Available online at www.sciencedirect.

com

Expert Systems with Applications

Expert Systems with Applications 34 (2008) 23702381 www.elsevier.com/locate/eswa

A dynamic and automatic trac light control expert system for solving the road congestion problem
W. Wen
Department of Information Management, LungHwa University of Science and Technology, Taiwan, ROC

Abstract Trac congestion is a severe problem in many modern cities around the world. To solve the problem, we have proposed a framework for a dynamic and automatic trac light control expert system combined with a simulation model, which is composed of six submodels coded in Arena to help analyze the trac problem. The model adopts interarrival time and interdeparture time to simulate the arrival and leaving number of cars on roads. In the experiment, each submodel represents a road that has three intersections. The simulation results physically prove the eciency of the trac system in an urban area, because the average waiting time of cars at every intersection is sharply dropped when the red light duration is 65 s and the green light time duration is 125 s. Meanwhile, further analysis also shows if we keep the interarrival time of roads A, B, and C, and change that of roads D, E, and F from 1.7 to 3.4 s and the interdeparture times at the three intersections on roads A, B, and C are equal to 0.6 s, the total performance of the simulation model is the best. Finally, according to the data collected from RFID readers and the best, second and third best trac light durations generated from the simulation model, the automatic and dynamic trac light control expert system can control how long trac signals should be for trac improvement. 2007 Elsevier Ltd. All rights reserved.
Keywords: Simulation system; Trac light control expert system; Radio Frequency identication; Trac congestion

1. Introduction Trac congestion has been causing many critical problems and challenges in most cities of modern countries. To a commuter or traveller, congestion means lost time, missed opportunities, and frustration. To an employer, congestion means lost worker productivity, trade opportunities, delivery delays, and increased costs. To solve congestion problems is feasible not only by physically constructing new facilities and policies but also by building information technology transportation management systems. A growing body of evidence proves that simply expanding a road infrastructure cannot solve trac congestion problems. In fact, building new roads can actually compound congestion, in some cases, by inducing greater

E-mail address: wenwu@mail.lhu.edu.tw 0957-4174/$ - see front matter 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2007.03.007

demands for vehicle travel demands that quickly eat away the additional capacity. Therefore, many countries are working to manage their existing transportation systems to improve mobility, safety, and trac ows in order to reduce the demand of vehicle use. By enhancing public transport, route guidance systems, trac signal improvements, and incident management, congestion can be improved greatly. Of course, construction of new private bus way, expressways, or subway to increase these growth for easy travel has not kept pace. From a recent analytical statistics of the US department of transportation (2007), it is estimated that roughly half of the congestion is what is known as recurring congestion caused by recurring demands that exist virtually every day, where road use exceeds existing capacity. The other half is due to nonrecurring congestion caused by temporary disruptions. Four main reasons of non-recurring congestion are: trac incidents (ranging from disabled vehicles to major crashes), work zones, weather, and special events. Non-recurring

W. Wen / Expert Systems with Applications 34 (2008) 23702381

2371

events dramatically reduce available capacity and reliability of the entire transportation system. Therefore, researchers have done many researches to increase capacity and remove bottlenecks. Schaefer, Upchurch, and Ashur (1998) developed a simulation model for evaluating freeway lane control signing. The simulation results show that lane control has little inuence on congestion. However, the region between heavy and medium trac ow is sensitive to lane control. Chen and Yang (2000) and Chen and Yang (2003) have created an algorithm to nd a minimum total time path to simulate the operations of trac light control in a city. Stoilova and Stoilov (1998) also built a simulation model to measure the best of trac lights to achieve low noise levels with optimal trac management and environmental pollution. Grau and Barcelo (1992) and Messmer and Papageorgiou (1994) discussed the minimum of queue lengths in dierent intersections. Meanwhile, to aid trac management systems, Nooralahiyan, Dougherty, Mckeown, and Kirby (1997) adopted a Time Delay Neural Network (TDNN) to classify individual traveling vehicles based on their speed-independent acoustic signature. Wen and Hsu (2005) designed a route navigation system with a new revised shortest-path routing algorithm and made a comparison of performance evaluation. Besides, many researches on how to avoid trac congestion by using the shortest-path algorithm have been published Chabini (1998), Chabini (1997), Hoyer and Jumar (1994), Ikeda, Hsu, and Imai (1994), Ikeda and Imai1 (1994) and Maniccam (2006). Moreover, the widespread use of information technology provides an opportunity to enhance the techniques of expert systems (ES), which help managers deal with fast changing environment at a human expert with high-quality performance. Expert systems have a variety of applications in many areas. Additionally, researchers have attempted to develop eective intelligent systems to assist managers in making decisions about how to solve various problems Liao (2002), Liu (1997), Maniccam (2006), Sheu (2006), Xia and Shao (2005), Yang and Recker (2005). Wangermann and Stengel (1998) proposed an intelligent aircraft/airspace system that provides better system performance, redundancy, and safety by using the overlapping capabilities of agents. Powerful and exible multiple agents with the function of principled negotiations are communicated each other. The system gives aircraft and airlines greater freedom to optimize their operations than their have now. Wen and Yang (2006) developed a dynamic and automatic trac light control system for solving the road congestion problem. They simulated a specic road, the Chung San North road in Taipei, Taiwan, to discuss whether a road simulation model can solve a congestion problem. Findle, Surender, and Catrava (1997) developed a exible and general on-line method to determine whether the phasing as an intersection in given trac ow scenarios needs a protected left-turn. In their study, a simulation model and was constructed to reproduce the eect of per-

mitted left-turns at an intersection. For comparison purposes, a number of experiments were carried out. Their new approach has been proven to produce better intersection performance than the 50,000 rule over a signicant range of trac ows. Eriksson (1996) utilized a two-tier architecture, a clientserver model, to build a straightforward expert system which was coded in HTML embedded Java Applet to communicate with a knowledge server in the back end. Some user-interface operations such as mouse dragging, display, and eld checking, were put into the front-end machine that can provide rapid response. Fay (2000) described a railway dispatching system, which has a knowledge base in fuzzy rules of the IFTHEN type. The system adopts fuzzy reasoning to obtain train trac control decisions. The study shows that by systematically making use of the knowledge of train dispatching, trac quality can be improved and operation costs can be reduced. The above descriptions show that using an expert system combined with a trac light control simulation model is a good idea for solving congestion problem. Therefore, our study focuses on trac signal improvements to improve trac congestion problem. The remainder of this paper is organized as follows. Section 2 introduces the framework for the dynamic and automatic trac light control expert system. Section 2.1 describes the description of a simulation model for controlling trac signals. Section 2.2 gives denitions and notations for the simulation model. In Section 3, simulation analysis and results for improving road trac problems are illustrated. Finally, some important conclusions and future work are discussed in Section 4. 2. The framework for the dynamic and automatic trac light control expert system The dynamic and automatic trac light control expert system (DATLCES) is composed of seven elements: a radio frequency identication (RFID) reader, an active RFID tag, a personal digital assistance (PDA), a wireless network, a database, a knowledge base, and a backend server (see Fig. 1). The RFID reader detects a RF-ACTIVE code at 1024 MHz from the active tag pasted on a car. The active tag includes a battery so that it can periodically

Fig. 1. A framework for dynamic and automatic trac light control expert systems.

2372

W. Wen / Expert Systems with Applications 34 (2008) 23702381

and actively transmit messages stored in the tag. Upon receiving the data, the reader will save all information in the PDA. After accumulating x size of data, the PDA with a wireless card will connect to the backend server and store them into the database in the server. The server uses the data stored in the database to calculate maximum ow, interarrival time, and average car speed. When all possible congestion roads and car speed are collected, these data will be used as the input parameters of the trac light control simulation model built in the server. Upon getting the simulation results, the DTLCES is able to automatically infer and provide dierent alternatives in terms of varieties of trac situations and then set red or green light duration via a trac light control interface for improving the trac congestion problem. For easily performing knowledge representation and reasoning, all rules in the knowledge base are presented as a set of if hantecedent clausesithen hconsequent clausesirules. Basically, there are two types of reasoning: backward chaining and forward chaining. Forward chaining, which is a data-driven approach, starts from a basic idea and then tries to draw conclusions. It checks the IF part of IFTHEN rules to nd out whether the antecedent clauses of the rule is matched. As each rule is tested, the program can inference one or more conclusions. Conversely, backward chaining is a goal-driven method. It starts with a goal to be veried as either true or false. It then examines all the THEN parts of IFTHEN rules. A rule that contains this goal in its consequent clause will be checked to know whether its antecedent clause is true. If fails, the program searches for another rule whose conclusion is the same as that of the rst one. This process continues until all possibilities have been tested. Our approach makes use of forward chaining. To clarify the forward chaining procedure, several example rules in the knowledge base for controlling trac signals are shown below: Rule: 1 if 3.0 < Interarrival_time then Singal_Type = 1 Rule: 2 if 1.7 < Interarrival_time 5 3.0 then Singal_ Type = 2 Rule: 3 if 0 5 Ineterarrival_time 5 1.7 then Singal_ Type = 3 Rule: 4 if Interarrival_time = Exception then Singal_ Type = 4 Rule: 5 if Singal_Type = 1 then Red_light_duration = 65 and Green_light_duration = 95 Rule: 6 if Singal_Type = 2 then Red_light_duration = 65 and Green_light_duration = 110 Rule: 7 if Singal_Type = 3 then Red_light_duration = 65 and Green_light_duration = 125 Rule: 8 if Singal_Type = 4 then Red_light_duration = Manual and Green_light_duration = Manual Rule: 9 if 0 5 Interarrival_time 5 1.7 then Interdeparture_ time = 0.6 The algorithm for controlling all the trac lights is as follows:

for(;;) { if(3.0 < Interarrival_time) { then Red_light_duration = 65 and Green_light_ duration = 95; } endif; if(1.7 < Inte0rarrival_time 5 3.0) { then Red_light_duration = 65 and Green_light_ duration = 110; } endif; if(0 5 Interarrival_time 5 1.7) { then Red_light_duration = 65 and Green_light_ duration = 125 and Interdeparture_time = 0.6; } endif; if(Interarrival_time = Exception) { then Red_light_duration = Manual and Green_ light_duration =Manual; } endif; exit(0); } Basically, the simulation model described in the next section will only be run once. Once we get three optimal alternatives, the best, second and third best trac light duration, and the interarrival time collected and computed from the server, the DTLCES uses the above algorithm to reason and control the trac light signals. Notice that the trac light duration on roads A, B, and C is the same as the green light duration on roads D, E, and F. 2.1. Description of the simulation model The 6-road trac control simulation model is composed of six submodels, A, B, C, D, E, and F, which simulate A, B, C, D, E, and F roads trac conditions as shown Fig. 2. Submodels A, B, and C are similar and Submodels D, E, and F are similar. Therefore, we only explain the processes of submodels A and D (see Figs. 3 and 4). The rest of submodels can be understood easily. In the model, we assume that the interarrival time for each car is 1.7 s, the interdeparture time for a car passing the stop line at each intersection is 1.2 s, and the time for passing a length of a car is 0.41 s, which are physically observed in average. We also suppose that the trac light signals at every intersection on a road are set to the same duration and color. In Fig. 3, there are four modules named Light control A1, Light control A2, Light control A3, and Car arrival A.

W. Wen / Expert Systems with Applications 34 (2008) 23702381

2373

Submodel D Submodel A
Traffic Light D1

Submodel E
Traffic Light E1

Submodel F
Traffic Light F1

Submodel B
Traffic Light A1

Traffic Light A2 Traffic Light D2

Traffic Light A3 Traffic Light E2 Traffic Light F2

Submodel C
Traffic Light B1 Traffic Light B2 Traffic Light D3 Traffic Light E3 Traffic Light B3 Traffic Light F3

Traffic Light C1

Traffic Light C2

Traffic Light C3

Fig. 2. A trac control simulation.

The process in the upper dash area of Fig. 3 controls the trac signal at the rst intersection on road A. Light control A1 generates an entity to control trac light signal. Assign for light A1 clock timegets the current simulation time. Prempt seizes the resource switch A1 that has rst priority to get the resource. Delay for red light A1 sets the duration for red light. Release switch A1 releases the resource switch A1 for allowing cars to seize the resource. Finally, Delay for green light A1 sets the duration for green light. Like the process of Light control A1, the processes of Light control A2 and Light control A3 are the same. So, it is easy to derive from the same processes. The lower dash area of Fig. 3 shows the moving process of cars in the rst segment of road A. Car arrival A generates each entity for representing a car based on the interarrival time. The block Queue presents the average number or waiting time of cars in the queue, which cannot pass through the rst intersection on road A. The block Seize catches the resource switch A1 for getting the right to passing through the rst intersection on road A. Seize space of the second segment A road takes the resource spaceA2 to check whether the second segment on the A road has space or not. Then, Assign for clock time passing intersection A1 uses an equation, TNOW + cross A1, to represent the time for passing the rst intersection on road A. Delay for passing intersection A1 is the time for a car passing the stop line at the rst intersection on road A. Release switch A1 for next car frees the resource switch A1 for allowing the next car to move. Assign for decreasing space A2 reduces by 1 to calculate the space in the second segment on road A. Delay for driving through the second segment A

road uses an equation, NA2*T, to represent how long it will take to pass through the remaining space in the second segment on road A. Upon the trac light at the rst intersection changing to red, a car entity holding switch A1 needs to release the resource switch A1 (i.e., the Preempt block in the upper dash area in Fig. 3 takes one unit of a resource switch A1 away from the Seize block in the lower dash area that originally seizes it). The interrupted car entity then will be sent to the Delay block (i.e., a dummy block whose value is 0). The purpose of the dummy block is to continue the process for going to Decide clock light A1 Eqt clock cross A1. Next, Decide clock light A1 Eqt clock cross A1 examines if clock light A1 is equal to cross A1, then delay the remaining time, which is 0, for passing the stop line at intersection A1, then go to Assign for decreasing space A2. If clock light A1 is not equal tocross A1, then go to the module Decide remaining time for passing intersection A1 Eqt 0. If the remaining time for passing the stop line at intersection A1 is equal to 0, then the car entity will be added into the Queue block with the rst priority by using the Insert block. Otherwise, the car entity will go to the Delay block on the right side of Decide remaining time for passing intersection A1 Eqt 0 and delay the remaining time. 2.2. Denitions and notations of the simulation model To illustrate how the system works, we have developed a six-road trac simulation model, which contains nine intersections, by using Arena. Before giving example, let us introduce the denition of notations as follows:

2374

W. Wen / Expert Systems with Applications 34 (2008) 23702381

Light controlA1
0

Assign for light A1clock time

Preempt
switch A1

Delay for red light A1

Release switch A1

Delay for green light A1

Light control A2

Assign for light A2 clock time

Preempt
switch A2

Delay for red light A2

Release switch A2

Delay for green light A2

Light control A3

Preempt

Dealy for red light A3

Release switch A3

Delay for green light A3

switch A3

Car arrival A

Queue

Seize
switch A1

A1Queue

Seize space of the second segment A road

Assign for clock time passing intersection A1

Delay for passing intersection A1

Release switch A1 for next car

Delay
0

Decide clock light A1 Eqt clock cross A1

0 True
Delay
RemainingTime for PassingIntersectionA
1

False

Assign for decreasing space A2

True

Insert

Decide remaining time 0 for passing intersection A1 Eqt 0

False

Delay for driving through the second segment A road

Queue
A2 Queue

Seize
switch A2

Seize space of the third segment A road

Release space of the second segment A road

Assing for increasing space A2

Assign for clock time passing intersection A2

Delay
0

Decide clock light A2 Eqt clock cross A2

0 True
Delay
RemainingTime for PassingIntersectionA2

Delay for passing intersection A2

0
True

False

Insert

Decide remaining time 0 for passing intersection A2 Eqt 0

Release switch A2 for next car

False

Delay for driving through the third segment A road

Assign for decreasing space A3

Seize for car departure at intersection A3

Release space of the third segment A road

Assigb for increasing space A3

Delay for passing intersection A3

Release switch A3 for next car

Cars leave system

Attributes
RemainingTime for PassingIntersectionA1 RemainingTime for PassingIntersectionA2 RemainingTime for PassingIntersectionA3

Delay
RemainingTime for PassingIntersectionA3

Fig. 3. The ow chart of submodel A for trac signal control simulation.

2.3. Variables T: The time for passing a length of a car. Nij: The remaining space of the jth segment of i road, which is measured by the number of

cars (in the model, the remaining space is set to 88), where j = 2, 3 i = A, B, C, D, E, F. cross i: The interdeparture time for a car passing the stop line at the rst intersection where i = A1, B1, C1, D1, E1, F1.

W. Wen / Expert Systems with Applications 34 (2008) 23702381

2375

Light control D1
0

Delay for green light D1

Assign for light D1 clock time

Preempt
switch D1

Delay for red light D1

Release switch D1

Light control D2

Delay for green light D2

Assign for light D2 clock time

Preempt
switch D2

Delay for red light D2

Release switch D2

Light control D3

Delay for green light D3

Preempt
switch D3

Delay for red light D3

Release switch D3

Car arrival D

Queue

Seize
switch D1

Seize space of the second segment D road

Assign for clock time passing intersection D1

Delay for passing intersection D1

Release switch D1 for next car

D1 Queue

Delay
0

Decide clock light A1 Eqt clock cross D1

0 True
Delay
RemainingTime for PassingIntersection D
1

False

Assign for decreasing space D2

True

Insert

Decide remaining time for 0 passing intersection D1 Eqt 0 False

Delay for driving through the second segment D road

Queue
D2 Queue

Seize
switch D2

Seize space of the third segment D road

Release space of the second segment D road

Assing for increasing space D2

Assign for clock time passing intersection D2

Delay
0

Decide clock light A2 Eqt clock cross D2

0 True
Delay
RemainingTime for PassingIntersection D2

Delay for passing intersection D2

0
True

False

Insert

Decide remaining time for 0 passing intersection D2 Eqt 0 False

Release switch D2 for next car

Delay for driving through the third segment D road

Assign for decreasing space D3

Seize for car departure at intersection D3

Release space of the third segment D road

Assigb for increasing space D3

Delay for passing intersection D3

Release switch D3 for next car

Cars D leave system

Attributes
RemainingTime for PassingIntersectionD1 RemainingTime for PassingIntersectionD2 RemainingTime for PassingIntersectionD3

Delay
RemainingTime for PassingIntersection D3

Fig. 4. The ow chart of submodel D for trac signal control simulation.

cross j: The interdeparture time for a car passing the stop line at the second intersection j = A2, B2, C2, D2, E2, F2. cross k: The interdeparture time for a car passing the stop line at the third intersection k = A3, B3, C3, D3, E3, F3.

clock light i: Simulation clock time for trac light signal at the rst intersection turning to red where i = A1, B1, C1, D1, E1, F1. clock light j: Simulation clock time for trac light signal at the second intersection turning to red where j = A2, B2, C2, D2, E2, F2.

2376

W. Wen / Expert Systems with Applications 34 (2008) 23702381

clock cross i: Current simulation time of the car permitting to leave the stop line at the rst intersection (TNOW) pluses cross I where i = A1, B1, C1, D1, E1, F1. clock cross j: Current simulation time of the car arriving at the stop line at the second intersection (TNOW) pluses cross jwhere j = A2, B2, C2, D2, E2, F2. green i: Green light duration at the rst intersection where i = A1, B1, C1, D1, E1, F1. red i: Red light duration at the rst intersection where i = A1, B1, C1, D1, E1, F1. green j: Green light duration at the second intersection where j = A2, B2, C2, D2, E2, F2. red j: Red light duration at the second intersection where j = A2, B2, C2, D2, E2, F2. green k: Green light duration at the third intersection where k = A3, B3, C3, D3, E3, F3. red k: Red light duration at the third intersection where k = A3, B3, C3, D3, E3, F3. 2.4. Resources switch i: A resource for controlling trac light signal or for determining whether a car permits to pass the stop line at the rst intersection where i = A1, B1, C1, D1, E1, F1. switch j: A resource for controlling trac light signal or for determining whether a car permits to pass the stop line at the second intersection where j = A2, B2, C2, D2, E2, F2. switch k: A resource for controlling trac light signal or for determining whether a car permits to pass the stop line at the third intersection where k = A3, B3, C3, D3, E3, F3.
Table 1 The average waiting times for the same light durations (roads A, B, and C) A1Queue r50g50 r65g65 r80g80 r95g95 r110g110 r125g125 1712.85 1685.33 1723.46 1702.55 1740.85 1743.5 A2Queue 49.0432 48.4127 49.2698 49.2105 49.3554 49.0452 A3Queue 37.8693 29.0645 46.4928 49.2501 46.5865 49.0754 B1Queue 1712.85 1685.33 1723.46 1702.55 1740.85 1743.5

2.5. Queues iQueue: A queue at the rst intersection to store all waiting cars, which do not have rights to pass the intersection where i = A1, B1, C1, D1, E1, F1. jQueue: A queue at the second intersection to store all waiting cars, which do not have rights to pass the intersection where j = A2, B2, C2, D2, E2, F2. kQueue: A queue at the third intersection to store all waiting cars, which do not have rights to pass the intersection where k = A3, B3, C3, D3, E3, F3.

3. Simulation analysis and results Using Arena, we have built a trac light control signal simulation model, which consists of six submodels, A, B, C, D, E, and F. During the simulation analysis, initially a warm-up time of 1800 s for reaching system stable and a run time of 10,000 s are set. Then, we set the green light duration with 50 s and the red light duration with 50 s as the rst group. The rest of groups are selected based on both red light and green light duration increased by 15 s each time (i.e., r = 50, g = 50; r = 65, g = 65; r = 80, g = 80; r = 95, g = 95; r = 110, g = 110; and r = 125, g = 125 s). So in Table 1, totally we inspect and analyze 6 cases in terms of dierent light durations. As we mentioned earlier, iQueue represents a queue at the rst intersection to store all waiting cars, which do not have rights to pass the intersection where i = A1, B1, C1, D1, E1, F1. According to the result of this simulation analysis in Table 1 the average waiting times in A1Queue, B1Queue, and C1Queue are very high (see Fig. 5). Therefore, most cars are waiting at the rst intersection on roads

B2Queue 49.0398 48.4121 49.2663 49.2096 49.3503 49.0138

B3Queue 37.9105 29.0378 46.5198 49.2302 46.6262 49.0889

C1Queue 1712.85 1685.33 1723.46 1702.55 1740.85 1743.5

C2Queue 49.0413 48.4125 49.2736 49.2021 48.3535 49.0352

C3Queue 37.9125 29.056 46.4816 49.2346 46.5965 49.0729

2000 1500 1000 500 0

ue ue ue ue ue ue ue ue u ue ue ue ue ue ue Q Q ue Q Q 2Q 1Q 3Q B1 B2 Q C1 C2 C3 Q ue ue e

r50g50 r65g65 r80g80 r95g95 r110g110 r125g125

Fig. 5. The average waiting times for the same light durations (Roads A,B, and C).

B3

W. Wen / Expert Systems with Applications 34 (2008) 23702381

2377

A, B, and C. For the second and third intersections, there are a few cars in the waiting queues. Therefore, based on the average waiting times of A1Queue, B1Queue, and C1Queue, we can nd that it is a better case when the green light duration is 65 s and the red light duration is 65 s and the average waiting time of A1Queue is 1685.33 s. Similarly, Table 2 presents that the average waiting times in D1Queue, E1Queue, and F1Queue are very high in the rst intersection. It also shows when the green light duration is 65 s and the red light duration is 65 s, the average waiting time of D1queue is 1656.2, which is an optimal solution

among these 6 cases (see Fig. 6). From the above description, thus, in the following, we further x the green light duration to 65 s and each time increase the red light duration by 15 s (i.e., g = 65, r = 50; g = 65, r = 65; g = 65, r = 80; g = 65, r = 95; g = 65, r = 110; and g = 65, r = 125 s) as shown in Figs. 7 and 8. Also, we x the red light duration to 65 s and each time increase the green light duration by 15 s (i.e., g = 50, r = 65; g = 65, r = 65; g = 80, r = 65; g = 95, r = 65; g = 110, r = 65 and g = 125, r = 65 s) as shown in Figs. 9 and 10. Table 3 illustrates the average waiting times in A1Queue are 1125.32,

Table 2 The average waiting times for the same light durations (roads D, E, and F) D1Queue r50g50 r65g65 r80g80 r95g95 r110g110 r125g125 1683.6 1656.2 1709.95 1693.3 1717.31 1685.4 D2Queue 32.3026 34.6355 32.1839 34.8785 32.1179 34.5409 D3Queue 29.5532 34.6568 35.0228 34.9261 32.1523 31.8162 E1Queue 1683.6 1656.2 1709.95 1693.3 1717.31 1685.4 E2Queue 32.2915 34.6356 32.1829 34.8729 32.1194 34.5437 E3Queue 29.5403 34.6853 35.0011 34.9295 32.1666 31.8157 F1Queue 1683.6 1656.2 1709.95 1693.3 1717.31 1685.4 F2Queue 32.3031 34.6323 32.1801 34.8844 32.1207 34.5425 F3Queue 29.5391 34.6492 35.0147 34.8996 32.1584 31.8154

2000 1500 1000 500 0

ue ue ue ue ue ue ue ue ue ue ue ue ue ue 3Q ue ue Q Q 1Q 2Q Q Q Q E2 F2 E1 E3 F1 F3 Q ue ue

r50g50 r65g65 r80g80 r95g95 r110g110 r125g125

Fig. 6. The average waiting times for the same light durations (Roads D, E, and F).

2800 2400 2000 1600 1200 800 400 0

ue ue e e ue ue ue ue ue ue u ue ue u ue ue ue B1 Q Q ue C1 Q Q 2Q 1Q 3Q B2 Q C2 B3 C3 Q ue ue

r50g65 r65g65 r80g65 r95g65 r110g65 r125g65

Fig. 7. The average waiting times for dierent red light durations and same green light durations (A, B, and C).

2400 2100 1800 1500 1200 900 600 300 0

r50g65 r65g65 r80g65 r95g95 r110g65 r125g65

ue
ue

ue u

ue

ue

ue

Q D1

D2 Q

Q E2

D 3Q

E3

E1

F1

Fig. 8. The average waiting times for dierent red light durations and same green light durations (D, E, and F).

F2

F3 Q

Qu

ue u

ue

eu

ue

ue

ue

ue

ue

ue

2378
2400 2100 1800 1500 1200 900 600 300 0

W. Wen / Expert Systems with Applications 34 (2008) 23702381

r65g50 r65g65 r65g80 r65g95 r65g110 r65g125

ue ue ue ue ue ue ue Q ue Q ue ue Q ue Q ue A 2Q ue A 3Q ue Q ue C3 ue

A 1Q ue

B1

B2

Q ue

C1

Fig. 9. The average waiting times for the same red light durations and dierent green light durations (A, B, and C).

3000 2700 2400 2100 1800 1500 1200 900 600 300 0
ue ue ue ue ue ue ue ue ue ue ue ue ue Q ue ue ue 2Q 1Q Q 3Q E1 Q Q Q Q ue ue

B3

C2

r65g50 r65g65 r65g80 r65g95 r65g110 r65g125

E2

Fig. 10. The average waiting times for the same red light durations and dierent green light durations (D, E, and F).

Table 3 The average waiting times for dierent red light durations and same green light durations (A, B, and C) A1Queue r50 g65 r65 g65 r80 g65 r95 g65 r110 g65 r125 g65 1125.32 1685.33 2124.22 2475.5 2786 3006.5 A2Queue 39.972 48.4127 57.0281 65.4119 73.8641 82.2903 A3Queue 22.457 29.0645 35.7112 42.22 48.7588 55.2945 B1Queue 1125.32 1685.33 2124.22 2475.5 2786 3006.5 B2Queue 39.9695 48.4121 57.0258 65.4136 73.841 82.3039 B3Queue 22.4688 29.0378 35.725 42.1975 48.7539 55.3223 C1Queue 1125.32 1685.33 2124.22 2475.5 2786 3006.5 C2Queue 39.9627 48.4125 57.0285 65.4142 73.8721 82.3243 C3Queue 22.4461 39.056 35.7176 42.1887 48.7676 55.3138

E3

1685.33, 2124.22, 2475.5, 2786, and 3006.5. We can know when the red light duration is 50 s and the green light duration is 65 s, the average waiting time of A1Queue is 1125.32, which is the lower (see Fig. 7). Furthermore, Table 4 shows that the average waiting times of A1Queue are 2233.75, 1656.2, 1271.3, 900.13, 643.9, and 381.17. The average waiting time of 381.17 is the lowest, which means that the red light duration is 125 s and the green light duration is 65 s (see Fig. 8). Therefore, compare Tables 3 and 4, the average waiting time of 381.17 is the minimum when the red light duration is 125 s and the green light duration is 65 s for roads D, E, and F. However, from Table 5, we

can nd an interesting situation, the average waiting times in A1Queue are 2259.11, 1685.33, 1291.6, 911.29, 656.05, and 387.73. The average waiting time of 387.73 is the lowest when the red light duration is 65 s and the green light duration is 125 s for roads A, B, and C (see Fig. 9). Table 6 also shows the average waiting times of D1Queue are 1110.2, 1656.2, 2097.05, 2461.29, 2746.94, and 3005.2 respectively (see Fig. 10). This means when the red light duration is 65 s and the green light duration is 125 s, the average waiting time, 387.73 s, is the lowest. Actually, this is because if the trac light signals on roads A, B, and C in the EastWest direction are green then those of roads D, E,

Table 4 The average waiting times for dierent red light durations and same green light durations (D, E, and F) D1Queue r50g65 r65g65 r80g65 r95g65 r110g65 r125g65 2233.75 1656.2 1271.3 900.13 643.29 381.17 D2Queue 41.2237 34.6355 26.7653 25.0126 20.3486 20.1356 D3Queue 38.1059 34.6568 29.3691 25.0368 20.3846 17.9425 E1Queue 2233.75 1656.2 1271.3 900.13 643.27 381.17 E2Queue 41.219 34.6356 26.7652 25.0112 20.3434 20.1355 E3Queue 38.0922 34.6853 29.3687 25.0462 20.3625 17.9729 F1Queue 2233.75 1656.2 1271.3 900.13 643.27 381.17 F2Queue 41.2204 34.6323 26.7623 25.0116 20.3472 20.1356 F3Queue 38.1232 34.6492 29.3538 25.0469 20.381 17.956

F1

F2

F3

W. Wen / Expert Systems with Applications 34 (2008) 23702381 Table 5 The average waiting times for the same red light durations and dierent green light durations (A, B, and C) A1Queue r65g50 r65g65 r65g80 r65g95 r65g110 r65g125 2259.11 1685.33 1291.6 911.29 656.05 387.73 A2Queue 60.1125 48.4127 42.337 37.3408 34.1988 31.3646 A3Queue 47.5528 29.0645 39.7984 37.3982 31.9347 31.3834 B1Queue 2259.11 1685.33 1291.6 911.29 656.05 387.73 B2Queue 60.1353 48.4121 42.3542 37.3557 34.1983 31.3609 B3Queue 47.5378 29.0378 39.7762 37.3734 31.9025 31.38 C1Queue 2259.11 1685.33 1291.6 911.29 656.05 387.73 C2Queue 60.123 48.4128 42.3568 37.3521 34.1993 31.3591

2379

C3Queue 47.5485 29.056 39.8164 37.3913 31.9155 31.3799

Table 6 The average waiting times for the same red light durations and dierent green light durations (D, E, and F) D1Queue r65g50 r65g65 r65g80 r65g95 r65g110 r65g125 1110.2 1656.2 2097.05 2461.29 2749.94 3005.2 D2Queue 27.5412 34.6355 41.6674 48.4843 56.1156 62.994 D3Queue 27.5616 34.6568 41.6899 48.5221 56.1541 63.013 E1Queue 1110.2 1656.2 2097.05 2461.29 2749.94 3005.2 E2Queue 27.5413 34.6356 41.6633 48.4877 56.1148 63.0042 E3Queue 27.5699 34.6853 41.7105 48.5158 56.1479 63.0129 F1Queue 1110.2 1656.2 2097.05 2461.29 2749.94 3005.2 F2Queue 27.544 34.6323 41.6664 48.4906 56.1085 63.9992 F3Queue 27.5764 34.6492 41.698 48.5129 56.1693 63.0467

and F in the NorthSouth direction are red. When the trafc duration for EastWest direction is set longer, in contrast, that of NorthSouth direction will be shorter. Consequently, according to the above analysis, we summary that if the red light duration is set to 65 s and the green light duration is set to 125 s, the average waiting time will be the lowest, which means that it is the best solution. When the red light duration is set to 65 s and the green light duration is set to 110, the average waiting time is the second lowest, which means that it is the second best solution. When the red light duration is set to 65 s and the green light duration is set to 95, the average waiting time is the third lowest, which means that it is the third best solution. These simulation results will be used for knowledge reasoning for
Table 7 The average waiting times for red = 65 and green = 125 (A, B, and C) A1Queue ABC_1.2 ABC1_0.6 ABC12_0.6 ABC_0.6 387.73 292.04 199.01 16.9542 A2Queue 31.3602 156.48 147.92 12.1409 A3Queue 31.3882 29.2196 156.48 12.8234 B1Queue 387.73 292.04 199.01 16.9542

trac improvement. At this circumstance, it does not seem worth to continually adjust the red/green light duration. Particularly, if we reduce the red light duration at an intersection, it will decrease the green light duration in the other side at the intersection, which will probably increase the average waiting time in the other side. To nd out the best trac performance, we further adjust the interarrival time and departure time for a car passing the stop line at an intersection. First, we assume that the trac on roads D, E, and F are very light so we change the interarrival time on roads D, E, and F from 1.7 to 3.4 s. This case is represented as ABC_1.2 s, the interdeparture time at the rst intersection on roads A, B, and C has been modied from 1.2 s to 0.6 s (i.e., it

B2Queue 31.3563 156.48 147.92 12.1384

B3Queue 31.4052 29.221 156.48 12.8046

C1Queue 387.73 292.04 199.01 16.9542

C2Queue 31.3582 156.48 147.92 12.1413

C3Queue 31.3926 29.2149 156.48 12.8249

600

400

200

0
ue ue ue ue ue ue ue ue ue ue ue ue ue ue ue ue Q Q Q Q 2Q 1Q 3Q B3 Q Q C3 ue ue

B1

B2

C1

Fig. 11. The average waiting times for red = 65 and green = 125 (A,B, and C).

C2

2380
180 160 140 120 100 80 60 40 20 0
ue

W. Wen / Expert Systems with Applications 34 (2008) 23702381

ue

ue

ue

ue

ue

D 1Q ue

Q ue

ue

ue

D 3Q ue

D 2Q ue

Q ue

Q ue

E2

Q ue

Q ue

E1

E3

Fig. 12. The average waiting times for red = 65 and green = 125 (D, E, and F).

Table 8 The average waiting times for red = 65 and green = 125 (D, E, and F) D1Queue ABC_1.2 ABC1_0.6 ABC12_0.6 ABC_0.6 152.6 152.6 152.6 152.6 D2Queue 70.7229 70.7286 70.7244 70.7269 D3Queue 66.8913 66.897 66.8862 66.8833 E1Queue 152.6 152.6 152.6 152.6 E2Queue 70.7222 70.7279 70.7246 70.7267 E3Queue 66.915 66.8934 66.8973 66.8921 F1Queue 152.6 152.6 152.6 152.6 F2Queue 70.7217 70.7295 70.727 70.7191 F3Queue 66.8833 66.8736 66.8944 66.8918

means that an extension lane is allowed so the average speed can be increased) and named it as ABC1_0.6. Then, that of the rst and second intersection on roads A, B, and C is changed to 0.6 and represented as ABC12_0.6. Finally, all departure times at the three intersections on roads A, B, and C are changed and called as ABC_0.6. Based on Table 7 and Fig. 11, we know that all models (e.g., ABC_1.2, ABC1_0.6, ABC12_0.6, and ABC_0.6) improve trac performance, however, the model for ABC_0.6, which presents all departure times at three intersections changed from 1.2 to 0.6, is the best alternative because the average waiting times of A1Queue, B1Queue, C1Queue, A2Queue, B2Queue, C2Queue, A3Queue, B3Queue, and C3Queue are sharply dropped. Fig. 12 and Table 8 also show the average waiting times of D1Queue, E1Queue, and F1Queue are greatly reduced and those of D2Queue, E2Queue, F2Queue, D3Queue, E3Queue, and F3Queue remain the same values as those in Table 6.

4. Conclusions and future work This paper proposes a new framework for dynamic and automatic trac light control expert systems for improving trac congestion problem. To know how to automatically and dynamically set the time duration of red (green) light signals, we also design a simulation model for improving trac problem in rush hours. In order to analyze system performance, we design a trac simulation model, which consists of six submodels. Each submodel represents a road that has three intersections. The simulation results physically prove the eciency of the trac system in an urban area, because the average waiting time of cars at every intersection is dropped down sharply when the red light duration is 65 s and the green light time duration is 125 s. Meanwhile, further analysis also shows if we keep the inter-

F1

arrival time of roads A, B, and C, and change that of roads D, E, and F from 1.7 to 3.4 s. Besides, four cases including ABC_1.2, ABC1_0.6, ABC12_0.6, and ABC_0.6 are adopted. The case of ABC_1.2 means that the interdeparture time at the three intersections is 1.2 s. The case of ABC1_0.6 presents that the interdeparture time at the rst intersection on roads A, B, and C is 0.6 s, and that of the second and third intersection on roads A, B, and C remains 1.2 s. ABC12_0.6 describes that the interdeparture time at the rst and second intersections on roads A, B, and C is 0.6 and remain that of the third intersection is 1.2 s. Finally, ABC_0.6 stands for the interdeparture time at the three intersections is 0.6. The result shows when the interdeparture times at the three intersections on roads A, B, and C are equal to 0.6 s, the total performance of the simulation model is the best. Although this paper presents and analyzes the DATLCES, there are still several aspects where we can further improve its functions. In particular, we can extend the simulation model to use two-way roads or allow cars turning left or right to let the model more close to the reality. In addition, because we can collect trac ow and average car speed by using RFID technology, the method of dynamically nding a best route or a second optimal route for road navigation systems will be also a major research issue in the future. References
Chabini, I. (1997). A new algorithm for shortest paths in discrete dynamic networks. In 8th IFAC/IFIP/IFOrS symposium on transportation systems, Tech Univ Crete, Greece (pp. 1618). Chabini, I. (1998). Discrete dynamic shortest-path problems in transportation applications. Transportation Research Record, 170175. Chen, Y. L., & Yang, H. H. (2000). Minimization of travel time and weighted number of stops in a trac-light network. Transportation Research B, 34, 241253.

F2

F3

Q ue

ue

W. Wen / Expert Systems with Applications 34 (2008) 23702381 Chen, Y. L., & Yang, H. H. (2003). Minimization of travel time and weighted number of stops in a trac-light network. European Journal of Operational Research, 144, 565580. Eriksson, H. (1996). Expert systems as knowledge servers. IEEE, Expert Systems and the Web, 1418. Fay, A. (2000). A fuzzy knowledge-based system for railway trac control. Engineering Applications of Articial Intelligence, 13, 719729. Findle, N. V., Surender, S., & Catrava, S. (1997). On-line decision about permitted/protected left-hand turns in distributed trac signal control. Engineering Applications Articial Intelligence, 10(3), 315320. Grau, R., & Barcelo, J. (1992). An experience in demand-responsive trac control. In Proceeding of ist Meeting of the Euro Working Group in the Urban Trac and Transportation, Landshut. Germany: Techniczl University of Munich. Hoyer, R., & Jumar, U. (1994). An advanced fuzzy controller for trac lights. Annual Review in Automatic Programming, 19, 6772. Ikeda, T., & Imai1, H. (1994). Fast a algorithms for multiple sequence alignment. IPSJ SIG Notes 94-AL-42-7, IPSJ. Ikeda, T., Hsu, M.Y., & Imai, H., 1994. A fast A., Algorithm for nding better routes by AI search techiques. In: Vehicle navigation and information systems conference proceedings, IEEE (pp. 291296). Liao, S. H. (2002). Problem solving and knowledge inertia. Expert Systems with Applications, 2131. Liu, B. (1997). Routing nding by using knowledge about the road network. IEEE Transactions on System, man, and Cybernetics-Part A: Systems and Humans, 27(4), 425430, July. Maniccam, S. (2006). Adaptive decentralized congestion avoidance in twodimensional trac. Physica A, 343, 512526. Messmer, A., & Papageorgiou, M. (1994). Automatic control methods applied to freeway network trac. In 12 IFAC World Congress, Australia, vol. 9 (pp. 233238).

2381

Nooralahiyan, A. Y., Dougherty, M., Mckeown, D., & Kirby, H. r. (1997). A eld trail of acoustic signature analysis for vehicle classication. Transport Research C, 5(3/4), 165177. Schaefer, L., Upchurch, J., & Ashur, S. A. (1998). An evaluation of freeway lane control signing using computer simulation. Mathematical and Computer Modelling, 27(911), 177187. Sheu, J. B. (2006). A composite trac ow modeling approach for incident-responsive network trac assignment. Physica A, 367, 461478. Stoilova, K., & Stoilov, T. (1998). Trac noise and trac light control. Transportation Research D, 3(6), 399417. US Department of transportation (2007) Congestion Mitigation. <http:// www.fhwa.dot.gov/congestionmitigation/congestionmitigation.htm>. Wangermann, J. P., & Stengel, R. F. (1998). Principled negotiation between intelligent agents: a model for air trac management. Articial Intelligent in Engineering, 12, 177187. Wen, W., & Hsu, H. W. (2005). A route navigation system with a new revised shortest-path routing algorithm and its performance evaluation. WIT Transactions on the Built Environment (Urban Transport), 77, 733743. Wen, W., & Yang, C. L. (2006). A dynamic and automatic trac light control system for solving the road congestion problem. WIT Transactions on the Built Environment (Urban Transport), 89, 307316. Xia, L., & Shao, Y. (2005). Modelling of trac ow and air pollution emission with application to Hong Kong Island. Environmental Modelling & Software, 20, 11751188. Yang, X., & Recker, W. (2005). Simulation studies of information propagation in a self-organizing distributed trac information system. Transportation Research Part C, 13, 370390.