Professional Documents
Culture Documents
ASCEJTE2012StochasticOptimization Ilsoo
ASCEJTE2012StochasticOptimization Ilsoo
net/publication/266310364
CITATIONS READS
53 448
All content following this page was uploaded by B. Brian Park on 09 June 2016.
Abstract: Existing state-of-the-practice traffic signal timing-optimization programs rely on macroscopic and deterministic models to
Downloaded from ascelibrary.org by UNIVERSITY OF VIRGINIA on 05/30/13. Copyright ASCE. For personal use only; all rights reserved.
represent traffic flow, including coordinated actuated traffic signal systems. One distinct shortcoming of such an approach is its inability
to account for the stochastic nature of traffic, such as the variability in traffic demand, driver behavior, vehicular interarrival times, vehicle
mix, and so forth. In addition, the existing traffic signal timing-optimization programs for coordinated actuated traffic signal systems still
focus on four basic traffic signal timing parameters (i.e., cycle length, green times or force-off points, offsets, and phase sequences). Studies
have shown that actuated signal settings such as minimum green time, vehicle extension, and recall mode are also important parameters in
traffic signal operations. This study presents the development of a stochastic-optimization method for coordinated actuated traffic signal
systems. The proposed method accounts for stochastic variability by using a well-calibrated microscopic simulation model, CORSIM, instead
of a macroscopic and deterministic model, and it simultaneously optimizes actuated signal settings and the four traffic signal timing param-
eters by adopting a genetic algorithm with special decoding schemes. The proposed method was applied to a real-world arterial network in
Charlottesville, Virginia. The performance of the proposed method was compared with that of an existing traffic signal timing-optimization
program, Synchro, using a well-calibrated microscopic simulation model, VISSIM. The results indicated that the proposed method outper-
forms the existing timing plan and the Synchro-optimized traffic signal timing for the tested arterial network. DOI: 10.1061/(ASCE)TE
.1943-5436.0000384. © 2012 American Society of Civil Engineers.
CE Database subject headings: Optimization; Traffic signals; Simulation; Calibration; Algorithms; Stochastic models.
Author keywords: Stochastic optimization; Traffic signal control; Microscopic traffic simulation; Calibration; Genetic algorithm.
Stochastic optimization has been applied extensively in transporta- This section describes a stochastic-optimization problem formu-
tion applications. Most stochastic-optimization approaches have lation for traffic signal control settings, and then presents the meth-
dealt with transportation network assignments. Transportation odology used in the development of a stochastic-optimization
and location selection problems were formulated in a stochastic- method.
optimization problem by LeBlanc (1977) and Franca and Luna
Problem Formulation
(1982). The former formulated a stochastic problem that
determines plant locations and shipment by minimizing expected A formulation of the stochastic-optimization problem is presented
holding and storage costs, and solved it using a heuristic algorithm, for traffic signal control settings of a coordinated actuated traffic
whereas the latter formulated it as a mixed-integer nonlinear signal system, including cycle length, green times, offsets, phase
program and solved it using the generalized Benders decomposi- sequences, minimum green time, vehicle extension time, and recall
tion method. Since Daganzo and Sheffi (1977) first formulated a mode. As shown in Eq. (1), a system delay (i.e., queue time or con-
stochastic user equilibrium model, many studies have been trol delay), dðs; rÞ, associated with a set of coordinated actuated
conducted. Details can be found in Sheffi (1985). However, there traffic signal control settings, s, and a set of random seeds, r,
have been relatively few research efforts in traffic signal timing can be determined by running a microscopic simulation program.
optimization using a relative stochastic-optimization approach. The objective function, an expected system delay on the basis of the
The GA has been widely applied in the area of transportation selection of traffic signal control settings and random number
engineering, including network design (Xiong and Schneider seeds, and its constraints are shown as follows:
1992, 1995), transit scheduling (Chakroborty et al. 1995), shortest
path finding (Soehodho 1998), and bilevel programming for Find s ¼ fC; θi ; ρi ; gij ; mgij ; extij ; rcij g min E½dðs; rÞ ð1Þ
network flow (Yin 2000). A few of the studies on traffic signal s
X
4 Development of Stochastic-Optimization Method
min C ≥ ðmgij þ CTij Þ ð3Þ
j¼1
Because of the stochastic nature associated with the objective
function, which can be obtained reliably by implementing a well-
X
8 calibrated microscopic simulation model, conventional optimiza-
min C ≥ ðmgij þ CTij Þ ð4Þ tion methods on the basis of a gradient were not considered. Thus,
j¼5 a GA was adopted on the basis of a study conducted by Park and
Yun (2006) in which three commonly used optimization methods,
X
4 X
4 X
8 X
8
including simulated annealing, a GA, and a commercial optimiza-
gij þ CTij þ gij þ CTij ¼ 2C ð5Þ tion program OptQuest Engine, were compared extensively in traf-
Downloaded from ascelibrary.org by UNIVERSITY OF VIRGINIA on 05/30/13. Copyright ASCE. For personal use only; all rights reserved.
where f C = fraction value, 0 ≤ f C ≤ 1. where f extj = fraction value, 0 ≤ f extj ≤ 1 and j = 1,3,4,5,7, and 8.
The vehicle extension time can be defined with narrower inter- simulation run time compared with other microscopic simulation
vals such as decimal seconds with any reasonable minimum and models, including VISSIM and PARAMICS (Park and Yun 2003).
maximum values. The selection of these values and intervals In addition, because the proposed method uses CORSIM during
should be determined by the control logics used in the microscopic optimization, the evaluation should be conducted with an indepen-
simulation models and/or the user’s decision. For coordinated ac- dent evaluator that could be considered as a surrogate to field
tuated signal control, force-off points and permissive periods can implementation. VISSIM was selected as an independent evaluator
be used instead of green times or splits in real-world controllers for the signal control settings optimized by the proposed stochastic-
and microscopic simulation models, including CORSIM and VIS- optimization method and that of the state-of-the-practice signal
SIM. The force-off points and permissive periods are determined timing-optimization program, Synchro. VISSIM can also ad-
from the green times calculated from Eq. (18). In this paper, the equately model various traffic signal control systems using external
force-off point of each phase is the end of its split duration or the programs, such as the vehicle actuated programming (VAP), the
end of the maximum green time in which the phase is allowed to NEMA emulator, or the ring barrier controller (RBC) emulator.
remain green (CORSIM Version 5.0). For example, CORSIM as- Both CORSIM and VISSIM simulation models were well cali-
sumes that the force-off points of coordinated phases 2 and 6 are brated using the field traffic data. This is because the traffic patterns
set to zero in which the local controller time is zero (this is com- represented in the microscopic simulation models are likely to be
monly known as the yield point or local zero). Thus, force-off critical in assessing the quality of both optimization and evaluation.
points should be calculated for every noncoordinated actuated In addition, the same effort was carried out to calibrate Synchro.
phase. For example, the force-off point for phase 3 is at the Even though Synchro is a macroscopic and deterministic model,
end of phase 3’s green time, counted from the local zero. The local it requires a series of input data and calibration. The process of data
zero is the local reference point from which all other coordination- collection and calibration efforts is presented in the following
related settings for this particular controller are measured, and is in sections.
turn dependent on the type of controllers and its operational strat-
egy (CORSIM Version 5.0). Permissive periods provide specific Site Selection and Data Collection
intervals when the controller can respond to demand calls for a An arterial network consisting of four signalized intersections on
certain phase. The first permissive period starts at the end of Emmet Street, which is Route 29 in Charlottesville, Virginia in the
the coordinated phase(s) green (i.e., the yield point or local zero). United States, was selected as a test site because of the convenience
This is the point at which the permissive window opens to respond of data collection and its familiarity to the authors. The test site is
to calls for service. In the example shown in Fig. 2, the end of located at Emmet Street between Hydraulic Road and Barracks
permissive period 1 is mathematically determined by subtracting Road. Along Emmet Street, there are several business establish-
the minimum green time of phase 3 and the intergreen time of ments and retail stores, such as Wachovia Bank, Best Buy, and
phase 2 (yellow plus red clearance time) from the force-off point Holiday Inn. Emmet Street intersects with Route 250, which serves
of phase 3. as major access roads to other areas, including Washington, DC and
the Richmond area. The test site includes four signalized intersec-
tions. Among them, two intersections are signalized interchanges,
Case Study which operate as overlaps of the other intersections.
Data-collection efforts were designed to provide inputs for
The proposed stochastic-optimization method is demonstrated Synchro, CORSIM, and VISSIM, and to performance measures
through a real-world network. The performances are compared for the calibration of these programs. While signal timings and link
with those of the state-of-the practice signal timing-optimization geometry attributes were obtained from the City of Charlottesville
program, Synchro. and the Virginia Department of Transportation (VDOT), other data,
including traffic counts, travel times, and travel speed, were
Selection of Microscopic Simulation Models: CORSIM collected directly from the test site on a weekday from 2:45 to
and VISSIM 4:15 p.m.
CORSIM was selected in this study as a microscopic simulation
Synchro, CORSIM, and VISSIM Network Building
model for optimization because of its long history of development
and support from the Federal Highway Administration (FHwA), its As noted, the CORSIM network was prepared for the stochastic
ability to model common U.S. traffic signal controllers, and its fast optimization, whereas the VISSIM network was used for the
Fig. 3. Microscopic simulation networks: (a) real-world CORSIM network; (b) real-world VISSIM network
Fig. 4. Calibration result of CORSIM and VISSIM networks: (a) histogram of travel times of section 1; (b) histogram of travel times of section 2
• Recall modes of four noncoordinated actuated phases for each • Uniform crossover with 0.5 probability and simple mutation
of the two intersections. with 0.03 probability, and
The GA-based optimizer and the optimization-simulation inter- • Stop at a predefined maximum number of generations.
face in the proposed stochastic-optimization method were devel- The objective function used in the proposed stochastic optimi-
oped using the MATLAB program. In the GA-based optimizer, zation is system delay, as shown in Eq. (1). Given that CORSIM
the following settings were used: provides delay measures of those vehicles that already left the
• 40 populations and 25 generations, subject link, the control delay from CORSIM does not account
• Normal geometric selection with elitist method, for vehicles remaining on the links. As such, the total queue time,
which considers both discharged and remaining vehicles, was used Given that microscopic simulation models at times show quite a
as an objective-function value. During the optimization of traffic significant variability in their performance measures, it is crucial to
signal control settings, certain sets of parameter settings could account for such variability during optimization. To accommodate
result in extreme congestion. When this happens, total queue time this variability, multiple objectives (i.e., a combination of total
can adequately reflect the effect of the congestion. queue time and its variability) or expectation values obtained from
GA tends to give a family of solutions instead of one best solution In Table 3, Synchro shows a 12.3% reduction in average
in complex optimization problems like traffic signal timing optimi- delay time over the base case, whereas the proposed stochastic-
zation. The convergence properties of all three trials are presented optimization method produces approximately 28.5% lower ave-
graphically in Fig. 5, which illustrates the best value of the total rage delay time than the base case. In the comparison between
queue time in vehicle-minutes in every generation. The average Synchro and the stochastic-optimization method, the stochastic-
computation time for each trial was approximately 2 h. The optimization method shows 18.4% lower average delay time than
computational time was measured on a personal computer with Synchro. Fig. 6 illustrates distributions of average delay times of
a Xeon CPU (3.60 GHz) and 1 GB of random-access memory the five cases obtained from the same 50 random-seeded VISSIM
(RAM), which is not an up-to-date high-performance computer.
simulation runs. As shown in the figure, the traffic signal control
Much faster computational time can be expected when the recent
settings from the stochastic-optimization method are able to reduce
high-performance computer with multiple CPUs is used in the
the variance of delay and the degree of delay.
stochastic-optimization. For example, when the stochastic-
In addition, the traffic signal control settings optimized by the
optimization method was executed in a dual-processor computer
stochastic-optimization method showed a significant reduction in
with two Xeon CPUs (3.60 GHz) and 1 GB of RAM, a 19%
the travel times of sections 1 and 2. Table 3 and Fig. 6 show that
reduction of computation time was achieved.
all three trials from the stochastic-optimization method consistently
and significantly outperform Synchro for all MOEs considered.
Evaluation of Traffic Signal Control Settings These results indicate that the stochastic-optimization method
shows its outstanding capability in optimization, as it can account
Table 2 shows the optimized traffic signal control parameters ob- for the stochastic nature of traffic using a well-calibrated micro-
tained by Synchro and the stochastic-optimization method in trial 1. scopic simulation model. It is apparent from Table 3 that the sto-
As shown in Table 2, green splits including transition interval chastic optimization successfully optimized the traffic signal
showed up to 7-s differences between them. In addition, offsets control settings, including the four basic traffic signal timing
showed 4 and 9 s difference. The cycle length was fixed as 107 s. parameters and the actuated signal settings for a coordinated actu-
The performance of signal control settings optimized by the ated signal control system. In addition, the stochastic-optimization
stochastic-optimization method is evaluated in this section using method consistently produced an acceptable solution.
In this study, the writers developed a stochastic-optimization This research was made possible through funding from the Center
method that combines a heuristic optimizer, GA, and a well- of Transportation Studies at the University of Virginia. The writers
calibrated microscopic simulation model, CORSIM. The main thank the staff members at the City of Charlottesville and the
feature of the proposed stochastic-optimization method was its abil- Virginia Transportation Research Council (VTRC) for their support
ity to optimize traffic signal timings within a realistic microscopic during data collection and network coding during the case study.
This work was also supported by a National Research Foundation
simulation environment by taking into account the stochastic vari-
of Korean grant funded by the Korean government [Ministry
ability commonly present in a transportation system. In addition,
of Education, Science and Technology (MEST)] (NRF-2010-
the proposed stochastic-optimization method was able to optimize
0029451).
simultaneously the four basic traffic signal timing parameters and
the actuated signal settings. On the basis of the case-study results,
the following conclusions were made: References
• The stochastic-optimization method outperformed a macro-
scopic optimization program, Synchro, and Abu-Lebdeh, G., and Benekohal, R. F. (1997). “Development of a traffic
• The stochastic-optimization method is able to consistently control and queue management procedure for oversaturated arterials.”
Transportation Research Record 1603, Transportation Research Board,
produce quality solutions.
Washington, DC, 119–127.
Even though this application was successful, there is a need for Bullock, D., and Catarella, A. (1998). “A real-time simulation environment
further research in the application of the stochastic- for evaluating traffic signal systems.” Transportation Research Record
optimization method. First, one of the significant drawbacks of 1634, Transportation Research Board, Washington, DC, 130–135.
the stochastic-optimization method is the huge computation-time Chakroborty, P., Deb, K., and Subrahmanyam, P. S. (1995). “Optimal
requirements. A reduction in computation time could be achieved scheduling of urban transit system using genetic algorithms.” J. Transp.
by using better hardware (e.g., faster CPUs and more RAM) and Eng., 121(6), 544–553.
Daganzo, C. F., and Sheffi, Y. (1977). “On stochastic models of traffic
parallel computing. If the state-of-the-art computer systems contin-
assignment.” Transp. Sci., 11(3), 253–274.
ues to evolve, the stochastic-optimization method could become Engelbrecht, R. J., and Barnes, K. E. (2003). “Advanced traffic signal
practical in the near future. Second, most traffic signal control log- control for diamond interchanges.” Transportation Research Record
ics embedded in the microscopic simulation models appear to be 1856, Transportation Research Board, Washington, DC, 231–238.
outdated when compared with those used in actual modern traffic Foy, M., Benekohal, R. F., and Goldberg, D. E. (1992). “Signal timing
controllers. Thus, stochastic optimization should be conducted determination using genetic algorithms.” Transportation Research
Record 1365, Transportation Research Board, Washington, DC,
using a software-in-the-loop simulation (SILS) (Bullock and
108–115.
Catarella 1998; Engelbrecht and Barnes 2003). This would ensure Franca, P. M., and Luna, H. P. L. (1982). “Solving stochastic transportation-
adequate evaluations of the advanced features available within the location problems by generalized Benders decomposition.” Transp.
modern traffic controllers during optimization. Sci., 16(2), 113–126.
3(2), 87–94. Rilett, L. R., and Kim, K. (2001). “Comparison of TRANSIM and
Little, J. D. C., and Kelson, M. D. (1980). Optimal signal timing for arterial CORSIM traffic signal simulation modules.” Transportation Research
signal systems, MAXBAND, Operations Research Center, Massachu- Record 1748, Transportation Research Board, Washington, DC, 18–25.
setts Institute of Technology, Cambridge, MA. Sheffi, Y. (1985). Urban transportation networks, Prentice-Hall, Upper
Memon, G. Q., and Bullen, A. G. R. (1996). “Multivariate optimization Saddle River, NJ.
strategies for real time traffic control signals.” Transportation Research Soehodho, S. (1998). “Hybrid model of taxonomy and genetic algorithms
Record 1554, Transportation Research Board, Washington, DC, 36–42. for finding shortest path in transportation systems.” J. Adv. Transp.,
Park, B., Messer, C. J., and Urbanik, T. (1999). “Traffic signal optimization 32(3), 353–368.
program for oversaturated conditions, genetic algorithm approach.” Stevanovic, A., Martin, P. T., and Stevanovic, J. (2007). “VISGAOST:
Transportation Research Record 1683, Transportation Research Board, VISSIM-based genetic algorithm optimization of signal timings.” Proc.
Washington, DC, 133–142. of the 86th Annual Meeting of the Transportation Research Board,
Park, B., Rouphail, N. M., Hochanadel, J., and Sacks, J. (2000). Transportation Research Board, Washington, DC.
“Evaluating the reliability of T7F optimization schemes.” 79th Annual Stevanovic, J., Stevanovic, A., Martin, P. T., and Bauer, T. (2008).
Meeting of the Transportation Research Board, Transportation Re- “Stochastic optimization of traffic control and transit priority settings
search Board, Washington, DC. in VISSIM.” Transp. Res. Part C, 16(3), 332–349.
Park, B., Rouphail, N. M., and Sacks, J. (2001). “Assessment of stochastic Texas Transportation Institute. (2002). PASSER V, College Station, TX.
signal optimization method using microsimulation.” Transportation Xiong, Y., and Schneider, J. B. (1992). “Transportation network design
Research Record 1748, Transportation Research Board, Washington, using a cumulative genetic algorithm and neural network.” Transpor-
DC, 40–45. tation Research Record 1364, Transportation Research Board,
Park, B., and Schneeberger, J. D. (2003). “Evaluation of traffic signal Washington, DC, 37–44.
timing optimization methodology using a stochastic and microscopic Xiong, Y., and Schneider, J. B. (1995). “Processing of constraints in trans-
simulation program.” Research Rep. No. UVACTS-5-0-4, Univ. of portation network design problem.” J. Comput. Civ. Eng., 9(1), 21–28.
Virginia, Charlottesville, VA. Yin, Y. (2000). “Genetic-algorithms-based approach for bilevel program-
Park, B., Won, J., and Yun, I. (2006). “Application of microscopic ming models.” J. Transp. Eng., 126(2), 115–120.