This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS 1

Back-Pressure Traffic Signal Control With Fixed and

Adaptive Routing for Urban Vehicular Networks
Ali A. Zaidi, Balázs Kulcsár, and Henk Wymeersch, Member, IEEE

Abstract—City-wide control and coordination of traffic flow methods, however, requires centralized decision making for all
can improve efficiency, fuel consumption, and safety. We consider intersections based on the traffic-related measurements. In ad-
the problem of controlling traffic lights under fixed and adap- dition to these traffic-adaptive signal control implementations,
tive routing of vehicles in urban road networks. Multicommodity
back-pressure algorithms, originally developed for routing and other centralized traffic signal control algorithms have recently
scheduling in communication networks, are applied to road net- been proposed [8]–[11] using different approaches from control
works to control traffic lights and adaptively reroute vehicles. The theory, such as linear quadratic regulator, robust control, and
performance of the algorithms is analyzed using a microscopic model predictive control. The traffic signal control problem has
traffic simulator. The results demonstrate that the proposed multi- also been studied under game theoretic formulations [12].
commodity and adaptive routing algorithms provide significant
improvement over a fixed schedule controller and a single- In contrast to the many centralized approaches for traffic
commodity back-pressure controller in terms of various perfor- signal control, the literature on decentralized solutions, which
mance metrics, including queue length, trips completed, travel would be very useful especially for large urban areas, is scarce.
times, and fair traffic distribution. Recently, researchers in the transportation and control commu-
Index Terms—Back pressure control, communication system nities have proposed different traffic-adaptive scalable and dis-
traffic control, fixed and adaptive vehicle routing, queueing analy- tributed methods [13]–[20], where the general idea is to solve
sis, traffic control, traffic signaling. a separate optimization problem for every intersection. These
per-intersection optimization problems are loosely coupled via
I. I NTRODUCTION real-time traffic conditions. The implementation of these con-
trollers requires either the knowledge of expected traffic load on

I N urban road networks, traffic congestion is a major problem

leading to time loss, pollution, and accidents [1]. Vehicle
flows in such networks are controlled by traffic lights and are af-
the links associated with the intersection during the next cycle,
or the difference between the traffic loads on the links asso-
ciated with the network. Many of these schemes are inspired
fected by route choices that the drivers make. Traffic conditions by scheduling and routing algorithms in wireless networks,
can thus be improved by the development of efficient traffic sig- in particular the well-known back-pressure scheme from [21].
nal control and route selection methods, see the papers e.g., [2], Back-pressure is a decentralized scheme that can provide maxi-
[3] for a survey to the field. Traditionally, traffic light controllers mum network throughput under the assumption that all links in
follow a pre-defined optimized schedule [4], which may result the network have infinite capacities (it is in fact optimal in the
in a poor performance under time-varying traffic conditions and sense of supporting maximum traffic arrival rates that guarantee
under very high traffic demands. This problem can be alleviated stability of queues in a stochastic sense). This idea was first
through adaptive traffic signal controllers, such as SCOOT, adapted to urban road networks in [15], where it was shown that
UTOPIA, SCATS, or RHODES [5]–[7]. In these adaptive traffic significant performance gains can be achieved in terms of net-
controllers, real-time measurements are collected using on-road work queue lengths by employing a back-pressure scheme for
detectors. Based on these measurements, either the parameters signal control. It was also shown to provide good performance
(splits, offsets, cycle-length) of the signal plans are adjusted on compared to the fixed time schedule controllers, when the links
a cycle-to-cycle basis or a best signal plan is selected from a have finite capacities. However, [15] does not dynamically re-
pre-defined set of signal plans. The implementation of these route vehicles, leading to local bottlenecks in the road network.
In the literature, there exist different methods for route selec-
tion based on different performance metrics such as shortest
Manuscript received November 9, 2014; revised April 5, 2015, July 20, 2015, path, shortest travel time, congestion minimization, etc. For
and October 19, 2015; accepted December 10, 2015. This work was supported vehicle routing, the fundamental challenge is that the traffic
by a SAFER grant on “enhanced traffic flow optimisation” and by the Transport demand and vehicle departure times at different links in a road
Area of Advance at Chalmers University of Technology, Gothenburg, Sweden.
It was also supported in part by the European Research Council under Grant network are not known a priori. However, the real-time traffic
258418 (COOPNET). The Associate Editor for this paper was W. Fan. information along with the historic traffic data can be used to
A. A. Zaidi is with Radio Access Technologies, Ericsson Research, 16480 anticipate traffic conditions and has been shown to be very
Stockholm, Sweden (e-mail: [email protected]).
B. Kulcsár and H. Wymeersch are with the Department of Signals and useful in devising route selection methods [22]. Recent works
Systems, Chalmers University of Technology, 41296 Gothenburg, Sweden on traffic-adaptive routing methods include [23], [24].
(e-mail: [email protected]; [email protected]). In this paper, we extend [15] by performing both traffic-
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. adaptive signal control and routing, under back-pressure based
Digital Object Identifier 10.1109/TITS.2016.2521424 control methods. In particular, rather than a single-commodity
1524-9050 © 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

back-pressure scheme with fixed routes as in [15], we apply a

multi-commodity (one commodity per destination) version of
the back-pressure scheme [25], [26] under both fixed and adap-
tive route selection. Our results demonstrate that the proposed
schemes can provide significant performance gains.
The remainder of this paper is organized as follows. In
Section II, we mathematically formulate the problem of traffic
signal control and adaptive routing of vehicles. The algo-
rithms based on back-pressure multi-commodity schemes are
proposed in Section III. Performance of these algorithms is
analyzed with a detailed discussion in Section IV. Finally, the
key findings are summarized in Section V along with directions
for future work on this topic.

II. P ROBLEM F ORMULATION

Fig. 1. A four-way junction with 8 roads and 12 possible traffic movements.
A. Road Network
Consider an urban road network comprised of N links/roads
and L junctions (signalized intersections). We model the
network as a directed graph G = (R, J ), where R =
{R1 , R2 , . . . , RN } is the set of links and J = {J1 , J2 , . . . , JL }
is the set of junctions in the road network. A vehicle exoge-
nously enters the network from a certain link (origin), travels
along one or more links in the network and finally leaves the
network at a certain link (destination). Thus, for each vehicle in
the network, there is an associated origin and destination pair.
All vehicles that have a common origin and destination pair
constitute a flow f . Let F be the set of all flows in the network
and let (o(f ), d(f )) be the origin—destination pair for a flow
f ∈ F , where o(f ), d(f ) ∈ R. Let λf (t) be the rate at which
vehicles associated with flow f exogenously enter o(f ) at dis-

crete time slots t ∈ N, with limT →∞ 1/T Tt=1 E[λf (t)] = λf .
We assume that the flow arrival processes are independent of
each other and also independent across time slots and have finite Fig. 2. Typical phases through a four-way junction.
second moments. At any time t, let Qab (t) be the number of
vehicles queued  in a link Ra to move to an adjacent link Rb and
let Qa (t) = b Qab (t) be the total queue length at link Ra .

B. Traffic Phase Switching

Each junction has certain traffic movements associated with
it. A traffic movement through a junction corresponding to the
vehicles exiting Ra and entering Rb is denoted by the pair
(Ra , Rb ). Let Mi be the set of all traffic movements through
a junction Ji . Consider an example of a four-way junction in
Fig. 1, where there are twelve possible traffic movements. The
set of all possible traffic movements for this four-way junction
Fig. 3. Typical phases through a three-way junction.
is given by
M = {(R3 , R8), (R3 , R1), (R7 , R4), (R7 , R6), (R3 , R6), (R7 , R1), phases in Fig. 3. Typical phases of the four-way junction are
given by
(R5 , R4),(R5 , R1),(R2 , R6), (R2 , R8), (R2 , R4),(R5 ,R8)} . (1)
p1 = {(R3 , R8 ), (R3 , R1 ), (R7 , R4 ), (R7 , R6 )}
A subset of traffic movements that can occur simultaneously
through a junction constitute a phase. Let Pi = {pi1 , pi2 , . . . , pil } p2 = {(R3 , R6 ), (R7 , R1 )}
be the set of all possible phases through a junction Ji . As
p3 = {(R5 , R4 ), (R5 , R1 ), (R2 , R6 ), (R2 , R8 )}
examples, consider a four-way junction with four possible
phases in Fig. 2 and a three-way junction with three possible p4 = {(R2 , R4 ), (R5 , R8 )} (2)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZAIDI et al.: TRAFFIC SIGNAL CONTROL WITH FIXED AND ADAPTIVE ROUTING FOR VEHICULAR NETWORKS 3

and for the three-way junction are given by follow fixed routes but ignoring the fact that all vehicles in
the network have different destinations (single-commodity back
p1 = {(R15 , R13 ), (R13 , R18 ), (R17 , R16 )} pressure scheme). We now propose to employ multi-commodity
p2 = {(R17 , R13 )} back-pressure schemes for traffic signal control with fixed as
well as adaptive routing of vehicles. In contrast to the single-
p3 = {(R14 , R16 ), (R14 , R18 )} . (3) commodity scheme where the vehicle queue length information
has to be known on a per-link basis, the operation of multi-
Furthermore, we assume that with every possible movement commodity schemes under fixed and adaptive routing requires
(Ra , Rb ) through a junction, there is a rate sab (t) with which queue length information on a per-flow and per-destination
vehicles can flow through the junction. That is, sab (t) is equal basis. Since the numbers of origins and destinations in a road
to the number of vehicles that can go from link Ra to Rb if a network are normally very big, it is not possible to maintain
phase p is activated, where (Ra , Rb ) ∈ p. physically separate vehicle queues on a per-flow and per-
destination basis. In order to tackle this issue, we utilize the
C. Routing of Vehicles concept of virtual queues, which is essential for the operation
of the proposed multi-commodity back-pressure traffic control
We consider that each vehicle that enters the network has a schemes in road networks.
fixed destination but the route it takes towards the destination
may be either fixed or variable. That is, we have two cases:
A. Virtual Queues and Virtual Vehicles
1) Fixed Routing: In fixed routing we assume that all vehicles
that have a common origin and destination, follow the Following the wireless networking approach in [25], [26],
same route. That is, for all vehicles belonging to a certain we introduce virtual traffic and virtual queues (referred to as
flow f in the network, the route is fixed. Let L(f ) be the shadow queues in [26]) in the road network. For each vehicle
set of links forming the route of flow f . that exogenously enters a link in the network, we generate a
2) Adaptive Routing: In adaptive routing, the route of each virtual vehicle with probability one and another virtual vehicle
vehicle is adapted with traffic conditions. Thus, the vehi- with probability  > 0. Hence, for any flow f in the network,
cles with a common origin and destination pair may not the arrival rate of virtual traffic is (1 + )λf (t). The reason
necessarily follow the same route. We consider that for for introducing  here is explained in Section III-D1. With
every vehicle, the route is dynamically updated at every the virtual traffic we can associate two virtual queues: we
junction. Whenever a vehicle enters a link Ra , its next denote the number of virtual vehicles of flow f on link Ra
movement (Ra , Rb ) through the upcoming junction is de- by Q̃fa ; similarly, we denote the number of virtual vehicles for
cided in real-time and the vehicle joins one of the possible destination d on link Ra by Q̃da . We note that virtual traffic and
queues (lanes) accordingly. For instance, in the example queues are merely counters, which form a fictitious queuing
shown in Fig. 1, when a vehicle enters R2 , it can join one system on which the signal control and route control algorithms
of the three possible vehicle queues {Q24 , Q26 , Q28 }. are based. The real queues Qab (t) containing real vehicles
are maintained on a per movement basis, for every possible
movement (Ra , Rb ) through a given junction.
D. Control Problem
At every junction Ji , there is a controller Ci that has to B. Signal Control Algorithm With Fixed Routing
perform the following tasks at every time slot t. The signal control algorithm for each junction is decentral-
1) Select a phase ∈ Pi (i.e., the traffic controller gives
pik (t) ized.1 At each junction Ji , the algorithm works based on the
the right of way to certain traffic movements in every time per flow virtual queue length information Q̃fa for all links Ra
slot). associated with the given junction. The algorithm works as
2) Make a routing decision for the vehicles (i.e., assign follows for each junction Ji :
queues to the vehicles) related to every flow f passing
through the given junction. 1) For all (a, b) such that (Ra , Rb ) ∈ Mi , determine the flow
with maximum back-pressure and then assign a weight to
The routing decisions are communicated to the corresponding
that flow:
vehicles and the vehicles in the network are assumed to follow  
the routing decision made by the traffic controller. 
fab (t) = arg max Q̃fa (t) − Q̃fb (t) (4)
f ∈Fi
  
f (t) f  (t)
III. P ROPOSED M ETHODS Wab (t) = max Q̃aab (t) − Q̃b ab (t), 0 (5)
In this section, we will describe two novel methods for signal
where Fi is the set of all flows passing through links Ra
control: one with fixed routing and one with dynamic routing.
and Rb .
These methods are based on the back-pressure algorithm [21],
originally invented for scheduling and routing of packets in 1 The algorithm is decentralized in the following sense. A controller at a junc-
wireless networks. In [15], a back-pressure scheduling algo- tion makes decisions based on local communication with vehicles associated
rithm was used for traffic signal control, assuming all vehicles with those links that are connected to the given junction.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

2) For each phase pik ∈ Pi , compute the pressure release as When a certain phase is activated, the real vehicles in the
 network move according to the given rates and the queues
Spik (t) = Wab (t)sab (t). (6) of real vehicles evolve accordingly. The virtual queues evolve
(Ra ,Rb )∈pik according to:

3) The controller Ci at junction Ji activates the phase pik Q̃da (t + 1) = Q̃da (t) − I{d (t)=d} sab (t)
ab
with the highest pressure release, i.e., it selects b:(Ra ,Rb )∈Mi

pik = arg max Spik (t). (7) + I{dca (t)=d} sca (t)
pik ∈Pi
c:(Rc ,Ra )∈Mi

When a certain phase is activated, the real vehicles in the + I{o(f )=a,d(f )=d}λ̃f (t), for a = d (13)
network move according to the given rates and the queues f ∈F
of real vehicles evolve accordingly. The virtual queues evolve
according to: where I{·} denotes the indicator function and λ̃f (t) is the num-
 ber of virtual vehicles associated with flow f that exogenously
Q̃fa (t + 1) = Q̃fa (t) − I{f =f  } sab (t) enter o(f ) at time t. We assume that Q̃dd = 0, i.e., a vehicle is
ab
c:(Rc ,Rb ) not counted in any queue when it enters its destination link.

d
+ I{f =fca } sca (t) 2) Adaptive Route Control Algorithm: Let σab (t) be the
c:(Rc ,Ra ) number of virtual vehicles transferred from link Ra to link
+ I{a=o(f )} λ̃f (t), for a = d(f ) (8) Rb for destination d under the above signal control algorithm
d
during the time slot t, σ̄ab its expected value in stationary
d
where I{·} is an indicator function (whose value is equal to 1 if regime, and σ̂ab (t) the estimate at time t of this expected value.
the statement in its argument is true otherwise its value is equal d
1) At every junction, compute σ̂ab (t) for every feasible
to 0) and λ̃f (t) is the number of virtual vehicles associated with
movement (Ra , Rb ) ∈ Mi associated with that junction
flow f that exogenously enter o(f ) at time t. We assume that
using an exponential averaging method:
Q̃fa (t) = 0 for all a = d(f ), i.e., a vehicle is not counted in any
queue when it enters its destination link. d
σ̂ab (t) = (1 − β)σ̂ab
d
(t − 1) + βσab
d
(t) (14)

C. Signal Control Algorithm With Adaptive Routing where 0 < β < 1 is a smoothing factor.
2) Compute the routing probabilities:
We follow [26], where an algorithm is proposed that de-
couples routing and scheduling in wireless networks. Adaptive d
σ̂ab (t)
routing operates by placing incoming vehicles in real queues
d
Pab (t) =  d (t)
. (15)
c:(Ra ,Rc )∈Mi σ̂ac
according to a probabilistic routing, which signal control is
based on back-pressure on virtual queues per destination. 3) A vehicle entering link Ra joins real queue Qab with prob-
1) Signal Control Algorithm: The signal control algorithm d
ability Pab (t). That is, the vehicle entering Ra destined for
for each junction is again decentralized. At each junction Ji , Rd will be routed to Rb with probability Pab d
(t) at time t
the algorithm works based on the per destination virtual queue through the junction Ji .
length information Q̃da for all links Ra associated with the given
junction. The algorithm works as follows for each junction Ji : The routing information is communicated to vehicles in
terms of probabilities or percentages. For example, consider
1) For all (a, b) such that (Ra , Rb ) ∈ Mi , determine the d
that for the four-way junction illustrated in Fig. 1 if [P24 (t) =
destination with maximum back-pressure and then assign d d
0.1, P26 (t) = 0.2, P28 (t) = 0.7], then among all those vehicles
a weight to that destination: that enter link 2 having destination d, approximately 10 percent
  should join queue Q24 , 20 percent should join queue Q26 , and
dab (t) = arg max Q̃da (t) − Q̃db (t) (9) 70 percent should join queue Q28 . In this setup, the routing
d
   probability is governed by the BP scheme, implicitly. In fact,
d (t) d (t)
Wab (t) = max Q̃aab (t) − Q̃b ab (t), 0 . (10) we first estimate the mean of the virtual vehicles transferred
from link a to b heading destination d, i.e. the local flows
2) For each phase pik ∈ Pi , compute the pressure release as to destination d. The estimation of the mean uses a recursive
 method, taking into account the latest known vehicle number at
Spik (t) = Wab (t)sab (t). (11) every instant t and the estimated mean from the previous sample
(Ra ,Rb )∈pik time. By means of the average valued local flows (towards a
destination d) in stationary regions, we split vehicles according
3) The controller Ci at junction Ji activates the phase with to the probability calculated. Moreover, we route vehicles in
the highest pressure release, i.e., it selects a local, decentralized context (V2I communication is required
though). Finally, BP only enables traffic phase activation, while
pik = arg max Spik (t). (12)
i pk ∈Pi the proposed routing solution distributes the virtual flows.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZAIDI et al.: TRAFFIC SIGNAL CONTROL WITH FIXED AND ADAPTIVE ROUTING FOR VEHICULAR NETWORKS 5

3) Enhancing the Performance of Adaptive Routing capacity region of a network under fixed flow routes cannot be
Algorithm: It will be shown in Section IV that the proposed larger than the capacity of the same network where routes of
back-pressure routing algorithm is suitable for heavily loaded the flows are not fixed, since capacity region cannot decrease
networks but can lead to unnecessarily long routes in a low by removing path constraints on flows.
load situation. This is also the case in wireless networks, where The traffic control and routing algorithms proposed above
several methods have been proposed to improve the delay per- are based on the algorithms for scheduling (rate allocation) and
formance of back-pressure routing [26], [27]. These methods routing of packets in a communication network. These algo-
are usually based on including bias terms in the calculation of rithms are throughput-optimal according to [25, Theorem 4]
queue backlogs. That is, if one wants to encourage (discourage) and [26, Theorem 1], where the authors show that if a network
traffic flow to a certain link, then one can add bias terms in the is stable under the back-pressure scheme based on real queues
calculation of queue backlog differences. For instance, one can (which is already known to be optimal) with flow arrival
modify (9) and (10) as follows: rate λf , then it is also stable (i.e., real queues are stable)
   under the back-pressure scheme based on virtual queues with
dab (t) = arg max Q̃da (t) − Q̃db (t) + α Vad − Vbd (16) flow arrival rate λf (1 + ) for any  > 0. In the proposed
d
  traffic control algorithm, the phase activation (activation of a
d (t) d (t)
Wab (t) = max Q̃aab (t) − Q̃b ab (t) set of movements) procedure is equivalent to scheduling over
   a communication network. In contrast to a wireless communi-
d (t) d (t)
+ α Va ab − Vb ab ,0 (17) cation network, this scheduling procedure is decentralized at
every junction because the activation of links associated with
where Vad is equal to the minimum number of links that one junction does not affect the activation of links at any
exist between link a and link d (destination d) and α is a other junction if one assumes links to be infinitely long. The
non-negative real number that can be optimized. A higher possible traffic movements associated with a given junction can
value of α forces the vehicles to follow shorter paths, which be interpreted as connections in a wireless network. Similar
is good for low-load situations but may not be good in a high to a wireless network where packets cannot be simultaneously
load situation, as shown in Section IV. transmitted with high rates over neighboring links due to cross-
link interference, here in the road network, some movements
at a junction cannot be activated simultaneously. With this
D. Stability and Optimality interpretation and equivalence in mind, we establish optimal-
1) Infinite Length Links: In this section we discuss opti- ity of the proposed traffic control and routing algorithms.
mality of the proposed methods in the sense of supporting The algorithms are optimal in the sense that they can stably
maximum traffic arrivals in a road network under the assump- support any flow arrival rate which is in the interior of the
tion that all links are infinitely long. Although in practice capacity region.
all links in a network have finite lengths, the BP scheme is 2) Finite Length Links: When a network has links with finite
originally inspired by its proven throughput optimality under lengths, the issue of stability (according to Def. 1) does not arise
the assumption of links with infinite lengths. The earlier papers because the queues can never be unstable due to finite length
[15], [18] that study single-commodity traffic signal control links. In this situation, stability corresponds to maintaining
under real queues guarantee optimality under the assumption bounded queue backlogs in the links where traffic is being input
of infinite length links. In order to complement the existing to the network (ingress buffers in the context of communication
literature, we also provide this discussion. To be precise, we networks [28]), assuming that origin links can be infinitely
first define stochastic stability and capacity region of a network long. It is not known if the back-pressure based schemes are
and then discuss the optimality. throughput-optimal in this context. In the following section,
Definition 1: A queue Q(t) with stochastic arrival and depar- we analyze performance of the proposed back-pressure based
ture processes is said to be strongly stable if algorithms over a network having links of finite lengths.

1
t−1
lim sup E [Q(τ )] < ∞. (18) IV. P ERFORMANCE A NALYSIS
t→∞ t τ =0
We analyze performance of the proposed algorithms in
Furthermore, if all queues in a network are stable, then the terms of queue lengths, trips completed, and travel time using
network is said to be stable. PTV VISSIM [29], which is a microscopic traffic simulator.
Definition 2: The capacity region Λ associated with the Within VISSIM, every vehicle is simulated individually and
network is the closure of the set of all flow arrival rates that can several useful properties related to every vehicle can be ac-
be stably supported by the network. That is, for a network to cessed dynamically. We will consider and compare four distinct
be stable, we must have {λf }f ∈F ∈ Λ. Furthermore, a scheme methods:
that can stabilize all flows that have arrival rates in the capacity
region is known as a throughput-optimal scheme. • Fixed time (FT) schedule signal controller: The possible
If the route of every flow f is fixed, then the capacity region phases at each intersection are activated in a predeter-
is the set of all flow arrivals that are supportable given the mined periodic fashion. All vehicles are assumed to fol-
set of flows and their corresponding routes. We note that the low shortest routes to their respective destinations.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

Fig. 4. Road network with 24 intersections, 84 links, 16 origins and 16 destinations.

• Single-commodity back-pressure (SC-BP) controller: As TABLE I

P HASE D ISTRIBUTIONS
proposed in [15], each junction i maintain queues Qa (t)
for all connected links. For each pair (Ra , Rb ) ∈ Mi , the
back pressure Wab (t) = Qa (t) − Qb (t) is computed. For
each phasepik ∈ Pi , the pressure release is computed as
Spik (t) = (Ra ,Rb )∈pi Wab (t)sab (t). Finally, the phase
k
giving rise to the maximum pressure release is selected. 4.76 meters × 1.5 meters. The maximum speed of all vehicles is
This approach is similar to Section III-B, but does not dis- set to 70 km/h, as some of the links on the boundary of the net-
tinguish between different flows. All vehicles are assumed work in Fig. 4 are highways. We assume that a car is in a queue
to follow shortest routes to their respective destinations. if its speed is below a certain threshold (here set to 5 km/h).
• Multi-commodity back-pressure control (MC-BP) with For the fixed time schedule control (FT), we assume the time
fixed routing: The method described in Section III-B. period of each cycle equal to 60 seconds at all intersections
Moreover, all vehicles are assumed to follow shortest (both three-way and four-way) according to the signal plan
routes to their respective destinations. (phase distribution) given in Table I.3 For the back-pressure
• Adaptive routing back-pressure control (AR-BP): The methods, we consider that a phase is activated after every
method described in Section III-C. 15 seconds. Moreover, all the traffic related measurements
are also taken after every 15 simulation seconds in order to
For the sake of simplicity, we assume that sab (t) = sab for update the signal phases and the routing decisions under the
all t, i.e., flow rate through a junction does not depend on time back-pressure methods. All simulations are performed for 7200
or any other state2 in the road network. However, the traffic simulation seconds (i.e., 2 hours). Within VISSIM, we have set
signal control schemes presented above also applicable to the the simulation speed to 10 simulation seconds per second and
situations where traffic movement rates are time varying. the simulation resolution is set equal to 1 in order to generate
fastest simulation results. Further details on how the simula-
A. Network and Simulation Parameters tions were performed using VISSIM are given in Appendix A.

The simulations are performed using a road network from

a central region in the Stockholm area, comprising 24 signal- B. Simulation Results and Discussions
ized intersections (16 three-way intersections and 8 four-way In Fig. 5, we fix vehicle arrival rate to 350 vehicles/hour at
intersections) and 84 links. The network is depicted in Fig. 4. all traffic origins and plot the evolution of queue length over
The lengths of the longest and the shortest links are approxi- time under different signal control methods. Here, queue length
mately 1980 meters and 333 meters. All links are assumed to refers to the total number of vehicles that are queued in the net-
have three lanes, where each lane is 3.5 meters wide. There work. An arrival rate of 350 vehicles/hour per origin means that
are 16 traffic origins {O1 , O2 , . . . , O16 } and 16 destinations approximately 5600 vehicles enter the network per hour since
{D1 , D2 , . . . , D16 } in the network. The traffic associated with there are 16 origins. As we observe in Fig. 5, AR-BP yields
an origin-destination pair Oi − Di forms a flow fi . Hence, there smallest queue length followed by MC-BP, SC-BP, and FT re-
are 16 traffic flows in total, {f1 , f2 , . . . , f16 }. We perform sim- spectively. Queue length under FT is increasing approximately
ulations with cars of dimensions 4.11 meters × 1.5 meters and
3 The time-loss due to amber or yellow signals is not considered in the
2 For example, the state of the network may change in case of an accident. simulations, however, it can be incorporated easily in VISSIM.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZAIDI et al.: TRAFFIC SIGNAL CONTROL WITH FIXED AND ADAPTIVE ROUTING FOR VEHICULAR NETWORKS 7

Fig. 5. Evolution of total queue length in the network over time under different
control methods with vehicle input rate equal to 350 vehicles/hour. Fig. 7. Average travel times computed over a 2 hours simulation period as a
function of the vehicle input arrival rate under different control methods.

Fig. 6. Average queue lengths over a 2 hours long simulation time period as a
function of the vehicle input arrival rate under different control methods.
Fig. 8. Average vehicle speeds computed over a 2 hours simulation time period
as a function of the vehicle input arrival rate under different control methods.
linearly over time, whereas for the BP based methods the queue
lengths remain bounded. It was already shown in [15] that SC- AR-BP a next hop route is chosen at every intersection. When
BP outperforms SCATS (where a signal plan is optimized over the network is under-saturated, the pressure terms (queue back-
a set of fixed signal schedules) in terms of queue lengths. An ad- log differences) are very low and a vehicle may traverse several
ditional observation is that the multi-commodity back-pressure links before arriving to its destination, thus taking a route that
schemes (MC-BP, AR-BP) have the potential to provide con- is unnecessarily long. However, it is this adaptive routing that
siderable improvement over single-commodity scheme. forces the vehicles to distribute in the network more uniformly
In order to investigate further, we plot the average queue and thus reduces congestion queue lengths when the network is
lengths and average travel times of vehicles from their origins heavily loaded. In a saturated network, although vehicles may
to their respective destinations as functions of vehicle arrival follow a longer route on average under AR-BP, the travel time
rates in Figs. 6 and 7 respectively. Here the averages are taken is significantly lower on average compared to the fixed routing
over simulation time in the case of queue length and over both methods as shown in Fig. 7. This reduction in average travel
simulation time and number of vehicles in the case of travel time happens due to a smaller queue lengths in the network as
time. According to Figs. 6 and 7, MC-BP is significantly supe- observed in Fig. 6.
rior to SC-BP in terms of both average queue length and average In Section III-C3 we presented a modified version of AR-BP
travel time at all vehicle arrival rates. The behavior of AR-BP is scheme that can force vehicles to avoid unnecessarily long
not straightforward—it provides relatively much smaller queue routes in a load situation. This modified AR-BP method can
lengths but the travel times are very high at low traffic volumes. be optimized for a given network. In Figs. 9 and 10, we plot
Normally, one expects that a larger queue length should lead to average travel time and average queue lengths under the modi-
a higher travel time. In order to study what makes the average fied AR-BP scheme with different values of α. (Note that α = 0
travel times so high under AR-BP, we must consider the aver- gives the original AR-BP scheme.) A higher value of α forces
age speed of vehicles under all schemes. Fig. 8 shows that the vehicles to follow shorter path, which is good for low-load
average vehicle speeds are always highest under AR-BP. This situations but may not be good in a high load situation. Accord-
implies that the vehicles travel longer distances on average ing to Figs. 9 and 10, there exists a value of α (equal to 1.5)
to reach their destinations under AR-BP, especially when the for the given network that provides good performance in both
vehicle arrival rates are low. Under MC-BP, a path for every ve- low load and high load scenarios in terms of travel time as well
hicle from its origin to its destination is pre-defined, whereas in as congestion.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

Fig. 12. Total number of vehicles that could not enter the network until the end
of 2 hour simulation period under different control methods.
Fig. 9. Average travel time computed over a 2 hours simulation time period as
a function of the vehicle input arrival rate under the modified AR-BP method.

Fig. 13. Total waiting time of all vehicles that could not enter the network
immediately during 2 hour simulation period under different control methods.

conserving4 and may cause congestion propagation to other

Fig. 10. Average queue length computed over a 2 hours simulation time period
as a function of the vehicle input arrival rate under the modified AR-BP method. links in a network. The deadlocks occur at very high traffic
loads depending on the network topology and especially when
there is a significant mismatch between lengths (or capacities)
of adjacent links. Note that our simulated network is quite
asymmetric in terms of lengths of different links and there-
fore it is also more susceptible to deadlocks. One way of
resolving deadlocks under SC-BP is to use normalized pres-
sure functions [18]. Interestingly, the proposed back-pressure
schemes MC-BP and AR-BP are robust against deadlocks
because their control decisions are based on virtual queues that
keep growing irrespective of the lengths (capacities) of their
corresponding links.
Finally, we analyze performance of the proposed methods
under the following two measures that are relevant in high load
Fig. 11. Total number of vehicles that exit the network, i.e., reach their situations: i) latent demand and ii) latent delay. Latent demand
destinations during a 2 hour simulation period under different control methods. refers to the total number of vehicles that are waiting till the
end of simulation to enter the network. Latent delay refers to the
Next we investigate the network throughput in terms of the total waiting time of all vehicles that are not able to immediately
number of vehicles exiting the network (number of completed enter the network. This also includes waiting time (outside the
trips) under different signal control methods. In Fig. 11, we network) of the vehicles which were later able to enter the net-
plot the total number vehicles that exit the given network in work before the end of the simulation time. In Figs. 12 and 13,
two hours when the traffic is continuously arriving at a fixed we have plotted latent demand and latent delay as functions of
rate. Interestingly, FT provides higher throughput than SC-BP vehicle arrival rates, respectively. These simulation results also
at very high input traffic load, despite the fact that FT always indicate the benefits of using virtual queues in back-pressure
gives rise to a higher time averaged queue length than SC-BP methods and adaptive routing in general for saturated networks.
according to Fig. 6. This happens due to the fact that when
4 A control is work-conserving if the two conditions, i) Q (t) > 0 and
back-pressure schemes are employed over a network with finite ab
ii) Qb (t) < Cb , are sufficient to ensure that the server of the junction works
length links, some links can experience deadlock situation, as during slot t, where Cb is capacity of link b. Non-work conservation is a sign
observed in [18]. Deadlocks make the controllers non-work of inefficiency [18].
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

ZAIDI et al.: TRAFFIC SIGNAL CONTROL WITH FIXED AND ADAPTIVE ROUTING FOR VEHICULAR NETWORKS 9

time, and the number of trips completed. In addition, multi-

commodity methods allow for relatively fair distribution of
vehicle queues associated with different origin-destination pairs
(flows) within the network. Due to the use of virtual queues,
MC-BP and AR-BP are more robust to deadlock situations than
SC-BP.
In SC-BP and MC-BP, all vehicles follow fixed routes. Fixed
routing is not appropriate when a road network is heavily loaded
with vehicles, since links will get more congested. AR-BP is
able to distribute vehicles more uniformly across the network,
thereby significantly improving congestion, throughput, and
travel times (on average). For low load situations, AR-BP may
lead to unnecessarily long route selections for some vehicles,
Fig. 14. Normalized queue lengths at vehicle arrival rate equal to 350 vehicles/
hour. giving very high travel times on average. In such situations, a
simple fixed-time control can have better performance. Alterna-
From the above discussion, we conclude that the multi- tively, a modified version of AR-BP can be used to reduce travel
commodity back-pressure control methods are significantly times by restricting route selection from a set of fewer paths.
superior to the single-commodity back-pressure traffic method In practice, the proposed routing method can be complemented
in terms of travel time, queue length, and trips completed. with additional intelligence to avoid unnecessarily long routes
Another advantage of employing multi-commodity schemes is in the case of very low traffic demand.
that it allows for relatively fair distribution of vehicle queues The implementations of MC-BP and AR-BP require commu-
associated with different origin-destination pairs (flows) within nication from every vehicle to the traffic controller located at
the network. As an example, in Fig. 14 we have shown the nor- the upcoming intersection and/or between adjacent controllers,
malized average queue lengths of all flows (origin-destination depending on how these schemes are realized in practice. For
pairs) in the network under SC-BP and MC-BP, with a traffic AR-BP, the controller also needs to broadcast routing informa-
arrival rate equal to 350 vehicles/hour. We can see that the tion comprised of routing probabilities (or turning percentages)
queue length distribution among different traffic flows is more to the vehicles. An interesting direction in future would be to
fair when the multi-commodity scheme is employed. investigate MC-BP and AR-BP schemes subject to uncertain
Remark on Communication Requirements: The performance and delayed queue information. It will be useful to devise
gains discussed under the proposed back-pressure based meth- suitable protocols for vehicle-to-infrastructure communication.
ods (fixed and adaptive routing) are achieved assuming perfect The proposed methods have been shown to be optimal in the
communication between vehicles and controllers. In MC-BP sense of supporting maximum traffic arrival rate while main-
and AR-BP, every vehicle has to communicate with upcoming taining stable queues under the assumption of infinite length
controller and/or the adjacent controllers, depending on how links. The stability regions of a general network under the BP
these schemes are realized in practice. Moreover, under AR-BP schemes are only known under the assumption of infinite length
the controllers have to communicate the routing information links. Thus, another useful direction would to characterize the
to the vehicles. The routing probabilities calculated according stability regions of a road network with finite-length links under
to (15) can be broadcast to all vehicles in the form of a different BP schemes. Here, stability of a network would mean
look-up table and the vehicles would then adapt their routes that all queues at the traffic origins are stable. All back-pressure
depending on the received routing probabilities. We believe based signal control algorithms proposed so far assume a fixed
that the results presented in this paper provide motivation for (pre-defined) signal phase duration. It will be interesting to
analyzing back-pressure schemes under imperfect vehicle-to- study how much we can gain by keeping both phase duration
infrastructure communication and exploring relevant commu- and phase activation as functions of queue length information.
nication protocols.
A PPENDIX A
V. C ONCLUSION AND F UTURE D IRECTIONS S IMULATION D ETAILS
We studied the problem of decentralized traffic signal control The simulations are performed by allowing data exchange
and adaptive routing of vehicles under different back-pressure between MATLAB and VISSIM using the following proce-
control schemes, namely, single-commodity back-pressure dure: Create the network file (.inp file) using VISSIM GUI,
(SC-BP), multi-commodity back-pressure (MC-BP), and adap- based on a network image taken from Google Earth. Drop
tive routing back-pressure (AR-BP). The proposed back- all relevant objects (vehicle inputs, controllers, routing deci-
pressure methods address network level traffic control by means sions, data collection points) in the network with desired initial
of the interacting queue dynamics on adjacent roads. Note, parameters/settings. Save these settings in a file (.ini file).
however, that the algorithms are applied locally, following Create a MATLAB script, where first activate the VISSSIM
decentralized control policies, relying on knowledge of adja- COM server, and then load the network file and the settings file.
cent queues. We observed that MC-BP always outperforms Within the MATLAB script, one can access all relevant objects
SC-BP in terms of average queue lengths, average vehicle travel and change the control signals on state of the traffic.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS

R EFERENCES [23] L. Xiao and H. K. Lo, “Adaptive vehicle navigation with en route stochas-
tic traffic information,” IEEE Trans. Intell. Transp. Syst., vol. 15, no. 5,
[1] H. Frumkin, “Urban sprawl and public health,” Public Health Rep., pp. 1900–1912, Oct. 2014.
vol. 117, no. 3, pp. 201–217, 2002. [24] H. F. Wedde and S. Senge, “BeeJamA: A distributed, self-adaptive vehicle
[2] H. Kamal, M. Picone, and M. Amoretti, “A survey and taxonomy of urban routing guidance approach,” IEEE Trans. Intell. Transp. Syst., vol. 14,
traffic management: Towards vehicular networks,” CoRR, abs/1409.4388, no. 4, pp. 1882–1895, Dec. 2013.
2014. [25] L. X. Bui, R. Srikant, and A. Stolyar, “A novel architecture for reduction of
[3] C. Chen, F. Zhu, and Y. Ai, “A survey of urban traffic signal control
delay and queueing structure complexity in the back-pressure algorithm,”
for agent recommendation system,” in Proc. IEEE ITSC, Sep. 2012, IEEE/ACM Trans. Netw., vol. 19, no. 6, pp. 1597–1609, Dec. 2011.
pp. 327–333. [26] E. Athanasopoulou, L. X. Bui, T. Ji, R. Srikant, and A. Stolyar, “Back-
[4] N. H. Gartner, J. D. C. Little, and H. Gabbay, “Optimization of traffic pressure-based packet-by-packet adaptive routing in communication net-
signal settings by mixed-integer linear programming,” Transp. Sci., vol. 9,
works,” IEEE/ACM Trans. Netw., vol. 21, no. 1, pp. 244–257, Feb. 2013.
no. 4, pp. 344–363, 1975. [27] M. J. Neely, “Dynamic power allocation and routing for satellite and
[5] P. B. Hunt, D. I. Robertson, R. D. Bretherton, and R. I. Winton, wireless networks with time varying channels,” Ph.D. dissertation, Lab.
“SCOOT—A traffic responsive method of coordinating signals,” TRL
Inf. Decision Syst., Massachusetts Inst. Technol., Cambridge, MA, USA,
Lab., Wokingham, U.K., Tech. Rep. 1014, 1981. 2003.
[6] V. Mauro and C. D. Taranto, “UTOPIA,” in Control, Computers, [28] L. B. Le, E. Modiano, and N. B. Shroff, “Optimal control of wireless
Communications in Transportation Research, J. P. Perrin, Ed. Oxford, networks with finite buffers,” IEEE/ACM Trans. Netw., vol. 20, no. 4,
U.K.: Pergamon, 1990, pp. 245–252.
pp. 1316–1329, Aug. 2012.
[7] P. R. Lowrie, “The Sydney coordinated adaptive traffic system— [29] M. Fellendorf, “VISSIM: A microscopic simulation tool to evaluate ac-
Principles, methodology, algorithms,” in Proc. Int. Conf. Road Traffic tuated signal control including bus priority,” presented at the 64th ITE
Signaling, 1982, pp. 67–70.
Annual Meeting, Dallas, TX, USA, 1994.
[8] C. Diakaki, M. Papageorgiou, and K. Aboudolas, “A multivariable regu-
lator approach to traffic-responsive network-wide signal control,” Control
Eng. Pract., vol. 10, no. 2, pp. 183–195, Feb. 2002.
[9] L. B. de Oliveira and E. Camponogara, “Multi-agent model predictive Ali A. Zaidi received the M.Sc. degree in wire-
control of signaling split in urban traffic networks,” Transp. Res. C, less systems and the Ph.D. degree in telecommu-
Emerging Technol., vol. 18, no. 1, pp. 120–139, Feb. 2010. nications from KTH Royal Institute of Technology,
Stockholm, Sweden, in 2008 and 2013, respectively.
[10] H. Ezawa and N. Mukai, “Adaptive traffic signal control based on vehi-
cle route sharing by wireless communication,” in Knowledge-Based and In 2014, he joined Ericsson Research, Stockholm,
Intelligent Information and Engineering Systems, ser. Lecture Notes in where he is currently working with research and
standardization of future radio access technologies.
Computer Science, R. Setchi, I. Jordanov, R. Howlett, and L. Jain, Eds.,
vol. 6279. Berlin, Germany: Springer-Verlag, 2010, pp. 280–289. He was a Postdoctoral Researcher with the Depart-
[11] T. Tettamanti, T. Luspay, B. Kulcsár, T. Peni, and I. Varga, “Robust control ment of Signals and Systems, Chalmers University
for urban road traffic networks,” IEEE Trans. Intell. Transp. Syst., vol. 15, of Technology, Sweden, and a Visiting Researcher
with the Department of Mathematics and Statistics,
no. 1, pp. 385–398, Feb. 2014.
[12] I. Alvarez, A. Poznyak, and A. Malo, “Urban traffic control problem a Queen’s University, Canada. He has coauthored more than 30 research articles
game theory approach,” in Proc. IEEE Conf. Decision Control, Dec. 2008, on various topics related to communication and control in wireless networked
pp. 2168–2172. systems.
[13] S. Lämmer and D. Helbing, “Self-control of traffic lights and vehicle flows
in urban road networks,” J. Statist. Phys., vol. 2008, no. 4, p. P04019,
2008. Balázs Kulcsár received the M.Sc. degree in traffic
[14] T. Le et al., “Decentralized signal control for urban road networks,” engineering and the Ph.D. degree in automatic
Transp. Res. C, Emerging Technol., vol. 58, pp. 431–450, Sep. 2015. control from Budapest University of Technology
[15] T. Wongpiromsarn, T. Uthaicharoenpong, Y. Wang, E. Frazzoli, and and Economics (BUTE), Budapest, Hungary, in
D. Wang, “Distributed traffic signal control for maximum network 1999 and 2006, respectively. He was a Researcher/
throughput,” in Proc. IEEE ITSC, Sep. 2012, pp. 588–595. Postdoctoral Researcher with the Department of
[16] P. Varaiya, “The max-pressure controller for arbitrary networks of sig- Control for Transportation and Vehicle Systems,
nalized intersections,” in Advances in Dynamic Network Modeling in BUTE; the Department of Aerospace Engineer-
Complex Transportation Systems, ser. Complex Networks and Dynamic ing and Mechanics, University of Minnesota,
Systems, S. V. Ukkusuri and K. Ozbay, Eds., vol. 2. New York, NY, Minneapolis, MN, USA; and Delft Center for Sys-
USA: Springer-Verlag, 2013, pp. 27–66. tems and Control, Delft University of Technology,
[17] J. Gregoire, E. Frazzoli, A. de La Fortelle, and T. Wongpiromsarn, “Back- Delft, The Netherlands. He is currently an Associate Professor with the Depart-
pressure traffic signal control with unknown routing rates,” presented at ment of Signals and Systems, Chalmers University of Technology, Gothenburg,
the IFAC World Congr., Cape Town, South Africa, 2014. Sweden. His main research interest focuses on traffic flow modeling and related
[18] J. Gregoire, X. Qian, E. Frazzoli, A. de La Fortelle, and T. Wongpiromsarn, controller design.
“Capacity-aware back-pressure traffic signal control,” IEEE Trans.
Control Netw. Syst., vol. 2, no. 2, pp. 164–173, Jun. 2015.
[19] A. Kouvelas, J. Lioris, S. A. Fayazi, and P. Varaiya, “Maximum pressure Henk Wymeersch (S’99–M’05) received the Ph.D.
controller for stabilizing queues in signalized arterial networks,” Transp. degree in electrical engineering/applied sciences from
Res. Rec., vol. 2421, pp. 133–141, 2014. Ghent University, Ghent, Belgium, in 2005. He was a
[20] B. Kulcsár, K. Ampountolas, and A. Dabiri, “Single region robust Postdoctoral Associate with the Laboratory for Infor-
perimeter traffic control,” in Proc. IEEE Eur. Control Conf., 2015, mation and Decision Systems, Massachusetts Institute
pp. 2628–2633. of Technology. He is currently an Associate Profes-
[21] L. Tassiulas and A. Ephremides, “Stability properties of constrained sor with the Department of Signals and Systems,
queueing systems and scheduling policies for maximum throughput in Chalmers University of Technology, Gothenburg,
multihop radio networks,” IEEE Trans. Autom. Control, vol. 37, no. 12, Sweden. He was an Associate Editor of IEEE C OM -
pp. 1936–1948, Dec. 1992. MUNICATION L ETTERS (2009–2013) and is currently
[22] S. Kim, M. E. Lewis, and C. C. White, “Optimal vehicle routing with real- an Associate Editor of IEEE T RANSACTIONS ON
time traffic information,” IEEE Trans. Intell. Transp. Syst., vol. 6, no. 2, W IRELESS C OMMUNICATIONS (2013–present) and Transactions on Emerging
pp. 178–188, Jun. 2005. Telecommunications Technologies (2011–present).