.154... 34%.»: r:
w (1%“

It ,
t’i‘.

. 9n.
ram?»
.3sz i
, $5513.}.
afiiifififi

I
s. I

:l

.921

)
4 lit-run!“ H
1 I. n. it.

 

3 LIBRARY
209-? Michigan State
University

This is to certify that the
dissertation entitled

DEVELOPMENT OF DYNAMIC REAL-TIME
INTEGRATION OF TRANSIT SIGNAL PRIORITY IN
COORDINATED TRAFFIC SIGNAL CONTROL SYSTEM
USING GENETIC ALGORITHMS AND ARTIFICIAL
NEURAL NETWORKS

presented by
MOHAMMAD SHAREEF GHANIM

has been accepted towards fulfillment
of the requirements for the

Ph.D. degree in Civil Engineering

5 , m i, 59/ OB/OI/Zoog

Major Professor’s Signature
J uw C9 I Q ”0%

Date

 

MSU is an affirmative-action, equal-opportunity employer

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested. ‘

 

DATE DUE DATE DUE DATE DUE

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

5/08 K:/Prolecc&Pres/ClRC/DaleDue.indd

DEVELOPMENT OF DYNAMIC REAL-TINIE INTEGRATION OF TRANSIT
SIGNAL PRIORITY IN COORDINATED TRAFFIC SIGNAL CONTROL SYSTEM
USING GENETIC ALGORITHMS AND ARTIFICIAL NEURAL NETWORKS

By

Mohammad Shareef Ghanim

A DISSERTATION

Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY
Department of Civil Engineering

2008

ABSTRACT
DEVELOPMENT OF DYNAMIC REAL-TIME INTEGRATION OF TRANSIT
SIGNAL PRIORITY IN COORDINATED TRAFFIC SIGNAL CONTROL SYSTEM
USING GENETIC ALGORITHMS AND ARTIFICIAL NEURAL NETWORKS
BY
Mohammad Shareef Ghanim
Many transit agencies are interested in employing transit signal priority (TSP) at
signalized intersections in urban areas to reduce travel time and achieve other Operational
benefits. However, state—of—the-art systems do not consider the influence of bus dwelling
activities at bus stops, especially when the stops are located near the intersection stop
line. They are also incapable of efficiently considering simultaneous priority requests and
implementing TSP within coordinated traffic networks. This research develops a dynamic
real-time control logic considering the influence of transit dwelling activities on priority
requests and surrounding traffic performance. The problem of accurately predicting
transit travel time in situations in which dwelling activities may take place before buses
reach the stop line is first solved through the development of an Artificial Neural
Network prediction model. The optimization of the prioritized and coordinated network
traffic signals is then performed by using Genetic Algorithms. The resulting signal
controller is tested in a simulated traffic network using the VISSIM microscopic traffic
simulation software. The simulation results show that the proposed signal control model
has the ability to improve transit and non-transit traffic operational performance. Test
results more specifically show an ability to reduce delays and the number of stops
incurred by transit vehicles, while improving schedule adherence and minimizing

negative impacts on traffic resulting from the provision of transit priority treatments.

DEDICATION

I dedicate this dissertation to my parents, Shareef and Amal, my first teachers. Without
their caring support and encouragement, I would have never been in the graduate school

pursuing the graduate studies.

I dedicate this work to my beloved wife Amal. The care, patience, and encouragement

she offered have a great influence in completing this dissertation.

To the person who does not know what is going on at this moment, my son Ahamd, I
dedicate this work. I would like to thank him for his understanding. Even though he was a

little baby when this research was completed, he knew when to quit disturbing me.

I also dedicate this work to my brother Ahmad, and my sisters, Heba, Waf’ a and Reem,

who are always happy for me and my success.

iii

ACKNOWLEDGMENT

During the work on this research, there is no doubt about who gets my sincere thanks. I
would like to thank my advisors, Dr. Francois Dion and Dr. Ghassan Abu-Lebdeh.
Needless to say that without their guidance and directions, this work would not have been

finished.

I would like also to thank Dr. William C. Taylor and Dr. Erik Goodman for joining the
Ph.D committee, and for the valuable advices and discussions we had throughout my

staying in MSU.

Obviously I would like to extend my thanks to Dr. Ron Harichandran, Mary Mroz, Laura
Taylor, and Margaret Conner for all the help they offered to go through the complications

I faced.

I am also grateful for the faculty and staff who helped me in my undergrad studies at
An-Najah National University for their support and encouragement, Dr. Sameer
abu-Eisheh, Dr. Khalid Al-Sahili, Dr. Najeh Tamirn, Dr. Abdelrazaq Tuqan, Dr. Ryad
Abdelkareem, Dr. Anan Jaioussi, Dr. Shaker Al-Bitar, Dr. Mohammad Ghazzal,
Dr. Hafiz Shaheen, Dr. Muneer Abdo, Dr. Nehaya Awartani, Hamees Etbeeleh, Haitham

Awadeh., Samia Al-Shanti, and Ahmad Al-Ahmad.

For many reasons that I cannot count, I am thankful to my friend Monther Dwaikat.

Although a few words are not enough to do so, but this is the available space I have.

iv

My sincere and deep thanks go to those whom we spent long time since childhood, Basel

Ghanim, Nidal Abdullah, Akram Yameen, Murad Faris, and Ahmad Tawfeeq.

I am thankful for Hui Chen and Abrar Siddiqi for making my stay enjoyable and active at
MSU. I appreciate the time we spent discussing and talking about several things outside

the academic college life.

There are many people I have met during my academic and social life, and the
conversations we had affect me in one way or another, therefore, I would like to extend
my sincere thanks to them, Mohammad Halloush, Osama Ababneh, Hussein Kurt, Suhaib .
Salawdeh, Nassar Odwan, Ma’en Srouji, Tareq Zeyad, Tareq Ayyad, Kais Salhout,
Mahmoud Dwaikat, Buran Khalid, Mahmood Khassawneh, Qasem Al-Zoubi, Salama
Fayyad, Fuad Abu-Jarad, Hassan Abdelkareem, Ahmad Masa’deh, Nazmi Arar, Jomana
Hamadenah, Ala’a Olian, Maya Abu-Elhayyat, Ibraheem N asrallah, Rami Shar’ab, Haya
Al-Aghbar, Jawad Al-Adham, Fawz Kobary, Maha Al-ruzzi, Samah Sarrawi, Ghorood

Awadeh, and Suha Abu-Ebaya.

II.

III.

IV.

TABLE OF CONTENTS

 

LIST OF TABLES ..................................................................... IX
LIST OF FIGURES -------- - -- - _---_X
INTRODUCTION........ ............ . ....... . ......... . ........ . ..... . ...... ......... l

PRIMER ON TRANSIT SIGNAL PRIORITY SYSTEMS .........6

2.1. Traffic Signal Control System .................................................................. 6
2.2. TSP Priority Objectives .......................................................................... 10
2.3. Transit Signal Priority System Components ........................................... 11
2.3.1. Transit Vehicle Detection System ............................................... 12
2.3.2. Communication System ............................................................... 18
2.3.3. Trafi‘ic Signal Control System ..................................................... 18
2.4. Priority Strategies .................................................................................... 19
2.4.1. Passive Priority Systems ............................................................. 20
2.4.2. Active Priority Systems ............................................................... 21
2.4.3. Adaptive (Real-Time) Systems .................................................... 26
2.5. Signal Recovery ...................................................................................... 32
PROBLEM DEFINITION 34
3.1. Bus Stop Location ................................................................................... 36
3.2. Variability of Transit Dwell Times ......................................................... 42
3.3. TSP Within Coordinated Traffic Signal Control Systems ...................... 44
3.4. Simultaneous TSP Requests ................................................................... 46
3.5. Summary ................................................................................................. 47

RESEARCH OBJECTIVES AND APPROACHES....................49

4.1. Research Objectives ................................................................................ 49
4.2. Research Approach ................................................................................. 50
4.2.1. Traffic Signal Optimization Methods .......................................... 51
4.2.2. Optimization Method Selection ................................................... 57
4.2.3. Transit Dwelling Activity Prediction Model ............................... 57
4.2.4. Simulation Environment .............................................................. 60
4.2.5. Overall Research Approach ........................................................ 63

vi

V. DWELL TIME IMPACT ON TSP OPERATIONS- -..--66

5.1.
5.2.
5.3.
5.4.
5.5.

 

Simulation Modeling .............................................................................. 66
Impacts of Dwell Times on Transit Operations ...................................... 74
Impacts on Traffic along Priority Route ................................................. 82
Impacts on Cross-Street Traffic .............................................................. 83
Summary ................................................................................................. 85

VI. D-SPORT COST FUNCTION F ORMULATION.......................86

6.1.
6.2.
6.3.

6.4.

Traffic Signal Control Assumptions ....................................................... 86
Decision Variables .................................................................................. 87
D-SPORT Controller Cost Function Formulation .................................. 93
6.3.1. Traffic Network Performance Index ........................................... 95
6.3.2. Transit Travel Time (TIT) ........................................................ 100
6.3.3. Schedule Adherence Penalty (SCH) .......................................... 104
Optimization and Simulation Integration .............................................. 106

VII. TRANSIT TRAVEL TIME PREDICTION MODEL...............108

7.1.
7.2.
7.3.

7.4.

7.5.
7.6.
7.7.

Experimental Setup ............................................................................... 109
Initial Candidate Model Parameters ...................................................... 112
Data Generation .................................................................................... 113
Selection of ANN Model Parameters ................................................... 116
ANN Prediction Model Training .......................................................... 118
Enhanced Active TSP Performance ...................................................... 123
Summary ............................................................................................... 129

VIII. D-SPORT FRAMEWORK INTEGRATION 130

8.1.

8.2.
8.3.

Genetic Algorithm Platform ................................................................. 130
8.1.1. Solution String Modeling .......................................................... 134
8.1.2. Genetic Algorithm Operations .................................................. 136
8.1.3. Genetic Algorithm Operational Performance .......................... 141
8.1.4. The GA Near-Optimal Solution Search Convergence .............. 143
Integration of D-SPORT within Simulation Environment ................... 148
Experimental Setup ............................................................................... 151

vii

IX.

SIMULATION RESULTS AND COMPARISONS .................. 156

9.1. Transit Travel Time Prediction ............................................................. 157
9.2. Overall Network Performance .............................................................. 159
9.3. Network Total Delay and Number of Stops ......................................... 162
9.4. Schedule Adherence and Transit Service Reliability ............................ 180
9.5. Summary ............................................................................................... 188
CONCLUSIONS AND RECOMMENDATIONS ...................... 189
10.1. Introduction ........................................................................................... 189
10.2. Major Conclusions and Research Contributions ................................... 192
10.3. Recommendations for Future Research ................................................ 194

APPENDIX A: ANN MODEL SOURCE CODE IN
MATLAB 6 ................... -- :.202

APPENDIX B: GA PLATFORM SOURCE CODE IN
FORTRAN 90/95 ............................................... 205

viii

Table 3.1
Table 5.1
Table 7.1
Table 7.2
Table 7.3
Table 7.4

Table 9.1

Table 9.2

Table 9.3

Table 9.4

Table 9.5

LIST OF TABLES

Comparative Analysis of Bus Stop Locations .......................................... 41
Average transit delay for the scenarios considered, in s/veh .................... 76
Variables considered for the ANN prediction model ............................. 113
Dependent and independent variables correlation matrix ....................... 117
Performance evaluation of various neural network configurations ........ 122
Measures of effectiveness for tested controllers ..................................... 127

Simulated transit delays and number of st0ps for different traffic
signal controls ......................................................................................... 165

Simulated traffic delay and number of stops for boarding rate of 50
passenger/hr ............................................................................................ 171

Simulated traffic delay and number of stops for boarding rate of 100
passenger/hr ............................................................................................ 174

Simulated traffic delay and number of stops for boarding rate of 150
passenger/hr ............................................................................................ 177

Transit lateness at bus stops for different traffic demand and a
boarding rate of 50 passenger/hr ............................................................. 182

ix

Figure 2.1
Figure 2.2
Figure 2.3
Figure 2.4
Figure 2.5
Figure 3.1

Figure 3.2

Figure 4.1

Figure 4.2
Figure 5.1

Figure 5.2

Figure 5.3

Figure 5.4

Figure 5.5

Figure 5.6

Figure 5.7

Figure 6.1

Figure 6.2

LIST OF FIGURES

Schematic sketch shows TSP system components. .................................. 12
Early green strategy .................................................................................. 23
Green extension strategy ........................................................................... 23
Phase rotation strategy .............................................................................. 25
Phase insertion strategy ............................................................................. 26
Schematic sketch for bus stops location ................................................... 38

Observed dwell time distribution at a bus stop in Kitchener, Ontario,

Canada ....................................................................................................... 43
Typical architecture of a one-hidden-layer ANN with i input

elements and s neurons follow. ................................................................. 60
Schematic sketch for the D-SPORT controller element integration ......... 63
Simulated traffic network geometric layout ............................................. 67

Sketch summarizes the scenarios that are examined under fixed and
variable dwell time priority setup ............................................................. 73

Likelihood of priority actions with respect to arrival window ................. 77

Average bus delay with respect to average dwell time and signal
cycle length ............................................................................................... 78

Efficiency of early green strategy under different level of service ........... 80

Efficiency of green extension strategy under different level of
service ....................................................................................................... 8 1

Traffic vehicle delays for scenarios considering stochastic dwell
times .......................................................................................................... 84

Schematic sketch for speed profile through the link ................................. 91

Time space diagram for transit vehicle traveling throughout the
system ..................................................................................................... 102

Figure 6.3

Figure 7.1

Figure 7.2

Figure 7.3

Figure 8.1
Figure 8.2
Figure 8.3

Figure 8.4

Figure 8.5

Figure 8.6
Figure 8.7

Figure 8.8

Figure 8.9

Figure 8.10

Figure 8.11

Figure 8.12
Figure 8.13

Figure 9.1

Figure 9.2

Flowchart for optimization process and simulation environment
integration ............................................................................................... 107

Experiments used to generate the datasets required for the ANN
model development ................................................................................. l 1 1

Regression analysis of predicted against observed travel times ............. 122

Comparison of simulated and predicted travel times between check-

in detector and stop line for the evaluation scenarios ............................. 126
The genetic algorithm optimization process flowchart ........................... 133
Binary system representation and decoding ........................................... 134
The examined genetic algorithms configurations ................................... 137

Average cost function value for tournament selection, for different
crossover and mutation rates ................................................................... 142

Average cost function value for roulette wheel selection, for

different crossover and mutation rates .................................................... 142
Number of function evaluations for different GA configurations .......... 144
Average CPU runtime for different GA configurations .......... L .............. 144

GA performance over number of generations for different crossover
rates and a mutation rate of 0.2% ............................................................ 145

GA performance over number function evaluations performed for
different crossover rates and a mutation rate of 0.2% ............................ 146

GA performance over number of generations with and without

Elitism ..................................................................................................... 147
The integrated system architecture for the VISSIM-COM Interface

within the simulation network framework .............................................. 150
Simulated network layout. ...................................................................... 152
Sketch summarizes all the possible combinations that are tested ........... 155

Comparison between simulated and predicted Transit Travel Time ...... 159

Overall traffic network cost function value under different traffic
demand and 50 passenger/hr transit demand .......................................... 161

xi

Figure 9.3

Figure 9.4

Figure 9.5

Figure 9.6

Figure 9.7

Figure 9.8

Figure 9.9

Figure 9.10

Figure 9.11

Figure 9.12

Figure 9.13

Figure 9.14

Figure 9.15

Figure 9.16

Figure 9.17

Overall traffic network cost function value under different traffic
demand and 100 passenger/hr transit demand ........................................ 161

Overall traffic network cost function value under different traffic
demand and 150 passenger/hr transit demand ........................................ 162

Simulated transit delay for different volume-to-capacity
combinations and a boarding rate of 50 passenger/hr ............................. 166

Simulated transit delay for different volume-to-capacity
combinations and a boarding rate of 100 passenger/hr ........................... 167

Simulated transit delay for different volume-to-capacity
combinations and a boarding rate of 150 passenger/hr ........................... 168

Simulated non-transit traffic delay for different We combinations and
a boarding rate of 50 passenger/hr .......................................................... 172

Simulated non-transit number of stops for different We combinations
and a boarding rate of 50 passenger/hr ................................................... 173

Simulated non-transit traffic delay for different We combinations and
a boarding rate of 100 passenger/hr ........................................................ 175

Simulated non-transit number of stops for different v/c combinations
and a boarding rate of 100 passenger/hr ................................................. 176

Simulated non-transit traffic delay for different We combinations and
a boarding rate of 150 passenger/hr ........................................................ 178

Simulated non-transit number of stops for different We combinations
and a boarding rate of 150 passenger/hr ................................................. 179

Transit lateness at bus stop 1 for scenarios and 50 passenger/hr
boarding rate ........................................................................................... 184

Transit lateness at bus stop 2 for scenarios and 50 passenger/hr
boarding rate ........................................................................................... 185

Transit lateness at bus stop 3 for scenarios and 50 passenger/hr
boarding rate ........................................................................................... 186

Transit lateness at bus stop 4 for scenarios and 50 passenger/hr
boarding rate ........................................................................................... 187

xii

CHAPTER I.

INTRODUCTION

 

n response to increasing traffic congestion, many transit agencies are considering
providing buses with preferential treatments along urban arterials. A frequent solution
is to provide priority of passage to transit vehicles at signalized intersections through
temporary signal timing adjustments designed to reduce incurred delays. Since
intersection delays typically account for 15% to 20% of bus travel times [1], such
adjustments offer, in theory, a significant potential for improving transit service along

urban arterials through reduced commute times.

The primary objective of Transit Signal Priority (TSP) systems, as preferential systems
are commonly referred to, is to reduce transit travel times by shortening or eliminating
the time spent by vehicles waiting for a green signal to appear or a queue to dissipate.
Achieved delay reductions can further lead to improved schedule adherence and help

better regulate inter-arrival times, and thereby, contribute to increases in reliability and

quality of service. By favoring the movement of high-occupancy vehicles over low-
occupancy vehicles, TSP also has the potential for improving overall intersection
efficiency on a per person basis. Local transit agencies can further benefit from induced
ridership if motorists decide to switch to transit as a result of the improved service. In
turn, such a modal shift may improve traffic operations at signalized intersections by
reducing the overall traffic demand. TSP may finally provide cost savings to transit
system operators by allowing a specific service to be offered with less person-hours of
work or through the use of fewer vehicles, not counting the possibility of reducing the

frequency of maintenance work on vehicles through reduced usage.

Although traffic signal enhancement strategies favoring the movement of transit vehicles
through signalized intersections have been considered and applied since the 19703 [2],
only recently has the use of such strategies increased. This is strongly linked to the
availability of new technologies allowing improved vehicle detection, reliable
communication between vehicles and roadside equipment, and enhanced traffic signal
control functionalities. The majority of TSP applications are currently primarily
developed in areas with extensive transit service (areas typically characterized by short
headways and/or heavy ridership) and where moderate-to-heavy traffic conditions
combined with the presence of traffic signals create significant impediments to transit
service. However, because of its potential operational benefits and reduced cost of
required technologies, TSP deployments are increasingly being considered for locations

where traffic impediments may not be as significant as past applications [3,4].

TSP is different from what is commonly referred to as signal preemption. While both

processes result in temporary signal alterations, a fundamental difference exists in how

2

these alterations are granted. Signal preemption is usually offered only to emergency
vehicles, such as ambulances and fire trucks, and trains approaching railroad crossings. In
this case, the signal alterations are generally implemented without consideration of the
potential impacts on other traffic. The primary objective here is simply to switch the
signal to green as soon as possible on the approach on which the emergency vehicle or
train is located to reduce the risks of collision with conflicting traffic. On the other hand,
signal priority is primarily provided to facilitate the movement of prioritized vehicles in
mixed-traffic conditions. In this case, signal timing alterations are only provided within a
certain set of conditions and when their impacts on other traffic are within tolerable
limits. In most systems, signal preemption calls from emergency vehicles would override

priority calls from transit vehicles.

While TSP has been shown to provide operational improvements to transit vehicles, it
may negatively impact other vehicles in the traffic stream or pedestrian safety. One of the
most commonly cited negative impacts is the possibility for vehicles on competing
approaches to incur significant delay increases as their green time is being reduced to
serve the approaching transit vehicle. This creates a need for TSP strategies to take into
consideration the challenge of balancing the specific needs of transit vehicles and other

users of the road.

This research focuses on using a genetic algorithm (GA) optimization method and
artificial neural network (ANN) model to develop a real-time TSP controller for use in
signal controllers operating within coordinated signalized urban networks. When
compared with existing practice, the objective of the research is to develop a TSP logic

providing enhanced operational performance when compared to current practice.

3

The remainder of this dissertation is divided into nine chapters organized as follows:

0 The next section, Chapter 2, provides general information about TSP systems,

including typical components and commonly used signal alteration strategies.

0 Chapter 3 focuses on the problems associated with the implementation of TSP

within signalized networks.

0 Chapter 4 presents the objectives of this research and the general approach that
has been adopted for solving the problem defined in Chapter 3. It also details the
optimization and prediction methods that will be followed, as well as the
simulation test-bed that will be used for developing and evaluating the proposed

TSP strategies.

0 Chapter 5 presents the results of simulation analyses that were conducted to assess
the need for the development of an improved bus arrival time prediction tool
considering situations in which bus stops located downstream of bus detection

points can create difficulties in accurately predicting intersection arrival times .

0 Chapter 6 presents the control variables that are considered in the provision of
TSP and details the mathematical derivation of the general cost function that will
be used by the genetic algorithms to determine near-optimal signal timings to

effectively serve both transit and traffic needs.

0 Chapter 7 presents a stop line arrival time prediction tool that has been deve10ped
based on the results of Chapter 5 for providing reasonably reliable information
about the expected arrival time of buses at an intersection stop line on approaches

4

on which dwelling activities downstream of a detection point may create

significant uncertainties in predicting stop line arrivals.

Chapter 8 details the experimental setup for evaluating the proposed traffic signal
control algorithm against pre-timed signal control with and without active TSP.
This Chapter also discusses the architecture of the optimization model that has

been developed.

Chapter 9 presents different measures of effectiveness that have been selected to
assess the benefits associated with implementing the proposed traffic signal
control algorithm against more traditional control strategies under various traffic

conditions and transit demands.

Chapter 10 highlights the main findings in this research and the main

recommendations for extending the work in future research.

CHAPTER II.

PRIMER ON TRANSIT SIGNAL
PRIORITY SYSTEMS

 

3 indicated in the introduction, TSP systems attempt to provide some priority
service opportunities to transit vehicles by implementing temporary signal timing
alterations designed to reduce vehicle wait time at signalized intersections. For readers
less familiar with such systems, this section presents an overview of traffic signal control
systems and current practice associated with TSP systems. The following sub-sections
detail more specifically the general objectives behind the provision of signal priority,

typical system components, and frequently used strategies for signal timing alterations.

2.1. Traffic Signal Control System

Traffic signal control is one of the most effective methods for managing traffic at
intersections. Signal control systems try to manage conflicts between various traffic
streams by allocating passage rights in a specific sequence. The sequence by which

vehicles are allowed to enter an intersection can be set to satisfy various objectives. The

6

most common optimization objectives are to minimize vehicle delay, minimize the

number of stops incurred, or maximizing vehicle throughput.

While traditional traffic signal control systems rely on historical data to develop optimal
signal timing plans, recent systems increasingly rely on real—time data to provide adaptive
control that automatically adjust to changes in traffic conditions. Regardless of the type
of system used, traffic signal control might be approached from three different levels:
isolated intersection level, arterial level, and network level. Controlling at the isolated
intersection level occurs when the traffic signals simply aims to operate the traffic
movements at one intersection, without any communicating with surrounding
intersections. Control at the arterial level occurs when the intersections along a corridor
are synchronized to operate together, either through the establishment of local
communication links or through the use of a centralized control unit. Control at the
network level finally occurs when intersections along different corridors are synchronized

to optimize traffic flow throughout an entire sector or region.

There are some terminologies that are commonly used in designing traffic signal control
systems that should be defined before presenting the different types of signal controls.

These terms are cycle length, phase, phase split, and offset:

0 Cycle length is the total time in seconds required for a complete sequence of
signal indication color.
0 Phase is the part of the cycle that is allocated for a specific traffic stream

movement, or a combined traffic streams.

0 Phase split is the ratio between the green time assigned for a given phase to the
non-green time within a particular cycle.

0 Finally, the offset is the time elapsed in seconds (or percentage) between the
beginning of a given phase in one signalized intersection, and the beginning of the

corresponding phase in the downstream signalized intersection.

Traffic signal control systems can vary in their degree of complexity and required
technology. This has important implication on how TSP can be provided at an

intersection. Existing systems can be classified under one of the following types:

Pre-timed Traffic Signal Control

In pre-timed signal control, the signal plan, such as cycle length, splits, and phases are
based on historical traffic data and do not react to the traffic fluctuation. Pre-timed signal

control is appropriate for areas where traffic demand is very predictable.

Actuated Signal Control

Actuated signal control systems allow the signal controller to provide some response to
fluctuations in vehicle demand and to actuations by pedestrians. This type of system
allows the controller to adjust the phase lengths or skip phases from a pre-timed signal
plans to better match prevailing demands. In this case, the green time is a function of the
number of service calls placed by approaching vehicles and pedestrians. The resulting
phase length could be anywhere between pre-defined minimum and maximum green
times. In all cases, the changes in phase duration are only temporary, as the system

always reverts back to a predefined baseline fixed-time plan at the end of each cycle. As

such, this approach thus only allows for adjustments around a predefined average

demand.

Actuated signal control systems can be categorized into two main groups: semi-actuated
and fully-actuated systems. In semi-actuated systems, all phases can be subject to
actuation except the main green phase, the one serving the approach with the highest
flow. These systems are commonly used along coordinated arterials, where it is desired to
maintain a certain pattern of green indications between intersections to preserve specific
flow progression patterns. In fully-actuated systems, all phases can have their duration
adjusted to traffic demands. These systems are used at isolated intersections with

balanced random demand from all approaches.

Traffic Responsive Signal Control

Traffic responsive signal control system uses information about current traffic conditions
at an intersection or a group of intersections to select appropriate pre-define timing plans
from a library of predefined plans. As such, these systems do not dynamically optimize
signal timings. They simply use available information to determine when might be an

optimal time to switch from one plan to another.

Adaptive (Real-Time) Signal Control

Adaptive (real—time) signal control systems are the most advanced types in operation
today. These systems heavily rely on traffic detection to provide signal timing plans that
vary with observed changes traffic demands. These systems collect information about
approaching traffic and use this information to develop short-term signal timing plans

that optimized for the expected flows. New signal timing plans can be developed by

either modifying an existing plan or generating an entire new plan. Re-optimization can
occur once every few minutes or once every few seconds, for periods extending as well
from a few seconds to a few minutes. The basic difference with actuated control systems
is that newly developed plans generally become the new baseline for the next

optimization. There is no reliance on background predefined plans.

2.2. TSP Priority Objectives

As indicated in the introduction, signal timing alterations are generally provided within
TSP systems with the objective to reduce the number of stops and/or duration of stops
incurred by buses due to the operation of traffic signals. While the reduction of delays
and travel times has historically been the primary goal behind the implementation of
early TSP systems, such systems are nowadays being implemented for a variety of

reasons that may include one or more of the following:

0 Reduction of transit delays at signalized intersections.

0 Reduction of overall transit travel times.

0 Provision of improved schedule adherence, service reliability and regularity.
0 Reduction of vehicular traffic demand by providing enhanced transit service.

0 Reduction of vehicle emissions and fuel consumption.

While priority can be given without consideration for cross-street traffic, such an
approach is usually not desirable, nor easily justified by traffic engineers, due to its
potentially significant negative impacts on overall network traffic performance. The

challenge with the implementation of TSP systems thus generally becomes to give

10

priority to transit vehicles with appropriate consideration for potential negative impacts

on other traffic.

2.3. Transit Signal Priority System Components

There are three major components to a TSP system:

0 Transit vehicle detection system.
0 Communication system.

0 Traffic signal control system.

The schematic in Figure 2.1 illustrates the relationship between these components. The
transit vehicle detection is the component responsible for alerting the signal control
system of an approaching bus. Upon detection of a vehicle, information regarding the
position of the vehicle is sent to the traffic signal controller in charge of the intersection,
either through a local or global communication system. After estimating the most likely
arrival time of the approaching bus, the signal controller then determines whether signal
timing adjustments may benefit the approaching bus and to implement these adjustments

if found beneficial.

The following subsections present a more detail overview of the role of each component

and of their interactions with other system components.

11

 

 

 

Figure 2.1 Schematic sketch shows TSP system components.
2.3.1. Transit Vehicle Detection System
The provision of priority to transit vehicles at signalized intersections relies on knowing
that a vehicle is approaching an intersection or has already reached the intersection stop
line. Once a bus has been detected, it is then the task of the signal controller to make a
decision regarding the appropriate signal timing alterations to implement to reduce transit

wait times while not overly negatively impacting other traffic.

The vehicle detection system is responsible for initiating requests for priority based on
predefined criteria. The criteria used for initiating requests may be static in nature, such
as automatically generating of requests for all buses on certain routes when they arrive at
a certain distance from an intersection, or dynamic, such as the generating requests only
for buses that are behind schedule by more than a certain amount of time or buses that

have a certain number of persons onboard.

12

Depending on the approach selected, transit detection can be made at the local level or at
the network level. At the local level, transit vehicles may be detected approaching an
intersection through the use of various sensing technologies that do not look beyond a
certain distance from the intersection (a few hundred feet). The most commonly used

technologies at this level currently include:

o Inductive loop detection systems.

0 Infrared-based detection systems.

0 Radio frequency-based detection systems.
0 Ultrasonic detection systems.

0 Hi gh-intensity light detection systems.

Detection at the network level is accomplished through an automatic vehicle location
(AVL) system, where vehicles communicate their position to a centralized transit or
traffic management center through use of the Global Positioning System (GPS) or other
vehicle location method. In this case, priority requests are generated based on the
reported position of each vehicle and the expected time at which the bus is forecasted to
reach the prioritized intersection. Appropriate TSP actions would then be selected either
at the central traffic management office and passed down to the local signal controller or
selected directly by the local controller after reception of information from the central

management center that a priority action has been requested and can be granted.

The following sub-sections provide more detailed explanations on how selective transit

vehicle detection is accomplished with each of the technologies mentioned above.

13

Inductive Loop Detection System

Inductive loops operate on the principle of inductance, i.e., the property of a wire or
circuit element to "induce" electric currents in isolated but adjacent conductive media. A
detector consists of an insulated electrical wire, placed on or below the road surface,
attached to a signal amplifier, a power source, and other electronics. Driving an
alternating current through the wire generates an electromagnetic field around the loop.
Any metallic mass passing through the field will absorb electromagnetic energy and
decrease the inductance and frequency of the loop. For most conventional installations, a
change in inductance or frequency beyond a preset threshold will then indicate that a

vehicle has been detected.

Many factors determine loop inductance, including wire size, wire length, the number of
turns, lead length, and insulation. However, detector manufacturers are increasingly
producing detectors having the capability of outputting a vehicle’s inductive signature.
When compared to a set of templates representing typical vehicle inductive signatures,
the extracted profile can be used for classifying passing vehicles and selectively detecting
specific vehicle types, such as transit vehicles. Older technologies generally relied on the
use of an onboard transponder that would emit a unique signal that would be recognized

by the system monitoring the inductive loop [5].

One of the primary disadvantages of such a detection system is its immobility. Since loop
detectors are embedded into the pavement, they cannot be moved easily. In particular,
any change in the location of detectors would imply making additional cuts into the
pavement and possibly installing new power and communication cables in the ground at

some COSIS.

14

Radio Frequency-Based Detection System

Radio frequency based system uses a low frequency radio signal to establish
communication between a transit vehicle and a local signal controller. Transit vehicles
are equipped with a radio beacon transmitter, while the local controller is equipped with a
receiver. When a transit vehicle approaches the intersection, the controller receives the
radio signal emitted by the vehicle. The signal is then decoded to verify the identity of the
bus, its travel direction and assess priority needs. Some of the disadvantages of this
technology include the potential for interferences from surrounding power lines, as well
as a need to control the signal range to avoid influencing nearby signals and systems

operating with similar technology.

Ultrasonic Detection System

Ultrasonic detection systems use sound energy to identify transit vehicles. In order for
this system to work, a set of phased-array microphones is mounted on the traffic signal
mast arm and set in a directional position parallel to the expected source of sound signal.
A vehicle is detected when the microphones pick up changes in sound levels that are
beyond a certain threshold. In this case, selective detection of transit vehicles is based on
the recognition of specific sound patterns associated with these vehicles. However, one of
the main potential shortcomings for this detection technology is noise interference from

the surrounding environment or weather conditions (wind noise, rain noise).

Infrared-Based Detection System

Infrared detection relies on two components: a transmitter that is mounted on transit
vehicles, and a receiver that is generally mounted on a traffic signal support. Detection

occurs when an infrared signal emitted by the transmitter is detected by the traffic signal

15

receiver. Upon reception of an infrared signal, the phase selector then interprets the data
encoded in the signal to identify the approaching vehicle, assess the distance at which the
vehicle is located, and the need for priority action. The main advantage of this system is
that the detection distance can be easily adjusted. However, similar to ultrasonic
detectors, infrared detectors are generally subject to weather and environment

interferences that may reduce their detecting ability and accuracy.

High-Intensity Light Detection System

High—intensity light detecting systems, also known as strobe light systems, are currently
one of the most commonly used detection systems for TSP [5, 6]. Similar to infrared
systems, this technology bases the detection of vehicles on the reception of a light signal
emitted from an emitter mounted on a transit vehicle. The light signal is modulated in
such a way as to provide a unique light signature different from other surrounding light
sources (vehicle headlights in particular), Upon reception of a signal, a priority request is
then sent to the signal controller, which will then decide whether a priority should be
granted or not. Similar to other detector systems, strobe light detectors are subject to
weather and environmental interference that may reduce their ability to detect an

approaching vehicle.

OPS-Based A VL Systems

The Global Positioning System (GPS) is a satellite-based navigation system that has
initially been developed by the US. Department of Defense for military purposes. The
system was gradually implemented throughout the 19803 and made available to the
public in the early 19903 following its completion and recognition of its potential

importance in civilian applications [7].

16

The system is based on a network of 24 satellites that are used for determining the
position of objects anywhere on the globe. Objects can be tracked by the system if they
are equipped with an appropriate receiver allowing them to communicate with satellites
that are in line-of-sight from their location. Communication with at least three satellites is
required for establishing position. Position is determined by triangulation methods using
the measured distance between the object and each satellite. Since distance is measured
by assessing the time spent by a signal to travel between the two objects, any interference
in the path of the signal may result in decreased measurement accuracy. Examples of
interferences include presence of particles in the atmosphere and reflected signals from

tall buildings.

Since there is no subscription for the system, the only costs for using the system are
mainly associated with the need to purchase GPS receivers and equipment to transmit the
assessed position to a dispatch center. There is in this case no need to deploy an extensive
field infrastructure. Agencies wishing to use a system providing increased position and
speed measurement accuracy can still choose to subscribe to a Differential GPS (DGPS)
service. DGPS services use base stations located at known positions for calculating signal
correction factors when calculating the time required by signals to travel between
satellites and receivers. These correction factors are then broadcast and used by DGPS

receivers to increase the accuracy of position measurements.

Because of low initial deployment costs and the ability to know at any time the location
of buses, transit agencies are increasingly relying on the use of GPS to track the

progression of their fleet of vehicles while on the road. Consequently, increasing efforts

l7

have been made to use the information provided by the resulting vehicle location system

to determine appropriate priority actions at signalized intersections.

2.3.2. Communication System

The communication system is responsible for linking the detection system to the traffic
signal control system. Its purpose is to transfer detection information to the control
system. It also allows communication from management center to local controllers.
Depending on the size of the communication system, it is important to specify the system
characteristics that are required for the provision of efficient TSP, such as system
capacity and speed. As an example, in an earlier TSP deployment within a SCOOT real-
time traffic signal control system, it was decided that TSP actions should be provided
locally, by the signal controllers themselves, and not by the central computer that is in
charge of SCOOT signal operations, on the simple grounds that communication of
information from the local intersection to the central computer and then back to the local
controller could take as much as four seconds. The rationale for this decision was based
on the assessment that a four-second lag was too long to allow efficient signal timing
adjustment to be calculated, particularly when considering that a bus may or may not be

traveling and changing position within this interval.

2.3.3. Traffic Signal Control System

The traffic signal control system is responsible for acting on the priority requests and
making any applicable changes to the signal indications, via a centralized or local traffic
signal controller, or a combination of both. In some systems, the local traffic signal
controllers may be able to perform this function completely. In other systems, a

centralized control system may instead be put in charge of granting priority requests. In

18

this case, the local controller would only act an executor of the decisions made by the
central traffic control system. In another variation, a centralized traffic control center may
be instructed to determine a set of constraints that the local controller should follow, but
leave the final decision as to which priority decision to implement to be made by the local

controller based on the established constraints.

Depending on the system setup, the traffic signal control system may or may not make
changes to the signal indications following the reception of a priority request. For
example, if a local policy limits the number of priority activations to one per cycle, a
second priority request received by the traffic signal control system would not result in
further changes to the signal indications. The traffic signal control system is also
responsible for ensuring that higher priority requests (e.g. emergency/railroad

preemption) override other requests in order of priority.

2.4. Priority Strategies

There are generally three classes of TSP strategies;
0 Passive strategies.
0 Active strategies.

0 Real-time strategies.

Passive priority strategies give priority to transit vehicles without the need for vehicle
detection. Conversely, active priority strategies provide priority only after a transit
vehicle has been detected and priority conditions are met. Real-time strategies finally
include strategies in which both the transit and traffic needs are determined

simultaneously and used in on-line signal timing optimizations. The selection of an

19

appropriate strategy to implement depends on the characteristics of the transportation
network, objectives of the transit agency, cost considerations, and factors associated with

the performance of the traffic signal controller, as outlined in the paragraphs that follow.

2.4.1. Passive Priority Systems

As indicated above, passive priority strategies do not require a transit detection system.
As such, these strategies can be described as static as they operate continuously
regardless of whether a transit vehicle is present or not. Passive strategies mainly consist
of signal tinting modifications generally favoring transit vehicles based on historical
traffic performance, but might also include geometric or infrastructure enhancements.

Examples of passive strategies include:

o Adjusting signal settings, such as cycle length, green times and offsets in a way
that favors transit movements.

0 Transit-based coordination: Replacement of traditional traffic-based traffic signal
coordination schemes by new schemes tailored upon transit needs.

0 Increasing priority phase split: Arbitrary increase in the duration of the green
phase serving transit vehicles to increase the probability that a transit vehicle
would reach an intersection during a green signal.

0 Phase splitting: Splitting priority phases into multiple phases and repeating these
phases within the signal cycle to reduce the amount of time that a vehicle has to
wait at a red signal.

0 Queue jumps: Provision of high occupancy lanes or transit lanes at intersections

to allow buses to bypass queues.

20

Passive priority strategies are considered efficient form of TSP in situations where transit
frequencies are high and traffic volumes are low. In such a situation, the high transit
service frequency will tend to produce bus arrivals in almost all cycles or every other
cycle. In addition, the low traffic volumes will tend to minimize impacts on general
traffic in cycles in which a transit vehicle is not present and in which prioritized timing is
not theoretically required. Passive priority strategies may also be efficient at intersections
where transit operations are predictable, as the resulting regularity of transit priority
needs will then facilitate the establishment of permanent signal timing, signal phasing or

road geometry alterations.

The main advantages of passive priority strategies are ease of implementation and low
cost, mainly because they do not require a detection system. However, disadvantages
may include increased delay to side-street traffic, excessive allocation of green time to
priority movement, and dissatisfaction from motorists on non-priority movements
experiencing increased stops and delay, particularly from motorists reaching the
intersection in cycles in which no transit vehicles are present. Furthermore, if the transit
headways are large, as is generally the case in small/medium size cities, these strategies
may induce unnecessary delay to the entire system, since the prioritized signal timing will

keep running even when buses are not there.

2.4.2. Active Priority Systems

Active priority strategies are dynamic signal timing enhancements in which the signal
timings are temporarily modified upon the detection of an approaching transit vehicle.
These strategies provides for an efficient operation by implementing signal changes only

when needed and possible.

21

Active priority strategies are classified into two types: unconditional and conditional.
Unconditional priority simply considers whether a bus has been detected or not. As such,
this type of priority offers considerable potential for express transit during off-peak
hours, especially when there are no limits on green extension or red truncation length.
Conditional priority provides an enhancement by awarding priority to a detected bus only
when specific conditions are met, such as the number of passengers onboard a vehicle,
the extent to which a bus is on or behind schedule, the time since the last priority was

awarded, or potential delays on other traffic.

Whether conditional or unconditional strategies are considered, the following signal

timing alterations are considered to provide preferential treatment to buses:

0 Early green (red truncation).
0 Green extension.
0 Phase rotation.

o Actuated transit phase (phase insertion).

Early Green (Red Truncation)

The early green strategy, also known as red truncation (RT), shortens the red time of the
preceding phases to expedite the return of the green signal serving the movement on
which a prioritized vehicle has been detected. This strategy is applied when the
prioritized vehicle is expected to arrive at an intersection near the end of a red signal.
Figure 2.2 illustrates as an example the changes occurring to a signal timing sequence
when a transit vehicle is expected to arrive a few second before the priority phase starts.

In this case, the traffic signal controller truncates the non-priority phase to allow the

22

transit vehicles to leave the intersection a few seconds earlier than its regular time. This
priority action is subject to minimum green requirements from other phases, as no green
phase is usually allowed to be shortened beyond a certain minimum, such as the time

required by pedestrians to cross an approach or a minimum queue serving time.

 

Normal Cycle
m
Transit Projected Arrival T
Red Truncated

Early Green Strategy

 

 

 

Figure 2.2 Early green strategy

Green Extension (Phase Extension)

The green extension (GE) strategy extends the green time allocated to a phase serving an
approaching bus. This strategy is only applied when an approaching bus is scheduled to
arrive a few seconds after the end of the green. Green extension is one of the most
effective forms of priority action since an extension not only does not require additional
clearance intervals and but also allows a transit vehicle to avoid a relatively long delay
corresponding to an entire red duration. Figure 2.3 explains the concept of green
extension, and what change would occur within a signal cycle when a priority is

requested.

 

Normal Cycle

.nn— .. x ' . -.n._-.-

   

Transit Projected Arrival

Green Extension

Green Extension Strategy l-—)

 

 

 

Figure 2.3 Green extension strategy

23

A3 is often the case, early green and green extension strategies may be applied together to
maximize the time within the signal cycle in which transit would be eligible for priority.
Depending on the system, early greens and green extensions can be offered within a fixed
cycle framework, i.e., without modifying the cycle length, or within a variable cycle
framework. A variable cycle framework provides the most flexibility as there is no need
to compensate phase extensions by reducing the duration of other phases. However, since
the use of variable cycles can disrupt traffic progression pattems in coordinated signal
systems, TSP systems implemented along urban coordinated arterials have historically
only be developed to operate within a fixed cycle setup. Nevertheless, as a result of the
development of more complex signal controllers, some arterial priority systems are now
starting to appear in which the cycle may be temporarily extended or shorted at individual
intersections to accommodate priority requests. In most cases, an extension or shortening
of the signal cycle is compensated by an opposite action in the next cycle to allow the

signal system to quickly reestablish progression patterns.

Phase Rotation

As illustrated in Figure 2.4, the order of signal phases can also be “rotated” to
accommodate approaching transit vehicles. In this case, a northbound left turn phase may
follow a general northbound through phase. A northbound left turning bus requesting
priority arriving before the start of the green phase serving through traffic could then
request that the left turn phase be initiated earlier. With the phase rotation concept, the
left turn phase could be instead displayed as a leading phase before the through phase
instead of as a lagging phase in order to expedite the passage of the transit vehicle. When

compared to the early green and green extension options. the phase rotation would have

24

the advantage of maintaining the amount of green initially dedicated to each traffic
movement. However, implementation of such a strategy is often not recommended as it

might confuse drivers and lead to increased crash risks.

 

Normal Cycle
Non-Priority Non-Priority
Phase 1 Phase 2 Priority Phase
_:—_—_—'__J
Transit Projected Arrival¢
Phase Rotation Strategy
Non-Priority Non-Priority
Phase 1 Priority Phase Phase 2

 

 

 

Figure 2.4 Phase rotation strategy

Actuated Transit Phases (Phase Insertion)

Actuated transit phases are only displayed when a transit vehicle is detected approaching
an intersection. As shown in Figure 2.5, an example would be an exclusive left turn phase
for transit vehicles. The left turn phase would only be displayed within the signal cycle
when a transit vehicle is requesting priority of passage. A potential problem is the fact
that additional lost time might be added into the cycle due to the need to add another
interval. Another element to consider is the possibility for buses to benefit from the
displayed phases. The potential for blockage by other traffic has led to the application of
this priority action only at intersections where buses travel on exclusive lanes or where

queue jumps have been added to allow a transit vehicle to bypass a queue of vehicles.

25

 

Normal Cycle

Non-Priority Non-Priority Non-Priority
Phase 1 Phase 2 Phase 3

     

Transit Projected Arrival¢

Phase Rotation Strategy
Non-Priority Priority Phase Non-Priority Non-Priority

Phaset Phase2 Phase3
.. . . - - - - -

Figure 2.5 Phase insertion strategy

 

 

 

2.4.3. Adaptive (Real-Time) Systems

Adaptive/real-time TSP strategies provide priority while simultaneously trying to
optimize traffic movements. Optimization is generally conducted using performance
criterion linked to general traffic operations. The selected criterion may include person
delay, transit delay, vehicle delay, and/or a combination of these criteria. Adaptive
strategies continuously seek to optimize the active timing plan based on real-time,
observed data. As such, they require early detection of both transit vehicles and traffic to
obtain advanced information about short-near-term vehicle arrivals. Adaptive systems
also often require the ability to update the transit vehicle’s arrival time, which can vary
due to the number of stops and traffic conditions. The updated arrival time can then be

fed back into the process of adjusting the signal timings.

Over the years, there have been many attempts to incorporate TSP within real—time traffic
signal control systems. Below is a summary how SCATS, SCOOT, OPAC, UTOPIA and
RHODES were designed or are proposing to handle TSP within a real-time operational

optimization context.

26

SCA TS

The SCATS (Sydney Coordinated Adaptive Traffic System) traffic adaptive signal
control system was originally developed for the New South Wales Roads and Traffic
Authority for application in Sydney and other Australian cities. Over the years, this
system has been implemented in over 80 cities worldwide, including Australia, New
Zealand, and East Asia. In the US, it has been deployed in Oakland County, Michigan,
Delaware and Minnesota. Its main operational objective is to adjust cycle time, splits and
offsets in response to real-time traffic demand with the goal to minimize overall stops and

delays [5].

The hierarchical control architecture in SCATS consists of two levels: a strategic level
and a tactical level. At the strategic level, a regional computer coordinates the timing of
signals in a network or subsystem consisting of one to ten signals. Subsystems can link
together, to form larger systems operating on a common cycle time, where each
subsystem has only one critical intersection. At the tactical control level, signal tinting
decisions are distributed among the individual intersections. Split plans are selected by
individual local controllers within the constraints imposed by the regional computer’s
strategic control. Each local controller bases its signal control decisions on data collected
from sensors monitoring how traffic utilizes the available green time. The objective here
is to maximize green time utilization by considering terminating a phase when the

demand for the green drops below a certain threshold.

With respect to TSP, SCATS is capable of providing both passive and active priority.
Passive priority treatments can be granted to transit vehicles by allocating a favorable

bias to the links traveled by these vehicles in the signal coordination plans considered by

27

the regional computer. On the other hand, active priority requests can be granted by local
controllers as part of their tactical control decisions. Upon detection of an approaching
bus, priority phases can be called to either clear the queue ahead of the vehicle or to
provide a temporary phase extension. Priority requests can be granted or denied
depending on the time of day, tidal flow determination based on traffic flows or on how
congested the intersection is. Following the granting of a priority request, the system
would automatically compensate for any negative impacts on traffic by reacting to
demands for longer green time by vehicular traffic on the approach for which the green

time may have been reduced.

Because SCATS’ stop line traffic detector cannot provide advanced information about
approaching transit vehicles, TSP is only possible if additional detectors located at some

distance upstream of the intersection are installed.

SCOOT

SCOOT (Split Cycle Offset Optimization Technique) is a centralized traffic signal
control system that was designed to automatically respond to fluctuations in traffic
demand through the use of detectors located at some upstream distance from the
intersection. The system uses the detectors to build a profile of traffic demand in the next
signal cycle and uses this cyclic flow profile to determine whether it may be beneficial to
change by a few seconds the time at which individual phases are scheduled to be
terrrrinated. Signal control decisions within SCOOT are taken not by individual signal
controllers, but by a central computer that is in charge of optimizing signal timings at all

intersections within a given region. This system is the predominant real-time traffic

28

signal control system in operation today, with installations throughout Europe, Asia and

North America [8, 9].

Similar to SCATS, SCOOT can provide either passive or active preferential treatments.
Passive treatments can be implemented by introducing biases for the links traveled by
transit vehicles or imposing fixed offsets between signalized intersections based on
average transit travel times. Active priority treatments can also be offered. Such
treatments are granted by local controllers as an override of signal timing decisions made
by the central computer, although some systems may have the central computer taking an
active role in the granting of priority requests [10]. Upon detection of an approaching
bus, a signal control first checks whether it has been granted permission to override
network control decisions. Permissions to override network control may for instance be
denied if traffic demands are too high and the intersection is operating near capacity. If
overrides are possible, the signal controller will then assess whether a green extension or
early green recall can be offered, within the limits set by the central computer, to provide

preferential treatment to the approaching bus.

To avoid bus arrival prediction problems caused by variable dwell times, it is generally
required within SCOOT systems that bus detection occurs downstream of bus stops. This
means that when SCOOT system is implemented, bus stops must be located on the far-
side of an intersection or sufficient upstream of an intersection approach to provide

advance warning of an approaching bus [9].

29

OPA C

Optimization Policies for Adaptive Control (OPAC) is a computational strategy for real-
time demand responsive traffic signal control [11] that was initially developed to control
isolated intersections but that has been expanded to network control. This system uses a
dynamic programming approach to determine the combination of signal switching
decisions that maximizes overall effectiveness or minimizes overall disutility. The
optimization algorithm tries to find a short-term local signal control solution that would
be close to near-optimal operation based on predicted and actual vehicle arrivals. OPAC
progressively selects a signal pattern from among a number of possible solutions for each
intersection individually. The selected patterns are recalculated every few seconds based
on updated traffic data [5]. TSP is implemented within OPAC by modifying the
algorithm so that signal settings are optimized based on number of passengers instead of

vehicles flow [12].

U TOPIA

UTOPIA is an urban traffic control system that was first developed and implemented in
Italy in 1985 with the objective of integrating real-time traffic control with transit priority
needs. The system places a strong emphasis on the decentralization of signal timing
optimization. It is a two-level control system in which signal controllers play a dominant
role in the development of signal timing strategies. The controllers determine the signal
settings to be implemented at each intersection according to local traffic demands, transit
priority needs, and coordination needs with adjacent intersections. The central computer

adds robustness and stability to the control strategies developed by the local controllers

30

by imposing timing constraints based on network-wide coordination objectives and a

network-based reference timing plan.

As indicated above, transit priority is primarily considered at the intersection level. In the
simplest setup, priority is granted on a simple call basis by local controllers, with no
direct consideration for the potential disruptions to other traffic. A more complex
implementation in the original system developed for the city of Turin in Italy had a more
complex bus arrival prediction system allowing individual signal controllers to get up to
8-min advance warning of an upcorrring bus arrival. Each time a transit vehicle is
detected, a message is sent to the central computer asking it to predict the arrival times of
the vehicle at successive downstream intersections for the next 8 minutes. These times
are updated each time a bus passes another detector [11, 13], and the updated information

is used to adjust the signal timings to better match the projected bus anival.

RHODES

RHODES was deve10ped based on the concept of predicting arrival of vehicles on all
approaches to each intersection and using the resulting predictions to explicitly optimize
phase durations to produce a given optimal performance according to a user-defined
measure, such as minimizing average delay or number or stops. Similar to UTOPIA,
RHODES features a hierarchical control structure. At the highest level, the system uses a
dynamic network loading model that captures general traffic characteristics variation.
The traffic characteristics considered at this level pertain to the network geometry, travel
demand between origins and destinations, and typical route selections of travelers.
Variations in these characteristics provide RHODES with general allocations of green

times for each different demand pattern. The resulting green time decisions are then

31

updated at the middle level of the hierarchy, referred to as “network flow control.” At this
level, traffic flow characteristics are measured in terms of platoons of vehicles and their
speeds. Signal timings generated at the upper level are adjusted here to adequately service
expected platoon arrivals. The resulting tirrrings are then passed down to the third level,
the “intersection control” level, which will then select the appropriate phase change
epochs based on observed and predicted arrivals of individual vehicles at each

intersection.

TSP is implemented in RHODES in two different ways referred to as the weighted buses
approach and the phase constrained approach. In the weighted buses approach, transit
vehicles are converted to a certain number of passenger cars equivalent based on their
lateness and their passenger load. In the phase constrained approach, the TSP is
implemented by scheduling the green time for transit vehicles according to the time each

transit vehicle is expected to arrive at an intersection [14-16].

2.5. Signal Recovery

Although signal recovery is not a particular TSP strategy, it is important to mention its
role in system performance. TSP actions do not always end when a transit vehicle passes
through an intersection and the priority phasing is terminated. For equity reasons, many
traffic engineers will require that signal controllers follow up on the priority actions with
a recovery operation, in which the signal controller would transition back to normal
signal operation. The idea behind such transition is to compensate for time lost by the
non-priority movements or maintain coordination along arterials. Examples of recovery

actions commonly considered include:

32

o Denial of priority for n cycles following the granting of a priority request.
0 Provision of extended green time to the non-prioritized traffic that was penalized

by the priority action.

The implementation of signal recovery plans is seen by traffic engineers as playing a vital
role in mitigating impacts on cross streets and helping to maintain coordination along
urban corridors. However, it must also be recognized that the implementation of recovery
methods can significantly limit the ability of a priority system to respond to the needs of
all buses requesting priority. For this reason, signal recovery plans must be carefully

designed if their use is warranted.

33

CHAPTER 111.

PROBLEM DEFINITION

 

umerous studies have already demonstrated that transit vehicles benefit from

traffic signal priority systems [17-20]. However, many of these studies have also
shown a potential for increased traffic delays as a result of the need to temporarily
shorten certain green phases or extend the signal cycle to accommodate approaching
buses. These delay increases can be particularly significant at busy intersections where
most of the available green time may already be required to serve existing traffic.
Another concerning element is the potential loss of progression patterns along
coordinated arterials, which may result in delay increases at a number of successive

intersections.

Due to the potential for negative impacts, systems providing effective preferential
treatments to buses while keeping traffic impacts at a minimum are generally sought.
This means developing systems that consider all potentially impacting parameters when

making signal timing alteration decisions. One parameter to consider is uncertainty in

34

predicting transit travel times. The importance of this parameter is linked to the
uncertainty regarding the time a bus will need to board and alight passengers. If dwell
times were perfectly predictable, stop line bus arrival times could always be accurately
predicted. However, dwell times generally exhibit some variability, which creates
uncertainty in arrival predictions. In turn, this may lead to the implementation of green
extensions and early recalls that may not be used by the transit vehicles at the source of

their implementation.

Based on observations from several US. cities, the Transit Capacity and Quality of
Service Manual, or TCQSM [21], indicates that bus dwell times are characterized by a
coefficient of variation ranging from 0.4 to 0.8. Among the impacting elements

responsible for such a variation are:

Passenger demand and loading, particularly the number of persons passing

through the highest-volume door and the proportion of alighting and boarding

passengers attempting to use the same door.

0 Bus stop spacing, as fewer bus stops lead to increased demands at individual
stops.

0 Fare payment procedure, as payment with exact change can take up to an
additional 1.5-3 per passenger than payment with a pass.

0 Vehicle type, as having to ascend or descend steps increases the time required to
serve each passenger.

0 In-vehicle circulation, as extra time may be needed to allow passengers to reach

the door or the fare box when persons are standing within the vehicle.

35

o Other specific needs, such as time required for wheelchair loading.

Specific problematic elements covered in this dissertation and outlining challenges to be

considered in the proposed research include:

0 The problem of bus stop location.

0 The problem of variable dwell times.

0 TSP within coordinated signal systems.

0 Simultaneous TSP requests.
3.1. Bus Stop Location
As the point of access to transit services, the bus stop is a critical element in a transit
system's overall goal of providing timely, safe, and convenient transportation. On this
note, TCRP Report No. 19 summarized several concerns of interest to transit service
users and riders regarding the design and location of transit stops [22]. These concerns

are I

0 Transit system performance: Travel time for a bus trip has four components: the
time it takes to walk to the bus stop, the wait time for the bus, the actual in-
vehicle travel time, and the time to walk to the destination. Within this context,
transit system performance is strongly affected by the location of bus stops and

the frequency at which buses service each stop.

0 Traflic flow impacts: Both the location and design of a bus stop affect the flow
and movement of other vehicles. In particular, a well-designed bus stop can allow
passengers to board and alight without the bus significantly impeding or delaying

adjacent traffic.

36

0 Safety: Safety is the freedom from danger and risk. In a transit environment, this
concept includes an individual's relationship to buses and general traffic, and the
relationship of a bus to other vehicles. Safety issues for pedestrians include
neamess of a bench to the flow of traffic on a busy street, or the ability to safely
cross a street to reach a bus stop. For buses, safe reentry into the flow of traffic is
a prime example of operational safety concern. In general, different safety levels
may be associated with different bus stop location and layout, often making this

evaluation context-specific.

0 Security: Security refers to an individual's feeling of well being. Security is
affected by lighting at bus stops, bus stop visibility from the street and nearby

land uses, and presence of hiding places at bus stops.

The above elements define functional and performance-related concerns for public
transportation systems. Each must be addressed to achieve the goal of timely, safe, and
convenient public transportation and to satisfy the needs of the service area. More
importantly, as was outlined, each area of concern is influenced by the selected location

and layout of bus stops.

For TSP systems, a particularly important element is the bus stop location relative to the
intersection. Bus stop locations can be categorized as far-side, near-side, and midblock.

Figure 3.1 illustrates the physical meaning of each of these categorizations.

37

 

l I I l Far-Side Bus Stop:

 

 

@21 Bus stops after passing
through the intersectionck

 

 

 

Near-Side Bus Stop:

 

 

if? Bus stops right before
passing through the
P intersection

 

 

 

 

 

 

 

Mldblock Bus Stop:

 

 

[ii Bus stops within the block

F

Figure 3.1 Schematic sketch for bus stops location

 

 

 

 

 

 

 

 

 

Each type of bus stops has its own advantages and disadvantages, which often lead
individual agencies to favor one type over the other. The TCRP report [22] suggests the
consideration of the following factors when selecting which type to implement at a

particular location:

0 Adjacent land use and activities.

0 Bus route (for example, is bus turning at the intersection).

0 Bus signal priority (e. g., extended green suggests far side placement).
0 Impact on intersection operations.

0 Intersecting transit routes.

0 Intersection geometry.

0 Parking restrictions and requirements.

0 Passenger origins and destinations.

38

o Pedestrian access, including accessibility for handicap/wheelchair patrons.
0 Physical roadside constraints (trees, poles, driveways, etc.).

0 Potential patronage.

0 Presence of bus bypass lane.

0 Traffic control devices.

Consideration of the above elements may result in a bus stop placement on the near-side
or far-side of an intersection, or at some distance upstream. For TSP, the most
problematic placement is a near-side stop, where the bus would stop at or very close to
the intersection stop line. This is problematic as such a placement creates difficulty in
predicting the exact moment at which a bus may need to enter the intersection due to the
uncertainty in knowing the exact amount of time that a bus will need to remain at the stop
to board and alight passengers. Since the bus stop is close to the stop line, there is very
little time to respond to a bus leaving the stop. Prediction models must then be used to
assess the most likely bus stop line arrival times. As will be detailed later, this creates
problems for the signal controller in correctly assessing the priority needs and the actions
that need to be undertaken. While moving the bus stop further upstream may allow bus
detection downstream of the stop and remove the problem of variable dwell time, such an
upstream location introduces new problems in the form of travel time uncertainties in
mixed traffic, where traffic interactions may delay the progression of a bus beyond what

may be expected.

As can be assessed, there is no clear ideal bus stop location. This assessment is further

emphasized by conclusions from TCRP Report 19, shown in Table 3.1, which summarize

39

the advantages and disadvantages associated with each bus stop type. Because of the
difficulty of accurately predicting bus arrivals with near-side stops, current practice with
respect to TSP systems is to prefer midblock or far-side stops, to allow bus detection
downstream of bus stops. Many TSP equipment providers even go as far as suggesting
moving near-side bus stops. However, consideration of cases with near-side stops should
not be ignored, as many transit agencies may prefer these st0ps for various reasons (such
as facilitating transfer, reducing street crossings) or may physically be unable to move
near-side stops to a “more suitable” alternate location. Moreover, the deployment of TSP
should be flexible enough to take into account the different types of bus stops. Therefore,
the need for a comprehensive TSP model that is insensitive to the bus stop type is

important.

40

Table 3.1

Comparative Analysis of Bus Stop Locations

 

Advantage

Disadvantage

 

Far-side Bus Stop

Minimize conflicts between right
turning vehicles and buses.

Provide additional right turn
capacity by making curb lane
available for traffic.

Minimize sight distance problems
on approaches to intersections.
Encourage pedestrians to cross
behind the bus.

Creates shorter deceleration
distance for buses since the bus
can use the intersection to
decelerate

Resulting in bus drivers being
able to take advantage of the
gaps between traffic flows that
are created at signalized
intersections.

May result in the intersection being
blocked during peak periods by
stopping buses.

May obscure sight distance for
crossing vehicles.

May increase sight distance
problems for crossing pedestrians.

Can cause a bus to stop far side
after stopping for a red light, which
interferes with both bus operations
and all other traffic.

May increase number of rear-end
accidents since drivers do not
expect buses to stop again after
stopping at a red light

Could result in traffic queued into
intersection when a bus is stopped
in travel lane.

 

Near-side Bus Stop

Minimize interferences when
traffic Is heavy on the far side of
the intersection

Allow passengers to access
buses closest to the crosswalk.

Results in the width of the
intersection being available for
the driver to pull away from curb.
Eliminates the potential of double
stopping

Allow passengers to broad and
alight while the bus is stopped at
the red light

Provides drivers with the
opportunity to look for oncoming
traffic, including other buses with
potential passengers.

Increases conflicts with right-tuming
vehicles

May result' In stopped buses
obscuring curbside traffic control
devices and crossing pedestrians
May cause sight distance to be
obscured for cross vehicles stopped
to the right of the bus

May block the through lane during
peak period with queuing buses
Increases sight distance problems
for crossing pedestrians.

 

Mid-Block Bus Stop

 

 

Minimizes sight distance
problems for vehicles and
pedestrians.

May result' In passenger waiting
areas experiencing less
pedestrian congestion.

 

Require additional distance for no-
parking restrictions.

Encourage patrons to cross street
at midblock (jaywalking).

Increases walking distance for
patrons crossing at intersections.

 

Source: Transit Capacity and Quality of Service Manual, 2nd Edition.
TCRP Report 100, ed. T.C.R. Program. 2003, Washington, D. C. .' TRB, National Research Council

41

 

3.2. Variability of Transit Dwell Times

As indicated earlier, it is often recommended that priority systems be only implemented
at intersections with far-side stops or where stops are sufficiently upstream to allow
vehicle detection downstream of the stop and to reduce uncertainties in predicting
intersection bus arrivals created by variable dwell times [19, 23-26]. While such
recommendation may satisfy the operational practices of many agencies, stop relocation

may not always be possible.

Dwell time is defined as the time a bus is immobilized for passengers boarding and
alighting. This only includes the time during which doors are opened. Another delay
related to dwelling activities is the bus stop clearance time. For buses stopping in a traffic
lane, this is the time needed to start up and leave the stop after the doors have closed [21].
For buses stopping in a bus stop, the time spent waiting for a gap of sufficient length in
the adjacent traffic is also included. For signal priority systems, dwell and clearance
times are often conveniently combined into a single operational parameter. Since signal
timing adjustments are primarily concerned with the total delay incurred by vehicles at
bus stops, combining both parameters allow system operators to consider a single

parameter reflecting what they are primarily concerned with.

To illustrate issues associated with dwell time variability consider an approach with a
near-side stop on which buses are detected when 20-s away from the bus stop and stop
line. If buses always dwell for 30-3 at the stop, signal timing alterations based on a
projected stop line arrival 50-3 after detection would always reflect reality. However, as
illustrated by the example of Figure 3.2, dwell times are rarely constant. Here, while a

10-3 dwell time would adequately characterize average bus operations, it would not

42

reflect the fact that individual stop durations vary between 6 and 16 seconds and that
about 58% of all buses were observed to not stop at all, thus creating a 16-3 overall

window of potential arrival times at the downstream intersection.

While the above example does not consider the effects of clearance times, as buses at the
observed stop board and alighted passengers while stopped in a traffic lane, situations in
which buses would have to seek gaps in adjacent traffic should exhibit even greater
variability. According to the TCQSM, a bus needs 10-3 to start up and clear a stop after
closing its doors. For buses stopping in a bus stop, field observations further indicate that
the average reentry delay can vary between 1 and 15 seconds, depending on the traffic
volume on the adjacent lane. Consideration of these statistics would thus clearly add to

the potential variability of total bus dwell times as considered by TSP systems.

 

 

 

 

 

 

 

 

 

 

60W — rrrrrrrrr - , -
50 —— - e —————————————————————————————————— E Empirical distribution —————————
I: Non stopping buses
40 __ r _______________________________ —o— LogNormal (2.253, 0.244) _________
5‘ 130 Total number of buses
g 58 Buses stopped (45%)
$.30" ' ””" ”’ 728usesdldnotstop(55%) "'
IL
20 _ ,,,,,, E L 2 E V
10 - 77777777777777
0 , ,-,----,-,-,-,
0 18 20 22 24
Dwell Time (seconds)

 

 

 

Figure 3.2 Observed dwell time distribution at a bus stop in Kitchener, Ontario,
Canada.

While several studies have examined the performance of transit signal priority systems
under various geometric, traffic, and transit conditions [17-20], these studies have

43

primarily focused on intersections with far-side stops or where detection can occur
downstream of the stop. Although it is possible to use an average dwell time, this rrright
have a significant impact on evaluating the proposed traffic signal controller [27, 28], and

further need is required to predict dwell times at bus stops.

Among the few studies that have considered near-side stops, a significant focus has been
on investigating bus arrival prediction techniques. Research examples include the use of
linear and nonlinear regression models to predict dwell times based on the number of
boarding and alighting passengers, the average headway between buses, schedule
adherence, the number of doors on the vehicle, the fare collection method, and bus type
[29, 30]. In other efforts, ANN-based models have also been proposed for predicting bus
arrival times at intersections through improved consideration of historical link travel time
and bus stops demand [31]. However, these models depend mainly on historical data to
develop a prediction model, which will reduce their efficiency in real-time predictions,
especially where a highly accurate and reliable dwell time prediction is required.
Furthermore, the transit priority system and traffic control performance highly depends
on the accuracy of dwell time prediction model, and its ability to go beyond data used to

develop the prediction model [27, 28].

3.3. TSP Within Coordinated Traffic Signal Control Systems

Implementing TSP at an isolated intersection or at intersections that are part of a
coordinated network presents different difficulty levels. At an isolated intersection, there
is no consideration for the need to maintain traffic progression patterns with adjacent
intersections. In such a case, there is great flexibility in selecting which signal timing

alterations can be implemented. However, a majority of TSP applications that have been

44

considered in recent years involve deployments along coordinated arterials. In these
cases, some constraints must be considered when determining the best priority actions to

implement in order to maintain decent or specific traffic progression patterns.

From a network point of view, there are generally two different approaches to providing
TSP. In its simplest form, TSP is provided one intersection at a time, without
coordination constraints. Following the detection of an approaching vehicle, the signal
controller in charge of the intersection at the end of the approach would consider
possibilities for providing priority to the approaching vehicle at the intersection under its
control without factoring potential effects on traffic patterns or priority opportunities at
intersections further downstream. Most TSP implementations are currently of this type
[32]. A particular advantage of this approach is the removal of the need to track vehicles
across multiple intersections, which greatly simplify system implementation, particularly
if vehicles are not already equipped with automated vehicle location systems. Handling
priority on an intersection-by-intersection basis also removes the need to develop

procedures for predicting vehicle arrivals at successive intersections.

In a network setting, however, intersections are often linked together through the
imposition of signal coordination constraints. In such systems, signal timings at one
intersection are function of the timings at adjacent intersections and determined not
necessarily to simply minimize stops and delay at the local intersection but also to
minimize stops and delays across a succession of intersections through the provision of
smooth traffic progression'pattems. The primary difficulty for TSP systems in such a
control environment is how to provide preferential treatments while maintaining signal

coordination and effective traffic progression patterns.

45

TSP issues within coordinated systems are not only linked to a potential for reduced
green time along main travel directions. Priority actions granted to buses traveling along
an arterial with the main flow of traffic could for instance significantly alter established
progression patterns by allowing traffic to be released earlier from an intersection. While
this may reduce delays at the intersection in question, it may also result in vehicles
reaching the following intersections earlier than expected. Since vehicles are reaching
downstream intersection earlier than expected, they could be forced to stop again, causing
additional delays at this intersection and potentially contributing to increasing queue

lengths and side-street blockage.

The potential for altering traffic progression patterns introduces a need to consider
operational constraints limiting the magnitude or type of signal timing alterations that can
be implemented. As an example, some transit agencies limit for instance green extensions
or early recalls to 5 seconds in order to minimize impacts on traffic progression patterns.
While such a constraint may seem arbitrary, its imposition is often the results of an
inability to quickly assess the traffic impacts that may arise at downstream intersections
within a coordinated system following priority decisions taken at a particular intersection.
In an ideal system, constraints to impose on priority actions should be tailored to the
actual potentials for traffic impacts, being more or less relaxed depending on the

situation.

3.4. Simultaneous TSP Requests

A particular difficulty with current TSP systems is what to do when two or more buses
are approaching an intersection simultaneously. In most cases, TSP systems would grant

priority on a first-come-first-serve basis. This is a carryover of past years, when

46

technology did not allow tracking vehicles or obtaining real-time information about
ridership or schedule adherence. Technological advances now allow controllers to
consider these elements, and thus, to theoretically provide priority to the bus yielding the
highest benefits, such as to the bus with the highest ridership or the bus that is the furthest
behind schedule. However, for these criteria to be effectively considered, a system must
exist to forecast needs over the next few seconds or cycle to allow optimum decisions to
be made and avoid situations in which a switch to a more efficient action is prevented by

constraints associated with earlier decisions.

3.5. Summary

This chapter focuses mainly on the issues that might happen when TSP is implemented.
One of the issues is the location of bus stops. Placing the bus stops on the far-side of an
intersection allow the controller to ignore variations in dwell time and to detect transit
vehicles from a certain distance from the intersection. One potential problem is however
the need to accurately predict bus travel time between the detection point and intersection
stop line in mixed traffic. Placing the bus stops on the near-side of an intersection reduces
the travel time variation but introduces the need to consider dwell time variations. Due to
specific operational factors, transit agencies may be required to use both types of bus
stops in order to reach their service objectives, such as having a bus stop to be a transfer

point to other transit routes, or minimizing passenger’s walking distance.

The uncertainty in predicting dwell times further affects the controller ability to predict
when a transit vehicle is expected to enter and leave its target intersection. As outlined

above, this issue is more serious at intersections with a near-side bus stop. Such an

47

intersection will leave a limited timeframe to the controller to implement a successful

TSP strategy.

Another potential problem regarding TSP implementation is the provision of preferential
treatment in coordinated traffic signal control systems. While many traffic signal control
systems are designed to consider network-based traffic needs and perform signal
optimization by considering the collective traffic needs within a given area, TSP systems
often only focus on providing preferential benefits to buses at individual intersections.
This lack of coordination between transit and traffic operations needs may lead to
decreased benefits from the implementation of TPS systems, either through increased
traffic stops and delays as a result of loss of coordination or through the imposition of
constraints unduly limiting signal timing alteration options for granting priority to buses

[33].

Finally, the inability to handle simultaneous and near-simultaneous priority requests
might prevent a TSP system from implementing truly optimal signal timing alterations.
The particular significant of this problem is that as TSP systems are increasingly being
deployed in urban areas, there is an increasing likelihood that the desire the priority
multiple transit routes within a given area will lead to conflict situations at intersections

crossed by multiple routes.

48

CHAPTER M.

RESEARCH OBJECTIVES AND
APPROACHES

 

his chapter briefly outlines the objectives at the base of the research project detailed

in this dissertation. The overall objective of the research is to seek ways to improve
TSP operations within coordinated arterial networks, more specifically with respect to the
various technical issues presented in Chapter 3. This chapter highlights the objective of
this research in details, introduces the optimization and prediction methods that will be
utilized for the development of the proposed TSP controller. The need for the use of a
microscopic simulation model to act as a virtual testbed for the development and

evaluation of the proposed traffic signal control algorithm is also presented.

4.1. Research Objectives

The objective of this research is to be achieved through the development of Dynamic
_S_ignal Priority Optimization in Real-Time Traffic Control System, (D-SPORT), which is

a traffic signal control that integrates TSP within the real-time signal control in a

49

coordinated traffic network. The development of the D-SPORT controller is to overcome
the challenges associated with implementing TSP in the coordinated system. These

challenges include:

0 Need for predicting bus intersection arrival times that is insensitive to the bus stop
type that is used, particularly those with near—side stops or where bus dwell times
may be highly variable. This need is driven by the fact that accurate information

about the expected bus arrival time is a key to provide efficient TSP control.

0 Need to minimize impacts on surrounding traffic performance of surrounding

intersections when providing preferential treatment to buses.

0 Need to efficiently handle situations in which conflicting priority requests may be

received from more than one approach.

The D-SPORT controller consists of prediction and optimization processes. While the
prediction process is used to provide the D-SPORT controller with the ability to keep
track of the location of transit vehicles in the near future, the optimization process is used
to find a network-wide signal timings to accommodate both transit and non-transit traffic

throughout the network.

4.2. Research Approach

To develop an efficient TSP decision model that has the potential to address all of the
problems discussed earlier, it is proposed to develop control algorithms to leverage novel
emerging optimization techniques. For this task, genetic algorithms and artificial neural

networks are used to develop a real-time traffic signal control and TSP integration model

50

that has the capability to be efficiently applied in a variety of control situations.
Application of these techniques not only requires developing a fundamental
understanding of their operations, but also how a traffic signal control problem can be
translated into an appropriate model suitable for each technique. This requires not only
identifying the variables to consider, such as bus detection times, traffic volumes and
expected bus dwell times, but also identifying how these variables can be used to

generate efficient traffic signal timing control decisions.

4.2.1. Traffic Signal Optimization Methods

As indicated in the problem definition, this research aims to develop a traffic signal
optimization methodology to find near-optimal signal timings at a coordinated network of
signalized intersections. Optimization can be defined as the process by which a “best”
solution to a problem under analysis can be found by means of application of scientific
methods and tools [34]. The following subsections present a quick review of available
optimization methods and a justification for the selection of the Genetic Algorithm
approach as the traffic signal optimization tool for the development of the D-SPORT

controller.

Calculus-Based Optimization

Calculus-based optimization relies on purely mathematical techniques to find the largest
or the smallest value a function would take in a specified interval. This optimization
method requires the development of mathematical formulation replicating as close as
possible the real process that is to be Optimized. As such, calculus-based optimizations
are only valid as long as it can be certified that the mathematical model used in the

optimization is adequately representing reality. This has been a significant problem in the

51

application of such an optimization technique to traffic control problems, as traffic
behavior is a highly complex and dynamic process that is difficult to represent by simple
mathematical equations. Other limitations of this method include the assumption that the
function used in the optimization should be continuously differentiable. Furthermore, this
method is unable to handle the fact that the control variables are restricted to a set of

control functions that are used to validate the control variables.

Dynamic Programming

Dynamic Programming (DP) deals with problems involving a finite number of stages but
in which the system being optimized may be in several possible states. The general aim
of DP is to determine an optimal policy that minimizes the total costs in all the stages. A
problem is solved recursively by finding, at each decision point, an action that rrrinirnizes
the sum of the current cost and the best future costs. A typical example of DP application
is to solve a sequential decision problem such as finding the shortest route to reach a

specific destination [35].

DP can deal with discrete variables, non-convex, non-continuous and non-differentiable
functions. It is considered as a powerful optimization tool due to its ability to take into
account the stochastic variability of the objective function. However, high computational
requirements could reduce its efficiency in optimization problems requiring solutions in
real-time. These computational problems are mainly related to the storage of information,
especially when there is a large set of decision variables involved in the optimization

problem.

52

Simulated Annealing

Simulated Annealing (SA) is a type of global heuristic search involving some random
element in the way the algorithm proceeds [36]. SA is based on an analogy with a highly
successful Monte Carlo simulation model for the physical annealing process of finding
low energy states of a solid. Physical annealing is the process of finding the ground state
of a solid, which corresponds to the minimum energy configuration by initially melting
the substance and then lowering the temperature slowly, spending long times at
temperatures close to the freezing point. In simulations of physical annealing processes,
the system is given a small temperature displacement in any iteration before the resulting
change in system energy is calculated. If the change results in lower energy, it is
accepted. If not, it is accepted with a probability that is a function of the energy change

and the temperature.

By analogy, the different solutions of an optimization problem can be seen as
corresponding to the different states of a physical substance, and the cost function to be
minimized as the energy of the system. Since the concept of temperature has no obvious
equivalent in optimization problems, it is converted into a control parameter having the
same units as the optimization function. In such a case, the search process will start at a
given high “temperature” and the temperature is lowered every iteration. At each
temperature, a certain number of changes will be considered. An acceptance function that
defines the probability of accepting a change in system status that result in a higher
objective function value will be used to accept or reject the change. Acceptance functions

are generally defined so that the probability of accepting a change triggering an increase

53

in objective function will be high at high temperatures but close to none when the

temperature approaches zero.

One of the main advantages of the simulated annealing approach is that it avoids getting
trapped at a local optimum by sometimes accepting a neighborhood move that increases
the value of the objective function. However, one of the disadvantages is that it does not
work in parallel, which means that the algorithm should finish the first stage completely

before going to a second one.

Genetic Algorithm

A genetic algorithm (GA) is a heuristic near optimal solution search method based on
natural genetics and mechanisms of natural selection [37]. Genetic algorithms are a
particular class of evolutionary algorithms that use techniques inspired by evolutionary
biology such as inheritance, mutation, selection, and crossover. As in biological
evolution, organisms in a certain population have different sets of genes. Each gene
corresponds to a decision variable of the optimization problem. As such, a set of genes
held by an organism can thus be seen as representing a potential solution to the

optimization problem at hand.

GA determines how well an individual’s decision variables solve the problem at hand by
evaluating a cost function (or objective function). Individuals with lower cost values (or
higher fitness values) are seen as representing better solutions. The idea behind the
genetic algorithm approach is to use an existing population of organisms to iteratively
generate new organisms that may exhibit lower cost (or higher fitness). This

improvement process is executed through the application of a series of genetic operators

54

creating a new population of organisms on the basis of an existing population. The goal
of this evolutionary approach is to produce diversity within the population and to explore
various gene combinations or solutions to the optimization problem. However, this is not
a blind search as the probability of reproduction of a specific organism is related to its
cost function value. In other words, better parents are generally made to produce more
children, thus theoretically increasing the chances of creating new organisms with lower

cost function values.

Typical genetic operators used in GA include selection, crossover and mutation. While
other operations may also be used, the three mentioned are the most commonly used. The

functions performed by these operators are described in more details below.

Selection

Selection is a process in which individual strings (chromosomes) are copied into the new
population according to their cost function (or objective function) value. Strings with a
lower cost function value have higher probability of contributing one or more offspring in
the next generation. The selection process is probabilistic. The probability for selection is
based on the chromosome normalized cost function value relative to the normalized cost
function values of all of individuals in the population. Selection begins by determining a
chromosome’s relative cost function value. This is done by dividing its cost function
value by the sum of all the cost function values for all the chromosomes in the
population. Then, a random number generator is used to select chromosomes for the
crossover phase. The probability for a chromosome to be selected during each spin of the
random number generator is equal to the chromosome’s relative cost function value. The

number of chromosomes selected is equal to population size, therefore, keeping the size

55

constant for every generation. Some chromosomes may also be selected more than once,
resulting in multiple copies of this chromosome to be used by the crossover operation

[38].

Crossover

Crossover is the process of combining the genes of one chromosome with those of
another to create offspring inheriting traits of both parents. The crossover rate is the
likelihood of a chromosome being selected for the crossover operation. The
chromosomes that are not selected will not have their genes changed before proceeding to
the mutation phase. Those that are chosen will be paired with another chromosome that
was also selected for crossover (a mate). From each pair, two offspring will be created to
replace their parents. For classical one-point crossover, to determine which genes are
inherited from parents, a random number between one and the total number of genes
minus one is generated. For the first offspring, the genes numbered between one and the
random number will be inherited from the first chromosome, and the genes numbered
between the random number plus one and the maximum number of genes will come from
the second chromosome. The genes for the second offspring will be inherited just like
those of the first offspring except that the genes that came from the first chromosome in
the first offspring will come from the second chromosome and those inherited from the

second chromosome will come from the first chromosome [38].

Mutation

Just as in nature, some chromosomes will have random mutations occur in their genes.
The mutation rate specifies the odds that a given gene in a chromosome will be mutated.

If a gene is selected for mutation then its value will be changed. In the case of binary

56

representation, the gene will simply be flipped (i.e., gene with a value of I changed to 0,

or a gene with a value of O changed to 1 [38].

4.2.2. Optimization Method Selection

In this research, genetic algorithm optimization technique is used to optimize the cost
function. The rationale behind this selection is due to the flexibility of representing a cost
function using Genetic Algorithms and the ability of GAs to optimize complex functions.
The GA allows more flexibility in formulating extremely complicated cost functions than
other optimization methods. This flexibility includes ability to represent cost functions
with discrete or continuous variables. Since the GA is only interested in the cost function
value for a string solution, its performance is superior in avoiding getting trapped in local
minima. Furthermore, the GA has the ability to deal with large computational problems
since it can be run in parallel, and hence, the required computational time can be
significantly reduced. For the problem in hand, if three cycles were optimized, there are
5.4566x1067 possible solutions, which required a long amount of time to find the optimal
solution, where the GA can provide a near-optimal solution within an acceptable period

of time.

4.2.3. Transit Dwelling Activity Prediction Model

This research requires developing a prediction tool for providing reasonably reliable
information about the time elapsed between the moment a bus is first enters the bus stop
and the moment it is ready to leave the bus stop, where dwell time activities may take
place and result in variation around its average value. This section describes the possible
approaches that can be adopted to develop the prediction tool and the rationale behind the

selection of an ANN prediction method for the purpose of this research.

57

Regression Analysis

Regression Analysis is a statistical approach of modeling the relationships among three or
more variables. It is used to predict the value of one variable (known as the criterion or
dependent variable) given the values of others variables (known as predictors or
independent variables). In this case, it is believed that variations in the values of the
predictor variables cause or agree with variations in the values of the criterion variable.
The objective of the regression is to evaluate the coefficients of an equation that is used
to predict values of the criterion variables based on a specific combination of one or more
predictor variables. The regression analysis assumes that the error, which is the difference
between the predicted and measured values of the criterion variables are independent or
each other, have a zero mean, have a constant variance across all values of the predictor
variables, and are normally distributed. The violation of any of those assumptions leads

to incorrect model structure [39].

In most cases, the equation that relates criterion and predictor variable to each other
should be formulated first. This potentially limit the usefulness of this prediction method
as results will be dependent on a priori choices regarding the set of predictor variables to
consider [39]. This may create significant difficulties in problem considering large

number of variables in which not only linear combinations are possible.

Artificial Neural Networks

Artificial Neural Networks, (ANN s) are non-linear statistical data modeling tools that are
used to model complex relationships between input and output data or to find patterns in
datasets. Instead of following a set of instructions to solve a problem or determine an

outcome, ANNs take a non-algorithmic approach to problem solving by attempting to

58

process information in a way similar to how the human brain operates. Figure 4.]
illustrates a typical ANN framework for a multilayer perceptron model [40]. ANNs have
been used in many transportation-related applications, such as travel time prediction,
crash rate modeling, and pavement distress prediction [41-43]. As illustrated, ANNs
typically consist of three layers. Similar to the human brain, each layer is composed of a
large number of highly interconnected processing elements called neurons. These

neurons work in parallel to solve a specific problem or make a prediction [40, 44].

Within the framework of Figure 4.1, an outcome is generated from a series of input
patterns X1...X,» subjected to transformation using functional information contained within
the neurons and the communication channels (links). For the ANN supervised learning
process, the target, or correct output has to be known. The goal of this type of network is
then to correctly map the input X1...X,- to the corresponding output. The training takes
place through repeated applications of all training input patterns and adjusting the
strengths of the connections until the network’s outputs approximately matches the
correct or target outputs. In that sense, ANNs learns from examples, as children learn to

distinguish dogs from cats based on examples of dogs and cats.

Since ANNs learn by example, the training examples must be selected carefully. Proper
training is important, as well-trained ANNs may exhibit some capability for
generalization beyond the training data, such as producing approximately correct results

for new cases that were not used for training.

59

 

 

 

 

 

Input Layer Hidden Layer Output Layer
ilnputs S Neurons One Target
W11 f
X1‘ n1 ———r
/7 ’ XI: Input numberi
X2 b1
1 a: Target or correct output
X3 \ n3: Neuron number s
: > n2 , f F \7 f: Transfer Function
' LC l : l . .
: l b '—‘ b.: BIas In neuron s
l
2
l 1 E W“: Adjusted connection
E j : weight between input iand
: j : neuron s
I l I
I w I ‘
l ’2 7 4
l .
l S n. :L' BF
1 1:.
xi ’5 1

 

 

 

 

 

 

 

 

 

 

 

Figure 4.1 Typical architecture of a one-hidden-layer ANN with i input elements
and s neurons follow.

Prediction Method Selection

For this research, the ANN model will be used as the basis for predicting the transit time
elapsed between entering the bus stop and the time it is ready to leave the bus stop. This
ANN prediction model is integrated within the D-SPORT controller. The selection of the
ANN to perform the prediction can be explained by the fact that, unlike univariate and
multivariate regression models, ANN are particularly well suited for approximating
complex functional relationships, can handle noisy data, do not require predefined

formulas or equations, and can handle intercorrelated predictor variables.

4.2.4. Simulation Environment
Use of computer simulation models is very common in the transportation field to analyze,
assess, and evaluate different transportation systems and their components. In the world

60

of traffic, simulation can be defined as a numerical technique for the conduct of
experiments on digital computers. These simulation models may include stochastic
characteristics, be microscopic, mesoscopic, or macroscopic in nature, and involve
mathematical models that describe the behavior of a transportation system over extended
periods of real-time [45]. The nature of this research requires evaluation of the
operational effectiveness of traffic signal control algorithms in a variety of traffic
situations. One approach would have been to evaluate the effectiveness of the proposed
model through a field test. However, such an approach would not only require the
purchase and installation of a variety of necessary and costly equipment, but may also
limit the number of scenarios that could possibly be considered. Because of this, a need
was identified for using a microscopic simulation environment to help develop and
evaluate the proposed TSP algorithms. In this case, use of a simulation would allow one
to quickly test the effectiveness of proposed algorithms, in addition to allow a wide

variety of traffic conditions to be considered.

There currently exist a variety of microscopic traffic simulation packages used by traffic
engineers to evaluate traffic signal control systems. Among the most commonly used are
CORSIM [46], AIMSUN/2 [47], PARAMICS [48], and VISSIIVI [49]. Each simulation
package has its own advantages and disadvantages. For the current research, a model
allowing its user to define custom traffic signal control functions is needed. This
immediately removes CORSIM from the list as this model does not allow its user to
specify custom controller logic. AIMSUN/2 allows the user to develop a built-in traffic
controller within the simulation package itself, however, the user is unable to attach an

external controller within the simulation environment. PARAMICS and VISSIM are the

61

only known simulation packages that allows the user to use an external controller
developed using specific programming languages such as C, C++, Visual Basic and
FORTRAN 90/95 [50]. In both cases, the user can use either the compiler provided with
the simulation packages or Dynamic-Link Library (DLL), which is a collection of
executable functions or data that could be shared between applications and are executed

when needed.

A literature review has shown that VISSIM has been widely used to assess and evaluate
traffic controllers implementing TSP algorithms [19, 51]. Based on this fact and the
model’s modeling capability, VISSIM was therefore chosen as the model that will be
used in this research to assist with the development and evaluation of the proposed TSP
model. Selection of this software is based more particularly on its ability to provide

detailed modeling of transit operations and priority systems, including:

Ability to model transit dwell times as a normal distribution with customizable
parameters, a user-defined empirical distribution, or a function of the number of
passengers waiting to board and alight at a particular stop.

0 Provision of transit signal priority functions similar to those found in the
real-world with the use of the National Electrical Manufacturers Association
(NEMA) controller modeling.

0 Ability to develop custom vehicle-actuated traffic signal control logic.

0 Possibility to assign detailed priority rules at intersections and around bus stops.

0 Ability to view the bus detections and resulting signal timing alterations.

0 Ability to implement external controller within the simulation using the DLL.

62

4.2.5. Overall Research Approach

The proposed controller will use genetic algorithms to search for a near-optimal signal
timing solution considering prevailing traffic conditions and fluctuations. As schematized
in Figure 4.2, VISSIM will be set up to provide the traffic signal controller with current
traffic and transit data in a real-time manner. This information will also be passed to the
ANN prediction model, which will use it to predict transit dwell times and provide the
result of this estimation to the controller. The controller will then use all available data to
determine near optimum phase splits, cycle length, and offsets with adjacent intersections

satisfying as best as possible both the transit and general traffic needs. Results of the

optimizations could then finally be passed to the simulation model to test their impacts on

transit and traffic operations.

 

/

’l

 

 

 

Start VISSIM-COM
Interface

 

 

Y

 

 

Call VISSIM Simulation
Environment

 

 

Y

FOR Sim = 1, Simulation
Period

Time to
call GA?

NO
1

 

 

Continue with the
Simulation

 

 

 

 

 

 

 

i

1 YES

SAVE Simulated Traffic Conditions
SAVE Signal Control Timings
SAVE Transit Related Data

Y

 

Call GA Optimization
Platform

 

 

 

 

 

Y
For Each
Intersection in
Network

Y
Save New Signal

 

 

 

 

>{ For Each Generatiow ,

 

Y
For Each String in ,_
Genertion

Y

 

 

 

Call ANN

 

 

 

 

Y

 

 

Calculate Fitness
Find Next Generation

 

 

 

 

 

 

 

Timing v
v Parameters Return Signal Timings
End
Figure 4.2 Schematic sketch for the D-SPORT controller element integration

63

 

To achieve the research objectives, the following sequence of research activities was

proposed:

Development of an artificial test network in the VISSIM microscopic traffic
simulation environment is used for evaluating the D-SPORT controller. VISSIM
is responsible for performing the traffic simulation, collecting and summarizing
the simulation results, and making these results available to the D-SPORT
controller upon request.

Development of the D-SPORT controller cost function that addresses the issues
discussed in Chapter 3. This work involves in delay and number of stops
modeling for transit and non-transit vehicles travel throughout the traffic network,
transit travel time modeling to predicting transit vehicle locations in the near
future transit arrival time at bus stops.

The use of an artificial neural network modeling approach to develop a tool for
predicting bus arrival times at signalized intersections that can consider situations
in which bus arrival times may be subject to significant variability due to either
bus stop location or variability in bus dwell times.

The use of a genetic algorithm optirrrization model to optimize the D-SPORT
controller cost function (which includes the ANN model) and to return the near-
optimal signal timings that can be applied at the signalized intersections operated
in coordinated networks.

Integration of the D-SPORT controller within the VISSIM microscopic traffic

simulation model to be used as a research simulation environment using the

64

VISSIM-COM interface to allow the D-SPORT to communicate with the
simulation environment.

Utilization of the microscopic simulation model to validate the transit travel time
prediction model proposed under the D-SPORT controller cost function, and to
quantify the potential benefits of the D-SPORT controller with respect to

providing TSP using current state-of-the-art techniques.

65

CHAPTER V.

DWELL TIME IMPACT ON TSP
OPERATIONS

 

his chapter reports on a simulation study that was done to investigate the
performance of a signal control system offering priority to buses approaching an
intersection with a near-side stop under various traffic conditions. The objective of this
work was to examine the potential impacts of dwell time variability on priority system
performance and assess the importance of these impacts for the development of effective

TSP control.

5.1. Simulation Modeling

This section describes the VISSIM simulation model that was developed to evaluate the
impacts of dwell time variability on the performance of traditional active TSP systems
offering red truncation (RT) or early green recall, and green extensions (GE) based on
detection of approaching transit vehicles. The following subsections detail more

specifically the geometry of the test network, the traffic signal control assumptions, the

66

modeling of traffic and transit demands, the simulated TSP logic, the evaluation setup

and the performance measures considered.

Test network

The test network is illustrated in Figure 5.1. This network includes two intersections
located 305 m (1000 ft) apart and linked together by an arterial segment featuring two
lanes per direction, ideal lane widths, no grade, no curb parking, left—tum bays, and a
speed limit of 65 km/h (40 mph). The cross-streets feature similar characteristics, except
for shared through-left lanes. Within this network, priority is to be provided exclusively
to buses reaching Intersection 1 from the west. Intersection 2 is modeled to generate

realistic platooned arrivals on the eastbound (EB) approach to Intersection l.

 

    
 
 

Check ln ; :, Signal Head

Detectors {

 

Bus Stop 5;

 

 

 

 

 

Figure 5.1 Simulated traffic network geometric layout
Traffic Demand
To remove truck effects, only passenger cars are simulated. Vehicles are generated at the
upstream end of each entering link following Poisson distribution and VISSIM’s default

vehicle characteristics. Depending on the scenario, traffic volumes are set to produce

67

equal volume-to-capacity (v/c) ratios of 0.50, 0.70, 0.80, 0.85, 0.90, and 0.95 on both the
arterial and cross-street approaches. In each case, the appropriate volumes to assign to

each approach were determined using the SYNCHRO 6 signal optimization software.

On arterial approaches with exclusive turning lanes, right and left turn flows are further
assigned so as to maintain a We ratio of 0.50 on each turning lane. For the cross-streets,
the left-tum and right-tum flows were simply assumed to correspond to 5% and 10%

respectively of the total approach volume.

Transit Service

Buses are assumed to travel only in the eastbound (EB) direction and serve a single stop
located in the right-tum bay of the eastbound approach to Intersection 1 with a normally
distributed frequency of one vehicle every S-min on average, and l-rnin standard
deviation. Both general traffic and buses share a desired 65 km/h (40 mph) travel speed.
Transit vehicles entrance frequency, the interactions with other vehicles and traffic
signals, in addition to dwell time variability, create fluctuations in the time individual
buses take to reach Intersection 2. Since not all buses arrive at the same moment within
the signal cycle, not all vehicles require a priority treatment. In the evaluated scenarios,
12 to 20 of the 23 buses generated over a 2-hour simulation end up requesting priority

actions.

For the study, it was determined that a bus stop located at the intersection stop line would
provide the most difficult priority case as such a stop would provide virtually no advance
warning of the readiness of a bus to enter an intersection once boarding is completed.

Starting from this premise, a stop located on an exclusive right-tum lane was then

68

selected to minimize the impacts of dwelling buses on through traffic. Such a placement
requires buses to wait for any queued vehicles on the right-tum lane to be cleared before
being able to reach the stop. A similar situation would have existed were buses required
to stop on a shared through/right-tum lane, expect that buses would in this case be
waiting within the queue and not on an adjacent lane. To minimize the impacts of gap-
seeking delays when leaving the bus stop, all buses are finally assumed to have priority of
movement over the adjacent traffic, which is an assumption that can be regulated in the
real world. While this facilitates the bus stop departure process, it does not completely
eliminate the need to seek gaps in the adjacent lane. For instance, buses still have to wait

for queues of vehicles on the adjacent lane to dissipate.

Scenarios featuring either stochastic or deterministic dwell times are considered. In the
deterministic scenarios, every bus is assumed to dwell for exactly n seconds. In the
stochastic scenarios, dwell time is instead a function of the number of individuals waiting
to board a bus at the time of its arrival. Here, boarding demand is modeled by simulating
the process of passenger arrivals at the bus stop. Depending on the scenario, a constant
arrival rate ranging between 25 and 150 passenger/hr is considered, with an increment of
25 passenger/hr. Individual bus dwell times are then determined by assuming that 3-3 are
required for each passenger waiting to board, with another 3-3 for door opening and
closing. Bus occupancies were randomly assigned to each bus individually, and
passengers are alighting in a rate of 10% of bus occupancy. Dwell time variations thus
result from randomness in passenger arrivals at the bus stop combined with the effects of

early/late bus arrivals.

69

Additional delays due to the need to wait for queued vehicles to clear the right turn lane
and to wait for acceptable gaps in the adjacent traffic are not included in the dwell time
estimates as such delays are more closely associated with traffic interactions than dwell
activities. Ideally, however, such delays should be accounted for in algorithms used to

predict transit travel time throughout the traffic network.

Traffic Signal Control

All traffic signals are controlled on a fixed-time basis. Each intersection is controlled
with a 90-3 cycle and two phases implementing a 50:50 green split between the east-west
and north-south travel directions. For each phase, clearance intervals are modeled by 4-3
amber followed by a 1-s all-red. A 20-3 minimum green is further imposed to ensure that
rrrinimum pedestrian crossing times are not jeopardized by priority actions. Signal
coordination between the two intersections is obtained from optimizing each experiment
using SYNCHRO 6, a software package for modeling and optimizing traffic signal
timings [52] by taking into account the traffic volume and associated with each

experiment.

Priority Setup

Transit movements are monitored through two detectors. A first detector is placed
upstream of the stop line to detect approaching buses when they are 3-3 away from the
stop line in situations in which their progression is not impeded by other traffic or a need
to serve boarding/alighting passengers. The second detector is placed just downstream of

the stop line to determine when a bus has entered the intersection and no longer requires

priority.

70

Priority actions are provided using Vehicle Actuated Programming, (VAP), which is a

VISSIM’s programming module. This module allows the building of custom controller

functions based on information provided by modeled traffic detectors. As indicated by

the following rules, the simulated TSP logic is designed to provide green extensions and

early recalls based on the expected stop line arrival time of detected buses, a maximum

allowed green extension, and minimum green constraints:

For the fixed dwell time scenarios, the signal controller predicts stop line bus
arrivals using a fixed dwell duration corresponding to the actual vehicle dwell
time.

For the stochastic dwell time scenarios, the signal controller is not assumed to
know the fixed dwell time of each bus. Priority needs are assessed by simply
considering an average dwell duration based on the average number of boarding
passenger expected at the bus stop under a perfectly regular service.

No priority action is taken for projected arrivals during the normal green.

For projected arrivals during the red, green extensions are granted only if buses
are arriving close to the end of the green, within the maximum allowed extension
window.

Red truncations (RT), or early recalls are considered for projected arrivals during
the red beyond the maximum green extension window, but only for arrivals close
enough to the end of the red. For a maximum extension of 10 s, an early recall
will for instance be considered only for arrivals within 10-3 of the end of red.
Green extensions (GE) and early recalls are granted in increments of 5-3, with a

maximum of 5 to 20 seconds depending on the scenario.

71

0 Any granted extension is terminated if a bus is detected to have crossed the stop
line.

0 Each bus is allowed only one priority request. A bus failing to cross the stop line
in time to benefit from a green extension will thus not be considered for an early

recall.

Evaluation Setup

Evaluations are based on two-hour simulations. For each simulation, the first 5 min is
neglected to allow network loading. Past this interval, traffic data are collected in five-
rrrinute intervals, while transit data are collected for individual buses. To account for the
stochastic nature of VISSIM’s simulation processes, 10 replications are made of each
scenario using different random number seeds. Unless otherwise noted, all results

reported are thus for an average of 10 replications.

Figure 5.2 summarizes all the possible experiment combinations that are examined under
each priority setup. The total of 360 scenarios are tested, 180 are assigned to the fixed
dwell time priority setup, and 180 for priority setup with variable dwell time. The
differences between these scenarios are by combining volume-to-capacity ratios along the
major street (i.e., the Eastbound (EB) and Westbound (WB) directions), with maximum
allowable signal alterations for green extensions and early green recalls, and the average

passenger demands at the bus stop.

72

 

 

 

 

 

 

 

 

 

 

Figure 5.2 Sketch summarizes the scenarios that are examined under fixed and
variable dwell time priority setup

Performance Measures

Two performance measures were retained to assess the impacts of dwell time variability
on transit and traffic operations. The first is the average delay per vehicle. This delay is
calculated by VISSIM as the difference between the actual travel time experienced by a
vehicle and the travel time that would have been observed if the vehicle had been able to
travel at its desired speed without any impedance. The same definition is used for buses,
except that dwell time is excluded from the calculations as this is considered a normal
transit activity. Delays associated with buses slowing down on the approach to a stop and

bus stop clearance times are considered as travel time delays.

The second measure is priority system efficiency. This parameter measures the ability of

buses requesting priority to benefit from the system based on the following criteria:

0 Buses arriving during the normal green are not considered as they do not require

priority.

73

o Buses requesting priority and entering the intersection during a provided green
extension (GB) or early recall (RT) are counted as successes.

o A bus requesting a green extension and arriving too late to benefit from it is
considered as a failure.

0 A bus requesting a green extension but arriving earlier than expected and for
which a priority request cancellation can be issued before the extension is
implemented is considered as a success since the system was able to adjust to the
early arrival.

o The awarding of early green recalls for buses that end up arriving during the
normal green is considered as a success as reduced delays may still result from

early queue dissipation.

At the end of a simulation, the efficiency is determined by the ratio of the number of
buses having benefited from the system over the total number of buses having requested a

priority action.

5.2. Impacts of Dwell Times on Transit Operations

Table 5.1 compares the average bus delays obtained under deterministic and stochastic
dwell times for various combinations of WC ratio, passenger boarding demand, and
maximum limits on green extensions/early recalls. For each combination of We ratio and
passenger boarding rate, it is first observed that allowing longer signal alterations tends to
reduce bus average delays for both the deterministic and stochastic dwell time
assumptions. Transit delay reductions reach up to 40 % when considering a 20—3 limit on

signal alterations with respect to no priority action. This was expected as longer

74

extensions provide more flexibility for accommodating buses. For a given alteration
limit, v/c ratio, and dwell time, it is further observed that increases in the number of
boarding passengers tend to decrease average delays. This is an effect directly associated
to the geometry of the test network and the traffic signal setup for the scenarios
considered. For instance, longer dwell times could push more buses to reach the

prioritized intersection during the normal green, when no delay is incurred.

The results of Table 5.1 indicate that dwell time variability clearly impacts the ability of
buses to benefit from a priority system. As observed, stochastic dwell scenarios produced
lower average bus delays than deterministic scenarios. This is contrary to what was
expected. These delay reductions can be explained by distributing bus arrivals at stop line
under variable dwell times, where the probability of successfully using a granted TSP

treatment.

To illustrate this point, consider the sketch of Figure 5.3. This sketch illustrates two
scenarios of potential bus arrival times. The first scenario illustrates a situation in which
buses arrive in the middle of the red signal after having dwelled for a certain time
upstream of the intersection. In the absence of variability, all buses would reach the
intersection at exactly the same moment, as exemplified here by the thick trajectory line.
This means that all buses would receive the same priority treatment and incur identical
delays. However, as dwell time variability spreads arrivals over a wider window, some
buses would start to arrive close enough to the red initiation to be able to benefit from a
green extension. Buses arriving closer to the end of the red would also incur less delay.
At the system level, these reduced delays due to increased variability of arrival would

then translate into lower average bus delays for the intersection.

75

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Table 5.1 Average transit delay for the scenarios considered, in s/veh
Passenger Fixed Dwell Time (3) Stochastic Dwell Time (s) Difference (7.)
’ Maximum GEIRT Alteration Maximum GEIRT Alteratlon Maximum GEIRT Alteration
H0!" 0 5 10 15 20 0 5 10 15 20 0 5 10 15 20
vlc: 0.50
25 38.0 35.4 31.6 29.2 25.8 33.2 31.3 27.7 25.3 22.5 -12.6 -11.6 ~12.5 -13.6 -12.5
50 32.7 30.2 26.5 23.4 20.5 29.2 26.6 23.3 20.5 18.1 -10.8 -11.9 -12.3 -12.7 -12.0
75 28.6 25.9 22.5 20.1 17.1 26.6 24.0 20.3 18.0 15.9 -6.8 -7.5 -10.0 -10.4 -7.5
100 26.3 21.9 19.5 17.0 15.0 24.3 20.5 18.4 16.5 15.3 -7.5 -6.7 -5.4 -3.1 1.5
125 22.7 20.3 17.7 15.7 15.0 21.6 19.0 17.1 16.0 15.5 -4.8 -6.5 -3.8 1.7 3.3
150 19.7 17.3 15.6 15.4 15.4 20.1 18.3 16.7 16.0 15.9 2.1 5.6 6.9 4.2 3.4
vlc: 0.70
25 40.3 36.7 33.5 30.3 27.1 36.5 32.4 29.2 26.0 23.5 -9.6 -11.6 -12.6 -14.2 -13.4
50 35.4 32.1 28.9 25.5 22.9 33.0 29.5 26.0 22.7 20.9 -6.9 -8.2 -9.9 -10.7 -8.7
75 32.2 29.2 25.2 22.7 19.9 32.0 26.5 23.3 21.4 19.4 -0.7 -9.1 -7.7 -5.6 -2.3
100 32.1 26.5 24.8 23.0 20.9 29.2 25.0 23.2 21.2 19.7 -9.2 -5.7 -6.6 -7.6 -5.8
125 29.5 26.0 24.8 23.0 20.9 26.9 24.5 22.6 23.3 21.3 -8.6 -6.1 -8.9 1.3 1.7
150 25.7 24.7 21.9 21.7 20.5 24.5 22.5 22.6 20.7 21.0 -4.8 -8.9 3.1 -4.5 2.1
vlc: 0.80
25 41.0 38.1 35.7 32.7 30.7 36.7 35.2 31.6 29.6 26.8 -10.7 -7.6 ~11.3 -9.4 -12.7
50 36.5 33.7 30.5 28.4 24.9 35.4 32.3 29.6 25.7 24.1 -3.0 -4.2 -2.9 -9.4 -3.4
75 35.3 30.6 28.0 26.3 23.6 34.8 32.1 26.9 24.8 21.2 -1.6 4.9 -3.8 -5.8 -10.4
100 35.5 32.8 29.0 25.9 23.8 33.6 32.0 27.6 24.7 23.1 -5.2 '2.3 -4.9 -4.7 -2.9
125 34.4 31.9 30.6 27.9 27.9 32.5 30.3 28.1 26.2 24.6 -5.6 -5.0 ~8.2 -6.3 -11.6
150 30.9 28.6 27.4 25.9 25.2 29.0 26.8 28.4 25.0 24.2 -6.3 -6.3 3.8 -3.3 -3.7
We: 0.85
25 42.1 39.1 36.3 33.5 30.3 40.3 35.8 32.4 30.0 27.2 -4.3 -8.5 -10.7 -10.3 -10.2
50 37.8 35.2 32.6 30.0 26.8 38.5 34.7 30.1 26.8 24.8 2.1 -1.3 -7.7 ~10.8 -7.4
75 37.2 34.3 30.6 27.5 26.7 38.9 33.7 30.0 27.3 23.7 4.5 -1.7 -1.8 -0.7 -11.2
100 40.0 36.0 32.3 28.9 28.0 36.6 34.6 31.4 28.1 26.6 -8.4 -3.7 -2.7 -2.8 -5.1
125 37.4 36.3 33.0 30.8 30.2 35.1 33.0 30.5 28.4 28.4 -6.3 -9.1 ~7.5 -7.6 -6.0
150 32.5 31.4 29.4 28.1 27.6 32.7 31.3 28.6 28.8 26.2 0.6 -0.3 -2.8 2.4 -4.9
We: 0.90
25 43.1 40.1 37.0 34.2 31.3 41.4 37.0 33.1 31.6 29.4 -3.8 -7.9 -10.5 -7.5 -6.0
50 39.5 36.1 34.1 30.6 28.5 42.0 35.4 31.3 28.5 26.0 6.3 -1.8 8.3 -6.9 -8.7
75 40.1 34.8 31.3 28.6 26.8 40.1 34.8 31.9 28.7 25.0 0.1 0.1 2.2 0.3 -6.5
100 44.3 35.4 32.8 30.7 28.3 38.9 35.4 31.6 28.5 27.3 -12.2 0.0 -3.4 -7.0 -3.2
125 38.8 37.3 32.2 30.8 28.9 33.9 33.6 31.4 29.9 28.6 -12.6 -10.0 -2.5 -3.1 -1.0
150 37.3 33.9 31.3 30.5 28.6 32.8 31.8 30.6 29.1 27.7 -12.1 -6.0 -2.4 -4.4 -3.1
We: 0.95
25 43.1 40.8 37.9 35.3 32.2 42.4 37.2 34.7 32.2 30.0 -1.8 -8.8 -8.3 -8.8 -7.1
50 41.6 37.4 34.4 30.6 29.2 42.3 37.7 32.0 29.6 27.5 1.8 0.8 -7.2 -3.5 -5.9
75 40.6 35.3 32.9 30.5 27.7 41.5 37.0 32.7 29.4 28.2 2.3 5.0 -0.5 -3.4 2.0
100 43.2 38.5 37.9 32.5 30.7 40.8 37.1 34.5 30.7 29.5 -5.4 -3.8 -9.0 -5.4 -3.8
125 40.1 38.9 35.2 35.6 31.4 37.2 36.8 31.6 29.8 29.0 -7.2 -5.3 -10.5 -16.3 -7.9
150 36.2 33.6 32.2 31.3 29.8 34.0 34.3 31.6 30.2 28.6 -6.2 2.2 -1.8 -3.5 -3.7

76

 

 

(1) Arrival window for early green

SPACE A (2) Arrival window for green extension
2 1 2 1
l—l—l l—l——l

 

 

— No TSP Action

 

 

 

 

 

 

Check-Out
Detector Green Extension
Early Green
1
\ Near—side
_ Bus Stop
Check-In
Detector
\
fir
TIME

(A) Arrival window is primarily within the window to initiate EG. NO TSP needed for early arrival

 

(B) GE most likely to be initiated, possibility for either EG (late arrival) or no TSP need

 

 

Figure 5.3 Likelihood of priority actions with respect to arrival window
Another effect is linked to the number of buses stopping to board passengers. Under fixed
dwell times, all buses are assumed to service the simulated bus stop. However, under
stochastic dwell times buses may skip the stop if no passenger is waiting to board. In
such a case, buses skipping the stop would then tend to arrive during the normal green

and incur no delay, thus reducing the estimated average incurred delays.

Figure 5.4 provides another look at the impacts of varying dwell times on bus delays. The
figure illustrates the bus delays that were obtained for fixed and stochastic dwell times
ranging virtually from 0 to 1.5 the cycle length. For each scenario, the standard deviation
and average of all observed delays are shown. It is first noticed that stochastic dwell
times tend to lead to higher average delays in some scenarios and lower delays in others.
As explained earlier, these effects are due to variations in the proportion of arrivals
during the normal green or at a moment for the application of priority treatment.

77

 

 

 

 

EStochastic dwell time I Fixed dwell time ' WC: 0-70

. . - TSleosflex
I—I Bus delay stdev for Stochastre dwell tIme _ 2-hr sim.
Bus delay stdev for Fixed dwell time _ 23 buses/Sim

60

Cycle1 Welt”
50 - 1
g 40 - : - :
z . ‘
£1 :
CD
0 —_. 1‘ ‘5
3 20— ‘ -

10—

Ol .
sasssaasaaassassaasss
ceasassessaxssssggggg

Average Dwell Tune, (5)

 

 

 

Figure 5.4 Average bus delay with respect to average dwell time and signal cycle

length

A particular observation is that a cyclic pattern appears to exist between the average bus
delay and the dwell duration with respect to the cycle length. Here, priority request
issuance and incurred delays are at a minimum when bus arrivals generally correspond to
the middle of the green signal. As bus arrivals shift towards the end of the green as a
result of longer dwell times, an increasing proportion of buses start to arrive during the
early portion of the red. While green extensions can initially accommodate early red
arrivals, eventually more buses start to arrive past the point for which an extension can be
provided and are thus forced to stop. This causes delay increases as the stopping buses
must either wait for a time when an early green recall can be implemented or the entire
red duration if a recall cannot be granted. Maximum delays occur when the average
arrival time coincides with the beginning of the red signal. As the arrivals further shift

towards the beginning of the green, an increasing proportion of buses start to arrive

78

during the green or close enough to the end of the red to fully benefit from early green
recalls, thus causing reductions in incurred delays. A perfectly symmetric impact cycle is
not necessarily obtained as green extensions and early recalls have different impacts on

delays.

With respect to system efficiency, Figure 5.5 and Figure 5.6 show that dwell time
variability tends to reduce efficiency. While deterministic scenarios produced systems
with 100% efficiency, corresponding stochastic scenarios produced efficiencies varying
between 40% and 95%. This was expected as dwell time variability creates situations in
which granted priority treatments may end up not being used by the buses that had
initially requested them. When no bus has requested a green extension or early green
request during the 2-hour simulation period, the system efficiency is reported as 0% since

the system efficiency calculations are not applicable.

For a given boarding rate, system efficiency does not appear to be significantly affected
by increasing the limit on timing alterations, especially when efficiency is already high.
This is explained by the implementation of green extensions and recalls that are shorter
than the maximum allowed for a number of buses. For these buses, increasing the
alteration limit does not lead to changes in priority actions. For a given alteration limit,
however, system efficiency tends to reduce with increasing boarding rates. This is again
due to a change in the proportions of buses arriving during the green and red portions of

the cycle.

79

Early Green Strategy Performance

Stochastic Dwell Time

Early Green Success Rate. v/c 0.70

 

Fixed Dwell Time

Early Green Success Rate, v/c 0.70

% ‘Koueroms

  

0m 8 s a 8

    

 

 

Maximum 0 Maximum 0 5
Allowable Allowable 10 15 m 1‘2 Bus stop
GE/RT, (s) GE/RT. (5) 2° N volume
perihr

Early Green Success Rate. v/c 0.80

 

Early Green Success Rate, v/c 0.80

   

   

 

-‘ 100 100
80 E 80 n
so '5’: ,so <3
m o
40 .§ 40 i
20 32 g 20 s
o 0
Maximum 0 Maxlmum 0
Allowable 15 m K’ Bus stop Allowable 15 20 u.) 3 Bus stop
GE/RT. (5) 2° N volume GE/RT, (s) N volume
per/hr per/hr
Early Green Success Rate, v/c 0.90

 

Early Green Success Rate, v/e 0.90

Maximum 0
Allowable
GEIRT, (s)

   
 

   

 

Maximum 0 - g3

Allowable " Bus stop

GE/RT, (3) volume
per/hr

 

 

 

 

 

Figure 5.5

* Zero values indicates that no bus was observed

Efficiency of early green strategy under different level of service

80

Green Extension Strategy Performance

 

 
    

   

 

 

 

      
   

   
   

 

 

Stochastic Dwell Time Fixed Dwell Time
Green Extension Success Rate, v/c 0.70 l Green Extension Success Rate, v/c 0.70
100 100
80 m 80 m
3 :s
60 2. 60 §
0 0
4o 3’ 40 e
20 °\° 20 32
0 0
Maximum 0 N Maximum 0 5'
Allowable 15 20 u, l\ " Bus stop Allowable 15 ,0 R Bus stop
GE/FiT, (s) N volume GEIRT, (3) 2° N volume
per/hr per/hr
Green Extension Success Rate, v/c 0.80 Green Extension Success Rate, v/c 0.80
100 100
80 5 80
E i B”.
60 (1' 3 60 E
o -. o
3 _‘ 8
4° .3 4° :<
20 32 20 .\°
0 0
Maximum 0 5 I ’ In N Maximum 0 5 . m N
Allowable 15 m l‘ " BUS stop Allowable 15 20 m N Bus stop
GE/RT, (s) 2° N volume GE/RT, (s) N volume
per/hr per/hr
Green Extension Success Rate, v/c 0.90 Green Extension Success Rate. v/c 0.90
100 _ 100
so m
:t 2
60 §.' 52
(D 2 m
3 5
40 a : 53
20 S“ a~°
' 0
Maximum 0 N Maximum 0 5 «I
Allowable 15 u, h " Bus stop Allowable 15 m E " Bus sIop
GE/RT, (s) 20 N volume GE/RT, (s) 20 N volume
per/hr per/hr

 

 

 

 

" Zero values indicates that no bus was observed

Figure 5.6 Efficiency of green extension strategy under different level of service

81

5.3. Impacts on Traffic along Priority Route

Figure 5.7 illustrates impacts on passenger car delays for vehicles traveling on the east-
west arterial. It is observed here that dwell time variability has relatively small impacts
on the general traffic. This is explained by the fact that dwell time variability does not
translate into significant changes in priority actions in the scenarios considered. While
variability leads to fluctuations in bus arrival times, priority needs are determined under
both the deterministic and stochastic dwell time assumptions using similar expected
average dwell time assumptions. In both cases, buses thus tend to request similar priority
actions even though buses with stochastic dwell times will not necessarily all fully

benefit from the granted green extensions and early recalls.

Figure 5.7 shows that increases in average boarding rate (i.e., increase in dwell time) lead
to increased traffic delays, and that increases in the maximum signal alteration limit tend
to reduce delays for a given dwell duration. In the later case, delay reductions are
primarily due to the increasing ability of buses requesting priority to lengthen the
effective green signal along the east-west arterial. In the former case, the delay increases
with longer dwell times are again attributed to changing proportions of buses arriving
during the green and red portions of the signal cycle. As the dwell time increases, more
buses arrive during the green, thus reducing the need for priority and the opportunities for

increasing the effective green along the arterial approaches.

Non-transit traffic tends to benefit the most with early recalls. When green extensions are
awarded, vehicles are able to avoid delays corresponding to an entire red duration.
However, only a few vehicles will benefit as the highest flows are generally observed at

the beginning of a green phase. For an early green recall, the potential delay savings are

82

limited to the duration of the recall, but all the vehicles in queue at the intersection at the
time of the action will benefit from an earlier departure. Thus, despite yielding lower
delays per vehicle, early recalls have a potentially greater impact on the overall traffic
through their ability to reduce delays for a greater number of vehicles. However, it should
be kept in mind that these results do not account for potential delay increases that may
result at downstream intersections from changes in progression patterns as the evaluation

only considered an isolated intersection.

5.4. Impacts on Cross-Street Traffic

Impacts on cross-street traffic were found to be generally opposite to those for vehicles
traveling along the prioritized bus route. While priority actions increased effective green
time and reduced delays along the east-west arterial, these actions reduced green
durations and increased delays on cross-streets. Whether a priority system will cause
overall traffic delays to increase or decrease will then depend on the proportions of

vehicles benefiting and suffering from the priority actions.

83

Stochastic Dwell Time

 

Fixed Dwell Time

 

EB Traffic Vehicles Delay, vlc 0.50

    

Traffic Delay, s/veh

pass/hr GE/RT, 5

EB Traffic Vehicles Delay, vlc 0.50

    

Traffic Delay, s/veh

pass/hr GE/RT, s

 

 

EB Traffic V hicles Delay, vlc 0.70

   

R3?

Traffic Delay, s/veh

150
125100 75 1510 5
Bus Stop Demand 50 25 20 Max. Allowable
pass/hr GE/FlT, 3

EB Traffic Vehicles Delay, vlc 0.70

Traffic Delay. s/veh

  

5 0

10
5° 25 20 ”Max. Allowable
G

75
Bus Stop Demand
pass/hr E/RT, s

 

 

EB Traffic Vehicles Delay, vlc 0.90
g 1

   
 

Traffic Delay, s/veh

    

 

 

EB Traffic Vehicles Delay, vlc 0.90

Traffic Delay. s/veh

     
 

 

 

150 150 0
1251 75 1510 5 125100 75 50 1510 5
Bus Stop Demand 50 25 20 Max. Allowable Bus Stop Demand 25 20 Max. Allowable
pass/hr GE/FtT, s pass/hr GE/RT. s
Figure 5.7 Traffic vehicle delays for scenarios considering stochastic dwell times

84

5.5. Summary

The simulation results indicate that dwell time variability impacts the ability of priority
systems to benefit transit operations. Contrary to expectations, scenarios with stochastic
dwell times resulted in lower average bus delays than scenarios with fixed dwell times. It
was expected that increased dwell time variability would decrease performance. In both
cases, no significant impacts were observed on the general traffic. The contrary impacts
on transit operations are attributed to changes in bus arrival patterns within the simulated
signal operation. In the simulated scenarios, dwell time variability pushes buses to arrive
at moments when they can fully benefit from the ability of a signal system to implement
green extensions and early green recalls, thus leading to increased benefits. In other
situations, however, dwell time variability may as well have opposite effects. For the
general traffic, the lack of noticeable impacts is the results of the fact that all
implemented TSP actions were based on expected average transit dwell times and not

actual dwell times.

Overall, the results of the simulation reported in this chapter indicate a need to retain
consideration of the potential effects parameter in any TSP algorithm. In particular, it can
be concluded that if the objective is to evaluate general traffic impacts associated with the
implementation of a priority system, simulations or algorithms considering average dwell
times can be used. However, if the objective is to determine the efficiency of selected
priority strategies or algorithms with respect to transit operations, it then becomes

important to consider dwell time variability in the TSP optimization.

85

CHAPTER V’l.

D-SPORT COST FUNCTION
FORMULATION

 

his chapter discusses in details the mathematical development of the cost function
that will be fed into the GA optimizer to find a near-optimal timing plans with TSP
actions. First, a brief description for the traffic signal control assumptions is introduced.
Next, the formulations for the decision variables that will be used in the cost function are
presented. The remaining sections show the mathematical derivatives for the general cost

function and the sub-cost functions.

6.1. Traffic Signal Control Assumptions

To support the evaluation of the D-SPORT controller performance, a custom traffic
signal controller will need to be developed for operation within the VISSIM microscopic
traffic simulation model since the functions provided with the software do not provide the
desired functionalities. As discussed earlier, the D-SPORT controller will have to provide

TSP treatments for buses traveling in an urban coordinated signalized network with other

86

traffic where some uncertainty may exist regarding true transit vehicle arrival times at the

intersection stop line.

In developing the D-SPORT cost functions, the following assumptions are made:

0 Transit vehicles are all equipped with a Global Positioning System allowing
continuous tracking of their movements.

0 All buses are equipped with an automatic passenger counter system.

0 A continuous communication system exists allowing the signal controllers to
obtain in real-time all relevant information about approaching buses (such as
position, speed, and number of passengers).

0 The D-SPORT controller is to operate in real-time.

0 Minimum green times must be granted for all phases in order to accommodate
pedestrians who are crossings the street.

0 Maximum green times must not be exceeded for all phases

0 The traffic signals operate with two phases only.

The traffic signals were set to operate with two phases only to simplify the problem at
hand and to furnish the background required to develop a model that can eventually be

improved to handle more complicated phasing schemes and traffic composition.

6.2. Decision Variables

The decisions variables upon which a signal controller operates are cycle lengths, green
times, and offsets. Each of these variables is described in more details in the following

paragraphs.

87

Cycle Length
The cycle length equals to the summation of all signal intervals within a complete signal

display sequence. This includes the summation of green times, yellow and all-red

intervals for all phases. Mathematically, this can be expressed by Equation 6.1 below:

CL,- = §(gm- + rrm- + Yni) ................................. Equation 6.1
n =1
Where:
CLi: Cycle length of intersection i, in seconds
gn i : Green duration for phase n at intersection i, in seconds
y ni3 Yellow duration for phase n at intersection i, in seconds, set to 4 s
rrni: All red duration for phase n at intersection i, in seconds, set to 1 s
(I): Total number of phase for intersection i
Green Time

For each traffic movement, the green time represents the interval during which traffic is
allowed to enter an intersection. For the optimization at hand, this parameter is bound by

both a maximum and a minimum value, as expressed in Equation 6.2:

gni E [8 min , 8 max 1 ................................. Equation 6.2
Where:
8min 1 Minimum allowable green time, in seconds, set to 20-s
gmax : Maximum allowable green time, in seconds, set to 40-s

88

A minimum green time must be imposed to ensure that pedestrians are always given
enough time to cross the intersection or to provide a guaranteed minimum traffic service
time. On the other hand, a maximum green time must be specified to avoid allocating
disproportionably long green intervals that may not be well received by motorists. To
determine the lower and upper limits of a green duration, a careful consideration of the
geometric layout of the intersection and the time required for pedestrians to cross the

intersection and traffic conditions is required.

Offsets

An offset is defined as the time elapsed between the beginning of a green phase at an
intersection, and the beginning of a corresponding green phase at the following
intersection. An efficient offset would be the one allowing the first vehicle of a platoon
leaving an intersection to reach the downstream intersection right when the tail of the
downstream queue has just cleared. Such an offset would in theory maximize utilization

of green time and contribute to reduced stops and delays.

To calculate the offset, information is required about the time taken by vehicles to travel
from one intersection to the next. This travel time is function not only of the desired
traffic speed but also of the initial speed at which vehicles enter the segment and the rate
at which vehicles can accelerate and decelerate. Assuming a linear relationship between
acceleration and speed, the following equation can be used to determine the maximum

acceleration rate of a vehicle traveling at a given speed v [53].

89

acc = U + 2' - v ................................. Equation 6.3

Where:

. . . 2
acc : Vehicles acceleration rate, in ft/s

v: Constant parameter representing the vehicle’s maximum feasible acceleration, in
ft/s2
2': Constant parameter representing the rate of reduction in acceleration with

. . . -1
increasmg speed, 2' < 0, in s

v: Speed of vehicle, in ft/s

This equation correctly represents that a vehicle’s acceleration capability tends to
decrease with increases in speed. While this model is not perfect, it represents a
commonly used assumption in the modeling of vehicle maximum acceleration
capabilities [53]. In this case, the equation is used to model the actual acceleration of a
vehicle and not necessarily the maximum acceleration capability. This is done by simply
assuming that the actual vehicle acceleration can be expressed as a constant fraction of

the maximum acceleration capability.

Equation 6.3 can be utilized to calculate the time required to travel across a link assuming
no initial queue in the downstream intersection. Figure 6.1 illustrates the speed change
over time for a vehicle traveling between intersections i and j. The vehicle leaves
intersection i with an initial speed v,- at to, and accelerates until it reaches the desired
speed V] at t]. After traveling at the desired speed, the vehicle reaches intersection j, at t2.

Therefore, the ideal offset between the two intersections should be synchronized with the

90

arrival of the vehicle. The following equations show how the link travel time can be

mathematically derived from Equation 6.3.

 

 

 

 

 

 

 

 

 

 

 

Int. i L Int. j
.<___>€ \
(f E
l 2
Speed, A
ft/s
v, : i
I i ’
to t1 t2
Time, s

 

 

 

Figure 6.1 Schematic sketch for speed profile through the link

V

‘ v+zv
fl=u+zv -) jdt=j dv 9 :,=—l-ln[ f]
dt 0 wv+zv 2' v+zv,-
l

 

 

 

1 l v+2v
t=-l V -) t =—l -) . 17:
n(u+zv)|. n(v+zv,-] (0+2v,)e v+zv

(’1 t1
v=%[(v+zvi)-e"—v] -) £ds=£ %[(v+zvi)-e"—v]-dt

 

 

 

 

Vf +11): 1) U+2Vf
3 9 [1: '——In
T T U'i‘TVi

 

& LinkTravelTime =11 + t2 -)

91

U+va]+ 2' 2' v+2v,- 9

LinkTravel Time = iln
B U + Wi Vf

 

 

L' ‘ v+ - v+2-v
LinkTravelTime= l—U + v, .1n[__f_

..................... Equation 6.4
V f T - vf v + 2' . v,-
Another parameter required to determine the ideal offset is the time required to clear a
queue of vehicles that may be present at the downstream intersection. This parameter is

important if it is desired to allow vehicles being released at an intersection not to be

stopped by a dissipating queue while traveling to the next intersection.

Equation 6.4 calculates the ideal offset based on the time required to travel between
intersections i and j under free flow conditions. However, this ideal offset should be
adjusted to take into account any queue presented in the downstream intersection. The
adjustment consists in subtracting from the ideal offset the time required to dissipate the
queue before the upstream platoon arrives. The queue dissipation time is a function of
number of vehicles in queue and the average discharge headway in saturated conditions.
For a one way corridor with no allowed turning movements, the ideal offset can therefore

be computed using Equation 6.5.

‘ ‘ v + » U 'l' T ' V _.
I_0ff,-j =—Ll—J-+ A -In —L —qj-h .................. Equation6.5
Vf 2'- Vf v + 2' - Vi
Where
I _0flij: Ideal offset between intersections i and j for any given cycle, in seconds

92

The initial speed at which the first vehicle in the platoon enters the

segment, in ft/s

vf : The desired speed reached by the first vehicle in the platoon, in ft/s
q j : Initial queue at intersection j during cycle c, in number of vehicles
l2- : Average saturation headway, in s/veh, set to 2.5 s/veh

6.3. D-SPORT Controller Cost Function Formulation

The cost function can be broken down into a set of sub-functions each representing a
specific objective. The combination of all these sub-cost functions into a single equation
would then create the general cost function. Since different agencies may have different
priorities and perspectives on what constitutes ideal TSP treatment, this approach would
allow the optimization process to be tailored to the specific needs of each agency. For
instance, if an agency is interested in implementing TSP in such a way that transit
vehicles experience less delay, then the optimization process would be defined with a
cost function minimizing delays. If another agency is interested in regulating transit
schedule, the cost function considering this parameter specifically could instead be used

within the same optimization framework.

To reflect the potential for considering alternate optimization objectives, the following
cost function has been coded for optimization by the GA. This equation gives a weighted
factor for each of the sub-cost functions. Attribution of a value of zero to any of the
weighting parameters will effectively cancel consideration of this parameter in the

optimization. Similarly, attribution of a value above zero will allow the parameter to be

93

considered with higher or lower importance in the optimization depending on the values

assigned to the individual weight parameters.

0’
COST =[Z PI 2m ZSCH1- 027:; ..................... Equation 6.6
“SCH
Where:
COST: Cost function value
PI : Traffic network overall performance index, (defined in Section 6.3.1)
TIT : Transit travel time (defined in Section 6.3.2)
SCH : Schedule adherence penalty (defined in Section 6.3.3)
apl : Traffic network performance index parameter sub-cost parameter, has a
value of 1.0
am. Transit travel time parameter sub-cost parameter, has a value of 10.0
a SCH : Schedule adherence sub-cost parameter, has a value of 10.0

In this research, the selection of the values for the sub-cost parameters (i.e.,apl ,am ,
and “SCH ) is based on the sensitivity of the cost function optimization to these

parameters. Specifically, weight values were assigned to produce a cost function in which
the magnitude of the performance index, travel time and schedule parameters are within

the same scale.

The following paragraphs describe the mathematical derivation for each of the

performance sub-cost considered in the general cost function.

94

6.3.1. Traffic Network Performance Index

The network performance index, (PI) is defined as the sum of the weighted linear
combination of estimated delay and number of stops per vehicle per unit time for all
signal phases under a given timing plan [54]. The PI will be used for optimizing the
signal timing to accommodate non-transit vehicles. For an isolated intersection, y/ is the
vector that expresses the signal timing parameters (i.e., the cycle length CL, green

time gq, , phases vectord> , and offset ofi‘). This can be expressed mathematically as:

l//=[CL gq) CI) off] ......................................................... Equation 6.7

For a given signal timingsl/I, the PI can therefore be mathematically expressed as a

function of:

PI: 2(Wd -D,° (w)+WnS -NS,-(w)) ................................. Equation 6.8
Vie INTX
Where:
PI : Performance Index
Di (l/l): Average delay per vehicle for phase group i, in s/veh
NS i (W): Average number of stops per vehicle for phase group i, in stops/veh
Wd : Weighted factor for average delay
WnS: Weighted factor for average number of stops
[NIX : Total number of intersections in the traffic network

For this research, Wd and Wns are selected to be 1.0 and 10.0 respectively. These values

were adopted from signal optimizations carried out using the SYNCHRO 6 traffic

95

optimization software [52]. These values were selected so that the number of stops and
average control delay are within the same scale, and to ensure that the GA will not be

biased toward minimizing delay over number of stops.

The following subsections discussing in details the formulations used for estimating

delays and the number of stops incurred by vehicles.

Average Control Delay Estimate

The HCM 2000 [55, 56] approach is used to calculate average control delay per phase
group per intersection. For a given intersection, the control delay is calculated using the

following formula:

D": dum- X PF + dim. + d,“ ......................................................... Equation 6.9
Where:
D": The average control delay per vehicle for phase group n, in seconds
PF : Uniform delay adjustment factor for quality of progression. According to

the HCM 2000 [55] guidelines, and for the purpose of this research, a

value of 1.0 is chosen

dam-2 Uniform delay assuming uniform arrival, in seconds
dine: Incremental delay due to traffic stochastic effects, assuming no residual

demand at the beginning of the start analysis period, in seconds

dm : Residual demand delay caused by the presence of queued vehicles at the

beginning of the analysis period, in seconds

96

Uniform Delay

The uniform delay is defined as the delay occurred in a lane group if vehicles arrive with
a uniform distribution and saturation does not occur in any cycle [56]. For each

intersection, the uniform delay for a specific lane group is defined as:

 

(I) 2
1- -/CL
dum- = 20.50 CL ( g' , ) ................................. Equation 6.10
i-l l—(gi/CL)[m1n(v/ci,l)]
Where:
v/ Ci : Volume to capacity ratio for phase i
g ,- : Effective green duration for phase i, in seconds
(I): Total number of phases for each cycle
Incremental Delay

The incremental delay takes into consideration that the arrival is random, not uniform,
and some cycles might overflow [56]. Mathematically, the incremental delay is expressed

as follows in the HCM 2000:

 

(I)
8k'l' / '
dim: Z 900T|:(v/ci —1)+\[(X,- --r)2 +——'—i—5‘—] .................. Equation 6.11
i=1 Ci
Where:
T: Analysis period duration, in hours

~o
I

Incremental delay factor based on controller settings. As per HCM 2000

guidelines [55], a value of 0.50 is assigned to this factor

97

I...

Upstream filtering metering adjustment factor accounting for the effect of
filtered arrival from the upstream. As per HCM 2000 recommendations

[55], a value of 1.00 is assigned to this parameter in the current research

Residual Delay

The residual delay occurs as a result of an initial unmet demand at the beginning of the
analysis period. For a given intersection, the calculation of residual delay done as per

Equation 6.12. [56].

 

 

(I)
1800 - 1— - t-
dres= Z Qb‘; u,) l ......................................................... Equation 6.12
i=1 cl
Where:
Qb, : Initial unmet demand at the start of period T for phase 2'
ti: Duration of unmet demand in T for phase i, (see Equation 6.13)
it,- : Delay parameter for lane group i, (Equation 6.14)
0.0 IF th' = 0.0
t' = - ............... uation 6.13
’ min Ti, Q,“ IF Qb, ¢ 0.0 Eq
Ci(l-mm(l- Xi»
0.0 [F ti < T
“i = 1 _ le' [F t' >T ............... Equation 6.14
l _

 

Ci(l-min(l-Xi))

98

Number of Stops Estimation

In this research, the number of stops is defined as the number of vehicles that arrive at a
downstream intersection during either the red interval or when queue exists [45]. For

simplicity, some assumptions are made in the execution of the calculations:

0 Only unsaturated traffic conditions are considered, (i.e., only situations with
traffic demand less that the capacity for each cycle).
0 All traffic demand is met during the cycle, with no overflow from the previous

cycle to the next.

Although the above assumptions seem to simplify the estimation of the number of stops,

these assumptions have been shown to fit well with simulation experiments [57].

For a given phase, the time required to dissipate a queue is a function of the red interval
duration, arrival rate, and saturation rate. The queue dissipation time can be expressed as

follows when considering average arrival, departure and saturation flow rates [45]:

TQl- =3“??? ..................................................................... Equation 6.15
Where:
TQi: Queue duration for phase i, in seconds
ri: Red interval for phase i, in seconds
S ,- : Saturation flow for phase i, in veh/s
xii: Vehicle arrival flow for phase i, in veh/s

99

The queue dissipation time represents the portion of the green interval that effectively
acts as a red signal for approaching traffic. The number of stops attributed to arrivals

while a queue is present can then be calculated using Equation 6.16. [45].

(D
NS ,, = 22, XTQi ......................................................... Equation 6.16
i =1
Where:
NS ,, : Number of stops during phase n, in number of stops/veh

6.3.2. Transit Travel Time (TTT)

Figure 6.2 illustrates the trajectory of a transit vehicle traveling through its route in the
network. Once a transit vehicle is detected at an intersection, it leaves the intersection and
starts traveling on the segment linking its current intersection to the next one. If a bus
stop is present between the two intersections, as in Figure 6.2, the vehicle may then
temporarily stop for boarding and alighting passengers, if needed. Once the transit
vehicle is done with dwelling, it resumes travel toward the downstream intersection,
where it might stop again for a while waiting for any existing queue to clear, or for green
time to start. Based on the above elements, Transit travel time, (TTT) can be expressed

using Equation 6.17:

E PT J
m = 7+ Zdwlstp + 2T1),- ................................. Equation 6.17
b stp=b i: j

Where:

100

[“I

dwl

stp :

PT;

TD,‘ :

stp 3

Estimated transit travel time from a transit current location to the end of

the network (i.e., the target destination), in seconds

Distance between the transit current location and the end of the network,
in ft

Transit desired speed, in ft/s

The predicted time elapsed between the moment a transit vehicle enters

the bus stop and the moment the bus will be ready to leave the bus stop stp
and merge with traffic, which includes dwell time (will be discussed in
detail in Chapter 7), in seconds

Bus stop ID

The first coming bus stop meets the transit vehicle

All the bus stops in front of a transit vehicle along its route

Expected delay for a transit traveling through intersection i

The first downstream intersection will meet a transit vehicle

The remaining intersections that a transit vehicle needs to travel through

101

 

Distance

’

A

 

 

 

 

 

 

 

 

 

 

L3
L3 lnt.2 E
_ = ) Bus
L 521+ ”22 Stopi
Int. 1 E
L1 . . , , .
Ll MA 1-1‘ 1/2 - L3 7
— TDl _1_ (“1 __1 TDZ —
—— [I lb 17) l'b l'b 7"”

TTT

 

 

 

Figure 6.2 Time space diagram for transit vehicle traveling throughout the
system

The simulation environment has the ability to provide the traffic controller with estimates
for the initial queue at each intersection. In reality, information about the number of
queued vehicles at any given time can be derived by individual signal controllers by
compiling vehicle detections on intersection approaches. This information could then be
communicated to adjacent signal controller if the controllers have the capability to
communicate with each other, either through a local direct communication link or

communication with a central traffic control center.

The first term in the equation predicting transit travel time is computed by simply
dividing the distance remaining for the transit vehicles to leave the network from its
current location by the vehicle’s desired speed. If the transit vehicle experiences no delay,
and does not dwell at any bus stop, the transit travel time would then equals the distance

to travel divided by a transit vehicle desired speed.

102

An ANN model is used to predict the transit dwell time parameter, dwl b , based on traffic,

signal, and transit information. The prediction of this parameter will be discussed in
Chapter 7. Note that this term represents the predicted time that a bus might take from the
moment it enters a bus stop to the moment the bus is ready to leave. If the bus stop is a
near side bus stop, the time the bus is ready to leave would then represent the stop line

projected arrival.

The last term required to estimate transit travel time is TDi. For a given analysis period,

the average delay is calculated as the summation of the total delay divided by number of
vehicles that pass within this period, which means that not all the vehicles are subjected
to delay. Therefore, it is important to separate the transit vehicle from the traffic stream,

and look at its expected delay individually.

If a vehicle arrives at the beginning of the red interval, it experiences the maximum
delay, since it remains still during the entire red interval. Vehicles arriving later during
the red are also forced to stop; however, their delay would not be as high as vehicles
arriving at the beginning of the red. When the signal turns green, the first vehicle starts
moving, followed by the second vehicle, until the queue is dissipated. Once the queue is
dissipated, any arrival vehicle does not have to stop, and therefore, does not face any

delay leaving the intersection.

The delay behavior in this case is linear. The maximum delay is when a vehicle arrives at
the beginning of the red interval, and the zero delay is when the vehicle arrives after the
downstream queue is dissipated. Therefore the expected delay can be calculated using

Equation 6.18:

103

0.0 IF TQ, < Arrival < CL

TDi= ZDn ><CL(1— Arr,) ......... Equation 6.18

 

IF 0 S Arrival S TQ,

2
TQi
Where:
Arty: The stop-line projected arrival time at intersection i relative to the cycle
local timing given that no is queue presents, in seconds
TDi : Expected transit delay at intersection I, in seconds

6.3.3. Schedule Adherence Penalty (SCH)

Buses are expected to arrive at each bus stop following a given schedule. However, due
to random traffic behavior, it is possible that individual buses will arrive earlier or later
than expected. In order to accommodate early or late arrivals, a penalty is introduced as a
function of the transit schedule and the anticipated bus stop arrival. The schedule
adherence penalty is added to the cost function for any solution in the search space for
which the absolute difference between the scheduled and projected arrival time of a
vehicle at a specific bus stop exceeds zero. For a specific bus stop, this penalty can be

expressed by Equation 6.19:

SCH 5,1,1, = SchArrsth, - EXPAr’stp,b ................................. Equation 6.19

Where:

SCH S’Pab: Bus b penalty due to early or late arrival at bus stop stp arrival according

to the schedule

SchArrwpb: Bus b scheduled arrival at bus stop sip, in seconds

104

Eprrrpr, : Bus b expected arrival at bus stop stp, in seconds

The transit vehicle expected am'val is computed by utilizing Equation 6.17, as presented

in the following equation:

I PT J
Eprrr=V—+ Zdwlstp + 2T1),- ................................. Equation 6.20
stp=b i: j

Where:

Z : Distance between the transit location and the target bus stop, stp when the
optimization process is called, in ft

j : The first intersection to meet the transit vehicle before reaching the target
bus stop, if any

J : The last intersection to meet the transit vehicle before reaching the target
bus stop, if any

b : The first bus stop to meet the transit vehicle while approaching the target
bus stop, if any

PT : The last bus stop to meet the transit vehicle while approaching the target
bus stop, if any

stp: Bus stop ID

For optimized signal timings in a network consisting of multiple links with multiple bus

stops, the cost function will be penalized by the sum of all the penalties applied at all bus

stops in the network, for all the buses currently in the network. A mathematical

expression for this calculation can be written as:

105

S B
SCH = Z ZSCHstJ) ............................................. Equation 6.21
VstpES VbEB

Where:

SCH : Schedule adherence penalty, in seconds

B : Number of transit vehicles that are currently in the network

S : Number of bus stops a transit vehicle did not pass yet along its route.

6.4. Optimization and Simulation Integration

Previous sections of this chapter represented the cost function that is to be fed to the
controller. As discussed in Chapter 3, the proposed algorithm is set to accommodate
problems associated with dwell time variability, implementing TSP within a coordinated
signalized network, and the need to handle simultaneous priority requests. Furthermore,
the D-SPORT controller needs to be able to optimize general traffic movements, in cases

in which transit vehicle may or may not be present in the network.

The flowchart shown in Figure 6.3, describes the integration of the optimizer within the
simulator by using the VISSIM-COM [58] interface. This interface allows
communicating to the simulation environment information to and from an external

controller. This integration process will be discussed in details in Chapter 8.

In the illustrated flowchart, the VISSIM—COM interface triggers the start of the
simulation process. The simulation moves forward until the time for calling the optimizer
has come, which is a function of signal controller’s cycle lengths and offsets. Once the
controller decides to call the optimizer, information such as traffic data, transit data, and
controller status is passed to the optimizer. The GA optimization platform, which is the

106

tool that is used to perform the optimization, then finds a near-optimal solution by
minimizing the D-SPORT cost function, and passes the signal timings back to the
controllers through the VISSIM-COM interface. This process is iterated until the 2-hr

simulation duration has elapsed.

 

 

 

SAVE Simulated Traffic Conditions
SAVE Signal Control Timings
SAVE Transit Related Data

Start VISSIM-COM
Interface

 

 

 

 

 

 

I
Call VISSIM Simulation {For Each Generation} .-
Environment

 

 

 

 

Y

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

. v
/ . _ . . Call GA Optimization 3 For Each String in
- FOR Slm - 1‘, Simulatlon YES Platform ‘ Gene rtion
’1 Period . , 4 .
‘ {t
. Call ANN '
Tlme to
call GA? V ,
For Each . ;‘
Intersection in , V T ‘
NO Network ; Calculate Fitness 1 :
v‘ 3 ; Find Next Generation .
Continue with the ' . ' .
Simulation 3.3”? New S'gna' a
Tlmlng v
v Parameters Return Signal Timings
End

 

 

 

 

 

 

Figure 6.3 Flowchart for optimization process and simulation environment
integration

107

CHAPTER VII.

TRANSIT TRAVEL TIME
PREDICTION MODEL

 

his Chapter presents the development of an ANN-based prediction tool for

providing information about the expected arrival time of buses at intersections
where dwell activities of a near-side bus stop may create uncertainties in the predictions.
As detailed in Chapter 6, an ANN prediction model is required for determining the transit
travel time term in the D-SPORT cost function. The ANN prediction model presented in
this chapter takes advantage of the properties of artificial neural networks. To ensure that
the model could be applied in real-world settings, its development is also based on
constraints stipulating that the travel time predictions are to be based on information that
is available to traffic signal controllers, traffic information systems, or that can be
obtained from current sensing technologies. The following sections successively describe
the experimental setup, the initial list of model parameters that were identified for

possible inclusion in the prediction tool, the process by which data simulating the bus

108

arrival processes were generated for the development of the prediction model, the process
by which a final selection of model parameters was conducted, the training of the

proposed neural network model, and the evaluation of the ANN model performance.

7 .1. Experimental Setup

The test network used for the model development is based on the network that was used
in Chapter 5 for the evaluation of dwell time variability impacts on TSP operations. This
network is illustrated in Figure 5.1. It is used in this part of the research to generate the
datasets for building and training the proposed ANN model. The only changes that were
made to the simulation model described in Chapter 5 are related to the traffic signal
control operations, traffic demand, and transit service demand. These input parameters
were adjusted to allow the execution of simulations covering a wide range of possible
combinations of traffic demand, transit demand and traffic signal control operations.
Below is a brief summary of the key features of the simulation modeling used in this

portion of the research.

Traffic Signal Control

Each intersection is controlled by a 90-s two-phase fixed-time signal operation providing
40-s of effective green time on each approach. A 20-s minimum green is further imposed
on each phase to ensure that priority actions do not unduly restrict pedestrian crossing
times. Based on the traffic demand assigned to each experiment, the signal coordination
between the two intersections is optimized using SYNCHRO 6 traffic signal timings

optimization software.

109

Traffic Demand

To remove truck effects, only passenger cars are simulated. Vehicles are generated at the
upstream end of each entry link using a Poisson distribution and VISSIM’s default
vehicle characteristics. To allow testing in a range of traffic conditions, scenarios with
eastbound (EB) and westbound (WB) volumes of 600, 800, 1000, 1200, 1400, and
1600 vph, and northbound/southbound volumes corresponding to 80% of the
eastbound/westbound volumes, were considered. For each scenario, additional sub-
scenarios were created by varying the proportions of right-turning and left-tuming
vehicles on the eastbound approach to Intersection 1 between 5% and 15%. Turning

percentages on all other approaches remained fixed at 15% right turns and 5% left turns.

Transit Service

Buses are assumed to travel in the eastbound (EB) direction only and to service a single
stop located in the right-tum bay of the eastbound approach to Intersection 1 with a
frequency of one bus every 5-min, with a standard deviation of 20-s. While buses are
generated to enter the network with a 20-s standard deviation and all share a desired
65 km/h (40 mph) travel speed, interactions with other vehicles and traffic signals create
more fluctuations in the time each vehicle takes to reach the bus stop. To minimize the
impacts of gap-seeking delays, all buses are modeled to have priority over the adjacent
traffic when leaving the bus stop, which can be regulated in the field. While this
facilitates their departure, it does not completely eliminate the need to seek gaps in the

adjacent traffic, particularly when a queue exists in the adjacent lane.

Passenger demand is modeled by simulating the process of passenger arrivals at the bus

stop. Passengers are modeled to arrive at an average rate varying between 25 and 200

110

passenger/h, with an increment of 25 passenger/hr. The generated passengers are
assumed to enter the bus stops by following the Poisson distribution, since the simulation
environment does not support other distributions for transit demand. For each bus, dwell
time is then determined by assuming that 3—s are required to allow each passenger to
board, with another 3-s for door opening and closing. Dwell time variations thus result
from randomness in passenger arrivals at the stop and the effects of early or late transit

arrivals.

Figure 7.1 summarizes the simulation experiments setup that are used to generate the data
required to train the ANN model. As can be seen in the sketch, different traffic volumes
are combined with different left and right turning volumes and different transit demand at
the bus stop. Based on the various combinations considered, a total of 360 scenarios were

built and each simulated for a 2-hr period.

 

 
 

Flight Turn % Transit Demand

  

Scenarios

‘1
U!

_s
N
01

 

titles

175

 

 

 

 

Figure 7.1 Experiments used to generate the datasets required for the ANN
model development

111

7 .2. Initial Candidate Model Parameters

Variables shown in Table 7.1 were initially selected as tentative input variables to the
ANN model. The time elapsed between detection and the moment a bus is ready to leave
the bus stop is the value to be predicted. All the selected variables quantify traffic
conditions along the approach on which the bus is traveling. Most of the selected
variables further represent average traffic conditions in S-min sampling intervals. The
information represented by these variables will not necessarily represent the average
traffic conditions in the S-min that preceded a detection, but rather the conditions in a
5-min window that may have ended anywhere between 0 and 5 min prior to the detection.
Observations over a 5-min interval are used to reduce the sensitivity of the prediction
model to short-term traffic variability and allow the model to focus on short-term trends
rather than responding to short-lived random fluctuations, and to make the model more
robust to handle various conditions. Also, a wider time interval is not used since
collecting data over a wider period of time, says 15-min, would cause a loss of real-time

data that are used to feed the dwelling activity prediction model.

112

Table 7.1

Variables considered for the ANN prediction model

 

Variable

Definition

 

input Variables

 

Pass (Persons/bus)

Number of passenger in the bus

 

 

:35) (Passenge ”b us Current number of passenger at the bus stop

THV (#veh) Number of through vehicles in previous 5-min interval

RTV (#veh) Number of right turn vehicles in previous 5-min interval

LTV (#veh) Number of left turn vehicles in previous 5-min interval

THTT (sec/van) Travel time of through vehicles in previous 5-min interval

RTI’T (sec/volt) Travel time of right turn vehicles in previous 5-min interval

LTTT (sec/veh) Travel time of left turn vehicles in previous 5-min interval

THD (sec/van) Average delay of through vehicles in previous 5-min interval

RTD (sec/veh) Average delay of right turn vehicles in previous 5-min interval

LTD (sec/veh) Average delay of left turn vehicles in the previous 5-min interval

THO (ft) Average queue of through vehicles in previous S-min interval

RTQ (ff) Average queue of right turn vehicles in the previous S-min interval

LTQ (It) Average queue of left turn vehicles in previous 5-min interval

THQM (ft) Maximum queue of through vehicles in previous 5-min interval

RTQM (ft) Maximum queue of right turn vehicles in previous 5-min interval

LTQM (ff) Maximum queue of left turn vehicles in previous 5—min interval

LCyc (s) The current local cycle time when bus is detected

Late Transit Lateness relative to the previous transit relative excluding average
tranSIt headway

Offset (s) thgggtegrbetween the upstream Intersection relatlve to the downstream

Predicted Variable
Arr (s) Time elapsed between bus detection and the moment the vehicle is ready to

leave the bus stop

 

7 .3. Data Generation

Similar to the simulations of Chapter 6, VISSIM is used to generate data for the
development of the ANN model. Simulated data were used due to difficulties in
collecting all the required data from real-world intersections, where typical existing
detection technologies may not allow easy collection of many of the parameters being

considered. This includes offsets, transit occupancies, and current local cycle. It is

113

furthermore difficult to control experiments in real field settings. Contrary to real field
experiments, simulation allows more flexibility in controlling the experimental setup and

performing sensitivity analysis.

For each simulated bus, input variables along with the dependent variable were stored.
These stored variables were used to generate the datasets for developing the ANN model.
Once the datasets were generated, input and output variables were normalized to have

their values only ranging between 0 and 1, as shown in Equation 7.1.

 

A X . _ X .
X ,- = l min ......................................................... Equation 7.1
MAX(X,-)-MIN(X,-)
Where:
If ,- : The normalized variable X ,-
X i: The variable i value

MIN (X i) : The minimum value for input X i in the collected data

MAX (X i) : The maximum value for input X ,- in the collected data

The transit time consumed at the bus stop (the dependent variable) was taken as the time
elapsed between the moment a transit enters a bus stop and the moment the transit has
closed its door and is ready to move. This includes the deceleration until complete stops,
the dwell time, a fixed 3-s for door opening, and another 3-s for door closing. Note that
the delay resulting from having the transit vehicles wait for a gap in the adjacent traffic to
enter the intersection (merge can occur in the intersection) is not part of the predicted

variable. This is because such delays are more closely associated with traffic interactions

114

than dwelling activities. Furthermore, non-transit vehicles are modeled to yield for transit
vehicles when they merge back in the traffic stream, which would reduces the
significance of gap-seeking delay impact on traffic behavior is minimal. Nonetheless, this

type of delay is accounted for in the transit travel time sub-cost function defined in

Equation 6.17 through the TDi term.

For non-dwelling buses, and for buses arriving during the normal green, the predicted
dwell time is the time elapsed between activation of the check-in and check-out detectors.
For buses arriving at the downstream intersection during the red the time, the predicted
time is the time elapsed between activation of the check-in detection and the moment the

bus is fully stopped.

The combination of 6 traffic volumes, 3 right turning percentage, 3 left turning
percentages, and 8 boarding rates resulted in 432 scenarios. For each scenario, the
consideration of a two-hour simulation period resulted in the generation of 23 buses. Out
of the 9936 potential datasets that could be produced, only 8871 were generated, as traffic
interactions prevented some buses from exiting the network before the end of the

simulation.

To ensure the validity of the ANN training process, three independent groups of datasets

were created by randomly assigning each of the 8871 datasets produced to a specific

group:

0 Training datasets: 5000 datasets used to train the ANN in predicting the

dependent variable.

115

0 Validation datasets: 2000 datasets used to monitor the performance of the ANN
during the training process and to keep the network from over-fitting, which
occurs when the ANN model may fit the noisy data, and not only the signal,
which reduces the ANN model ability to be generalized beyond the training
datasets. The use of an independent group of datasets for validation supervises the
training process and increases the prediction ability of an ANN for cases not seen
when submitting training datasets.

0 Testing datasets: 1871 datasets used to examine the ability of ANN to predict the

dependent variable in cases independent from the training and validation datasets.

7 .4. Selection of ANN Model Parameters

To help with the final selection of model variables, the training datasets were used to
generate the correlation matrix of Table 7.2. This matrix shows the result of correlation
analyses conducted between each pair of parameters. These analyses led to the rejection
of the variables shown in the shaded columns due to their high colinearity with other
variables. The rejected variables are travel time for through, right, and left movements,
(TH'IT, R'I'IT, L'I'IT), and the maximum queue for left, through, and right movements,
(LTQM, THQM, RTQM). Although some of the retained variables also appear correlated
to others, these were not excluded from the training process since their inclusion in the
ANN model was found to improve its prediction performance, mainly by reducing the
root mean square error (RMSE) and improving the coefficient of determination (R2). A
similar argument was also invoked to retain some variables with low correlation with the

travel time variable to predict.

116

0000.050 0. Boos zz< 3...: 05 can... 000.860.. 0.03 mcano 00.29.59:

 

83

0.0.9 08..

000.0 .000- 83

000.0- 00.0- 000.0 08..

000.0- .0..0. 000.0 000.0 83

.000. 00.0- 0000 000.0 000.0 83

000.0- 00.0. 000.0 000.0 000.0 000.0 83

08.0. 00.0- 00.0 000.0 00.0 000.0 000.0 83

000.0- 3.0- 000.0 .000 000.0 000.0 .000 000.0 83

000.0 000.0. 000.0 000.0 000.0 000.0 000.0 0.00 000.0 003

000.0 000.0 000.0 .000 00.0 00.0 00.0 00.0 0.0 00.0 003

0.0.0 0000 0.0.0- .000 000.0 0.0 000.0 000.0 0.00 000.0 00.0 003

000.0 000.0. 000.0 000.0 .000 000.0 .000 0.0.0 0000 03 00.0 000.0 000..

000.0 000.0 000.0 000.0 000.0 0.0 00.0 00.0 00.0 00.0 000.0 00.0 00.0 83

0.00 000.0 .00- 000.0 000.0 000.0 00.0 .000 000.0 03.0 0.0 03 03.0 00.0 003

000.0 .00- 0.0.0 0.00 000.0 000.0 .000 000.0 .000 0.00 00.0 000.0 000.0 00.0 0.00 000..

.000 000.0 000.0- 000.0 000.0 0.0.0 0.0 00.0 0.0 0.0.0 0.0 000.0 00.0 03.0 .000 000.0 003

000.0 000.0 000.0 000.0 000.0 000.0 000.0 000.0 0.00 000.0 5.0 0.0.0 0000 00.0 000.0 00.0 00.0- 003

.000 3.0.0- 0000 .00.? .000- 000.9 000.0- 000.9 000.0- 000.0- 00.0 0.0.0- 000.? 00.0 0.0.0- 0000- 000.0- 0.0.0. 003

0.00 00.0 000.0- 000.9 00.? .009 0.0. 00.0- 00.? 00.0- 00.0 0.0.0- 00.? 8.0 0.0.? 0.0.0. 0.0.0- .000 000.0 83
000.0 0.0.0- 0.0.0 00.0- 0:0- 0:0- 00.0- 090. 00.0- 3.0- 00.0 000.0. 3.0. 00..0 000.? 000.0- 000.0- ...0.0. 00.0 0.00 000..
.0000 0.0.. 000. .200... .2010 .200 A 00... GE 05 00.. 000. 0:0 . F... FE 5:0 _. >5 >0... >5 000.. 0000 0..

illlrr

J

 

03.50:. 5.50—2.80 00302.0“, 2.205965 35 2.0—Eamon

NS 030,—.

117

7 .5. ANN Prediction Model Training

A feedforward backpropagation learning algorithm (BP) was used to train different ANN
models. This is because BP learning algorithms are currently the most commonly used
learning tools for multi-layer ANNs [59]. The underlying theory behind BP algorithms is
the utilization of non-linear perceptron-like systems. Training is achieved by minimizing
the mean-squared-error (MSE) between the actual and desired outputs of the multi—layer

feedforward perceptron.

Equation 7.2 shows the objective function that was used for ANN training. As explained
above, the executed training sought to minimize the MSE between the predicted and

observed dwell time values.

£=-—1— int, — p, )2 ................................. Equation 7.2
"tr r=l
Where:
8: MSE over entire sets of neurons
or: Actual value of the dependent variable for dataset r
p r : Predicted value of the dependent variable for dataset r
n,,: Number of training datasets

Learning takes place through the repeated modification of the network weights based on
their relative contribution to the error term. Specifically, the derivative of the error

function with respect to any weight in the network is computed and used to adapt that

118

particular weight to meet the objective of the function to minimize the error value, as

shown in the following equation:

 

88
AW," ——77 ..................................................................... Equation 7.3
J aWij
Where:
AW,'j : Change of weight between nodes i and j
77: Learning rate, which is the rate at which the weights between neurons are

adjusted

There are two possible learning methods to train an ANN, the batch training and the
incremental learning. In the batch learning, the weight adjustments are accumulated over
the whole training sets, and updated at the end of each iteration in the ANN learning
process, which is also known as an epoch. In the incremental training, the weights are
instead adjusted for every dataset (pattern) before the next pattern is considered. For this
research, batch training is used since it could result in a high precision data mapping [60,

61].

Batch learning was used to train different ANN architectures for the development of an
ANN model able to accurately predict the desired output variable. The tested ANNs
differ by the number of hidden layers and neurons, the transfer functions, and the learning
rate used. Based on results of the learning process, it was determined that 500 epochs,
which is the number of cycles where the training process loops over the training dataset,
was sufficient to build an ANN model, as all the examined architectures were trained
before reaching the 500 epochs.

119

The different ANN architectures to test were developed using MATLAB 6, which is a
numerical computing environment and programming language [62]. MATLAB 6
contains a neural network toolbox that allows users to build, train, and test customized
ANN architectures. The MATLAB code that is used to develop the ANN model can be

found in Appendix A.

As shown in Table 7.3, the different architectures involved different numbers of
processing layers (3 or 4), hidden layers (1 or 2) and hidden neurons (8 to 23). The
learning rate was changed between 0.10 and 0.50. Three types of transfer functions were
further used: logsigmoidal, tansismoidal or purelinas. These are described in Equation

7.4, Equation 7.5, and Equation 7.6 respectively.

logsig (X i)=—l—__ ............................................. Equation 7.4
—X'
1+ e l
. 2 .
tansrg(Xi)=—T — 1 ............................................. Equation 7.5
1+e— ’
purelin (X i)=X i ............................................. Equation 7.6

Where X ,- is the input variable.

For each ANN model, the coefficient of determination R2 was obtained by comparing
travel time predictions with the simulated values. This parameter assesses how much of

the observed travel time variance is explained by the model considered.

120

The best architecture was found to be ANN Model 10. This model exhibits the highest R2
across the training, validation and testing datasets. This model has one hidden layer with
8 neurons and a learning rate of 0.10. The Iogsigmoidal transfer function was used
between the input and the hidden layer, while the purelin transfer function was applied
between the hidden and the output layers. While other models showed equivalently high
R2 values, they had longer training times. The validity of the selected model is
demonstrated by the plots of Figure 7.2, which compare predicted and simulated
(observed) travel times for the various datasets considered. As shown, the comparisons
produced clear linear relationships between the predicted and simulated data, in addition

to a relatively low RMSE.

However, there are a few points that do not follow the expected linear behavior. This is
because the model did not consider all the factors can be associated with the dependent
variable prediction, such as driver’s behavior, vehicle characteristics, and lane changing
models. All these factors had some interaction effects in the simulations from which the
data used to develop the ANN model were retrieved. While excluding the interaction
between these factors and the predicted variable could potentially lower the prediction
accuracy, these factors are difficult to measure in reality, and thus, difficult to incorporate

into the prediction model.

121

 

 

 

 

Table 7.3 Performance evaluation of various neural network configurations
Mpgel L a: er ,, Learartileng Strric'it'i‘re“ Transfer Function Type 82

Input Hidden Hidden

Jr 4r 4’ Train Validate Test
Hidden Hidden Output

1 4 0.50 14-( 15-8 )-1 logsig tansig purelin 0.972 0.956 0.966

2 4 0.50 14-( 15-5 )-1 logsig tansig purelin 0.970 0.958 0.968

3 4 0.50 14-( 10 )-1 logsig tansig purelin 0.972 0.958 0.966

4 3 0.50 14-( 8 )-1 logsig ------- purelin 0.970 0.953 0.960

5 3 0.50 14-( 8 )-1 tansig ------- purelin 0.970 0.956 0.964

6 4 0.50 14-( 8-2 )-1 logsig logsig purelin 0.958 0.949 0.955

7 4 0.50 14-( 8-2 )-1 logsig purelin purelin 0.960 0.943 0.953

8 3 0.25 14-( 8 )-1 logsig ------- purelin 0.972 0.956 0.964

9 3 0.15 14-( 8 )-1 logsig ----- - purelin 0.970 0.955 0.962

10 3 0. 10 14-( 8 )-1 logsig --— purelin 0.994 0.994 0.998

 

t

1".

Number of layers means the number of input, hidden, and output layers
ANN 14 - (10-5) - 1 represents the structure of a neural network has 14 inputs, two hidden
layers with 10 and 5 hidden nodes respectively, and 1 output

 

 

Training Datasets

 

 

 

 

 

 

 

 

 

Validation Datasets

 

 

 

 

 

 

 

 

 

 

 

120 - ' r . . 120
100 ‘ ' 100 '
U) 80 l- - U «- (n 80 l-
'0‘ 0 'o‘
0) .- o
g 60 - I U ‘N E, w r-
‘O . '0
9 .' ‘2
Q 40 i- i. 3’ 0 II Q 40 l-
(Riv
20 - ,5 r 20 -
0 J I I I I o I I I I I
0 20 40 60 100 120 0 20 40 60 80 100 120
Target, 3 Target, 8
Testing Datasets
120 I I I I I 1
Model if 10 in Table 7.3
‘00 Learning Rate 0.10
80 ANN Architecture “Inputs
'0. 8 hidden neurons
8 1 output
3% 60 Transfer Function logsig -> purelin
9 0 Data point
‘1 4o
————-— Linear Best Fit
20
o -------- Target = Predicted
0 I I I I I
0 20 40 60 80 100 120
Target, 3
Figure 7.2 Regression analysis of predicted against observed travel times

122

7.6. Enhanced Active TSP Performance

Following the development of the ANN model, the research focused on integrating active
TSP system and the ANN model within the control system, and examining the
performance of the enhanced TSP system using the VISSIM simulation environment. The
communication between the enhanced TSP signal control system and the simulation
environment is done by using the VISSIM-COM interface according to the flowchart of

Figure 6.3 presented earlier.

Within the integrated model, input variables corresponding to the retained variables listed
in Table 7.2 (variables not highlighted) are sent to the ANN model each time the
upstream check-in detector is activated by a bus. Upon reception of the information, the
prediction model uses it to determine when the approaching bus is most likely to be ready
to leave the bus stop and sends this prediction to the signal controller, which in turn

decides if a priority action is needed, and what TSP action should be granted.

The appropriate priority action is determined at each decision point based on the

following rules:

0 No action is taken for buses projected to arrive during the green.

0 For buses arriving soon after the end of green, a green extension is granted if the
projected arrival is within the maximum allowed extension. Extensions are
awarded in increments of l 3, up to a maximum of 20 s.

0 Green extensions are terminated when a bus crosses the check-out detector.

123

0 Early recalls are implemented for buses projected to arrive during the red beyond
the maximum green extension window but close enough to the end of the red to
be within the limit of permitted alterations.

0 Any phase cannot be shortened beyond its specified minimum green.

0 Each bus is allowed only one priority request. A bus failing to enter the
intersection in time to benefit from a green extension will thus not be considered
for an early recall.

- Buses arriving during the normal green are not considered as they do not require
priority.

- Buses requesting priority and entering the intersection during a provided green
extension (GB) or early recall (RT) are counted as successes.

0 A bus requesting a green extension and arriving too late to benefit from it is
considered as a failure.

0 A bus requesting a green extension but arriving earlier than expected and for
which a priority request cancellation can be issued before the extension is
implemented is considered as a success since the system was able to adjust to the
early arrival.

0 The awarding of early green recalls for buses that end up arriving during the
normal green is considered as a success as reduced delays may still result from

early queue dissipation.

To assess any incremental benefits that may be associated with the use of the ANN
model, the performance of the enhanced active TSP that integrates the ANN model

within the controller is first compared to the performance of a fixed-time traffic signal

124

control system offering no preferential treatments. A second series of comparisons is then
made with a controller making a standard TSP decisions by considering only an average

dwell time and an average travel time between the check-in and check—out detectors.

Impacts of priority actions on transit operations are evaluated by comparing the resulting
average bus delays and bus travel times. Traffic impacts are for their part assessed by
comparing average traffic delay along the east/west and north/south bounds. For buses,
the delays estimated by VISSIM includes bus stop clearance times but exclude all dwell
times as dwelling activities are viewed as normal transit activities. Another performance
measure examines system efficiency by calculating the ratio of buses having benefited
from the system to the total number of buses having requested a priority action.
Successes in responding to priority requests are assessed according to the rules explained

earlier in this Section.

Four evaluation scenarios are finally considered by assuming volume-to-capacity ratios
of either 50% or 85% and average boarding rates of either 30 or 180 passenger/hr.
Similar to previous simulations, each scenario was run for two hours to allow observation
of system performance over a number of buses. For each scenario, 10 replications with
different random seeds were also executed to reduce the impacts of simulation stochastic

effects on the evaluation results.

Figure 7.3 compares the predicted times required by buses to travel from the check-in
detector to the intersection stop line to the actual travel times that were observed in the
simulations. As can be observed, a reasonable agreement exists between the observed and

predicted travel times, which is an indication of the general validity of the developed

125

ANN mode. Moreover, An R2 of 0.8832 is achieved, which means that the ANN model is
explaining over 80% of the variance of the observed results. However, there are some
points that appear to be out of order. This observation can be explained by the fact that
there are many other factors that can affect the accuracy of travel time prediction that
need further investigation, such as lane changing models and other simulation modeling

effects.

 

Time Spent between Check-in and Check-out Detectors

80 . , M-

     
 

Y = 19996".
R2 = 0.8832

0)
O
1

Predicted Transit Travel
J:
O

N
O

 

 

0 20 40 60 80 100 120 140
Simulated Transit Travel Time, s

 

 

 

Figure 7.3 Comparison of simulated and predicted travel times between check-in
detector and stop line for the evaluation scenarios

Table 7.4 further summarizes the measures of effectiveness (MOE’s) obtained for the
four scenarios considered. As expected, transit vehicles benefit from the provision of
priority, with 95% statistically significant delay reductions of 7.0 to 13.0 seconds, or 35%
to 50 % compared to the other two examined signal control systems. The simulation

results further indicate that use of the ANN prediction model generally leads to an

126

additional 1.0 to 3.0 seconds reduction in average bus delay and 4.0 to 10.0 seconds in
average travel time when compared to a priority system considering only average dwell
and travel times. These results are attributed to the better implementation of priority
actions to the needs of approaching vehicles that result from the ANN model’s enhanced

travel time prediction capabilities.

 

 

 

 

 

Table 7.4 Measures of effectiveness for tested controllers
Measure 0 1' (1) (2) Change (3) Change Change
Scenario Effectiveness No TSP Standard (1 )-) (2) Enhanced (1 )-) (3) (2)-9 (3)
TSP 00 TSP 00 0%)
Bus Delay (8) 31.26 18.16 -41.9 ' 15.88 «49.2 ' -12.6 ‘
S a Bus Travel Time (5) 50.29 36.60 -27.2 ‘ 31.36 -41.6 ' 49.8 '
2 g TSP Efficiency (°/.-) ---- 97.00 -.... 99.00 -»- ..
y 3 Traffic Delay (5) 13.60 11.77 -13.5 ' 11.42 -16.1 ' -3.0
Crossing Delay (5) 14.35 17.07 19 " 16.12 12.3 ' -5.6 ‘
0 Bus Delay (5) 19.02 14.26 -25.0 " 12.44 -34.6 ‘ 42.8 '
3 E Bus Travel Time (5) 77.74 72.77 -6.4 ' 63.42 ~23 " -17.7 ‘
3 g TSP Efficiency (°/.) 49.00 ---- 88.00 ----- —-
3 3 Traffic Delay (5) 13.66 12.94 -5.3 ‘ 12.67 -7.2 ' -2.1
F Crossing Delay (5) 14.35 15.25 6.3 ' 14.74 2.7 A —3.3 '
Bus Delay (5) 30.75 18.25 -40.7 ' 15.85 «48.5 " -13.2 '
3 i Bus Travel Time (5) 49.81 36.72 -26.3 ' 30.44 -43.5 ' -23.3 '
2 g TSP Efficiency (%) ----- 98.00 100.00 ---- ---
3 3 Traffic Delay (3) 15.86 14.18 -10.6 ‘ 14.16 -10.7 ' .0_1
Crossing Delay (5) 12.58 14.53 15.5 ' 14.20 12.9 " -2.3 A
0 Bus Delay (3) 19.86 13.45 -32.3 " 12.38 -37.7 ‘ 80 ‘
8 E Bus Travel Time (3) 78.83 72.11 -8.5 ' 66.72 -15.4 " ~7.5 '
i: g TSP Efficiency (%) ----- 64.00 ----- 82.00 ---
3 g Traffic Delay (8) 15.83 14.90 -5.9 ' 15.23 -3.8 ' 2.2
F Crossing Delay (5) 12.65 13.69 8.2 ' 13.35 5.5 ' -2.5

 

 

 

 

Statistically significant at 99% confidence level compared to Case 1, No TSP

Statistically significant at 95% confidence level compared to Case 1, No TSP

127

 

Table 7.4 also shows that buses tend to benefit more from the provision of modified TSP
controller when the transit demand is low, which results in a lower dwell time. This is
due to a reduction in the accuracy of predictions with higher boarding rates. Under low
boarding rates, the standard deviation in the distribution of passenger arrivals could be of
a few persons only. This translates into small dwell time variations and a small window
of potential arrival times at the downstream intersection. If the boarding rate is high, a
similar standard deviation results in a higher variation in the number of persons to serve

and, eventually, wider impacts on dwell time duration and the potential arrival window.

For a certain boarding rate, it is further observed that traffic congestion level only has a
slight impact on transit benefits. This observation is expected, as buses are assumed to
have priority to merge in the traffic stream when leaving the bus stop. While this priority
helped the buses reentering traffic in the simulations, it did not completely eliminate
associated delays, as buses still had to wait on average between 1.0 and 16.0 seconds to

reenter traffic.

Results further show that system efficiency is the highest when the enhanced active TSP
is implemented. This is due to the higher probability of successfully implementing
priority actions from which the buses can fully benefit as a result of a higher accuracy in

predicting bus arrival times.

Signal priority is finally observed to reduce arterial traffic delay between 4% and 16%
and increase cross-street delay between 3% and 13% when ignoring the type of priority
algorithm used. However, it can be observed that use of the enhanced TSP tends to

produce slightly higher arterial delays and lower cross-street delays when compared to

128

scenarios with no TSP or standard TSP treatments. In this case, a controller relying only
on average values implements a certain number of signal timing alterations that may end
up not used by approaching buses. Use of the ANN model enhances prediction
capabilities then provides more opportunities to ignore priority requests that may end up
not benefiting approaching buses. This translates into fewer green extensions and early
recalls. Therefore, less effective green time is lost, less delays for the traffic traveling
along the bus route, and less disruptions for the cross-street traffic is experienced in the

traffic network.

7 .7. Summary

An ANN prediction model is developed to predict the stop line arrival time of buses
traveling on intersection approaches with near-side bus stops. Results of integrating the
ANN model with active early green and green extension TSP algorithm shows an ability
to improve the benefits that transit signal priority systems can provide to buses by
enhancing the ability to correctly predict bus arrival times at intersections and reduce
unnecessary signal timing adjustment. In the test scenarios, this enhanced prediction
capability resulted in greater delay reductions for buses and lower traffic impacts than a

priority system simply basing its decision on average dwell and travel times.

129

CHAPTER VIII.

D-SPORT FRAMEWORK
INTEGRATION

 

his chapter describes the proposed optimization model that has been developed
within the VISSIM simulation software for implementing TSP within
coordinated networks of intersections. This development is based on the use of
the VISSIM-COM interface to allow communications between the GA optimization

process, the ANN prediction model, and the simulation environment.

8.1. Genetic Algorithm Platform

The first step in the development of the optimization process was to develop a GA-based
optimization procedure utilizing the general cost function that has been developed in
Chapter 6. The GA platform and the D-SPORT controller cost function were fully written
using the FORTRAN 90/95 programming language. The written code was developed by
taking into account the need for compatibility with the simulation VISSIM-COM

interface and the simulation environment, which restricts the use of pre-coded GA tools,

130

such as the GA toolbox in MATLAB. The selection of the FORTRAN 90/95
programming language over other more commonly used languages is based on the ability
of FORTRAN 90/95 programs to effectively handle number crunching operations. For
the problem at hand, it was assessed that the use of the FORTRAN 90/95 language would
provide powerful computational performance and faster computational speeds compared
to other programming languages [63]. The FORTRAN 90/95 code that was developed

can be found in Appendix B.

Figure 8.1 illustrates a flowchart for the GA process that was implemented. The figure
represents the GA optimization platform that appears in Figure 6.3. Once the controller
starts the GA optimization process, the GA parameters, such as population size,
maximum generations, mutation rate, and crossover rate, are passed to the GA platform.
Following reception of these parameters, the GA platform generates the initial population
of potential solutions. Traffic and transit data collected from the simulation environment
are then fed into the cost function and used to evaluate each individual within the current
population of solutions. Following this evaluation, the GA platform then starts the
selection process and initiates the generation of the next generation of solutions by apply
GA operations, such as mutation, crossover, and elitism, to the current population of

solutions.

The above process is iterated until the GA platform reaches the maximum allowable
number of generations, or the GA termination criterion is met. Once a near-optimal
solution is found, the solution is transmitted to the D-SPORT controller, which

implements it in the simulation environment.

131

The following subsections discuss in more details the modeling and operation of the GA
optimization process. Attention is first put on the binary coding and decoding for
population strings. This is followed by examinations of the roulette-wheel and
tournament selection methods. The GA performance under different mutation and
crossover rates is then addressed, and the influence of elitism on the GA operation is

discussed.

132

 

 

F Start

1
READ GA Parameters

 

 

 

 

Y

 

 

READ Delay Data

READ Queue Data

READ Travel Time Data
READ Traffic Volume Data
READ Signal Control Data
READ Transit Location
READ Bus Stop Data

 

 

 

Y

 

Create initial Random
Population

 

 

 

l

V
DEOR gen = 1, generationw
v
{FOR pop = 1. population )1

i
Y

Calculate Delay
Calculate Stops

 

 

 

 

 

 

Y
Calculate Delay
Calculate Stops

 

Pop++

 

 

 

  

is TSP Enabled?

  

 

 

 

 

 

~77 Gen++

 

 

 

 

 

 

 

 

 

 

No
Yes
1
Call ANN 5
, Y
i TTT = 0.00
Pity = 0.00
Calculate TTT
Calculate Pity
v'

 

Calculate Fitness L

 

 

 

 

 

   

MAX(fitness)

 
 

Mi N(fitness)
< 10 ES

    
 

NO

Y

 

Select 8. Mate Parent
Chromosomes

 

 

 

V
Crossover
Mutation
Elitism

 

 

 

 

 

 

-j YES
V
New
Population
is
Generated

 

 

 

 

 

       

Decode Best
Chromosome

  

Y

 

SAVE outputtxt

 

 

 

 

Population: Population Size

Generation: Maximum Number of generation

Cross __p: Cross over probability

Mutate_p: Mutation Probability

Gen, pop: Counters

Fitness: String fitness resulted from the
objective function

 

 

 

 

Figure 8.1

The genetic algorithm optimization process flowchart

133

 

8.1.1. Solution String Modeling

In GA, each possible solution is represented in a chromosome (also called string). In a
classical GA, a binary coding is used to represent the chromosome. Each chromosome
consists of a number of bits that could have a value of either 0 or 1. Individual decision
variables are represented by a number of bits that is determined based on the potential
range of the variable range and its increment. As an example, Figure 8.2 shows a
schematic sketch for a chromosome modeling a number of decision variables. In this
example, the variables DVI, DV; and DV3 represent the green time assigned to a specific

approach. For the problem at hand, the green time can range between 20.0 and 40.0

seconds, at an increment of‘ 0.5

sufficient to cover the all the possible values that the variable modeling green time could

take.

seconds. Based on this information, 6 bits are found to be

 

 

 

[0110111000000

 

 

 

 

 

 

k J\ J
Y Y

0v1 ov2

 

 

 

011011<l—:>

543210

DV,: Decision Variable i

111111 ------------------------ [001001
\ Y J H__J
DV3 DVi

0x(25) + 1x(24) +1x(23) +0x(22) +1x(2‘) +1x(20)
0x(32)+ 1x(16) +1x(8) +0x(4) +1x(2) +1x(1)
O + 16 + 8 + 0 + 2 + 1

27

111111 <:> Binmax= 63.0 (j:> DVmam
011011 <:’> BinDV1=27 C:> E9V1
000000 <:> Binmin= 0.00 <:(> DVmin

 

 

Figure 8.2 Binary system representation and decoding

134

 

Conversion from a binary system to a decimal representation is done using the following

equation:
P
Bin DVi = Zbin x 21) ............................................................ Equation 8.1
p=0
Where:
Bin DVi : Decimal representation for decision variable i
bin p : The bin value in place p
p : The bin relative location within the chromosome part that represents
decision variable i

P: The total number of bins occupied to represent decision variable i

In Figure 8.2, all the bits associated with the decision variable DV2 have a value of 0.
This corresponds to a decimal representation of a 0.00. For the variable DV3, the
assignment of a value of l to all the bits associated to the variable results in a decimal
representation of 63.00. The sequence of Os and Is associated with the variable DV;

finally results in a decimal value of 27, as illustrated in the figure.

As defined in Chapter 6, the lowest value a decision variable can take is the minimum
allowable green. A value of 0 for the variable DVZ is therefore out of range as the pre-
defined minimum allowable green is 20 seconds. For the variable DV3, 3 decimal value
of 63.0 is also out, of range, being beyond the maximum allowable green time of 40
seconds. To ensure that only valid green times are considered, interpolation is used to

determine the actual value of the decision variable value, rounded to the nearest 0.5

135

second, based on the variable’s minimum and maximum value constraints. Equation 8.2

shows the equation used to perform this calculation.

 

Bin .-Bin -
INT 2X[ . DV' . min X(DVmax ‘DVmin)+DVmin]
Bmmax — Bmmin

 

DVi = 2 Equation 8.2

Where:

DVi: Decision variable 1' actual value

Binmin: The minimum decimal representation a decision variable can ever take,
when all the bits are 0’s. It has a value of 0.00 with 6 bins per decision
variable

Binmax : The maximum decimal representations a decision variable can ever take,
when all the bits are l’s. It has a value of 63.00 with 6 bins per decision
variable

DVmin : The minimum allowable value a decision variable can take, which is equal
to the minimum allowable green of 20 seconds

DVmax :

The maximum allowable value a decision variable can take, which is equal

to the maximum allowable green of 40 seconds

8.1.2. Genetic Algorithm Operations

This section details the selected operations for the GA platform that has been developed.
Elements described include GA search termination criteria, population size, the selection
method, the crossover operator, the mutation operator, the elitism operator, the

population size, and the number of generations considered.

136

For the evaluations, the maximum number of generation was set to 1000 generations, as
initial experiments showed that executing longer runs had negligible improvement effects
on the cost function values. The population size was further set to 300 chromosomes as
initial experiments also showed that a population size of 300 with a maximum of 1000
generations generally provide good progress in the solution search process for the

problem at hand.

Figure 8.3 shows the combination of GA operational parameters that were tested prior to
selecting a set of preferred parameter values. Parameters considered include two selection
methods, five crossover rates, and five mutation rates. Selection of the parameter values
to adopt is based on performance of the optimization process on a Dell Dimension 3000
desktop with a central processor unit (CPU) speed of 2.80 GHz and a random access

memory (RAM) of 256 MB.

 

 
 

Crossover Rate

50% i ,. .
Configurations
% . '

70%

   

Selection Method

 

 

Tournament

 

 

90%

 

 

 

 

Figure 8.3 The examined genetic algorithms configurations

137

Termination Criteria

Termination criteria were added to the GA search process to avoid extensive
computations yielding little solution improvements. In this case, the GA was instructed to
terminate the solution search process if (a) the maximum number of generations is
reached or if (b) the absolute difference between the maximum and the minimum cost

function value for any generation is found to be less than 105.

Selection Method

Selection is the process by which individual solutions (strings) are selected for
contribution to the generation of next batch of solutions. In this research, two selection
methods are examined, the roulette wheel selection and the tournament selection

methods.

The roulette wheel selection algorithm is utilized to perform the parent’s string selection
process. For each string, the probability of its selection is equal to the ratio of its own cost
function value to the sum of cost function values of all strings within the same
generation. Once the probability of selection for each string is computed, the cumulative
probability is then computed for each string in the generation in the order of their
existence within the same generation. The next step involves generating a random
number varying between 0.0 and 1.0. This number is used to identify strings with a
maximum cumulative cost that is less than or equal to the generated random number. The
identified string is then selected and placed in the mating pool. This operation is iterated

until all the parents are selected.

138

For the tournament selection, strings are randomly grouped in pairs, and then from each
group, the string with the lowest value for the cost function is chosen for inclusion in the
mating pool to generate offspring for the next generation of solutions. For the problem at
hand, the binary tournament selection was examined, where each group contains two

strings that are selected randomly.

Crossover Rate

Crossover is defined as the genetic process of combining the genes of one chromosome
with those of another to create offspring inheriting traits of both parents. For this
research, a simple crossover technique is used in which genes are exchanged between two
chromosomes around one selected point. For the two selected parents from the mating
pool, a random number between 0.0 and 1.0 is first generated. Crossover then only occurs

if the generated random number is less than a pre-defined crossover rate.

In this research, different crossover rates are tested for the given population size of 300
and the maximum number of generations of 1000, along with the roulette wheel and
tournament selection methods discussed earlier. A total of 5 crossover rates are
examined. These rates vary between 50% and 90% with an increment of 10%. Based on
the sensitivity analysis of the cost function value to the crossover rate, a value of 80%

crossover was selected. The performed sensitivity analysis is discussed in Section 8.1.3.

Mutation Rate

The mutation operation is helpful in a GA search to avoid being trapped within a local
optimal solution. Mutation occurs when a randomly selected gene swaps its value, from 0

to l or from 1 to 0. For the problem in hand, a bit-wise mutation is used, where the

139

mutation rate identifies the probability of mutated genes, and all the genes in a population
have an equal chance of being mutated. In order to determine which bit (gene) will swap
its given value, a random number between 0.0 and 1.0 is generated. If the generated
random number is less than the mutation rate, then one gene in the population is
randomly selected and subjected to the mutation operation. This process is iterated until
the expected number of mutated genes is reached. The expected mutated genes can be
simply calculated as the total number of bits in the population times the bit-wise mutation

rate.

A total of 5 mutation rates were examined. The rates that were examined varied between
0.2% and 1.0%, with an increment of 0.2%. As shown in Figure 8.3, these mutation rates
are tested along with the crossover rates and the selection methods that are mentioned
earlier in order to determine the GA parameters that will be used to optimize the D-
SPORT controller cost function. However, the 0.2% mutation rate is selected as

discussed Section 8.1.3.

Elitism

Elitism is defined as keeping the best string of a given generation when developing the
next generation of strings. Elitism seeks to retain unchanged one or more strings having
the lowest cost function value within a given population to the next population. This is to
ensure that the best solution is not lost when going from one generation to the next. For
the developed GA platform, elitism is applied based on observations that it results in

better GA performance, as discussed in Section 8.1.3.

140

8.1.3. Genetic Algorithm Operational Performance

Figure 8.4 and Figure 8.5 show the cost function value averaged across 50 runs with
different random seeds for different selection methods, crossover and mutation rates,
where the objective is to minimize the cost function value. The selected number of runs
returns a statistical power of at least 80% for all the pairwise comparisons. These figures
show a statistically significant difference between the two selection methods, with a
confidence level that exceeds 95%. The results show that the roulette wheel selection
methods resulted in lower cost function values compared with the values that resulted
from tournament selection. The final GA parameters that are used are the ones associated
with the best (lowest) cost function values, which are those associated with the use of

roulette wheel selection, with 80% crossover rate, and 0.2% mutation rate.

141

 

Cost Function Value Using Tournament Selection
Population Size: 300 Maximum Generations: 1000 # Runs: 50

7750« ~ , 7 , 7 v 7 ,, 4

30.2% Mutation Rate
00.4% Mutation Rate
00.6% Mutation Rate
00.8% Mutation Rate
I 1.0% Mutation Rate

 

 

Cost Function Value

 

 

 

 

 

 

 

50% 60% 70% 80% 90%
Crossover Rate, %

 

 

 

Figure 8.4 Average cost function value for tournament selection, for different
crossover and mutation rates

 

Cost Function Value Using Roulette Wheel Selection
Population Size: 300 Maximum Generations: 1000 # Runs: 50

 

 

7750 - 77 7 77 7 7 7 777 7 7777
7700 7 7 7 7 77 777777777 7 77
o
2
g 30.2% Mutation Rate
7650 - 77 7 7 7—7 7 77 7 777777777
5 I0.4% Mutation Fiate
E 0 0.6% Mutation Fiate
E 00.6% Mutation Rate
7600 7 7 7 77777777 7 77 77 .
7;; I 1 .0% Mutation Rate
0
o
7550 - 7
7500 7

 

 

50% 60% 70% 80% 90%
Crossover Rate, %

 

 

 

Figure 8.5 Average cost function value for roulette wheel selection, for different
crossover and mutation rates

142

8.1.4. The GA Near-Optimal Solution Search Convergence

The computational performance of the GA platform using the roulette wheel selection is
discussed in this section. Figure 8.6 summarizes the total number of function evaluations.
For each examined GA configuration, the number of function evaluations is averaged
across the 50 GA runs with different random seeds. The number of function evaluations
increases with increasing crossover and mutation rates, because the probability of
producing and evaluating new individuals is a function of these two rates. As shown in
Figure 8.6, the GA configuration with a crossover rate of 90% and a mutation rate of 1%
has the highest number of function evaluations, and the configurations with a crossover

rate of 50% and a mutation rate of 0.2% has the lowest number of function evaluations.

On the other hand, and as shown in Figure 8.7, there is no clear relationship between the
CPU runtime and the number of function evaluations performed. However, it can be seen
that with increasing the crossover rate, the runtime tends to be less sensitive to the
variation in mutation rate. This is because the probability of evaluating more new
individuals increase with higher crossover rate, and applying mutation to these new
individuals does not change whether or not they need to be evaluated. In other words, the
GA needs to evaluate these new individuals regardless of whether they are mutated or

not.

The influence of the crossover rate on the GA search convergence progress versus
number of generations and number of function evaluations is examined. The relationship
between improvement of the cost function values and the number of generations, and
between the cost function values and number of function evaluations performed, are

shown in Figure 8.8 and Figure 8.9.
143

 

Average Number of Function Evaluations

 

265000 ,
240000 «
215000 .
190000 1

165000 i

 

140000j

 

Average Number of Function Evaluations
Population Size: 300 Maximum Generations: 1000 # Runs: 50

Selection Method: Roulette Wheel

00.2% Mutation Rate

I 0.4% Mutation Rate
00.6% Mutation Rate

00.8% Mutation Rate
I 1.0% Mutation Hate

 

 

 

 

 

 

 

50% 60% 70% 80% 90%
Crossover Rate, %

 

Figure 8.6

Number of function evaluations for different GA configurations

 

Average CPU Runtime, 5

 

65-

62-

 

 

50% 60% 70% 80% 90%
Crossover Rate, %

Average CPU Runtime
Population Size: 300 Maximum Generations: 1000 # Runs: 50

Selection Method: Roulette Wheel

5 0.2% Mutation Fiate
I0.4% Mutation Rate

00.6% Mutation Rate
00.8% Mutation Fiate
I 1 .0% Mutation Rate

 

 

 

Figure 8.7 Average CPU runtime for different GA configurations

144

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Cost Function Value vs. Number of Generations
Population Size: 300 Maximum Generations: 1000 # Runs: 50
8600 - 7 F — — — # -' # #FSeilectionflMfethod: Roulette Wheel
Mutation Rate: 0.2%

8400— — 4 4 i ,_E,__ —~— ~———
0
3
'5
>
c 8200 7
.g — 50% Crossover Rate
cs» 7 - 60% Crossover Rate
1.1: 8000 ~ 7- - 70% Crossover Rate
3 80% Crossover Rate
0 7800 .. —— 90% Crossover Rate

7600 J7

7400 I l I T l

O 200 400 600 800 1000
Generation Number
7550 7 g a.
l .

7545 \7
.3 \t‘ , \*
a l u ‘
> ~. . \
8 7540 \\ \
g: (5) xx .. \(1) 1 ——~- 50% Crossover Rate
E K 1, - , 2 77 60% Crossover Rate
5 7535 \i 2 \\X 3 70% Crossover Rate
8 \(. ) ' » (3) ' 4 80% Crossover Rate
0 7530 if ‘ ‘ 5 ~ 90% Crossover Rate

Lx‘ »
\_“\-‘
7525 \‘ _.\
(4) “
7520 1 r r . 1
0 200 400 600 800 1000
Generation Number
Figure 8.8 GA performance over number of generations for different crossover

rates and a mutation rate of 0.2%

145

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Cost Function Value vs. Number of Evaluations
Population Size: 300 Maximum Generations: 1000 # Runs: 50
86006 e — , E , — - ,._0 -— -_.__
Selection Method: Roulette Wheel
Mutation Rate: 0.2%
8400~ ~ 4 ~ , f e v#«-u~—--—
0
2
I
>
382004— v,, ~4ffi*v
z — 50% Crossover Rate
:3: 60% Crossover Rate
E; 8000 7 # ’fi 7 1’ f — ——' —— — ‘f” 7777 70% Crossover Rate
3 80% Crossover Rate
0 —-— 90% Crossover Rate
0 50000 100000 150000 200000
Number of Function Evaluations
7550 -i , - \ AAAAAAAA —— —— -— AA
'2- \\x
g 7545 , ~ 4: K A — -——— , e _ , ~ ~
— K \
'>" \ \
S 7540-4 ~~ ~~x \ -- .222, -
3': \K N) 1 —— 50% Crossover Rate
g , fix \ 2 - 60% Crossover Rate
2: 7535 “T 7 7A__;\ \.\\ ’ 7f \ " ’ ”'” ’7 " " 7 3 70% Crossover Rate
a (2) \\ 4 80% Crossover Rate
0 7530 - ,,,,,, ' . ‘j‘\¥\;v“ ,_ A , _ _ , 5 ——-- 90% Crossover Rate
\.\\\ (3)
7525 .2, - E—E— v - — 4— 757x. ‘47
«.. “a. (5)
(4)
7520 i . i .
0 50000 100000 150000 200000
Number of Function Evaluations
Figure 8.9 GA performance over number function evaluations performed for

different crossover rates and a mutation rate of 0.2%

146

Figure 8.8 and Figure 8.9 show that a GA configuration with 80% crossover rate and
0.2% mutation rate converges to a near-optimal solution in early generations. However,

and for later generations, there is some improvement observed for that GA configuration.

Figure 8.10 shows the cost function improvement over number of generations with and
without applying the elitism operation, when 80% crossover and 0.2% mutation rate are

used. The results show that applying elitism has improved the GA convergence.

In conclusion, and according to the previous discussions regarding the performance of the
GA under different operational schemes, a GA using roulette wheel selection, 80%
crossover rate, 0.2% mutation rate, and elitism was integrated within the simulation

environment to perform the required traffic signal timing optimizations.

 

Cost Function Value vs. Number of Generations
Population Size: 300 Maximum Generations: 1000 # Runs: 50

 

 

8600 — —— ~— _~ ~ ~- - 2....
Selection Method: Roulette Wheel
Mutation Rate: 0.2%
— —— —-——-- -— — - —————- — Crossover Rate: 80%
-With Elitism

-— Without Elitism

 

Cost Function Value

 

 

 

 

 

74 00 r r f r r T r r r 1
O 1 00 200 300 400 500 600 700 800 900 1 000

Generation Number

 

 

 

Figure 8.10 GA performance over number of generations with and without
Elitism

147

8.2. Integration of D-SPORT within Simulation Environment

The GA executable file resulted from compiling the FORTRAN 90/95 coding was
integrated within the VISSIM microscopic simulation environment. The integration of
the GA optimization model was achieved through the software’s Component Object
Model (VISSIM-COM) interface, which is an external module that enables
communication and dynamic object creation between the simulation environment and
external processes [58]. This interface facilitates extracting traffic volumes, queuing
conditions, bus stop demand, and transit vehicle global position, current signal

parameters and local timings, similar to what can be extracted under real-life conditions.

The data required to calculate the cost function value are collected from the simulation
environment through the VISSIM-COM interface and then made available to the GA
platform. When it is time to call the GA, the VISSIM-COM calls the GA routine, which
then reads the inputtxt file and uses data contained in this file to generate a signal tinting
solution for the next three signal cycles. Once the signal timings are generated, the
VISSIM-COM passes the signal timings to the signal controllers in the simulation

environment. The simulation is then continued until it is time to call the GA again.

In order to determine when the GA platform should be called, a countdown counter is
used. This counter starts with a time that is equal to the minimum cycle length of the
near-optimal solution resulted from a given optimization call. The counter starts counting
down till it reaches the zero, in which the GA optimization is called again. This process is
iterated throughout the simulation duration. This philosophy is adopted to take into
account that the stochastic nature of traffic conditions may create a need for updated

solutions in a relatively short term period.

148

The flowchart shown in Figure 8.11 illustrates the integration process between the
D-SPORT model and the simulation environment. The controller collects traffic and
transit data, and once it is time to trigger the GA, the controller passes the collected data
to the optimizer. The optimizer finds a near-optimal solution based on the passed data,
and returns the solution to the controller, which implements the solution within the

simulation environment.

149

Figure 8.11

 

 

Start VISSIM—COM interface

 

 

 

+
Call VISSIM Simulation
Environment

 

 

 

 

v
,. , { FOR Sim=1,Simulation Period ]

 

Time to _' E__,
call GA? 8

i YES i
l v
i SAVE Simulated Traffic Conditions

1
SAVE Signal Control Timings l
SAVE Transit Related Data 2

 

i l
I .

v

 

i Cali GA Optimization Piatforrn

l
r i l
. / . . 1
. For Each Intersection 1n 7
l

Network

 

 

 

 

 

177+

 

 

i
. l
; i i ;
.. i V l
- Save New Signal [

Timing Parameters

 

l
i
|
r

v

 

Continue with the Simulation <~#~4

 

 

 

L__-——< ,=W___ 4 , H a, 272...-“

 

 

 

The integrated system architecture for the VISSIM-COM Interface
within the simulation network framework

150

8.3. Experimental Setup

This section describes the VISSIM simulation model that was developed to study the
performance of the D-SPORT controller. Elements described in this section include the
geometric layout for the network, traffic and transit demands, and the alternative traffic

control logic considered.

Geometric Layout

The network used to test the model is shown in Figure 8.12. This network consists of 7
intersections and 14 links. It models two intersecting one-way streets with coordinated
signals. Link along the two modeled streets features two travel lanes and an exclusive
left-tum bay. The Links modeling minor cross-streets have the same features, except that

they do not include exclusive turn bays.

Intersections at the upstream ends of the network (intersections 1 and 7) are used to
minimize the influence of the boundary conditions and shape arriving traffic into realistic
platoons. For this reason, stops and delays incurred at these intersections are excluded
from the analyses detailed later in the dissertation. For each approach, detectors are
placed at the beginning and the end of the approach to collect traffic volume, delay, and

queue length data.

151

 

 

 

 

 

 

 

 

 

 

50011 . 700 ft 1 800 ft J 800 ft . 500 ft .
r21; '33 iii
L: ,3: gr pf. “-
Int. 1 51 Int. 2 : Bus Int. 3 3
iii “:1
-1 : Stopl :3 750 ft
, a a
. 500 it
1:1 East Bound Tran Sit Route
- North Bound Transit Route Egem . l-
'- BUS Stops 85011
Speed Limit 55 km/hr (35 mph)
N di—
W E 500 ft
5 1.
Figure 8.12 Simulated network layout.
Traffic Signal Control

Each intersection is assumed to be controlled by a two-phase signal providing 4-s of
amber and a 1-s all red within each phase. The amount of green time assigned to each
phase will depend on the type of controller being tested and can be either fixed of
variable. Depending on the traffic conditions, the cycle length is allowed to vary between
50 and 90 seconds, and the green time is allowed to vary between 20 and 40 seconds. A
50-second minimum cycle with minimum green restrictions is chosen to ensure that the

timing plans being considered do not unduly restrict pedestrian crossing times. Green

152

interval duration for each phase and offsets might vary from cycle to cycle based on the

cost function optimization results.

The amount of green time allocated to each phase depends on the type of signal controller

being considered. To evaluate the effectiveness of the D-SPORT, scenarios considering

various types of controllers were developed:

1.

Pre-Timed, which is a pre-timed signal control type hat has an offline
optimization signal timing plan with no TSP implementation. This controller is

optimized using SYNCHRO 6 for each traffic volume combinations.

Pre-Timed TSP, which is a pre-timed signal control, similar to the Pre-Timed
controller. However, this controller offers early green and green extension active
TSP treatments that are implemented locally. The traffic signal timing plans are
optimized using SYNCHRO 6. The standard TSP logic, which only considers
average dwell time is the one used to offer green extension and early green active

priority treatments, as described in Section 7.6.

Real-Time without TSP: this signal control is similar to the proposed GA-Based
optimizer while the TSP sub-cost function is inactive. In other words, the
parameters apl , 0777 and “SCH are set to 1.0, 0.0, and 0.0 respectively in
Equation 6.6, the equation that represents the D-SPORT controller cost function

formula.

D-SPORT, the real-time signal control optimization with integrated preferential

treatments for transit vehicles.

The first two scenarios represent traditional fixed-operations. They implement signal

timings generated before the start of a simulation based on average traffic conditions and

that do not change with fluctuations in traffic demand, except for the provision of priority

actions to accommodate buses. Signal timings plans in both of these two scenarios were

153

developed using SYNCHRO 6, which is a commonly used traffic signal optimization

software.

The scenario against which all other scenarios will be compared is the first one in the list,
(i.e., the one featuring fixed-time control without TSP). Comparisons to this scenario will
allow to evaluate how buses benefit from the provision of TSP and then to determine the

incremental benefits provided by the D-SPORT controller.

Traffic Demand

To remove truck effects, only passenger cars are simulated. Vehicles are generated at the
upstream end of each entry link using a Poisson distribution and VISSIM’s default
vehicle characteristics. To allow testing in a range of traffic conditions, traffic volumes
covering low, moderate and heavy traffic conditions along the two major corridors are
considered. This meant developing scenarios with volume to capacity ratios (v/c) of 50%,
75%, and 95%. In all scenarios, traffic volumes along the minor streets are kept at a fixed
v/c ratio of 50%. The proportion of traffic that turns left or right on each approach is also

assumed to be constant, at 15% of the approach volume.

Transit Demand

Two transit routes are assumed traveling along the network, one in the eastbound (EB)
direction and one in the northbound (NB) direction. This will allow for the evaluation of
the ability of the D-SPORT controller to generate efficient solutions at the intersection of
the two one-way streets when conflicting priority needs exist. For each transit route, a
transit vehicle is generated to enter the system following a normal distribution with an

average headway of S-minutes and a standard deviation of 20-seconds to take into

154

account the variability associated with the bus entrance to the network. Two bus stops are
placed along each transit routes bus routes. Furthermore, it is assumed that transit
vehicles are equipped with automatic passenger counts in order to help predict the
duration of the dwell time at bus stops, and a Global Positioning System (GPS) to track
the location of the transit vehicles while traveling along the network. It is further assumed
that all the bus stops have identical boarding rates following a Poisson distribution.
Boarding rates considered across various scenarios include rates of 50, 100, and 150
passenger/hr. An alighting rate of 10% of transit vehicle occupancy was finally

considered to simulate individuals getting of individual vehicles.

Figure 8.13 summarizes all the possible experiment combinations that were examined.
For the total of 72 scenarios, the difference between them is by changing the signal
control, the traffic demand along the eastbound and the northbound, and the average

passenger demand at bus stops.

 

Signal Control Type

  
 

50%-50%

Pre-Tlmed TSP
RT w/o TSP
D-SPOHT

75°/o-50°/o

75%-75°/o

 

95%-50%
95%-75%
95%-95%

Figure 8.13 Sketch summarizes all the possible combinations that are tested

 

 

 

 

1

U!

5

CHAPTER. IX.

SIMULATION RESULTS AND
COMPARISONS

 

n this chapter, the results of the simulation experiments described in Chapter 8 are
presented and discussed. The results are broken into different MOEs. Total delay
and number of stops are discussed for both traffic and transit movements. Service
reliability and schedule adherence are further considered for simulation results pertaining
to transit operations. The four signal control algorithms and test network described in
described in Section 8.3 are used to evaluate the improvement achieved by the D-SPORT
controller over current signal timing optimization practice. To account for stochastic
effects in simulations, all the results summarized in this section are based on 10
replications with different random seeds, with each replication covering a 2-hr simulation

period.

The simulation results show that the proposed D-SPORT controller is successfully able to

respond to the changes in traffic conditions, to provide effective priority to buses while

156

 

maintaining coordination across adjacent intersections, and to solve difficulties associated
with dwell time variability and conflicting priority requests. This success is expressed by
reduced transit and general traffic delays and number of stops, as well as by better

schedule adherence and service reliability.

9.1. Transit Travel Time Prediction

As described in Section 6.3.2, transit travel time is predicted through macroscopic
analysis and ANN predictions as the summation of free flow travel time, dwell time, and
delay encountered by a bus along its travel path. To quantify the performance of the
developed transit travel time prediction model, the simulation network, traffic volumes
and transit demands described in Section 8.3 are used to generate traffic scenarios and

measure transit travel times along the two modeled transit routes within each scenario.

While field testing would be an ideal evaluation tested, simulation is used to validate the
transit travel time prediction model due to the difficulties in setting up a validation
environment in the real world. Conducting evaluations in the real world would require
equipping a bus and traffic signal controller with the necessary detection and signal
control altering equipment, which is beyond the resources available for this research

project.

To ensure the validity of transit travel time predictions, the model was tested with
random pre-selected cycle lengths and splits. For each controller, different cycle lengths
and splits were randomly generated and assigned for the 2-hour simulation duration.
Although the cycle lengths and splits are randomly chosen, they all respect the minimum

and maximum allowable green times for individual phases defined in Section 8.3. These

157

parameters were introduced into the VISSIM simulation environment using the VISSIM-

COM interface.

While the simulation is running, a snapshot is taken every 60 seconds. When each
snapshot is taken, relevant traffic and transit data is collected for each transit vehicle in
the network. These data include transit identification number, route, location, speed, data
described in Chapter 7 required to predict dwell time activity, and data used to assess
travel time volatility. For each transit vehicle in the network when the snapshot is taken,
the transit travel time prediction model is used to predict the anticipated time required by
the transit vehicle to reach the end of the network. This prediction is then compared with
the actual time taken by the vehicle to leave the network to assess performance of the

prediction algorithm.

As illustrated in Figure 9.1, predicted transit travel times were found to generally match
actual simulated transit travel times. A coefficient of determination (R2) value of 0.9693
is further computed. Within this context, it is assessed that the model is capable of
providing the cost function with a reliable and accurate value for transit travel time

prediction model.

158

 

 

 

 

 

 

 

 

 

 

 

Simulated vs. Predicted Transit Travel Time
270 - _ _ __ __. ._ N __ __.- _4 __. _z..._ -.. . ._ _ __ _____ __ __ __
i .

24o+———~-w~re+v~—+e * w— . H10—
0) 3’ 0’:
a; -- I __I __ f__ f X I _ __.“ __._4
g 210 . ’.
'— J . O
a 180 »— — — « ~————#———————~. u; ——»—
5 “ ° y- 1 0182x
t 150 l— w * —~ ~— §——————w {4mm --
5. .o’ ,. R =0.9693
grzoa ~ F— l, __a_- —*———--—-—'——-
I: , , g)
3 90~—-— 7.. — -— -————~—~-—~——e
.13 °’
8 604 ~—— ,- ’ __w___ ———~ _IX_W vfl—w—
6: i o

30 / -—- fl *5 — —~ ~* — ————--

O " I I I I I I I I —l
0 3O 60 90 120 1 50 180 21 O 240 270
Simulated Transit Travel Time, s
Figure 9.1 Comparison between simulated and predicted Transit Travel Time

9.2. Overall Network Performance

An overall look at network performance is performed by linearly combining average
traffic delay, average number of stops per vehicle, average transit travel time, and
average transit lateness at bus stops. These parameters are considered based on their
inclusion in the D-SPORT controller cost function that is optimized by the GA

optimization model (see Equation 6.6) when determining appropriate signal control

strategies.

In this research, network performance is estimated by using

generated at the end of each 2-hr simulation period. To illustrate how overall network

performance is computed, consider the following example. First, it is assumed that the

simulation statistics

following statistics summarize simulation results from a given scenario:

159

 

0 Average vehicle delay for each phase group (total of 14 phase groups): 15.0 s/veh
0 Average number of stops for each phase group: 1.2 stops/veh

0 Average transit travel time for each transit route (total of 2 transit routes): 250.0 s
0 Average transit lateness for each bus stop (total of 4 bus stops): 50.0 s

From the statistics described above, the overall network performance can be computed

as:
ar31
COST=[ZPI 2m ZSCHl- am
“SCH
1.0
COST=[14x15.0+10x(14><1.2) 2x250.0 4x50.0]-10.0
10.0

Overall Network Cost Function Value = 7378.0
Similarly estimated network cost function values for different levels of transit and traffic
demand are shown in Figure 9.2, Figure 9.3, and Figure 9.4. As can be seen in the
figures, the proposed D-SPORT controller produced the lowest cost function value in all
cases, which is what was desired. This observation is consistent with the GA
minimization process, which looks for a solution that minimizes traffic delays, number of
stops per vehicle, transit travel time, and transit lateness at bus stops. Although this
conclusion is based on an aggregate value of all MOEs considered, the following sections
focus more on specific each measure of effectiveness used to calculate the overall

network performance.

160

 

Network Cost Functlon Value

 

 

 

Overall Netwrok Cost Function Value
Boarding Rate, 50 Passenger/hr

   
 
 

 

' Pre-Trmed
I Pre-Tlmed TSP

0 Real Time w/o TSP
Cl D—SPOFlT

 

 

Different Volume-to-Capaclty Comblnatlons Scenarlos

 

 

Figure 9.2

Overall traffic network cost function value under different traffic
demand and 50 passenger/hr transit demand

 

Network Cost Function Value

 

 

 

Overall Netwrok Cost Function Value
Boarding Rate, 100 Passenger/hr

   
   

 

B Pre—Timed
I Pre—Trmed TSP

[3 Real Time w/o TSP
D D-SPORT

 

 

leferent Volume-to-Capaclty Comblnatlons Scenarlos

 

 

Figure 9.3

Overall traffic network cost function value under different traffic
demand and 100 passenger/hr transit demand

161

 

Overall Netwrok Cost Function Value
Boarding Rate, 150 Passenger/hr

 

0

3

Tu

>

C

2

E I Pre-Tlmed

“:3 I Pro-Timed TSP

.8. 0 Real Time w/o TSP
o D D-SPORT

 

 

 

 

 

Different Volume-to-Capaclty Combinations Scenarios

 

 

 

Figure 9.4 Overall traffic network cost function value under different traffic
demand and 150 passenger/hr transit demand

9.3. Network Total Delay and Number of Stops

The first measures of effectiveness discussed in this section are the total delay and
number of stops for transit and non—transit vehicles. Estimates of traffic delays and
number of stops are important to emphasize the ability of a system to function efficiently
under different traffic conditions, especially during peak-hour. As an example, the level
of service (LOS) for a signalized intersection, which is a qualitative measurement of

traffic signal operational performance, is measured based on expected incurred delay.

Table 9.1 summarizes transit delays and number of stops averaged across 10 replications
under various transit and traffic demand levels and identical signal control alternatives for
both the eastbound and northbound routes of the network shown in Figure 8.12, it also

shows pair-wise statistical comparisons with scenarios using the pre-timed controller. On

162

the other hand, Figure 9.5, Figure 9.6, and Figure 9.7 visually compare transit delay when
different signal controllers are used. Note that the selection of the 10 replications has
ensured that the statistical power is at least 80% (i.e., the probability of failing to reject
the assumption that delays and number of stops estimates from any two signal control

scenarios are the same, when they are actually different is 20%).

In most scenarios, the results show a significant reduction in transit delay and number of
stops per vehicle when the D-SPORT controller is implemented compared to the other
tested signal control alternatives. While transit delay reduction varies between 20% and
70%, the reduction in number of transit stops varies between 10% and 85%. This finding
is associated to the fact that D-SPORT interacts with traffic demands, allowing the
D-SPORT controller to adjust in real-time the amount of green time allocated to each
signal phase in response to demand fluctuations in a manner that benefits both transit and
general traffic. The pro-timed signals, on the other hand, do not react to any fluctuation in
traffic demand, with the exception of the provision of early greens and green extensions

to accommodate approaching transit vehicles when the TSP logic is active.

Compared to the real-time controller without TSP, the results also show that the
D-SPORT controller has resulted in a better transit operational performance in most of
the scenarios. These reductions in the amount of delay and number of stops per vehicle
are associated with a minimization of the negative influence of transit vehicles on the
non-transit vehicles by providing the traffic stream with better signal timings and better

traffic progression.

163

 

In some scenarios, the implementation of the D-SPORT controller resulted in higher
delay for one transit route compared to the other scenarios, especially in cases in which
there is a high difference in traffic congestion between the eastbound and northbound
directions of the simulated network. This is because the D-SPORT controller attempts to
improve overall system performance by acting on a set of sub-cost benefits. For instance,
the D-SPORT controller may not find an ideal timing plan to reduce transit delay, but
could develop a plan that significantly improve non-transit traffic or improve schedule
adherence by imposing a short additional delay to the transit vehicles. By considering
network-level traffic, taking into account all the transit routes in the network, and all the
sub-objectives considered, such as service reliability and schedule adherence, it can be
easily noticed that the D-SPORT controller model performs better than the alternative

signal control approaches considered.

As an example, consider the scenario with an eastbound v/c ratio of 95%, and northbound
v/c ratio of 50%. In this case, higher eastbound transit delays are observed compared to
other scenarios. However, the average transit delay obtained by considering the two
transit routes is lower than other scenarios. This can also be observed within the context
of the cost function evaluation, where the controller that produced the lowest overall
network cost is the D-SPORT controller. For instance, the cost function value for the
D-SPORT controller is 5206.04, while cost function values of 6361.31, 6389.70, and
5959.64 are associated with the pro-timed, pre-timed with TSP, and real-time without

TSP, respectively, for the transit demand level of 50 passenger/hr.

164

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Table 9.1 Simulated transit delays and number of stops for different traffic
signal controls
I PreTlmed I Pretlmed TSP I Rea-Tlme wlo TSP D-SPORT
“3:239 (1) (2) (3) (4)
Pass e nger Ih r I Delay it Stops Delay it Stops I Delay # Stops Delay # Stops
s/veh stops/veh s/veh stops/veh s/veh stops/veh s/veh stops/van
' vlc EB 50% - vlc N3 50%
o 'ETTransit 70.56 2.52 71.95 2.55 51.66 * 1.33 * 45.80 ' 1.13 '
“’ NB Transit 18.26 0.89 18.16 0.89 7.81 * 0.26 * 7.20 * 0.25 *
8 EB Transit 22.29 0.88 29.53 ' 1.58 * 31.20 * 1.92 ' 26.70 * 1.70 *
'- NB Transit 8.96 0.36 8.27 ' 0.32 ' 7.80 * 0.26 * 7.63 * 0.30 *
8 EB Transit 66.59 2.34 66.38 2.33 25.00 ' 1.00 * 11.60 ' 0.10 '
'- NB Transit 17.45 0.85 17.38 0.84 7.10 * 0.22 ' 5.44 ' 0.25 '
vlc EB 75% - vlc N8 50%
0 EB Transit 73.62 2.68 73.5 2.54 * 51.94 * 1.41 * 48.50 " 1.29 "
"’ NB Transit 20.99 1.05 21.79 A 1.08 13.27 ' 0.49 1 10.30 ' 0.34 1
8 EB Transit 30.69 1.43 39.48 ' 1.78 ' 36.50 ' 2.00 ' 27.10 ' 1.70 "
'- NB Transit 9.33 0.38 10.83 * 0.52 * 12.40 * 0.42 * 9.35 0.38
3 EB Transit 68.31 2.44 68.53 2.46 26.30 ' 1.00 * 13.40 * 0.40 '
'- NB Transit 19.79 0.98 19.99 1.00 11.40 * 0.36 * 5.97 * 0.30 '
vlc EB 75% - vlc N8 75%
c, EB Transit 72.68 2.64 83.97 * 2.69 53.22 ' 1.45 ' 51.70 * 1.42 *
"’ NB Transit 21.23 1.04 21.57 1.13 ' 13.05 * 0.47 " 11.40* 0.42 *
8 EB Transit 33.51 1.55 39.33 ' 1.80 * 26.70 * 2.08 * 26.40 ' 1.70 '
" NB Transit 9.71 0.38 10.71 * 0.49 ' 12.60 ‘ 0.48 " 6.88 " 0.36
8 EB Transit 67.83 2.43 68.09 2.46 28.80 " 1.21 " 14.00 " 0.50 "
'- NB Transit 20.15 0.98 20.30 0.99 13.10 ' 0.44 ' 6.60 ' 0.36 '
vlc EB 95% - vlc N8 50%
o EB Transit 113.37 2.59 79.80 * 1.79 * 49.93 ' 1.34 ' 53.20 ' 1.38 *
'” NB Transit 16.62 1.03 18.33 * 1.16 * 18.20 * 0.71 ' 16.70 0.66 *
8 EB Transit 79.64 2.17 60.00 ' 1.20 * 33.40 ' 2.00 * 27.60 ' 1.40 *
'- NB Transit 9.84 0.44 11.61 ' 0.54 * 21.30 * 0.99 ' 14.92 ' 0.72 '
3 EB Transit 107.73 2.57 106.76 2.53 24.80 ' 0.79 * 14.50 * 0.10 '
'- NB Transit 16.09 0.99 16.17 0.99 18.4 ' 0.71 * 12.65 ' 0.65 '
vlc EB 95% - vlc NB 75%
0 EB Transit 66.35 2.52 60.44 ‘ 2.47 53.54 ' 1.53 " 49.00 ' 1.54 "'
"’ NB Transit 20.35 1.29 21.09 A 1.36 18.59 * 0.74 * 15.90 ' 0.58 '
8 EB Transit 73.25 2.11 45.36 * 1.19 ' 35.20 * 2.33 * 23.40 ' 1.60 *
'- NB Transit 15.15 0.78 17.01 * 0.92 ' 18.80 ' 0.84 * 13.31 * 0.76
8 EB Transit 64.45 2.46 64.02 2.43 25.20 * 1.04 ' 12.70 * 0.40 '
'- NB Transit 19.62 1.23 1_9.81 1.25 18.50 0.72 * 10.74 ' 0.54 *
vlc EB 95% - vlc NB 95% -
c, EB Transit 50.05 2.48 63.14 ' 3.33 A 61.77 * 1.77 * 59.00 * 1.67 *
"’ NB Transit 27.28 1.76 32.43 * 2.13 ' 18.98 * 0.74 ' 16.90 ' 0.64 '
8 EB Transit 52.19 1.76 59.26 ' 3.05 ‘ 29.20 ' 1.75 30.90 ' 1.70
" NB Transit 28.52 1.92 37.11 " 2.93 ' 19.60 ' 0.75 " 16.05 * 0.70 *
8 EB Transit 49.2 2.37 51.3 2.68 27.40 * 1.17 ' 15.00 ' 0.20 '
'- NB Transit 26.97 1.75 27.82 1.82 17.30 * 0.63 ' 9.72 ' 0.44 '

 

 

 

 

 

 

 

Statistically significant at 99% confidence level compared to case (1 ), pre-timed with no active TSP
A Statistically significant at 95% confidence level compared to case (1 ), pre-timed with no active TSP

165

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

East Bound Route Transit Vehicle Delay Comparison
Boarding Rate, 50 Passenger/hr
120.00 . e 7 7e 7 e7 -7 e7 7
100.00- ‘7777—55577 Wiv7~
S
O
§80.00- 7* 7 i V, eiie7
>1 flPre-Timed
E: 60 00 _ IPre-Tlmed TSP
o ‘ — Cl Real Time w/o TSP
a UD-SPORT
5 40.00 -
.:
20.00 -
0.001
Different Volume-to-Capaclty Combinations Scenarios
North Bound Route Transit Vehicle Delay Comparison
Boarding Plate, 50 Passenger/hr
120.007 7777 7 7 777 777 77 777 7 #77—
10000 -
.5
g 00.00 7
>1 IPre-Timed
2 I Pro-Timed TSP
8 600° :1 Real Time w/o TSP
'3 CID-SPORT
C 40.00 -
E
l-
20.00 7
0.00 <
Different Volume-to-Capaclty Combinations Scenarios

 

 

 

Figure 9.5 Simulated transit delay for different volume-to-capacity combinations
and a boarding rate of 50 passenger/hr

166

 

East Bound Route Transit Vehicle Delay Comparison
Boarding Rate, 100 Passenger/hr

 

 

 

 

 

 

 

 

 

 

 

 

120.007 7 77 7 7 7 7 7 7 7 7 7 i i I

100.0077 7 7 7 7 7 77 777 77,7I77,l
5
5 80.00 77 7 7 7 7 7 7 7 7 -7 77,7 41; .7,
>2 SPre-Timed
% 60 00 ‘ IPre-Timed TSP
o ' Ci Real Time w/o TSP
'3'; [JD-SPORT
g 40.00 7 7
a

20.00 -
0.00 7
Different Volume-to-Capaclty Combinatlons Scenarios
North Bound Route Transit Vehicle Delay Comparison
Boarding Flare, 100 Passenger/hr

120.00 7 7 7 7 77 77 7 .7, 7 If

100.00 7 7 77 777 7 777 7-7, 7 , ,7
5
> 7
1 EPre-fimed
g 7 IPre-Timed TSP
a Cl Real Time w/o TSP
7" CID-SPORT
E 7
l-

 

 

Different Volume-to-Capaclty Combinations Scenarios

 

 

 

Figure 9.6 Simulated transit delay for different volume-to-capacity combinations
and a boarding rate of 100 passenger/hr

167

 

 

 

 

 

 

 

 

 

 

 

 

East Bound Route Transit Vehicle Delay Comparison
Boarding Rate, 150 Passenger/hr
120.00 7 7 77 7 7 7 7
100.001 7 7 7 7 77 77 —
S
0
5 80.00 7 7 7 7 77 7 7 77 .
>1 E Pre-Trmed
2 I Pre-Timed TSP
' ea ime w o
3 5000 DR lT‘ / TSP
g 7 . 1:1 D-SPORT
g 40.00 77 7
I:
20.00 «7
0.00 <
Different Volume-to-Capaclty Combinations Scenarios
North Bound Route Transit Vehicle Delay Comparison
Boarding Rate, 150 Passenger/hr
120.00 77 7 7 7- 7 777777
100.00 7»
'5 80 00
i ' 1 BPre-Timed
% 60 00 _ IPre-Timed TSP
n ' E] Real Time We TSP
’5 CID-SPORT
g 40.00 7 7
E
20.00 7
0.00 7
Different Volume-to-Capaclty Combinations Scenarios

 

 

 

Figure 9.7 Simulated transit delay for different volume-to-capacity combinations
and a boarding rate of 150 passenger/hr

168

 

While performance evaluation results for non-transit traffic are summarized in Table 9.2,
through Table 9.4, Figure 9.8 through Figure 9.13 show the trend of these selected MOEs
for different traffic and transit demand levels. These results show that the D-SPORT
controller is able to significantly reduce traffic delay and number of stops along the two
major arterials compared to the other signal control scenarios considered. While the
reductions in traffic delay under D-SPORT control varies between 20% and 80%,
reduction in number of traffic stops varies between 10% and 80%. The reductions in
traffic delay and number of stops are attributed to the enhanced ability of the D-SPORT
controller to respond to traffic fluctuation in real-time and better tailor the signal timing
to the current traffic and transit needs. It is also observed that the D-SPORT controller
impact on traffic delay and number of stops on the streets crossing the two main corridors

varies between non-noticeable to slightly negative impacts.

In cases in which the pre-timed controller is used, the simulation results show a lower
delay for the busiest direction when the v/c ratio difference for the eastbound and
northbound major directions is wide. For instance, when the eastbound vlc ratio is 0.95,
and the northbound v/c is 0.5, the eastbound transit delay varies between 10717 seconds,
while the northbound transit delay varies between 20-35 seconds. Although these results
seem to be out of order, they can be explained if the optimization analogy in coordinated
network is clarified. As indicated in Section 8.3, the pre-timed signal timings where
obtained by using SYNCHRO 6. In this model, the optimization process generally favors
the direction with the higher volume by providing this direction with a wider green
window, whether the low-volume approach is progressed or not. Therefore, this often

results in poor coordination along the low-volume approach.

169

In the above example, the signal timings produced by SYNCHRO 6 resulted in more
green time being - allocated to the eastbound approach at intersection 4 than the
northbound approach since the former has a higher volume. A problem then occurs when
the northbound traffic at intersection 5 along the northbound corridor is being allocated
more green time than the northbound traffic at intersection 4, because the eastbound
volume is higher than the northbound volume. When the northbound green time at
intersection 5 is fully utilized, the signal at intersection 4 will then be unable to process

all the vehicles arriving at the intersection within a cycle, causing additional delay.

The D-SPORT controller shows a superior performance compared to the other controllers
through its ability to react to traffic demand fluctuation between the two major
approaches, which can be translated into a reduction of both transit and general traffic
delay and number of stops. In this case, the simulation results show that the approach
with higher traffic volume is associated with higher transit delay, which is a logical
result. This is because of the better consideration of traffic progression throughout the
network and the ability to reduce the disturbance caused by transit vehicles to the traffic
stream, which is achieved by the near-optimal green times allocation throughout the

network that takes into account both transit and non-transit vehicles movements.

Because of the system-based optimization approach, the D-SPORT controller does not
necessarily penalizes traffic on the crossing streets when providing priority treatments to
transit vehicles along the two main corridors. The controller attempts instead to
synchronize the green tinting with the transit expected arrival at bus stops. If green times
synchronizing is not possible, more un-utilized green times could be observed, which

could potentially lead to unnecessary traffic delay and number of stops.

170

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Table 9.2 Simulated traffic delay and number of stops for boarding rate of 50
passenger/hr
Pre- Timed Pre- Timed TSP Rea-Tlme wlo TSP D-SPORT
(1) (2) (3) (4)
Delay if Stops Delay # Stops Delay # Stops Delay # Stops
s/veh stop/veh s/veh stop/veh s/veh stop/veh s/veh stop/veh

EBv/c 50% - NBylc 50%

EB Traffic 18.26 0.89 18.16 0.89 7.81 * 0.26 * 7.20 * 0.25 *

NB Traffic 14.25 0.7 0 14.82 ' 0.71 5.32 "' 0.16 " 5.00 * 0.16 *

Crossing EB 9.77 0.54 9.91 0.54 11.98 * 0.55 A 11.33 * 0.52

Crossing NB 9.74 0.54 9.77 0.54 11.58 * 0.55 11.03 " 0.53
Eavlc 75% ‘ "By/c 50%

EB Traffic 20.99 1.05 21.79 A 1.08 13.27 * 0.49 * 10.30 * 0.34 *

NB Traffic 17.68 0.79 16.9 * 0.76 * 5.42 * 0.16 * 5.00 * 0.15 *

Crossing EB 12.32 0.6 12.3 0.6 12.23 0.56 * 11.66 * 0.54 *

Crossing NB 10.08 0.54 10.35 0.55 1 1.53 " 0.55 11.1 ' 0.54
Ech 75% - NBv/c 75%

EB Traffic 21.23 1.04 21.57 1.13 * 13.05 " 0.47 * 11.4 * 0.42 "

NB Traffic 17.92 0.84 29.71 " 1.42 * 9.32 * 0.29 * 8.00 * 0.23 *

Crossing EB 12.31 0.60 12.42 0.60 12.09 0.56 * 11.51 " 0.54 *

Crossing NB 12.1 0 0.60 12.17 0.60 11.60 * 0.55 " 11.10 ' 0.53 *
EBv/c 95% - NBy/c 50%

EB Traffic 16.62 1.03 18.33 * 1.16 * 18.20 * 0.71 * 16.70 0.66 *

NB Traffic 35.61 1.04 39.46 * 1.20 * 5.47 * 0.16 * 5.00 * 0.14 *

Crossing EB 25.31 0.82 25.35 0.82 12.28 " 0.56 * 11.73 * 0.54 *

Crossing NB 12.78 0.54 14.09 * 0.58 * 11.56 * 0.55 11.14 * 0.53
EBy/c 95% - Nch 75%

EB Traffic 20.35 1.29 21.09 A 1.36 18.59 * 0.74 " 15.9 * 0.58 *

NB Traffic 26.20 1.05 27.61 * 1.16 " 8.94 * 0.27 * 8.10 * 0.24 *

Crossing EB 22.60 0.78 22.69 0.79 12.32 " 0.57 * 11.71 * 0.55 "

Crossing NB 17.70 0.68 18.17 0.69 11.77 * 0.56 * 11.16 * 0.54 *
Eka; 95% - NBv/c 95%

EB Traffic 27.28 1.76 32.43 * 2.13 * 18.98 * 0.74 " 16.9 * 0.64 *

NB Traffic 28.01 1.95 43.9 * 3.51 " 13.9 * 0.48 * 12.7 * 0.41 *

Crossing EB 22.69 0.79 22.70 0.79 12.14 * 0.56 * 11.75 * 0.55 "

Crossing NB 22.39 0.78 22.79 0.79 11.73 * 0.56 * 11.1 * 0.54 *

 

 

 

 

 

 

" Statistically significant at 99% confidence level compared to case (1 ), pre-timed with no active TSP
A Statistically significant at 95% confidence level compared to case (1 ), pro-timed with no active TSP

171

 

 

 

   

 

 

 

 

   

 

 

 

 

   

 

 

Non-Transit Traffic Average Delay Non-Transit Traffic Averageo Delay
EBv /c. 50% - NBv/c: 50° .6 EBv /c. 75% - NBv/c
50 i 7* W . 50 7 7- 7
5 5
> 7 7 7 7 7 7 7 7 7 7 7 7 40
n 4° 5
5 30 7 7 7 7 7 7 E 30 7
0
o 8
20 7 W i 7 i i i 7 7 7 20
13 3
E 10 E 10
9 >"
0 0
EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
I Pro-Timed I Pro-Timed TSP I I Pro-Timed I Pre- Timed TSP
Cl RT wlo TSP D D-SPOR'T 1:! RT w/o TSP Cl D-S SPORT
Non-Transit Traffic Average Delay Non-Transit Traffic Average Delay
50 EBv/c: 75% - NBv/c: 75 5% 50 EBv/c .' 95% - NBv/c : 50%
.5 Z
5 4° §4°
5' so a“ so
0
o 3
20 20
§ 3
E 10 E 10
B s
0 0
EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
I Pro-Timed I Pro-Timed TSP I Pre-Tirned I Pre-Tlrned TSP
Ci RT wlo TSP D D-SPORT El RT wlo TSP D D-SPORT
Non-Transit Traffic Average Delay Non-Transit Traffic Average Delay
E /c : 95% - NBv/c : 75% 50 . EBv/c_: 95% - NBv/_c:
5 s
40
5 5
53°
0 a 20
§ 3
E E 10
g >‘
0
EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
I Pre7Tlmed I Pro-Timed TSP I Pro-Timed I Pro-Timed TSP
III RT w/o TSP D D-SPOFlT D RT wlo TSP D D-SPORT

 

 

 

 

 

Figure 9.8 Simulated non-transit traffic delay for different v/c combinations and
a boarding rate of 50 passenger/hr

172

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Non-Transit Traffic Average Number of Stops Non-Transit Traffic Average Number of Stops
EBv v/:c 50% - NBv/c EBv/c: 75% - NBv/c : 50%
:4 :4’ ”’ " ’ 77 _,
° 2
i 3 a 3 7 7 7 7 7 7 7
n’ n‘
a. D.
2 8
:02 r02 77 7777777 7 7 77
.- ‘-
o o
3 1 g
E
3 3
z o 2
EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
I Pro-Timed I Pro-Timed TSP I Pro-Timed I Pro-Timed TSP
El RT wlo TSP Ci D-SPORT Ci RT wlo TSP D D-SPORT
Non-Transit Traffic Average Number of Slope Non-Transit Trafflc Average Number of Stops
EBv/c 7:596 - NBv/c. EBv/c: 9596 - NBv/c
.: 4 T _ z 4
O I
>
5 3 7 7 7 7 7 7 77 7 s 3
a“ u‘
o. o.
.9 8
m 2 7 7 7 7 7 7 7 o) 2 7
'6 ‘8
.. a
5‘ i‘
3 3
z o z 0
EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
I Pro-Timed I Pro-Timed TSP I Pro-Timed I Pro-Timed TSP
Cl RT w/o TSP D D-SPORT Cl RT wlo TSP D DSPOR’T
Non-Transit Traffic Average Number of Steps Non-Transit Traffic Average Nurrber of Stops
EBv/c: 95% - NBv/c: EBv/c: 95% - NBv/c: 95%
.r: 4 ' 'T' .I: 4 '
£3 , _ 7 _ £3
a“ o”
3 5
m 2 77 7 77 7 7 G 2
'6 .0
:1 77 _. 7-77"“ 21
E g I g E
:f 3
z . I z 0
EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
I Pro—Timed I Pro-Timed TSP ’ I Pro-Timed I Pro-Timed TSP
[1 RT w/o TSP CI D-SPORT D RT wlo TSP El D-SPORT

 

 

 

 

 

 

Simulated non-transit number of stops for different v/c combinations
and a boarding rate of 50 passenger/hr

Figure 9.9

173

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Table 9.3 Simulated traffic delay and number of stops for boarding rate of 100
passenger/hr
Pre- Timed Pro- Timed TSP Rea-Time wlo TSP D-SPORT
(1) (2) (3) (4)
Delay If Stops Delay # Stops Delay if Stops Delay ff Stops
s/veh stop/veh s/veh sfop/veh s/veh stop/veh s/veh stop/veh
EBv/c 50% - Nch 50%
EB Traffic 8.96 0.36 8.27 * 0.32 * 7.80 * 0.26 * 7.63 * 0.30 *
NB Traffic 4.07 0.16 5.27 " 0.22 * 4.90 "’ 0.16 4.97 * 0.16
Crossing EB 9.84 0.54 10.09 0.55 11.94 * 0.55 10.84 * 0.55
Crossing NB 9.70 0.54 9.90 0.55 11.86 ‘ 0.56 * 11.01 ' 0.54
EBv/c 75% - NBylc 50%
EB Traffic 9.33 0.38 10.83 ' 0.52 * 12.40 * 0.42 * 9.35 0.38
NB Traffic 6.65 0.23 8.32 " 0.35 * 5.50 * 0.17 ' 5.34 * 0.19 *
Crossing EB 12.34 0.60 12.59 0.61 12.36 0.58 * 11.79 * 0.55 '
Crossing NB 10.06 0.54 10.81 " 0.55 A 11.43 * 0.54 10.77 * 0.53
Bey/c 75% - new 75%
EB Traffic 9.71 0.38 10.71 * 0.49 * 12.6 * 0.48 "’ 6.88 " 0.36
NB Traffic 6.05 0.18 9.21 * 0.39 * 11.00 * 0.37 * 10.01 " 0.33 *
Crossing EB 12.32 0.60 12.67 A 0.61 12.2 0.57 " 11.43 * 0.55 "
Crossing NB 12.08 0.60 12.51 * 0.61 11.59 * 0.55 * 10.85 * 0.55 *
EBv/c 95% ' "By/c 50%
EB Traffic 9.84 0.44 11.61 * 0.54 * 21.3 * 0.99 ‘ 14.92 * 0.72 *
NB Traffic 20.04 0.46 27.61 * 0.69 * 5.10 * 0.14 * 5.52 * 0.20 *
Crossing EB 25.35 0.82 25.78 0.82 12.01 * 0.56 ‘ 10.53 " 0.54 *
Crossing NB 12.73 0.55 15.42 * 0.6 * 11.67 "’ 0.55 11.04 "' 0.55
53m; 95% - NBv/c 75%
EB Traffic 15.15 0.78 17.01 " 0.92 * 18.8 * 0.84 ‘ 13.31 * 0.76
NB Traffic 16.88 0.57 18.80 * 0.74 * 8.7 * 0.27 * 10.24 ‘ 0.36 *
Crossing EB 22.67 0.78 28.60 1.18 12.44 * 0.58 * 10.46 * 0.56 *
Crossing NB 17.70 0.68 24.58 1.11 11.59 " 0.55 * 10.85 * 0.54 "
Eka: 95% " NBv/c 95%
EB Traffic 28.52 1.92 37.11 * 2.93 * 19.6 * 0.75 * 16.05 * 0.70 *
NB Traffic 20.61 1.30 39.38 " 3.1 * 12.2 * 0.39 * 15.27 * 0.62 *
Crossing EB 22.74 0.79 35.71 1.69 11.54 * 0.54 * 10.45 * 0.53 *
Crossing NB 22.44 0.78 35.88 1.68 11.5 * 0.54 * 10.82 * 0.54 *

 

 

 

 

 

 

* Statistically significant at 99% confidence level compared to case (1 ), pre-timed with no active TSP
A Statistically significant at 95% confidence level compared to case (1 ), pro-timed with no active TSP

174

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Non-Transit Traffic Average Delay Non-Transit Traffic Average Delay
EVE /c. 50% - NBv/c: 50% EBv/ /:c 75% - NBV/ /c:
50 50
5 5
5 4o 7 7 77 7 7 7 a 40
5 30 7 7 7 5 30
a“ 8
20 7 7 7 7 7 7 7 7 7 7 20
1! §
E 10 7 E 10 7
g g
0 0 4
EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
I Pre-Tlmed I Pro-Timed TSP I Pro-Timed I Pre-Tlmed TSP
[3 RT wlo TSP Ci D—SPORT CI RT wlo TSP El D—SPORT
Non-Transit Traffic Average SDeIay Non-Transit Traffic Average Delay
EBv/c: 75% - NBv/c. EBv/c: 95% - __NBV Bv/c: 50%
50 50 7
.C .c
€40 54°
5" 30 3‘30
8 20 20
a 5
E 10 5 10
5’ §
0 0
EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
I Pro-Timed I Pro-Timed TSP I Pro-Timed I Pro-Timed TSP
D RT w/o TSP D D-SPORT Ci RT wlo TSP U D—SPORT
Non-Transit Traffic Average Delay Non-Transit Traffic Average Delay
EBv/c : 95% - NBv/c : 7593a EBv/c .' 95% - NBv/c : 95}:
i E
3 a
.8. 5
E E
g ;
EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
5 Pro-Timed a Pro-Timed Tsrj a Pro-Timed I Pro-Timed TSP
Cl RT wlo TSP D D—SPORT D FiT wlo TSP U D-SPORT

 

 

 

Figure 9.10 Simulated non-transit traffic delay for different v/c combinations and
a boarding rate of 100 passenger/hr

175

 

 

 

Non-Transit Traffic Average Number of Stops
EBv/c.

50%-NEW

 

 

b

Non-Transit Traffic Average Number of Stops
EBv/c. 50%

'75% - NBv/c.

 

 

 

 

 

 

 

 

g s
i 3 '5 3
e' u'
D. O.
2 2
in 2 m 2
‘6 '6
u ..
E‘ E‘
:l 3
z 0 z 0
EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
I Pro-Timed I Pre-T'Iomed TSP I Pro-Timed I Pro-Timed TSP
Ci RT wlo TSP 0 RT wlo TSP D D-SPORT
Non-Transit Traffic Average Number of Steps Non-Transit Traffic Average Number of Stops
EBv /c: 75% - NBv/c: 75% EBv/c .9596 - NBv/c
s ’ = ‘ ’ "
5 £3
6 of
n. a.
2 8
tn w 2
‘6 ‘6
h h
i ‘3‘
:I 3
z o z 0
EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
I Pro-Timed I Pie-Timed TSP H Pro—Timed I Pro-Timed TSP
D RT wlo TSP Ci D-SPORT [:1 RT wlo TSP D D-SPORT

 

 

Non7Translt Traffic Average Number of Stops
EBv/c :95% - NBv/c. 75

Number of Stops, if/veh

NB Traffic

Crossing EB Crossing NB
Traffic Movement Direction

 

I Pro-Timed I Pie-Timed TSP
[3 RT wlo TSP D D~SPORT

 

 

 

 

Number of Stops, erh

0

_a

:95% - NBv/c:

Non-Transit Traffic Average Number of Stops
rchBv

 

 

EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction
I Pro-Timed I Pro-Timed TSP
C! RT wlo TSP Cl D—SPORT

 

Figure 9.11

Simulated non-transit number of stops for different We combinations

and a boarding rate of 100 passenger/hr

176

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Table 9.4 Simulated traffic delay and number of stops for boarding rate of 150
passenger/hr
Pre- Timed Pre- Tlmed TSP Rea-Time wlo TSP D-SPORT
(1) (2) (3) (4)
Delay if Stops Delay if Stops Delay if Stops Delay # Stops
s/veh stag/veh s/veh stop/veh s/veh stop/veh s/veh stgp/veh
EBv/c 50°/o NBy/c 50%
EB Traffic 17.45 0.85 17.38 0.84 7.10 " 0.22 " 5.44 * 0.25 "
NB Traffic 13.21 0.65 13.30 0.65 5.40 * 0.17 * 4.86 * 0.16 *
Crossing EB 9.81 0.54 9.82 0.54 12.3 * 0.56 * 11.37 ' 0.54
Crossing NB 9.74 0.54 9.74 0.54 11.36 * 0.53 A 11.26 * 0.55
EBv/c 75% NBvlc 50%
EB Traffic 19.79 0.98 19.99 1.00 11.4 " 0.36 * 5.97 " 0.30 "
NB Traffic 16.50 0.73 16.70 0.75 5.80 * 0.17 * 5.16 * 0.17 *
Crossing EB 12.31 0.60 12.33 0.60 11.80 * 0.54 * 12.08 0.57 *
Crossing NB 10.08 0.54 10.13 0.54 11.43 " 0.54 11.10 * 0.56 *
EBvlc 75% NBv/c 75%
EB Traffic 20.15 0.98 20.3 0.99 13.1 " 0.44 ' 6.60 * 0.36 ‘
NB Traffic 16.84 0.78 17.11 0.79 10.3 * 0.34 * 10.00 " 0.32 *
Crossing EB 12.31 0.60 12.32 0.60 11.79 * 0.55 * 9.96 * 0.52 *
Crossing NB 12.12 0.60 12.14 0.60 11.69 * 0.55 "‘ 11.12 * 0.55 *
EBv/c 95% "By/c 50%»
EB Traffic 16.09 0.99 16.17 0.99 18.4 * 0.71 * 12.65 * 0.65 "
NB Traffic 33.92 0.98 33.91 0.98 5.8 * 0.16 * 6.23 * 0.19 *
Crossing EB 25.32 0.82 25.34 0.82 12.17 * 0.56 "' 10.58 * 0.54 '
Crossing NB 12.77 0.54 13.05 0.55 11.8 * 0.57 * 10.45 * 0.54
EBv/c 95% "By/c 75%
EB Traffic 19.62 1.23 19.81 1.25 18.5 0.72 * 10.74 " 0.54 *
NB Traffic 25.15 1.00 25.59 1.04 10.00 * 0.33 * 8.33 * 0.28 *
Crossing EB 22.6 0.78 22.62 0.78 11.94 * 0.56 * 10.18 * 0.52 *
Crossing NB 17.69 0.68 17.84 0.69 11.62 ‘ 0.56 * 10.92 * 0.54 ‘
Eka; 95% "By/c 95%
EB Traffic 26.97 1.75 27.82 1.82 17.30 * 0.63 * 9.72 * 0.44 *
NB Traffic 27.14 1.87 30.16 2.18 15.00 * 0.53 * 15.21 " 0.57 *
Crossing EB 22.71 0.79 22.72 0.79 11.95 * 0.54 * 10.95 * 0.54 *
Crossing NB 22.43 0.78 22.49 0.78 11.76 * 0.55 * 10.32 " 0.52 *

 

 

 

 

 

 

A

Statistically significant at 99% confidence level compared to case (1 ), pre-timed with no active TSP
Statistically significant at 95% confidence level compared to case (1 ), pro-timed with no active TSP

177

 

 

Non-Transit Traffic Average Delay
E /c : 50% - :50%

NBv/c ,

 

 

(If
C

Non-Transit Traffic Average Delay
EBv/c.‘ 75% - NBv/c : 50%

 

 

5 5
s 5 4°
: a 30
a 3 20
5 =8
.r: .r:
g g 10
0
EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
Pro-Timed I Pro-Timed TSP E Pro-Timed I Pro-Timed TSP
C] RT w/o TSP D D-SPORT El RT w/o TSP D D—SPORT

 

 

Non-Transit Traffic Average5 Delay
EBv/c: 75% - ’NBV

Vehllce Delay. slveh

 

 

Vehllce Delay. alveh

Non-Transit Traffic Average Delay
EBv/c: 95% - NBv/c:

 

 

EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
I Pro-Timed I Pro-Timed TTSP I Pro-Timed I Pro-Timed TSP
E1 RT wlo TSP D D—SPO [1 RT wlo TSP D D-SPORT

 

 

Non-Transit Traffic Average Delay
EBv/c:95% - ,,~BV/°;§’L 7 A

 

 

 

 

 

50
1: .:
e a
S $40
5' 53°
8 820
.3. 13
E E10
g 3

O

Non-Transit Traffic Average Delay
EBv/c : 95% - NBv/c 5 95%

 

EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
I Pro-Timed I Pro-Timed TSP I Pre<Timed I Pre~TImed TSP
Ci RT wlo TSP D D—SPORT El RT wlo TSP E1 D—SPORT

 

 

Figure 9.12 Simulated non-transit traffic delay for different We combinations and
a boarding rate of 150 passenger/hr

178

 

Non-Transit Traffic Average Number of Stops
EB We. 50% - NBv/c:

b

EB Traffic NB Traffic Crossing EB Crossing NB

Traffic Movement Dlmction

 

 

&

.C .C
3 2
:3 :3
a’ a’
2 3
:02 «>2
‘6 '6
a ..
31 £1
E

3 3
20 z0

Non-Transit Traffic Average Number of Stops
Bv/c: 75% - NBv/c: 50%

 

EB Traffic NB Traffic Crossing EB

Traffic Movement Direction

Crossing NB

 

I Pro-Timed I Pro-Timed TSP ‘ I Pro-Timed I Pro-Timed TSP
[3 RT wlo TSP El D—SPORT El RT w/o TSP El D-SPORT

 

 

Non-Transit Traffic Average Number of Steps
EBv /c. 75% - NBv/c

EB Traffic NB Traffic Crossing EB

Traffic Movement Direction

Crossing NB

 

 

r. =4
>

. as
s s
6 52
'6 '6
h h
f E‘
3 5
z z

Non-Transit Traffic Average Number of Stops
v/:c 95 % - NBv/c :5096

 

EB Traffic NB Traffic Crossing EB Crossing NB

Traffic Movement Direction

 

I Pro-Timed I Pro-Timed TSP I Pro-Timed I Pro-Timed TSP
E1 RT wlo TSP D DSPORT E] RT wlo TSP C] D-SPORT

 

 

 

Non-Transit Traffic Average Number of Steps
EBv/c. 95% '. NBv/c: 75L 7

 

 

 

 

.5 2
z E
‘3 "a
a a“
2 2
0:2 :02
"6 3
h h
31 £1
E

3 3
Z Z

Non-Transit Traffic Average Number of Stops
7 EBv/c. 95% - NBv/ /:c

 

 

EB Traffic NB Traffic Crossing EB Crossing NB EB Traffic NB Traffic Crossing EB Crossing NB
Traffic Movement Direction Traffic Movement Direction
Pro-Timed I Pre-Timed TSP I Pro-Timed I Pro-Timed TSP
[3 RT w/o TSP El D-SPORT CI RT wlo TSP El D-SPORT

 

 

 

Figure 9.13 Simulated non-transit number of stops for different vlc combinations
and a boarding rate of 150 passenger/hr

179

 

9.4. Schedule Adherence and Transit Service Reliability

Improving the schedule adherence and service reliability is part of the D-SPORT
controller cost function. In order to quantify the service reliability for transit service
users, transit lateness relative to the scheduled arrival for each transit vehicle is recorded
in each simulation experiment. As defined in Section 6.3.3, the lateness is computed for
each transit vehicle as the difference between the actual and scheduled arrival times at
each bus stop, averaged across all the simulated buses in the 2-hr simulation replications.
For a given bus stop, the scheduled arrival is calculated as the transit headway plus the
expected travel time from the moment a transit vehicle enters the network until it reaches
the bus stop while traveling at a desired speed of 35 mph (55 km/hr) without any
impedance. For each scenario, the lateness is then averaged across all the transit vehicles
that have dwelled at a bus stop over the 2-hr simulation duration. The standard deviation
for lateness at given bus stop is finally computed based on all the recorded transit lateness

over all the replications associated to a given scenario.

Table 9.5 summarizes the average and standard deviation of transit lateness for each
simulated bus stop in each of the scenarios considered. The results show that the
D-SPORT controller generally produced the lowest transit lateness compared to the other
controllers, with average lateness varying between 30 to 80 seconds. Compared to the
other controllers, lateness was reduced by 5% to 55% in most of the cases. Furthermore,
variations in lateness from implementing the D-SPORT controller are lower by 50%-
90%, with lateness standard deviations varying between 4.2 and 8.8 seconds. In this case,
a low value of standard deviation means a better and more reliable service for passengers

because of the lower probability of observing late transit arrivals at bus stops.

180

The results further show no statistical difference at a 95% confidence level between the
pre-timed scenarios with or without active TSP. In this case, there is no statistical
evidence to suggest that implementing TSP plans within pre-timed controllers would
yield any significant improvement on schedule adherence. This finding is associated with
the fact that active TSP is implemented locally while the MOEs are measured at the
corridor and the network level. Active TSP does not take into account the projected
arrival of a transit vehicle at the downstream intersection. Because of this, priority
treatments granted at one intersection may result in more delay before it reaches the

downstream bus stop because such an action could disturb traffic progression pattern.

Furthermore, the simulation results show that application of the D-SPORT controller
reduced the average lateness of transit vehicles at bus stops when compared to real-time
controller without TSP. This is due to the fact that the disabling the TSP sub-cost
function removes the ability of the controller to keep track of the trajectory of transit
vehicles. Although the real-time controller without TSP actions controller has resulted in
comparable transit delays from the D-SPORT controller in some experiments, the
D-SPORT controller provides a better transit service reliability for these scenarios. In
terms of service reliability, the results of Table 9.5 also show a relatively lower lateness
standard deviation resulted from implementing the D-SPORT controller, which means a

more reliable schedule to the transit riders and the service users.

181

 

 

 

 

 

 

 

Table 9.5 Transit lateness at bus stops for different traffic demand and a
boarding rate of 50 passenger/hr
Average Transit Lateness, Seconds
S ce n a fl 0 Pre- Timed Pre- 11212,)“ TSP Flea-Time wlo TSP D-SPORT
EBv/c _ NBv/c Lategess 5;: Lategess Deg: Lategess DSetg Lategess [is
S S S S

50 - 50 58.09 3.96 57.41 2.35 58.09 3.96 28.25 * 5.82

,_ 75 - 50 66.80 11.92 66.68 11.81 66.8 11.92 32.67 * 5.88
3 75 - 75 66.73 12.06 66.85 12.25 66.73 12.06 30.96 * 5.57
g 95 - 50 70.94 13.92 71.12 14.25 70.94 13.92 57.79 * 5.55
“3 95 - 75 88.83 9.42 88.88 9.47 88.83 9.42 57.04 * 6.04
95 - 95 86.15 6.39 86.34 5.52 86.14 6.41 54.00 * 4.20

50 - 50 77.86 3.58 77.73 1.58 77.86 3.58 51.71 * 6.55

N 75 - 50 90.93 16.65 91.30 17.03 90.93 16.65 55.38 * 6.15
3 75 - 75 90.93 16.71 91.58 17.06 90.93 16.71 51.13 * 7.32
g 95 - 50 84.80 13.74 90.24 * 16.02 84.80 13.74 80.17 * 7.05
m 95 - 75 97.63 12.89 100.25 16.18 97.63 12.89 79.88 ‘ 8.81
95 - 95 98.44 16.15 113.00 * 21.98 98.65 16.50 76.79 * 5.57

50 - 50 93.48 23.98 94.75 23.13 93.48 23.98 55.67 " 7.88

n 75 - 50 91.57 16.56 88.71 16.67 91.57 16.56 54.17 * 7.62
3 75 - 75 91.32 17.61 92.68 17.67 91.32 17.61 59.46 * 7.38
g 95 - 50 112.70 20.25 88.27 * 22.30 112.70 20.25 57.58 " 6.46
m 95 - 75 98.47 20.06 87.99 * 21.54 98.47 20.06 60.88 ' 8.19
95 - 95 101.86 16.70 109.16 A 27.39 101.08 16.35 80.25 * 6.76

50 - 50 71.58 16.62 72.68 16.07 71.58 16.62 31.96 ' 5.97

Q 75 - 50 66.60 12.51 64.05 12.09 66.60 12.51 31.46 ' 5.33
fa} 75 - 75 67.22 12.20 67.97 12.29 67.22 12.20 35.25 * 5.66
g 95 - 50 72.45 16.58 64.90 * 12.01 72.45 16.58 31.83 * 6.59
‘0 95 ~ 75 70.24 3.76 67.67 " 3.94 70.24 3.76 36.50 * 5.35
95 - 95 74.12 16.94 75.03 17.79 73.58 16.08 59.71 * 5.48

 

 

 

 

 

 

 

A

Statistically significant at 99% confidence level compared to case (1 ), pre-timed with no active TSP
Statistically significant at 95% confidence level compared to case (1 ), pro-timed with no active TSP

182

 

Figure 9.14 through Figure 9.17 summarizes the transit lateness fluctuation over time at
each bus stop for different traffic demand levels and a transit demand of 50 passenger/hr.
The results show that D-SPORT has performed better than the other controllers in most
of the experiments. Most of the transit vehicles observed during the 2-hr simulation
period have lower lateness compared to other experiments, with lateness standard
deviation not more than 8.8 seconds. For the other controllers, the figures show a cyclic
pattern in transit lateness over time. This cyclic pattern is due to the fact that transit
lateness values are auto-correlated and not totally independent from each other. In this
case, an early transit vehicle arrival at a bus stop may influence the arrival of other transit
vehicles. For instance, fewer passengers might board the first vehicle as a result of its
early arrival. This may then leave more passengers to board the next arriving vehicle,
causing it to spend more time dwelling and to potentially being late arriving at its next
stop. As can be seen in the figures, the D-SPORT controller was able to reduce the transit
lateness auto-correlation by reducing the transit lateness variance, regulating the transit

service, and improving the transit schedule adherence.

183

 

U)

Lateness

'1PEETTrriéEF ”—
+ Real Time w/o TSP

1 4O
1 2O
1 00

80

40
20

1

Transit Lateness Profile at Bus Stop 1
EB V/C 50°/o - NB V/C 50°/o

 

3 5 7 911131517192123
Transit Vehicle ID

 

-8- D-SPORT

+ Pre-Timed TSP— T

Transit Lateness Profile at Bus Stop 1
EB VIC 75% - NB VIC 50%

 

    

 

 

 

140- a s; ———“—— ——— —
.,, 1201~ — -— —— -~— — ————————
g1oo~ , -— , ——— —— _ ———————
2 so- —
601

40‘ ”I

20 ‘1 1 a 1 1 1”. 1.1113 me". ‘1 ma

13 5 7 911131517192123
Transit Vehicle ID
:PTé-Timed + Pre-Trmed TSP T T
5;— Real Time wlo TSP 3 D-SPORT

 

 

Lateness

 

Transit Lateness Profile at Bus Stop 1
EB VIC 75% - NB VIC 75%

 

 

 

 

 

 

MfIIITFII‘IIIITIIIIIIllllli

1

3 5 7 911131517192123
Transit Vehicle lD

 

 

__ + Pre—Timea
4.— Real Time w/o TSP

431- D-SPORT

+ Pre-Timed TSP ‘ "T

Transit Lateness Profile at Bus Stop 1
EB VIC 95% ' NB VIC 5070

 

 

 

 

 

 

 

 

13 5 7 911131517192123
Transit Vehicle ID
”—4. Pre—Tmed + Pre-T'med TSP

-x— Real Time wlo TSP ~9- D-SPOFlT

 

 

 

Lateness, 5

Transit Lateness Profile at Bus Stop 1
EB vlc 95% - NB v/c 75%

 

 

 

 

 

 

140 -

120 —

100 7

20 I I I I I I I I I I I I I I r Ti I I I I I l
13 5 7 911131517192123
Transit Vehicle 10
:15re-Timedmnu“ :— Pre-Timed TSP __
—><~ Real Time wlo TSP 9- D-SPORT

 

 

Transit Lateness Profile at Bus Stop 1
EB vlc 95% - NB vlc 95%

 

 
 
  

 

 

 

 

 

IJ
firrrrrrrrrrTITIrrrrrrrfi

13 5 .7 911131517192123
Transit Vehicle ID

 

 

; Pre-Timed TSP ‘ __2

-><- Real Time w/o TSP 4+ D—SPORT

 

Figure 9.14 Transit lateness at bus stop 1 for scenarios and 50 passenger/hr
boarding rate

184

 

 

 

140
120
100
80
60
40
20
13 5 7 911131517192123
Transit Vehicle ID
+ Pre-Timed' + Pre-Timed TSP
as Real Time w/o TSP 413- D-SPORT

Lateness, 9

Transit Lateness Profile at Bus Stop 1
EB VIC 50% - NB v/c 50°/o

 

 

Lateness. s

-J'I-"Pre-Trmed

Transit Lateness Profile at Bus Stop 1
EB V/C 75Vo - NB V/C 50°/o

 

_. _J

-J ' L ‘ "
IIIIIITIITIIIIIIITIIIIIW

13 5 7 911131517192123
Transit Vehicle ID

 

 

 

4— Pre-Timed TSP

+:— Real Time w/o TSP ~9- D-SPORT

 

 

Lateness, s

 

Transit Lateness Profile at Bus Stop 1
EB VIC 75% - NB VIC 75%

 

 

 

 

 

20[VIII-IFIITITIITTITIIIITT—l
1 3 5 7 9 11 13151719 2123
Transit Vehicle ID
+“PrTJTime’dTT I"_"‘—1_P’re~TirnTedTTS‘P”—""T
+1.— Real Time w/o TSP 43- D-SPORT

_ I; “P £11636

Transit Lateness Profile at Bus Stop 1
EB VIC 95% - NB v/c 50°/o

 

 

 

 

 

 

 

IIIIIIIIIIIITIITIIIIIIII

13 5 7 911131517192123
TransitVehiclelD

 

+ Pre-Timed TSP

+1- Real Time w/o TSP 43— D-SPORT

 

 

 

Lateness. s

":‘Pié-Tirfiéd
+£— Real Time w/o TSP

140
12o ~~

 

Transit Lateness Profile at Bus Stop 1
EB vlc 95% - NB We 75%

 

I.
IIITIIIIIIIIIIIIIIIIIIII

 

2O

13 5 7 911131517192123
Transit Vehicle 10

””T‘I'Pr‘érimed TSP
4+ D-SPORT

 

 

Lateness. 3

Transit Lateness Profile at Bus Stop 1
EB VIC 95% - NB vlc 95%

 

140 1 — -»
120

 

 

 

 

 

20

IFIIIIIIIIIIIITTIIIRIIIj

13 5 .7 911131517192123
Transit Vehicle ID

 

‘iP'réifiFed
-x- Real Time wlo TSP

+ PPS-Timed TSP—T T
4+ D-SPORT

 

Figure 9.14 Transit lateness at bus stop 1 for scenarios and 50 passenger/hr
boarding rate

184

 

 

Transit Lateness Profile at Bus Stop 2 Transit Lateness Profile at Bus Stop 2

 

 

 

 

 

 

 

 

EB v/c 50°/o - NB VIC 50% EB We 75% - NB v/c 50°/o
m 120 *3
m’
(I)
m
C
(D
‘6
_r __.... __-- ____._.

20 ‘FI—TTTTITI’TIIIITIIIIITITTII 201IIIfTT—TIIITIIIIFTTTITT—‘I
1357911131517192123 1357911131517192123
Transit Vehicle ID Transit Vehicle ID
T+ PrE-TirTer TTT "TTT—1TPreiTTrr‘nZdTTSPTTT 'Tfinhfi— TTTTEEJTTBQT’STPT T
+3 Real Time w/o TSP +3- D-SPORT +1— Real Time w/o TSP 43— D-SPORT
Transit Lateness Profile at Bus Stop 2 Transit Lateness Profile at Bus Stop 2
EB VIC 75% - NB vlc 75% EB VIC 95% - NB VIC 50%

 

 

 

 

 

 

 

   

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

ZOTIIITFTIIITIIIIIIIIIIIITTI
1357911131517192123 1357911131517192123
Transit Vehicle ID Transit Vehicle ID
r T;- Pre-Timed T + Pre-Tirned TSPTT' + PreTTned + Pre—Timed TSP TTTT
—><— Real Time w/o TSP «a- D-SPORT 4(- Real Time wlo TSP «e— D-SPORT
Transit Lateness Profile at Bus Stop 2 Transit Lateness Profile at Bus Stop 2
EB VIC 95% - NB VIC 75% EB VIC 95% - NB VIC 95%
140 7
m 120
g“ 100 ~~ -~
9 .0. =
+- 60 «J
3 4o «1 - -———-— —
20 IIIIIIIIIIIIIIIIIIIIITI1
1357911131517192123 1357911131517192123
Transit Vehicle ID Transit Vehicle ID
:Préified TT mafia T +Pre-Trmed TTaTPTreT-TTirfiédTSPT T
+£— Real Time w/o TSP «P- D-SPORT -><— Real Time w/o TSP .3— D-SPORT

 

Figure 9.15 Transit lateness at bus stop 2 for scenarios and 50 passenger/hr
boarding rate

185

 

 

 

 

Transit Lateness Profile at Bus Stop 3
EB VIC 50% - NB v/c 50°/o

 

 

Transit Lateness Profile at Bus Stop 3
EB vlc 75% - NB v/c 50°/o

 

 

 

 

 

 

 

 

1405~-——— —— ——

0,120+“ ——~— — ———-—~ ——

3‘ 100 — ,_. — ,- »—~ ~— ~A~ —

g 80 _, __ ‘ __v \ A ' __.. A

~ 60-5 “—7 - a—m ._ ~— — 2

3 4o - -.__' __ _ __ __ .._._____" _

20 TTTIITITTIIITIITTIIIIIIITfi—l 20 ITIIIITITIII‘ITIITIIITTTTI
1357911131517192123 1357911131517192123

Transit Vehicle ID Transit Vehicle ID
TT+PE~Timed + PE-TTiFrerTTSP' T ' + Pre-Timed + Pre-TirhTeB‘T’SPT ’ T
as Real Time w/o TSP -e- D-SPORT +1— Real Time w/o TSP ~9- D-SPORT

 

 

Transit Lateness Profile at Bus Stop 3
EB VIC 75% - NB VIC 75%

 

 

 

 

 

 

Transit Lateness Profile at Bus Stop 3
EB We 95% - NB We 50%

 

 

 

 

q, 120
8’
O
C
.9
3
20 TTT—ITITIIIITT—FTIIIITTfifiIfi
1357911131517192123 1357911131517192123
Transit Vehicle ID Transit Vehicle ID
T #PrTe-‘Iimed + Pre-Timed TSP 4» Pro-Timed + Pro-Timed TSP
4.- Real Time w/o TSP ~9- D-SPORT -.\- Real Time wlo TSP «e— D-SPORT

 

 

 

Transit Lateness Profile at Bus Stop 3
EB VIC 95% - NB vlc 75%

 

 

Transit Lateness Profile at Bus Stop 3
EB vlc 95% - NB vlc 95%

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

140
a, 120 -
g 100 ~
2 so—
9 so—
-‘ 4o«
20 TIIIIIIIIIIIIIIIITIIITIIT mIIITIITTfiTIIIIIIIFIITTfi—T
1357911131517192123 1357911131517192123
Transit Vehicle ID Transit Vehicle ID
+ Pre—Timed + Pre-Tirned TSP +Pre-Trmed + Pro-Timed TSP
+£— Real Time w/o TSP +7— D-SPORT -x— Real Time wlo TSP +1 DSPORT

 

Figure 9.16 Transit lateness at bus stop 3 for scenarios and 50 passenger/hr

boarding rate

186

 

 

 

Transit Lateness Profile at Bus Stop 4
EB VIC 50% - NB vlc 50°/o

  

 

 

 

Transit Lateness Profile at Bus Stop 4
EB VIC 7570 ' NB V/C 50%)

 

  

 

 

 

 

 

 

 

 

 

  

 

 

 

 

 

140 1" — ——~-— — — — — —— ——— 140 — --~-—— # ————— ——
m120«-—~—— ——~———-— ~-——~— m120-~~——~——————-——-——
a; a; 100 +s—~—~- —— .————-—~ —-—~ ——»—— __-—.—
5 3‘.»

3 3
1357911131517192123 1357911131517192123
Transit Vehicle ID Transit Vehicle ID
lPrETTmédTT + Pre-Timed TSPT" T + Pre-Timed +PreTimTédTT'SP T T
31:— Real Time w/o TSP 9— D-SPORT ->:— Real Time w/o TSP -e- D-SPORT
Transit Lateness Profile at Bus Stop 4 Transit Lateness Profile at Bus Stop 4
EB v/c 75%-NB VIC 75% EB v/c 95%-NB VIC 50%
(I)
of
(I)
d)
C
2
3
1357911131517192123 1357911131517192123
Transit Vehicle ID Transit Vehicle ID
TJ-TPreTTTiniéd + PreTimédTTTSP TTT ’ +PTe-Timed + Pre-Tirned TSPT T
—v.— Real Time w/o TSP +3— D-SPORT -x— Real Time w/o TSP ~8— D-SPORT

 

 

Transit Lateness Profile at Bus Stop 4
EB VIC 95% - NB v/c 75%

140
a) 120 T

U)

Latenes

 

 

 

ITTIIIIIITWIIIIIITIIIIII

3 5 7 911131517192123
Transit Vehicle ID

20
1

41- Pre—Timed

—><— Real Time w/o TSP «a- D-SPORT

_T—JPrmTTSP TA

 

 

Lateness, s

"‘T—a-THe-‘Trir—ned

Transit Lateness Profile at Bus Stop 4
EB VIC 95% - NB VIC 95%

 

 

 

 

 

 

 

IIIIIIIIIIIIIIIIIIIIIII1

3 5 7 911131517192123
Transit Vehicle ID

1

 

 

 

+ Pre-Timed TSP T

—><— Real Time wlo TSP .3, D-SPORT

i

 

 

Figure 9.17 Transit lateness at bus stop 4 for scenarios and 50 passenger/hr
boarding rate

187

9.5. Summary

This chapter summarizes the results of simulation experiments that were performed to
assess traffic network performances under different traffic control approaches in a range
of traffic and transit demand scenarios. The results show a significant reduction in transit
delay and number of stops when the D-SPORT controller is implemented. The results
further show that D-SPORT is able to react to traffic fluctuations by allocating green
times at traffic signals along the network to accommodate both general traffic and transit
movements. The simulation results further show that the D-SPORT controller has the
ability to improve and regulate transit schedule adherence at bus stops by reducing transit
lateness, as well as to improve transit service reliability by reducing the variation of

transit arrivals at bus stops.

188

CHAPTER X.

CONCLUSIONS AND
RECOMMENDATIONS

 

fforts were made in this research to develop and integrate TSP and a real-
time traffic signal control model for an arterial network. The importance of
this work is to provide practical solutions for several issues related to the
integration of the needs of both transit vehicles and general traffic. This chapter
summarizes the work performed in this research, the major findings of the research, and

recommendations for future work.

10.1. Introduction

Implementing TSP in coordinated networks of signalized arterials presents several issues
linked to the differing operational requirements associated with transit vehicles and the
general traffic. The primary problem is that transit vehicles often do not maintain

progression with the general traffic stream due to their need for accommodating

189

passenger boarding and alighting at bus stops. Another issue is related to the nature of
TSP operations. TSP operations aim to provide preferential treatments to transit vehicles
approaching individual intersections. While granted preferential treatments at an
intersection may benefit transit vehicles, such actions may negatively impact local traffic
and have further negative repercussions at other surrounding intersections if network-
based operational objectives are not considered. For instance, providing transit vehicles
with preferential treatment might lead to an increased reliability in schedule adherence,
which would be an obvious benefit to transit riders. However, temporary changes in
signal timings could also interfere with the established traffic progression patterns. An
ensuing loss of progression could potentially increase delay at downstream intersections
for non-transit vehicles, thus potentially negating all the transit benefits from a system
standpoint. Furthermore, the reception of conflicting or simultaneous priority requests is
generally not handled appropriately in existing systems. Priority requests are handled on
a first-come-first-serve basis. Better results could be achieved if attempts were made to
better utilize potentially available information about transit operations and traffic

conditions.

The objective of this research was to develop a dynamic signal priority thimization in
real-time traffic control system (D-SPORT), which is capable of providing efficient
transit priority at signalized intersections operated within a coordinated network.
Efficiency was defined as the ability to benefit transit vehicles, in terms of reducing
transit delay and improving schedule adherence by reducing transit lateness, while

improving non-transit traffic conditions by reducing traffic delay and number of stops.

190

 

To allow the proposed signal controller to consider all the performance measures of
interest, a cost function was developed to combine all of the aforementioned objectives in
one mathematical formula, and express them in relation to variables that can be measured
by available traffic surveillance and vehicle tracking technologies, such as traffic
volumes, transit vehicle location, and signal control timing parameters. Within this cost
function, estimates of delays incurred by the general traffic are obtained by utilizing the
delay equations defined in the HCM 2000, while the number of stops incurred by

individual vehicles is computed using macroscopic queuing analysis models.

The D-SPORT controller bases its priority actions on the anticipated time a transit
vehicle will need to reach a specific point along its route (such as a bus stop or a stop
line). This travel time is a function of the distance between a transit vehicle’s current
location and the selected target point, the desired transit speed, the expected transit delay
before reaching the target point, and the time consumed to board and alight passengers at
bus stops located between their current location and the target point are both considered
as well. Similar to traffic performance parameters, the expected transit delay is calculated
based on HCM 2000 delay equations. The prediction of dwell time at bus stops is on the
other hand achieved using an ANN model that was developed in this research and which

was demonstrated to produce good results.

The determination of appropriate signal timing actions is finally achieved using a GA
model seeking to minimize the cost function value. The value of the cost function is a

function of delay, the number of stops, transit travel time, and transit vehicle schedule.

191

Testing of the D-SPORT controller was performed using a virtual network coded in the

VISSIM microscopic simulation model. The tests examined the performance of the

proposed signal control logic compared to alternative traffic signal control approaches

commonly used in practice. Performance of the proposed model was compared more

specifically with (l) pre-timed signal control optimized offline using SYNCHRO 6,

(2) pre-timed signal control with active TSP offering green extensions and early green

recalls, and (3) a real-time signal control that does not provide any preferential treatments

for transit vehicles.

10.2. Major Conclusions and Research Contributions

The major findings of this research are summarized below:

Dwell time variability affects signal controller’s ability to provide transit vehicles
with efficient priority treatments. The danger is for the controller to provide a
specific treatment that will end up not being used because of a longer or shorter
than expected dwell time. Such wasted treatments could negatively affect traffic
signal control at the prioritized intersection and cause unduly increases in vehicle
delays. This is particularly significant issue at intersections with near-side bus
stops, where buses stopping near the stop line provide a limited timeframe to

implement TSP strategies following indication that a bus is ready to leave.

Although dwell time variability might impact the ability of transit vehicles to
benefit from priority systems, dwell time variability can be accommodated by
changing transit vehicle arrival patterns at an intersection relative to the traffic

signal state. The primary action would be to increase the percentage of transit

192

 

vehicles arriving within a time window for which they is a higher probability to
benefit from an early green recall or a green time extension, i.e., less sensitivity to

changes in actual arrival time.

While the potential effects of dwell time variability in any TSP algorithm is an
important key in transit priority planning and operations, its consideration
depends on the objective of the study. If the objective is to better model and
evaluate general traffic impacts associated with land development, or roadway
improvements, then experiments using average dwell time can be used. If the
objective is instead to determine the efficiency of selected priority strategies or
algorithms with respect to transit operations, or the efficiency of signal control
with TSP treatment, it then becomes important to consider dwell time variability

in the TSP optimization.

An ANN prediction model was successfully used to predict the stop line arrival
time of buses traveling on an intersection with near-side bus stops. Results of
model testing show an ability to improve the operation of transit signal priority
systems by enhancing the ability to correctly predict bus arrival times at
intersections and reduce unnecessary signal timing alterations. In the test
scenarios considered, this enhanced prediction capability resulted in lower delay
for buses and lower impacts on general traffic when compared to a priority system

simply basing its decisions on average dwell and travel times.

The GA was successfully able to optimize multiple cost functions concurrently

and to return a near-optimal signal timings for a real-time traffic control system to

193

 

accommodate both transit and non-transit traffic streams at the traffic network
level. The use of the GA has strongly promoted the real-time implementation of
the D-SPORT controller in the field due to the short time that is required to

perform the optimization process.

0 The performance of the D-SPORT controller was examined under different traffic
and transit demands levels, and was compared to other common-used signal
control systems. The simulation experiments have resulted in a significant
reduction in transit delays and number of stops. The simulation results also show
the D-SPORT controller ability to react to traffic fluctuations and to re-distribute
the green times in the network to accommodate general transit and non-transit
needs. In terms of schedule adherence, the D-SPORT implementation reduced

both transit lateness at bus stops and variations in bus stop arrival times.

0 The D-SPORT controller was able to solve several issues related to implementing
TSP in a coordinated traffic system. More specifically, the D-SPORT controller is
insensitive to the bus stop classifications, the variability of dwell times and its
impact on accurately predicting transit travel times, implementing TSP in a

coordinated traffic networks, and the issue of simultaneous priority recalls.

10.3. Recommendations for Future Research

Although this research has addressed, discussed, and solved several issues related to
integrating TSP within a real-time signal control coordinated traffic network, other issues
can be addressed and examined. The following summarizes some of the most important

issues that might need further and more detailed investigation:

194

Further investigations to improve the proposed model applicability by taking into
account more complicated network layouts and geometric designs such as grid
networks, phasing schemes, turning volume, traffic composition, saturated traffic

conditions, and two-way traffic networks.

Considering the pedestrian movements for more comprehensive modeling that

balances the need of transit vehicles, the general traffic, and pedestrians.

Sensitivity analysis of the D-SPORT controller cost functions to the microscopic

delay, queue and number of stops models.

The consideration of a car-following model in deveIOping the cost function, and
determining the D-SPORT controller efficiency and the traffic network

performance.

Adapt the dynamic speed concept as a part of the D-SPORT cost function.

Determine the D-SPORT controller’s ability to respond to unexpected traffic
events, such as traffic incidents, and the effect of sink-and-source traffic

movements.

Actual field experiments to validate the applicability of the D-SPORT controller
to real-world conditions, where other factors, such as weather and difficult to

model driver behavior, might influence the traffic network performance.

195

REFERENCES

 

Baker et al., An Overview of Transit Signal Priority. 2002, Washington, DC: ITS
America.

Vincent, R.A.a.H., K., Public Transport Priority as a Signalized-Controled
Junction: An Experiment in Bern, Switzerland. Traffic Engineering & Control,
1970.

Teng, H., et al., Simulation Testing of Adaptive Control, Bus Priority, and
Emergency Vehicle Preemption in New York City. in Transportation Reseach
Board, 82nd Annual Meeting. 2003. Washington, DC.

Wattleworth, J.A., K.G. Courage, and CE. Wallace, Evaluation of Bus-Priority
Strategies on Northwest Seventh Avenue in Miami. Transportation Research
Record, 1977(626): p. 32-35.

Klein, L.A., Sensor Technologies and Data required for ITS, book. 2001, London:
Artech House Publishers.

Parviainen, A., An Overview of available and Developing Highway Vehicle
Electronic Technology and Research Program. Transportation Technology and
Energy Branch, Ministry of Transportation. 1990, Ontario, Canada.

Drane, C. and C. Rizos, Positioning Systems in Intelligent Transportation
systems. 1997, London: Artech House Publishers.

196

 

10.

ll.

12.

l3.

14.

15.

l6.

17.

18.

19.

20.

Pearson, R. Traflic Signal Control. [cited 3/15/2007]; Available from:
http://www.calccit.ongitsdecision/serv and tech/Traffic signal control/traffsigre
Lprinthtmfiadaptive.

Bretherton, R.D., N.B. Hounsell, and B. Radia. Public Transport Priority in
SCOOT. in Proceedings of the 3rd World Congress on Intelligent Transport
systems. 1996.

Cranshaw, V. and M.J. Shaw. Vehicle Detection in Norwich in Highways and
Transportation. 1995.

Mathan H. Gartner, OPAC: A Demand Responsive Strategy for Trafi‘ic Signal
Control. Transportation Research Record, TRR, National Research Council,
Washington, DC, 1983. 906.

Jacobson, K.L. Transit Signal Priority Treatments in the Puget Sound Region. in
Pacific Rim Transtech Conference. 1993. Seattle.

Donati, f., et al., A Hierarchical decentralized trafiic light control system, the first
realization. IFAC 9th World congress, 1984. 11(1 lG/A-l ).

Head, KL. and PB. Mirchandani, A real-time traffic signal control system:
architecture, algorithms, and analysis. Transportation Research. Part C:
Emerging Technologies, 2001. 9(6): p. 415-432.

Head, K.L., P.B. Mirchandani, and S. Shelby, Estimation and prediction in a real-
time traffic adaptive signal control logic: the RHODES prototype. Informs
meeting, San Diego, 1997.

Mirchandani, PB. and DE. Lucas, Integrated Transit Priority and
Rail/Emergency Preemption in Real-Time Traffic Adaptive Signal Control.
Intelligent Transportation systems, 2004. 8: p. 101-115.

Dion, F., H. Rakha, and Y. Zhang., Evaluation of Potential Transit Signal Priority
Benefits along a F ixed-Time Signalized Arterial. Journal of Transportation
Engineering, 2004. 130(3): p. 294-303.

Deshpande, V. and J. Collura. ransit Signal Priority: Green Extension Benefits in
Congested Corridor. in Preprint CD-ROM, 83rd Annual Meeting of the
Transportation Research Board. 2003. Washington, DC.

N gan, V., T. Sayed, and A. Abdelfatah, Impacts of Various Trafiic Parameters on
Transit Signal Priority Efiectiveness. Journal of Public Transportation, 2004. 7(3):
p. 7 1-93.

Chang, G.-L., M. Vasudevan, and CC. Su, Bus Preemption under Adaptive

Control Environments. Transportation Research Record 1494, TRB, National
Research Council, Washington, DC, 1995: p. 146-154.

197

 

21.

22.

23.

24.

25.

26.

27.

28.

29.

30.

31.

32.

Transit Capacity and Quality of Service Manual, 2nd Edition. . TCRP Report 100,
ed. T.C.R. Program. 2003, Washington, DC: TRB, National Research Council.

TOWNES, M.S., et al. TCRP Report 19. Guidelines for the Location and Design
of Bus Stops. 1996, Washington, DC: Transportation Research Board,
Transportation Research Board.

Huffman, B., et al. Signal Priority Demonstration in Suburban Chicago. in
Proceedings, 1998 Bus Operations, Technology and Management Conference.
May 1998.

Chada, S. and R. Newland, Efifectiveness of Bus Signal Priority. Report NCTR—
416-04. National Center for Transit Research, University of South Florida. 2002,
Tampa, FL.

Feng, Y., J. Perrin, and RT. Martin. Bus Priority of SCOOT Evaluated in a
VISSIM Simulation Environment. in Preprint CD-ROM, 82nd Annual Meeting of
the Transportation Research Board. 2003. Washington, DC.

Byme, N., et al. Using Hardware-in-the-Loop Simulation to Evaluate Signal
Control Strategies for Transit Signal Priority. . in Preprint CD-ROM, 84th
Annual Meeting of the Transportation Research Board. 2005. Washington, DC.

Ghanim, M., F. Dion, and G. Abu-Lebdeh. Artificial Neural Network Bus Arrival
Time Prediction Tool for Transit Signal Priority with Nearside Bus Stop. in 86th
Annual Meeting of the Transportation Research Board. 2007. Washington, DC.

Dion, F. and M. Ghanim. Impact of Dwell Time Variability on Transit Signal
Priority Performance at Intersections with Nearside Bus Stop. in 86th Annual
Meeting of the Transportation Research Board. 2007. Washington, DC.

Kim, W. and L. Rillett. An Improved Transit Signal Priority System for Networks
with Nearside Bus Stops. in Preprint CD-ROM, 84rd Annual Meeting of the
Transportation Research Board. 2005. Washington, DC.

Rakham, H. and Y. Zhang, Sensitivity Analysis of Transit Signal Priority Impacts
on the Operation of a Signalized Intersection. Journal of Transportation
Engineering, 2004. 130(6): p. 796-804.

Chien, S., Y. Ding, and C. Wei, Dynamic Bus Arrival Time Prediction with
Artificial Neural Networks. Journal of Transportation Engineering, 2002. 128(5):
p. 429-437.

Al-Sahili, K.A., and Taylor, W.C. , Evaluation of Bus Priority Signal Strategies in
Ann Arbor, Michigan. Transportation Research Record 1554, Transportation
Research Board, Washington, DC, 1995: p. 74-79.

198

 

33.

34.

35.

36.

37.

38.

39.

40.

41.

42.

43.

44.

45.

Duerr, P.A., Dynamic Right-of- Way for Transit Vehicles: An Integrated Modeling
Approach for Optimizing Signal Control on Mixed Traffic Arterials.
Transportation Research Record, TRB, National Research Council, Washington,
DC, 2000. 1731: p. 31-39.

Sun, W. and Y.-x. Yuan, Optimization Theory and Methods, Nonlinear
Programming. Vol. 1. 2006, New York: Springs.

Robinett, R.D., et al., Applied Dynamic Programming for Optimization of
Dynamic Systems’. 2005, Philadelphia, PA: Society for Industrial and Applied
Mathematics.

Murty, K.G., Operations Research: Deterministic Optimization Models. 1995,
Englewood Cliffs, NJ: Prentice Hall.

Holland, J.H., Adaptation in Artificial and Natural Systems. University of
Michigan Press. 1975, Ann Arbor.

Goldberg, D.E., Genetic Algorithms in Search, Optimization, and Machine
Learning. 1989, Boston, MA, USA Addison-Wesley Longman Publishing Co.,
Inc.

Ayyub, B.M. and RH. McCuen, Probability, Statistics, & Reliability for
Engineers. 1997, New York: CRS Press.

Haykin, S., Neural Networks: A Comprehensive Foundation. 1994, New York,
NY: Macmillan.

Singh, A. and G. Abu-Lebdeh, Conditional Independence Graph-based State-
Space Neural Network Modeling for Short-Tenn Travel Time Prediction on
Urban Arterials. Presented at Artificial Neural Networks in Engineering
(ANNIE), St. Louis, MO, 2006.

Mandavilli, S., Y Najjar, and G. Abu-Lebdeh, Modeling Crash Rates For the
Kansas Rural Expressway Network. Artificial Neural Networks in Engineering
(ANNIE), 2005. 15: p. 721-730.

Ahmed, K., G. Abu-Lebdeh, and R.W. Lyles, Prediction of Pavement Distress
Index with Limited Data on Causal Factors: An Autoregression Approach.
International Journal of Pavement Engineering, 2006. 7(1): p. 23-35.

Bailer-Jones, D. and M. Bailer-Jones. C.A.L. Modelling data: Analogies in neural
networks, simulated annealing and genetic algorithms. In Model-based
reasoning: Scientific discovery, technological innovation, values,
Kluwer/Plenum., 2001. in Model-based reasoning: Scientific discovery,
technological innovation, values. 2001. Kluwer/Plenum.

May, A.D., Trafiic Flow Theory. 1990, New Jersy: Prentice-Hall.

199

 

46.

47.

48.

49.

50.

51.

52.

53.

54.

55.

56.

57.

58.

59.

60.

CORSIM user manual version 1.01. August 1996: Federal Highway
Administration.

AIMSUN 2 User Manual. V0]. 4.0. November 2000: Transport Simulation
systems.

Paramics trafiic simulation nrodeller V3.0: Reference manual. 1999, Edinburgh:
Quadstone Ltd.

Planung, P., VISSIM 4.10 Users Manual. PT V Planung Transport Verkehr AG.
2005, Karlsruhe, Germany.

Chapman, S.J., Fortran 90/95 for Scientists and Engineers. 1 ed. 1998, Boston,
USA: McGraw-Hill Companies.

KIM, W., An Improved Bus Signal Priority System for Networks with Nearside
Bus Stops, in Office of Graduate Studies. 2004, Texas A&M University.

Trafficware. SYNCHRO 6. Traffic Signal Simulation Software User Manual.
2005. Albany, CA.

Rakha, H., M. Snare, and F. Dion, Vehicle Dynamics Model for Estimating
Maximum Light Duty Vehicle Acceleration Levels. Transportation Research
Record, 2004. 1883: p. 40-49.

Ceylan, H., Developing Combined Genetic Algorithm — Hill-Climbing
Optimization Method for Area Traffic Control. Journal of Transportation
Engineering, 2006. 132(8): p. 663-671.

Highway Capacity Manual, Sepcial Report 209, in Transportation Research
Board. 2000, National Research Council: Washington, DC.

Garber, NJ. and LA. Hoel, Traffic and Highway Engineering. Third ed. 2002,
Pacific Grove, CA: Broose/Cole.

Kang, Y.-S., Delay, Stop and Queue Estimation for Uniform and Random Traffic
Arrivals at F ixed-Time Signalized Intersections, in Civil and Environmental
Engineering. 2000, Virginia Polytechnic Institute and State University.

Planung, P., VISSIM 4.30-00 COM Interface Manual. PTV Planung Transport
Verkehr AG. 2007, Karlsruhe, Germany.

Schizas, C.N., et al., Artificial Neural Networks in Forecasting Minimum
Temperature. Second International Conference on Artificial Neural Networks,
1991(18-20 Nov 1991): p. 112-114.

Battiti, R., F irst-and second-order methods for leaming: between steepest descent
and Newton '5 method. Neural Computation, 1992. 4(2): p. 141-166.

200

61.

62.

63.

Hagen, MT. and MB. Menhaj, Training feedforward networks with the
Levenberg-Marquardt algorithm. IEEE Trans. on Neural Networks, 1994. 5(6): p.
989-993.

Demuth, H. and M. Beale, Neural Network Toolbox For Use with MATLAB,
User’s Guide. Vol. 3. 1998, Natick, MA: The MathWorks.

Moreira, J.E., SP. Midkiff, and M. Gupta, A comparison of Java, C/C++, and
FORTRAN for numerical computing. Antennas and Propagation Magazine, IEEE,
1998. 40(5): p. 102-105.

201

APPENDIX A: ANN MODEL SOURCE
CODE IN MATLAB 6

 

%with normalization tansig purelin,50—3

% Size of original data m is the number of cases and n is the number of
% attributes ( including the three targets)

[m,n]=size(all);

disp('Original Data dimensions [Attributes Casesl');

diSptln m1);

m_dim
n_dim

I“;
I1;

% Normalization

p = all(:,4:l7); p = p';

t = all(:,1); t = t';
[pn,minp,maxp,tn,mint,maxt] = premnmx(p,t);

% Preparing data for Training, Validation and Test, 1/2 data for
training
% and 1/4 for validation and 1/4 to test

iitr = 1:1:5000;
iival 5001:1z7000;
iitst 7001:1:m;

val.P = pn(:,iival); val.T = tn(:,iival);
test.P = pn(:,iitst); test.T = tn(:,iitst);
ptr = pn(:,iitr); ttr = tn(:,iitr);

202

% Setting up Lambert-Marrquardt FF Backpropagation Network. (More
details at 'Neural Network Toolbox: Backpropagation: Levenberg-
Marquardt (trainlmI')

disp('Setting up network ...');

% Network Architecture and Sigmoid Functions
net = newff(minmax(ptr), [8 l],{'logsig' 'purelin'},'trainlm');

% Training Parameters
net.trainParam.epochs = 500;

net.trainParam.goal = le—4; % MSE to reach 0.00000001
net.trainParam.max_fail = 10 ; % Five Validations to check when
error rises

net.trainParam.mem_reduc = l; % Full Jacobian Calculated, no

memory restrictions
net.trainParam.min_grad = le-8;
net.trainParam.mu = 0.0001;

net.trainParam.mu_dec = 0.1;
net.trainParam.mu_inc = 10;
net.trainParam.mu_max

= lelO;
net.trainParam.show = 1;
net.trainParam.time = inf;

% Learning Rate
net.trainparam.lr = 0.1;

% Performance 2 'Mean Square Error’
net.perforchn = 'mse';

% Training
disp('Training ...');
[net.tr]=train(net,ptr,ttr,[],[],val,test);

% Ploting the training results

figure(10);

title('Plot of Training, Validation and Test by Epochs'); hold on
plot(tr.epoch,tr.perf,tr.epoch,tr.vperf,tr.epoch,tr.tperf)
legendI'Training','Validation','Test',-l);

ylabel('Mean Squared Error'); xlabel('Epoch')

hold off

% Simulation and Post training analysis

an = sim(net,pn);
a = postmnmx(an,mint,maxt);

wts = net.iw;

% Regression Analysis
figure(l)
title('Regression Analysis of Results'); hold on

203

[m,b,r] = postreg(a(1,:),t(l,:)); hold off

% Results - Network Training

an2 = sim(net, ptr);

a2 = postmnmx(an2,mint,maxt);

figure(2);

title('Regression Analysis of Training'); hold on
[m,b,r] = postreg(a2(l,:),t(:,iitr)); hold off

% Results - Network Validation

an4 = sim(net, val.P);

a4 = postmnmx(an4,mint,maxt);

figure(4);

titlef'Regression Analysis of Validation'); hold on
[m,b,r] = postreg(a4(l,:),t(:,iival)); hold off

% Results - Network Test

an3 = sim(net, test.P);

a3 = postmnmx(an3,mint,maxt);

figure(3);

title('Regression Analysis of Testing'); hold on
[m,b,r] = postreg(a3(l,:),t(:,iitst)); hold off

format short;
disp('Network Architecture');
disp('Number of Layers');
disp(net.numlayers+l);
nneurons = 7;
slayer = size(net.layers);
for i=1:slayer(l,1)
nneurons =[nneurons net.layers{i}.size];
end

predicted = a3';
actual = t(7000:m_dim)';

204

.......

ROGRAM

APPENDIX B: GA PLATFORM
SOURCE CODE IN FORTRAN 90/95

 

Genetic Algorimth Code for Traffic Signal Controller USE.
Written By:
Mohammad Shareef Ghanim
Civil Engineering Dept.
Michigan State University

June lst 2008
Genetic

IMPLICIT NONE

INTEGER

REAL, ALLOCATABLE, DIMENSION(:,:)::

REAL::
INTEGER

REAL, ALLOCATABLE, DIMENSION(:,:)::
REAL, ALLOCATABLE, DIMENSION(:,:)::
REAL, ALLOCATABLE, DIMENSION(:)::

INTEGER

INTEGER::
INTEGER::
INTEGER::
INTEGER::
INTEGER:

REAL::
INTEGER
REAL::
REAL::
REAL::
REAL::C
REAL::

, ALLOCATABLE, DIMENSION(:,:):: ipop, offspring, child
m_rnd
s_rnd

, ALLOCATABLE, DIMENSION(:):: parent
var, x,
bestever
randsel, minmax
, ALLOCATABLE, DIMENSION(:):: counter
cumcounter

popsize, generation,
bestindex, elite
cross, mutate
:i, n, j, igen
str_mutate
i_mutate
minfit, maxfit,
minnorm, maxnorm,
best
ross_p,

cross_i,

eval

dv, str

sumfit,avgfit

sumnorm, avgnorm
check_cross, check_mutate
mutate_i

mutate_p,

205

REAL, ALLOCATABLE, DIMENSION(:):: fit, freq
REAL, ALLOCATABLE, DIMENSION(:)::nevals
INTEGERz: eval_count

REAL:: time_in, time_out

REAL:: 5

REAL:: fitval

CHARACTER:: dummy

CHARACTER:: selection, Check_elitism
INTEGER, DIMENSION(l:l):: rand_seed
INTEGERzz icase

eval_count = l
I Constants
dv

str

II II
1.5
[\J

ALLOCATE(minmaX(1:2))
I Genetic Algorithm Parameters

OPEN ( UNIT = 100, FILE = 'C:\Final Project\Project
Network\ga_parameters.txt', STATUS = 'UNKNOWN')

READ(100, *) dummy, selection I Selection Method, R for Roulette
Wheel, T for Tournament Selection

READ(100, *) dummy, popsize ! Population Size, Should be EVEN
number

Number of Generations
Crossover Probability
Mutation Probability
Applying Eltisim? 'Y' or 'N'
Random seed

Minimum Green

Maximum Green

READ(100, *) dummy, generation
READ(100, *) dummy, cross_p
READ(100, *) dummy, mutate_p
READ(100, *) dummy, Check_elitism
READ(100, *) dummy, rand_seed
READ(100, *) dummy, minmax(l)
READ(100, *) dummy, minmax(2)
CLOSE(100)

oooooooooooo

I This is to activate MultiRun with different Random Seeds

icase = 100

call cpu_time(time_in)
rand_seed = icase
CALL RANDOM_SEED(put = rand_seed)

ALLOCATE(ipop(l:popsize,1:dv*str))
ALLOCATE(offspring(l:popsize,l:dv*str))
ALLOCATE(child(l:popsize,l:dv*str))
ALLOCATE(var(l:popsize, 1:dv))
ALLOCATE(x(l:popsize, 1:dv))
ALLOCATEIeval(1:popsize,l:3))
ALLOCATE(randsel(lzpopsize))
ALLOCATE(parent(l:popsize))
ALLOCATE(bestever(l:generation, 1:2))
ALLOCATE(counter(l:generation))
ALLOCATE(fit(l:generation*popsize))
ALLOCATE(freq(l:generation*popsize))
ALLOCATE(nevals(lzgeneration))
counter = 0

I Boundary Limits I
IIIIIIIIIIIIIIIII

OOOOOOOOOOOOOOOOO

I minmax(l)= 20.
I minmax(2)= 25.

206

I Generate Initial Strings I
IIIIIIIIIIIIIIIII

IIIIIIIIIIIIIIIII

ALLOCATE(m_rnd(l:popsize, 1:dv*str))
CALL RANDOM_NUMBER(m_rnd)

ipop = NINT(m_rnd)
DEALLOCATE(m_rnd)

I Starting The GA Process I

ooooooooooooooooo

DO igen = 1, generation

WRITE(* *)'::::==:::=::=:::===:::'
WRITE(*, *)'STARTING GENERATION NO: ', igen
WRITE(* *)'::===::::===::::::::=:'
WRITE(*,*)' .................................................. '
IIIIIIIIIIIIIIIII
I Decode the string to variables I
IIIIIIIIIIIIIIIII
! WRITE(*,*)'::::::::==::=:=:='
I WRITE(*,*)'Decoded Variables'
! WRITE(*'*)'::::::==::::::::='
var=0.
x=O.
DO 1 = l,popsize
DO n = 1,dv
DO j=l+(n-l)*str, str*n
var(i,n)=var(i,n) + ipop(i,j) * ( 2.** (( str * n ) - j) )
END DO
x(i,n) = (var(i,n)/(2.**str)) * ( minmax(2)- minmax(l) )+ minmax(1)
END DO
END DO

IDO i = 1. popsize
IDO n = 1, dv

Ix(i,n) (var(i,n)/(2**str—1))*(minmax(2)-minmax(l))+minmax(1)
IEND DO

IEND DO

IIIIIIIIIIIIIIIII

I Calculate the Fitness I
IIIIIIIIIIIIIIIII

! WRITE(*,*)'=================='

I WRITE(*,*)'Variables Fitness'

' WRITE(*,*)'======:==========='

I Rounding to the nearest 5—second green duration
x = NINT(X)

eval=0.

ICalculate Fitness of strings

DO i = l,popsize

CALL OPT(x(i,:), fitval)

eval(i,l)=fitval

fit(eval_count) = fitval
fitval = O.

eval_count = eval_count+l
END DO

sumfit=sum(eval(:,l))
avgfit=sumfit/popsize
minfitzminval(eval(:,l))
maxfit=maxval(eval(:,l))

INormalizing Fitness

IF (maxfit==minfit) exit
DO i = l,popsize

207

eval(i,2)= (eval(i,1)—minfit)/(maxfit—minfit)
END DO

OPEN ( UNIT = 123, FILE = 'C:\Final Project\Project
Network\evaluations.txt', STATUS = 'UNKNOWN')

DO 1 = 1, popsize

WRITE(123,*) -1*eval(i,l)

END DO

IFinidng Sum, Average, Minimum and Maximum NormalizedFitness
sumnorm=sum(eval(:,2))

avgnorm=sumnorm/popsize

minnorm=minval(eval(:,2))

maxnorm=maxval(eval(:,2))

ICalculate Cumilative Normalized cost
eval(l,3)=eval(l,2)/sumnorm

DO i = 2,popsize

eval(i,3)=eval(i,2)/sumnorm+eval(i—l,3)

END DO

IWRITE(*,*)' .................................................. '
I WRITE(*,*) 'Sum = ', sumfit, sumnorm

I WRITE(*,*) 'Avg = ', avgfit, avgnorm

I WRITE(*,*) 'Min = ', minfit, minnorm

I WRITE(*,*) 'Max = ', maxfit, maxnorm

I WRITE(*,*)' .................................................. '
IIIIIIIIIIIIIIIII

I Roullette Wheel Selection I
IIIIIIIIIIIIIIIII

IF (selection == 'R') THEN

ALLOCATE(m_rnd(l:popsize, 1))

CALL RANDOM_NUMBER(m_rnd)

randsel(:) = NINT(m_rnd(:,l) * 100000000.)/100000000.
DEALLOCATE(m_rnd)

parent=0

DO i = l, popsize
DO j=1, popsize
IF (eval(j,3)>=randsel(i)) THEN

EXIT
END IF
END DO

parent(i)=j
END DO
IIIIIIIIIIIIIIIII
I Tournament Selection I
IIIIIIIIIIIIIIIII
ELSE IF (selection == "T') THEN

ALLOCATE(m_rnd(l:popsize, 2))
CALL RANDOM_NUMBER(m_rnd)
parent=0
m_rnd = NINT(m_rnd*(popsize-l)) + 1
DO i = 1, popsize
IF ( m_rnd(i,l)>= m_rnd(i,2)) THEN
parent(i) = m_rnd(i,1)

ELSE

parent(i) = m_rnd(i,2)
END IF

208

 

END DO
DEALLOCATE(m_rnd)

ELSE

WRITE(*,*) 'Please select a Selection Method'

END IF

I WRITE(*,*)' .................................................. '
IIIIIIIIIIIIIIIIIIII!

I Selecting New Generation I
IIIIIIIIIIIIIIIIIIIII

[WRITE(*,*)'::::::::::::::::::=::::='

IWRITE(*,*)'Selecting New Generation'
!WRITE(*,*)'========================'

offspring=ipop(parent,:)

IDO i = l,popsize

IWRITE(*,40) parent(i), offspring(i,1:8) , offspring(i,9:l6),
offspring(i,l7:24), offspring(i,25:32)

IEND DO

I40 FORMAT(I3, ' ', 8I2,‘ ', 812, ' ', 812, ' ', 812)

I Applying CrossOver I

ooooooooooooooooooooo

[WRITE(*,*)'::::::::=:::::::::'
IWRITE(*,*)'Applying CrossOver'
[WRITE(*,*)':::::::::::::::=::'

child: 0
DO i = 1, popsize, 2

CALL RANDOM_NUMBER(s_rnd)
check_cross=s_rnd

IF (cross_p >= Check_cross ) THEN

counter(igen) = counter(igen) + 2

CALL RANDOM_NUMBER(S_rnd)

cross= NINT(s_rnd*(str*dv—1—2))+2
DO j = 1, cross

child(i, j)=offspring(i,j)

child(i+1, j)=offspring(i+l,j)

END D0
D0 j = cross+1, dv*str
child(i,j) = offspring(i+1, j)
child(i+l,j) = offspring(i, j)
END DO
else

DO j = l, dv*str

child(i,j)=offspring(i,j)

child(i+1,j)=offspring(i+l,j)
END DO

END IF
END DO

209

 

 

IDO i = l,popsize

IWRITE(*,SO) Child(i,l:8) , child(i,9:16), Child(i,l7z24),
Child(i,25:32) ’

IEND DO

I50 FORMAT(8I2,' ', 8I2, ' ', 8I2, ' ', 812)

IWRITE(*,*)' .................................................. '

I Starting Mutation I
lllllllllllllllll

3WRITE(*,*)':::::::::::::::::'
IWRITE(*,*)' Applying Mutation'
'WRITE(* *)'::::::=::::=:::::'
DO i = 1, popsize*dv*str
CALL RANDOM_NUMBER(s_rnd)
Check_mutate = s_rnd
IF(check_mutate <= mutate_p ) THEN
counter(igen) = counter(igen) + 1
CALL RANDOM_NUMBER(s_rnd)
mutate = NINT(s_rnd*(dv*str-l) +1)
CALL RANDOM_NUMBER(str_mutate)
i_mutate = NINT(str_mutate * (popsize—l)+l)

IF (child(i_mutate,mutate)==O) THEN
child(i_mutate,mutate)=l
else
child(i_mutate,mutate)=0
END IF
END IF
END DO

IDO i = l,popsize
IWRITE(*,GO) Child(i,l:8) , child(i,9:16), Child(i,l7:24),
child(i,25:32)

IEND DO

160 FORMAT(8I2,' ', 812, ' ', BIZ, ' ', 812)

IWRITE(*,*)' .................................................. '
IIIIIIIIIIIEIIEII

I Ellitism I
I!!!!!!!!!!!!!!!!

fiWRITE(*'*)'==::::::==::::::='

IWRITE(*,*)'Applying Ellitism'
[WRITE(*,*)'z::::::::::::::::'

best = eval(l,1)
bestindex=1

DO 1 = l, popsize
IF (eval(i,l)>= best) THEN
best=eval(i,l)
bestindexzi

END IF

END DO

counter(igen) = counter(igen) +1
IF (Check_elitism == 'Y') THEN

CALL RANDOM_NUMBER(s_rnd)
elite=NINT(s_rnd*(popsize-l)+1)
child(elite,:) = ipop(bestindex,:)

ELSE IF (Check_elitis == 'N') THEN

210

 

I Do Nothing
ELSE

WRITE(*,*)‘Please Make an Elitism Selection "Y" if yes, "N" if no.
Elitism cannot be left empty'

END IF

IDO i = l,popsize

IWRITE(*,70) Child(i,l:8) , child(i,9:16), Child(i,l7z24),
child(i,25:32)

IEND DO

I70 FORMAT(812,‘ ', 812, ' ', 812, ' ', 812)

IWRITE(*,*)' .................................................. '

I Preparing For Next Generation I

I GoTo Next Generation I
WRITE(*,*)'::::==:::===::========'
WRITE(*,*)'END OF GENERATION No: ', igen
WRITE(*,*)'=::=::::::::=:=:::::::'

WRITE(*,

*)' .............................................................

bestever(igen, eval(bestindex,1)

MIA
IIH

bestever(igen, avgfit
END DO
IIIIIIIIIIIIIIII!
I Number of Function Evaluations I
IIIIIIIIIIIIIIII!
fit = -f1t
DO 1 = l, popsize * generation
DO j = 1, i
IF ( fit(i) == fit(j)) THEN
freq(i) = freq(i) + 1
END IF
END DO
WRITE(*,*) i
END DO
5 = 0.
DO i = 1, generation

DO j = (i—1)*popsize+l, i*popsize
IF (freq(j) <> 1 )THEN
s = s + 1.

END IF

END DO

nevals(i) = popsize - s

s = 0.
WRITE(*,*) 1
END DO
I!!!!!!!!!!!!!!!!
I Final Resutls I
EIIIIIIIIIIIIIIII
WRITE(*,*)':::::::::::::=::==::::'

WRITE(*,*)'THE FINAL RESULTS ARE '
WRITE(*,*)'::::::::::::::::::::::'
best=eval(l,1)

211

 

bestindex=1

DO i = 1, popsize

IF (eval(i,l)>= best) THEN
best = eval(i,1)

bestindex = 1
END IF
END DO
OPEN ( UNIT = 200, FILE = 'C:\Final Project\Project
Network\bestever.txt', STATUS = 'UNKNOWN')

cumcounter = 0
DO 1 = 1, igen-1
cumcounter = cumcounter + counter(i)

WRITE(200,*) i, bestever(i, 1) , bestever(i, 2), counter(i),
cumcounter, INT(nevals(i))

END DO

WRITE(*,*)' .................................................. '
WRITE(*,*) 'Best Index =', bestindex

OPEN ( UNIT = 300, FILE = 'C:\Fina1 Project\Project Network\output.txt',

STATUS = 'UNKNOWN')

WRITE(BOO, 300) x(bestindex, 1:7 ), x(bestindex, 8:14)
WRITE(300, 300) x(bestindex, 15:21), x(bestindex, 22:28)
WRITE(BOO, 300) x(bestindex, 29:35), x(bestindex, 36:42)
call cpu_time(time_out)

write(300,*) time_out-time_in

300 FORMAT(7(F4.1, ' '), ’* ', 7(F4.l, ' '))

OPEN ( UNIT = 201, FILE = 'C:\Final Project\Project Network\sumary.txt',
STATUS = 'UNKNOWN')

call cpu_time(time_out)

WRITE(201,*) cross_p, mutate_p, INT(icase/lOO), bestever(igen-l,1),
time_out—time_in, INT(SUM(nevals))

WRITE(*,*) 'Closing ............................... '

DEALLOCATE(ipOp)
DEALLOCATE(offspring)
DEALLOCATE(child)
DEALLOCATE(Var)
DEALLOCATE(X)
DEALLOCATE(eval)
DEALLOCATE(randsel)
DEALLOCATE(parent)
DEALLOCATE(bestever)
DEALLOCATE(counter)

write(*,*) time_out
STOP

END PROGRAM

212

 

   

IIIEIIIII‘IILIIIIIIIIIIIILILIILILIII