Page 1

Traffic signal timing optimisation based on genetic

algorithm approach, including drivers? routing

Halim Ceylana,*, Michael G.H. Bellb

aDepartment of Civil Engineering, Engineering Faculty, Pamukkale University, Denizli 20070, Turkey

bDepartment of Civil and Environmental Engineering, Imperial College, Exhibition Road, SW7 2BU London, UK

Received 3 July 2002; received in revised form 6 December 2002; accepted 13 February 2003

Abstract

The genetic algorithm approach to solve traffic signal control and traffic assignment problem is used to

tackle the optimisation of signal timings with stochastic user equilibrium link flows. Signal timing is defined

by the common network cycle time, the green time for each signal stage, and the offsets between the

junctions. The system performance index is defined as the sum of a weighted linear combination of delay

and number of stops per unit time for all traffic streams, which is evaluated by the traffic model of

TRANSYT [User guide to TRANSYT, version 8, TRRL Report LR888, Transport and Road Research

Laboratory, Crowthorne, 1980]. Stochastic user equilibrium assignment is formulated as an equivalent

minimisation problem and solved by way of the Path Flow Estimator (PFE). The objective function

adopted is the network performance index (PI) and its use for the Genetic Algorithm (GA) is the inversion

of the network PI, called the fitness function. By integrating the genetic algorithms, traffic assignment and

traffic control, the GATRANSPFE (Genetic Algorithm, TRANSYT and the PFE), solves the equilibrium

network design problem. The performance of the GATRANSPFE is illustrated and compared with mu-

tually consistent (MC) solution using numerical example. The computation results show that the GA

approach is efficient and much simpler than previous heuristic algorithm. Furthermore, results from the test

road network have shown that the values of the performance index were significantly improved relative to

the MC.

? 2003 Elsevier Ltd. All rights reserved.

Transportation Research Part B 38 (2004) 329–342

www.elsevier.com/locate/trb

*Corresponding author. Tel.: +90-258-2134030; fax: +90-258-2125548.

E-mail address: halimc@pamukkale.edu.tr (H. Ceylan).

0191-2615/$ - see front matter ? 2003 Elsevier Ltd. All rights reserved.

doi:10.1016/S0191-2615(03)00015-8

Page 2

1. Introduction

In an urban road network controlled by fixed-time signals, there is an interaction between the

signal timings and the routes chosen by individual road users. From the transportation engi-

neering perspective, network flow patterns are commonly assumed fixed during a short period and

Nomenclature

L

N

M

m

c

cmin

cmax

h

/

Ii

/min

set of links on a road network, 8a 2 L

set of nodes, 8n 2 N

numbers of signal stages at a signalised road network

numbers of signal stages for a particular signalised junction, 8m 2 M

a road network common cycle time

minimum specified cycle time,

maximum specified cycle time

vector of feasible range of offset variables

vector of duration of green times

intergreen time between signal stages

minimum acceptable duration of the green indication for signal stage w ¼ ðc;h;/Þ

whole vector of feasible set of signal timings

vector of feasible region for signal timings

vector of the average flow qaon link a

vector of stochastic user equilibrium link flows

set of origin–destination pairs

set of paths each origin–destination pair w, 8w 2 W

vector of origin–destination flows

vector of all path flows

link-path incidence matrix

OD-path incidence matrix

vector of expected minimum origin–destination cost

gðq;wÞ vector of path travel times

c0

vector of free-flow link travel times

cðq;wÞ vector of all link travel times

dU

a

uniform delay at a signal-controlled junction

dro

a

random plus over saturation delay at a signalised junction

Kmatrix of link choice probabilities

k

the number of signal timing variables on a whole road network, the dimension of the

problem is k ¼PN

NN population size

l

total number of binary bits in the string (i.e., chromosome)

ttgeneration number

X0

q

q?ðwÞ

W

Pw

t

h

d

K

y

i¼1miþ N

Xtt

potential solution matrix of dimension ½NN ? l? for the GA random search space

330

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342

Page 3

control parameters are optimised in order to improve some performance index. On the other

hand, from the transportation planning perspective, traffic assignment models are used to forecast

network flow patterns, generally assuming that capacities decided by network supply parameters,

such as signal settings, are fixed during a short period. The mutual interaction of these two

processes can be explicitly considered, producing the so-called combined control and assignment

problem.

The TRANSYT model proposed by Robertson (1969) has been widely recognised as one of the

most useful tools in studying the optimisation of area traffic control. On the other hand, many

traffic assignment models have been developed in order to find the link flows and path flows given

origin–destination trip rates in an urban road network. One of these assignment models is the

Path Flow Estimator (PFE). It has been developed by TORG (Transport Operation Research

Group), in Newcastle University, to find link and path flows based on stochastic user equilibrium

routing. However, it has been noted (Allsop, 1974; Gartner, 1974) that a full optimisation process

needs to be applied where both problems are relevant; the area traffic control optimisation and

user routing. The combined optimisation problem can be regarded as an Equilibrium Network

Design Problem (ENDP) (Marcotte, 1983). Genetic Algorithms (GAs), first introduced by

Goldberg (1989), have been applied to solve the ENDP (Lee and Machemehl, 1998; Cree et al.,

1999; Yin, 2000).

A number of solution methods to this ENDP have been discussed and good results have been

reported in a medium sized networks. Allsop and Charlesworth (1977) found mutually consistent

traffic signal settings and traffic assignment for a medium size road network. In their study, the

signal settings and link flows were calculated alternatively by solving the signal setting problem

for assumed link flows and by carrying out the user equilibrium assignment for the resulting signal

settings until convergence was achieved. The link performance function is estimated by evaluating

delay for different values of flow and then fitting a polynomial function to these points. The

resulting mutually consistent signal settings and equilibrium link flows, will, however, in general

be non-optimal as has been discussed by Gershwin and Tan (1979) and Dickson (1981).

The first appearance of the GA for traffic signal optimisation was due to Foy et al. (1992), in

which the green timings and common cycle time were the explicit decisional variables and the

offset variables were the implicit decisional variable in a four-junction network when flows remain

fixed. In the optimisation process, a simple microscopic simulation model was used to evaluate

alternative solutions based on minimising delay. The results showed an improvement in the

system performance when the GA was used and suggested that the GA has the potential to

optimise signal timing. The results, however, were not compared with what could be achieved

using existing optimisation tools. It was also concluded that the GA model may be able to solve

more difficult problems than traditional control strategies and search methods in terms of con-

vergence and that good convergence were reported in that study.

In this paper, for the purpose of solving the problem, a bi-level approach has been used. The

upper level problem is signal setting while the lower level problem is finding equilibrium link flows

based on the stochastic effects of drivers? routing. It is, however, known (see Sheffi and Powell,

1983) that there are local optima. It is not certain that the local solution obtained is also the global

optimum because equilibrium signal setting is generally a non-convex optimisation problem.

Hence, the GA approach is used to globally optimise signal setting at the upper level by calling

TRANSYT (Vincent et al., 1980) traffic model to evaluate the objective function.

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342

331

Page 4

2. Formulation

The network performance index (PI) is a function of signal setting variables w ¼ ðc;h;/Þ and

equilibrium link flows q?ðwÞ. The objective function is therefore to minimise PI with respect to

equilibrium link flows q?ðwÞ subject to signal setting constraints. This gives the ENDP problem as

the following minimisation problem:

Minimise

w2X0

PIðw;q?ðwÞÞ ¼

X

8

>

>

a2L

ðWwaDaðw;q?ðwÞÞ þ KkaSaðw;q?ðwÞÞ

cmin6c6cmax

06h6c

/min6/6/max

Pm

ð1Þ

subject to

wðc;h;/Þ 2 X0;

cycle time constraints

values of offset constraints

green time constraints

8m 2 M;8n 2 N

i¼1ð/iþ IiÞ ¼ c

>

>

>

>

>

>

<

:

where q?ðwÞ is implicitly defined by

Minimise

q

Zðw;qÞ

t ¼ Kh;

ð2Þ

subject toq ¼ dh;

hP0

Then the fitness function (i.e., objective function for ENDP) becomes

Maximise FðxÞ ¼

1

PIðw;q?ðwÞÞ

ð3Þ

where PIðw;q?ðwÞÞ is the value of the performance index of the network which is a function of

equilibrium flow pattern q?ðwÞ and signal settings w. All control variables are expressed in integer

seconds, and x is a set of chromosomes that represents w 2 X0, and F is a fitness function for the

GA, to be maximised.

2.1. GA formulation for the upper-level problem

Suppose the fitness function ðFÞ takes a set of w signal timing variables, w ¼ ðc;h1;/1;...;

hn;/nÞ: Rk! R. Suppose further that each decision variable w can take values from a domain

X0¼ ½wmin;wmax? ? R for all w 2 X0. In order to optimise the objective function, we need to code

the decision variables with some precision. The coding process is illustrated as follows:

Decision variables

Mapping

ChromosomeðstringÞ x ¼ j01010101jj01010111;...;10101011jj10101010;...;01010010j

Then, the mapping from a binary string ðbl1bl0...b0Þ representation of variables into a real

numbers w from the range ½wmin;wmax? is carried out in following way:

w ¼

#

jcj

#

jh1;h2;...;hnj

#

j/1;/2;...;/nj

#

332

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342

Page 5

(a) Convert the binary string ðbl1bl0...b0Þ from base 2 to base 10:

(b)

ðbl1bl0...b0Þ2¼

X

li

j¼li?1

bj2j

!

10

¼ Ui

i ¼ 1;2;3;...;k

ð4Þ

(c) Find a corresponding real number for each decision variable for a particular signal timing:

wi¼ wi;minþ Ui

wi;max? wi;min

2li? 1

i ¼ 1;2;3;...;k

ð5Þ

where Uiis the integer resulting from (4). The decoding process from binary bit string to the real

numbers is carried out by means of (5).

The following transformations are carried out for each signal timing variable for use in signal

timing optimisation and traffic assignment purposes.

2.1.1. For common network cycle time

c ¼ cminþ Uiðcmax? cminÞ

2li? 1

i ¼ 1

ð6Þ

2.1.2. For offsets

hi¼ Ui

c

2li? 1

i ¼ 2;3;...;N

ð7Þ

Mapping the vector of offset values to a corresponding signal stage change time at every junction

is carried out as follows:

hi¼ Si;j

where Si;jis the signal stage change time at every junction.

i ¼ 1;2;...;N; j ¼ 1;2;...;m

2.1.3. For stage green timings

Let p1;p2;...;pibe the numbers representing by the genetic strings for m stages of a particular

junction, and I1;I2;...;Imbe the length of the intergreen times between the stages.

The binary bit strings (i.e., p1;p2;...;pi) can be encoded as follows first;

pi¼ pminþ Uiðpmax? pminÞ

2li? 1

i ¼ 1;2;...;m

where pminand pmaxare set as cminand cmax, respectively.

Then, using the following relation the green timings can be distributed to the all signal stages in

a road network as follows second:

/i¼ /min;iþ

pi

k¼1pi

Pm

c

?

X

m

k¼1

Ik?

X

m

k¼1

/min;k

!

i ¼ 1;2;...;m

ð8Þ

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342

333

Page 6

2.2. The formulation for the lower-level problem

2.2.1. The (PFE) as a stochastic user equilibrium assignment (SUE)

The underlying theory of the PFE (see for details Bell et al., 1997) is the logit SUE model based

on the notion that perceived cost determines driver route choice. The basic idea is to find the path

flows and hence links flows, which satisfy an equilibrium condition where all travellers perceive

the shortest path (allowing for delays due to congestion) according to their own perception of

travel time.

The logit model assumes a particular distribution, the Gumbel distribution, for perceived travel

times, which has the great advantage of allowing the formulation of a convex mathematical

program whose solution is unique in the path flows. The equilibrium path flows are found by

solving an equivalent minimisation problem. The attraction of the logit model is that it allows

SUE flows and costs to be calculated by solving the convex mathematical program. The following

minimisation problem

minimise

ZðhÞ ¼ hTðlnðhÞ ? 1Þ þ a

X

a2L

ZqaðhÞ

0

caðxÞdx

ð9Þ

subject tot ¼ Kh;

hP0

where all the notation is as previously stated, due to Fisk (1980), leads to a logit path choice

model. Provided that the link cost functions are monotonically increasing with flows and as-

suming separable link cost functions, then ZðhÞ is strictly convex and, since the constraints are

also convex, it can be proved that there exists one unique solution to the program. The Kuhn–

Karush–Tucker optimality conditions are

rZðhÞ þ KTu?P0;

Since all paths are used

h?P0 and

ðrZðhÞ þ KTu?ÞTh?¼ 0

rZðhÞ ¼ ln h þ ag

ln h?¼ ?ag?? KTu?

It can be shown that this implies the following logit path choice model

expð?ag?

P

where twis the demand for origin destination pair w in W, u is the dual variable and its definition

can be obtained in Bell and Iida (1997). u may be associated with equilibrium link delays when it is

divided by a, the dispersion parameter.

At the optimum

As h?> 0

h?

p¼ tw

pÞ

p2Pwexpð?ag?

pÞ

ð10Þ

rZðh?Þ þ KTu?¼ 0

where rZðh?Þ ¼ ln h?þ ag?

At optimality

ln h?þ ag?þ KTu?¼ 0

334

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342

Page 7

implying

hp¼ expð?agp? uwÞ:

From this logit model

expð?uwÞ ¼

tw

P

p2Pwexpð?agpÞ

so

?uw¼ lntw? ln

X

p2Pw

expð?agpÞ

Hence

?u ¼ ln t þ ay

PFE consists of two loops; an outer loop, which generates paths, and an inner loop, which assigns

flows to paths according to logit path choice model (10). The SUE path flows are found by solving

an equivalent optimisation problem (9) iteratively. An outer loop generates paths and an inner

loop assigns flows to paths according to a logit path choice model.

2.3. The GATRANSPFE solution of the ENDP

A decoded genetic string is required to translate into the form of TRANSYT and PFE inputs,

where TRANSYT model accepts the green times as stage start times, hence offsets between signal-

controlled junctions, and the PFE requires the cycle time and duration of stage greens for that

stages. The assignment of the decoded genetic strings to the signal timings is carried out using the

following relations in the GATRANSPFE.

1. For road network common cycle time

c pði;jÞ

i ¼ 1; j ¼ 1;2;3;...;NN

where p represents the corresponding decoded parent chromosome, j represents the population

index, and i represents the first individual in the chromosome set.

2. For offset variables

hnði;jÞ pði;jÞ;

Since there is no closed-form mapping for offset variables, it is common to map these values to the

interval ð0;cÞ, hence offset values are mapped using (7). The decoded offset values are in some

cases higher than the network cycle time due to the coding process in the GA. In this case, the

remainder of a division between pði;jÞ and the c (i.e., modulo division) is assigned as a stage

change time as follows:

hnði;jÞ MODðpði;jÞ;cÞ; i ¼ 2;3;4;...;N j ¼ 1;2;3;...;NN

3. For green timing distribution to signal stages as a stage change time is

hn;mði;jÞ ¼ hn;m?1ði;jÞ þ ððI þ /Þn;mði;jÞÞ6c;

i ¼ 2;3;4;...;N; j ¼ 1;2;3;...;NN;

8n 2 N; 8m 2 M; i ¼ 1;2;3;...;M

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342

335

Page 8

The solution steps for the GATRANSPFE is:

Step 0. Initialisation. Set the user-specified GA parameters; represent the decision variables w as

binary strings to form a chromosome x by giving the minimum wminand maximum wmax

specified lengths for decision variables.

Step 1. Generate the initial random population of signal timings Xtt; set tt ¼ 1.

Step 2. Decode all signal timing parameters of Xttby using (6)–(8) to map the chromosomes to

the corresponding real numbers.

Step 3. Solve the lower level problem by way of the PFE. This gives a SUE link flows for each

link a in L.

At Step 3, the link travel time function adapted for the PFE is the sum of free-flow travel time

under prevailing traffic conditions (i.e., c0

a) and average delay to a vehicle at the stop-line at a

signal-controlled junction by simplifying the offset expressions for the PFE link travel time

function, where the appropriate expressions for the delay components can be obtained in Ceylan

(2002), as follows:

caðqa;wÞ ¼ c0

aþ dU

aþ dro

a

Step 4. Get the network performance index for resulting signal timing at Step 1 and the corre-

sponding equilibrium link flows resulting in Step 3 by running TRANSYT.

Step 5. Calculate the fitness functions for each chromosome xjusing the expression (3).

Step 6. Reproduce the population Xttaccording to the distribution of the fitness function values.

Step 7. Carry out the crossover operator by a random choice with probability Pc.

Step 8. Carry out the mutation operator by a random choice with probability Pm, then we have a

new population Xttþ1.

Step 9. If the difference between the population average fitness and population best fitness index

is less than 5%, re-start population and go to the Step 1. Else go to Step 10.

Step 10. If tt¼maximal generation number, the chromosome with the highest fitness is adopted

as the optimal solution of the problem. Else set tt ¼ tt þ 1 and return to Step 2.

3. Numerical application

A test network is chosen based upon the one used by Allsop and Charlesworth (1977) and

Chiou (1998). Basic layouts of the network and stage configurations for GATRANSPFE are

given in Fig. 1a and b, where Fig. 1a is adapted from Chiou (1998) and Fig. 1b adopted from

Charlesworth (1977). Travel demands for each origin and destination are those used by

Charlesworth (1977) and also given in Table 1. This numerical test includes 20 origin–destination

pairs, and 21 signal setting variables at six signal-controlled junction.

The GATRANSPFE is performed with the following user-specified parameters:

Population size is 40.

Reproduction operator is binary tournament selection.

336

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342

Page 9

1

A

C

G

F

E

D

B

26

3

5

4

6

14

5

11

12

13

10 17

21

8

9

18

20

154

23

22

19

7

16

3

1 2

Legend

Origin-Destination

Junction

N

Junction

Stage 1Stage 2 Stage 3

1

16

1

2

19

3

15

23

20

4

14

5

11

12

12

10

5

6

13

8

9

8

17

21

7

18

22

1

2

3

4

5

6

(a)

(b)

Fig. 1. (a) Layout for Allsop and Charlesworth?s test network. (b) Stage configurations for the test network.

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342

337

Page 10

Crossover operator is uniform crossover, and the probability is 0.5.

Mutation operator is creep mutation operator, and the probability is 0.02.

The maximal number of generation is 100.

The signal timing constraints are given as follows:

3.1. GATRANSPFE solution for Allsop and Charlesworth’s network

Although the bi-level problem (1) is non-convex and only a local optimum is expected to be

obtained, in this numerical test, the GATRANSPFE model is able to avoid being trapped in a bad

local optimum. The reason for this is that the model starts with a large base of solutions, each of

which is pushed to converge to the optimum. If there is no more improvement on the population

best fitness and population average fitness for the current generation, the GATRANSPFE

re-starts the population. This has the effect of jumping from the current hill to different hills.

The method applied is not dependent on the initial assignments of the signal setting variables.

The random number seed given controls the initial sets of solutions within the population size.

Unlike the GATRANSPFE solution of the network, the MC solution requires the initial as-

signment.

The application of the GATRANSPFE model to Allsop and Charlesworth?s network can be

seen in Fig. 2, where the convergence of the algorithm and improvement on the network per-

formance index and hence the signal timings can be seen. The model calculates the fitness of each

individual chromosome xjin the population. The maximum fitness value found in the current

generation is noted, then for each population pool, the selection, crossover and mutation oper-

ators are applied. When the differences between the population average fitness and population

best fitness of the current generation is less than 5% then the algorithm re-starts with new ran-

domly generated parents, whilst keeping the best fit chromosome from the previous population.

The reason for this is to improve the speed of the model towards the optimum.

cmin;cmax¼ 36;120 s

hmin;hmax¼ 0;120 s

/min¼ 7 s

I1?2;I2?1¼ 5 s

Common network cycle time

Offset values

Minimum green time for signal stages

Intergreen time between the stages

Table 1

Travel demand for Allsop and Charlesworth?s network in vehicles/h

Origin/destinationABDEFOrigin totals

A

C

D

E

G

Destination totals

–250

20

250

130

450

1100

700

200

30

130

50a

200

900

100

20

20

1240

1180

1290

800

480

1250

5000

40

400

300

550

1290

–

30a

170

1100

–

60

270

aWhere the travel demand between O–D pair D and E are not included in this numerical test which can be allocated

directly via links 12 and 13.

338

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342

Page 11

In Fig. 2, there are no improvements on the best fitness value on the first few generations. The

reason for this is that in the first iterations, the algorithm finds a chromosome with very good

fitness value which is better than average fitness of the population. The algorithm keeps the best

fitness then starts to improve population average fitness to the best chromosome while improving

the best chromosome to optimum or near optimum. The considerable improvement on the ob-

jective function usually takes place in the first few iteration because the GA start with randomly

generated chromosomes in a large population pool. After that, small improvements to the ob-

jective function takes place since the average fitness of the whole population will push forward the

population best fitness by way of genetic operators, such as mutation and crossover.

Model analysis is carried out for the 75th generation, where the difference between the popu-

lation average fitness and population best fitness is less than 5%, and network performance index

obtained for that generation is 712.5 £/h. The model convergence can be seen in Fig. 3. The re-

start process began after the 75th generation and there was not much improvement to the pop-

ulation best fitness previously found as can be seen in Fig. 2.

Table 2 shows the signal timings and the final value of the performance index in terms of £/h

and veh-h/h. The common network cycle time resulting from the GATRANSPFE application is

77 s and the start of greens for every stage in the signalised junctions are presented in Table 2.

300.0

500.0

700.0

900.0

1100.0

1300.0

1500.0

1700.0

1900.0

2100.0

2300.0

1917253341

Generation Number

495765738189 97 105

Performance index (£/h)

Avg (£/h)

Best(£/h)

Fig. 2. The application of GATRANSPFE model to the test network.

960.0

980.0

1000.0

1020.0

1040.0

1060.0

1080.0

1100.0

1120.0

1140.0

1160.0

147 10 13 16 19 22 25 28 31 34 37 40 43 46 49 52 55 58

Iteration number

Performance Index (£/h)

PI(£/h)

Fig. 3. The convergence behaviour of MC calculation.

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342

339

Page 12

3.2. MC solution for Allsop and Charlesworth’s network

The MC calculations were carried out with the initial set of signal timings given in Table 3,

where the signal timing are equally distributed to the signal stages. For this initial set of signal

timings, along with equilibrium link flows resulting from the PFE, the initial performance index

and its corresponding value of veh-h/h is given in Table 3.

As can be seen in Fig. 3, for the first iteration after performing a full TRANSYT run with the

corresponding equilibrium flows, the value of the performance index increased from 1024 £/h to

1100 £/h, an increase of 7%. In the second iteration the MC soluton also increases the system

performance index. Thereafter, alternately carrying out the two separate procedures of traffic

assignment and TRANSYT optimisation of the signal timings, the final value of the performance

index is 1075 £/h. This shows that the MC solution increases the system performance index by 5%

when it is compared to the initial value of 1024 £/h. Fluctuations of the value of the performance

index from iteration to iteration is obvious, which shows the non-optimal characteristic of the

mutually consistent signal settings and equilibrium flows for the solution of the bi-level problem.

The total number of iterations in performing the MC calculations in Fig. 3 is 60. The maximum

degree of saturation is 0.97.

Table 4 shows the final values of the start of green timings for each signalised junction and

performance index resulting from the MC solution. The network common cycle time is 82 s.

As for the solution of the MC calculation, Fig. 3 showed that the MC calculation increases

the system performance index. In terms of the convergence, the MC is dependent on the ini-

Table 2

The final values of signal timings derived from the GATRANSPFE model

Cycle time c

(s)

£/hveh-h/h

Performance indexJunction

number i

Start of green in seconds

Stage 1 Si;1

Stage 2 Si;2

Stage 3 Si;3

712.5 75.4771

2

3

4

5

6

032

25

60

72

5

6

–

–

–

59

13

44

64

47

20

30

–

Table 3

Initial signal timing assignment for use in the MC

Performance indexCycle time c

(s)

Junction

number i

Start of green in seconds

Stage 1 Si;1

£/hveh-h/h

Stage 2 Si;2

Stage 3 Si;3

1024.0 110.0701

2

3

4

5

6

0

0

0

0

0

0

35

35

35

23

23

35

–

–

–

46

46

–

340

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342

Page 13

tial assignment. Various sets of initial signal timings were used as a starting point for the MC

(see Ceylan, 2002). Only one set of the initial solutions converged to the predetermined threshold

value that is presented in Fig. 3.

4. Conclusions

1. Allsop and Charlesworth?s example network was used as an illustrative example for showing

the performance of the GATRANSPFE method in terms of resulting values of performance

index for the whole network and the degree of saturation on links. The performance of the pro-

posed method in solving the non-convex bi-level problem showed in that the differences be-

tween resulting values of performance index in all cases were negligible at the 75th

generation. Furthermore, none of the degree of saturation, resulting from the GATRANSPFE

model, were over 90%. The GATRANSPFE model showed good improvement over the MC

calculation in terms of the final values of performance index, with a 34% improvement over

the MC solution of the problem, at the 75th generation, and an improvement in terms of con-

vergence in all cases.

2. The MC solution of the problem was dependent on the initial set of signal timings and its so-

lution was sensitive to the initial assignment. Depending on the initial signal timings, the con-

vergence of the MC solution was not guaranteed. Note that applying the MC solution to Allsop

and Charlesworth?s road network caused the network performance index to increase compared

with the initial performance index, whilst the GATRANSPFE model converged to the optimal

solution (i.e., at the 75th generation) irrespective of the initial signal timings.

3. As for the computation efforts for the GATRANSPFE model, performed on PC 166 Toshiba

machine, each iteration for this numerical example was less than 16.5 s of CPU time in Fortran

90. The total computation efforts for complete run of the GATRANPFE model run was 18.4 h.

On the other hand, the computation effort for the MC solution on the same machine was per-

formed for each iteration in less than 20 s of CPU time and the complete run did not exceed 1 h

on that machine.

4. In this work, the effect of the stage ordering to a network performance index is not taken into

account due to the coding procedure of the GATRANSPFE. Future work should take into

account the effect of the stage orders by appropriately representing the stage sequences as a

suitable GA code.

Table 4

The final values of signal timings resulting from the MC solution

Performance indexCycle time c

(s)

Junction

number i

Start of green in seconds

Stage 1 Si;1

£/hveh-h/h

Stage 2 Si;2

Stage 3 Si;3

1075.0116.0821

2

3

4

5

6

32

15

52

2

27

80

72

66

20

32

62

46

–

–

–

59

5

–

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342

341

Page 14

Acknowledgements

The author would like to thank to David Carroll who has provided me some part of the source

code of the genetic algorithm. The work reported here was sponsored by the scholarship of

Pamukkale University, Turkey. We thank the anonymous referees for their constructive and

useful comments on the paper.

References

Allsop, R.E., 1974. Some possibilities of using traffic control to influence trip distribution and route choice. In:

Transportation and Traffic Theory (Proceedings of the Sixth International Symposium on Transportation and

Traffic Theory). Elsevier, Amsterdam, pp. 345–374.

Allsop, R.E., Charlesworth, J.A., 1977. Traffic in a signal-controlled road network: an example of different signal

timings including different routings. Traffic Engineering Control 18 (5), 262–264.

Bell, M.G.H., Iida, Y., 1997. Transportation Network Analysis. John Wiley and Sons, Chichester, UK.

Bell, M.G.H., Shield, C.M., Busch, F., Kruse, G., 1997. Stochastic user equilibrium path flow estimator.

Transportation Research 5C, 197–210.

Ceylan, H., 2002. A genetic algorithm approach to the equilibrium network design problem. Ph.D. Thesis, University of

Newcastle upon Tyne.

Charlesworth, J.A., 1977. The calculation of mutually consistent signal settings and traffic assignment for a signal-

controlled road network. In: Proceedings of the Seventh International Symposium on Transportation and Traffic

Theory. Elsevier, Kyoto, pp. 545–569.

Chiou, S.-W., 1998. Optimisation of area traffic control for equilibrium network flows. Ph.D. Thesis, University

College, London.

Cree, N.D., Maher, M.J., Paechter, B., 1999. The continuous equilibrium optimal network design problem: a genetic

approach. Proceedings of the Fourth Euro Transportation Meeting. pp. 163–174.

Dickson, T.J., 1981. A note on traffic assignment and signal timings in a signal-controlled road network.

Transportation Research 15B (4), 267–271.

Fisk, C., 1980. Some developments in equilibrium traffic assignment. Transportation Research 14B, 243–255.

Foy, M.D., Benekohal, R.F., Goldberg, D.E., 1992. Signal timing determination using genetic algorithms. In:

Transportation Research Record 1365 TRB. National Research Council, Washington, DC, pp. 108–115.

Gartner, N.H., 1974. Area traffic control and network equilibrium methods. In: Floran, M. (Ed.), Traffic Equilibrium

Methods. Springer-Verlag, Berlin, pp. 274–297.

Gershwin, S.B., Tan, H.N., 1979. Hybrid optimisation: optimal static traffic control constrained by drivers? route choice

behaviour. Laboratory for Information and Decision System Report LIDS-p-870, Massachusetts Institute of

Technology.

Goldberg, D.E., 1989. Genetic Algorithms in Search, Optimisation and Machine Learning. Addison-Wesley, Harlow,

England.

Lee, C., Machemehl, R.B., 1998. Genetic algorithm, local and iterative searches for combining traffic assignment and

signal control. Traffic and transportation studies. In: Proceedings of ICTTS 98. pp. 489–497.

Marcotte, P., 1983. Network optimisation with continuous control parameters. Transportation Science 17 (2), 181–197.

Robertson, D.I., 1969. TRANSYT: a traffic network study tool. RRL Report, LR 253, Transport and Road Research

Laboratory, Crowthorne.

Sheffi, Y., Powell, W., 1983. Optimal signal settings over transportation networks. Journal of Transportation

Engineering 109 (6), 824–839.

Vincent, R.A., Mitchell, A.I., Robertson, D.I., 1980. User guide to TRANSYT, version 8. TRRL Report LR888,

Transport and Road Research Laboratory, Crowthorne.

Yin, Y., 2000. Genetic-algorithm-based approach for bilevel programming models. Journal of Transportation

Engineering 126 (2), 115–120.

342

H. Ceylan, M.G.H. Bell / Transportation Research Part B 38 (2004) 329–342