2001 IEEE Intelligent Transportation Systems Conference Proceedings - Oakland (CA), USA - August 25-29, 2001

A SIMULATION-BASED SIGNAL OPTIMIZATION ALGORITHM WITHIN A DYNAMIC

TRAFFIC ASSIGNMENT FRAMEWORK
By: Akmal S. Abdelfatah" and Hani S. Mahmassani'*

integrated assignment program. They also

ABSTRACT developed a solution algorithm and established its
This paper considers the issues that arise in seeking convergence to the desired solution.
to optimize network performance through Allsop (9) appears to be the first to suggest, in
simultaneous determination of optimal routes and the published scholarly literature, that traffic
optimal signal settings. It presents numerical engineers should take explicit account of the long
experiments for a simulation-based algorithm on a
run effect of signal settings on the traffic
realistic moderately large network (a portion of the
assignment. While his approach has been shown
Fort- Worth area) and a hypothetical network to
underscore the potential of additional improvement
to be non-convergent, it laid the groundwork for
through joint consideration of signal control and several subsequent contributions. Charlesworth
route assignment compared to either one of these (10) obtained mutually consistent traffic
decisions alone. assignment and signal settings through an
I. BACKGROUND REVIEW iterative procedure in which the well known
TRANSYT software is used to optimize the
Most prior research that has recognized the signal settings. Dickson(l1) gave an example
interaction between signal timing and traffic where these procedures may lead to an increase in
assignment has addressed static demand the total network travel time. He attributed this
conditions, i.e. constant 0 - D trip rates and increase in the travel time to the fact that these
constant average flow rates through the network. approaches assumed no simultaneous changes in
Furthermore, previous work has focused equilibrium flows due to signal setting
primarily on setting signals in a manner adjustments. Sheffi and Powell (12) presented a
consistent with users' choice of route, i.e. with a mathematical programming formulation for the
descriptive assignment rule, typically resulting in problem and an algorithm that could be used for
a Wardrop User Equilibrium (UE), rather than very small networks, though the problem
with seeking to optimize performance through formillation may be non-convex. Another
System Optimal (SO) route assignment jointly iterative procedure was presented by Gartner and
with signal control. This review addresses Al-Malik (1 3) to provide a simultaneous solution
selected contributions under static assignment of the assignment and the control problem.
conditions, followed by a brief overview of the B. Time-Dependent Assign men t
limited body of work dealing with time-dependent
conditions. As noted previously, research addressing signal
A . Signal Tinzing under Static Assignment timing in conjunction with time-dependent
network flow assignment is very limited
Some of the most important theoretical
compared with that reported for the static
contributions to the problem of signal control
assignment problem. Smith and Ghali (14,15)
and UE traffic assignment are due to Smith
presented a mathematical formulation that
(2,3,4,5) who derived conditions that guarantee
extends their work on the static case, and
the existence of an equilibrium as well as
suggested an algorithm that uses an incremental
conditions for the uniqueness and stability of the
assignment logic. Ghali and Smith (16)
traffic equilibrium when there is interaction
implemented an iterative procedure using
between signal setting and users' route choice
CONTRAM and examined its convergence
decisions. He also showed the properties of a
pattern.
local control policy, named Po, which ensure the
existence of a traffic equilibrium consistent with Smith and Van Vuren (17) outlined the steps for
the control policy. Smith et al. (6,7) defined a an iterative algorithm which could be considered
principle which is equivalent to Wardrop's (8) as an extension of Allsop's algorithm to the
first principle such that the allocation of flows time-varying case. Gartner and Stamatiadis (18)
and green times can be combined into one recently presented a general conceptual framework

*Assistant Professor of Civil Engineering, The American University of Sharjah, UAE

** Professor of Civil Engineering and Information Systems, The University of Teaxas-Austin, USA
0-7803-7194-1/01/$10.00 02001 IEEE 428

Authorized licensed use limited to: Xi'an Jiaotong-Liverpool University. Downloaded on October 24,2022 at 11:47:32 UTC from IEEE Xplore. Restrictions apply.
 for the implementation of a combined solution which is the number of upstream or downstream
for dynamic traffic assignment and signal control, intersections to be considered to delineate the
but did not report implementation of a specific local area. The shown Figure shows an example
algorithmic procedure. of the local area for K= 1 .
In a previous paper (I), we considered a central
controller seeking to optimize system
performance by determining time-varying route
guidance and signal control policies. The paper
presented the mathematical formulation and
solution algorithm for the combined problem of
dynamic assignment (SO assignment) and
responsive signal control. The approach is
applicable to a general traffic network which has
different control types (i.e. some intersections
might have stop or yield sign, while some others
may have signal control with any phasing
pattem). Finally, the solution algorithm was The steps for the signal optimization are the
implemented and applied an actual network model following :
using Webster’s (19) formulas to optimize the
signal settings. This paper presents numerical
Step 0: Initialization; Go= set of green times in
results using a more powerful simulation method
For signal optimization within the same S, (S, and R, are the set of signal settings and
algorithmic framework presented previously, and routing for iteration i in the main algorithm).
shows the robustness of the solution algorithm Define the local area for each signalized
under different traffic congestion levels and intersection . Set m=I.
conditions. For iteration m
For n EN, (set of signalized
11. PROBLEM FORMULATION AND
SOLUTION ALGORITHM intersections), f o r each time interval t :
do;
The reader is referred to reference (1) for a detailed Steu I : Calculate the total delay at intersection n
discussion of the problem formulation and the t m-1
during time interval t ( Dn (Gn ), R, )), by
main solution algorithm for the problem . The
objective function is to minimize the total travel resimulating the local area for intersection n

n -
time in the network considering the path flows using ~m (set of green times at intersection
and the signal setting parameters as the decision
variables. The main objective of this paper is to n for iteration m- I), R,
present the simualtion-based signal optimization Ster, 2 : Find the new green times ( gt”)
algorithm. nl
(Hooke and Jeeves’ method)
A . Signal Optimization
step 2.1: Initialize Z = G =~ G ~ - ’ ,
This section provides a detailed description of a n n
simulation-based signal optimization procedure 6 >O, O<cx<l .O and B=l
using an implementation of Hooke and Jeeves’ Step 2.2 : Exploratory Search, for each phase j
method (21). The simulation method for signal do,
optimization uses a modified version of the a. k z
= +B&, ,simulate, and check if
DYNASMART simulation model to optimize
the signal settings. In this method, we define a
“local area” for each intersection which comprises
other neighboring intersections that affect or ace
z z I
If yes, update = and check next phase.
If no, go to step 2.2b
affected by the delay at the intersection of b. if B=l, set B=-1 and go to 2.2.a,
interest. Similar to the local area definition for otherwise, check next phase.
decentralized route guidance proposed by Hawas Step 2.3 : Pattem Search
and Mahmassani (22), we define the “depth” K,

429

Authorized licensed use limited to: Xi'an Jiaotong-Liverpool University. Downloaded on October 24,2022 at 11:47:32 UTC from IEEE Xplore. Restrictions apply.
 time proportions were considered for each
signalized intersection in the network and the
put G F = Z and cycle length is considered to be the only decision
variable for each intersection during each signal
setting interval.
4. Time Intervals
go to step 2.2, otherwise, go to step 2.3.b. The total loading time in these experiments is 35
b. if 6 is “sufficiently” small, stop with minutes and the network is simulated for as long
Gm “optimal”. Otherwise, put as is required for all vehicles to reach their
n
n
’
k0.56, j = l , Z = Gm - and go to step 2.2.
destination. While the path assignments are
obtained for the loading time only, the signal
settings are obtained for the whole simulated

’
SteD 3 : Check for convergence :
interval. The loading time is the same as the
If gk;” G gk;” - for all I, if all signalized assignment period, and is subdivided into 5-
minutes assignment intervals. For the signal
intersections are optimized, stop an return to the setting interval, we considered two values, 5
main algorithm, otherwise start for a new minutes, and 10 minutes.
signalized intersection.
B. Experimental Results
Else, G
=: current set of green times, set
Two sets of experiments were performed on each
m=m+l and go to step 1 network. The details and discussions of these
111. NUMERICAL ILLUSTRATION sets of experiments are provided in the following
two subsections respectively.
?be results reported in reference (1) were obtained 1. Results for Network 1
by implementing Webster’s formulas for signal
For the first set of experiments, the green time
optimization. In this section, the solution
for each phase was considered to be a decision
zlgorithm is implemented, using the simulation
variable and the time interval for signal setting
method for signal optimization, on both an
changes to be equal to the assignment interval of
actual network and a hypothetical network under
the SO module (5 minutes). To reduce the
different congestion levels. The main objective
computational time, some assumptions were
OF th-se experiments is to illustrate the potential
made for this set of experiments. First
for improving network performance when
assumption, only a depth k=O was considered
solving for the combined SO assignment and
(i.e. assuming isolated intersections). The
signal control problem beyond that obtained
second and most restrictive assumption is the use
through SO assignment onIy.
of one iteration with a fixed step size, if a
A . Experimental Design reduction in the delay at the intersection is
1. Test Networks achieved, then the signal optimization algorithm
Figure 2 depicts a part of Fort-Worth street stops, otherwise, it implements the same step
network (Network 1) and it consists of 178 nodes size in the other direction. The step size is set to
(including 6 1 signalized intersections) and 441 be 0.20. The improvements achieved in this set
directed links. Figure 3 illustrates the of experiments may be considered as the lower
hypothetical network consists of 22 nodes bound for the possible improvement in the
(including 14 signalized intersections) and 68 network performance.
directed arcs.
Table 1 shows the average travel time, over all
2. Congestion Levels
vehicles in the system, expressed as a percentage
The experiments are conducted under three
of the System Optima1 traffic assignment only
congestion levels (6000, 12000 and 17000
solution at convergence, for the unconstrained
vehicles) for both networks.
cycle length case for set number 1 for network
3. Cycle Length
Table 1 shows the lower bound for the
Two different cases are considered for this
improvement which is promising. The
important factor. Under the first case, the cycle
percentage of improvement is ranging from about
length is unconstrained, the green times are
5% to about 7 % and appears to be decreasing
subject to upper and lower bound constraints.
with the increase in congestion level.
Under the second case, fixed time-dependent green

430

Authorized licensed use limited to: Xi'an Jiaotong-Liverpool University. Downloaded on October 24,2022 at 11:47:32 UTC from IEEE Xplore. Restrictions apply.
 proportion is considered to be the same as the
ratio of the critical phase flow to the sum of
critical flows over all phases. The assumption
for the time interval for the signal setting is to
set the signals every 10 minutes (i.e. double the
assignment interval). Convergence was attained
in all experiments at iteration number 5,8 and 11
for loading levels 1,2 and 3 respectively. Table 2
shows the results for the second set of
experiments for network 1 , for the two depth
levles (0 and 1).

FIGURE 2 : Fort Worth Network

Configuration (Network

Level (1) Level (2) Level (3)

SO Onlv 100% 100% 100%

Solution Algorithm 92.85% 91 . l o % 95.50%

Destination

The second set of experiments were performed FIGURE 3: Hypothetical Network

using a differcnt set of assumptions First, the Configuration (Network 2)
signal optimization procedure was allowed to
iterate until reaching convergence; two levels for As shown in Table 2, better performance under
the depth (K) are tested 0 and 1. The green time all experiments can be observed from these
proportions are fixed and the only decision results, though recognizing that there are still
variable at each intersection is the cycle length. some assumptions that limit the achievable
The green time proportions are calculated by improvements. As for the comparison of the
aggregating the traMic counts for the signal base case (SO solution) to different cases, these
setting interval, finding the critical movements base cases were almost the same as the ones
for each phase and finally the green time obtained from experimental set 1 for Network
l(the difference is less than 1%).

431

Authorized licensed use limited to: Xi'an Jiaotong-Liverpool University. Downloaded on October 24,2022 at 11:47:32 UTC from IEEE Xplore. Restrictions apply.
 TABLE 2 : Average Travel Time (as percentage at iteration number 5, 11 and 17 for loading
of SO assignment only at convergence) for levels 1, 2 and 3 respectively. The results for
ifferent depth level (K= 0 or I). this set of experiments are shown in Table 4.
I I I 1 The results in Table 4 show that the algorithm
results in significant improvement in the
5 0 Assignment network performance under incident conditions.

+-+-
TABLE 4 : Average Travel Time (as percentage
of SO assignment only at convergence) for
Solution
90.02 % 91.35 % 93.45 Yo different de th level K= 0 1 d2)
Algorithm K=O
Level (1)

Solution
Algorithm K=l
88.86% 89.22 % 92.53 % 100.00 % ( 100.00 %I
Solution 95.42 %

2. Results for Network 2

The first set of experiments, for Network 2, uses
the same assumptions as the second set for
Network 1. The signal optimization step using
the simulation-based method is implemented for
Solution
Algorithm
94.37 Yo
A
89.16 % 87.15 %

IV. CONCLUSION
a local area depth of 0, 1 and 2. Table 3 shows
the results for this set of experiments. All This paper presented a heuristic simulation-based
experiments reached convergence at iteration procedure to obtain time-varying assignment of
number 5, 11 and 17 for loading levels 1 , 2 and 3
0 - D vehicles to paths jointly with time-varying
signal settings that seek to optimize the total
rcspectively. It can be noted from Table 3 that
the achieved improvement increases with the travel time in the network. The algorithm was
successfully implemented on an actual network
loading level. This may be due to the fact that
from the Fort-Worth area and a hypothetical
all the intersections in the network are controlled,
aiid are considered as part of the decision network. The algorithm performed succ~ssfdly
and reached convergence achieving better netwoi k
variables.
performance in all cases.
TABLE 3 : Average Travel Time (as percentage The principal substantive conclusions from the
of SO assignment only at convergence) for application to the test networks are:
1. Joint optimization of routing and signal
setting can lead to improved overall
network performance, relative to either
routing or signal setting only.
I I I I
2. The potential for additional
improvements through signalization
(beyond that achieved through optimal
routing for given initial settings)
Algorithm appears to decrease with higher
Solution 93.50% 87.23% 85.00% congestion levels and increase with the
Algorithm increase in the percentage of controlled
intersections.
The second set of experiments for the small 3 . The algorithm showed to be robust
network assumes occurrence for an incident on under incident conditions.
link (16-18). The incident starts at time 10.0
Furthermore, the application clearly reflected the
minutes and ends at time 55 minutes (relative to
fact that the objective function is not globally
the start of the simulation) and has a severity of
convex , and that the algorithms are likely to
90 % (the severity is the percentage of reduction
discover local optima, which nonetheless may
in the link capacity). All experiments converged

432

Authorized licensed use limited to: Xi'an Jiaotong-Liverpool University. Downloaded on October 24,2022 at 11:47:32 UTC from IEEE Xplore. Restrictions apply.
 offer better network performance than the base 12. Sheffi and Powell, “Optimal Signal Setting
solution. over Transportation Networks,” Joumal of
Transp. Eng., 109, pp. 824-839 (1983).
REFERENCES 13. Gartner and Al-Malik, “A Combined Model
1. Abdelfatah and Mahmassani, “System Optimal for Signal Control and Route Choice in Urban
Time-Dependent Path Assignment and Signal Networks,” TRR # 1554, pp. 27-35 (1996)12.
Timing in Traffic Network,” TRR #1645, pp. 14. Smith and Ghali, “The Dynamics of Traffic
185-193 (1998). Assignment and Traffic Control : A Theoretical
2. Smith, “The Existence, uniqueness and Study,” Transp. Res., 24B, pp. 409-422 (1990).
Stability of Traffic Equilibria,” Transp. Res. 15. Smith and Ghali, “Dynamic Traffic
13B, pp. 295-304 (1979). Assignment and Dynamic Traffic Control,” Paper
3. Smith, ”The Existence of an Equilibrium presented at the 11” Intemational Symposium on
Solution to the Traffic Assignment Problem Transp. and Traffic Theory, Yokohama, Japan
when there are Junction Interactions,” Transp. and published in the proceedings, pp. 273-290
Res., 15B, pp. 443-451 (1981). (1990).
4. Smith, “A Local Traffic control Policy which 16. Ghali and Smith, “Traffic Assignment,
Maximizes the Overall Travel Capacity of an Traffic Control and Road Pricing,” Proceedings
Urban road Network,” Traffic Eng. and Control, of the 121h International symposium on the
?p. 298-302 (1980). Theory of Traffic Flow and Transp., California,
5. Smith, ”Properties of a Traffic Control Policy pp. -169 (1993)
which Ensure the Existence of a Traffic 17. Smith and Vuren, “Traffic Equilibrium with
Equilibrium Consistent with the Policy,” Responsive Traffic control,” Transportation
Transp. Res. 15B, pp. 453-462 (1981). Science, Vol. 27, No. 2, pp. 118-132 (1993).
5. Van Vuren, Smith, and Vliet, “The Interaction 18. Gartner and Stamatiadis, “Integration of
Between Signal-Setting Optimization and Dynamic Traffic Assignment with Real Time
Xeassignment : Background and Preliminary Traffic Adaptive Control,” Paper presented at the
Results,” TRR # 1142 (1988). 761h annual TRB meeting, (I 997)
7. Smith, Vuren, Heydecher, and Vliet, “The 19. F.V. Webster, “Traffic Signal Settings,”
Interaction Between Signal Control Policies and Road Research Technical Paper No. 39, HMSO,
Route Choice,” Paper presented at the loth London (1 958).
Intemational Symposium on Transp. and Traffic 70. Mahmassani and Peeta, “System Optimal
Theory, and printed in the proceedings, pp. 319- Dynamic Assignment for Electronic Route
338, Elsevier, NY (1987). Guidance in Congested Traffic Networks,”
E. Wardrop, “Some Theoretical Aspects of Road Proceedings of 2”d International Capri Seminar
Traffic research,” Proceedings of the Institute of on Urban Traffic Networks (1992).
Civil Engineers, Part 11, pp. 325-378, 1952 21. Hooke and Jeeves, “Direct Search Solution of
9. Allsop, “Some Possibilities for Using Traffic Numerical and Statistical Problems,” J. Ass.
Control to Influence Trip Distribution and Rout Comp. pp. 212-229 (1961)
Choice,” Paper presented at the 6Ih International 22. Hawas and Mahmassani, “Comparative
Symposium on Transp. and Traffic Theory, Anaiysis of the Robustness of Centralized and
Sydney, Australia (1974), and printed in the Distributed Network Route Control Systems in
proceedings, pp. 345-374. Incident Situations,” TRR # 1537, pp. 83-90
10. Charlesworth, “The Calculation of Mutually (1996)
Consistent Signal-settings and Traffic
Assignment for a Signal Controlled Network,”
Paper presented at the 7Ih Intemational
Symposium on Transp. and Traffic Theory,
Kyoto, Japan (1977), and printed in the
proceedings, pp. 545-569
1 1. Dickson, “A Note on Traffic Assignment and
Signal Timing in a Signal-controlled Road
Network,” Transp. Res. 15B, pp. 267-271
(1981).

433

Authorized licensed use limited to: Xi'an Jiaotong-Liverpool University. Downloaded on October 24,2022 at 11:47:32 UTC from IEEE Xplore. Restrictions apply.