Transpn. Res.-B. Vol. 288, No. 3, pp.

197-212, 1994
Copyright 0 1994 Elsevier Science Ltd
Pergamon Printed in Great Britain. All rights reserved
0191-2615&l $6.00 + .OO

PRICING OF TRACK TIME IN RAILROAD OPERATIONS:

AN INTERNAL MARKET APPROACH

PATRICK T. HARKER and SUNGWOOK HONG

Fishman-Davidson Center for the Study of the Service Sector, The Wharton School,
University of Pennsylvania, Philadelphia, PA 19104-6366, U.S.A.

(Received 5 October 1992; in revisedform 15 July 1993)

Abstract -This paper presents a computable equilibrium model of an internal market for track
resources in a railroad. The problem of estimating the value to each train of track capacity, which
in turn is used to create the actual train schedules, is formulated as an N-player, noncooperative
game with nondisjoint strategy sets. In this model, the effects of other traffic on a given train
schedule (the mean and variance of total travel time) are represented by a line delay model for a
scheduled railroad on a partially double track rail line. The generalized Nash equilibrium for the
resulting game-theoretic model is found as a solution to a quasi-variational inequality problem.
The goal of this model is to ascertain how close the prices from the internal market system (the
game-theoretic model) comes to globally optimal prices. Data from a major Class I railroad are
used to explore this issue in detail.

1. INTRODUCTION

Traditionally, railroads have provided low-cost, low-quality transportation service for

bulk commodity movements. However, due to the recent movement toward the produc-
tion of high-valued goods and subsequent demand for highly reliable and efficient trans-
port service, railroads need to reinvest in modern technology and restructuring of their
management practices in order to respond to present market demand. The demand of
higher quality transport service is accompanied by the desire for better management of
rail track resources along with major organizational changes.
Under the United States rail industry’s present structure, the railroads operate as
privately owned carriers over privately owned track, with each network reserved for the
near exclusive use of its owner-carrier (in most cases, only limited trackage rights are
provided to other carriers). Each railroad company can serve only as far as its own tracks
extend. In contrast, the typical structure of other transportation modes (e.g., highway,
water, and air) is privately owned carriers operating over a publicly owned, jointly used
fixed way (e.g., highways, ports, and airports). Although a broad rang of service quali-
ties and prices are available in the other modes, the present railroad structure often
discourages diversity and innovation.
To overcome these limitations, this paper proposes the creation of an internal market
for track resources in a railroad. A decentralized managerial structure of rail operations,
in which a railroad company consists of one railroad authority and several divisions (e.g.,
one for each type of train: passenger, intermodal, etc.), is proposed. Each of these
divisions is responsible for a particular customer group. Because the responsibility for
operating and tactical decisions are assigned to each self-contained division, the rail
authority would be able to evaluate the overall performance of the organization rather
than become absorbed in the affairs of the functional parts. The rail authority is princi-
pally concerned with strategic decisions, involving long-term planning, appraisal, and
control, including the allocation of resources among the competing divisions. The rail
authority manages the rail track spaces and charges a track user fee to each division,
because the rail track resource involves one of the most expensive components in a
railroad. The main issue in this paper is how to estimate the track user fee which the
authority should charge for use of the track.
An additional train on a track may slow down the other trains which are running on
197
198 P. T. HARKERand S. HONG

the same track at that time. The delay of trains imposes a time cost on the movement of
cars and constitutes a large portion of the total cost for the system, when one includes
customer revenue impacts. Each unit of train service is delayed perhaps by a very small
amount, but summed over all traffic in the rail network, the impedance imposed is much
larger and will be proportionally greater with a higher density of traffic. The delay of a
high-valued product train with a tight schedule would be more costly than that of a
low-valued train: each type of train has a different value for transit time and schedule
adherence. Thus, the problem of how to estimate the adequate and competitive user fee
for limited track capacity is formulated in a game-theoretic model because each division’s
independent decisions affect the interest of all. In an N-player, noncooperative game
with nondisjoint strategy sets, each division tries to maximize its utility (i.e., minimizing
deviations from their ideal train schedules). Actual train schedules are found as a solution
to a quasivariational inequality problem, which represents how the actual operating group
would implement these train schedules. The rail track prices are determined based on the
dual price of train travel time constraints which represent the value to each train of using
existing track capacity.
Note that the proposed pricing scheme differs substantially from traditional conges-
tion (or peak-load) pricing schemes in that the spatio-temporal aspects must be factored
into an oligopolistic-like market structure. The original approach of estimating conges-
tion tolls by Pigou (1920) and the resulting literature in this area (Vickrey, 1969; Walters,
1968; Levin, 1981; Carlin & Park, 1970; De Vany & Saving, 1980; Small, Winston, &
Evans, 1989; Morrison, 1986) essentially ignores the imperfectly competitive nature which
is present in the railroad problem under investigation. While peak-load pricing (Crew &
Kleindorfer, 1987; Panzar, 1976; Morrison, 1983) does deal with this market phenome-
non, it typically does not handle the spatial and temporal aspects in a simultaneous
fashion. The one exception is Viton’s (1980) model of equilibrium short-run peak load
pricing using a binomial logit mode choice model. The model is applied to peak load
travel on the San Francisco-Oakland Bay bridge, and tolls for autos, buses, and trucks
are calculated.
Similar types of organizational structures as that proposed in this paper have been
discussed by Overbey (1982) for the railroad industry and Williamson (1975) for general
market hierarchies. Overbey (1982) proposed the Free Enterprise Alternative which is
explained as follows:

All rail roadways in a region would be owned by a regional roadway company, and
several roadway companies would form the national roadway network. Many carriers
of several different types would share use of the roadway network. . . . Joint use of
roadways offers carriers better routes, lower costs, and improved efficiency. (p. 147)

Faced with the types of internal operating problems that emerge as the unitary form
(or U-form) company increases in size and complexity, Williamson (1975) proposed the mul-
tidivisional (or M-form) structure. His concept involves substituting quasi-autonomous
divisions for the functional divisions of the U-form structure as the principal basis for
dividing up tasks and assigning responsibilities.
Here, we are not proposing the roadway/carrier separation as in Overbey (1982),
although this is precisely the approach taken in the deregulation of the Swedish rail
system (Hansson & Nilsson, 1991). We are considering the internal market of a railroad
with an M-form structure. The proposed internal rail track market structure has three
advantages. First, we can estimate the track user fees for each different user group based
on the marginal benefit of using a unit of track time. Second, when a carrier wants to
add a train on the rail track with a certain train schedule, the game-theoretic solution can
estimate the new user’s track fee and the effect (e.g., additional congestion or track user
fee) on the existing users. Third, each division can respond better for each customer
group. Because the organization has an internal (within each division) rather than external
(within total system) control mechanism to make detailed evaluations of the performance
 Pricing of track time 199

of each of its operating parts, each division can make ex post adjustments more easily in
response to varying market demands and evidence of performance failures.
The proposed model considers the case where a railroad authority is an owner of
track spaces, a scarce and expensive resource, and acts as the leader of the rail operation
game. That is, the railroad authority sets a vector of optimal track user fees for using a
unit of track time as its strategy and predicts the users’ reactions. These internal user fees
will be imposed on each train. A crucial point is that we do not assume any social welfare
function for the rail authority. Rather, the goal of the authority is to set track prices so
that capacity on each rail line is not exceeded. Another way to state this point is to make
the utility function for the authority equal to zero if the system is feasible, and minus
infinity if infeasible.
By solving a quasivariational inequality formation of the generalized Nash equilib-
rium game, the optimal user fee for a unit of track time which a railroad authority can
charge to each train is derived from duality theory. The dual variable (or price) of the
track time constraint will be the user fee for a unit of track time. For example, the higher
priority trains usually run with tighter schedules and require a high level of service; thus,
they must pay a higher track user fee. On the other hand, more delays will be imposed on
the trains carrying agricultural products or coal which, in turn, will pay lower track user
fees at the expense of service quality (less reliable schedules with more delays) because
they have a lower value of time. The dual variables will be the discriminatory track user
fees of each train based on its value of time.
The ultimate goal of this paper is to compare the results of two models: (a) the
market-drive (game-theoretic) train scheduling model and (b) the system optimal train
scheduling model. The results of the first model, where each division minimizes its sched-
ule deviations may not be globally optimal for the railroad; however, this global optimum
is unlikely to be achieved due to the necessity of a tightly controlled, U-form organization
structure in this case. The solution to the system optimal model will be used to benchmark
the performance of the proposed decentralized structure. Thus, we consider the system
optimal train scheduling model where one operations department in a railroad (a single
player in a system) controls every movement of trains and tries to minimize the total train
schedule deviations from the ideal train schedules. A nonlinear mathematical program-
ming problem is solved to obtain the optimal train schedule for every train within a
system. Using data from a major U.S. railroad, the question of the potential efficiency
of a decentralized management structure will be explored in detail.

2. LITERATURE REVIEW

This section provides a brief review of the analytical tools that are used in the
development of the internal market model in Section 3.

2.1 Line delay models

To capture the impact of various divisions on the performance of the railroad as a
whole, one must first provide a representation of the congestion which occurs in rail
operations. These delays that occur along a rail line have been studied by several research-
ers in analyzing train movements and dispatching activities. The early notable models of
delay over lines are those by Petersen (1974, 1975) in which the average line delay for a
single and a partially double-tracked railroad are computed. The models set priority rules
for meets and overtakes for different speeds of trains and calculate the value for the
expected delay, based on assumptions of equally spaced sidings and independently and
identically distributed trains over a given time period.
Chen and Harker (1990) relax the strong assumption in the Petersen line delay
models that the trains are uniformly distributed over time (i.e., Poisson arrivals) and
provide single track delay functions given train schedules and distributional information
concerning operational uncertainties. The mean and the variance of the total travel time
of trains over a single track are estimated, which enables one to measure the reliability of
a given set of schedules.
200 P. T. HARKER and S. HONG

Harker and Hong (1990) extended the results of Chen and Harker (1990) to a par-
tially double track situation. The expected running time of a train in Harker and Hong
(1990) is the sum of the following items:

1. Train’s free running time between two yards;

2. Probability of a meet x meet delays by the trains from the opposing direction;
3. Probability of an overtake x overtake delays by the faster trains from the same direc-
tion.

Note that the delay is encountered by each train as a result of random interference
with other trains and is a function of its own schedules as well as the other trains’
schedules. Therefore, the scheduling decisions of each train will directly affect each other
because of the delay (or congestion) generated by operating additional trains in the
system. Given the other trains’ schedules, one may reduce the expected meet and overtake
delay by changing one’s own train schedule. Furthermore, the priorities of the various
trains that are embodied in these models define the rail authority’s operating policy. That
is, these priorities are ultimately what dispatchers and schedulers use to operate the rail
lines and are clearly in the control of the authority. The Harker-Hong delay model will
be used to represent the impact of various divisions and the authority on overall schedule
performance in the sequel.

2.2 Variational inequality, game theory, and transportation networks

The other important analytical device that is used herein is the theory of variational
inequalities. A variational inequality (VI) problem is to find a vector x* in a closed and
convex set X c (R” such that for the function F: X + CR”, the following inequality
holds:
F(x*)~(~ - x*) 2 0 vy E X. (1)

The quasivariational inequality (QVI) problem, when X is replaced with a point-to-set

mapping X: Glm -+ &lm,is to find an x* E X(x*) such that

F(x*)~(Y - x*) 2 0 vy E X(x*). (2)

Harker and Pang (1990) presented a comprehensive discussion of the theoretical

and computational issues surrounding finite-dimensional variational and quasivariational
inequality problems, as well as listing numerous applications of the methods.
The notion of a generalized Nash equilibrium arises whenever a player in a game can
affect both the utilities and feasible sets of other players. As shown in the next section,
this interaction of the feasible sets will arise due to congestion externalities. Harker (1991)
studied the properties of a generalized Nash game, which is an N-person noncooperative
game with nondisjoint strategy sets. The relationships between variational and quasivaria-
tional inequality theory were presented and the practical implementations of a generalized
Nash game were given. In particular, the important results from Harker (1991) are:

1. A Stackelberg or principal agent-like problem can be solved using the generalized Nash
equilibrium concept and the quasivariational inequality (QVI) formulation.
2. The QVI solution set contains the solution of a related VI problem.
3. In a special case (it is also the case in this paper), any VI algorithm can be employed in
order to calculate a solution to the model; thus, the existence of a QVI solution is
established.
4. By solving the QVI formulation, the optimal strategy of a leader in a Stackelberg-like
game is provided based on the dual variables of the model.

We shall use these facts extensively in what follows.

 Pricing of track time 201

3. THE GAME-THEORETIC EQUILIBRIUM MODEL

In this section, a mathematical programming model of the train scheduling problem

is structured as an N-player, noncooperative game with nondisjoint strategy sets by as-
suming that a railroad company consists of a rail authority and several divisions based
on each type of commodity.
The mechanism of an internal rail track market can be explained by considering a
three-stage game for rail operations. In the first stage, a railroad authority determines
the relative importance weight (in dollar terms) for each traffic class and each subdivision
submits its ideal schedules to the rail authority. In the second stage, the rail authority,
who is the leader of the game, charges an adequate user fee to each subdivision based on
its marginal benefit of using a unit of track time. In setting these user fees, the authority
assumes that each division behaves in order to minimize its own schedule deviations with
no consideration of the effect of this action on other network users or the overall well
being of the system. The important point to note here is that the railroad authority is
ultimately unable to directly control the behavior of trains; i.e., the only “lever” the
authority possesses is the track user fee. The strategy sets of each division are nondisjoint
because division k’s train j must know the other trains’ schedules to determine its own
feasible schedule set, but the others cannot determine their feasible strategies without
knowing trainJ% schedule. An additional train or change of an existing train’s schedule
may slow down other trains and these additional delays impose a time cost on the move-
ment of trains. In the third stage, each subdivision revises its ideal train schedules based
on the individual division’s marketing strategies and the user fees charged by the rail
authority. Then, these revised train schedules will be used at the first stage of the next
cycle. In this Stackelberg-like (leader-followers) game, we develop the pricing mechanism
which would be used in the second stage in order to charge an adequate rail track user
fee. This fee will explicitly recognize the externalities in the market and the desire by each
division to “push” the operating department to have its trains perform as close to its ideal
schedules as possible.
In a standard normal form, noncooperative game, it is usually assumed that players
can only affect the utilities of the other players but not their feasible sets. In the general-
ized Nash game, the other players influence player J’S utility level as well as his or her
feasible strategies. This type of problem has several names in the literature: social equilib-
ria games, pseudo-Nash equilibria games, and generalized Nash equilibria (GNE) games
(Ichiishi, 1983; Harker, 1991). Each division chooses its own strategies so as to maximize
its utility, defined as minimizing deviations from the ideal train schedules.
Given the ideal train schedules, which consist of the arrival and departure times at
major points on the railroad, the purposes of the model are twofold. First, the model
provides the railroad’s optimal track user fees per unit of track time for each train. These
user fees are inferred from the dual variables of associated track time constraints for
each train; therefore, fees are based on the relationships among the value of time cost
and transit time. The time cost imposed by the other trains is: (a) much larger for trains
with tighter schedules and carrying higher-valued products, and (b) proportionally greater
the higher the density of traffic. Second, the model develops train operating plans that
will achieve the times stated in the ideal schedule as best as possible given the delays
encountered by each train as a result of interference from other trains. This second
outcome is a representation of how the operating department would respond to the
competitive desire of the division to achieve its ideal schedules. From this representation
of the operation, the user fees will emerge.
In this section, we shall first list the assumptions and then define the notation used
herein. Note that all of the assumptions of the delay model in Harker and Hong (1990)
are adopted because the impedance which train k imposes on the other trains is the major
factor when the division in charge of train j determines its own schedule. The additional
assumptions for this game-theoretic model can be stated as the following:

l Each division behaves in a noncooperative manner in which each division takes the
other divisions’ train schedules as given when deciding on its own schedule.
TR(B) 28:3-8
202 P. T. HARKERand S. HONG

l Because of the schedule delay imposed on each other, division k must know the others’
schedules to determine his or her own feasible schedule set Fk(x), but the others cannot
determine their feasible schedules without knowing division k’s schedule. Therefore,
the strategy sets of each division in the train scheduling game are nondisjoint and the
other divisions, K\k, can restrict the feasible schedule set and affect the utility level of
division k.
l Each division strives to have the operating group make his or her train schedule as
close to the ideal schedule as possible.
l The rail authority’s payoff (utility) is based solely on making the system feasible; i.e.,
to assure that capacity is never exceeded.
l The weight parameters for train schedule tightness and priority are based on overall
railroad operation policy defined by long-term planning issues.
l Every train has its own ideal schedule (arrival and departure time at every major point
on the network). Minor points on the railroad are assumed not to be scheduled.
l Train schedules within a finite planning horizon are considered.
l Yard delay or capacity is not included in this model and hence, constant minimum
yard dwelling times are used as an input. This assumption can easily be relaxed by the
inclusion of a yard model similar to the line delay model described in Harker and Hong
(1990).

To state the train scheduling model for division k, let us define the following set of
indices:
K = set of railroad divisions,
k = an index for a division, k E K,
K\k = set of divisions except division k,
T = set of trains that travel on the system during the planning horizon,
t E an index for a train, t E T,
tic = a train run by division k, t, E T,,
Tk = set of trains run by division k,
Tk\tk = set of trains run by division k except for train t,,
TK\Tk = set of trains run by divisions other than division k, K\k,
I = set of major yards or reporting stations where trains are scheduled,
i = a departing yard, i E I,
j = an arrival yard, j E I,
V = set of track links on the railroad,
V E a track link between i and j, v = (i,j) E V.
The input data required by the model are:
ZAjtc I the ideal train arrival time at yard j for division k’s train tk (defined by each
division),
ID,,, = the ideal train departure time at yard i for division k’s train tk,
W& = the relative importance weight parameter for the closeness of the actual arrival
time of train tk at yard j to IA, (defined by the rail authority),
Wfk = the relative importance weight parameter for the closeness of the actual depar-
ture time of train tk at yard i to ZDifk,
D Wi, = the minimum yard dwelling time that a train t must spend at yard i.

Values for the following variables will be generated by the model:

A, = the actual arrival time of train t at yard j,
Air,\*, = the arrival time of a set of trains T,\t, at yard j,
= (Aj, : t E Tk, t + tk),
D, = the actual departure time of train t at yard i,
DiT& = the departure time of a set of trains Tk\tkat yard i,
= (Dir: t E Tk, t # tJ,
AjTK\t = the arrival time of a set of trains TK,k at yard j,
= (Aj, : t E TK, t B Tk),
 Pricing of track time 203

=
DiTK,k the departure time of a set of trains TK,k at yard i,
= (Di, : t E TK,t B Tk).
The expected running time of a train tk E Tk between yards i and j, Ecij)lt, is a
function of following train schedules:

1. (Ajlk, Dir,): arrival and departure times of a train, tk, between yards i and j,
2. (&&’ DiT,\,,): arrival and departure times of division k’s trains except train tk, Tk\tk,
between yards i and j,
3* @jT,,, sDiTx,J: arrival and departure times of the set of trains, TK,k, between yards i
and j.

Harker and Hong (1990) provided the formulation for the expected train running
time on a partially double track line between two yards. It is a sum of (a) the free running
time, (b) probabilistic meet delays, and (c) probabilistic overtake delays. Rewriting the
formulation for the expected running time with the notation used in this paper, one has:

(3)

3.1 Model formulation

It is assumed that each division behaves noncooperatively and acts in relation with
the operating group so as to minimize its own schedule deviations (or travel cost) with no
consideration of the effect of his or her action on the other network users or the overall
well-being of the system. On the other hand, a railroad authority wants to minimize the
total, weighted sum of train schedule deviations from the ideal schedules. The game-
theoretic model is formulated to capture both the noncooperative behavior of the divi-
sions and the overall system optimization which is desired by the railroad authority.
The solution is based on the postulate that each division makes decisions which
affect the value of the objective function when the other divisions do not change their
feasible decision set via the generalized Nash equilibrium concept. Given all the other
divisions K\k’s schedules, a division k chooses his or her train schedules so as to maximize
its utility. The Nash equilibrium solution will be found when no division has an incentive
to change its train schedule because no better schedule can be obtained through unilateral
changes. It is important to note that no division can individually determine his or her
schedule because division k must know the others’ schedules in order to know his own
feasible schedule set C&(x&.
Division k E K faces the following two operational constraints: one for yard dwelling
time and the other for line travel time. First, if a train tk arrives in yard j, it is required to
stay for at least an amount of time DWj,, before it departs at Dj,,. This constant minimum
dwelling time must be long enough for inbound inspection, classification, assembly, and
outbound inspection. The constraint on the minimum dwelling time is:

Dj,, - Ajtk 1 D Wj,, V, E Ik, tk E Tk. (4)

Second, the arrival time of train tk at yard j must be greater than or equal to the
departure time at yard i plus the expected train running time from yard i to yard j. The
expected running time is the sum of the free running time and the delays that are incurred
through interference with other trains:

(5)

The constraint (5) exhibits why the strategy sets of divisions are nondisjoint. For
example, consider the meet delay between an east bound train E from yard East and a
west bound train W from yard West. When a free running time is 2 hours for train E and
2 and a half hours for train W, respectively, the departure time differences between two
trains must be at least 2 hours in order to avoid a meet delay. Because the expected
204 P. T. HARKER and S. HONG

running time is the sum of free running time and delays incurred from other trains, a
train W (or E) influences a train E (or W) by affecting train Es running time between
yard East and West. Therefore, in the constraint (5), train t, must know the schedule of
the other divisions’ trains, TK,p, in order to know his or her own feasible schedule set.
Due to this issue, a variational inequality formulation of the standard Nash game
cannot be used; one must utilize the generalized Nash game concept and its quasi-
variational inequality representation. To state this formulation, let us define:

A$ = (* * * ,Ajlk, * - *)T, D,, = (* * * ) I&, * * *)T,

4 = W,It, E TdT, Q = (QkItk E Tk)T,
A = (- . . ,Ak, - . .)T, D = (. . . ,& . . .)T
x, = (&Dd=, XK\k = tx,, * * ’ ,xk-I, xk+l, ’ . . ,xK)

x = (* * * ,x,, * * *)T.

The feasible train schedule set of a division k E K is defined by:

nk(XK\k) = {xk = (AkiD,) I(4) and (5) hold}.

Division k’s schedule deviation minimization problem (its utility maximization) be-
comes:
[GAME]

Minimize & ( xk;xK\X_)

subject to: xk = (A,; Dk) E Qk(xKlk). (6)

The first term of the objective function (6) measures the square of the difference between
the actual arrival time and the ideal arrival time, and the second term measures the
difference between the actual departure time and the ideal departure time. Note that any
convex function can be used for the objective function; we use a quadratic function
merely for convenience. The ideal train schedules, ZAjrk and ID+ are the input of the
model and are determined by each division based on its market strategy and operations
history. The relative importance of the trains, described by the weight parameters wA
and wD, defines the overall operating policy determined by the rail authority. Thus, the
smaller the weight parameter, the lower the priority that the train would experience in
real-time operations. Assuming the other trains’ schedules as given, each division k is
able to reduce the expected congestion delays by rescheduling. We penalize both the
earliness and lateness of train schedules (arrival and departure time); in other words,
early arrival, along with late departure, is considered a less desirable train operation.

3.2 Quasivariational inequality formulation

Each division k chooses a strategy & belonging to a closed convex set X, E Bm,
where X = IIk,,Xk G Bkm and X represents the full Cartesian product of strategy sets.
Let Ok : X -+ X, be a point-to-set mapping that represents the ability of the other divisions
K\k to affect the feasible strategies of division k. That is,

n,(x) c X, vx E X.

Then, a vector x* = (* * * , xx, * * *) E X of strategies is said to be a generalized,

pseudo- or social equilibria if no division is able to obtain better train schedules by
unilaterally modifying the chosen train schedule, where X,* E 0, for all k E K.
Division k’s schedule deviation minimization problem is to find an x,$ E n,(x) where
 Pricing of track time 205

rI,(* * * 3-c,,Gx*k+,, * * *) I I&(* * * ,x*k-l,xk,x*k+*, * * *) vx, E&(X). (7)

The first order condition of (6) can be stated as:

where V,&(x,*, x,,,) is the gradient of I& with respect to division k’s strategy vector xt.
The equivalent quasivariational inequality formulation is the summation of the indi-
vidual divisions’ first order conditions (8):
find x* E Q(x*) = I&@(x&) such that

c v,~k(x*,,G\k)T(xk-
ksK
6) 2 0x E Q(x*),

or
F(x*)T(x - x*) 2 0 vx E iI( (10)

where Fk(x*) is the gradient of II&,*, xK\k) and F(x*) = (. . * , Fk(xk*)r, . - *)‘. The
vector x* is a generalized Nash equilibrium if it is feasible with respect to the correspon-
dence s2 and satisfies (10). Writing the gradient of IIk explicitly, the quasivariational
inequality formulation for the generalized Nash equilibrium problem can be written as:
[QVII

c
ksK
c[c w;J&I,
r,eT, jd
- ZAj,,)(Ajtk - A7tk) + C wfk(D:tk-
El
zDirk)(Ditk -
1
O*itkt,>

1 0 V(A,D) E Q(A*,D*). (11)

3.3 Derivation of track prices

The model described by (9) only involves the divisions and not the rail authority.
When solved, this model generates equilibrium arrival and departure times x* as well as
dual variables for all the constraints (4) and (5). Given this model and solution, how are
track prices defined?
Consider the delay constraint (5). Each division on the railroad requests arrival and
departure times and it is the rail authority that, through its operating policies, generates
the actual times. These operating policies are embodied in the delay function described in
Section 2.1 and thus, are used to define the right-hand side of (5). That is, constraint (5)
defines the authority’s response to the arrival and departure time requests, generating the
actual, feasible, times. Put another way, the policies of the authority define the capacity
of the rail lines via the delay function.
The utility function for the authority, as stated in Section 3, is to make each line
feasible; i.e., have (5) hold. Because the divisions do not operate the railroad, they never
know the feasibility of each line; only the authority has this information. The authority
in turn chooses track prices to force (5) to hold. If this is the case, where are the prices in
the model?
As shown by Harker (1991), this type of model leads to a solution without explicitly
placing the prices into each divisions’ utility/payoff function. First, assume that (5) holds
and solve the resulting equilibrium problem. This solution will obviously satisfy (5), and
generate an associated dual price for (5). If this constraint were now removed from each
divisions’ problem (6), but the dual prices times the track occupancy times were added to
their payoff functions II, no division would have an incentive to change. That is, the dual
prices can be used to “price out” the capacity/delay constraint in equilibrium. Thus, the
dual prices are exactly the prices that the rail authority would need to charge to the
divisions to assure that (5) holds in equilibrium. The reader is referred to Harker (1991)
for the technical details of this situation, and Hong and Harker (1992) for an application
of this concept to airport pricing.
206 P. T. HARKER and S. HONG

4. THE SYSTEM OPTIMIZATION MODEL

Unlike the previous model in Section 3, the main consideration of the model in this
section is to minimize the overall sum weighted train schedule deviations without taking
accounting for constraints imposed on operation by the posited divisional structure of
the railroad. The problem is seen from the viewpoint of a central operations department.
In this case, the authority not only assures feasibility, but chooses all the schedules in
order to optimize system-wide performance. Of course, such a single entity would achieve
a better (least cost) solution for the system as a whole. However, the basic thesis herein is
that such a global optimality may not be achievable in practice due to organizational
restrictions. Thus, this model will not serve as an operational model, but rather, as a
benchmarking tool to judge how close a market-based, divisional structure can come to
global optimality.
To state the system optimization train scheduling model, one can use the same
notation as in Section 3, except that the divisional index k is dropped throughout. Using
these notations, a nonlinear programming problem can be formulated to obtain the
system optimization train scheduling model. The first constraint is the dwell time restric-
tion.

Dir - Ai, 2 D Wit Vi E 1, t E Ty (12)

and the second constraint is the expected running time restriction:

Ajt - Di, 2 Ecij,t(Ajt, Djt, A,,\,, DlT\,) VI E 1, t E T. (13)

Then, the feasible set for the single operation department is defined by:

n = {x = (A; D)l (12) and (13) hold}.

The formulation is:

[SYSI

Minimize H(x)
Total Schedule Deviation from the “Ideal” Train Schedule

subject to: x = {(A;D) EL?}. (14)

Note that, as in the last section, we use a quadratic function for the objective
function because it is convenient; however, any convex function can be used.

5. SOLUTION PROCEDURE

To solve the system optimization problem, any nonlinear program system that can
handle nonlinear constraints, such as the MINOS (Murtagh & Saunders, 1985) package,
can be employed. Thus, this problem is straightforward to solve. With respect to the QVI
problem (ll), the result of Harker (1991) showed that any algorithm for a variational
inequality, such as Newton’s method, projection or diagonalization algorithms, can be
employed for solution of this special case of the quasivariational inequality. The ap-
proach for solving [QVI] of (11) is to use a diagonalization algorithm in conjunction with
a nonlinear programming package such as MINOS (Murtagh & Saunders, 1985); see
Harker and Pang (1990) for details.
 Pricing of track time 207

6. APPLICATION TO A MAJOR U.S. RAILROAD

This section reports on the application of the methodology just developed to a data
set from a major Class I railroad. The purpose of this exercise is to show the applicability
of the rail track capacity pricing and scheduling model to realistic situations and to
provide some insights into the results which the model generates. This application illus-
trates the nature of the coupling and interaction between the division in railroad opera-
tions. As explained in Section 3, a railroad in this study consists of a rail authority and N
operating divisions (i.e., one for each type of commodity). Thus, in our game-theoretic
context, there are N + 1 players. For this application, there are six divisions (Amtrak,
double stack intermodal, single stack intermodal, priority, secondary, and regional
freight trains) and a rail authority. The train scheduling decisions of each division are
directly related to each other due to congestion externalities. The rail authority manages
the rail track space and charges user fees to each division.
The rail network for our empirical study consists of five major yards and four links
which connect the yards. A total of 31 trains (by their ID number) are running in the
time period under investigation: 16 westbound trains and 15 eastbound trains. In this
case study, we use the abbreviations for each train class: Amtrak for Amtrak trains,
DSTK for double stack intermodal trains, INTERMOD for single stack intermodal
trains, PRIFRT for priority freight trains, SECFRT for secondary freight trains, and
REGNFRT for regional freight trains. Hong (1991) contains a detailed description of the
data for and results of this application.
It is assumed that the ideal train schedules are determined by each division at the
beginning of the operating game. The train class priorities and relative schedule impor-
tance parameters are provided by the rail authority based on extensive empirical (albeit
confidential) analysis. Thus, the priority and cost numbers used herein are disguised.
Amtrak (passenger train) has the highest priority (l), and REGNFRT (regional freight
train) has the lowest priority (0.5). Trains with higher priorities and tighter schedules are
less likely to be delayed and thus, have more incentive to pay for a higher transport
service quality because they put a higher value on their travel time. Train routes (origins
and destinations) are not altered in this study.
Two different results (one for the game-theoretic solution and the other for the
system optimization solution) are compared. Variations of these two approaches, caused
by altering certain track capacities and train schedules, is discussed in the following
subsections. In particular, three different cases of internal markets for track resources
are studied: the first case with the existing capacity and train schedules, the second case
defined by varying track capacities, and the third case which is created by changing the
ideal train schedules. The track user fees and the actual train schedules are derived, and
the game-theoretic and system optimization solutions are compared for each case.

6.1 Case I: The basic case

The first case is to estimate the rail track capacity price based on the given train
schedules and track capacity data presented in Hong (1991); Table 1 summarizes the
results. The results are consistent with what one would expect: (1) Amtrak and DSTK

Table 1. Summary of the basic case

Objective Value Track Prices

Train Type System Game System Game

Amtrak 2369.00818 2369.02119 4405.71399 4404.45990

DSTK 239.92891 240.05180 991.54893 977.03893
INTERMOD 77.99300 78.00885 381.56627 379.91672
PRIFRT 0.00113 0.00113 O.OOOOO O.OOOOO
SECFRT 0.00050 0.00050 O.OOOOO O.OOOOO
REGNFRT 1.05793 0.96891 O.OOOOO O.OOOOO
Total 2687.98967 2688.05238 5778.82919 5761.41555
208 P. T. HARKERand S. HONG

pay higher track user fees because one would expect a relatively high value of time for
their movements, (2) PRIFRT, SECFRT, and REGNFRT trains are paying nearly zero
track prices because they can make their ideal schedules by using the slack (slack =
scheduled running time minus free running time) in their schedules and have a low value
of time. While Amtrak, DSTK, INTERMOD trains are scheduled with little slack (e.g.,
the scheduled running time of one Amtrak train is 34 minutes fess than its free running
time), PRIFRT, SECFRT, and REGNFRT trains usually run with a great deal of schedule
slack (e.g., 5 hours and 37 minutes slack for a priority freight train). PRIFRT, SECFRT,
and REGNFRT can more easily make their ideal schedules than Amtrak and DSTK and
thus, have little incentive to pay for high-valued track time. Although the objective value
of the game-theoretic solution is a little higher than that of the system optimization
solution, the DSTK division is paying less track user fees in the game-theoretic approach.
That is, the DSTK division can do better in the internal market than in the system
optimization solution.

6.2 Case 2: The effects of varying track capacities

The effects of track capacity changes are considered here. The number of sidings on
the single track sections in each of the four lanes are reduced by three in this case;
however, the proportion of single and double track remains the same. Thus, the net-
work has a total of 12 less sidings. A summary of the results of this case can be found in
Table 2.
Both the objective values and the track prices increase from the basic Case 1 due to
the reduced track capacity (5778.83 to 5998.57 for the system optimal solution and
5761.41 to 6306.07 for the game-theoretic solution). Amtrak pays almost the same track
user fees in both situations, DSTK pays about $14 less in the game-theoretic approach,
and the amount of track user fees for INTERMOD in the game-theoretic solution are
approximately $323 higher than in the system optimization solution. Thus, INTERMOD
shows the biggest track price increase from the capacity change and the biggest track
price difference between the system optimization solution and the game-theoretic solu-
tion among the six operating divisions.

6.3 Case 3: The effects of schedule changes

The third case involves studying the effect of train schedule changes. After the basic
train schedules were reviewed, it was recognized that (a) some trains are running with
more than 4 hours slack; and (b) some trains can never make their schedules because
their scheduled running time is less than their free running time. Although the same track
capacity as in the basic case is used in the current case, the train schedules are modified
(depending on their train type) according to the following rule: 20 minutes slack for
Amtrak, 30 minutes slack for DSTK, 40 minutes slack for INTERMOD, 60 minutes
slack for PRIFRT, 75 minutes slack for SECFRT, and 90 minutes slack for REGNFRT.
In this case, the scheduled running times of each train are equal to the free running time
plus this slack time. Thus, the new ideal schedules for each division are most likely
feasible (see Table 3).
The objective value of Case 3 is the highest among the three cases presented because
the schedule changes in this case create the most congested network. As in Cases 1 and 2,

Table 2. Summary of the effects of varying track capacities

Objective Value Track Prices

Train Type System Game System Game

Amtrak 2410.54213 2470.46983 4500.50028 4499.09992

DSTK 271.58101 270.72500 1072.06946 1058.23486
INTERMOD 90.66630 90.4693 1 426.00093 748.73363
PRIFRT 0.00161 0.00161 O.OOOOO O.OOOOO
SECFRT 0.00070 0.00069 O.OOOOO O.OOOOO
REGNFRT 0.89194 2.35887 O.OOOOO O.OWOO
Total 2833.68429 2834.02531 5998.57067 6306.06741
 Pricing of track time 209

Table 3. Summary of the effects of schedule changes

Objective Value Track Prices

Train Type System Game System Game

Amtrak 1138.27923 1135.64944 4409.88142 2773.85482

DSTK 890.60489 941.03970 4268.90664 4166.96699
INTERMOD 794.79615 901.43556 5638.11683 3120.07254
PRIFRT 26.07643 161.16208 105.68595 525.39929
SECFRT 147.17914 145.98686 167.12148 136.59936
REGNFRT 2.85454 123.08000 O.OOOOO 15.13531
Total 2999.79038 3408.35365 14587.71232 10738.02831

the objective value of the game-theoretic solution is higher than that of the system
optimization solution. In these cases, the rail network was not very congested, so that the
objective values were not much different between the system optimization solution and
the game-theoretic solution. By reducing the slack times (i.e., tighter train schedules),
trains are experiencing more schedule delays so that every train operating group, except
the REGNFRT division, pays positive track user fees in the system optimization ap-
proach. Amtrak, DSTK, INTERMOD, and SECFRT pay less in the game-theoretic
solution, while PRIFRT and REGNFRT pay more. Amtrak and INTERMOD manage to
pay much lower track user fees in the game-theoretic approach. In particular, Amtrak
pays the lowest track price in Case 3. Note that Amtrak had no slack time in the previous
two cases and even minus slack time (scheduled running time is less than free running
time) for some sections. Thus, the Amtrak division takes advantage of the proposed
schedule modifications.

6.4 Practical implications of the results

What are the implications of the empirical results which were just presented? First, a
comparison of the game-theoretic and system optimization solutions (Table 4) shows
how close the market mechanism (the game-theoretic approach) can come to global
optimality. Because the system (or global) optimum solution is unlikely to be achieved in
practice due to the necessity of a tightly controlled, U-form structure, the system (or
global) optimization solutions are used to benchmark the performance of the proposed
market mechanism.
In Table 4, the objective values of the game-theoretic solution are always higher
because, in this approach, each group behaves noncooperatively and minimizes his or her
own travel cost with no consideration of the effect of this action on the other network
users or the wellbeing of the system as a whole. Thus, the minimization of system-wide
travel cost is not equivalent to economically rational decisions on the part of the operating
divisions. When more than half of the trains are paying zero track user fees in Cases 1
and 2, the objective values are not much different between the game-theoretic solution
and the system optimization solution. That is, the market-based solution nearly obtains
global optimality in Cases 1 and 2. In Case 3, when most trains are paying track user fees
(experiencing a more congested network than in Cases 1 and 2), the game-theoretic
solution has a much higher objective value (system optimization solution = 2999.79,
game-theoretic solution = 3408.35); i.e., a 13% increase from the system optimization

Table 4. Summary of results

Objective Value Sum of Track Prices

Train Type System Game System Game

Case 1 2687.98967 2688.05238 5778.82919 5761.41555

Case 2 2833.68429 2834.0253 1 5998.57067 6306.06741
Case 3 2999.79038 3408.35365 14587.71232 10738.02831

TR(6) 28:3-C
210 P. T. HARKER and S. HONG

objective value. Thus, the market mechanism seems to work better in Cases 1 and 2.
However, a 13% difference may be quite acceptable given the high coordination costs
associated with the U-form structure needed to implement the globally optimal solution.
In Case 1, the sum of the track user fees are almost same in both approaches. In
Case 2, the sum of track user fees are higher in the game-theoretic solution. Trains which
pay track user fees in Case 1 pay more in Case 2, and trains which pay zero track user
fees in Case 1 also pay zero track user fees in Case 2. Although track capacities are
reduced in Case 2, PRIFRT, SECFRT, and REGNFRT can make their ideal schedule
easily (without schedule delays). Because trains (PRIFRT, SECFRT, and REGNFRT)
with zero track user fees have a lot of slack time, they barely miss their scheduled time.
Moreover, they have a lower value of travel time. It is worth noting that the relationship
between the value of time and adequate track user fees persists in these examples.
The upshot of the discussion and evidence presented in this paper is that the proposed
market mechanism (the game-theoretic method) with a decentralized organizational
structure is both feasible and desirable. The game-theoretic solution is unlikely to result
in a significant difference in the objective values and the track user fees from the system
optimization solution (global optimality), which would be nearly impossible to achieve in
practice. Thus, the results in this paper imply that market mechanism (game-theoretic)
solution can be a good approximation of the system optimal solution. Except for Case 3,
which is the most congested case, the objective values of game theoretic approach in
Cases 1 and 2 are very close to the objective values of the system optimization approach.
The results show that true time costs of traffic on a line, if properly charged to the
users, can be used to plan schedules. Each division can modify its ideal train schedules to
either pay less track user fees or to experience fewer meet and overtake delays. For
example, Amtrak pays higher track user fees in Cases 1 and 2 than in Case 3 because
Amtrak has very tight train schedule (i.e., scheduled running times of Amtrak trains in
certain sections are less than their free running time) and high value on its travel time.
Thus, Amtrak has an incentive to either shift its schedules to less congested times or to
modify its schedule running times appropriately. In Case 3, Amtrak increased its sched-
uled running time (i.e., free running time plus 20 minutes slack). Then, Amtrak actually
managed to pay the least amount of track user fees among three cases due to these
schedule changes. (Note that these schedule changes in Case 3 are based on the results of
Case 1.) If we assume that each division can easily change its schedules, they could pay
zero track prices by constantly updating their train schedules (e.g., adding schedule slack
or moving its schedules to less congested time period). However, this would not always
happen due to market constraints. For example, when a train needs to be reassembled at
a yard, it can depart only after other trains arrive. Thus, the marketing division may
have special reasons to keep a certain train schedule, and incur positive track user fees.
The proposed model can also be used for track price and schedule negotiations with
a new user. A new user comes to the rail authority with its train schedule, and the rail
authority can estimate how much additional delay will occur and how high a track price
the new user needs to pay. However, this track price is not based on fully allocated costs
but rather, only the costs associated with time. In other words, we ignored track-wear
maintenance costs, crew costs, fuel costs, and capital costs. The proposed model only
considers the cost associated with the additional travel time imposed on existing users by
an additional user. In order to be used in practice, the additional, nontime related costs
must be included in the final computation of track user fees.
Finally, our results do not deal with the implementation details of a track pricing
scheme; only the desirability of such a system has been shown through example. Whether
an auction-like system or a monthly/weekly/daily/etc. pricing method should be used to
implement the market mechanism remains to be addressed.

7. CONCLUSION

In this paper, a series of realistic examples were presented to illustrate the applicabil-
ity of market-based mechanism for track pricing show how close it could come to the
system optimal solution. These examples arose from a realistic situation on a major Class
I railroad.
 Pricing of track time 211

We have estimated the track user fees charged to each traffic class based on their
marginal benefit of using a unit of track time. When a new user wants to use the rail
track with a certain train schedule, the model can provide the new user’s track fee and the
effect (i.e., additional congestion) on the existing users. Track user fees will explicitly
recognize the externalities in the internal rail track market. An optimal pricing rule would
account for the congestion externality associated with various demand levels. The service
qualities obtained by each user are endogenous; they depend on how many other users
want to use the facility and have the same or higher track valuation. A user with a high
value on using a facility usually pays a high user charge and, at the same time, may pay a
low user charge if the demand for the facility by other users is low.
Possible extension to the work presented in this paper must first involve a more
careful implementation of track prices. Track prices were determined based on the as-
sumption that the most important cost is that associated with increased travel time by an
additional track user. If we include track-wear maintenance costs, crew costs, fuel costs
and capital costs in the model, we would get more practical and realistic track prices.
Secondly, a multiple use train is not considered in this paper. When a train carries more
than one type of commodity (e.g., agricultural products and industrial products), each
customer group (by its commodity type) has different values on its travel time. Then, a
question of how much the rail authority charges a multiple use train remains to be
answered. Finally, when track prices are set in practice, the time frame for schedule
changes is important. If each division could react (i.e., change train schedules) within a
day, the rail authority would change the track prices for each division every day. Because
this may be impossible in practice, the question of how often each division can change its
schedules and the rail authority its prices also remains to be answered.

Acknowledgements-This work was supported by the National Science Foundation Presidential Young Investi-
gator Award ECE-8552773 and by the Burlington Northern Railroad. The comments and advice of W. Bruce
Allen and Edward K. Morlok are gratefully acknowledged.

REFERENCES
Carlin A. and Park R. E. (1970). Marginal cost pricing airport runway capacity. American Economic Review
60, 310-319.
Chen B. and Harker P. T. (1990). Two moments estimation of the delay on single-track rail line with scheduled
traffic. Transportation Sci. 24, 261-275.
Crew M. A. and Kleindorfer P. R. (1987). The economics of public utility regulation. Cambridge, MA: The
MIT Press.
De Vany A. and Saving T. R. (1980). Competition and highway pricing for stochastic traffic. Journal of
Business 53, 45-60.
Hansson L. and Nilsson J. E. (1991). A new Swedish railroad policy: Separation of infrastructure and traffic
production. Transpn. Res. 25A, 153-159.
Harker P. T. (1984). A variational inequality approach for the determination of oligopolistic market equilib-
rium. Mathematical Programming 30, 105-l 11.
Harker P. T. (1991). Generalized Nash games and quasi-variational inequalities. European Journal of Operation
Research 54, 81-94.
Harker P. T. and Hong S. (1990). Two moments estimation of the delay on a partially double-track rail line
with scheduled traffic. Journal of Transportation Research Forum 31, 38-49.
Harker P. T. and Pang J.-S. (1990). Finite-dimensional variational inequality and nonlinear complementarity
problems: a survey of theory, algorithms and applications. Mathematical Programming 48, 161-220.
Hong S. (1991). Incentive-compatible capacity pricing for congested transportation facilities: A game-theoretic
approach. Unpublished Ph.D. dissertation, Regional Science Department, University of Pennsylvania, Phil-
adelphia, PA.
Hong S. and Harker P. T. (1992). Air traffic network equilibrium: toward frequency, price and slot priority
analysis. Transpn. Res. 26B, 307-323.
Ichiishi T. (1983). Game theory for economic analysis. New York: Academic Press.
Levin R. C. (1981). Railroad rates, profitability, and welfare under deregulation. Bell Journal of Economics
12, l-26.
Morrison S. A. (1983). Prices and investment level for airport runways. In T. E. Keeler (ed.) Research in
Transportation Economics 1, 103-130.
Morrison S. A. (1986). A survey of road pricing. Transpn. Res. 20A, 87-98.
Murtagh B. A. and Saunders M. S. (1985). MINOS 5.0 user’s guide. Systems Optimization Laboratory, Stan-
ford University, Stanford, CA.
Overbey D. L. (1982). Railroads: Thefreeenterprise alternative. Westport, CT: Quorum Books.
Panzar J. C. (1976). A neoclassical approach to peak load pricing. Bet/ Journal of Economics 7, 521-530.
Petersen E. R. (1974). Over-the-road transit time for a single track railway. Transportation Sci. 8, 65-74.
Petersen E. R. (1975). Interference delays on a partially double-tracked railway with intermediate signalling.
Transportation Research Forum, Toronto, Canada.
212 P. T. HARKERand S. HONG

Pigou A. C. (1920). The economics of welfare. London: MacMillan.

Small K. A., Winston C. and Evans C. A. (1989). Road work. Washington, DC: The Brookings Institute.
Vickrey W. (1969). Congestion theory and transportation investment. American Economic Review 59, 251-
260.
Viton P. A. (1980). Equilibrium short-run marginal cost pricing of a transport facility: The case of the San
Francisco Bay Bridge. Journal of Transportation Economics and Policy 14, 185-203.
Walters A. A. (1968). The economics of road user charges. Baltimore, MD: Johns Hopkins Press.
Williamson 0. E. (1975). Market and hierarchies: Analysis and antitrust implications. New York: The Free
Press.