibration and validation, and
the type of network considered (urban, motorway).
Nevertheless, some general remarks can be made. For one, the application
of microscopic (simulation) models will in general be more time-consuming,
both in the sense of computation time needed to perform the simulations (long
computation time per simulation due to detailed representation of dynamic
processes) and requirement to do multiple runs to get statistically valid results
in case of stochastic microsimulation models.
Moreover, calibration and validation of microscopic models may be a labo-
rious task. This can be explained by noticing that these models aim to mimic
human behavior in real-life traffic (not in contrived “car-following experi-
ments”), which is hard to observe, measure and validate (cf. Daganzo, 1994).
This is problematic, given the observed nonlinear behavior of the collective
traffic flow: the microscopic details have to be just right for the simulation to
realistically describe and predict for instance stop–start waves in traffic flow.
In Brackstone and McDonald (1999) it is convincingly argued that suitable
data (e.g., pair-wise vehicle trajectories collected by instrumented vehicles,
or remote-sensing; cf. Hoogendoorn et al., 2002) must be used in the model
calibration, and that the models are to be disassembled and tested in a step-by-
step fashion. In general, the lack of “microscopic” data results in macroscopic
calibration that cannot produce the optimal parameters since the number of
degrees of freedom is too large.
Macroscopic models are generally suited for large scale, network-wide ap-
plications, where macroscopic characteristics of the flow are of prime interest.
Macroscopic models are generally too coarse to correctly describe microscopic
details and impacts, for instance caused by changes in roadway geometry.
Macroscopic models are assumed to describe macroscopic characteristics of
traffic flow more accurately. Calibration of macroscopic models is relatively
simple (compared to microscopic models), for instance using loop detector
data (see Cremer and Papageorgiou, 1981; Helbing, 1997). Mostly, speed-
density relations derived from observations are required. In a recent paper,
Kerner et al. (2000), it is shown that traffic jam dynamics can be described and
predicted using macroscopic models that feature only some characteristic vari-
ables, which are to a large extent independent on roadway geometry, weather,
etc. This implies that macroscopic models can describe jam propagation reli-
ably, without the need for in-depth model calibration.
732 M. Papageorgiou et al.

3 Route guidance and information systems

3.1 Introduction

Advances in data processing, sensor and communications technologies have

made it possible to provide travelers with information on network conditions
based on real-time measurements. Better information should enable indi-
viduals to make better travel decisions. Moreover, as significant numbers of
travelers respond to such information, network conditions will themselves be
affected. Of particular interest here are systems intended to improve route
choice decisions, either by providing data on network conditions or by recom-
mending specific routes to a destination. These are called Route Guidance and
Information Systems (RGIS). A number of factors justify interest in these sys-
tems.
To begin with, travelers are frequently unaware of all the options available to
them. This is clearly the case for those unfamiliar with an area; the increasing
popularity of GPS-based navigation systems that provide turn-by-turn direc-
tions to a destination, attests to the value of basic way-finding information.
In addition, there is considerable evidence that even individuals who con-
sider themselves familiar with an area, have only a limited knowledge of travel
options. In Jeffery (1981), for example, it is estimated that, with better infor-
mation on travel options and conditions, even habitual drivers in an area could
reduce their average distance and travel time by around 7 percent.
Moreover, congestion is to a significant extent an unpredictable phenom-
enon. It has been estimated, for example, that roughly 60 percent of congestion
delays on US urban highways result from specific unpredictable events such as
accidents, vehicle breakdowns, and the like (Lindley, 1987). Recurrent con-
gestion, resulting from traffic levels that are systematically high relative to
available roadway capacity over a particular time period, also has a random
component that derives from variability in travel demand levels and network
performance. Because of the randomness in travel conditions, prior experience
can be an imperfect basis for travel decisions, and supplementing it with more
up-to-date information could result in better individual decisions as well as,
perhaps, improved network conditions.
A number of researchers have assessed the likely network-level impacts
and benefits of RGIS. Notable studies include Koutsopoulos and Lotan
(1989), Mahmassani and Jayakrishnan (1991), Al-Deek and Kanafani (1993),
Emmerink et al. (1995), and Hall (1996). Under a variety of assumptions and
approaches, these studies have investigated the likely reduction in total travel
time, the distribution of travel time savings between guided and unguided
drivers, and the variation of these benefits as a function of the RGIS market
penetration rate (fraction of drivers receiving guidance).
On the other hand, operational experiences with RGIS to date have gen-
erally been on too limited a scale and for too brief a time to allow strong
 Ch. 11. ITS and Traffic Management 733

empirical conclusions about its network level impacts to be drawn. Possible ex-
ceptions include high-volume corridors equipped with variable message signs,
and urban areas with real-time traffic condition reports. For example, using
data from the Washington, DC traffic information system (Wunderlich et al.,
2001) travelers with and without access to information on prevailing link tra-
versal times were simulated. Travelers were assumed to have a desired arrival
time at their destination, and to determine their path and departure time ac-
cordingly. The simulated travel experiences, in terms of travel times, on-time
arrival reliability, risk of lateness, and early and late schedule delays were com-
piled. It was found that guidance improved the various measures of travel time
reliability without significantly affecting average travel time itself.
This section reviews current knowledge about route guidance and informa-
tion systems. Although a distinction is sometimes made between prescriptive
guidance and descriptive information, both kinds of data will generally be re-
ferred to here as guidance. A particular set of guidance data disseminated at
a particular time will be called a message. Objectives and technological con-
straints that influence message content are discussed in Section 3.2.
Messages may be derived from static or dynamic information about the net-
work. Static systems provide fixed information about the network and may be
of use in tasks such as way-finding or preliminary trip planning; however, they
do not recognize actual traffic conditions. Static systems will not be further con-
sidered here. Dynamic RGIS can be classified as nonpredictive and predictive
systems. The former base the messages provided to drivers on measurements
or estimates of prevailing network conditions, while the latter derive messages
from forecasts of future network states. The two kinds of system can involve
quite different issues, and will be discussed separately in Sections 3.3 and 3.4.
Data on the effects of guidance on individual traveler behavior are available
from laboratory experiments with driving simulators and, to a more limited
extent, from surveys and observations of travelers who use RGIS. Knowledge
of these effects is important to develop predictive guidance, and ultimately to
evaluate the economic benefits of RGIS. Current knowledge of traveler re-
sponse to information is discussed in Section 3.5.
Finally, Section 3.6 identifies some areas of current research.

3.2 Overview of route guidance and information systems

3.2.1 Guidance objectives

In traffic networks subject to congestion, a flow pattern that optimizes
a system-level objective is not generally the same as one that results when each
driver independently chooses her preferred route. For example, the flow pat-
tern that minimizes the total travel time of all vehicles on the network generally
differs from the pattern obtained when each driver attempts to minimize her
individual travel time. The two types of pattern are referred to as system and
user optimal flow patterns, respectively.
734 M. Papageorgiou et al.

It is sometimes suggested that network operators could use route guidance

as a tool to force the network towards a system optimal flow pattern. This may
be an appropriate policy in exceptional circumstances such as emergency evac-
uation situations, but it is unlikely to be successful in the long run for routine
situations. Drivers are free to ignore any guidance messages that they per-
ceive as incompatible with their own decision criteria. Moreover, systematic
attempts to influence drivers’ decisions by providing misleading information,
even if well intentioned, are likely to lead over time to large scale driver rejec-
tion of the RGIS.
The focus here will be on guidance derived from user optimality objectives.
As argued in Hall (1996), such guidance can be usefully viewed as a way of
correcting the dis-equilibrium travel behavior that results from lack of infor-
mation. However, it may not be feasible to provide individual drivers with
messages precisely matched to their particular route choice decision criteria.
A common alternative approach is to base messages solely on travel times,
i.e., to report link or path traversal times (descriptive information) or to rec-
ommend minimum travel time paths (prescriptive guidance). Traversal times
prevailing at the time of guidance generation are often used for this purpose;
these are called instantaneous travel times. However, in a dynamic network,
with link traversal times that vary over time, instantaneous times may be dif-
ferent from the experienced times that drivers incur when making a trip. While
guidance based on experienced times is arguably closer to drivers’ own choice
criteria, generating such guidance requires the use of a predictive model, and
is considerably more complex than simply measuring and reporting the pre-
vailing traversal times.
Guidance is an attempt to improve the information available to drivers for
their route choice decision, yet the guidance itself will rarely be perfect. Data
collection, processing, and communications systems constrain the quality and
quantity of data that can be generated and transmitted, and humans are limited
in their ability to process information, in particular while driving. These factors
affect the type of guidance that an RGIS can provide. A poorly designed RGIS
can exacerbate rather than improve congestion. In Ben-Akiva et al. (1991) and
Kaysi et al. (1995) RGIS design issues are discussed including possible counter-
productive effects of poorly-designed systems. These include concentration, in
which guidance reduces the normal dispersion of driver behaviors and leads to
increased congestion on a smaller number of routes; and overreaction, in which
drivers’ response to guidance shifts congestion or leads to oscillations in flows
on different routes.

3.2.2 RGIS functional characteristics

The principal features that characterize different RGIS designs are briefly
identified here. Although the features are presented separately, many of them
are, in fact, interrelated in the sense that a choice for one constrains the feasi-
ble options for others.
 Ch. 11. ITS and Traffic Management 735

Basis for guidance. As was indicated above, a major distinction is between

nonpredictive systems that generate guidance based on measurements or es-
timates of prevailing network conditions, and predictive systems that utilize
forecasts for this purpose.
Local/area-wide focus. Guidance generation may focus on either a local or
a wide area traffic network. A local focus considers conditions on individual or
perhaps small contiguous groups of network elements (road segments or junc-
tions); the guidance will typically be disseminated only over that area. A wide
area focus considers conditions throughout the network in determining the
messages to disseminate.
Transmission range. The RGIS dissemination technology determines the
distance from a guidance source over which the messages may be received.
Line-of-sight, small area and wide area technologies are possible. In the case
of short- and medium-range technologies, the locating of the guidance dissem-
ination sources, and the resulting coverage of the network and its flows, are
important system design decisions.
Collective/individual dissemination. The RGIS dissemination technology also
determines whether messages can be received by all vehicles in transmission
range, or only by vehicles equipped with suitable receivers. Examples of the
former include roadside variable message signs and highway advisory broad-
casts over standard radio frequencies; examples of the latter include coded
infrared, microwave, FM sub-carrier, and cellular packet radio transmissions.
One-way/two-way communications. In an one-way communication system,
drivers receive messages from the RGIS but do not provide any informa-
tion to the system. In a two-way system, drivers notify the system about their
travel desires and receive messages that are tailored to their specific trip needs.
Moreover, the data acquired about travel times and drivers’ trip intentions and
choices can be incorporated in the system’s state estimates and forecasts.
Pre-trip/en-route access to guidance. Guidance received prior to beginning
a trip may influence the decision to travel or not, the destination(s), the time
of departure, the mode of travel as well as the particular route to follow. Guid-
ance received while en-route will generally only affect the subsequent choice of
path. Driver response to messages may also be different in the two situations:
a pre-trip decision is a choice without immediate antecedent, whereas the en-
route decision to switch routes may involve a reluctance to abandon a prior
choice and so exhibit hysteresis or a threshold effect.
Message dissemination and guidance update intervals. Guidance messages
generally relate to a period of time rather than to a single instant, and so are
maintained or retransmitted during that period. The message dissemination in-
terval is the period of time during which a disseminated guidance message does
not change. The length of this period may be dictated by technological con-
straints or by human factors. The guidance update interval refers to the time
between successive computations of the messages to disseminate. In a com-
plex guidance system, messages cannot be continually recomputed because of
736 M. Papageorgiou et al.

delays inherent in collecting and processing data, and in generating and dis-
seminating the messages.
Message design. The final issue concerns the syntax and semantics of the
guidance messages themselves, including their medium of delivery, format,
content, and precision. The distinction between prescriptive guidance and de-
scriptive information was mentioned above. Visual or audio messages intended
for direct reception by drivers cannot be overly complex because of the diffi-
culty of assimilating them while driving. Messages that will be processed by an
in-vehicle unit and conveyed to the driver in a schematic visual form might per-
haps have a higher data content. The available communications bandwidth or
message display capabilities may also constrain message complexity and preci-
sion.

3.3 Nonpredictive and related systems

In a nonpredictive dynamic RGIS, messages are derived from estimates of

the network conditions that prevail at or before the time the guidance is gen-
erated. To the extent that these instantaneous conditions are a good indication
of what a driver will actually encounter during a trip, then route choice deci-
sions made using nonpredictive guidance should be well founded. Conversely,
the potential limitation of nonpredictive guidance is that network conditions
may change significantly during a trip and so invalidate decisions based solely
on conditions around the time of departure.
Descriptive nonpredictive guidance consists of information on estimated
prevailing traffic conditions, by processing measurements collected from var-
ious kinds of traffic sensors. Details of such data collection and processing
methods are beyond the scope of this paper. Prescriptive nonpredictive guid-
ance consists of route recommendations based on estimated prevailing condi-
tions. Numerous methods have been proposed for generating such guidance.
In Papageorgiou (1990), Bolelli et al. (1991), and Charbonnier et al. (1991),
for example, nonpredictive route guidance approaches based on methods from
control engineering have been proposed.
Most currently operational dynamic RGIS are nonpredictive. They typically
disseminate messages using variable message signs, public radio and televi-
sion, or telephones. In-vehicle dynamic nonpredictive RGIS have mostly been
limited-scale experimental prototypes. For example, the TravTek system (Rilett
et al., 1991), deployed in Orlando, Florida, included an in-vehicle unit that re-
ceived by radio coded updates of prevailing link travel times and computed the
minimum time path to destinations selected by the driver.

3.4 Predictive systems

3.4.1 Predictive guidance and consistency

Predictive route guidance and information systems base guidance on fore-
casts of future network conditions. The messages disseminated to a driver
 Ch. 11. ITS and Traffic Management 737

reflect expectations of what conditions will be at network locations at the

time the driver will actually be there, and so are arguably closer to driver’s
actual decision criteria than nonpredictive messages based on prevailing or
historical conditions. In Ben-Akiva et al. (1996) the conditions under which
nonpredictive and predictive guidance result in the same flow patterns were
analyzed. In Pavlis and Papageorgiou (1999) it was showed that, in densely
meshed networks, nonpredictive guidance can have the same effects as predic-
tive guidance. In general, however, nonpredictive and predictive guidance will
be different and will have different impacts on network conditions.
Predictive guidance messages are derived from forecasts of future network
conditions. However, when these messages are disseminated and drivers re-
act to them in some way (for example, by changing departure times or paths),
future conditions are likely to be affected, possibly invalidating the forecasts
and rendering the messages irrelevant or worse. Within the context of a model
system, predictive guidance is said to be consistent if the forecasts on which
it is based are the same, within the limits of modeling accuracy, as those pre-
dicted to result after drivers receive the guidance and react to it. Consistency
is a generalization of the concept of traffic equilibrium. Unlike conventional
equilibrium, which assumes that travelers are perfectly informed about net-
work conditions, consistency accounts for the specific characteristics of avail-
able travel information and driver response to it.

3.4.2 Approaches for predictive guidance generation

One approach to generating predictive guidance simply ignores the con-
sistency issue. Guidance messages are based on condition forecasts that are
extrapolations from prevailing and historical conditions. The limitation of this
approach is that it does not take account of the effects of the guidance itself
on future network conditions: it does not ensure guidance consistency. This
may not be important at low levels of driver participation in the RGIS, but
is likely to lead to incorrect guidance messages when the number of vehicles
responding to guidance is sufficient to affect network flows and conditions.
Computation of consistent predictive route guidance requires application of
a dynamic traffic network model in order to forecast network conditions. Such
models are high-dimensional nonlinear systems; they can be quite complex and
require considerable amounts of data to identify. (Again, the necessary data
collection and processing tasks are not considered here.) They take the form
of mathematical models when analytical tractability is important, and of sim-
ulation models when the involved relationships defy convenient mathematical
expression.
Such a model is driven by exogenous time-dependent origin–destination
(OD) demands. The model includes components that predict how these de-
mands will choose paths and propagate along them. (Some models also predict
travelers’ choice of departure times.) The propagation of demand over its cho-
sen paths is determined by time-dependent link traversal times, and these
738 M. Papageorgiou et al.

traversal times are in turn affected by the congestion that results from the
movement of demand along the links.
Conventional (nonguidance) dynamic traffic assignment models compute
time-dependent equilibrium flows and traversal times under the assumption
that demand has perfect information about present and future congestion lev-
els and chooses paths accordingly. Guidance-oriented traffic network models,
on the other hand, must explicitly consider the availability and nature of travel
information, as well as driver behavior in the presence and absence of such
information. In Ben-Akiva et al. (1991) and Watling and van Vuren (1993) the
features that dynamic network models require for route guidance applications
are considered.
Guidance-oriented traffic models have been less studied than conventional
traffic assignment models. A high-level formal representation of a network
model for predictive guidance generation can be obtained using three time-
dependent problem variables and three maps (that are implemented as models
and algorithms) that relate them. The variables are: C, the network conditions;
M, the guidance messages; and P, the path splits (fraction of trips going to
a particular destination via each available path or subpath) at trip origins and
en-route decision points. The maps are:
• the network loading map S : P → C, which determines the network con-
ditions that result from the movement of exogenous time-dependent
OD demands over the network in accordance with a particular set of
path splits;
• the routing map D : M → P, which determines the path splits that
result from a particular set of guidance messages. The routing map gen-
erally incorporates a model of driver response to guidance messages;
and
• the guidance map G : C → M, which represents the response of the
RGIS, in the form of guidance messages, to a given set of network
conditions. (Note that messages output by this map for a given set of
conditions are not necessarily consistent, since driver reaction to the
messages may result in network conditions different from the inputs.)
Composite problem maps can be obtained by combining the network load-
ing, routing and guidance maps in different sequences. Each composite map
transforms an element of one problem variable into another element of the
same variable. There are three such maps (the symbol ∗ denotes functional
composition):
• a composite map D ∗ G ∗ S : P → P from the domain of path splits into
itself;
• a composite map S ∗ D ∗ G : C → C from the domain of link conditions
into itself; and
• a composite map G ∗ S ∗ D : M → M from the domain of messages into
itself.
 Ch. 11. ITS and Traffic Management 739

In terms of these composite problem maps, predictive guidance consistency

means that a map’s time-dependent inputs (i.e., the time trajectory of a prob-
lem variable) coincide with its time-dependent outputs. Equality of the map’s
input and output values means that the value is a fixed point of the map. (If
T : X → X is a one-to-one map, x∗ ∈ X is a fixed point of T if x∗ = T (x∗ ).)
Thus, consistency in the context of predictive route guidance can be computed
and studied in terms of fixed points of the problem maps, or of approximations
to them.
Fixed point approaches under perfect information assumptions have been
investigated in Kaufman et al. (1998) as a basis for dynamic traffic assignment
and in Engelson (1997) in the context of driver information systems. In Yang
(1998) and Bovy (1999) guidance models using fixed point approaches were
also considered. In Bottom (2000) additional details on the predictive guidance
generation formulation presented here are provided.

3.4.3 Operational predictive guidance systems

Operational experience with predictive systems is currently very limited.
The LISB system in Berlin (Hoffman, 1991) and the Autoguide system in
London (Catling, 1989) were early prototypes of systems with in-vehicle and
infrastructure-based components. Communications between the two compo-
nents served to establish travel times as well as to disseminate guidance. The
guidance consisted of next turn recommendations derived from minimum path
calculations using simple link traversal time predictions. These predictions,
in turn, were derived from historical link traversal time patterns and recently
measured traversal times. The usage of these systems did not attain levels that
would create consistency issues.
In the Netherlands, dynamic route information panels (DRIPs) can display
route recommendations for simple network topologies based on traffic predic-
tions and optimal control laws (Hoogendoorn and Bovy, 1998).
At least one traffic data company in the US sells network condition pre-
dictions that are based on extrapolations that take account of current traffic
measurements, typical patterns of link condition variations over time, and
other factors such as weather and special events. The details of the extrapo-
lation method are considered secret.
Systems that provide consistent predictive guidance at the level of an urban
or regional network are not yet operational. There have been limited experi-
mental deployments in traffic control centers of software systems (for example,
DYNASMART-X and DynaMIT) that are designed to generate such guidance.
The guidance generation logic in DynaMIT is explicitly based on the fixed
point approach described in the preceding section.

3.5 Driver response to RGIS

Understanding driver response to RGIS is required to develop effective

guidance systems that meet drivers’ information needs and contribute to con-
740 M. Papageorgiou et al.

gestion relief. Predictive guidance, in particular, requires the ability to forecast

driver responses to different possible messages in order to ensure guidance
consistency. Moreover, the economic evaluation of guidance systems also re-
quires a knowledge of the range of traveler responses to travel information.
Data on driver response to RGIS come from laboratory experiments with
travel choice simulators and, to a lesser extent, from observations of trav-
eler interactions with operational systems. The following paragraphs briefly
summarize what is currently known and not known regarding the impacts of
information on various travel-related decisions.
RGIS awareness and access decisions. Awareness, willingness to pay, and us-
age rates can be obtained by conducting travel surveys in areas where RGIS
services are available (Polydoropoulou and Ben-Akiva, 1999). The large-scale
panel surveys conducted every 3 years in Seattle-area by the Puget Sound Re-
gional Council provide data on the evolution over time of awareness and usage
decisions for different RGIS types (Peirce and Lappin, 2002).
Decision to travel or not. Relatively little information is available regard-
ing the effects of RGIS on the decision to travel or not; however, it is not
inconceivable that information about sufficiently bad travel conditions could
induce travelers to cancel their intended trips, particularly discretionary trips.
In Khattak et al. (1999) evidence are cited for this effect among noncommuters
from surveys carried out as part of the San Francisco-area TravInfo project.
Choice of destination or destinations. Similarly, relatively little information is
available in the literature regarding the effects of RGIS on destination choice,
or on the decision to visit several destinations and accomplish several purposes
in one trip through trip-chaining. Trips offering a choice among multiple des-
tination alternatives are likely to be for shopping (see, for example, Kraan et
al., 2000) or personal purposes. Opportunities to group multiple purposes and
destinations into a single trip-chain are more varied and difficult to character-
ize.
Departure time choice. In Mannering et al. (1994) results from Seattle-area
surveys about commuters who receive travel information from radio, televi-
sion, and telephone services are presented. Of the commuters surveyed, 40%
indicated that they had some flexibility in scheduling and selecting the route
for their morning commute trip; 23% indicated no flexibility. However, 64%
responded that they rarely changed their departure time because of pre-trip
information.
Mode choice. Little detailed information is available about the mode choice
impacts of RGIS, although there is some evidence for this effect. As reported
in Yim and Miller (2000), less than 1% of the early callers to San Francisco’s
TravInfo service asked to be rerouted to the transit menu after learning about
bad traffic conditions from the traffic menu. However, as experience with the
system increased over the duration of the TravInfo field test deployment, it was
found that up to 5% of the callers asked to be rerouted to the transit menu. Of
those who accessed transit information, 90% ultimately chose transit for their
travel mode.
 Ch. 11. ITS and Traffic Management 741

Route choice. Many surveys and travel choice simulator studies have demon-
strated the ability of RGIS to influence route choice. Based on analysis of
driver route choice responses to both VMS and radio information, it has been
suggested in Emmerink et al. (1996) that some people have an innate propen-
sity to use traffic information of any kind and from any source. Nonetheless,
there is considerable evidence that the nature of the guidance information, and
the conditions experienced prior to its dissemination, can strongly affect driver
route choice response to it.
Drivers’ perceptions of the accuracy and reliability of the messages is a key
determinant of their response. It has been found (Kantowitz et al., 1997) that
there exists an accuracy “threshold”, beneath which drivers will simply ignore
RGIS messages. Factors that increase drivers’ confidence in the accuracy of
the messages tend to increase the likelihood that the drivers will react to them.
In the context of route choice, such factors include observation of congestion
prior (and particularly just prior) to receiving the message, and favorable expe-
riences with the RGIS in prior uses. Drivers appear to be tolerant of a certain
amount of error in RGIS messages, although drivers familiar with an area will
expect a higher degree of accuracy from the information system.
Some drivers express a strong preference for descriptive information
on traffic conditions, while others prefer prescriptive recommendations of
a particular route to take (Khattak et al., 1996; Polydoropoulou et al., 1996).
Combining a prescriptive recommendation to change routes with descriptive
information justifying the recommendation has been found in some travel
choice simulator experiments to result in the highest route switching compli-
ance rates (Bonsall and Palmer, 1999).
A number of generally idiosyncratic factors condition a driver’s route choice
response to RGIS messages. For example, a motorway bias has been observed
in several studies. Because of this bias, drivers receiving messages that suggest
diverting from a nonmotorway to a motorway facility are considerably more
likely to comply than those who receive the opposite message, other things
being equal. As mentioned above, habit also plays a significant role in travel
decisions.
Learning. The day-to-day dynamics of commuter pre-trip departure time
and route choices as well as en-route path switching for morning commutes
were analyzed in Mahmassani and Liu (1999). Factors affecting route choice
behavior include: (1) arrival time flexibility, (2) user characteristics, and (3) in-
formation reliability. In Ozbay et al. (2001) the use of a stochastic learning
algorithm to analyze drivers’ day-to-day route choice behavior is proposed.
This model addresses the learning behavior of travelers based on experienced
travel time and day-to-day learning.

3.6 Areas of current research

Further development of route guidance and information systems will re-

quire better understanding of a number of issues, many of which have only
742 M. Papageorgiou et al.

recently begun to receive attention. This section describes ongoing research in

RGIS architecture, real-time computing, stochasticity and driver behavior.
System architecture. Nonpredictive guidance systems are relatively simple in
conception and robust in operation. Although they use data on instantaneous
network conditions, these systems may sometimes succeed in attaining objec-
tives based on experienced conditions, but need not do so in general. Basing
guidance on instantaneous conditions may sometimes exacerbate rather than
improve traffic problems. Predictive systems, based on experienced conditions,
depend on the availability and reliability of complex models of traveler be-
havior and network performance. Furthermore, they may be sensitive to high
levels of noise in model predictions, and are computationally demanding. Is
there a guidance system architecture that combines the better features of the
two approaches while avoiding their drawbacks? For example, multilevel con-
trol system designs have been developed for traffic control systems, but have
been less investigated in the context of RGIS.
Real-time response. Predictive guidance generation for a realistic network
requires considerable amounts of computation, yet it must be done quickly and
accurately enough for the guidance to be timely and of use to drivers. Parallel
and distributed computation environments are of interest in this regard, as are
fixed-point solution heuristics.
Stochasticity. Any of the individual maps involved in the predictive guidance
generation problem may be stochastic. The composite problem map will then
be stochastic as well, and its output when evaluated will be a realization of
a stochastic process rather than the deterministic time trajectory of a problem
variable. Nonetheless, the fixed-point interpretation of guidance consistency
continues to apply in this case, with the understanding that consistency now
means that problem map inputs and outputs are both stochastically equiva-
lent realizations of the same stochastic process. Markov chain Monte Carlo
techniques such as Gibbs sampling may be used to compute problem variable
statistics to any desired degree of accuracy. However, this approach is very
computationally demanding.
In practice, most stochastic guidance modeling efforts have adopted a “noisy
map” approximation to address the effects of stochasticity in the computation
of guidance solutions. Implicitly or explicitly, these approaches treat model
outputs not as a realization of a general stochastic process but rather as a time
trajectory of deterministic values affected by noise. Stochastic approximation
procedures are then applied to compute the fixed point. No rigorous justifica-
tion has yet been provided for this approach.
Driver behavior modeling. Applications of RGIS require the development
of reliable models of driver behavior and, in particular, of their response to
guidance messages. An important aspect of this is the development of better
models of the ways in which travelers form new perceptions from their most
recent experience, the guidance they received and their earlier experiences.
These efforts will benefit from advances in the understanding of the psycholog-
ical and cognitive processes involved in decision-making. Of particular interest
 Ch. 11. ITS and Traffic Management 743

are studies of decision-making under the time pressure of driving situations,

and studies of the ways in which spatial and network knowledge affect driver
response to RGIS.

4 Urban network traffic control

4.1 Introduction

Optimum management and control of traffic in urban networks is an impor-

tant requirement for city authorities as they seek efficient, safe and sustainable
transport. In addition there is an increasingly wide range of demanding objec-
tives for transport policy makers to achieve, such as public transport priority,
improved conditions for vulnerable road users, real-time traffic information;
emergency and incident management and restraining traffic in sensitive ar-
eas. As a response to these issues, Urban Traffic Management and Control
(UTMC) systems have been introduced in many cities around the world to pro-
vide the tools to support efficient and effective network management to meet
needs of current and future traffic problems. Fundamentally UTMC systems
are conceived as modular, open systems that incorporate and build on existing
functionalities of existing signal control and other traffic management systems
as illustrated in Figure 3. An important point to note is that the Urban Traffic
Control (UTC) systems are often at the heart of UTMC and provide a better
migration path so that improvements in UTC are utilized to the full in UTMC.
UTC refers to the control of traffic in urban areas using traffic signals, which
are linked to operate in a coordinated way. Such linked signal systems may
be used to achieve a variety of policy objectives, which relate to efficiency of
traffic operations, improved safety, reduced atmospheric pollution, priority for
specific road user groups, access control to maintain or enhance urban envi-
ronments, and to mitigate the effects of irregular events such as accidents or
road closure. UTC systems use historic or (more commonly) real-time knowl-
edge of network conditions to determine the control strategy most appropriate
for the conditions, and signal infrastructure to inform and control road users.

4.2 UTC systems: general requirements

Early systems in the 1950s and 1960s were based on fixed-time traffic control
providing signal coordination or progression for traffic on an arterial, through
the optimization of offsets between adjacent sets of signals. UTC was therefore
justified on there being a sufficient density of traffic signals to make signal co-
ordination worthwhile, compared to the alternative of operating traffic signals
in isolation. Whilst relatively effective for traffic co-ordination in “predictable”
conditions, the inability of fixed-time systems to adjust to changing traffic con-
ditions has been a drawback in this approach. The desire for traffic signaling
744 M. Papageorgiou et al.

Fig. 3. Schematic illustration of a UTMC system (Source: Department of the UK Environment, Trans-
port and the Regions (1999))

to be more responsive to changing traffic conditions has led to the develop-

ment of a range of semi or fully traffic responsive UTC systems. The improved
performance of these systems has generally justified their additional cost (e.g.,
detection, maintenance, etc.).
A variety of methods for UTC have evolved over the last decades, re-
sponding to the needs of individual cities/countries, the existing research and
development base and advances in detection, communications and control
technology. These traffic-responsive UTC systems are continuously upgraded
to meet with current requirements. Quality attributes of a UTC system play
a major role in its architecture. These may include attributes such as the
speed of system response to recurring congestion and incidents (i.e., respon-
siveness), feedback philosophy, ability of integration, functional and spatial
extendability, wider range of control strategies, robustness, installation and
maintenance costs, etc. Flexibility of the system to incorporate enhancements
as policies/technologies advance is a further key attribute.
Criteria for installing a real-time UTC system are now much wider than the
need for efficient signal coordination for traffic. For example, the UTC com-
munications infrastructure and processing capabilities give a powerful tool for
 Ch. 11. ITS and Traffic Management 745

the network manager, including such functions as traffic information, auto-

matic incident detection (AID), and congestion management.

4.3 Fundamentals of UTC

4.3.1 Principles of traffic signal control

Traffic signals operate by giving sequential priority to movements, including
pedestrian/cycling stages and other priorities. Sufficient separation of stages is
essential to ensure that conflicts between movements do not occur and there
are a variety of regulations and guidance documents provided by government
agencies. The regulations are similar across many countries, and differences re-
flect local driver behavior/expectation, enforcement regimes, and national atti-
tudes to guidance, regulation, and safety. In general, the smaller the amount of
time lost to road users when changing signal priority, the greater the capacity
and hence the shorter the delay.
Traffic signal controlled intersections may operate in isolation or be linked
to one or more adjacent intersections as part of a coordinated approach,
i.e., UTC. An isolated traffic signal is usually set by estimating an optimum
cycle time and green splits (the green times allocated to each separate move-
ment), i.e., those which minimize the delay. The optimum times are deter-
mined from a knowledge of traffic demand and the maximum (or “saturation”)
flows, also taking account of the time lost to traffic movements when the signals
are red to all traffic and the time lost as flows build up and fall from maximum
discharge rates. Usually delay is minimized when the degree of saturation (i.e.,
the ratio of demand to capacity, for the key movement) is about 0.85. Capacity
is a key parameter affecting performance and is determined as the saturation
flow multiplied by the effective green time per hour available for that flow.
When linking signals, the offset (i.e., the time delay in the start of the down-
stream cycle) is the crucial third variable to be considered with green splits
and cycle time. The need for linking signals usually relates to their proximity
to each other and the extent to which linking allows “platoons” of traffic to
proceed through adjacent junctions more efficiently.
For any specific condition of traffic demand a range of algorithms may be
used to optimize signal settings. In the early UTC systems, the database of
traffic demand was assumed fixed and signal timings were determined off-line.
However, traffic demand exhibits substantial short-term variability as well as
longer term changes in levels of flow and movement patterns. This has led to
increasingly sophisticated approaches to UTC which rely on substantial de-
tector input. Broadly, UTC systems may be categorized as fixed-time using
historic databases or demand responsive using on-line traffic data inputs. The
latter may be subdivided into centralized or decentralized systems. The char-
acteristics of these systems are outlined below in Sections 4.3.2 and 4.3.3 and
further details of specific systems are given in Section 4.4.
746 M. Papageorgiou et al.

4.3.2 Fixed-time systems

In fixed-time systems, off-line optimization is undertaken using demand lev-
els which are assumed constant for the period over which each fixed-time plan
is intended to run. Up to 10–15 plans may be developed to represent the com-
plete set of traffic conditions to be found on the network at different times.
A network is considered to operate as a series of different regions of groups of
linked signals, within each of which different signal plans will run. To ensure
that the arrival patterns of successive platoons of vehicles arrive consistently at
the downstream signals:
(1) a single cycle time must apply across the region and
(2) time offsets of the starts of successive cycles at one intersection from the
next must be the same.
The regional cycle time is based on that required for the busiest intersection.
Thus, a region may be bounded either by road links along which the benefits
of linking are small or where a common cycle time is very inappropriate.
Fixed-time plans may be readily used to create green waves, give pre-
determined priorities, and respond to special events which can be predicted,
such as football matches. They cannot respond to unplanned incidents such as
traffic accidents or unplanned road works. Plans may be set to change at pre-
determined times or changes may be triggered by flow or queue measurements
taken at key locations. There are also several systems which generate new plans
on-line, i.e., using very recent historic data.
Fixed-time systems require a considerable amount of traffic data to be col-
lected to set up and to keep up-to-date. Fixed-time plans can age rapidly,
particularly where traffic growth is high, and the benefits of linking may be lost
in three to four years if the plans are not updated. A further problem occurs
when plans change and discontinuities in flow patterns occur. This limits the
number of plans which can be used. Both the above points can be addressed
using traffic-responsive systems.

4.3.3 Traffic-responsive systems

Traffic-responsive systems use on-line detector measurements to optimize
signal timings on a cycle-to-cycle basis to better meet demand. Such sys-
tems may be coordinated largely from a central computer, e.g., SCOOT
(Bretherton, 1998) or have distributed intelligence and be coordinated largely
at a local level, e.g., UTOPIA (Donati et al., 1984). Centrally controlled sys-
tems use less intelligent local controllers, whilst with decentralized systems
each controller is more capable of taking local decisions, with some coordina-
tion between adjacent controllers. A wide range of traffic responsive systems
are now available with varying degrees of central and local control and key
systems are described in Section 4.4.
If a system is to respond to changes in traffic conditions, comprehensive de-
tection must be available. Detectors must be accurate, located appropriately
for the characteristics of the UTC system, and the information must be reli-
ably sent to the appropriate control center. In general, the more sophisticated
 Ch. 11. ITS and Traffic Management 747

the system the more comprehensive the detector requirements and the more
susceptible it is to detector failure. Many systems have default values for the
controllers based on time of day which are implemented if loss of detectors or
communication occurs.
Increasingly, a wide range of detectors are available. Traditionally, ground-
based systems using inductive loops to measure the presence of a vehicle have
formed the basis of most UTC detection. Other ground-based systems include
magnetometers which measure changes in the earth’s magnetic field brought
about by the presence of a vehicle. Above ground detectors include microwave
systems, radar, infra-red, video, and laser systems. Each has specific charac-
teristics to capture different aspects of vehicle behavior. Image-based systems
can be installed without costly and disruptive installation works, but have yet
to reach their full potential. Using vehicles themselves as detectors is an ap-
plication being considered for the future. Overall, the quality, quantity, and
reliability of future information will encourage more sophisticated UTC con-
trol strategies.
Table 2 provides a summary of the main advantages and disadvantages of
different types of UTC systems.

4.4 System summaries

4.4.1 Fixed-time systems

TRANSYT (TRAffic Network StudY Tool). TRANSYT (Robertson, 1997) is
the most well-developed and widely-used fixed time UTC system. It is an off-
line program for calculating optimum coordinated signal timings in a network
of traffic signals. For each distinct traffic stream it assumes that the flow rate
averaged over a specified period is known and constant and that the saturation
flows for each link are also known. TRANSYT consists of two main elements
called the “traffic model” and the “signal optimizer”, as shown in Figure 4.
The traffic model represents traffic behavior in a highway network and pre-
dicts a performance index (PI) for the network for a given fixed-time plan and
average set of flows on each link. The PI measures the overall cost of traffic
“congestion”, which is usually a weighted combination of the total delay and
the number of stops made by vehicles.
Cyclic Flow Profiles (CFP’s) showing the distribution of flows entering each
link are used with a “platoon dispersion” model to estimate patterns of vehicle
arrivals at the downstream junction. “Uniform” delay is calculated in a similar
way as for SCOOT, illustrated in Figure 5, supplemented by formulae to repre-
sent random delays and oversaturated delays when the junction is over-loaded
(i.e., the queue does not clear in the green period). Signal optimization involves
an iterative “hill climbing” process to adjust the signal timings to achieve an op-
timum PI. Specific links may be given extra weighting by the user to implement,
for example, green waves on a corridor. Other city/country specific fixed-time
UTC systems are used around the world, but are not as widespread nor as well
documented as TRANSYT, so they are not described further here.
748 M. Papageorgiou et al.

Table 2.
Summary of advantages and disadvantages of different types of UTC systems

UTC system Advantages Disadvantages

Fixed-time 1. Cheaper to install and 1. Large amount of data to

maintain. be collected and updated.
2. Can be implemented using 2. Signal plans may require
noncentrally controlled updating.
equipment.
3. Familiarity with settings 3. Disruption of plan
for regular users. changing.
4. Green waves more easily 4. Operator reaction to
implemented. incidents required.
5. Can favor specific vehicle 5. Can not deal with
types easily. short-term traffic
fluctuations.
Responsive plan 1. Can deal with some day to 1. Requires more data than
selection day fluctuations. fixed-time systems.
2. Plan change time could be 2. Detector failures possible.
more appropriate.
3. Might be valuable on 3. Needs discussions on
arterial routes. thresholds for plan change.
4. Cheaper than fully 4. Plan may change for a
responsive systems. wrong reason.
5. Difficult to foresee all plan
needs.
Fully responsive 1. Less data needed to be 1. Detector failures possible.
collected in advance.
2. Plan evolves, so avoids 2. More expensive to install
problems with plan changing and maintain.
and updating.
3. Can deal with short and 3. Requires some central
long term traffic fluctuations. control.
4. Automatic reaction to 4. Maintenance critical.
incidents.
5. Monitors traffic situation
throughout the area.

4.4.2 Traffic-responsive systems: Centralized

With centralized control, traffic detector information is sent to the UTC
center where it is processed and “optimum” timings are calculated for all the
traffic signals within the UTC system. These timings are then sent back to each
traffic signal controller on-street. Intelligence is therefore retained at one lo-
cation (the UTC center). The costs of on-street controllers can then be less,
although communication costs will usually be higher than decentralized sys-
tems with distributed intelligence. Five such systems are summarized below.
 Ch. 11. ITS and Traffic Management 749

Fig. 4. Traffic model and signal optimizer in TRANSYT.

Fig. 5. Principles of the SCOOT traffic model (Source: Department of the UK Environment, Transport
and the Regions (1999)).
750 M. Papageorgiou et al.

SCOOT (Split Cycle Offset Optimization Technique). SCOOT (Hunt et al.,

1981; Bretherton, 1998) was developed in the United Kingdom and is op-
erational in many cities around the world. It operates based on three main
principles, namely the measurement of cyclic flow profiles (CFP), the updat-
ing every 4 seconds of its on-line traffic model of queues and delays on each
link and incremental optimization of signal timings. SCOOT uses detectors at
the upstream end of links to measure demand and CFP’s in real time. The up-
stream detection also allows any congestion on the link to be monitored (i.e.,
when the queue reaches the upstream detector) and the possible exit block-
ing effects of this congestion on upstream links. The SCOOT model predicts
downstream arrival patterns using a calibrated link cruise speed with some dis-
persion. The saturation flow rate for each signal stop line is validated when the
system is commissioned; this allows the growth and clearance of queues to be
estimated accurately. The on-line traffic model is used in real time by the signal
optimizer. SCOOT has three optimization procedures by which it adjusts sig-
nal timings (Department of the UK Environment, Transport and the Regions,
1999). These are the cycle time, green splits, and offsets, each optimized using
a different procedure at different frequencies. By the combination of relatively
small changes to traffic signal timings, SCOOT can respond to both short-term
local peaks in traffic demand, as well as following trends over time and thus
maintain a constant coordination of the signal network.
In addition to the optimization from the basic SCOOT model, the operation
has considerable flexibility to override values and set parameters for different
regions and different times. These may include gating strategies to protect an
area from excessive levels of traffic, bus priorities, etc. In addition to network
management, SCOOT has a substantial data base facility for storing, manipu-
lating, and presenting traffic data including flows, journey times, and queues.
Further facilities have been included in the latest releases (Bretherton et al.,
2003) with further research ongoing (Bretherton et al., 2004).

SCATS (Sydney Coordinated Adaptive Traffic System). SCATS (Lowrie, 1982)

was developed in Australia and has been implemented in many cities around
the world. It operates at two basic levels known as the “upper” level, which in-
volves offset plan selection and the “lower” level, which involves optimization
of junction parameters. The upper level generates offset plans by time of day
from historic data while the lower junction level optimizes green splits, cycle
times, and offsets between signalized junctions using an incremental feedback
process based largely on detectors situated at the stop lines. SCATS calculates
green splits based on the flow in the previous cycle and so is not fully respon-
sive to unpredictable arrival flows. It differs from systems such as SCOOT and
UTOPIA in that it does not have a traffic model and uses stop line detectors
to estimate departure rates, rather than arrival rates modeled from upstream
detectors.
SCATS is basically a modular system largely run by regional computers ca-
pable of handling a large number of intersections, with significant intelligence
 Ch. 11. ITS and Traffic Management 751

within local controllers. A central computer may also be used to improve man-
agement functions. SCATS differs from many other systems in that the network
manager has a more direct involvement in setting up the system, i.e., it does not
have a model. The degree of operator understanding increases with the level
of simplicity of a system and this would lead to corridor operations being ad-
dressed most beneficially.

RHODES (Real-Time Hierarchical Optimized Distributed and Effective System).

RHODES (Head et al., 1992) was developed in United States. The RHODES
architecture is based on three levels of hierarchy. The highest level assigns
traffic to the network to determine base levels of traffic across the network
which takes into account both evolving traffic demand and current network
geometry. At the next level down, RHODES operates as a more typical UTC
system based on predicted platoon arrival patterns. At the intersection level
the movements of individual vehicles are modeled.
Basically there are two processes in RHODES namely “estimation and pre-
diction” and “decision system”. The first process takes the upstream detector
data and estimates the actual flow profiles in the network and the subsequent
propagation of these flows. On the other hand, the second process is where
the phase durations are selected to optimize a given objective function (min-
imization of average delay per vehicle, average queue lengths, numbers of
stops, etc.), the optimization being based on dynamic programming and de-
cision trees. Recently RHODES has been updated for the integration of bus
priority measures (Mirchandani et al., 2001).

MOTION (Method for the Optimization of Traffic Signals in Online Controlled

Networks). MOTION (Busch, 1996) is a recent UTC system developed
in Germany, with some limited implementation in some European cities.
MOTION is basically based on four functional levels namely “data acquisition
and pre-processing”, “traffic modeling and analysis”, “optimization of control
variables”, and “decision and transfer of signal programs”. The first module
receives the dynamic information from detection equipment via the central
traffic computer and may perform some data processing functions like deter-
mination of origins and destinations. The second module uses most important
individual traffic streams to determine actual O–D streams within the network
and turning movements at intersections, which is necessary to calculate green
splits and minimum cycle times for each intersection. Selection of a common
network cycle time for coordination, determination of link progression speeds,
and optimal offsets between individual intersections to minimize delay and
stops are determined in the third module. On the fourth level new signal pro-
grams are evaluated and transferred via the central traffic computer.

TUC (Traffic-Responsive Urban Control). TUC (Diakaki et al., 2000) has

been developed recently in Greece and has also been implemented in a few
other European cities, particularly in the context of EC-funded demonstration
752 M. Papageorgiou et al.

projects. TUC operates by modifying nominal signal timings using a multivari-

able regulator on-line. Nominal starting values for signal timings are based
on historic levels of demand. The regulator is based on the formulation of the
urban traffic control operation as a linear–quadratic control problem. The con-
trol objective is to minimize and balance link queues taking into account link
storage capacity. This formulation is potentially particularly useful in dealing
with oversaturated conditions. TUC has been expanded in recent years to allow
for public transport priority (Diakaki et al., 2003).

4.4.3 Traffic-responsive systems: decentralized

With decentralized (distributed) control, more intelligence for signal op-
timization is distributed to local traffic signal controllers. This can increase
flexibility and reduce communications costs, but controllers are usually more
costly. Three distributed systems are summarized below.

UTOPIA (Urban Traffic Optimization by Integrated Automation). UTOPIA

(Donati et al., 1984) was originally developed in Italy. The system is structured
as a hierarchical system organized on three levels known as the local level, the
area level and the town supervisor level. UTOPIA’s intelligent local controllers
can communicate with each other as well as with a central computer. The local
outstation level applies a microscopic model to estimate the state of the inter-
section directly collecting data from detectors located at the start of each link.
Local queue and turning percentage estimation, saturation flows and delay cal-
culations are performed by the local “observer”. The next level uses a historic
traffic database to validate the local detection, checking changes in the traffic
data or making comparisons of data upstream and downstream of the con-
gested links. The final level integrates the congestion information with data
from other systems like public transportation. The macroscopic model used at
this level has the advantage of collecting different sources of information and
having the coverage of the whole city.

PRODYN. PRODYN (Farges, 1990) was developed in France and has been
implemented in some other European cities. It uses an intersection open-loop
optimal feedback algorithm for traffic signal control. As with SCOOT and
UTOPIA, detectors are located at the upstream end of each link and where
appropriate at 200 m and 50 m upstream. The detectors collect occupancy
data. The system operates in 5 sec steps and the demand for each period is
estimated from that in the previous period. A time horizon for prediction is
75 sec. Optimization seeks to minimize the sum of the delays over the horizon.
A forward dynamic programming procedure is used for optimization. Intersec-
tion controllers simulate the outputs over the horizon using the link outputs
and off-line determined turning proportions. Intersection controllers commu-
nicate with each other to achieve a better arrival forecast for the downstream
intersection. The control structure at the network level is a decentralized one.
 Ch. 11. ITS and Traffic Management 753

OPAC (Optimization Policies for Adaptive Control). OPAC (Gartner, 1991)

was first developed in the United States using dynamic programming to gen-
erate optimal control strategies. It provides the computation of signal timing
without requiring fixed cycle time, split, and offset in the conventional sense,
and it is constrained only by minimum and maximum green times. OPAC cal-
culates, in real time, near-optimal signal timings using on-line data that is
typically readily available from upstream detectors at local level and OPAC
supports system-wide coordination at the network level. Many developments
have been carried out over the years (Valdes and Paz, 2004).

4.4.4 Performance
A variety of studies have been undertaken in different locations seeking
to compare the performance of alternative control systems. Early compar-
isons were between isolated and coordinated forms of control. Results would
be expected to be highly dependent on network characteristics, so that, co-
ordination should be most favorable on arterial routes with closely spaced
traffic signals. Probably the most detailed surveys were undertaken in Glas-
gow, where fixed time co-ordination was found to reduce vehicle journey times
by some 16% on average compared to isolated control (Holroyd and Hillier,
1979).
Further comparisons by the UK Transport Research Laboratory have found
that the SCOOT UTC system offers delay savings of around 12% compared to
up-to-date fixed-time plans and up to 40% in peak periods in networks oper-
ating under isolated vehicle actuated control (McDonald and Hounsell, 1991).
A 4% annual increase in delay has also been reported for fixed-time plans
if not updated (Bell and Bretherton, 1986), so that the potential benefits of
traffic-responsive systems would then be higher.
Performance of the other systems described in Sections 4.4.1 and 4.4.2 have
also generally been evaluated through “before-and-after” studies. For exam-
ple, surveys of UTOPIA in Turin gave reductions in journey times of 20% for
public transport vehicles and 10–15% for other vehicles. Good results have
also been reported for SCATS and PRODYN. However, there is very little ev-
idence of the comparative performance of the systems described on the same
network.

4.5 Discussion

4.5.1 Operational Research (OR) techniques

The significant increase in real-time information availability on traffic states
in recent years, driven by advances in technology for detection, communica-
tions, and data processing, opens up exciting opportunities for further OR
applications. It is beyond the scope of this paper to discuss these in detail,
but opportunities are evident even from a sample of OR-related techniques
already being used, such as:
• Short-term prediction/forecasting (e.g., evolution of traffic states).
754 M. Papageorgiou et al.

• Data fusion (data increasingly available from different sources).

• Closed and open-loop control theory applications.
• Dynamic programming for optimization.
• Advanced control theory applications.
• Prediction methods for platoon dispersion, time-dependent queuing,
etc.
• Real-time simulation modeling and network analysis.
• Optimization methods (e.g., for signal timings to optimize against in-
creasingly diverse objective functions).
• Applications of fuzzy-logic for modeling/optimization.
• Artificial intelligence and “expert” systems.
Perhaps a key challenge in the coming years will be how to select and use
OR techniques most effectively against a background of changing optimiza-
tion criteria and data provision which, whilst rich, will be inevitably variable in
quantity, quality, and coverage.

4.5.2 Concluding comments

The optimum use of traffic signals for urban network traffic management
and control will continue to be a key issue for City Authorities. This section
therefore concludes with some comments on some of the opportunities and
challenges which can be identified for the coming years.
• Accurate and timely data will remain a key requirement of UTC sys-
tems: In general, the less timely the data, the poorer the control
(Bretherton et al., 2004). However, this can impose a considerable cost
burden on detection/communications, and advantage is yet to be fully
taken of new above-ground systems. Good real-time modeling of con-
gested situations remains an important component of effective UTC
(Jhaveri et al., 2003). This would seem to be a priority area for scien-
tific research, given the increase in congestion occurring in many towns
and cities.
• UTC systems will increasingly have to provide flexibility in control
strategy selection, including priorities for public transport, pedestrians,
and other road user groups. This will have an implication for optimiza-
tion methods and criteria.
• Integration of UTC with other physical and ITS-related urban traffic
management systems can offer significant benefits for the network and
will be a key requirement for the coming years.

5 Motorway traffic control

5.1 The control loop

Controlling the motorway traffic flow process is a highly complicated task

which may involve a variety of spatially distributed control measures such as
 Ch. 11. ITS and Traffic Management 755

ramp metering, route guidance, variable speed limits, etc. The way the con-
trol measures behave and act on the traffic process stems from the specific
design of the control strategy used. The control strategy employed determines
the control actions, and the specific response to the prevailing traffic condi-
tions, through the available control actuators, is based on its design and on
pre-specified goals.
Figure 6 depicts the general control loop for the motorway network traffic
process which includes all technical and physical phenomena that should be
influenced according to the specific goals. The evolution of the traffic process
depends upon the control inputs and the process disturbances. The control in-
puts are directly related to corresponding control devices such as traffic lights,
variable message signs, variable direction signs, etc., and may be selected from
an admissible control region subject to technical, physical, and operational
constraints. The process disturbances cannot be manipulated, but may pos-
sibly be measurable (e.g., demand) or detectable (e.g., incident) or predictable
over a future time horizon with appropriate algorithms. Typical disturbances in
motorway traffic are traffic demands, origin–destination patterns, the drivers’
compliance to variable message signs, environmental conditions, and incidents.
The process outputs are quantities chosen to represent the performance as-
pects of interest, e.g., total time spent, queue lengths, etc. The estimation of the
traffic state and the prediction of the various traffic quantities are performed
based on real-time measurements taken from the traffic process, and are sub-
sequently fed to the control strategy. The control strategy determines, based

Fig. 6. Motorway traffic flow process under control.

756 M. Papageorgiou et al.

on the measured, estimated, and predicted quantities, the appropriate control

inputs which are fed to the traffic process so as to meet the specified goals
despite the impact of various disturbances.

5.2 Control strategy design

5.2.1 Ramp metering control strategies

General remarks. Ramp metering is the most direct and efficient way to con-
trol and upgrade motorway traffic. Various positive effects are achievable if
ramp metering is appropriately applied:
• Increase in mainline throughput due to avoidance or reduction of con-
gestion.
• Increase in the served volume due to avoidance of blocked off-ramps
or motorway interchanges.
• Utilization of possible reserve capacity on parallel arterials.
• Improved traffic safety due to reduced congestion and safer merging.

Fixed-time ramp metering strategies. Fixed-time ramp metering strategies are

derived off-line for particular times-of-day, based on constant historical de-
mands, without the use of real-time measurements. They are usually based on
simple static models (Wattleworth, 1965; Schwartz and Tan, 1977).
As an objective criterion, one may wish to maximize the number of served
vehicles (which is equivalent to minimizing the total time spent), or to max-
imize the total travel distance, or to balance the ramp queues. These formu-
lations lead to linear programming or quadratic programming problems that
may be readily solved by use of broadly available computer codes. An exten-
sion of these methods that renders the static model dynamic by introduction of
constant travel times for each section was suggested in Papageorgiou (1980).
The main drawback of fixed-time ramp metering strategies is that their set-
tings are based on historical rather than real-time data. This may be a rude
simplification because:
• Demands are not constant, even within a time-of-day.
• Demands may vary at different days, e.g., due to special events.
• Demands change in the long term leading to “aging” of the optimized
settings.
• Turning movements are also changing in the same ways as demands; in
addition, turning movements may change due to the drivers’ response
to the new optimized signal settings, whereby they try to minimize their
individual travel times.
• Incidents and further disturbances may perturb traffic conditions in
a nonpredictable way.
In addition, fixed-time ramp metering strategies may lead (due to the ab-
sence of real-time measurements) either to overload of the mainstream flow
(congestion) or to underutilization of the motorway.
 Ch. 11. ITS and Traffic Management 757

5.2.2 Reactive ramp metering strategies

Reactive ramp metering strategies are employed at a tactical level, i.e., in
the aim of keeping the motorway traffic conditions close to pre-specified set
values, based on real-time measurements.

Local ramp metering. Local ramp metering strategies make use of traffic mea-
surements in the vicinity of a ramp to calculate suitable ramp metering values.
The demand-capacity strategy (Masher et al., 1975), quite popular in North
America, reads
qcap − qin (k − 1) if oout (k)  ocr 
r(k) = (23)
rmin else
where (Figure 7) k is the discrete time index, qcap is the motorway capacity
downstream of the ramp, qin is the motorway flow measurement upstream of
the ramp, oout is the motorway occupancy measurement downstream of the
ramp, ocr is the critical occupancy (at which the motorway flow becomes maxi-
mum), and rmin is a pre-specified minimum ramp flow value. The strategy (23)
attempts to add to the measured upstream flow qin (k − 1) as much ramp flow
r(k) as necessary to reach the downstream motorway capacity qcap . If, how-
ever, for some reason, the downstream measured occupancy oout (k) becomes
overcritical (i.e., a congestion may form), the ramp flow r(k) is reduced to the
minimum admissible flow rmin to avoid or to dissolve the congestion.
Comparing the control problem in hand with Figure 6, it becomes clear that
the ramp flow r is a control input, the downstream occupancy oout is an output,
while the upstream motorway flow qin is a disturbance. Hence, (23) does not
really represent a closed-loop strategy but an open-loop disturbance-rejection
policy (Figure 7(a)) which is generally known to be quite sensitive to various
further nonmeasurable disturbances.
The occupancy strategy (Masher et al., 1975) is based on the same philos-
ophy as the demand-capacity strategy, but it relies on occupancy-based esti-
mation of qin , which may, under certain conditions, reduce the corresponding
implementation cost.
An alternative, closed-loop ramp metering strategy (ALINEA) (Figu-
re 7(b)), suggested in Papageorgiou et al. (1991), reads
r(k) = r(k − 1) + KR ô − oout (k)  (24)
where KR > 0 is a regulator parameter and ô is a set (desired) value for the
downstream occupancy (typically, but not necessarily, ô = ocr may be set, in
which case the downstream motorway flow becomes close to qcap ). In field ap-
plications, ALINEA has not been very sensitive to the choice of the regulator
parameter KR .
Note that the demand-capacity strategy reacts to excessive occupancies oout
only after a threshold value (ocr ) is exceeded, and in a rather crude way, while
ALINEA reacts smoothly even to slight differences ô − oout (k), and thus it
may prevent congestion by stabilizing the traffic flow at a high throughput level.
758 M. Papageorgiou et al.

(a)

(b)
Fig. 7. Local ramp metering strategies. (a) Demand–capacity, (b) ALINEA.

The set value may be changed any time, and thus ALINEA may be embedded
into a hierarchical control system with set values of the individual ramps being
specified in real time by a superior coordination level or by an operator.
Comparative field trials have been conducted in various countries to as-
sess and compare the efficiency of local ramp metering strategies (see, e.g.,
Papageorgiou et al., 1998), such as the demand-capacity, ALINEA, and the
occupancy strategy. The field results clearly show ALINEA’s superiority for all
employed performance criterions.

Multivariable regulator strategies. Multivariable regulators for ramp metering

pursue the same goals as local ramp metering strategies: they attempt to op-
 Ch. 11. ITS and Traffic Management 759

erate the motorway traffic conditions near some pre-specified set (desired)
values. While local ramp metering is performed independently for each ramp,
based on local measurements, multivariable regulators make use of all avail-
able mainstream measurements oi (k), i = 1     n, on a motorway stretch, to
calculate simultaneously the ramp volume values ri (k), i = 1     m, for all
controllable ramps included in the same stretch (Papageorgiou et al., 1990).
This provides potential improvements over local ramp metering because of
more comprehensive information provision and because of coordinated con-
trol actions. Multivariable regulator approaches to ramp metering have been
reported in Yuan and Kreer (1968), Young et al. (1997), and Benmohamed
and Meerkov (1994). The multivariable regulator strategy METALINE may
be viewed as a generalization and extension of ALINEA, whereby the metered
on-ramp volumes are calculated from
r(k) = r(k − 1) − K1 o(k) − o(k − 1) + K2 O − O(k)  (25)
where r = [r1      rm ] is the vector of m controllable on-ramp volumes,
o = [o1      on ] is the vector of n measured occupancies on the motorway
stretch, O = [O1      Om ] is a subset of o that includes m occupancy lo-
cations for which pre-specified set values O = [O1      Om ] may be given.
Note that for control-theoretic reasons the number of set-valued occupancies
cannot be higher than the number of controlled on-ramps. Typically one bottle-
neck location downstream of each controlled on-ramp is selected for inclusion
in the vector O. Finally, K1 and K2 are the regulator’s constant gain matrices
that must be suitably designed via an LQ procedure, see Papageorgiou et al.
(1990), and Diakaki and Papageorgiou (1994), for details.

Nonlinear optimal ramp metering strategies. Reactive ramp metering strategies

may be helpful to a certain extent, but, first, they need appropriate set values,
and, second, their character is more or less local. What is needed for motorway
networks or long stretches is a superior coordination level that calculates in
real time optimal set values from a proactive, strategic point of view. Such an
optimal control strategy should explicitly take into account:
• Demand predictions over a sufficiently long time horizon.
• The current traffic state both on the motorway and on the on-ramps.
• The limited storage capacity of the on-ramps.
• The ramp metering constraints regarding maximum queues allowed.
• The nonlinear traffic flow dynamics, including the infrastructure’s lim-
ited capacity.
• Any incidents currently present in the motorway network.
Based on this comprehensive information, the control strategy should deliver
set values for the overall motorway network over a future time horizon so as
• to respect all present constraints,
• to minimize an objective criterion such as the total time spent in the
whole network including the on-ramps.
760 M. Papageorgiou et al.

Such a comprehensive dynamic optimal control problem may be formulated

and solved with moderate computation time by use of suitable solution algo-
rithms. The nonlinear traffic dynamics may be expressed by use of suitable
dynamic models in state space form, where the state vector comprises all traffic
densities and mean speeds of motorway segments, as well as all ramp queues;
the control vector comprises all controllable ramp volumes; the disturbance
vector comprises all on-ramp demands and turning rates at bifurcations. The
problem’s constraints include the ramp metering constraints and the queue
constraints, see Section 5.3.
Thus, for given current (initial) state from corresponding measurements and
given demand predictions, the problem consists in specifying the ramp flows
r(k), k = 0     K − 1, where K is the considered horizon, so as to minimize
the total time spent (or some other criterion) subject to the nonlinear traffic
flow dynamics and the constraints, see Section 5.3.
This problem or variations thereof was considered and solved in various
works (Blinkin, 1976; Kotsialos et al., 2002). Although simulation studies indi-
cate substantial savings of travel time and substantial increase of throughput,
advanced control strategies of this kind have not been implemented in the field
as yet. Section 5.3 contains a simulation study for such an advanced coordi-
nated ramp metering control strategy.

5.2.3 Link control strategies

Link control may include one or a combination of the following actions:
• Variable speed limitation.
• Changeable message signs with indications for “keep lane”, or conges-
tion warning, or environmental warning (e.g., information about the
pavement state).
• Lane control.
• Incident warning.
• Reversable flow lanes (tidal flow).
There are many motorway stretches, particularly in Germany and in the
Netherlands, employing a selection of these measures. It is generally thought
that control measures of this kind lead to a homogenization of traffic flow (i.e.,
more homogeneous speeds of cars within a lane and of average speeds of dif-
ferent lanes) which is believed to reduce the risk of falling into congestion at
high traffic densities and to increase the motorway’s capacity. Very few sys-
tematic studies have been conducted to quantify the impact of these control
measures (see, e.g., Zackor, 1972; Smoulders, 1990) and corresponding val-
idated mathematical models are currently lacking. This is one of the reasons
why the corresponding control strategies of operating systems are of a heuristic
character (e.g., Bode and Haller, 1983; Zackor and Balz, 1984).

5.2.4 Route guidance control strategies

A route guidance system may be viewed as a traffic control system in the
sense of Figure 6. Based on real-time measurements, sufficiently interpreted
 Ch. 11. ITS and Traffic Management 761

and extended within the surveillance block, a control strategy decides about
the routes to be recommended (or the information to be provided) to the road
users. This, on its turn, has an impact on the traffic flow conditions in the net-
work, and this impact is reflected in the performance indices. Because of the
real-time nature of the operation, requirements of short computation times
are relatively strict (for more details see Section 3).

5.2.5 Integrated motorway network traffic control

As mentioned earlier, modern motorway networks may include different
types of control measures. The corresponding control strategies are usually de-
signed and implemented independently, thus failing to exploit the synergistic
effects that might result from coordination of the respective control actions. An
advanced concept for integrated motorway network control results from suit-
able extension of the optimal control approach outlined above. More precisely,
the dynamic model of motorway traffic flow may be extended to enable the
inclusion of further control measures, beyond the ramp metering rates r(k).
Formally r(k) is then replaced by a general control input vector u(k) that
comprises all implemented control measures of any type. Such an approach
was implemented in the integrated motorway network control tool AMOC
(Advanced Motorway Optimal Control) in Kotsialos et al. (1999), where ramp
metering and route guidance are considered simultaneously with promising re-
sults, see also Moreno-Banos et al. (1993), Ataslar and Iftar (1998), Bellemans
(2003), Hegyi et al. (2003), and Hegyi (2004).

5.3 An advanced example

5.3.1 The motorway network traffic model

The efficiency and the amelioration potential of nonlinear optimal ramp
metering strategies may be demonstrated by means of simulation for a large-
scale network with the use of the AMOC generic motorway network control
tool. In this case AMOC does not consider routing control measures, but only
ramp metering control actions.
The network is represented by a directed graph whereby the links of the
graph represent motorway stretches. Each motorway stretch has uniform char-
acteristics, i.e., no on-/off-ramps and no major changes in geometry. The nodes
of the graph are placed at locations where a major change in road geometry oc-
curs, as well as at junctions, on-ramps, and off-ramps.
The time and space arguments are discretized. The discrete-time step is
denoted by T . A motorway link m is divided into Nm segments of equal
length Lm . Each segment i of link m at time instant t = kT , k = 0    K, is
characterized by the macroscopic variables traffic density (see also Section 2)
ρmi (k) (veh/lane-km), mean speed vmi (k) (km/h), and traffic volume or
flow qmi (k) (veh/h). The basic equations used for their calculation for each
segment i of link m at each time step, are (this is a time-space discretized
762 M. Papageorgiou et al.

Payne-like model, see Section 2.1.6)

T [qmi−1 (k) − qmi (k)]
ρmi (k + 1) = ρmi (k) +  (26)
Lm λm
qmi (k) = ρmi (k)vmi (k)λm  (27)
T {V [ρmi (k)] − vmi (k)}
vmi (k + 1) = vmi (k) +
τ
T [vmi−1 (k) − vmi (k)]vmi (k)
+
Lm
νT ρmi+1 (k) − ρmi (k)
−  (28)
τLm ρmi (k) + κ
  
1 ρmi (k) am
V ρmi (k) = vfm exp −  (29)
am ρcrm
where vfm denotes the free-flow speed of link m, ρcrm denotes the critical
density per lane of link m (the density where the maximum flow in the link
occurs), λm its number of lanes, and am is a parameter of the fundamental
diagram (Equation (29)) of link m. Furthermore, τ, a time constant, ν, an an-
ticipation constant, and κ, are constant parameters same for all network links.
Additionally, it is assumed that the mean speed resulting from (27) is limited
from below by the minimum speed in the network vmin .
In order for the speed calculation to take into account the speed decrease
caused by merging phenomena and the speed reduction due to weaving phe-
nomena, resulting from lane drops in the mainstream, two additional terms are
added to (27), see Messmer and Papageorgiou (1990).
For origin links, i.e., links that receive traffic demand and forward it into the
motorway network, a simple queue model is used.
wo (k + 1) = wo (k) + T do (k) − qo (k)  (30)
where wo (k) is the queue length (veh) in origin o during period k, do (k) is the
demand (veh/h) at o at the same period, and qo (k) is the flow (veh/h) that en-
ters the mainstream. The outflow qo (k) is determined by the traffic conditions
on the mainstream link and possible ramp metering control measures applied.
If ramp metering is applied, then the outflow q̂o (k) that is allowed to leave o
during period k is a portion po (k) of the outflow that would leave o without
control.
qo (k) = po (k)q̂o (k) (31)
where po (k) ∈ [pmino  1] is the metering rate for the origin link o, i.e., a con-
trol variable. If po (k) = 1, no ramp metering is applied, else po (k) < 1. For
q̂o (k) we have
 
q̂o (k) = min q̂o1 (k) q̂o2 (k)  (32)
 Ch. 11. ITS and Traffic Management 763

with
wo (k)
q̂o1 = do (k) +  (33)
 T
Qo if ρμ (k) < ρcrμ 
q̂o2 = ρ (k)−ρcrμ (34)
Qo 1 − μ1
ρmax −ρcrμ if ρμ (k)  ρcrμ 

where Qo is the on-ramp’s capacity (veh/h), and ρmax (veh/lane-km) is the

maximum density in the network. Thus the maximum outflow q̂o (k) is deter-
mined by the current origin demand if q̂o1 < q̂o2 (see (32), (33)), or the
geometrical ramp capacity Qo if the mainstream density is undercritical, i.e.,
ρμ1 (k) < ρcrμ (see (34)), or the reduced capacity due to congestion of the
mainstream, i.e., ρμ1 (k) > ρcrμ (see (34)).
Motorway bifurcations and junctions (including on-ramps and off-ramps)
are represented by nodes. Traffic enters a node n through a number of input
links and is distributed to the output links according to

Qn (k) = qμNμ (k) (35)
μ∈In
qm0 (k) = βmn (k)Qn (k) ∀m ∈ On  (36)
where In is the set of links entering node n, On is the set of links leaving n,
Qn (k) is the total traffic volume entering n at period k, qm0 (k) is the traffic
volume that leaves n via outlink m, and βm n (k) is the portion of Qn (k) that
leaves the node through link m. βm n (k) are the turning rates of node n and
are assumed to be known for the entire time horizon. Equations (35) and (36)
provide qm0 (k) required in (26) for i = 1.
The upstream influence of density and the downstream influence of speed at
network nodes are taken under consideration by appropriate static models that
provide the required terms in (26) and (27) for i = 1 and i = Nm (Messmer
and Papageorgiou, 1990).

5.3.2 The constrained optimal control problem

The coordinated ramp metering control problem is formulated as a dy-
namic optimal control problem with constrained control variables which can
be solved numerically over a given time horizon. The general discrete-time
formulation of the optimal control problem reads:


K−1
minimize J = ϑ[K] + ϕ x(k) u(k) d(k) (37)
k=0
subject to
x(k + 1) = f x(k) u(k) d(k)  x(0) = x0  (38)
uimin  ui (k)  uimax  ∀i = 1     m (39)
764 M. Papageorgiou et al.

where K is the considered time horizon, x ∈ n is the state vector, u ∈ m

is the vector of control variables, d is the vector of disturbances acting on the
traffic process, and ϑ, ϕ are arbitrary, twice differentiable, nonlinear cost func-
tions.
Based on the previous section, it may be seen that by substituting (27), (35),
and (36) into (26); (29) into (27); (31)–(34) into (30), the traffic flow model
equations take the form of Equation (38). In this case the state vector x consists
of the densities ρmi , the mean speeds vmi of every segment i of every link m,
and the queues wo for every origin o. The control vector u consists of the ramp
metering rates po of every on-ramp o under control, with pomin  po (k) 
10 according to (39). Finally, the disturbance vector consists of all demands at
each origin of the network and all turning rates at the network’s bifurcations.
The chosen cost criterion aims at minimizing the Total Time Spent (TTS)
of all vehicles in the network (including the waiting time experienced in the
network queues). The cost criterion is as follows.
  
J=T ρmi (k)Lm λm + wo (k)
k m i o
  
2 2
+ af po (k) − po (k − 1) + aw ψ wo (k)
o o (40)
with
 
ψ wo (k) = max 0 wo (k) − womax  (41)
where the first two term in (40) account for the TTS while af , aw are weighting
factors. The term with weight af is included in the cost criterion to suppress
high-frequency oscillations of the control trajectories. The last additional term
is a penalty term included in the cost criterion in order to enable the control
strategy to limit the queue lengths at the origins if and to the level desired. The
parameters womax are pre-determined constants and express the maximum
permissible number of vehicles in origin o’s queue.
A powerful numerical solution algorithm is used to solve this constrained
discrete-time optimal control problem, see Papageorgiou and Marinaki (1995).

5.3.3 Application results

The previously described approach to network-wide optimal ramp metering
has been applied to the Amsterdam ring-road with the use of AMOC.
The Amsterdam Orbital Motorway (A10) is shown in Figure 8. The A10 si-
multaneously serves local, regional, and inter-regional traffic and acts as a hub
for traffic entering and exiting North Holland. There are four main connec-
tions with other motorways, the A8 at the North, the A4 at the South-West, the
A2 at the South, and the A1 at the South-East. The A10 contains two tunnels,
the Coen Tunnel at the North-West and the Zeeburg Tunnel at the East.
For the purposes of our study only the counter-clockwise direction of the
A10, which is about 32 km long, is considered. There are 21 on-ramps on
 Ch. 11. ITS and Traffic Management 765

Fig. 8. The Amsterdam ring-road.

this motorway, including the connections with the A8, A4, A2, and A1 mo-
torways, and a total number of 20 off-ramps, including the junctions with A4,
A2, A1, and A8. It is assumed that ramp metering may be performed at each
on-ramp, whereby the maximum permissible queue length for the on-ramps
is set to 20 vehicles, while storage of 100 vehicles is permitted on each of the
motorway-to-motorway ramps of A8, A4, A2, and A1.
The model parameters for this network were determined from validation
of the network traffic flow model against real data taken from the motorways
(Kotsialos et al., 1998).
The ring-road was divided in 76 segments with average length 421 m. This
means that the state vector is 173-dimensional (including the 21 on-ramp
queues). Since ramp metering is applied to all on-ramps, the control vector
is 21-dimensional, while the disturbance vector is 43-dimensional. With a time
step T = 10 s we have, for a horizon of 4 h, K = 1440 which results in a large-
scale optimization problem with 279,360 variables.

5.3.4 The no-control case

The ring-road was studied for a time horizon of 4 hours, from 16:00 un-
til 20:00, using realistic historical demands from the site. This time period
includes the evening peak hour. In absence of any control measures, the ring-
road is subject to recurrent congestion that is formed downstream of the
junctions of A10 with A2 and A1 in A10-South. This congestion propagates
backwards causing severe traffic delays in the A10-West. Figure 9(a) depicts
the density propagation along the motorway segments (segment 0 is the first
segment of A10-West after the junction of A10 with A8). The formation of
large queues at the on-ramps can be seen in Figure 9(b) (on-ramp 0 corre-
sponds to A8). As a result, the total time spent over the 4-h-horizon is equal to
13,226 veh h.
766 M. Papageorgiou et al.

(a)

(b)
Fig. 9. No control: (a) Density, (b) on-ramp queues.

5.3.5 The control case

When ramp metering is performed at all on-ramps, the congestion is vir-
tually lifted from the network (Figure 10(a)). The control strategy succeeds
in establishing optimal uncongested traffic conditions on the A10-South and
A10-West by applying ramp metering mainly at A1 and A2 at an early stage.
In Figure 10(b), the queues are mainly occurring at A2 and A1 because these
ramps have larger maximum permissible queues (100 vehicles). The control
trajectories are depicted in Figure 10(c). The resulting total time spent is
8833 veh h, which is a 33.2% improvement compared to the no-control case.
A further improvement to the total time spent could be reached with larger
maximum permissible queues. Had there been no queue constraints at all, the
density profile of Figure 10(a) would be completely flat. In fact, the control
strategy performs a trade-off between the queue lengths and the existence
of congestion inside the network. Stricter queue constraints result in more
degraded traffic conditions inside the motorway due to accordingly reduced
control maneuverability.
 Ch. 11. ITS and Traffic Management 767

(a)

(b)

(c)
Fig. 10. Optimal control. (a) Density, (b) on-ramp queues, (c) optimal ramp metering rates.

The computation time required to obtain the optimal solutions is moderate

and depends upon the search method used. The main part of the improvement
is typically achieved very fast. The computation time for the 4-h-horizon is
20 min for the bulk of the 33.2% improvement (more than 30.6%) on a Sun
Ultra5 with a Sparc IIi-360 MHz processor workstation.
768 M. Papageorgiou et al.

5.4 Future directions

As in many other engineering disciplines, only a small portion of the signif-

icant methodological advancements in motorway network control have really
been exploited in the field. It is beyond our scope to investigate and discuss the
reasons behind this theory–practice gap, but administrative inertia, little com-
petitive pressure in the public sector, the complexity of traffic control systems,
limited realization of the improvement potential behind advanced methods by
the responsible authorities, and limited understanding of practical problems
by some researchers may have a role in this. Whatever the reasons, the major
challenge in the coming decade is the deployment of advanced and efficient
traffic control strategies in the field.
Regarding motorway networks, operational control systems of any kind are
the exception rather than the rule. With regard to ramp metering, the main
focus is frequently not on improving efficiency but on secondary objectives of
different kinds. Most responsible traffic authorities and the decision makers
are far from realizing the fact that advanced real-time ramp metering systems
(employing optimal control algorithms) have the potential of changing dra-
matically the traffic conditions on today’s heavily congested (hence strongly
underutilized) motorways with spectacular improvements that may reach 50%
reduction of the total time spent.

References

Ahmed, K.I., Ben-Akiva, M.E., Koutsopoulos, H.N., Mishalani, R.G. (1996). Models of freeway lane
changing and gap-acceptance behaviour. In: Lesort, J.B. (Ed.), Transportation and Traffic Theory:
Proceedings 13th International Symposium of Transportation and Traffic Theory. Pergamon/Elsevier,
Lyon, France, pp. 505–515.
Al-Deek, H., Kanafani, A. (1993). Modeling the benefits of advanced traveler information systems in
corridors with incidents. Transportation Research C 1 (4), 303–324.
Ataslar, B., Iftar, A. (1998). A decentralized control approach for transportation networks. In: Preprints
of the 8th IFAC Symposium on Large Scale Systems, vol. 2. Patra, Greece, pp. 348–353.
Bando, M., Hasebe, K., Nakayama, A., Shibata, A., Sugiyama, Y. (1995). Dynamical model of traffic
congestion and numerical simulation. Physical Review E 51, 1035–1042.
Bell, M.C., Bretherton, R.D. (1986). Ageing of fixed-time traffic signal plans. In: Proceedings of the 2nd
International Conference on Road Traffic Control. Institution of Electrical Engineers, London, UK.
Bellemans, T. (2003). Traffic control on motorways. PhD thesis, Katholieke Universiteit Leuven, Leu-
ven, Belgium, May.
Ben-Akiva, M., De Palma, A., Kaysi, I. (1991). Dynamic network models and driver information sys-
tems. Transportation Research A 25 (5), 251–266.
Ben-Akiva, M., De Palma, A., Kaysi, I. (1996). The impact of predictive information on guidance effi-
ciency: An analytical approach. In: Bianco, L., Toth, I. (Eds.), Advanced Methods in Transportation
Analysis. Springer-Verlag, New York, pp. 413–432.
Benmohamed, L., Meerkov, S.M. (1994). Feedback control of highway congestion by a fair on-ramp
metering. In: Proceedings of the 33rd IEEE Conference on Decision and Control, vol. 3. Lake Buena
Vista, Florida, pp. 2437–2442.
Blinkin, M. (1976). Problem of optimal control of traffic flow on highways. Automation and Remote
Control 37, 662–667.
 Ch. 11. ITS and Traffic Management 769

Bode, K.R., Haller, W. (1983). Geschwindigkeitssteuerung auf der A7 zwischen den Autobahn-
dreiecken Hannover-Nord und Walsrode. Strassenverkehrstechnik 27, 145–151 (in German).
Bolelli, A., Mauro, V., Perono, E. (1991). Models and strategies for dynamic route guidance, Part
B: A decentralized, fully dynamic, infrastructure supported route guidance. In: Proceedings of the
DRIVE Conference in Advanced Telematics in Road Transports, Brussels, Belgium, pp. 99–105.
Bonsall, P., Palmer, I. (1999). Route choice in response to variable message signs: Factors affecting
compliance. In: Emmerink, R., Nijkamp, P. (Eds.), Behavioural and Network Impacts of Driver Infor-
mation Systems. Ashgate Publishing Company, pp. 181–214. Chapter 9.
Bottom, J. (2000). Consistent anticipatory route guidance. PhD thesis, Dept. of Civil and Environmen-
tal Engineering, Massachusets Institute of Technology, Cambridge, MA, USA.
Bovy, P.H.L., Hoogendoorn, S.P. (2000). Ill-predictability of road-traffic congestion. In: Bell, M.G.H.,
Cassir, C. (Eds.), Reliability of Transport Networks. Research Studies Press, pp. 43–54.
Bovy, P.H., van der Zijpp, N.J. (1999). The close connection between dynamic traffic management and
driver information systems. In: Emmerink, R., Nijkamp, P. (Eds.), Behavioural and Network Impacts
of Driver Information Systems. Ashgate Publishing Company, pp. 355–370.
Brackstone, M., McDonald, M. (1999). Car-following: A historical review. Transportation Research F 2,
181–196.
Bretherton, D., Bowen, G., Wood, K. (2003). Effective urban traffic management and control – recent
developments in SCOOT. In: Proceedings of the 82nd TRB Annual Meeting, Washington, DC.
Bretherton, D., Bodger, M., Baber, N. (2004). SCOOT – the future. In: Proceedings of the 83rd TRB
Annual Meeting. Washington, DC.
Bretherton, R., Wood, K., Bowen, G.T. (1998). SCOOT version 4. In: Proceedings of 9th International
Conference on Road Transport Information and Control. Institution of Electrical Engineers, London,
UK.
Busch, F. (1996). Traffic telematics in urban and regional environments. In: Proceedings of the Intertraffic
Conference. Amsterdam, The Netherlands.
Catling, I. (1989). AUTOGUIDE – electronic route guidance in the UK. In: Perrin, J.-P. (Ed.), Con-
trol, Computers and Communications in Transportation. Pergamon Press, Paris, France, pp. 269–276.
Selected Papers from the IFAC/IFIP/IFORS Symposium.
Chandler, R.E., Herman, R., Montroll, E.W. (1958). Traffic dynamics: Studies in car following. Opera-
tions Research 6, 165–184.
Charbonnier, C., Farges, J.-L., Henry, J.-J. (1991). Models and strategies for dynamic route guidance,
Part C: Optimal control approach. In: Proceedings of the DRIVE Conference in Advanced Telematics
in Road Transport, Brussels, Belgium, pp. 106–112.
Cremer, M., Papageorgiou, M. (1981). Parameter identification for a traffic flow model. Automatica 17,
837–843.
Daganzo, C.F. (1994). The cell transmission model: A dynamic representation of highway traffic con-
sistent with the hydrodynamic theory. Transportation Research B 28 (4), 269–287.
Daganzo, C.F. (1995). Requiem for second-order fluid approximations of traffic flow. Transportation
Research B 29, 277–286.
Daganzo, C.F. (1997). A continuum theory of traffic dynamics for freeways with special lanes. Trans-
portation Research B 31 (2), 83–102.
Daganzo, C.F. (2002a). A behavioural theory of multi-lane traffic flow. Part I: Long homogeneous free-
way sections. Transportation Research B 36 (2), 131–158.
Daganzo, C.F. (2002b). A behavioural theory of multi-lane traffic flow. Part II: Merges and the on-set
of congestion. Transportation Research B 36 (2), 159–169.
Department of the UK Environment, Transport and the Regions (DETR) (1999). The ’SCOOT’ urban
traffic control system. Traffic Advisory Leaflet 7/99, London, UK.
Diakaki, C., Papageorgiou, M. (1994). Design and simulation test of coordinated ramp metering control
(METALINE) for A10-west in Amsterdam. Internal report 1994-2, Dynamic Systems and Simula-
tion Laboratory, Technical University of Crete, Chania, Greece.
Diakaki, C., Papageorgiou, M., McLean, T. (2000). Integrated traffic-responsive urban corridor control
strategy in Glasgow, Scotland – application and evaluation. Transportation Research Record 1727,
101–111.
770 M. Papageorgiou et al.

Diakaki, C., Dinopoulou, V., Aboudolas, K., Papageorgiou, M., Ben-Shabat, E., Seider, E., Leibov, E.
(2003). Extensions and new applications of the traffic signal control strategy TUC. Transportation
Research Record 1856, 202–211.
Donati, F., Mauro, V., Roncoloni, G., Vallauri, M. (1984). A hierarchical decentralized traffic light
control system-the first realisation: Progetto Torino. In: Proceedings of the 9th World Congress of the
International Federation of Automotive Control, Budapest, Hungary, pp. 2853–2858.
Edie, L.C., Foote, R.S. (1958). Traffic flow in tunnels. Highway Research Board Proceedings 37, 334–344.
Edie, L.C., Foote, R.S. (1960). Effect of shock waves on tunnel traffic flow. Highway Research Board
Proceedings 39, 492–505.
Emmerink, R., Axhausen, K.W., Nijkamp, P., Rietveld, P. (1995). The potential of information provision
in a simulated road transport network with non-recurrent congestion. Transportation Research C 3
(5), 293–309.
Emmerink, R., Nijkamp, P., Rietveld, P., Van Ommeren, J.N. (1996). Variable message signs and radio
traffic information: An integrated empirical analysis of drivers’ route choice behavior. Transporta-
tion Research A 30 (2), 135–153.
Engelson, L. (1997). Self-fulfilling and recursive forecasts – an analytical perspective for driver infor-
mation systems. In: Preprints from the 8th Meeting of the International Association for Travel Behavior
Research (IATBR ’97). Workshop on Dynamics and ITS Response, Austin, TX, USA.
Esser, J., Neubert, L., Wahle, J., Schreckenberg, M. (1999). Microscopic online simulations of urban
traffic. In: Ceder, A. (Ed.), Transportation and Traffic Theory. Proceedings of the 14th International
Symposium of Transportation and Traffic Theory, Jerusalem, Israel. Pergamon/Elsevier, pp. 517–
534.
Farges, J.L., Khoudour, K., Lesort, J.B. (1990). PRODYN: On site evaluation. In: 3rd International
Conference on Road Traffic Control. Institute of Electrical Engineers, London, UK, pp. 62–66.
Ferrari, P. (1989). The effect of driver behaviour on motorway reliability. Transportation Research B 23
(2), 139–150.
Forbes, T.W., Zagorski, H.J., Holshouser, E.L., Deterline, W.A. (1958). Measurement of driver reac-
tions to tunnel conditions. Highway Research Board Proceedings 37, 345–357.
Gartner, N.N. (1991). Road traffic control: Demand responsive. In: Papageorgiou, M. (Ed.), Concise
Encyclopedia of Traffic and Transportation Systems. Pergamon Press, pp. 386–391.
Hall, R.W. (1996). Route choice and advanced traveler information systems on a capacitated and dy-
namic network. Transportation Research C 4 (5), 289–306.
Head, K.L., Mirchandani, P.B., Sheppard, D. (1992). Hierarchical framework for real time traffic con-
trol. Transportation Research Record 1360, 82–88.
Hegyi, A. (2004). Model Predictive Control for Integrating Traffic Control Measures. PhD thesis, Delft
University of Technology, Delft, The Netherlands.
Hegyi, A., De Schutter, B., Hellendoorn, J. (2003). MPC-based optimal coordination of variable speed
limits to suppress shock waves in freeway traffic. In: Proceedings of the 2003 American Control Con-
ference. Denver, CO, USA, pp. 4083–4088.
Helbing, D. (1997). Verkehrsdynamik – neue physikalische Modellierungskonzepte. Springer-Verlag (in
German).
Helbing, D., Hennecke, A., Shvetsov, V., Treiber, M. (2001). MASTER: Macroscopic traffic simulation
based on gas-kinetic, non-local traffic model. Transportation Research B 35 (2), 183–211.
Helly, W. (1959). Simulation of bottlenecks in single lane traffic flow. In: Proceedings Symposium on
Theory of Traffic Flow, pp. 207–238.
Henn, V. (1995). Utilisation de la logique floue pour la modelisation microscopique du trafic routier.
La Revue du Logique Floue (in French).
Herman, R., Montroll, E.W., Potts, R., Rothery, R.W. (1959). Traffic dynamics: Analysis of stability in
car-following. Operations Research 1 (7), 86–106.
Hockney, R.W., Eastwood, J.W. (1988). Computer Simulations Using Particles. Hilger, Bristol, NY.
Hoffman, G. (1991). Up-to-the-minute information as we drive – how it can help road users and traffic
management. Transport Reviews 11 (1), 41–61.
Holroyd, J., Hillier, J.A. (1979). The Glasgow experiment: PLIDENT and after. Technical Report
LR384, Transport and Road Research Laboratory, Crowthorne, UK.
 Ch. 11. ITS and Traffic Management 771

Hoogendoorn, S.P., Bovy, P.H.L. (1998). Optimal routing control using variable message signs. In: Bovy,
P.H.L. (Ed.), Motorway Flow Analysis: New Methodologies and Recent Empirical Findings. Delft Univ.
Press, pp. 237–263.
Hoogendoorn, S.P., Bovy, P.H.L. (1999). Multiclass macroscopic traffic flow modelling: A multilane
generalisation using gas-kinetic theory. In: Ceder, A. (Ed.), Transportation and Traffic Theory. Pro-
ceedings of the 14th International Symposium of Transportation and Traffic Theory, Jerusalem,
Israel. Pergamon/Elsevier, pp. 27–50.
Hoogendoorn, S.P., Bovy, P.H.L., van Lint, J.W.C. (2002). Short-term prediction of traffic flow condi-
tions in a multilane multiclass network. In: Taylor, M.A.P. (Ed.), Transportation and Traffic Theory
in the 21st Century: Proceedings of the 15th International Symposium on Transportation and Traffic
Theory. Pergamon/Elsevier, Oxford, pp. 625–651.
Hunt, P.B., Robertson, R.D., Bretherton, R.D., Winton, R.I. (1981). SCOOT – a traffic responsive
method of co-ordinating signals. Technical Report 1014, Transport and Road Research Laboratory.
Jeffery, D.J. (1981). The potential benefits of route guidance. Technical Report LR997, Transport and
Road Research Laboratory.
Jhaveri, C.S., Perrin, J., Martin, P.T. (2003). SCOOT adaptive signal control: An evaluation of its ef-
fectiveness over a range of congestion intensities. In: Proceedings of the 82nd TRB Annual Meeting.
Washington, DC.
Kantowitz, B.H., Hanowski, R.J., Kantowitz, S.C. (1997). Driver acceptance of unreliable traffic infor-
mation in familiar and unfamiliar settings. Human Factors 39 (2), 164–176.
Kaufman, D., Smith, R., Wunderlich, K. (1998). User-equilibrium properties of fixed points in iterative
dynamic routing/assignment methods. Transportation Research C 6 (1–2), 1–16.
Kaysi, I., Ben-Akiva, M., De Palma, A. (1995). Design aspects of advanced traveler information systems.
In: Gartner, N., Improta, G. (Eds.), Urban Traffic Networks: Dynamic Flow Modeling and Control.
Springer-Verlag, pp. 59–81.
Kerner, B.S. (1999). Theory of congested traffic flow: Self-organization without bottlenecks. In: Ceder,
A. (Ed.), Transportation and Traffic Theory. Proceedings of the 14th International Symposium of
Transportation and Traffic Theory, Jerusalem, Israel. Pergamon/Elsevier, pp. 147–172.
Kerner, B.S., Rehborn, H. (1997). Experimental properties of phase-transitions in traffic flow. Physical
Review Letters 79 (20), 4030–4033.
Kerner, B.S., Aleksic, M., Rehborn, H. (2000). Automatic tracing and forecasting of moving traffic jams
using predictable features of congested traffic flow. In: Schnieder, E., Becker, U. (Eds.), Control in
Transportation Systems 2000: Proceedings of the 9th IFAC Symposium on Control in Transportation
Systems. Braunschweig, Germany, pp. 501–506.
Khattak, A., Polydoropoulou, A., Ben-Akiva, M. (1996). Modeling revealed and stated pre-trip travel
response to ATIS. Transportation Research Record 1537, 46–54.
Khattak, A.J., Yim, Y., Stalker, L. (1999). Does travel information influence commuter and noncom-
muter behavior? Results from the San Francisco Bay Area TravInfo project. Transportation Research
Record 1694, 48–58.
Kikuchi, C., Chakroborty, P. (1992). Car following model based on a fuzzy inference system. Transporta-
tion Research Record 1365 (1992), 82–91.
Klar, A., Wegener, R. (1999a). A hierarchy of models for multilane vehicular traffic I: Modelling. SIAM
Journal of Applied Mathematics 59 (3), 983–1001.
Klar, A., Wegener, R. (1999b). A hierarchy of models for multilane vehicular traffic II: Numerical
investigations. SIAM Journal of Applied Mathematics 59 (3), 1002–1011.
Kotsialos, A., Pavlis, Y., Middelham, F., Diakaki, C., Vardaka, G., Papageorgiou, M. (1998). Modelling
of the large scale motorway network around amsterdam. In: Preprints of the 8th IFAC Symposium on
Large Scale Systems, vol. 2. Patra, Greece, pp. 354–360.
Kotsialos, A., Papageorgiou, M., Messmer, A. (1999). Optimal coordinated and integrated motorway
network traffic control. In: Ceder, A. (Ed.), Transportation and Traffic Theory. Proceedings of the
14th International Symposium of Transportation and Traffic Theory (ISTTT), Jerusalem, Israel.
Pergamon/Elsevier, pp. 621–644.
Kotsialos, A., Papageorgiou, M., Mangeas, M., Haj-Salem, H. (2002). Coordinated and integrated con-
trol of motorway networks via nonlinear optimal control. Transportation Research C 10 (1), 65–84.
772 M. Papageorgiou et al.

Koutsopoulos, H.N., Lotan, T. (1989). Effectiveness of motorist information systems in reducing traffic
congestion. In: Proceedings of the Conference on Vehicle Navigation and Information Systems (VNIS),
Toronto, Canada, pp. 275–281.
Kraan, M., Mahmassani, H.S., Huynh, N. (2000). Interactive survey approach to study traveler re-
sponses to ATIS for shopping trips. In: Proceedings of the 79th Transportation Research Board Annual
Meeting. Washington, DC, USA.
Kühne, R.D. (1991). Traffic patterns in unstable traffic flow on freeways. In: Brannolte, U. (Ed.),
Highway Capacity and Level of Service. Proceedings of the International Symposium on Highway
Capacity, Karlsruhe, 24–27 July 1991. Balkema, Rotterdam, pp. 211–223.
Lebacque, J.P. (2002). A two-phase extension of the LWR model based on the boundness of traffic
acceleration. In: Taylor, M.A.P. (Ed.), Transportation and Traffic Theory in the 21St Century. Proceed-
ings of the 15th International Symposium on Transportation and Traffic Theory, Adelaide, Australia.
Pergamon/Elsevier, pp. 697–718.
Leutzbach, W. (1988). An Introduction to the Theory of Traffic Flow. Springer-Verlag, Berlin, Germany.
Leutzbach, W. (1991). Measurements of mean speed time series autobahn A5 near Karlsruhe, Ger-
many. Technical report, Institute of Transport Studies, University of Karlsruhe, Germany.
Leutzbach, W., Wiedemann, R. (1986). Development and applications of traffic simulation models at
the Karlsruhe Institute fuer Verkehrswesen. Traffic Engineering and Control 27 (5), 270–278.
Lighthill, M.J., Whitham, G.B. (1955). On kinematic waves II: A traffic flow theory on long crowded
roads. Proceedings of the Royal Society of London Series A 229, 317–345.
Lindley, J. (1987). Urban freeway congestion: Quantification of the problem and effectiveness of po-
tential solutions. ITE Journal 57(1).
Logghe, S. (2003). Dynamic modelling of heterogeneous vehicular traffic. PhD thesis, Catholic Univer-
sity of Leuven, Belgium.
Lowrie, P.R. (1982). The Sydney coordinated adaptive traffic system; principles, methodology, algo-
rithms. In: Proceedings of Institute of Electrical Engineers Internacional Conference on Road Traffic
Signalling London, pp. 67–70.
Ludmann, J. (1998). Beeinflussung des Verkehrsablaufs auf Strassen – Analyse mit dem fahrzeugorien-
tierten Verkehrssimulationsprogramm PELOPS. Technical report, Schriftenreihe Automobiltech-
nik, Institut für Kraftfahrwesen, Aachen, Germany (in German).
Mahmassani, H., Jayakrishnan, R. (1991). System performance and user response under real-time in-
formation in a congested traffic corridor. Transportation Research A 25 (5), 293–307.
Mahmassani, H., Liu, Y.H. (1999). Dynamics of commuting decision behavior under advanced traveler
information systems. Transportation Research C 7 (2–3), 91–107.
Mannering, F.L., Kim, S.-G., Barfield, W., Ng, L. (1994). Statistical analysis of commuters’ route, mode
and departure time flexibility. Transportation Research C 2 (1), 35–47.
Masher, D.P., Ross, D.W., Wong, P.J., Tuan, P.L., Zeidler, A., Peracek, S. (1975). Guidelines for design
and operating of ramp control systems. Technical Report NCHRP 3-22, SRI Project 3340, Standford
Research Institute, SRI, Menid Park, California, USA.
May, A.D. (1990). Traffic Flow Fundamentals. Prentice Hall, Englewood Cliffs, NJ.
McDonald, M., Hounsell, N.B. (1991). Road traffic control: TRANSYT and SCOOT. In: Papageorgiou,
M. (Ed.), Concise Encyclopedia of Traffic and Transportation Systems. Pergamon, pp. 400–408.
Messmer, A., Papageorgiou, M. (1990). METANET: A macroscopic simulation program for motorway
networks. Traffic Engineering and Control 31 (8), 466–470; erratum 31 (9), 549.
Michaels, R.M. (1963). Perceptual factors in car following. In: Proceedings of the 2nd Symposium on
Theory of Road Traffic Flow. Paris, France, pp. 44–59.
Minderhoud, M.M. (1999). Supported Driving: impacts on motorway traffic Flow. PhD thesis, Delft
University of Technology, Delft, The Netherlands.
Mirchandani, P., Head, L., Knyazyan, A., Wu, W. (2001). An approach towards the integration of bus
priority and traffic adaptive signal control. In: Proceedings of the 80th TRB Annual Meeting. Wash-
ington, DC.
Moreno-Banos, J.C., Papageorgiou, M., Schaffner, C. (1993). Integrated optimal flow control in traffic
networks. European Journal of Operations Research 71, 317–323.
 Ch. 11. ITS and Traffic Management 773

Munjal, P., Pahl, J. (1969). An analysis of the Boltzmann-type statistical models for multi-lane traffic
flow. Transportation Research 3, 151–163.
Nagel, K. (1996). Particle hopping models and traffic flow theory. Physical Review E 53 (5), 4655–4672.
Nagel, K. (1998). From particle hopping models to traffic flow theory. Transportation Research
Record 1644, 1–9.
Nelson, P., Sopasakis, A. (1998). The Prigogine–Herman kinetic model predicts widely scattered traffic
flow data at high concentrations. Transportation Research B 32 (8), 589–604.
Nelson, P., Bui, D.D., Sopasakis, A. (1997). A novel traffic stream model deriving from a bimodal kinetic
equilibrium. In: Papageorgiou, M., Pouliezos, A. (Eds.), Transportation Systems 1997. Proceedings
of the 8th IFAC Symposium on Transportation Systems, Chania, Crete, Greece. Pergamon, pp. 799–
804.
Newell, G.F. (1961). A theory of traffic flow in tunnels. In: Herman, R. (Ed.), Theory of Traffic Flow, pp.
193–206.
Ozbay, K., Datta, A., Kachroo, P. (2001). Modeling route choice behavior using stochastic learning
automata. In: Proceedings of the 80th Transportation Research Board Annual Meeting. Washington,
DC, USA.
Papageorgiou, M. (1980). A new approach to time-of-day control based on a dynamic freeway traffic
model. Transportation Research B 14, 349–360.
Papageorgiou, M. (1990). Dynamic modeling, assignment, and route guidance in traffic networks. Trans-
portation Research B 24, 471–495.
Papageorgiou, M., Marinaki, M. (1995). A feasible direction algorithm for the numerical solution of
optimal control problems. Internal report 1995-4, Dynamic Systems and Simulation Laboratory,
Technical University of Crete, Chania, Greece.
Papageorgiou, M., Blosseville, J.M., Hadj-Salem, H. (1990). Modelling and real-time control of traffic
flow on the southern part of Boulevard Périphérique in Paris Part II: Coordinated on-ramp meter-
ing. Transportation Research A 24, 361–370.
Papageorgiou, M., Haj-Salem, H., Blossville, J.M. (1991). ALINEA: A local feedback control law for
on-ramp metering. Transportation Research Record 1320, 58–64.
Papageorgiou, M., Haj-Salem, H., Middelham, F. (1998). ALINEA: A local ramp metering: Summary
of field results. Transportation Research Record 1603, 90–98.
Paveri-Fontana, S.L. (1975). On Boltzmann-like treatments for traffic flow: A critical review of the basic
model and an alternative proposal for dilute traffic analysis. Transportation Research B 9, 225–235.
Pavlis, Y., Papageorgiou, M. (1999). Simple decentralized feedback strategies for route guidance in
traffic networks. Transportation Science 33 (3), 264–278.
Payne, H.J. (1971). Models for freeway traffic and control. In: G.A. Bekey, (Ed.), Mathematical Models
of Public Systems 1, pp. 51–61.
Peirce, S., Lappin, J. (2002). Evolving awareness, use, and opinions of Seattle region commuters con-
cerning traveler information: Findings from the puget sound transportation panel survey, 1997 and
2000. In: Proceedings of the 82nd Transportation Research Board Annual Meeting.
Pignataro, L.J. (1973). Traffic Engineering – Theory and Practice. Prentice-Hall, Englewood Cliffs, NJ.
Pipes, L.A. (1953). An operational analysis of traffic dynamics. Journal of Applied Physics 24 (1), 274–
287.
Polydoropoulou, A., Ben-Akiva, M. (1999). The effect of Advanced Traveler Information Systems
(ATIS) on travelers behavior. In: Emmerink, R., Nijkamp, P. (Eds.), Behavioral and Network Im-
pacts of Driver Information Systems. Wiley.
Polydoropoulou, A., Ben-Akiva, M., Khattak, A., Lauprete, G. (1996). Modeling revealed and stated
en-route travel response to ATIS. Transportation Research Record 1537, 38–45.
Pozybill, M. (1998). Ist Verkehr chaotisch? Strassenverkehrstechnik 10, 538–545 (in German).
Prigogine, I., Herman, R. (1971). Kinetic Theory of Vehicular Traffic. Elsevier, New York.
Rekersbrink, A. (1995). Mikroskopische Verkehrssimulation mit Hilfe der Fuzzy Logic. Strassen-
verkehrstechnik 2, 68–74 (in German).
Rilett, L.R., van Aerde, M., MacKinnon, G., Krage, M. (1991) Simulating the TravTek route guidance
logic using the INTEGRATION traffic model. In: Proceedings of the Conference on Vehicle Naviga-
tion and Information Systems. SAE Conference Proceedings No. 253, Part 2. SAE, Warrendale, PA,
USA, pp. 775–787.
774 M. Papageorgiou et al.

Robertson, D.I. (1997). The TRANSYT method of coordinating traffic signals. Traffic Engineering and
Control 38 (2), 76–77.
Schwartz, S.C., Tan, H.H. (1977). Integrated control of freeway entrance ramps by threshold regulation.
In: Proceedings IEEE Conference of Decision and Control, pp. 984–986.
Smoulders, S. (1990). Control of freeway traffic flow by variable message signs. Transportation Research
B 24, 111–132.
Tampére, C., van Arem, B., Hoogendoorn, S.P. (2002). Gas kinetic traffic flow modelling including
continuous driver behaviour models. In: Bovy, P.H.L. (Ed.), Proceedings of the 7th Annual TRAIL
Conference. Delft Univ. Press, Delft, The Netherlands.
Treiterer, J., Myers, J.A. (1974). The hysteresis phenomena in traffic flow. In: Buckley, D.J. (Ed.),
Proceedings of the 6th International Symposium on Transportation and Traffic Flow Theory. A.M. &
A.W. Reed Private Limited, Artarmon, NSW, pp. 13–38.
Valdes, D., Paz, A. (2004). An integration of adaptive traffic control and travel information. In: Pro-
ceedings of the 83rd TRB Annual Meeting. Washington, DC.
van Aerde, M. (1994). INTEGRATION: A model for simulating integrated traffic networks. Technical
report, Transportation Systems Research Group, Queens University, Canada.
van Arem, B., Hogema, J.H. (1995). The microscopic simulation model MIXIC 1.2. Technical Report
1995-17b, TNO-INRO, Delft, The Netherlands.
Verweij, H.D. (1985). Congestion warning and traffic flow operations. Technical report, Dutch Ministry
of Traffic and Transportation, Delft, The Netherlands (in Dutch).
Watling, D., van Vuren, T. (1993). The modelling of dynamic route guidance systems. Transportation
Research C 1 (2), 159–182.
Wattleworth, J.A. (1965). Peak-period analysis and control of a freeway system. Highway Research
Record 157, 1–21.
Wu, N., Brilon, W. (1999). Cellular automata for highway traffic flow simulation. In: Ceder, A. (Ed.),
Transportation and Traffic Theory. Proceedings of the 14th International Symposium of Transporta-
tion and Traffic Theory, Jerusalem, Israel. Pergamon/Elsevier, pp. 1–18 (abbreviated presentations).
Wunderlich, K.E., Hardy, M.H., Larkin, J.J., Shah, V.P. (2001). On-time reliability impacts of Advanced
Traveler Information Services (ATIS): Washington, DC case study. Technical report, Mitretek Sys-
tems.
Yang, H. (1998). Multiple equilibrium behavior and advanced traveler information systems with en-
dogenous market penetration. Transportation Research B 32 (3).
Yim, Y., Miller, M.A. (2000). Evaluation of the TravInfo field operational test. Technical report, Insti-
tute of Transportation Studies, University of California, Berkeley. California PATH Program.
Young, P.C., Taylor, J., Chotai, A., Whittaker, J. (1997). A non-minimal state variable feedback ap-
proach to co-ordinated ramp metering. In: Kotsialos, A. (Ed.), DACCORD Deliverable D 6.1 -
Cordinated Control, vol. 6.1. European Commission, Brussels, Belgium.
Yuan, L.S., Kreer, J.B. (1968). An optimal control algorithm for ramp metering of urban freeways.
In: Proceedings of the 6th IEEE Annual Allerton Conference on Circuit and System Theory. Allerton,
Illinois, USA.
Zackor, H. (1972). Beurteilung verkehrsabhangiger Geschwindigkeitsbeschrankungen auf Autobah-
nen. Forschung Strassenbau und Strassenverkehrstechnik 128, 1–61.
Zackor, H., Balz, W. (1984). Verkehrstechnische Funktionen des Leitsystems. Strassenverkehrstech-
nik 28, 5–9.
Zhang, H.M. (2003). Driver memory, traffic viscosity and a viscous vehicular traffic flow model. Trans-
portation Research B 37 (1), 27–41.
 Subject Index

A ant colony algorithm 380, 384, 397

approximate dynamic programming 300,
a priori optimization 412, 587, 610 318, 319, 329, 330, 332, 348
actionable attribute
– resource 310 – monotonicity axiom 578
– time 301 – transition function 309
active automatic incident detection (AID) 745
– guided evolution strategy 384 AVL 590
– resource 290
adaptive
B
– memory 383, 394
– method 448 backward reachable set 340
– route selection 587, 588, 590, 593 barge 198, 200, 209, 220, 235, 261, 262, 272,
agent 338, 340 273
air – scheduling 222
– cargo 289 basis function 333, 334
– mobility command 293, 327 batch process 341, 356
– taxi 449, 450 berth scheduling 476, 501
– traffic control (ATC) 4, 8, 9, 14, 16, 17, 20, bilevel program 691, 701
23, 24, 29, 38–40 booking 195, 264, 270
– – separation requirement 7 bounded penalty model 414
– traffic control system command center Braess paradox 644, 684, 694
(ATCSCC) 24, 32, 33 branch-and-bound 99, 100, 102, 108, 112,
– traffic flow management (ATFM) 4, 7, 12, 135, 142, 149, 164, 371
19, 21–25, 28, 29, 33, 37, 38, 60, 61 branch-and-cut 135, 152, 374, 391, 412
– traffic management (ATM) 5, 6, 8, 11, 14, branch-and-cut-and-price 375
18, 19, 20, 23, 39, 61 branch-and-price 48, 55, 100, 102, 103, 106,
– traffic service provider 24–26 160, 391
airborne holding 25, 26 bulk carrier 190, 197, 198, 204
aircraft busy fraction 458, 460
– recovery 41–45, 48
– sharing 429 C
airport capacity 2, 7, 15
airspace capacitated location problem 378
– capacity 2 capacity 15, 21
– sector 20, 21, 38 – constraint 368, 392
allocation and dispatching of yard cranes and – coverage chart (CCC) 16, 17
transporters 517 – envelope 10–12, 15
ALOHA 547, 560 – of airspace sectors 17
ambulance – of runway 7, 8, 10
– fleet management 454 – of runway systems 7, 12
– location 454 – uncertainty 29, 30, 32
AMOC 761, 764 CDM 25, 30, 32, 33, 35–39

775
776 Subject Index

chance constrained programming 411 deadheading 89, 90, 113, 118, 119
charter 198, 199, 220, 223, 237, 238, 240, 242, – capacity 20
252, 265, 267 – class 287–289, 306, 345
COA 236, 267 – function 305, 307, 323
coast guard 262, 263 – set 306, 321
collaborative decision making (CDM) 4, 24, deep-sea 200, 263, 264, 279
30, 32, 60 demand
column generation 49, 55, 76, 85, 98, 99, – management 19, 716
102–104, 106, 107, 159, 160, 162, 164, – satisfaction model 577
168, 176, 178, 238, 240, 242, 257, 389, – uncertainty 29, 32
674, 676–678, 683 dense
conditional risk 575, 579 – network 156, 157, 161
congested assignment 83, 84, 86, 89 – railway 165
consolidation density–speed relation 716, 717, 725, 730,
– operations 475 731
– terminal 481 deployment 196, 199, 200, 210, 211, 222,
– transportation 472 257–261, 267
constraint programming 102, 103 deterministic
container 190, 194, 196, 198–200, 205, 209, – annealing 380, 383
211–213, 215–219, 221, 257, 268–270, – dynamic model 298, 328
279 dial-a-flight 430, 448
– intermodal transportation 470 dial-a-ride 429, 430, 439
– port terminal 475 direct transfer 477
containership loading 269 dispatching 70, 95, 109, 110, 120
continuous value function approximations distributional forecast 308, 313
333 disutility model 576, 577
contract double deck 134, 135, 139
– evaluation 195, 206, 211, 212, 221, 279 double-horizon 447
– of affreightment 221, 267 driver
contracted cargo 237, 244 – scheduling 70, 95, 100
coverage 455, 458 – – duty 94
CPP 173, 174, 179 – support system 724
crew duty 173, 175, 176, 178–180
– duty 177 – generation 177
– planning 172 – scheduling 100, 104, 106
– recovery 41, 45–48 dynamic
– rostering 108, 109, 173–175, 178–180 – assignment problem 350
– scheduling 91, 102, 103, 120, 173–176, 178– – information process 302, 341, 350
180, 182 – model 285, 294, 302, 308, 341, 355, 362
critical – programming 90, 102, 168, 169, 294, 295,
– density 718, 762 308, 316, 324, 330, 331, 752
– occupancy 757 – relocation 461
curse of dimensionality 316, 330, 332, 333 – resource 291
customized transportation 471 – – transformation problem 296
cyclic timetable 135, 141, 149, 150, 182 – route information panel (DRIP) 739
– TTP 149 DynaMIT 739
DYNASMART-X 739
D
E
d-day policy 402
danger circle 558, 563, 566, 569, 570, 604 edge risk 544, 569–572
dangerous goods 539 efficient set 585, 586, 589
Dantzig–Wolfe 239, 241, 249 elementary shortest path 388, 391
 Subject Index 777

emergency FTA 556

– evacuation 734 full shipload 201, 206, 223, 224, 226, 227,
– vehicle 429, 431 228, 236, 270
empty balancing 495 fundamental diagram 717, 718, 726, 762
endogenous information process 306
environmental 196, 200, 257, 264–267 G
– routing 195
gap 672, 676, 677, 679, 680, 704
equilibrated 640, 641, 643, 646
– function 667
– flow 641
gas-kinetic model 728, 730
equilibrium 623, 624, 626–630, 632, 633,
Gaussian plume model 558, 559, 563
635–639, 641–650, 652–654, 656–658,
GDP 28–30, 32–34, 37, 61
660, 662, 663, 665–668, 674, 676, 678,
generalized
680–684, 687–691, 694, 696, 697, 700–
– assignment problem 378
702
– order constraints 443
– condition 625, 628, 639, 640, 648, 649, 660
genetic algorithm 77, 78, 104, 395
– speed 726, 727
geographic substitution 345
equipment
Gibbs sampling 742
– assignment 477
Gini coefficient 594
– cycle 162, 165
GIS 549, 550, 559, 568, 594, 599, 605–608
equity 543, 580, 584, 594–597, 602–604, 607,
global route planning 580, 581, 593–596, 608
608
GPS 591, 610, 732
ETA 556
gradient 704
event tree analysis 555, 556
– mapping 666, 670, 678, 702
evolution strategy 395, 396
– property 635, 651
exogenous information 298
granular tabu search 382, 384
– process 296, 305, 306, 308, 326, 346
GRASP 396
expected crew cost 57, 58
ground delay program (GDP) 17, 24, 28, 33,
40, 43
F
ground holding 24–26, 28–31
facility location and transportation 599 guided local search 396, 397
fault tree analysis 556
feeder 216 H
FETA 556
hazardous materials 539–542, 544–547, 564,
FIFO 589, 592
565, 573, 603, 611
fleet
heterogeneous
– management 495
– line system 134
– mix 209, 263
– resource 294, 295
– size 196, 199, 201, 205–207, 209–212, 221,
hub and spoke 212, 216, 217, 472, 483
223, 257, 261, 278, 279
flexible cargo size 223, 232, 234, 240, 242,
I
244, 248
flight plan 1, 4, 6, 25, 40 incompatibility graph 154
flow decomposition 633, 634, 645 incompatible
FN-curve 553, 554 – pair 151, 392
forecast 302, 308 – path inequalities 392
Frank–Wolfe 666, 669, 670, 672, 674–677, – routes 151, 153
679, 680, 683, 699, 701 indirect transfer 477
free speed 720, 722, 730, 762 individual risk 553, 555, 560, 594, 600–602,
freight transportation 285, 286, 290, 301 604
frequency 133–135, 138, 139, 155, 486, 487 information 285, 293, 297–299
– analysis 552 – class 287–290
– of service 473 – process 286, 296, 298, 301, 306, 342, 344
– setting 86, 87, 94 – state 295
778 Subject Index

infrastructure LP relaxation 145, 149, 150, 152, 157, 159,

– manager 130, 142, 144 160, 162, 164, 176, 179
– planning 130 LPP 135, 136
inland waterways 191, 193, 200, 222, 261, 279 LRP 600
integer LTL trucking 489
– L-shaped method 412
– multicommodity flow model 156, 158, 162, M
166 macroscopic model 731
integrated control 761 – kinematic wave model 725–728
INTEGRATION 725 – viscous model 728
intermodal 288 macroscopic variable
– terminal 474 – density 717, 725–727, 729, 761
– transportation 467 – flow 717, 725, 761
inventory 196, 204, 205, 243–256, 265, 274, – mean speed 717, 725, 727, 729, 761
278, 294 – speed variance 725
– routing problem 398 maintenance
inverse optimization 692 – requirement 161, 162
irregular operations 38 – routing 165, 166
– scheduling 70, 95, 111
K Markov
– chain 742
kinematic wave 725–728 – decision
knowable time 301 – – model 413
– – process 407, 408
L – property 337
mathematical program with equilibrium con-
label-setting algorithm 581
straints (MPEC) 691, 694, 695, 701
lagged information 296
mean–variance model 577, 587
– process 300, 301
mean-risk 584
Lagrangian relaxation 74, 98, 103, 106–108,
memetic algorithm 384
136, 143, 146, 174, 177, 387
metaheuristic 77, 78, 104, 379, 392, 394
lane
microscopic model 730, 731
– changing 719, 720 – action point 723
– control 760 – car-following 720, 722, 723, 727, 728, 731
layered resource 291, 293, 306 – – asymptotic stability 721
learning mechanism 384 – – local stability 721
less-than-truckload 288, 290, 341, 356, 357, – – optimal speed model 722
472 – – psycho-spacing model 722
letter service 489 – – safe-distance models 720
line – – stimulus-response models 720, 721, 725
– capacity constraint 143, 145, 146 – cellular automata 724
– planning 133, 134 – – rule 724
linear value function approximation 338 – particle hopping 724
liner 190, 195, 198–201, 204, 205, 209, 211, Minty 678
212, 213, 216–218, 221, 222, 257, 260, – variational inequality 641
267, 270, 271, 279 MIXIC 724
link control strategy 760 Monte Carlo techniques 742
LISB 739 MOTION 751
local motor carrier 472
– route planning 580, 581, 583 move-up crew 58–60
– search 142, 174, 379, 381 multi-attribute 646–648, 678, 680, 693, 694,
location–routing 600 701
locomotive 286, 288, 289, 291, 321, 336, 338, multi-mode 646, 647, 678
339, 351 multiagent 86, 116, 340
 Subject Index 779

multicommodity 97–99, 145, 294, 295, 313, perceived risk 568, 575, 582, 604
317, 336–338, 341, 481, 595 periodic event scheduling problem (PESP)
multilayered resource 294 136, 149, 150
multiobjective 543, 550, 551, 577, 583, 585– persistence 271, 275, 276
587, 590, 594, 604 Poisson process 574, 576
multiple population
– cargoes 223, 228, 232, 235 – exposure 563, 567, 574, 575, 576, 580, 582,
– products 192, 234, 235, 243, 251 586–589, 594, 603, 604
– stakeholders 543, 583 – search 380, 383
myopic port dimensioning 493
– model 308, 313, 314, 324 post-decision state variable 300, 317, 329
– policy 313 postal service 472, 489
pre- and post-decision state variable 300,
N 317, 329
precedence constraint 433, 434
Nash equilibrium 624, 642
predecessor inequality 443
naval 195, 198, 199, 202, 208, 222, 262, 263,
price of anarchy 643, 644
265
primitive resource 293
neuro-dynamic programming 415
PRODYN 752
noncyclic 182
– timetabling 141
– TTP 141, 143 Q
normal equilibrium 640, 641, 643 qualitative risk assessment 551
noxious facility 599 quantitative risk assessment 546, 552, 608
quay crane 475
O – allocation 476, 501
ocean shipping line 472 – scheduling 507
OD (origin–destination) 737, 738, 751, 755 queuing model 716
off-policy 333
OPAC 753 R
operational 192, 195, 196, 201, 206, 207, 209, rail 286, 287, 314, 341, 489
221, 263–265, 267, 270, 271, 275, 279 – transport 582
– planning 70, 91, 94, 95, 104, 109
railway 472
operations recovery 40
– CPP 175
optimality principle 578, 581
– crew
optional cargo 222, 223, 236, 237, 270
– – rostering 176
order matching constraint 434
– – scheduling 174, 176
– passenger transportation 132
P
ramp metering 755, 756
Pareto optimal 577, 583–587, 590, 596, 603 – fixed-time 756
parking 70, 94, 95, 109–111, 120 – – aging 756
particle model 725 – local 757, 759
passenger – – ALINEA strategy 757–759
– assignment 70, 71, 73–75, 77–80, 82, 84, 85, – – demand-capacity strategy 757, 758
87, 88 – – occupancy strategy 757, 758
– path assignment 83 – multivariable regulator strategies 758
– recovery 41, 48, 49 – – METALINE 759
– transportation 129, 132 – nonlinear optimal strategy 759, 761
path risk 572, 574, 575, 578, 579 – optimal control 763
PC*HazRoute 548 – reactive 757
PC*Miler 547 ration-by-schedule (RBS) 33–36
PELOPS 724 reaction time 720–722, 724, 726, 727
pendeling 723 reactive tabu search 395
780 Subject Index

real-time 362 – emergency evacuation 734

– control 69, 70, 112, 118, 120, 131, 182 – – consistency 737, 739
– update 593 – – descriptive 736
recourse 411 – – non-predictive 733, 735–737, 742
redeployment 461, 462 – – predictive 733, 735–737, 739, 740, 742
resource 286, 289–291, 294, 306, 309, 310, – – prescriptive 736
330 – market penetration 732
– class 287–289, 291, 292 – predictive guidance mapping 738
– dynamics 310, 352 – static system 733
– layer 289, 291, 292 – transmission range 735
– management 471 – travel decision 740, 741
– state 294, 340, 351, 353 – traveler response 733
– – variable 322 routing 191, 194, 196, 200, 201, 206, 209, 215,
– vector 320 216, 222, 223, 228, 238–243, 245, 247,
restricting schedule 4, 19, 20 249–253, 255, 257, 263–267, 274, 276–
RHODES 751 278
routing and scheduling 544, 550, 580, 581,
ride time 430, 439–443
583, 587
risk
runway configuration 12, 14, 15
– assessment 545–549, 551–553, 559, 567, 583,
605, 608, 609 S
– aversion 566–568, 576
Ro–Ro 198 saturation flow 745
road pricing 716 SCATS 750
robust schedule 473, 486
– aircraft routing 53, 54 – coordination 21
– airline scheduling 50 – planning 4, 5, 50, 53
– crew – recovery 5, 40
– – pairing 57 schedule-based transit assignment 82
– – scheduling 56 scheduling 191, 194, 196, 200, 206, 222, 223,
226–228, 230–232, 234–238, 240–244,
– fleet assignment 51–53
248, 249, 251, 257, 261–265, 267, 268,
– schedule design 51
270, 272, 274, 276–278, 439
robustness 195, 271–275
SCOOT 746, 748–750, 752
rolling 314
SDOT algorithm 593
– horizon 308, 326
security 549, 550, 609, 610
– – procedure 313
sensitivity 684, 686–691, 699–701
– stock circulation 133, 154, 156, 157, 165,
– analysis 700
167, 169, 182 – variational inequality 686
rolling stock circulation problem (RSCP) separable value function approximations
154, 155, 157, 160 337
rolling stock management 132 separation requirement 8, 9, 11–14
roster 172, 173, 175, 178, 179 sequencing 8, 10–12, 25, 39
route guidance and information system – aircraft 11
(RGIS) 732, 755, 760 service 486
– area focus 735 – network 487
– communications system 735 – – design 478, 485
– dissemination 735 shift 144, 146, 151, 153
– – cellular radio 735 ship
– – FM sub-carrier 735 – design 195, 196, 201–203
– – highway advisory broadcast 735 – loading 195, 196, 254, 264, 268, 269
– – infrared 735 shipment 198, 200, 201, 204, 205, 237, 243,
– – microwave 735 252, 257, 265
– – variable message sign 735 shipper 200, 204, 205, 211, 221–223, 243,
– driver response 739, 740 251, 252, 265, 270, 276, 278, 280
 Subject Index 781

shockwave theory 716 stowage 196, 269, 270

short-sea 200, 204, 209, 216, 264, 273, 279 – planning 509
short-turning 89, 113, 119 – sequencing 476, 509
shunting 151, 155, 166–172 strategic 195, 196, 201, 205, 206, 209–211,
shuttle service 212, 220 220, 221, 257, 260, 261, 263, 269, 271,
SIMONE 724 272, 278, 279, 657, 659–661, 665, 706
simplicial decomposition 674, 677, 683 – equilibrium 659, 660
simulated annealing 77, 381 – model 657, 660, 663, 682, 684
simulation 220, 221, 257, 261, 272 – planning 69, 70, 78, 131, 136
– – of multimodal systems 520
– model 4, 15, 38, 39
strategy 657, 659–661, 666, 668, 677, 678,
single
682, 706
– commodity 294, 337
stretch 144
– deck 134, 135, 138, 139
strictly monotone 638, 639, 645, 649, 651,
– link 357–359 668, 673, 680, 703, 704
social amplification of risk 543 strongly
societal risk 553, 555, 567 – acyclic 633, 634, 636, 637, 645
space–time network 492 – monotone 673, 677, 678, 689, 703, 704
space-allocation 477 – regular 685
– problem 514 STV 587–593
sparse successor inequality 443
– network 155, 156, 161 supply chain 195, 201, 220–222, 242, 243,
– railway 165 256, 257, 278–280
spatial distribution of risk 580, 593, 594, 602, swap of duties 166
607 switching times 162
speed system
– limit 720 – design 478
– selection 195, 196, 264, 267, 275 – optimal 643, 646, 658, 665, 691–693, 697,
spot 199, 206, 221, 223, 236–238, 240, 242, 698, 701, 733, 734
265, 272, 278, 279 – toll 701
– charter 252
T
stability 684
state tabu search 77, 104, 380, 381, 394, 446
– space 294, 316, 371 tactical 195, 196, 201, 206, 210, 220, 222, 243,
– variable 294, 295, 298, 300, 316, 317, 339 256, 257, 260, 261, 264, 269, 271, 272, 279
static – planning 86, 91
– resource 290, 291 tanker 189, 190, 197, 198, 200, 204, 208, 209,
– service network design 487 234–236, 257, 268
stochastic temporal
– approximation 332, 338 – substitution 345
– difference 338
– customer 410, 413, 415
terrorist attack 543, 609, 610
– demand 410, 414
threshold-accepting algorithm 380
– dominance 584, 585
tidal flow 760
– gradient 332, 334, 335
time window 192, 200, 223–225, 227–229,
– model 302 231, 232, 236, 238, 239, 241–246, 248,
– optimization 132, 150 249, 251, 263, 268, 273, 274, 430, 431,
– programming 308, 411 439–441, 445
– – with recourse 411 time–space
– time-varying network 581 – graph 158
– travel time 410, 416 – network 163
– vehicle routing problem 410 time-dependent service network design 492
stop-skipping 117, 118 timetable 133, 141–144, 146, 149, 150, 153,
storage activities in the yard 514 156, 161, 167, 172, 182
782 Subject Index

timetabling 91, 92, 94, 120 travel time

toll 654, 655, 665, 684, 691–696, 699, 701 – experienced 734
TPP 150–152 – instantaneous 734
traditional risk model 573, 574, 579, 580 TravInfo 740
traffic trip 144, 149, 150, 155–160, 162–166, 172–
– assignment 623, 624, 657, 678, 738, 739 177
– equilibrium 623–625, 627, 635–637, 645, 650, truckload 290, 313, 341, 350
652, 653, 657, 666, 667, 676, 678, 681, – trucking 287, 471
683, 684, 686, 687, 690, 697, 699–701, TTP 142, 143, 145, 152
705, 737 TUC 751
traffic flow TUSP 168
– hysteresis 718, 727
– metastable condition 718 U
– model
uncertainty 192–194, 200, 252, 261, 262, 264,
– – deterministic 719
265, 271, 272, 274, 279, 280
– – gas-kinetic 719
uncongested assignment 80, 85
– – macroscopic 717, 719, 725
(un)coupling constraints 156
– – mesoscopic 717, 719, 725
United States Emergency Medical Services
– – microscopic 717, 719
Act 457
– – particle 719
urban courier service 429, 430, 445
– – stochastic 719
US DOT 539–541, 543, 547, 555, 564, 573,
– – submicroscopic 723
575
– stationary conditions 717, 722
user
– synchronized 718
– equilibrated 641
– transient state 718
– equilibrium 625, 626, 629, 640–644, 666,
– unstable condition 718
684
– wave dynamics 716
– optimal 733
traffic signal
– optimized 640–643
– aging 746
UTOPIA 746, 752
– cycle time 745, 746
– cyclic flow profile 747, 750
V
– effective green time 745
– fixed-time 743, 745–748 value function
– green split 745 – approximation 318, 336, 353
– green wave 746 – direction sign 755
– offset 745, 746 – message sign 733, 735, 736, 755
– traffic-responsive 745, 746, 748 – neighborhood search 397
train – speed limit 755
– composition 155–157, 159 variational 645, 703
– operator 130, 136, 140, 142, 154, 156, 167, – inequality 624, 627, 630, 631, 636–638, 642,
172, 173 645, 647, 649–651, 660, 666, 667, 672,
– platforming 182, 150 673, 677, 678, 684, 685, 687, 688, 692,
– – problem 150 695, 697, 702, 703, 705
– shift 144 vehicle 91, 104
– timetable 145 – and crew scheduling 120
– timetabling 141 – and duty scheduling 95, 104
– unit 140, 155–160, 162, 166–172 – holding 112–116
– – shunting 167 – routing 430
– – – Problem (TUSP) 167 – – problem 367
tramp 195, 198–200, 206, 211, 221–223, 236– – – – with pickup and delivery 430, 431
238, 240–243, 251, 257, 264, 268, 270, – – – with time windows 385
271, 277–279 – scheduling 70, 94, 95, 104–106, 119, 120,
transportation on demand 429 430
TRANSYT 747, 749 – – and duty scheduling 104
 Subject Index 783

vendor managed inventory 398 waypoint 6, 18, 26, 27, 29

voyage 192, 199–201, 203, 209–213, 215–219, weather 193, 194, 200, 213, 265, 273, 274
235, 237, 252, 254, 255, 257, 258, 260,
264–266, 274 Y

W yard crane 475, 517

Wardrop equilibrium 624–627, 629, 632, Z

639–646, 652, 653, 655, 656, 658, 664,
666, 683, 694, 697 zone scheduling 89, 90