14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022].

See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Computer-Aided Civil and Infrastructure Engineering 21 (2006) 321–333

Bi-level Programming Formulation and Heuristic

Solution Approach for Dynamic Traffic
Signal Optimization
Dazhi Sun
Department of Civil and Architectural Engineering, Texas A&M University–Kingsville, MSC194, Kingsville,
TX 78363, USA

Rahim F. Benekohal∗
1205, Newmark Civil Engineering Laboratory, Department of Civil and Environmental Engineering,
University of Illinois at Urbana Champaign, Urbana, IL 61801, USA

&

S. Travis Waller
ECJ 6.204, Department of Civil Engineering, University of Texas at Austin, Austin, TX 78715-0278, USA

Abstract: Although dynamic traffic control and traffic ables. ILA is developed to find user optimal flow pattern
assignment are intimately connected in the framework of at the lower level, and CTS is implemented to propagate
Intelligent Transportation Systems (ITS), they have been traffic and collect real-time traffic information. The per-
developed independent of one another by most existing re- formance of the HSA is investigated in numerical appli-
search. Conventional methods of signal timing optimiza- cations in a sample network. These applications compare
tion assume given traffic flow pattern, whereas traffic as- the efficiency and quality of the global optima achieved
signment is performed with the assumption of fixed sig- by Elitist GA and Micro GA. Furthermore, the impact
nal timing. This study develops a bi-level programming of different frequencies of updating information and dif-
formulation and heuristic solution approach (HSA) for ferent population sizes of GA on system performance is
dynamic traffic signal optimization in networks with time- analyzed.
dependent demand and stochastic route choice. In the
bi-level programming model, the upper level problem rep-
resents the decision-making behavior (signal control) of
the system manager, while the user travel behavior is rep- 1 INTRODUCTION
resented at the lower level. The HSA consists of a Genetic
Algorithm (GA) and a Cell Transmission Simulation Dynamic traffic control is strongly bonded to dynamic
(CTS) based Incremental Logit Assignment (ILA) proce- traffic assignment (DTA) in the practice of Intelli-
dure. GA is used to seek the upper level signal control vari- gent Transportation Systems (ITS), mainly because ITS
provides real-time traffic control and routing guid-
∗ Towhom correspondence should be addressed. E-mail: rbenekoh@ ance with an ultimate goal to optimize transportation
uiuc.edu. system performance. However, most existing research


C 2006 Computer-Aided Civil and Infrastructure Engineering. Published by Blackwell Publishing, 350 Main Street, Malden, MA 02148, USA,
and 9600 Garsington Road, Oxford OX4 2DQ, UK.
 14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
322 Sun, Benekohal & Waller

treats these two topics as if they are independent. Con- capable of dealing with the inherent complexity of this
ventional methods of signal timing optimization take problem including: nonconvexity, nonderivability, non-
predefined traffic flow patterns for granted and traffic as- continuity of objective and constraint functions, stochas-
signment is only performed under the condition of fixed ticity of user route choice, time-dependency of demand
signal timing. Very limited efforts are made for dealing and control, etc.
with dynamic signal control taking into account the DTA A critical review of relevant literatures is presented
simultaneously, and vice versa. in Section 2, which includes two primary areas: the com-
In a transportation system, the interplay between traf- bined optimization problem for MC solution and bi-level
fic managers and users is a typical asymmetric game, programming models and the corresponding solution ap-
well-known as a sequential game (Gartner and Al-Malik, proaches. In Section 3, the upper and lower level ob-
1996). In a sequential game, the leader (traffic manager) jectives and corresponding constraints are presented to
determines system control settings to optimize the over- give a complete bi-level programming formulation. The
all system performance; the follower (user) reacts to the framework of the proposed solution approach is pre-
leader’s action in a way to maximize his or her own ben- sented in Section 4. Meanwhile, the development of GA
efits, which eventually leads to a new equilibrium flow and CTS-ILA is presented in this section. Section 5 illus-
pattern. This problem can be viewed as a combined op- trates numerical applications of the GA and CTS-ILA
timization problem seeking a mutually consistent (MC) approach in a sample network. Section 6 concludes this
solution, which aims to find the mutual consistence be- study.
tween the optimal signal control setting and traffic as-
signment. Although many previous studies were per-
formed to find the MC solution through different algo- 2 LITERATURE REVIEW
rithms, MC solutions did not guarantee a global optimal
solution. In addition, MC solutions usually depend on Gartner and Stamatiadis (1997) presented the combined
the initial signal setting and initial flow pattern. optimization problem in a framework with two interac-
As bi-level programming provides a good framework tive loops. One loop represents the procedure to deter-
to model asymmetric games, this problem can be formu- mine the traffic assignment pattern satisfying the user
lated as a bi-level problem as well. In the bi-level prob- equilibrium (UE) for the current signal setting condition.
lem, signal timing optimization is regarded as the up- The other loop represents the procedure to determine
per level problem, while equilibrium traffic assignment the signal settings to optimize the system performance
is regarded as the lower level problem. Bi-level program- under the current traffic assignment pattern. Traffic as-
ming problems generally are difficult to solve, because signment procedures are used to get the flow pattern
the evaluation of the upper level objective involves solv- which is resulted from the existing signal settings. On the
ing the lower level problem for every feasible set of up- other hand, signal timings are set to optimize the system
per level decisions. In addition, the intrinsic complexity performance. These two procedures interact with each
comes from the nonconvexity of objective functions and other when determining the link performance, which de-
constraints at both levels, which give rise to a challenging pends on a combination of flow and signal delay.
problem for global optimization. An efficient algorithm, Most of the existing literature on this combined op-
which is capable of finding the global optima or near- timization problem is for static traffic demand or fixed
global optima of the upper level control variables and at signal timing. For example, Smith (1979, 1980, 1981a,b)
the same time finding the user-decided flow pattern for studied the uniqueness and stability of the equilibrium
the lower level problem, is still needed. between the traffic assignment and traffic control, and
Most previous studies on this problem focused on the revealed some properties of a local control policy to en-
static traffic assignment under fixed demand or fixed sig- sure the existence of the equilibrium. Smith et al. (1987)
nal timing, which constrained their potential employ- presented the important principle “less pressurized sig-
ment for a real-world problem. nal phases receive no green,” which makes it possible
The primary motivation of this study is to formulate to combine the allocation of flow and green time into
this problem as a bi-level problem and propose a solu- one integrated assignment program. Sheffi and Powell
tion approach to deal with dynamic signal control and (1983) presented a mathematical programming formu-
DTA under time-variant demand. The proposed solu- lation for this problem and an algorithm to deal with a
tion approach will take advantage of both the parallelism nonconvex formulation for a small network, along with
and globality of Genetic Algorithm (GA) in the search two heuristic algorithms for a large-scale network appli-
process and the flexibility and fidelity of Cell Transmis- cation. Gartner and Al-Malik (1996) introduced another
sion Simulation (CTS) based on the Incremental Logit iterative procedure to achieve the simultaneous solution
Assignment (ILA) method. This solution approach is of the assignment and control in a network with two
 14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Bi-level programming formulation and HSA 323

fixed phases signal control strategy. The important com- tions of MC methods usually depend on the initial signal
mon feature of the above research is that they all con- timing and initial flow pattern.
sidered the signal timing optimization under static traffic Another group of studies on this problem employed
assignment. the bi-level programming method. In the bi-level pro-
Compared with the static assignment case, very lim- gramming model, the upper level problem represents the
ited research has been reported on traffic control consid- decision-making behavior of transportation planner or
ering time-dependent traffic assignment. A conceptual system manager, while the user travel behavior is repre-
solution framework consisting of two interactive loops sented in the lower level problem. Heydecker and Khoo
for the combined optimization problem was reported (1990) formulated this problem as a bi-level problem
by Gartner and Stamatiadis (1997). But neither real im- and reported that the bi-level formulation improved the
plementation nor numerical application was reported system performance when compared with the MC so-
on using the proposed solution framework. Abdelfatah lution in their network. The most well-known solution
and Mahmassani (1998) presented a mathematical for- approach for a bi-level problem is the sensitivity analy-
mulation and simulation-based solution algorithm for sis based (SAB) approach (Chan and Lam, 1998; Wong
the combined signal control and DTA problem. They and Yang, 1997; Yang and Yagar, 1994; Yang et al., 1994;
conducted numerical experiments on the simulation- Yang and Lam, 1996; Yang and Bell, 1998). The under-
based algorithm over a realistic, moderately large lying idea of SAB is to: (1) formulate local linear ap-
network. They implemented their algorithm on a mod- proximation of the upper level nonlinear objective and
erately large signalized traffic network using well-known constraints using the derivative information from sensi-
Webster’s formulas to optimize signal settings. Chen and tivity analysis for equilibrium flow; (2) solve the resul-
Ben-Akiva (1998) introduced an integrated framework tant linear programming problem by simplex method.
to combine dynamic control and assignment. They devel- Therefore, SAB methods deal with a sequence of lin-
oped a game-theoretic methodology to model the com- ear approximations of the original problem (Yang and
bined problem as a noncooperative game between the Yagar, 1994). However, the solution of SAB cannot guar-
traffic authority and traffic users. To find efficient coordi- antee global optima. In addition, the sensitivity analysis
nated timing plans in a large network, Cheng et al. (2004) of the equilibrium network flow is a very complex and
applied the game-theoretic paradigm of fictitious play to difficult task for most practitioners, even for a medium
find the local optimal coordinated timing plan. The sig- size network (Yin, 2000).
nificant merit of their algorithm is that only one simula- Although many attempts have been made to solve the
tion is required per iteration, and therefore it would be bi-level programming problems in the past decades, un-
robustly scalable for networks of realistic sizes. fortunately most of them are either incapable of find-
Abdelfatah and Mahmassani (2001) extended their ing global optima or computationally very intensive
1998s work by replacing Webster’s formula by a for problems of realistic size (Yang and Bell, 1998).
simulation-based signal optimization, using the same Yin (2000) proposed a framework to use the Genetic
solution algorithm framework. Algorithm Based (GAB) method to solve bi-level pro-
Most of the previously discussed research can be cat- gramming and he applied the GAB method to solve
egorized as an MC solution method, because their solu- optimal road pricing and reserve capacity problems un-
tion algorithms aimed to find the MC solution between der static traffic assignment, in which the F–W algorithm
optimal signal control and traffic assignment. MC may was applied to solve the lower level optimization prob-
not be the best term for this group solution method. One lem. The research work by Abu-Lebdeh and Benekohal
alternative term for MC is equilibrium solution (ES), be- (2000) and Girianna and Benekohal (2001) showed that
cause MC actually represents the equilibrium state that GA can successfully find optimal traffic control strate-
no player has incentive to deviate from, and their actions gies for a congested network. But they did not use traf-
are sometimes referred to as “mutual best response.” It is fic assignment as a part of traffic management. Ceylan
worth noting that the concept of equilibrium is different and Bell (2004) developed a GA approach to optimize
from global optimality. Therefore, the intrinsic problem traffic signal timing considering drivers’ routing. They
of the MC method is that the MC solution (ES) may not used the Path Flow Estimator to solve the lower level
be the global optima, especially for nonconvex objective problem. To simplify the GA chromosome, they used a
functions and constraints where local optima may exist. common network cycle time for all signalized intersec-
Ceylan (2002) discussed this problem by illustrating the tions, which may not be suitable for a realistic network.
MC approach in two-dimensional contours, which illus- In addition, further improvement on their models is re-
trates that the MC solution of signal control and traffic quired to deal with time-dependent demand and time-
assignment may not be the global optimal solution, in variant control strategies. It entails our study to develop
terms of system performance. In addition, the final solu- a bi-level programming formulation and propose a new
 14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
324 Sun, Benekohal & Waller

solution approach to deal with a dynamic network signal subject to

control conjunction with stochastic route choice under
time-variant demand. ul ≤ u ≤ uu (3a)
where A is the set of links in the network, va is the flow
3 FORMULATION on link a, ω(va , u) is link cost function defined by system
managers, given u, ul is the lower bound of the control
3.1 The general bi-level programming formulation vector, uu is the upper bound of the control vector, and
u is the control vector.
One of the transportation system manager’s objectives ω(va , u) is a general cost function; for example, a delay-
is to optimize system overall performance by adjusting based function is formulated as:
control settings or management strategies, which can be
stated as a function of one or a combination of the fol- ω(va , u) = α D(va , u) (4)
lowing factors: total delay, throughput, traveling time,
where ω(va , u) is the link cost function, D(va , u) is the
energy consumption, noise and emission, etc. The system
average link travel time, and α is the constant factor.
user is supposed to respond to maximize his/her own
Given the time-dependent demand and control fea-
benefit, for example, choosing the shortest path.
tures, the upper level objective function will be expressed
The problem of traffic control optimization conjunc-
as:
tion with DTA can be mathematically formulated as a   
bi-level programming model: min vat ω vat , ut (5)
u
t a
min F(u, v(u)) (1)
u subject to
subject to
ult ≤ ut ≤ uut , for all t (5a)
G(u, v(u)) ≤ 0 (1a)
where vat is the flow on link a during the time period of
where v(u) is obtained by solving: t, ω(vat , ut ) is link cost function during the time period of
t, given u, ult is the lower bound of the control vector for
min f (u, v) (2) the time period of t, uut is the upper bound of the control
v
vector for the time period of t, and ut is the control vector
subject to
of time period t.
g(u, v) ≤ 0 (2a) In a simulation-based DTA framework, the average
travel time of each link can be computed by:
where F(u, v(u)) is the upper level objective function of
system managers, G(u, v(u)) is the constraint set of the T1 + T2 + · · · + Tvat
α · Dt (va , u) = α · (6)
upper level problem, f (u, v) is the lower level objective vat
function of system users, g(u, v) is the constraint set of
the lower level problem, u is the upper level control vec- where Ti is the travel time of vehicle i on the link a at
tor (signal timing vector), and v is the lower level control time interval t.
vector (flow pattern).
System managers seek the optimal control setting u∗ , 3.3 The formulation for the lower level problem
through solving minu F(u, v(u)); meanwhile, they have
to take into account the corresponding flow pattern v∗ (u) The lower level problem deals with users’ route choice
resulting from users’ response to the given control vari- behavior; therefore the decision variables are the set of
able u, which is solved by the lower level problem. the flow resulted from user optimal route choice. To get a
realistic reflection of the traffic dynamics, the lower level
3.2 The formulation for the upper level problem problem was modeled as a Reactive Dynamic Stochastic
User Optimal (RDSUO) problem. The RDSUO prob-
The upper level problem represents the decision-making lem is to determine the vehicle flow at each instant of
behavior of the transportation planner or system man- time using the perceived minimum cost routes based on
ager to optimize the system performance. The decision the current traffic information of the system.
variables in this level are the set of signal timing variables, Many previous studies have been focused on Static
including cycle length, green ratio, and offset. Therefore, Deterministic User Equilibrium (SDUE), such as: Sheffi
the upper level objective function can be expressed as: (1985) provided a mathematical formulation on the

min va ω(va , u) (3) deterministic UE problem, without link capacity con-
u straints. Yang and Yagar (1994) and Yang et al. (1994)
a∈A
 14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Bi-level programming formulation and HSA 325

proposed a deterministic UE formulation for a link ca- Therefore, the probability of choosing route r for O–D
pacity constrained problem with a static O–D demand. pair w can be computed by:
One important difference between the SDUE assign-  
exp −θgrt
ment and Static Stochastic User Equilibrium (SSUE) pr,w =
t    (9)
assignment is that the former assumes that every user exp −θgit
has the perfect information on the link performance and i∈Rw
there is no difference between the perceptions of differ- the path flow assigned to route r from O–D demand Dw
ent users, whereas the latter deals with perception er- is calculated by:
rors of different users for route choice. The most known
choice model is Logit, because it has closed form and
t
hr,w = D tw · pr,w
t
(10)
equivalent mathematical formulation for a Stochastic
where Urt is the disutility of route r at time interval t, grt
UE problem. Logit models were widely used in trans-
is the average user-perceived path travel time for route
portation literature (Fisk, 1980; Chen and Alfa, 1991;
r at time interval t, θ is the dispersion parameter, pr,w t
Bell et al., 1993; Leurent, 1995; Damberg et al., 1996).
is the probability of choosing route r for O–D pair w at
The difference between the formulation of SSUE and t
time interval t, hr,w is the path flow assigned to route r
SDUE is that an entropy term was introduced to the
for O–D pair w at time interval t, D tw is the O–D demand
objective function of SSUE to deal with the stochastic
for pair w at time interval t, Rw is the set of routes for
route choice behavior.
O–D pair w.
An equivalent mathematical formulation of RDSUO
Although the above formulation gives a very straight-
using logit model given time-variant demand is given as:
forward description of a time-dependent demand traffic
 
min Z = Z t1 + Z t2 (7) assignment problem, link flow may not be the simple lin-
t ear combination of all path flow using it from the same
 vat
time interval. The link flow of each short period may
Z t1 = ξ (t)−1 ca (x, ut ) dx (7a) contain some flow assigned at the previous time intervals
a∈A 0 on different paths. The mix-up of link flow makes it diffi-
cult for conventional mathematical solution approaches;
1   t therefore we resort to a simulation-based solution. The
Z t2 = t
h ln hr,w (7b)
θ w∈W r ∈Rw r,w reason for using a simulation-based method is the cur-
rent lack of analytical representation of traffic flow that
subject to can adequately replicate the dynamic interaction among
 vehicles and signal control, and yield a well-behaved
t
hr,w = D tw , w∈W (7c)
r ∈Rw
mathematical formulation (Peeta and Ziliaskopoulos,
2001). What is more important, is that a simulation-based
t
hr,w ≥ 0, ∀t, r, w (7d) model has the capability to study a more complex system
with multiple information availabilities, different user
where Z t1 is the deterministic term of the objective func- groups, varying signal controls, management strategies,
tion denoting the total travel time, Z t2 is the entropy term etc.
t
designed for logit model, hr,w is the path flow assigned
to route r for O–D pair w at time interval t. ca (x, ut ) is
the user-perceived travel cost on link a at time interval 4 SOLUTION APPROACH
of t, θ is the dispersion parameter, D tw is the demand
for O–D pair w ∈ W during time interval of t, Rw is the 4.1 The solution framework
set of routes for O–D pair w, hr,w t
is the flow assigned
to route r for O–D pair w during time interval of t, and The heuristic solution approach (HSA) consists of two
t
hr,w t
ln hr,w = 0 for hr,w
t
= 0, ξ (t) is a time-variant ratio critical components: GA and CTS-ILA. GA is used to
of the current traffic volume in the system to the total find the optima or near-optima for system control strat-
loaded traffic volume into the system (a term defined for egy based on the evolutionary mechanism; CTS-ILA is
ILA approach), Equations (7c) and (7d) are flow conser- developed to capture the dynamics of traffic flow and
vation constraints and nonnegative path flow constraint. route choice behavior in networks with time-variant con-
In a logit route choice model, the demand is propor- trol and demand.
tionally distributed to paths according to their corre- The mechanism of the proposed solution approach is
sponding disutility Ur , which is calculated as: outlined in Figure 1.
  The GA application for the bi-level programming
Urt = exp −θgrt (8) problem is outlined as:
 14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
326 Sun, Benekohal & Waller

Generate initial population

of upper level control vector u

Apply dynamic traffic assignment to find

lower level problem
corresponding flow v i for each u i

Evaluate the upper level objectives of

each control pair ( u i v i ) upper level problem

Apply selection, crossover, mutation

to generate new generation

N
Termination?

Return the best control set u *

Fig. 1. Solution framework.

1) Decide the GA coding system to represent the sys- Nihan et al. (1995) applied logit model and the method
tem control variables, and specify all critical param- of successive average (MSA) for calculating route choice
eters of GA. probability and equilibrium flow to study the interaction
2) System initialization to generate the first gen- between driver information, route choice, and optimal
eration. signal timing for a simple system with only two routes
3) Decode the generated signal control variables in and one “T” intersection and fixed O–D demand.
each chromosome and deploy these signal control The CTS-ILA module first generates the paths for
variables to the studied network. Solve the lower each O–D pair, then dynamically calculates path travel
level problem by CTS-ILA to get the stochastic user time, and incrementally assigns the demand to corre-
optimal flows. sponding paths using a logit model and immediately
4) Calculate the upper level objective function us- updated path information at every time interval. The
ing the given signal control variables and the user- new travel information on each link—therefore on each
decided flows; therefore the fitness values of each path—can be dynamically calculated based on the sim-
population can be determined. ulated traffic performance.
5) Perform selection, crossover, and mutation to gen- The ILA module first enumerates all paths for each
erate the next generation. O–D pair and calculates the utility of each path using
6) Repeat Steps 3–5 until the predetermined maxi- the measured path cost. Then, ILA computes the path
mum number of generations has been achieved or use probability using the information of path utility and
convergence criteria has been satisfied. loads the demand into the network interval by inter-
val. Within each time interval, an iterative procedure is
A CTS model is programmed to propagate traffic, applied to find the user optimal path flow at the current
perform DTA, and evaluate the fitness at Step 4 as interval. A logit model is applied to calculate the path
well. use probability using the path travel time measured by a
 14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Bi-level programming formulation and HSA 327

CTS simulator, and load the interval demand onto cor- fic distribution. In this study, we assumed that the time
responding paths and run CTS again. The logit traffic as- plan varies from time interval to time interval; however,
signment procedure is to be performed iteratively, and the time plan within a given period will remain constant.
the interval path flow is updated using an MSA method The binary string representing control variables over all
until some convergence criteria are satisfied. periods is shown in Figure 2.
The binary string will be first translated into a set of
real numbers which are used to present the value of cy-
4.2 GA cle length, green split, and offset. A simple variant of
GAs-based optimization techniques have been widely the above string representation is to assume a common
and successfully implemented in many engineering ap- cycle length for all intersections along the same arterial
plications in the past decades. These techniques are or route. For a large-scale network, this helps shorten
based on the mechanism of natural selection and nat- the string length.
ural genetics (Goldberg, 1989), and are considered as In the numerical applications, Elitist GA and Micro
emerging optimization tools due to their simplicity, min- GA, both of which use binary tournament selection
imal problem restriction, parallelism, and global per- and two-point crossover, are applied in the solution
spective. Recently, Simple GA (SGA), Micro GA, and approach.
Parallel GA (PGA) were implemented for intersection
signal optimization by Park (1998), Abu-Lebdeh and 4.3 System architecture
Benekohal (1999, 2000), Park et al. (2000, 2001), Ceylan The solution system consists of a control module, ILA
(2002), Girianna and Benekohal (2001), and Ceylan and module, evaluation module, and a GA component. The
Bell (2004). SGA principally runs in a sequential envi- architecture of the solution system is shown in Figure 3.
ronment in which a set of possible solutions (population)
is generated and improved using genetic operators of
selection, crossover, and mutation. SGA converges at 4.3.1 Control module. The control module receives the
O (n log n) function evaluations for n population size signal control information from GA, and implements
(Goldberg and Deb, 1991). The trade-off between the them to the corresponding intersections (signalized cells
high quality of the solution and computation costs (con- in the Cell Transmission Simulator). Therefore, the sig-
vergence time) is a major issue associated with SGA. In nalized cells will have time-dependent capacity. The
this study, we adopted an Elitist GA, which replicates the function of the control module is to dynamically change
best individual of the current generation to the next gen- the signalized cell’s capacity (discharge rate), and coor-
eration. This is the only difference between this Elitist dinate their operations according to the given control
GA and a normal SGA. strategy.
Compared with regular (full size) GA, Micro GA 4.3.2 ILA module. Incremental assignment methods di-
takes a small population (n), converges within the small vide the O–D demand matrix to a set of the submatrix.
population, then keeps the best individual, and randomly Each submatrix corresponds to the O–D demand of one
generates another n − 1 individuals so that the new pop- time interval. MSA is used in each interval to find the
ulation is formed and so on. equilibrium flow, and the algorithm is shown as follows:
System decision variables in this problem are cycle
length, offset, and green split for each signalized inter- (1) n ← 1
section in the system over multiple periods. The control (2) Apply k-shortest path algorithm to find k̄rs (n)

plan varies due to the varying demand and dynamic traf- based on path flow pattern h = i−1
k=1 hk + hi

intersection 1 intersection i intersection 1 intersection i

cycle offset green split cycle offset green split cycle offset green split cycle offset green split

010010 010101 010010 ... 010010 010101 010010 ... 010010 010101 010010 ... 010010 010101 010010 ...

period 1 period i
Fig. 2. GA chromosome structure.
 14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
328 Sun, Benekohal & Waller

evaluate fitness
Evaluation Module

evaluate system performance

update apply control
control
GA Control Module Network

assign traffic update path cost

ILA Module

GA Traffic Assignment

Fig. 3. Simulator architecture.

(3) p ← ψ(g ∈ k̄rs (n)) (logit model for calculating time of link a as the no. (i − 1) vehicle leaves link a; Ta,i
path probability) is the travel time of the no. i vehicle on link a. Therefore,
(4) hi ← ti p (get auxiliary incremental path flow) link cost at time t equals ca,k, given the no. k vehicle is
(5) Apply MSA hi ← hi (1 − 1/n) + hi (1/n) the latest vehicle that leaves link a at time t.
(6) n ← n + 1. In Equation (14), the newest experienced costs dom-
inate the updated link cost, indicating that the imme-
If convergence is insufficient, then return to Step 2.
diate traffic condition dominates the value of link cost
Else stop.
function. Meanwhile, Equation (14) can control random
In this HSA, the demand matrix is first divided into impact through the averaging.
submatrices:
 
Dw = D tw (11) 4.3.3 Cell transmission simulation. In CTS, the links are
w∈W w∈W t divided into three types of cells: normal cell, merg-
the submatrix of time t is distributed into different paths ing/diverging cell, and signalized cell. The normal cell
according to logit assignment: represents the normal section in the highway, the only
function of which is to move the vehicle forward. The
hrt = φr,w
t
D tw , w∈W (12) diverging cell assigns the left/right turn vehicles to the
φr,w
where t
are the weighting factors such that ≥ φr,w
t exclusive left/right turn lanes, and moves the through
0, and r ∈Rw φr,w
t
= 1 and φr,w
t
is equal to the P(r ) cal- t traffic to the next through traffic cell. Similarly, the merg-
culated by Equation (9). Path costs are the sum of the ing cell will accept the vehicles from multiple upstream
costs of the constituent links, the cost of path r is: cells based on the given rules. The signalized cell is a
 variant of the normal cell, which has the time-varying
grt = δar cat (13) supply. The length of cell depends on the speed limit and
a
the length of simulation time step. For example, the cell
where cat
is the updated cost of link a at the beginning length used in this study is 88 ft using a speed limit of
of the time interval t, δar is an element of the link-path 30 mph and 2-second time step. For more details on the
incidence matrix. To reflect the newest traffic condition, cell transmission model, please refer to Daganzo (1994,
the link cost is updated once a new vehicle finishes its 1995).
trip on that link. Therefore, the link cost is calculated as:
ca,i−1 Ta,i 
i−2
Ta,i− j Ta,1 + Ta,2
ca,i = + = j+1
+ (14) 5 NUMERICAL APPLICATION OF HSA
2 2 j=0
2 2i−1
where ca,i is the newest updated travel time of link a as The proposed HSA is implemented in a sample net-
the no. i vehicle finishes its trip on link a, ca,i−1 is travel work slightly modified from the often cited Nguyen and
 14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Bi-level programming formulation and HSA 329

1 3
A 1 2 4
10
17 18
2
4 6
8 12
3 5 7 9 13

19 21 16
20

5 7 9 13
6 8 10 14 B

22
23
15
11
14
11 12

Fig. 4. A sample network with 10 signalized intersections.

Dupius’s test network with 10 signalized intersections. bility of the signal timing found by the HSA in a real
The topology of the sample network is shown in Figure 4. transportation network and (2) investigate the compa-
In this sample network, there are two O–D pairs: from rability of the system performances generated by HSA
A to B and from C to D. There are 29 alternative paths and obtained from a realistic network operation.
for O–D pair A–B and 13 paths for O–D pair C–D. In HSA was applied to find the signal timings for this sam-
this sample network, we consider one-way traffic which ple network. Then, the signal timing and corresponding
makes a straightforward control plan for the signalized traffic assignment found by HSA were used as the input
nodes. Each of them has a two-phase plan and the move- to CORSIM. CORSIM simulates network traffic opera-
ment of each phase is shown in Figure 4. In this sample tion using the signal timing generated by HSA, and its
network, the speed limit is 30 mph and the saturation operational performance was compared with that gener-
flow is 1,800 veh/h. The demands assigned to each O–D ated by the simulation module of HSA. The validation
are changed every 15 minutes, and correspondingly the procedure is depicted in Figure 5.
signal timings are updated every 15 minutes. A paired t-test was performed for the hypothesis that
First, HSA is validated with the prevailing commercial the MOEs (e.g., no. of vehicles discharged and average
simulation software CORSIM to (1) study the applica- link travel time) obtained from HSA and CORSIM come

MOE
validation

operational operational
performances performance

network with signal timing

HSA traffic assignment
CORSIM
static demand

Fig. 5. The validation of HSA.

 14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
330 Sun, Benekohal & Waller

from distributions with equal means. The difference 5.1 Information updating frequency
between them is assumed to come from a normal distri-
The frequency of updating traffic information will im-
bution with unknown variance. The results show that no
pact the accuracy of information, and therefore impact
significant difference exists in the network-level MOE
drivers’ route choice decision. In the numerical applica-
(no. of vehicles discharged) generated by the two ap-
tions, two cases are studied. One is to update information
proaches. At the link-level, the travel time on the major-
every 5 minutes and the other updates information ev-
ity of links (over 83%) are comparable. The validation
ery 1 minute. The information to be updated includes the
shows that the signal timing obtained from HSA can be
link travel time and path travel time, which is fed back to
applied into a “real” network and it can provide a real-
the ILA module for choosing the shortest path. Figure 6
istic presentation of traffic operation.
shows the convergence curves of these two cases.
To address the advantage of utilizing the dynamic sig-
In Figure 6, the solid lines represent the best fitness
nal timing found by HSA to a network with time-varying
of each generation and the dashed lines represent the
traffic, we compared the performance of dynamic signal
average fitness of each generation. Once the difference
timing and static signal timing on the same network. In
between the best fitness and average fitness of one gen-
the real world, static signal timings are widely used in
eration is less than 5%, it is considered converged in this
places where traffic flow changes over time. To simu-
study. The convergence curves showed that the best fit-
late this phenomenon, the static signal timing was first
ness and the average fitness continue to be improved for
found by HSA assuming constant traffic flow. Then, the
a certain number of generations after the first occurrence
static signal timing was applied to two time periods, TP1
of convergence. Numerical applications showed that it is
and TP2, which have different traffic flow. Second, dy-
adequate to stop the GA searching process at generation
namic signal timing was found by HSA based on time-
of 100, as a stable convergence state had been achieved.
varying traffic condition and implemented to the same
It is observed that the system performance is improved
network. The comparison results showed that utilizing
by 18% as the traffic information is updated from every 5
the dynamic signal timing plan found by HSA can reduce
minutes to every 1 minute. This indicates that shortening
the link travel time by 5–8% on average, and sometimes
the time interval for updating information and assigning
it can reach 14%.
traffic helps drivers avoid some congested links while
After the validation, HSA is applied in the sample
choosing a route and reducing travel time. In addition,
network to address:
the numerical application with shorter interval showed
better convergence features in terms of the number of
1. the impact of implementation of different frequen- generations before the first time of convergence.
cies for updating traffic information,
2. the efficiency and quality of searching the global
5.2 Elitist GA versus Micro GA
optimum by Elitist GA and Micro GA, and
3. the impact of population size of GA on searching Elitist GAs population size started at 50 and in-
quality. creased until no significant improvement was observed.

240

220
fitness (travel time in seconds)

200 min fitness (5)

avg fitness (5)
180
min fitness (1)
160 avg fitness (1)

140

120

100

80
0 10 20 30 40 50 60 70 80 90 100
generation

Fig. 6. Comparison of different frequency of updating information.

 14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Bi-level programming formulation and HSA 331

240

220

200
Population=50
180 Population=100
travel time(sec.) Population=200
160
Population=300
140

120

100

80

60
0 10 20 30 40 50 60 70 80 90 100
generation

Fig. 7. Elitist GA with different population sizes.

Given the computation time is 5,000 fitness evaluation, 95.15 (Elitist GA, population of 100 and generation of
Micro GA with a population size of 5 was compared with 50), 97.78 (Elitist GA, population of 200 and genera-
Elitist GAs with population sizes of 50, 100, 200, and 300, tion of 25), 104.5 (Elitist GA, population of 300 and
respectively. The Micro GA ran 1,000 generations, the generation of 17), and 96.6 (Micro GA, population of
Elitist GAs ran 100, 50, 25, and 17 generations depend- 5 and generation of 1,000). Under the same amount of
ing on population size. The convergence curves of the computational time (5,000 fitness evaluations), Micro
Elitist GA and Micro GA are shown in Figures 7 and 8, GA is comparable to the Elitist GA when population
respectively. size was between 100 and 200. With the computational
The average fitness values generated by differ- time of 5,000 fitness evaluation, the population size of
ent GAs after 5,000 fitness evaluations are: 102.45 100 obtained the best average fitness value. However,
(Elitist GA, population of 50 and generation of 100), with computational time of 10,000 fitness evaluations,

170

160

150

min fitness
140
avg. fitness
travel time (sec)

130

120

110

100

90

80
0 100 200 300 400 500 600 700 800 900 1000
generation

Fig. 8. Micro GA (population size = 5).

 14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
332 Sun, Benekohal & Waller

the average fitness value of the population size of 100 the cache in the GA component of the HSA saved the
(at generation of 100) is 94.5; the average fitness value total computation time by 15–20%.
of population size of 200 (at generation of 50) is 91.05; As GA is a search tool for near-optimal solution and
and for population size of 300 (at the generation of 34) some small perturbations may result in substantially dif-
is 94.4; therefore population size of 200 generates the ferent responses by different players in a game, the near-
best result given the 10,000 times of fitness evaluation. optimal solution found for an individual problem may
It is also observed that a population size of 50 results not guarantee the global optima and a globally efficient
in premature convergence and a population size of 300 outcome unless many runs of the GA with different ini-
generates similar results as a population size of 200. This tial populations are made and compared.
indicates that using a population size of 200 is adequate The research presented in this article can be further
to generate a satisfactory result for this midsize network. extended into several areas, such as to (1) Explore PGA
With a larger population size, computational efficiency and self-adaptive GA to improve the computational ef-
is sacrificed without effective improvement being made. ficiency and search quality; (2) Investigate a distributed
In this HSA, the fitness value of each individual is signal system instead of a centralized signal system; (3)
evaluated by the simulation, which takes more than Explore how to integrate the off-line traffic signal con-
10 seconds for this midsize network. Therefore, 5,000 trol optimization into a real-time online control system.
times of fitness evaluations cost about 14 hours, which In addition, further comparison between the computa-
was run on a PC with Pentium 1.8 GHz processor. As a tional efficiency and effectiveness of GA and MC meth-
matter of fact, many individuals are repeated in the later ods would be one area for future explorations.
generations before the GA search terminates. To avoid
reevaluating the fitness value of the same individual, a
cache option is integrated into the GAs, which keeps the REFERENCES
record of the individuals whose fitness value has already
been evaluated. By using the cache option, the compu- Abdelfatah, A. S. & Mahmassani, H. S. (1998), System Opti-
tation time is found to be significantly reduced by over mal Time-Dependent Path Assignment and Signal Timing in
20% for the Elitist GA with a population size of 100, and Traffic Network, TRR 1645, pp. 185–93.
Abdelfatah, A. S. & Mahmassani, H. S. (2001), A simulation-
more than 15% for the Micro GA application. based signal optimization algorithm within a dynamic traffic
assignment framework, in Proceedings of 2001 IEEE In-
telligent Transportation Systems Conference, Oakland, CA,
6 CONCLUSION pp. 428–33.
Abu-Lebdeh, G. & Benekohal, R. F. (1999), Convergence vari-
ability and population sizing in micro-genetic algorithms,
This study develops a bi-level programming formulation Computer-Aided Civil and Infrastructure Engineering, 14,
and HSA for the problem of signal control optimization 321–34.
in networks with stochastic route choice and time-variant Abu-Lebdeh, G. & Benekohal, R. F. (2000), Genetic Algo-
demand. In this solution approach, GA is used to find the rithms for Traffic Signal Control and Queue Management of
upper level signal control variables, ILA is developed Oversaturated Two-Way Arterials, Transportation Research
Record, No. 1727, pp. 61–67.
to find the user optimal flow pattern in the lower level Bell, M. G. H., Lam, W. H. K., Ploss, G. & Inaudi, D. (1993),
problem, and CTS is implemented to propagate traffic Stochastic user equilibrium assignment and iterative bal-
and collect real-time traffic information. ancing, in Proceedings of the 12th International Symposium
The validation of HSA shows that it can provide a on Transportation and Traffic Theory, Berkeley, California,
realistic representation of traffic operation in both net- pp. 427–39.
Ceylan, H. (2002), A genetic algorithm approach to the equi-
work and link level operational performance. The im- librium network design problem, Ph.D. thesis, University of
plementation of HSA in the sample network shows that Newcastle upon Tyne.
the system performance is improved by 18% as traffic Ceylan, H. & Bell, M. G. H. (2004), Traffic signal timing op-
information for choosing route is updated from every timization based on genetic algorithm approach, includ-
5 minutes to every 1 minute. In the numerical exper- ing drivers’ routing, Transportation Research, 38B(4), 329–
42.
iments, Elitist GA and Micro GA are compared. No Chan, K. S. & Lam, H. K. (1998), A sensitivity analysis based
significant differences are observed between the perfor- algorithm for determining the optimal detector density of
mance of Elitist GA and Micro GA with a small amount the network with variable message sign, in Proceedings of
of fitness evaluation, for example, a 2,500 fitness evalu- the 3rd Conference of the Hong Kong Society for Transporta-
ation. However, using large amounts of fitness evalua- tion Studies, Hong Kong Polytechnic University, Hong Kong,
pp. 247–56.
tion, for example, a 5,000 fitness evaluation, Micro GA Chen, M. & Alfa, A. S. (1991), Algorithms for solving Fisk’s
is more likely to obtain better results than Elitist GA, stochastic traffic assignment model, Transportation Re-
given the same computational load. In addition, using search, 25B, 405–12.
 14678667, 2006, 5, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/j.1467-8667.2006.00439.x by Xi'An Jiaotong-Liverpool, Wiley Online Library on [25/10/2022]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Bi-level programming formulation and HSA 333

Chen, O. J. & Ben-Akiva, M. E. (1998), Game-theoretic formu- intersections, Transportation Research Record, 1727, 32–
lations of the interactions between dynamic traffic control 41.
and dynamic traffic assignment, Transportation Research Park, B., Rouphail, N. M. & Sacks, J. (2001), Assessment of
Record, 1617, 179–88. a stochastic signal optimization method using microsimula-
Cheng, S. F., Epelman, M. A. & Smith, R. L. (2004), Cosign: tion, Transportation Research Record, 1748, 40–45.
A Fictitious Play Algorithm for Coordinated Traffic Signal Peeta, S. & Ziliaskopoulos, A. (2001), Foundations of dynamic
Control, IOE Technical Report 04-08. University of Michi- traffic assignment: The past, the present and the future, Net-
gan, Ann Arbor, MI. works and Spatial Economics, 1(3/4), 233–66.
Daganzo, C. F. (1994), The cell transmission model: A dynamic Sheffi, Y. (1985), Urban Transportation Networks: Equilibrium
representation of highway traffic consistent with the hydro- Analysis with Mathematical Programming Methods, Pren-
dynamic theory, Transportation Research B, 28(4), 269–87. tice Hall, Englewood Cliffs, New Jersey.
Daganzo, C. F. (1995), The cell transmission model. Part II: Sheffi, Y. & Powell, W. B. (1983), Optimal signal setting over
Network traffic, Transportation Research, 29B(2), 79–93. transportation network, Journal of Transportation Engineer-
Damberg, O., Lundgren, J. T. & Patriksson, M. (1996), An al- ing, 109, 824–39.
gorithm for the stochastic user-equilibrium problem, Trans- Smith, M. J. (1979), The existence, uniqueness and stability
portation Research, 30B, 115–31. of traffic equilibrium, Transportation Research, 13B, 235–
Fisk, C. (1980), Some development in equilibrium traffic as- 304.
signment, Transportation Research, 14B, 243–55. Smith, M. J. (1980), A local traffic control policy which maxi-
Gartner, N. H. & Al-Malik, M. (1996), Combined Model for mizes the overall travel capacity of an urban road network,
Signal Control and Choice in Urban Networks, TRR 1554, Traffic Engineering and Control, 21, 298–302.
pp. 27–35. Smith, M. J. (1981a), The existence of an equilibrium solution
Gartner, N. H. & Stamatiadis, C. (1997), Integration of Dy- to the traffic assignment problem when there are junction
namic Assignment with Real-Time Traffic Adaptive Control, intersections, Transportation Research, 15B, 443–51.
TRR 1644, pp. 150–56. Smith, M. J. (1981b), Properties of a traffic control policy which
Girianna, M. & Benekohal, R. F. (2001), Dynamic Signal Co- ensure the existence of traffic equilibrium consistent with the
ordination for Networks with Oversaturated Intersections, policy, Transportation Research, 15B, 453–62.
Transportation Research Record, No. 1811, pp. 122–30. Smith, M. J., Van Vuren, T., Heydecher, B. G. & Vilet, D. V.
Goldberg, D. E. (1989), Genetic Algorithm in Searching, Opti- (1987), The interaction between signal control policies and
mization and Machine Learning, Addision-Wesley, Reading, route choice, Paper Presented at the 10th International Sym-
MA. posium on Transportation and Traffic Theory, Cambridge,
Goldberg, D. E. & Deb, K. (1991), A comparative analy- MA, pp. 319–38.
sis of selection schemes used in genetic algorithms, in G. Wong, S. C. & Yang, H. (1997), Reserve capacity of a signal-
Rawlins (ed.), Foundations of Genetic Algorithms, Morgan controlled road network, Transportation Research, 31B(5),
Kaufmann, San Mateo, CA, pp. 69–93. 397–402.
Heydecker, B. G. & Khoo, T. K. (1990), The equilibrium net- Yang, H. & Bell, M. G. H. (1998), Models and algorithms for
work design problem, in Proceedings of AIRO’90 Confer- road network design: A review and some new development,
ence on Models and Methods for Decision Support, Sorrento, Transportation Review, 18(3), 257–78.
Florida, pp. 587–602. Yang, H. & Lam, H. K. (1996), Optimal road tolls under con-
Leurent, F. M. (1995), Contributions to the logit assignment ditions of queuing and congestion, Transportation Research,
model, Journal of Transportation Research Record, 1493, 30A(5), 319–32.
207–12. Yang, H. & Yagar, S. (1994), Traffic assignment and traffic con-
Nihan, N., Hamed, M. & Davis, G. (1995), Interactions between trol in general freeway-arterial corridor systems, Transporta-
driver information, route choice, and optimal signal timing tion Research, 28B(6), 463–86.
on a simple network, Journal of Advanced Transportation, Yang, H., Yagar, S., Iida, Y. & Asakura, Y. (1994), An algorithm
29(2), 163–82. for the inflow control problem on urban freeway networks
Park, B. (1998), Development of genetic algorithm-based sig- with user-optimal flows, Transportation Research, 28B(2),
nal optimization program for oversaturated intersections, 123–39.
Dissertation, Texas A&M University. Yin, Y. (2000), Genetic-algorithms-based approach for bilevel
Park, B., Messer, C. J. & Urbanik, T. (2000), Enhanced genetic programming models, Journal of Transportation Engineer-
algorithm for signal-timing optimization of oversaturated ing, 126(2), 115–20.