See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/245558581

System Optimal Time-Dependent Path Assignment and Signal Timing in Traffic

Network

Article in Transportation Research Record Journal of the Transportation Research Board · January 1998
DOI: 10.3141/1645-23

CITATIONS READS
63 181

2 authors:

Akmal Abdelfatah Hani S. Mahmassani

American University of Sharjah Northwestern University
39 PUBLICATIONS 1,033 CITATIONS 523 PUBLICATIONS 18,018 CITATIONS

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Evolving Mobility Services: Shared Mobility, Autonomous Vehicle Fleets, and Business Models View project

Vaccine Logistics and Distribution View project

All content following this page was uploaded by Hani S. Mahmassani on 08 July 2015.

The user has requested enhancement of the downloaded file.

TRANSPORTATION RESEARCH RECORD 1645 Paper No. 98-1349 185

System Optimal Time-Dependent Path

Assignment and Signal Timing in
Traffic Network
AKMAL S. ABDELFATAH AND HANI S. MAHMASSANI

Optimization of system performance in congested traffic networks is desires) is presented, and a heuristic solution algorithm is proposed,
one of the main goals of intelligent transportation systems. A formula- along with a computationally more efficient variant. Numerical exper-
tion and solution algorithm for the combined signal control and dynamic iments with the algorithm on a realistic moderately large network (a
traffic assignment problem is presented. The solution algorithm is
portion of the Fort Worth area) underscore the potential of additional
implemented and tested on an actual network, illustrating the benefits
that could be attained through joint optimization of signal control and improvement through joint consideration of signal control and route
route guidance decisions. assignment compared to either one of these decisions alone.

The performance of a traffic network can be influenced through sev- BACKGROUND REVIEW
eral types of actions or decision variables. Some of these pertain to
changing the load pattern on the network, through demand manage- Most prior research that has recognized the interaction between
ment actions, including attempts to route vehicles optimally through signal timing and traffic assignment has addressed static demand
the network; others pertain to how traffic flow is controlled through conditions—constant origin-destination (O-D) trip rates and con-
the network components, such as junction utilization through signal stant average flow rates through the network. Furthermore, previous
control (supply management). Although the potential of explicitly work focused primarily on setting signals in a manner consistent
combining both types of actions, especially joint signal control and with users’ choice of route, that is, with a descriptive assignment
route assignment, long has been suggested, most of the research and rule, typically resulting in a Wardrop user equilibrium (UE), instead
virtually all the approaches used in practice have followed one of of seeking to optimize performance through system optimal (SO)
two schemes: (a) considering the signal control to be fixed and using route assignment jointly with signal control.
traffic assignment as decision variables (traffic assignment models),
and (b) considering traffic assignment (loading pattern) to be fixed
and using signal control as decision variables (signal optimization Signal Timing Under Static Assignment
models).
Nonetheless, several researchers have recognized the interaction Some of the most important theoretical contributions to the problem
between signal control and traffic assignment. However, the difficulty of signal control and UE traffic assignment were made by Smith
of the problem in realistic networks has precluded the development of (1,2), who derived conditions that guarantee the existence of an
general robust procedures; most results are limited to idealized for- equilibrium as well as conditions for the uniqueness and stability of
mulations or to static network conditions. Furthermore, only recently the traffic equilibrium when there is interaction between signal set-
have developments in intelligent transportation systems (ITS) made ting and users’ route choice decisions (which determine the link
the prospect of route control through dynamic route guidance a likely flows). Smith (3,4) also showed the properties of a local control pol-
possibility in the foreseeable future. It should be noted that situations icy, named Po, which ensure the existence of a traffic equilibrium
in which joint optimization through signal control and route control consistent with the control policy.
would be called to correspond to congested networks, and congestion Van Vuren et al. (5) and Smith et al. (6) defined a principle that is
is a time-varying phenomenon. Therefore, practically relevant for- equivalent to Wardrop’s (7) first principle (more costly routes carry
mulations of the problem should be time dependent. The considera- no flow) such that the allocation of flows and green times can be
tion of dynamic route assignment and signal control in a time-varying combined into one integrated assignment program. The equivalent
context is in its infancy and virtually no work is available to guide principle is “less pressurized signal phases receive no green,” where
developments in this area. the pressure is a measure that would take several forms, such as max-
This report considers the issues that arise in seeking to optimize imum (flow to saturation flow ratio) or the negative value of the ratio
network performance through simultaneous determination of optimal of the flow to the change in delay with respect to green time propor-
routes and optimal signal settings. A mathematical formulation of the tion. They also developed a solution algorithm and established its
problem for given time-dependent origin-destination demand (trip convergence for desired solution.
Allsop (8) appears to be the first researcher to suggest, in the pub-
lished scholarly literature, that traffic engineers should take explicit
A. S. Abdelfatah, Department of Civil Engineering, and H. S. Mahmassani,
Department of Civil Engineering and Department of Management Science account of the long-run effect of signal settings on the traffic
and Information Systems, ECJ 6.2, University of Texas, Austin, TX 78712. assignment. He proposed an iterative solution procedure for the UE
186 Paper No. 98-1349 TRANSPORTATION RESEARCH RECORD 1645

static assignment problem in a pretimed signal-controlled network. PROBLEM FORMULATION

For simplicity, all intersections in the network are assumed to be
signalized and no special treatment is provided for the opposed In the formulation, a central controller has complete a priori infor-
turning movements at signalized intersections. Although this spe- mation on time-varying O-D desires for the entire assignment dura-
cific approach has been shown to be nonconvergent, it laid the tion. The controller assigns a path to the desired destination to each
groundwork for several subsequent contributions. Charlesworth (9) vehicle as it enters the network, to optimize some indicator of the sys-
obtained mutually consistent traffic assignment and signal settings tem performance subject to the applicable constraints. In addition, the
through an iterative procedure in which the well-known TRANSYT controller decides on the signal settings for each signalized intersec-
software is used to optimize the signal settings. Dickson (10) gave tion to optimize the system performance. Because the O-D desires are
an example in which these procedures may lead to an increase in time-varying, the routing information and signal settings should be
the total network travel time. He attributed this increase in the time dependent. The time period of interest is discretized into small
travel time to the fact that these approaches assumed no simultane- intervals (assignment intervals) for which the path assignments and
ous changes in equilibrium flows due to signal setting adjustments. signal settings are determined.
Sheffi and Powell (11) presented a mathematical programming for- The formulation incorporates dynamic link-path incidence vari-
mulation for the problem and an algorithm that could be used for ables that relate link flows to path flows. The fundamental diffi-
very small networks (several dozen links and nodes), although the culty solving the problem is that the dynamic incidence variables
problem formulation may be nonconvex. They also introduced two and optimized green time proportions are themselves functions of
the assignment. Ghali and Smith (15) showed that the dynamic for-
heuristic approaches for possible application to large networks,
mulation for the problem will lead to nonconvex link performance
although no implementation or application results have been reported
functions. Therefore, the resulting formulation, which involves
with these approaches.
nonlinearities in the objective function as well as in the constraints,
Another version of the iterative procedures was presented by
yields generally undesirable mathematical properties that preclude
Gartner and Al-Malik (12), in which they sought a simultaneous
the guarantee of global optimality. In the static assignment case,
solution of the assignment and the control. They considered a net-
flows assigned to a path exist on all links along that path at the same
work with two-phase pretimed signal control at all intersections and
time, leading to constant link-path incidence and green time pro-
link performance functions that explicitly relate the link travel time
portions. Such assumption in the dynamic case clearly is flawed,
to the green time allocated to that approach. because vehicles starting along a path at a given time are not simul-
taneously present on all links of the path. The incidence matrix and
green time proportions change dynamically, considerably compli-
Time-Dependent Assignment cating vehicle conservation constraints and travel time calculations
(including the delay at signalized intersections). The formulation is
As noted, research addressing signal timing in conjunction with an extension of Mahmassani and Peeta’s (18) work for the SO
time-dependent network flow assignment is very limited com- dynamic assignment problem. The formulation details are as
pared with that reported for the static assignment problem. Smith follows:
and Ghali (13,14) presented a mathematical formulation that
extends their work on the static case and suggested an algorithm Given:
that uses an incremental assignment logic (i.e., loading a small
proportion of the demand, adjusting the signals, then repeating the rijτ , ∀ i, j and τ = 1, K, T
process until the total demand is loaded). They also proposed two
algorithms for the dynamic assignment problem when the signal Objective function:
settings are fixed (ordinary dynamic assignment problem), which
are not of direct relevance to the scope of this review. Ghali and
Smith (15) implemented an iterative procedure using CONTRAM
Min ∑ ∑ ∑ ∑ rijkτ 䡠 Tijkτ ( R, S)
τ i j k
and showed the convergence pattern for it.
Smith and Van Vuren (16) outlined the steps for an iterative algo- Subject to:
rithm that could be considered an extension of Allsop’s algorithm
to the time-varying case. Pn

Gartner and Stamatiadis (17) presented a general conceptual ∑ λtln = 1 − Ltn Cnt ∀n, t (1)
l =1
framework for the implementation of a combined solution for
λtln (Cnt − Ltn ) ≥ glnmin
dynamic traffic assignment and signal control, but they did not
∀n, t (2 )
report implementation of a specific algorithmic procedure.
In this report, a central controller who is seeking time-varying route
guidance and signal control to optimize the system performance is rijτ = ∑ rijkτ , ∀ i, j, τ (3)
k
considered. Mathematical formulation and solution algorithm are pre-
sented for the combined problem of the dynamic assignment (SO
assignment) and responsive signal control. A general traffic network ∑ d tb = ∑ mtc + Int − Ont , ∀ t, n, b ∈B(n), c ∈C(n) ( 4)
b c
is considered that has different control types (i.e., some intersections
may have stop or yield signs, and some others may have signal con- x ta = x t −1a + d t −1a + m t −1a , ∀ t, a (5)
trol with any phasing pattern). Finally, we implement the solution
algorithm to a representation of an actual network and show the poten-
tial for improving overall system performance.
x ta = ∑ ∑ ∑ ∑ (rijkτ 䡠 δ ijkτta ), ∀ t, a (6)
k τ i j
Abdelfatah and Mahmassani Paper No. 98-1349 187

Tijkτ = ∑ ∑ [δ ijkτta 䡠 ∆ ], ∀ i, j, k , τ ( 7) The objective function states that the total travel time is the aggre-
t a gate of the product of the number of vehicles assigned to a particular
path (for a given O-D pair at a particular time) and the correspond-
τta
δ ijk [
= F (rijkτ , S ), ∀ i, j, k , t ] ∀ i, j, k , τ, t, a (8) ing path travel time, which is a function of the set of path assign-
ments R and signal settings S in the network. This assumption is
d ta = ∑ ∑ ∑ ∑ dijkτta , ∀ t, a ( 9) realistic when assignment intervals are reasonably small. The non-
linearity of the objective function arises because the travel time on
k τ i j
the path itself is a complicated nonexplicit function of the number
m ta = ∑ ∑ ∑ ∑ mijkτta , ∀ t, a (10) of vehicles assigned to various paths in the network, via the inci-
k τ i j dence matrix, and the time-varying green time proportions assigned
to each approach.
τ ≤t (11) Constraint 1 states that the green times and lost times allocated to
all approaches at any signalized intersection during a specified time
τta
δ ijk = 0 or 1, ∀ i, j, k , τ, t, a (12) interval should sum up to the cycle length for that intersection. Con-
straint 2 states that the green time for any phase should not be less
All variables ≥ 0 (13) than a minimum value. Constraint 3 is a definitional constraint stat-
ing that the assigned flows to various paths should satisfy the given
where demand. Constraint 4 represents the conservation of vehicles at inter-
mediate nodes. Constraint 5 represents the conservation of vehicles
r τij = number of vehicles that wish to depart from i to j in
on a link.
period t;
Constraints 6, 7, and 8 represent the time-dependent link-path inci-
r τijk = number of vehicles that wish to depart from i to j in period
dence relationships that fundamentally characterize the dynamic
t assigned to path k;
assignment problem. Constraint 6 represents the dynamic relation-
T τijk = experienced path travel time for vehicles going from i to j
ship between link and path assignments; Constraint 7 illustrates the
that are assigned to path k at time t;
R = {rijkτ }, ∀ i, j, k and τ ;
calculation of path travel time, and Constraint 8 states that the
dynamic link-path incidence variables are function of the assignment
S = {λtln }, ∀ t, l, n; (path assignments) and signal settings.
Pn = number of phases at intersection n; Constraints 9 and 10 are definitional constraints for the number
λ lnt = green time proportion for phase l at intersection n during of vehicles entering and exiting links in the various time intervals.
time interval t; Constraint 11 defines temporal correctness. Constraint 12 restricts
L nt = lost time at intersection n during time interval t; dynamic incidence variables to take values 1 or 0. Constraint 13
C nt = cycle length at intersection n during time interval t; represents the nonnegativity requirements.
xta = total number of vehicles on link a at the beginning of
period t;
dta = total number of vehicles that enter link a in period t; SOLUTION ALGORITHM
mta = total number of vehicles that exit link a in period t;
I nt = number of vehicles generated at node n in period t; Nonlinearities are present in the objective function and the con-
O nt = number of vehicles exiting the network through node n in straints for the problem at hand. Furthermore, the experienced path
period t; travel times are the net result of the complex nonlinear spatial and
C(n) = set of links incident to node n; temporal interactions among vehicles in the system over a period of
B(n) = set of links incident from node n; time, virtually precluding the ability to evaluate the objective func-
δ τta
ijk = time-dependent link-path incidence indicator, equal to 1 tion analytically. In addition, analytical evaluation would require
if vehicles going from i to j assigned to path k at time τ are correct representation of the various dynamic traffic flow phenom-
on link a in period t, that is, ena (queue formation and discharge, congestion build-up, and dis-
[δijkτta = 1, if rijkτ is on arc a during period t, sipation), which has yet to be derived in analytical mathematical
= 0, if arc a does not belong to path k , formulation. It also is important to note that for specifying the
= 0, if τ > t, dynamic link-path incidence variables explicitly in any solution pro-
= 0, if rijkτ is not on arc a during period t ]; cedure would give rise to a problem of huge dimension. Therefore,
standard nonlinear programming (NLP) solvers cannot be used
∆ = length of a time interval;
directly to solve the above formulation. Simulation is a suggested
ijk = number of vehicles going from i to j assigned to path k in
mτta
candidate for evaluating the objective function.
period τ, which exit link a in period t; and
The solution algorithm proposed for the formulation and tested in
ijk = number of vehicles going from i to j assigned to path k in
d τta
this study builds on the work of Mahmassani and Peeta (18) for the
period τ, which enter arc a in period t.
time-dependent assignment problem. It is an iterative simulation-
The decision variables in the above formulation consist of the set R based algorithm, which uses the DYNASMART (Dynamic Network
of path assignments r τijk (number of vehicles that wish to depart from Assignment Simulation Model for Advanced Road Telematics) sim-
i to j in period t assigned to path k) and the set S of signal settings λ lnt ulation model as an evaluation tool for the network status. For a
(green time proportion for phase l at intersection n during time inter- given set of time-dependent path assignments and signal settings, the
val t), noting that both sets of decision variables are defined for each model evaluates the objective function and provides time-dependent
assignment interval. information about link performance, including numerical approxi-
188 Paper No. 98-1349 TRANSPORTATION RESEARCH RECORD 1645

mations of the derivatives, which are then used to define a search τ,i
auxiliary path assignments (y odk ) (the number of vehicles that depart
direction along which the next iteration of the path assignments and at time t and are assigned to auxiliary path k between origin o and
signal control are determined. The structure of the algorithm is destination d for iteration i).
depicted in Figure 1. In addition, a streamlined variant of this algo- • Step 3b. Use the method of successive averages to update the
rithm was considered, labeled Algorithm 2 hereafter; the original path assignment variables as follows:
version is referred to as Algorithm 1, in which an attempt is made to
reduce the computational time. The structure of this variant (Algo-
rithm 2) is depicted in Figure 2. The main difference is that Step 2
τ , i +1
rodk (
τ,i
= rodk 1−
1
i +1 ) τ,i
+ yodk
1
( )
i +1
of the original algorithm is skipped in Algorithm 2; one does not re-
simulate after obtaining the signal settings and instead the time- Details of Steps 3a and 3b are provided by Mahmassani and Peeta (18).
dependent signal settings are updated in the same manner as the path • Step 4. Check for convergence. Are
assignment (by using the method of successive averages). The steps
τ ,i +1 τ ,i
of Algorithm 1 are as follows: rodk − rodk ≤ ⑀1 and λtln,i +1 − λtln,i ≤ ⑀ 2

• Preparation step. Starting with any signal settings (a fixed cycle If yes, stop. Otherwise go to Step 5.
and equal splits for all signalized intersections), simulate using • Step 5. Resimulate by using DYNASMART with the new Ri+1
DYNASMART, which has an assignment component. The resulting and Si+1. Set i = i + 1, and go to Step 1.
flows from DYNASMART are used as the initial time-dependent
path assignment R0. Calculate the optimal signal settings for R0; the
result will be the initial signal settings S0. Signal Optimization
• Step 0, Initialization. Starting with feasible initial signal set-
tings S0 and time-dependent path assignments R0, simulate using A detailed description of Step 1 in the above algorithm is provided
DYNASMART. Set i = 0. For iteration i do here. Two methods for signal optimization, reflecting different lev-
• Step 1, Signal Optimization. Calculate the time-dependent sig- els of computational effort, are suggested for this in Step 1 of the
nal settings Si+1 (which is λ lnt and C nt for all signalized intersections solution algorithm. First, the well-known Webster’s formulas (19)
during all time intervals). may be used to calculate the optimal cycle length and green time
• Step 2. Resimulate the network by using DYNASMART with splits. Second, a simulation-based implementation of Hooke and
Ri and Si+1 (obtained in Step 1). Evaluate the time-dependent link Jeeves’ method (20) is proposed.
travel times T ta,i (travel time on link a at time t for iteration i) and
the set of path assignments r τiodk (the number of vehicles that depart
at time t on path k between origin o and destination d for iteration i). Webster’s Formulas
• Step 3a. By using the output of Step 2 (T ta,i), calculate the time-
dependent link marginal travel times and the set of least cost paths In this case, Webster’s formulas are used to calculate the optimal cycle
(labeled as the set of auxiliary paths). Perform all-or-nothing time- length and green time splits for each intersection during each time
dependent assignment of O-D desires to auxiliary paths to obtain the interval. The steps for the signal setting calculation are as follows:

FIGURE 1 Solution algorithm (Algorithm 1).

Abdelfatah and Mahmassani Paper No. 98-1349 189

FIGURE 2 Variation of solution algorithm (Algorithm 2).

For each time interval t, for each signalized intersection n, t ( Dnt (Gnm −1 ), Ri )
• Calculate the traffic volume for each turning movement on
each approach. by resimulating the local area for intersection n using Gnm −1 (set of
• Calculate the volume-to-capacity ratio per lane for each approach green times at intersection n for iteration m − 1), Ri.
and find the critical lane for each phase k [p(k)]. The volume is defined • Step 2. Find the new green times ( gnlt, m ) (Hooke and Jeeves’
as the number of vehicles that exit the link during the current time method).
interval (mta). • Step 2.1. Initialize
• Calculate the cycle length C = (1.5 × L + 5)/[1 − Σp(k)], where Z = Gnm = Gnm −1 , d > 0, 0 < a < 1.0 and B = 1
L is the lost time.
• Calculate the green time for each phase j:

g( j ) = (C − L ) × p( j ) ∑ p(k )

Simulation Method

The simulation method uses a modified version of the DYNASMART

simulation model to optimize the signal settings. In this method, a
local area is defined for each intersection that makes up other neigh-
boring intersections that affect or are affected by the delay at the inter-
section of interest. Similar to the local area definition for decentralized
route guidance proposed by Hawas and Mahmassani (21), the depth
K, which is the number of upstream or downstream intersections, is
defined to delineate the local area. Figure 3 shows an example of the
local area for K = 1.
The steps for signal optimization are as follows:

• Step 0, Initialization. G0 = set of green times in Si (Si and Ri are

obtained from iteration i in the main algorithm). Define the local
area for each signalized intersection. Set m = 1.
• For iteration m, for n ∈ Ns (set of signalized intersections), for
each time interval t do
• Step 1. Calculate the total delay at intersection n during time
interval FIGURE 3 Local area for intersection (depth K = 1).
190 Paper No. 98-1349 TRANSPORTATION RESEARCH RECORD 1645

• Step 2.2, Exploratory Search. For each phase j do

) )
(a) Z = Z + Bδej, simulate, and check if Dnt ( Z , Ri ) < Dnt ( Z , Ri ). If yes,
)
update Z = Z and check next phase. If no, go to Step 2.2b.
(b) If B = 1, set B = −1 and go to Step 2.2a; otherwise, check next
phase.
• Step 2.3, Pattern Search.
(a) If
) )
Dnt ( Z ) < Dnt (Gnm −1 ), put Gnm = Z and Z = Gnm −1 + α(Gnm − Gnm −1 )
and go to Step 2.2. Otherwise, go to Step 2.3b.

(b) If d is sufficiently small, stop with G nm optimal. Otherwise, put

δ = 0.5 δ, j = 1, Z = Gnm −1 , and go to Step 2.2.
• Step 3. Check for convergence: If

gnlt, m ≅ gnlt, m −1

for all l, if all signalized intersections are optimized, stop and return
to the main algorithm. Otherwise, start for a new signalized inter-
section. Else, G nm = current set of green times, set m = m + 1 and go
to Step 1.

NUMERICAL ILLUSTRATION

In this section, the solution algorithm and its streamlined variant are
implemented on an actual network under different congestion levels.
The main objective of these experiments is to illustrate the potential
for improving network performance when solving for the combined
SO assignment and signal control problem beyond that obtained
through SO assignment only.

Experimental Design

Test Network

Figure 4 depicts the test network configuration. The network is part

of the Dallas–Fort Worth street network and consists of 178 nodes
(including 61 signalized intersections) and 441 directed links. A free-
way facility (I-35) runs through the middle of the network, giving it
features of both freeway corridor as well as urban street network.

Congestion Levels FIGURE 4 Fort Worth test network configuration.

The experiments are conducted under three congestion levels:

1. Loading Factor 1, 6,160 vehicles, themselves subject to upper- and lower-bound constraints. Under the
2. Loading Factor 2, 12,248 vehicles, and second case, a fixed cycle length is used for all intersections in the
3. Loading Factor 3, 16,717 vehicles. network, a restriction that commonly is imposed in many of the pre-
vious studies of this topic. Four different cycle lengths are tested: 80,
The loading factor denotes the ratio of the number of vehicles loaded 100,120, and 140 sec.
into the network over the duration of interest relative to the base-
level loading.
Time Intervals

Cycle Length The total loading time and loading pattern in these experiments is
35 min. The loading pattern, which exhibits a peaking characteris-
Two cases are considered for the important factor of cycle length. tic, is shown in Figure 5. To evaluate a particular assignment and
Under the first case, the cycle length is unconstrained and is obtained signal setting solution, the network is simulated for as long as is
simply by adding the green times provided by the solution algorithm, required for all vehicles to reach their destinations. Although the
Abdelfatah and Mahmassani Paper No. 98-1349 191

loading levels and the number of iterations was 4, 8, and 11 for load-
ing levels 1, 2, and 3, respectively. Table 1 shows the average travel
time, over all vehicles in the system, expressed as a percentage of
the SO traffic assignment only solution at convergence, for the
unconstrained cycle length case. Figure 6 illustrates the convergence
patterns for the SO assignment and the two joint signal setting-
assignment algorithms under the unconstrained cycle length case,
for loading level 3. Table 2 shows the evolution of the cycle length
through the iterations of the algorithm for each time interval at one
of the intersections for Algorithm 1 and Algorithm 2. Table 3 shows
the results for the constrained cycle length cases, for cycle lengths
of 80, 100, 120, and 140 sec.
The results presented in Table 1 indicate that both algorithms
resulted in some improvement (over the SO assignment solution
only) in the average travel time in the network for all cases, except
when the streamlined variant (Algorithm 2) was applied to the very
FIGURE 5 Demand loading pattern. high congestion case (loading factor 3). The increase in the travel
time for this case may be because Webster’s formulas do not perform
path assignments are obtained for the loading time only, the signal well in the oversaturation conditions that develop under this loading
settings are obtained for the whole simulated interval (i.e., until all level, and because Algorithm 2 calculates the signal setting and the
vehicles reach their destinations). assignment in parallel (i.e., there is not much information about the
The loading time is the same as the assignment period and is sub- updated signals when the path assignments are updated). Table 2
divided into 5-min assignment intervals (the time interval for which illustrates the difference between the two algorithms with regard to
the path assignment and signal setting decision variables are defined). the convergence of the cycle length. The results in Table 3 exhibits
the same general pattern as for Table 1, namely the success of Algo-
rithm 1 in obtaining better solutions than SO assignment for fixed
Experimental Results signal settings, the decreasing opportunity for improvement as the
congestion level increases in the network, and the unreliable perfor-
The experiments were performed for Algorithms 1 and 2 by using mance of the streamlined variant algorithm (Algorithm 2), which
Webster’s formulas to calculate the signal settings (green splits or should be used only with considerable caution. Algorithm 1 is much
cycle length). The algorithms have achieved convergence under all preferred in this regard.

TABLE 1 Average Travel Time (as Percentage of SO Assignment Only at Convergence) for
Unconstrained Cycle Length

FIGURE 6 Convergence patterns for loading level 3 (unconstrained

cycle).
192 Paper No. 98-1349 TRANSPORTATION RESEARCH RECORD 1645

TABLE 2 Cycle Lengths at Intersection 10 for Algorithm 1 and Algorithm 2

CONCLUSIONS The principal substantive conclusions for the application to the

Fort Worth network are that joint optimization of routing and signal
This study presented a formulation and heuristic simulation-based setting can lead to improved overall network performance, relative
procedure to obtain time-varying assignment of O-D vehicles to to either routing or signal setting only, and the potential for additional
paths jointly with time-varying signal settings that seek to opti- improvements through signalization (beyond that achieved through
mize the total travel time in the network. The algorithm and a optimal routing for given initial settings) appears to decrease with
streamlined variant were implemented successfully on an actual higher congestion levels.
network from the Dallas–Fort Worth area. The main version of Furthermore, the application clearly reflected the fact that the
the algorithm performed successfully in all cases, but the stream- objective function is not globally convex and that the algorithms are
lined version proved unreliable, especially under high congestion likely to discover local optima, which nonetheless may offer better
levels. network performance than the base solution.

TABLE 3 Average Travel Time (as Percentage of SO Assignment

Only at Convergence) for Different Cycle Lengths
 Abdelfatah and Mahmassani Paper No. 98-1349 193

REFERENCES 11. Sheffi, Y., and W. B. Powell. Optimal Signal Setting over Transporta-
tion Networks. Journal of Transportation Engineering, Vol. 109, 1983,
pp. 824–839.
1. Smith, M. J. The Existence, Uniqueness and Stability of Traffic Equilib-
12. Gartner, N. H., and M. Al-Malik. Combined Model for Signal Control
ria. Transportation Research 13B, 1979, pp. 295–304. and Route Choice in Urban Networks. In Transportation Research
2. Smith, M. J. The Existence of an Equilibrium Solution to the Traffic Record 1554, TRB, National Research Council, Washington, D.C.,
Assignment Problem When There Are Junction Interactions. Trans- 1996, pp. 27–35.
portation Research 15B, 1981, pp. 443–451. 13. Smith, M. J., and M. O. Ghali. The Dynamics of Traffic Assignment and
3. Smith, M. J. A Local Traffic Control Policy which Maximizes the Over- Traffic Control: A Theoretical Study. Transportation Research 24B,
all Travel Capacity of an Urban Road Network. Traffic Engineering and 1990, pp. 409–422.
Control, 1980, pp. 298–302. 14. Smith, M. J., and M. O. Ghali. Dynamic Traffic Assignment and Dynamic
4. Smith, M. J. Properties of a Traffic Control Policy which Ensure the Traffic Control. Presented at 11th International Symposium on Traffic and
Existence of a Traffic Equilibrium Consistent with the Policy. Trans- Transportation Theory. Yokohama, Japan, 1990.
portation Research 15B, 1981, pp. 453–462. 15. Ghali, M. O., and M. J. Smith. Traffic Assignment, Traffic Control and
5. Van Vuren, T., M. J. Smith, and D. V. Vliet. Interaction Between Signal- Road Pricing. Proc., 12th International Symposium on Theory of Traffic
Setting Optimization and Reassignment: Background and Preliminary Flow and Transportation, California, 1993.
Results. In Transportation Research Record 1142, TRB, National 16. Smith, M. J., and T. Van Vuren. Traffic Equilibrium with Responsive Traf-
Research Council, Washington, D.C., 1988, pp. 16–21. fic Control. Transportation Science, Vol. 27, No. 2, 1993, pp. 118–132.
6. Smith, M. J., T. Van Vuren, B. G. Heydecher, and D. V. Vliet. The 17. Gartner, N. H., and C. Stamatiadis. Integration of Dynamic Traffic Assign-
Interaction Between Signal Control Policies and Route Choice. Pre- ment with Real-Time Traffic Adaptive Control. Presented at 76th Annual
sented at 10th International Symposium on Transportation and Traffic Meeting of the Transportation Research Board, Washington D.C., 1997.
Theory. Elsevier, New York, 1987. 18. Mahmassani, H. S., and S. Peeta. System Optimal Dynamic Assignment
7. Wardrop, J. G. Some Theoretical Aspects of Road Traffic Research. for Electronic Route Guidance in Congested Traffic Networks. Proc.,
Proc., Institute of Civil Engineers, Part II, 1952, pp. 325–378. 2nd International Capri Seminar on Urban Traffic Networks, 1992.
8. Allsop, R. E. Some Possibilities for Using Traffic Control to Influence 19. Webster, F. V. Traffic Signal Settings. Road Research Technical Paper
Trip Distribution and Route Choice. Presented at 6th International 39, HMSO, London, 1958.
Symposium on Transportation and Traffic Theory, Sydney, Australia, 20. Hooke, R., and T. A. Jeeves. Direct Search Solution of Numerical and
1974. Statistical Problems. J. Ass. Comp., 1961, pp. 212–229.
9. Charlesworth, J. A. The Calculation of Mutually Consistent Signal- 21. Hawas, Y. E., and H. S. Mahmassani. Comparative Analysis of Robust-
Settings and Traffic Assignment for a Signal Controlled Network. Pre- ness of Centralized and Distributed Network Route Control Systems in
sented at 7th International Symposium on Transportation and Traffic Incident Situations. In Transportation Research Record 1537, TRB,
Theory, Kyoto, Japan, 1977. National Research Council, Washington, D.C., 1996, pp. 83–90.
10. Dickson, T. J. A Note on Traffic Assignment and Signal Timing in a
Signal-Controlled Road Network. Transportation Research 15B, 1981, Publication of this paper sponsored by Committee on Transportation Network
pp. 267–271. Modeling.

View publication stats