FRAMEWORK AND MODELS
FOR THE PROVISION OF REAL-TIME DRIVER
INFORMATION
pV
ISAM ADNAN KAYSI
B.S. Civil Engineering
Concordia University
(1985)
M.S. Civil Engineering
Massachusetts Institute of Technology
(1988)
Submitted to the Department of Civil Engineering
In Partial Fulfillment of the Requirements
for the Degree of
Docwor 3 Philosophy in Transportation Systems
at the
Massachusetts Institute of Technology
February, 1992
© Massachusetts Institute of Technology, 1992
Signature redacted
Signature of Author
devartméfit of Civil Engineering
November §, 1991
Signature redacted.
Certified by
Moshe Ben-Akiva
Profgssor of Civil Engineering
Wnted <i Thesis Supervisor
Signature redacted
Accepted by
Eduardo Kausel
Chairman, Departmental Committee on Graduate Studies
ARCHIVES Department of Civil Engineering
MASSACHUSETTS INSTITUTE
(OF TENE NY
MAR O 0 1992
LIBRARIES
FRAMEWORK AND MODELS
FOR THE PROVISION OF REAL-TIME DRIVER
INFORMATION
NY
[SAM ADNAN KAYSI
Submitted to the Department of Civil Engineering
on November 8, 1991 In Partial Fulfillment of the Requirements
for the Degree of Doctor of Philosophy in Transportation Systems
ABSTRACT
More efficient utilization of existing urban road networks may be achieved through
dynamic traffic management schemes. Such schemes are needed to alleviate the
increasing level of traffic congestion and the detrimental effect it has on mobility,
accessibility, and safety in urban areas. The provision of drivers with real-time
information that is responsive to emerging traffic conditions has the potential to reduce
both recurrent and incident congestion. However, the level of benefits to be derived from
such advanced driver information systems (ADIS) depends to a large degree on the
quality of the guidance being provided. The possible occurrence of adverse impacts due
to improved information is demonstrated in this dissertation.
In order to insure the effectiveness of real-time driver information systems, there
is a need for an ADIS implementation framework that is based on a dynamic network
modeling approach capable of assessing accurately network performance and forecasting
traffic conditions that may exist in the near future. A framework that would satisfy these
requirements is adopted in this dissertation consisting of a surveillance system which
collects relevant traffic data, a congestion prediction module which provides the guidance
system with congestion information, and a control and routing module which determines
the guidance advice. The major principles embodied in this framework are:
Principle 1: Routing information should be based on projected traffic conditions.
Principle 2: There exists a need for an advanced COP module that makes use of current
traffic conditions, predicted demand levels, guidance being provided, and anticipated
driver response to guidance in making its predictions of future congestion in the network.
The use of a dynamic traffic assignment (DTA) model is suggested in order to provide
such a congestion prediction capability.
Principle 3: Consistency has to be maintained between the guidance being provided to
drivers and the predicted traffic conditions.
Page 2
A framework that is based on the above principles has the potential to minimize
the overall effect of congestion, generate useful and valid guidance, improve travel time
reliability, and reduce the potential adverse impacts of improved information.
The modelling needs of the constituent elements of the framework are analyzed,
with emphasis being placed on the dynamic network performance and driver behavior
modelling aspects of the DTA that is suggested for congestion prediction. Methods of
obtaining the O-D predictions required by the DTA are also discussed. It is concluded
that the state of the art in these areas is not yet prepared to support ADIS functions and
that significant research is still required.
In this dissertation we also analyze the value of the proposed framework vis-a-vis
other information provision scenarios and under various conditions using simulation. The
scenarios that were considered include guidance based on last reported travel times,
instantaneous travel times, and travel times that are predicted using a DTA. We present
the results of a case study that was used to test the principles embodied in the proposed
framework. The case study also analyzed the impact on guidance effectiveness of various
parameters associated with congestion prediction, control and routing, and well as system
demand and supply characteristics.
Thesis Supervisor: Dr. Moshe Ben-Akiva
Title: Professor of Civil Engineering
apt 3
Acknowledgements
The financial support of the University Transportation Centers Program, the MIT
Center for Transportation Studies, the UPS Foundation, and the Charles Stark Draper
Laboratory is appreciated.
Pacz 4
Table of Contents Darn
Chapter 1 Background 12
i.1 Motivation  2
1.2 The Context of ADIS
1.3 Objectives of this Dissertation
1.4 Summary of Major Findings
1.4.1 ADIS Implementation Framework
1.4.2 Modelling Needs of Framework Components "9
1.4.3 Case Study Results 9
1.5 Dissertation Outline 20
Chapter 2 Driver Information Systems: Principal Characteristics
- -
and Potential Impacts 3oud
2.1 Overview of Driver Information Systems 25
2.1.1 System Characteristics 25
2.1.2 Existing Pilot Projects 7
2.2 The Impact of Driver Information Systems 30
2.2.1 Potential Benefits of Improved Information 20
2.2.2 Potential Adverse Impacts of Improved Information 32
Chapter 3 Integrated Framework for the Implementation of ADIS 33
nz
3.1 The Proposed Framework )
hp!
3.2 System Interactions
3.3 Principles Underlying the Proposed Framework ‘9
3.4 Control and Routing (CAR) Strategies
3.4.1 CAR Logic
3.4.2 Temporal Update Frequency
3.4.3 Spatial Update Frequency
A
3.5 Possible COP Schemes
3.5.1 Types of Information Provided by COP to CAR 43
3.5.2 COP Models 4.
3.5.3 Maintaining Consistency with Actual Traffic Conditions 51
3.6 The Role of COP and CAR in Reducing the Adverse
Impacts of Improved Information 51
3.6.1 Possible COP/CAR Scenarios 51
3.6.2 Dealing with Overreaction and Concentration 52
3,7 Evaluation Criteria 55
~
 | A
"4  3
Chapter 4 Dynamic Traffic Assignment (DTA):
Background, Literature Review, and Proposed Framework 62
Need for Dynamic Network Models 63
Review of Dynamic Traffic Assignment Methodologies “7
Proposed Approach and Elements of DTA
— =
o
Dynamic Network Performance Modelling
4.4.1 Flow Computation Difficulties J
4.4.2 Dimensions of Dynamic Network Performance Models
-
4.4.3 Illustration ’4
4.5 Self Calibration of the DTA
4.6 Updating the 3-Dimensional O-D Matrix 4
~
4.6.1 O-D Updating Methodologies J
~r
4.6.2 Illustration <}
Chapter 5 Dynamic Driver Behavior Modelling 20
5.1 Overview 90
5.2 Modelling Imperfect Driver Information 22
5.3 Perception Revision and Information Integration 95
5.3.1 A Simple Information Integration Model 76
5.3.2 The Use of Bayesian Updating 28
5.4 Existing Models of Travel Behavior Adjustment
in Response to Acquired Information 100
5.5 Incorporating Psychometric Data 104
5.6 Integrated Framework for Modelling Dynamic Driver Behavior 105
5.6.1 Dynamic Driver Behavior: The Basic Concepts 105
5.6.2 The Hierarchy of Dynamic Driver Choices 107
5.6.3 Components of the Modelling Framework 109
5.7 Implications for Information Provision Policies 114
5.8 Design of Data Collection Program 115
5.8.1 Potential Sources of Data on Dynamic Driver Behavior 115
5.8.2 Types of Data to be Collected 116
5.9 Concluding Remarks and Future Research Directions 118
Pac- J
Chapter 6 Case Study Experimental Design “|15
6.1 Objectives and Outline of Case Study 126
6.2 Elements of Case Study 127
6.3 Experimental Factors 131
6.3.1 Guided Probability 131
6.3.2 CAR Factors [131
6.3.3 COP Factors 133
6.3.4 Role of the Surveillance System 136
6.3.5 Demand and Supply Characteristics 137
6.4 Measures of Effectiveness 137
6.5 Structure of the Simulation 139
6.6 The Efficiency of Scenarios 2 and 3: First Insights 141
Chapter 7 Analysis of Simulation Results 147
7.1 Base Case Input Data 147
7.2 Simulation Results 152
7.2.1 Guided Probability 152
7.2.2 COP Factors 152
7.2.3 CAR Factors 155
7.2.4 DTA-Specific Issues 158
7.2.5 Demand and Supply Characteristics 159
7.3 Delays of Guided and Unguided Vehicles 160
7.4 Delays by Departure Time 160
7.5 Statistical Analysis of Simulation Results 160
7.6 General Conclusions 162
dye I
Chapter 8 Concluding Remarks and Directions for Further Research 119
8.1 Contributions 198
8.2 Major Findings 199
8.2.1 ADIS Implementation Framework 199
8.2.2 Modeling Needs of Framework Components 199
8.2.3 Case Study Results 200
8.3 Directions for Future Research 202
8.3.1 Possible Extensions to Case Study 202
8.3.2 Possible Directions for Future Research 203
References 205
Appendix A 113
pe Wg. 8
J»
List of Figures Page
Fig. 1.1 Transport System Interactions 23
Fig. 1.2 Dynamic Traffic Management Overview 24
Fig. 3.1 Proposed Framework 58
Fig. 3.2 Real-Time System Operation 59
Fig. 3.3 Framework Implementation oJ
Fig. 3.4 No Driver Behavior Modelling 61
Fig. 3.5 Degenerate Case 52
Fig. 4.1 Sample Network RQ
Fig. 5.1 Taxonomy of Attribute Values 120
Fig. 5.2 Drivers’ Decision Process 121
Fig. 5.3 Pre-Trip Driver Behavior 122
Fig. 5.4 En-Route Driver Behavior 123
Fig. 5.5 Modelling Framework 124
Fig. 5.6 Role of Psychometric Data 125
Fig. 6.1 The Networks 143
Fig. 6.2 Simulation Structure 144
Fig. 6.3 Generic COP/CAR Functions 145
Fig. 6.4 COP/CAR with a DTA 146
Fig. 7.1a FLAT Demand Pattern 66
Fig. 7.1b SHARP Demand Pattern 167
Fig. 7.2a Impact of COP Information on Average Delay 168
Fig. 7.2b [mpact of COP Information on Guidance Validity 169
Fig. 7.2c Impact of COP Information on Maximum Delays 70
Fig. 7.2d [mpact of COP Information on Fraction with Long Delays 71
Fig. 7.2¢ Maximum Delays of Unguided Vehicles 172
Fig. 7.2f Average Delays for External Flows 173
Fig. 7.3a Impact of Demand Variability on Average Delays (74
Fig. 7.3b [mpact of Demand Variability on Guidance Validity 175
Fig. 7.3c [mpact of Demand Variability on Maximum Delays 176
Fig. 7.3d [mpact of Demand Variability on Fraction with Long Delays 7
Fig. 7.4a [Impact of Demand Pattern on Average Delay 178
Fig. 7.4b Impact of Demand Pattern on Maximum Delay 79
Fig. 7.5a Impact of CAR Logic on Average Delay 180
Fig. 7.5b [mpact of CAR Logic on Maximum Delay 181
Fig. 7.5¢c Impact of CAR Logic on Guidance Validity 182
Fig. 7.6a [mpact of Spatial Update Frequency on Average Delay 183
Fig. 7.6b Impact of Spatial Update Frequency on Maximum Delay
dupe  J)
Fig. 7.7a [Impact of Temporal Update Frequency on Average Delay 185
Fig. 7.7b Impact of Temporal Update Frequency on Maximum Delay 186
Fig. 7.7c Impact of Temporal Update Frequency on Guidance Validity 187
Fig. 7.8a [mpact of Incident Detection Delay on Average Delay 188
Fig. 7.8b {mpact of Incident Detection Delay on Guidance Validity 189
Fig. 7.8c Impact of Incident Detection Delay on Maximum Delay 190
Fig. 7.9a Average Delays for Low Congestion Case 191
Fig. 7.9b Maximum Delays for Low Congestion Case 192
Fig. 7.9¢ Guidance Validity for Low Congestion Case 193
Fig. 7.10 Impact of Accident Probability on Average Delay 194
Fig. 7.11a Average Delay of Guided and Unguided Vehicles 195
Fig. 7.11b Maximum Delays of Guided and Unguided Vehicles 196
Fig. 7.12 Delays by Departure Time for Route R1 197
8,
Foie a
 nl oo
List of Tables dd, Ag~
Table 1.1 Traffic Operations and Control Management Variables I!
Page 11
CHAPTER 1 BACKGROUND
1.1 Motivation
Urban traffic flows have increased in recent years, causing high levels of
congestion. The four factors that have contributed the most to the increasing levels of
congestion are (Pisarski, 1987):
» more travel in metropolitan areas
» more travel by car
» more travel to and from locations dispersed throughout the metropolitan region
» more travel in areas where the necessary highway capacity has not been provided
A recent study by the Federal Highway Administration projected an increase in
total delay due to urban freeway congestion in central cities of 360% between 1985 and
2005 (FHWA, 1987). Such traffic congestion has become a major problem in most urban
areas, with detrimental effects on the mobility of people and the movement of freight.
This large increase in traffic volumes is expected to have negative impacts on local
traffic, economic growth, community access, quality of life, safety, and environmental
quality (ITE, 1989).
Moreover, it is not expected that the construction of new road facilities will be
able to alleviate the anticipated congestion and related problems. Transportation planners
and traffic managers are hence looking for ways to improve the utilization of existing
facilities with enhanced traffic management and operations schemes. The use of more
innovative solutions which would make use of the latest technologies has been suggested
as a possible means of achieving such improved utilization of existing facilities.
Possible Future Traffic Systems
It has been suggested that during the coming decades, efficient operation of
existing road networks may be achieved through dynamic traffic management schemes
which make use of the available and anticipated advanced technologies (US DOT, 1989).
Within this context, Intelligent Vehicle-Highway Systems (IVHS) are currently being
developed. These IVHS systems envision the linking of road infrastructure, vehicles, and
drivers using advanced communication technology, computers, information display
equipment, and traffic control systems.
Two-way communication between centers controlling highway operations and
vehicle operators is expected to improve significantly the utilization of existing road
networks. Agencies operating highways will be able to monitor traffic conditions on the
network and provide more effective and dynamic management of network facilities.
ag. 12
Two IVHS schemes which are expected to provide such improvements are
Advanced Driver Information Systems (ADIS) and Advanced Traffic Management
Systems (ATMS). Other developments within the IVHS scope include Automated
Vehicle Control Systems (AVCS) which are expected to help drivers perform certain
vehicle control functions. The implementation of these schemes using advanced
technologies is envisioned to produce an integrated vehicle/roadway transportation system.
However, such implementation will require major efforts aimed at a unified design
and development of the required roadway infrastructure, vehicle instrumentation, and
central control strategies and methodologies related to information provision and traffic
control. In addition, the impact of the new systems on driver behavior and the potential
human factors problems which may develop also need to be addressed within the scope
of the comprehensive studies aimed at implementing IVHS systems.
The Potential Role of Advanced Driver Information Systems
In this context, Advanced Driver Information Systems (ADIS) based on modern
information technology may play an important role in reducing traffic congestion and
improving traffic flows and safety. ADIS have the potential to reduce delays due to both
incident and recurrent congestion by providing information to motorists with respect to
alternative paths to their destinations and actual traffic conditions on links of interest
using a combination of roadside signals and onboard systems.
Specifically, two interrelated techniques have considerable potential for alleviating
congestion and increasing network throughput. They involve (i) optimizing driver route
selection and (ii) making this selection responsive to real time road and traffic conditions,
i.e., by providing opportunities for dynamic routing. The potential benefits from the
implementation of these two schemes by ADIS are discussed in detail in Chapter 2.
1.2 The Context of ADIS
The overall objective of Advanced Driver Information Systems (ADIS) is to use
available network capacity more efficiently. These new systems differ from conventional
driver information sources in that advanced information and computation technologies
now enable us to implement optimal traffic guidance strategies that respond in real-time
to emerging conditions in the transportation network and to anticipated demand levels.
In this section we provide a brief overview of the interactions between the different
elements in such a system.
Static Demand-Supply Interactions
Traditionally, urban traffic flow has been viewed as an equilibration between the
demand for travel and the supply (capacity) of transportation facilities. This interaction
Jac. 3
has been viewed in a "static" context; i.e., it is assumed that both the O-D travel demands
and the link capacities are fixed for a certain period of time. When long-range changes
in demand or capacity are envisioned we call this process transportation planning. When
only short-range changes in demand or capacity are considered we call it transportation
operations management.
The interactions between the different components of the transportation system
are illustrated, via twin feedback loops, in Figure 1.1 (OECD, 1987). According to
Gartner et al. (1980), the user loop consists of an assessment of the existing (or
perceived) traffic conditions, a user decision process in light of the user preferences and
objectives, and a resulting O-D demand for travel. The user objectives are thought to be
the individual optimization of utility/disutility (e.g., minimization of travel time), hence
the term "user-optimization".
The system (or control) loop, on the other hand, consists of a managing traffic
authority (the "system") which assesses the same traffic conditions, but from a system
point of view. The system objective is, generally, to optimize collective measures of
performance, e.g., minimize total delay time, maximize average speed of travel, reduce
energy consumption and pollutant emissions, etc. (hence, the notion of "system-
optimization"). In view of these objectives the authority can "manipulate" the available
transport capacity to produce the desired effects. Such manipulation includes various
urban traffic control and transportation system management strategies. Decision variables
available to the authority can be banded into three groups: traffic signalization, traffic
operations improvements, and preferential treatments (see Table 1.1).
The resulting interaction among user- and system-optimization objectives is a
stationary point of operation, i.e. it produces fixed (equilibrium) volumes and performance
measures. The dynamics of this interaction are slow-changing. Based on experiences
acquired by either the user or the authority, over an extended period of time, adjustments
are made in the decisions and another equilibrium point is reached. Such systems do not
have the capacity to interact with the users on a real-time basis or try to influence their
behavior based on actual or predicted occurrences (e.g., incidents, congestion flareups,
etc.).
Dynamic Traffic Management
Advanced information and computation technologies can be brought to bear upon
urban transportation networks to produce a dynamic traffic management system. A
schematic diagram of such a system is illustrated in Figure 1.2 (OECD, 1987). In the
Information Loop, an advanced driver information system (ADIS) can be used to affect
the travel demand. Such information could consist of:
1) Trip planning advice: travelers’ routes, mode of travel, and departure
times may be influenced if they can be given advance warning of current
Tegape 4
and anticipated conditions on the road network;
(ii) Route following advice: the actual routes followed by drivers may be
influenced to avoid congestion sites and improve travel times as well as
to reduce wasted travel that is caused by inefficient route choice.
In the Control Loop, direct traffic control is employed to affect network supply
by influencing the link capacities. Control categories include: traffic signal systems,
ramp metering, priority/reversible lane assignments, etc. Advanced traffic management
systems (ATMS) may eventually be implemented as elements of this loop. Together, the
Information and Control Loops, through a dynamic interaction can improve the
performance of the transportation network. (A third loop is shown in Figure 1.2, which
combines the interactions of demand and supply in the case of public transportation and
fleet management systems.) This, in principle, represents the structure of an advanced
traffic management system combined with an advanced driver information system
(ATMS-ADIS).
1.3 Objectives of this Dissertation
ADIS are expected to provide the facility for generation and dissemination of
driver information that can be used for real-time diversion of traffic. Guidance provided
to drivers has to possess a number of desirable properties if it is to be useful for dynamic
traffic management and if it is to be adopted by drivers.
To be useful for dynamic traffic management, guidance has to be able to respond
in real-time to emerging traffic conditions and anticipated demand levels. The real-time
response has to be devised in such a way that it reduces the overall extent of congestion.
Finally, since driver information may prove to be counterproductive in some cases,
guidance should be provided in such a way that it reduces the potential adverse impacts
of improved information.
From the drivers’ point of view, guidance should be useful: it should convey to
the driver the traffic information he is most interested in obtaining. In addition, guidance
should be valid in the sense that, for a given departure time, the travel time a driver
experiences on the route he was guided to should be shorter than that on alternate routes.
Moreover, guidance has to improve the reliability of travel times experienced by drivers
by reducing variability and eliminating long travel times.
If drivers receive guidance which is not characterized by these desirable properties,
they will not have confidence in the information system. As a result, drivers may not
comply with the guidance being provided, in which case the dynamic traffic management
potential of the information system would not materialize.
2282 15
The achievement of all these properties represents a significant challenge to
developers of ADIS for several reasons. First, the implementation of route guidance
necessitates a systems-based treatment of the guidance problem including the
identification of modular functions that have to be developed and the options available
for each, as well as the isolation of the many parameters that enter into the problem.
Moreover, modelling tools required to implement the modular functions referred to above
have to be specified and developed.
The aim of this dissertation is to propose an ADIS implementation scheme that
will be capable of achieving the above-mentioned properties and subsequently to conduct
prototypical evaluation analyses of the value of the proposed scheme. As such, the
objectives of the dissertation may be stated as follows:
i. To formulate a framework for ADIS implementation
A dynamic network modeling approach is critical to the effectiveness of real-time
driver information systems. Such a modeling approach is needed to assess accurately
network performance as well as to forecast traffic conditions that may exist in the near
future in order to develop real-time diversion strategies to alleviate both recurring and
non-recurring congestion conditions.
A proper framework for ADIS implementation should represent an operational
system that provides all the desirable properties referred to above. Such a framework is
proposed in this dissertation with the major components being the following:
» a surveillance system which collects relevant traffic data
« a Control And Routing module (CAR) which determines the guidance advice
» a COngestion Prediction module (COP) which provides CAR with congestion
information
| This framework is described in detail in Chapter 3 and is designed to achieve the
desirable properties outlined above. The principles embodied in this framework are:
Principle 1: Need for Projection of Traffic Conditions
A dependable look-ahead capability is central to the success of guidance systems
since it permits the advice to be based on traffic conditions that are most likely to
materialize in the future thus being more responsive to such conditions and providing
more opportunities for reducing overall congestion levels. Guidance with such a
capability is likely to be more useful to drivers since it attempts to anticipate future traffic
conditions that they will actually experience. Moreover, guidance based on projected
traffic conditions has the best potential for improving the travel time reliability of drivers.
In contrast, guidance based solely on the use of historical data or available measurements
from surveillance equipment is not able to anticipate future variations in demand levels
-—
Cadea
or the buildup of congestion at specific locations in the network.
Principle 2: Nature of Projection Methodology
There exists a need for an advanced COP module that provides accurate
projections of future traffic conditions. Such an advanced module would make use of
current traffic conditions, predicted demand levels, guidance being provided, and
anticipated driver response to guidance in making its predictions of future congestion in
the network. An advanced congestion prediction capability results in guidance which is
more responsive to (a) future shifts in demand patterns, (b) upcoming variations in
network capacity due to traffic control actions, and (c) the potential adjustments in driver
travel decisions. Therefore, the system would be in a better position to guide drivers to
routes that will be less heavily congested when they use them.
However, predicting congestion in a dynamic network is a non-trivial undertaking.
The advanced congestion prediction scheme that we propose to use consists of a Dynamic
Traffic Assignment Model (DTA) that requires projections of the temporally changing
demand levels and the detection of changes in network capacities due to incidents, for
example. The DTA also involves modelling of driver behavioral rules and a forecast of
the travel decisions drivers may make when faced with the predicted traffic conditions.
Such modelling of driver behavior, including reaction to information provision, is
expected to reduce the adverse effects associated with improved information.
The DTA differs significantly from what is currently being used in demonstration
projects as a basis for the guidance being provided to motorists. In such demonstration
projects, the guidance is either based on current traffic conditions or on traffic conditions
that are predicted based on some heuristic methodologies. In none of these projects or
analyses being conducted by researchers has there been consideration of the impact of
guidance on motorists when setting the guidance.
Principle 3: Consistency Requirements
Based on principle 2, consistency has to be maintained between the guidance being
provided to drivers and the predicted traffic conditions. That is, the information system
has to check that the guidance being provided will prove to be optimal to guided drivers
based on a prediction of the future traffic conditions. This should prove to be quite
helpful in maintaining high levels of guidance validity.
ii. To analyze the modelling needs of the constituent elements of such a framework
A second objective of this thesis is to analyze the modelling needs of the
constituent elements of the framework, namely, the O-D updating, COP, and CAR
modules. Most of the emphasis is placed on the proposed COP module (the DTA
referred to above) which would require modelling dynamic network performance as well
Mg 17
as dynamic driver behavior. Identifying and analyzing these modelling needs is thought
to be a significant issue as it clarifies whether the state of the art in these fields is
prepared to support ADIS or still requires extensive research and development.
iii. To conduct prototypical evaluation analyses
The final objective of the dissertation is to analyze the value of the proposed
framework vis-a-vis other information provision scenarios and under various conditions.
Specifically, a case study will be conducted that will:
» compare the proposed framework with other information provision scenarios
» study the impact of various parameters on guidance effectiveness
This objective will be accomplished using a simulation model in conjunction with
an experimental design that will focus on the issues and parameters that are likely to
determine the effectiveness of ADIS.
1.4 Summary of Major Findings
This section describes the major findings of the research reported in this
dissertation. It is subdivided into three major subsections that relate to the framework,
component models, and simulation results.
1.4.1 ADIS Implementation Framework
It was concluded that the actual benefits that may be realized from driver
information systems depend heavily on the quality of guidance that is being provided.
The possible occurrence of adverse impacts from improved information was demonstrated.
This suggests the need for an ADIS implementation framework that is designed to provide
drivers with guidance characterized by a set of desirable properties.
In order to insure the effectiveness of real-time driver information systems, the
ADIS implementation framework has to be based on a dynamic network modeling
approach. Such a modeling approach is needed to assess accurately network performance
as well as to forecast traffic conditions that may exist in the near future in order to
develop real-time diversion strategies to alleviate both recurring and non-recurring
congestion conditions. A framework that would satisfy these requirements is adopted in
this dissertation and consists of a surveillance system which collects relevant traffic data,
a congestion prediction module which provides the guidance system with congestion
information, and a control and routing module which determines the guidance advice.
Page 1
1.4.2 Modelling Needs of Framework Components
Chapters 4 and 5 of this dissertation analyze the modelling requirements associated
with the implementation of a DTA as the COP module. The analysis focusses on the
dynamic network performance and dynamic driver behavior elements of the DTA. The
basic conclusion that was reached is that the state of the art in these two fields is not yet
prepared to support ADIS functions and that significant research and development is still
required. Specifically, it was found that existing dynamic traffic assignment models
suffer from several shortcomings, including not-truly dynamic assignments or the
existence of inconsistencies in network performance representation. Moreover, there is
a significant need for research on developing theoretically sound methodologies for
dynamic O-D estimation.
Moreover, significant difficulty lies in the dynamic driver behavior modelling
aspect of the problem. In this respect, models in the literature fail to consider the fact
that in an ADIS context drivers continually update their perceptions of travel
characteristics and revise their travel choices. When a Driver Information System is
available, a central requirement of the DTA is that it take into account the impact of any
guidance information on driver decisions when projecting future traffic conditions. In
such a case, a driver’s pre-trip choices and en-route diversion decisions will not depend
solely on his own experience with, and observation of, traffic conditions, but will also be
affected by information being provided concerning downstream traffic conditions on the
pre-planned route as well as on alternate routes. Thus, the DTA has to include models
of pre-trip and en-route driver behavior which are sensitive to the drivers’ characteristics
and their access to traffic information and route guidance.
Modelling- driver behavior in such a dynamic context requires significant
enhancements to existing models. It was indicated that psychometric data may be of help
in explaining some latent factors that enter into such dynamic behavioral decisions.
Finally, the data required to estimate such models has to be collected in a manner that
differs from traditional data collection efforts.
1.4.3 Case Study Results
Based on the case study, the following conclusions were reached:
Proposed Framework
Guidance based on projected traffic conditions offers advantages over other
guidance scenarios under the following conditions:
» when O-D demand levels are not highly variable form day-to-day
» when good predictors of daily O-D demand levels are available
» with the existence of fast incident detection capabilities
Nage 10
» when network topology does not provide "cheap" opportunities for re-routing
Even with no real-time updating of the O-D demands, and relying solely on the
average historical levels, projective guidance is still to be preferred over other information
provision scenarios if O-D demand patterns display significant peaking occurring at
different times of the analysis period, an effect which may not be captured by other
guidance scenarios. Moreover, even with such absence of real-time O-D updating
capabilities, projective guidance offers better results in terms of reductions in maximum
delay experience by guided vehicles.
[he occurrence of adverse effects from guidance at high guided fractions due to
overreaction was eliminated when the following actions were adopted:
» guidance is based on projected traffic conditions (Principle 1)
» a DTA is used for congestion prediction (Principle 2)
» an advanced CAR logic (consistency check) is resorted to (Principle 3)
3 the CAR’s temporal update frequency is increased ensuring that only a small
number of drivers receive the same guidance advice
These actions resulted in average delays that did not increase at high guided
probabilities and in significant improvements in guidance validity.
General
Guidance (in general) is needed most when the traffic conditions are highly
stochastic, as when accidents occur frequently. In such cases, guidance offers a more
significant advantage by guiding drivers away from locations that are highly congested
as a result of accidents. This also means that guidance is more effective in dealing with
incident (non-recurrent) congestion than with recurrent congestion when the network is
congested.
The benefits obtained from guidance are sensitive to the level of market
penetration. The maximum reductions in average delay that were obtained by using
guidance were modest and occurred at market penetration rates that are less than 50%.
However, some of the other performance measures indicated that guidance does offer
other more significant advantages. Specifically, the reductions in maximum delays were
found to be more significant and occurred for the full range of guided probabilities.
Unguided vehicles were also found to receive some benefits from guidance. Such
benefits, however, were smaller than those experienced by guided vehicles under the same
conditions.
Tradeoffs
—- 3; eoid
the overall framework for implementation of real-time driver information systems. For
instance, a clear tradeoff emerged between the spatial and temporal update frequencies
associated with the CAR element on the one hand and the need for an advanced COP
module on the other hand. Moreover, there was another tradeoff between the use of an
advanced CAR logic or an increased temporal update frequency for the purpose of
controlling overreaction. These issues are to be considered carefully, together with their
implementation implications (cost, hardware, and computational requirements) when
deciding on the implementation most suitable for specific situations.
1.5 Dissertation Outline
This dissertation consists of 8 chapters.
Chapter 2 provides an overview of Driver Information Systems and describes the potential
benefits and adverse impacts of improved information.
Chapter 3 presents in detail the proposed framework, including the possible COP and
CAR schemes. It also analyzes potential ways of reducing the adverse impacts of
improved information, and the consistency principle embodied in the proposed framework.
Chapter 4 concentrates on the proposed COP scheme, the DTA. It reviews existing DTA
procedures, discusses dynamic network performance issues, and proposes guidelines for
the DTA necessary as a part of the proposed framework. In addition, some O-D updating
methodologies are presented.
Chapter 5 analyzes in detail the modelling requirements for dynamic driver behavior
models required as part of the DTA. It presents several ideas concerning imperfect driver
information, perception revision, and travel adjustment in response to acquired
information. This chapter also discusses the role that psychometric data on driver
behavior may play in modeling dynamic driver behavior and provides a design for a data
collection effort that would be needed to assess and model such behavior.
Chapter 6 describes the design of the case study. The experimental factors, elements of
case study, measures of effectiveness, and structure of the simulation are presented.
Chapter 7 provides the input data and results associated with the case study. These
results are analyzed and specific conclusions are drawn.
Chapter 8 provides some concluding remarks and suggests directions for further research.
Page 21
Table 1.1 Traffic Operations Control and Management Variables’
Traffic Signalization
Cycle time
Greentime splits
Offsets (Throughbands)
Signal phasing
Phase sequencing
Variable-message signs (route diversion)
Traffic Operations Improvements
Channelization of traffic
One-way streets
Metering access to freeways
Reversible (tidal) traffic lanes
Other traffic engineering improvements
Preferential Treatment
Reserved or preferential lanes on freeways and city streets
Exclusive lanes to bypass congested points
Conversion of selected downtown streets to exclusive bus use
Exclusive access ramps to freeways
Bus preemption of traffic signals
Special turning lanes
Truck routes
Parking controls
Source: Gartner et al., 1980.
Page 22
figure 1.1 Transport System Interactions
~~ Socio0-
Demand [ DEMAND/ Capacity TRANSPORT
ECONOMIC SUPPLY SYSTEM
SYSTEM INTERACTION
User TSH
Choices Strategies
USER TRANSPORT |
DECISION SYSTEM
PROCESS MANAGEMENT
Observed
State
User
User Hsaan System
Objectives &) Objectives{
USER SYSTEM
ASSESSMENT >, ASSESSMENT
PROCESS PROCESS
ser Public
Preferences Policy
Page 23
Figure 1.2 Dynamic Traffic Management Overview
TRAFFIC
MANAGEMENT
CENTER
DIRECT
DRIVER TRAFFIC
INFORMATION CONTROL
—
Control
Information vJ
Loop Loop od
TRAVEL NETWORK
DEMAND SUPPLY
TRAFFIC
CONDITIONS
LL PUBLIC TRANSPORT «
AND FLEET
| MANAGEMENT
SYSTEMS
lage 24
CHAPTER 2 DRIVER INFORMATION SYSTEMS:
PRINCIPAL CHARACTERISTICS
AND POTENTIAL IMPACTS
Chapter Objectives and Layout
The objectives of this chapter are two-fold, namely:
» to present a general overview of driver information systems that includes a
summary of their most important characteristics and a description of major
European, Japanese, and U.S. systems and demonstration projects (section
2.1)
» to demonstrate the potential beneficial and adverse impacts that may be
experienced due to the implementation of such systems (section 2.2)
2.1 Overview of Driver Information Systems
2.1.1 System Characteristics
Classes of Driver Information and Guidance Systems
Recent technological developments offer new opportunities for road information
provision. Reviews of the types of driver information services and traffic management
systems which are currently in use or under development are presented in OECD (1987)
and US DOT (1989). The various types of possible driver information and guidance
systems may be classified according to their functions and capabilities as follows:
* Pre-trip route planning systems which provide drivers with information regarding
a particular route or the optimum path to their destinations at a specific time.
Such information is usually based on a historical or current data base of network
conditions and could be provided using a telephone or a remote terminal (at home,
for example);
» Autonomous navigation aids which help the driver locate his position on the
network. Such aids do not need a communication link and may be implemented
by on-board map display systems;
» Local or area broadcasting systems and roadside variable message signs used to
communicate traffic information to drivers over a limited or wide area; and.
» Automatic route guidance systems which provide drivers with updated guidance
at equipped junctions to direct them to the shortest routes to their destinations;
Page 25
such systems, however, require extensive roadside infrastructure in addition to the
in-vehicle units. When two-way communication is established between equipped
vehicles and the control center, the vehicles themselves would provide the control
center with useful information concerning prevailing traffic conditions and travel
times.
The last class of driver information systems is the only one capable of providing
drivers with real-time "customized" guidance and is better equipped to distribute traffic
more evenly throughout the network. The addition of two-way communication between
equipped vehicles and the control center would facilitate dynamic traffic management and
continuous monitoring of the effects of driver information provision and control actions.
It would also facilitate the revision and modification of such information provision and
control actions in order to utilize the road network in a more efficient manner.
In this dissertation we focus on the third and fourth classes of guidance systems
which are capable of providing drivers with en-route information, thus realizing a form
of guidance that is more adaptive in its response to emerging traffic conditions compared
to the first two classes.
Approaches to Driver Information Provision
In the case of the more advanced motorist information systems being developed
now, two different types of approaches are being followed (in the US, Japan and Western
Europe). In the first approach, drivers receive information from the system about traffic
conditions which they use to figure their best routes. In the second approach, the system
ascertains the vehicle’s present location and desired destination and provides specific
route directives to the driver. In one version of this approach, the system informs the
driver about the route with the shortest travel time under the prevailing network
conditions.
Temporal Nature of Information
[Information provided to drivers may conceptually fall into one of three categories:
Historical Information - information which describes the state of the
transportation system during previous time periods.
Current Information - the most up-to-date information about current traffic
conditions.
Predictive Information - information concerning expected traffic conditions
during subsequent time periods when travel can occur.
Page 26
2.1.2 Existing Pilot Projects
In what follows we present a review of some of the existing ADIS systems and
projects in Europe, Japan, and the U.S. We will focus on describing the major
characteristics of such systems as well as the way guidance is set and provided (CAR
module) and the specific congestion information it is based on (COP module). The
review is based on papers by French (1990), Catling and McQueen (1991) and Rillings
and Betsold (1991).
European Programs
ALI-SCOUT and AUTOGUIDE are two European IVHS projects associated with
the program EUREKA. ALI-SCOUT is a route guidance system developed in Germany
which uses infrared transmitters and receivers to transfer navigation information between
roadside beacons and on-board displays in equipped vehicles. Early versions of this
system were tested along a 60-mile stretch of the German autobahn in the North Rhine-
Westphalia region. The more advanced ALI-SCOUT system is being tested in Berlin
under the name LISB where beacons installed at 250 selected intersections and at other
key locations transmit location, map, and shortest path routing "trees" to the 700 equipped
vehicles as they pass the beacons. The central control system provides new
recommendations every 5 minutes (Hoffman, 1989). The in-vehicle units select from the
tree the route associated with the destination entered by the driver and subsequently issue
real-time guidance instructions using simple graphics in conjunction with audio messages.
The route guidance is based on predicted journey times downstream. Inputs to such
predictions are provided by information flowing from the vehicles to the central control
system via road-side beacons. At every beacon vehicles transmit their journey times on
just completed network sections. These journey times are compared in the control center
to standard (historical) journey time patterns to come up with updated predictions of
journey times. More details concerning the prediction methodology used in LISB are
presented in Chapter 3.
AUTOGUIDE is another route guidance system developed in England based on
the ALI-SCOUT technology. A pilot test was undertaken in the Westminster section of
London and in a corridor between London and Heathrow Airport. This pilot test was
followed by a large scale pilot installation which is now underway in London in which
1000 vehicles are expected to be equipped, and between 200 and 300 beacons installed.
The system is expected to be available by 1993.
The current AUTOGUIDE system in London provides route recommendations
based on real-time information of area-wide traffic conditions. As is the case with the
Berlin system, traffic information which is needed for route guidance is collected
automatically and continuously by the system itself. The time taken to travel between
adjacent roadside units is known once a guided vehicle completes a section of its trip.
This information is sent at regular intervals to the central computer system which uses it
= LE
-§y 0
1.a ]
-
to update a database containing typical ranges of journey times by time of day and day
of week for each section of the network. This traffic database is used to predict the most
likely journey times that will occur in the near future when vehicles reach subsequent
portions of their journeys.
In both demonstration projects described above, the routing strategy (CAR module)
consists of providing drivers with route directives that guide them to the shortest route
to their destination. Such route directives are based on projections of network travel
times based on a heuristic methodology (COP module).
Japanese Programs
In the CACS pilot experiment carried out in Japan between 1977 and 1979, guided
vehicles were provided with shortest travel time paths calculated on the basis of predicted
route costs. The predictions were based on a heuristic-statistical methodology and were
updated every 15 minutes.
Japan now has two major IVHS development programs: Advanced Mobile Traffic
Information and Communication System (AMTICS) and Road-Automotive
Communication System (RACS). AMTICS is oriented towards surface street operations
while RACS is more concerned with expressways. The AMTICS system uses roadside
beacons to transmit real-time traffic information from a traffic control center to an in-
vehicle navigation system. AMTICS does not suggest any specific route for the driver
to follow. Another component of the system uses in-vehicle compact disc systems (CD-
ROM) and in-vehicle video display terminals to display road maps, local traffic
regulations, the location of parking lots, hospitals, gas stations, tourist information, etc.
The system has been demonstrated on a small number of vehicles.
The RACS system is composed of roadside communication units, car units, and
a control center. The available system functions include navigation (autonomous dead
reckoning by the car unit only) and roadside information (by means of communication
between the roadside units and the car units). The system is designed for exchange of
information between the control center and individual automobiles, making it possible to
monitor the location of particular cars as well as the collection of detailed road traffic
data. Three tests of the system were conducted between 1987 and 1989 with the latest
involving the three RACS system functions of on-board navigation, real time traffic
information, and individual communication.
In 1990 it was announced that the two systems would be integrated under the
name VICS, a process that would provide for an integrated traffic information system for
both freeways and surface street systems.
Guidance provided by both AMTICS and RACS is based on current measurements
of traffic volume and speed , and does not involve any prediction of future traffic
Na
Tek - IR
conditions (COP module). Moreover, drivers are not guided to any specific route but are
rather provided with the real-time traffic information directly through their in-vehicle units
(CAR module). Future developments in these systems envision the implementation of
dynamic on-board navigation whereby shortest route directives are determined by the in-
vehicle computer based on the available real-time traffic information.
U.S. Programs
In the U.S., a number of pilot projects have been implemented while others are
in the development stage. The Program on Advanced Technology for the Highway
(PATH) which was established in 1986 by the University of California and the California
Department of Transportation (CALTRANS) focuses on the development of automobile
automation and navigation technologies with less emphasis being placed on driver
information and traffic management systems.
Moreover, the FHWA along with CALTRANS and General Motors are sponsoring
a field evaluation of in-vehicle urban freeway navigation and information systems. The
site for this evaluation is the Santa Monica Freeway Corridor in the Los Angeles area
(Smart Corridor). Sensors on the freeways and arterial streets in this corridor provide
information on traffic conditions which is used to make control decisions and provide up-
to-date congestion information to drivers in 25 specially equipped vehicles. Guidance
provided to drivers is based on current measures of traffic density and vehicle speeds
(COP module) and consists of congestion information only with no route directives being
provided (CAR module).
Travtek is a cooperative project involving AAA, GM, the FHWA, and the Florida
DOT to demonstrate motorist information systems which reflect actual traffic conditions
in the Orlando, Florida area during 1991-93. One hundred cars equipped with navigation
and driver information systems linked to traffic centers by radio data communication will
participate in the demonstration. After a driver selects his destination, the system will
determine the best route and use graphic displays as well as audio to provide instructions.
In addition, the in-vehicle units can display maps of the Orlando area (including service
information) and information regarding traffic incidents and traffic congestion locations.
In the case of Travtek route directives will be provided (CAR module) based on current
traffic conditions (COP module).
A large demonstration project is planned for the Chicago area under the name
ADVANCE and will involve Motorola, the Illinois DOT, and the Illinois Universities
Transportation Research Consortium. The stated objectives of the project are to examine
potential approaches to driver information provision, acquisition of dynamic traffic
information, and the use of forecasting to relieve traffic congestion. Almost 4,000
vehicles are expected to be equipped with in-vehicle units connected to a traffic
management center through two-way radio communication. Real-time, traffic-dependent
guidance is expected to be provided to drivers.
Page 29
2.2 The Impact of Driver Information Systems
2.2.1 Potential Benefits of Improved Information
Travel Choices and Driver Information
When making travel choices, drivers constantly combine various sources of
information to form perceptions and expectations of traffic conditions. Conventional
sources of information available to drivers include personal experience, word of mouth,
and media messages. Drivers who rely solely on such information are likely to have a
partial and inaccurate knowledge of traffic conditions on the network and consequently
may follow sub-optimal routes.
Reasons for Sub-Optimal Travel Choice
For any trip, a departure from optimum routing may occur for any of the following
reasons (King, 1987):
» Use of route selection criteria that do not lead to an optimum route
Inadequate trip planning skills
 >» Information necessary for optimum trip planning is unavailable, inadequate,
inaccurate or out of date
Route following (i.e., trip plan implementation) skills are inadequate
Inadequacy in the highway information system
Inefficiencies Resulting from Sub-Optimal Travel Choice
It has been estimated that excess travel, due to sub-optimal route selection amounts
to 6.4% of all United States travel by non-commercial vehicles (King, 1987). The cited
study addressed only steady state route selection and did not estimate the adverse
consequences of the failure of adjusting pre-selected routes in response to changes in
highway and/or traffic conditions. Moreover, an analysis of a sample of trips in the U.K.
found that the average inefficiency of trips taken over unfamiliar roads is about 20-25%,
a figure that translates to about a 2% potential reduction in overall mileage (Jeffery,
1988).
The Potential Role of Driver Information Systems
in Reducing Delays and Wasted Mileage
It is widely accepted among the developers of route guidance systems that such
driver information systems offer potential for reducing delays and wasted mileage (see,
for example, Boyce 1988a). The basic rationale behind this belief is that many drivers
possess little or no reliable information concerning travel and route alternatives in general
Page 30
and may be uninformed of road conditions on any specific day. Such unawareness leads
to misperceptions on the part of the drivers as to the relative desirability of alternative
travel decisions. Driver information systems thus have the potential to reduce or
eliminate poor route choices and consequently excess travel distance and time incurred
by unaware or uninformed drivers.
Empirical and theoretical results reached by different researchers in different
countries, under different assumptions, and for different driver information system designs
seem to support this belief. An experiment to establish the level of benefit that could be
achieved by implementing the traffic responsive AUTOGUIDE system in London was
also reported by Jeffery (1988). This experiment involved driving vehicles at different
times of the day over a number of alternative routes used by drivers in London to
establish typical journey times on each and to identify the routes which would have been
indicated by AUTOGUIDE to be the best at that time. The results indicated that
AUTOGUIDE should, on average, reduce the time a driver spends on a main road journey
by between 9% and 11%. However, this result is based on the assumption that the
system can perfectly predict future traffic congestion. Moreover, Smith (1989) presented
the results of a study by the UK Transport and Road Research Laboratory entitled
"Estimates of AUTOGUIDE Traffic Effects in London". In this study both "guided" and
"unguided" vehicles are assigned to routes on a highway network. However, unguided
vehicles are assumed to base route choice on average travel demand and corresponding
network performance whereas guided vehicles are assigned to the network based on
"actual" conditions which are simulated for each day by introducing a random
perturbation to the trip matrix. The fraction of guided vehicles was varied from 10% to
100%. The results of this simulated experiment indicated tangible improvements in
distances, travel times, and speed on the network for the whole range of fraction of
guided vehicles. For instance, total travel time on the network was reduced by 2.5% with
10% of the vehicles being guided whereas the reduction was up to 6.2% with all vehicles
receiving route guidance. This experiment also indicated that while guided vehicles
receive higher benefits, unequipped vehicles are better off too. Finally, the results of the
CACS route guidance experiment in Tokyo indicated savings in travel time to equipped
users of the system ranging from 9 to 14% (Tsuji et al., 1985) with a small penetration
rate.
A recent study by Koutsopoulos and Lotan (1990) focused on evaluating the
expected effectiveness of motorist information systems in reducing recurrent traffic
congestion. The modelling approach used in this study consists of stochastic traffic
assignment in which users are assumed to have different perceptions of travel time costs
on which they base their travel decisions. When traffic information is provided to
motorists, it is assumed that their perceptions improve and become more concentrated
around the true travel times. The level of information provided to motorists was
incorporated into the model by different values of the coefficient of variation of perceived
travel time. The model was applied to a relatively small network (and subsequently to
the Boston network) and results indicated modest reductions in travel times (about 4.4%)
Yor
. ry Ja
could be achieved by the introduction of motorist information systems.
Fricker and Tsay (1984) used traffic assignment models and simulation techniques
to evaluate the effect of advance notice to drivers of a link blockage in a freeway
network. The study found that if 20 percent of drivers diverted, in response to a traffic
advisory, a queue would clear 32 minutes earlier and diverted traffic delay would be
reduced by 79 percent.
Similar results were demonstrated in studies conducted in connection with the
IMIS project (Gartner and Reiss, 1987). Average delays for various traffic incident
scenarios were reduced by 16.2% (with some as high as 75%), and the average time
required to clear congestion after an incident was reduced by 56.4%. These
improvements were accomplished using only quasi-dynamic routing and control strategies.
Thus, there is considerable potential in the application of real-time driver information
systems.
Finally, Koutsopoulos and Yablonski (1991) present the results of a study of
design parameters of ADIS for the case of incident congestion and small market
penetration. The study indicated that while benefits to equipped vehicles due to the
availability of ADIS were small with respect to average travel time reductions, the
improvements in travel time reliability were significant. Reductions of up to 80% in the
standard deviation of travel time were observed.
The driver information systems also offer a variety of intangible benefits such as
the satisfaction from the greater reassurance of selecting a good route and from being
better informed about traffic incidents and the subsequent reduction in uncertainty
concerning overall traffic conditions. Safety benefits may also result from a decrease in
total distance travelled on the road network, as well as from advance warnings as to
inclement weather conditions or incidents ahead on the road.
2.2.2 Potential Adverse Impacts of Improved Information
In considering the potential benefits of alternative driver information systems it
is also essential to evaluate the potential adverse effects that improved information may
bring. Three behavioral phenomena that may negate some of the beneficial effects of
improved information are discussed below.
Oversaturation - Consider a situation in which a small number of drivers receive route
guidance information. The information corresponds to various signals which reduce or
eliminate uncertainty. When a driver is able to use this information efficiently he or she
is better off. However, a driver may be unable to process this information to select the
optimal route or may be distracted by the large amount of available information. In this
case we say that oversaturation occurs.
Page 32
Oversaturation is mainly a man-machine interaction problem. Lunenfeld (1989)
considers oversaturation as one of the human factor aspects of motorist information
systems and points out that whether drivers can use the information effectively cannot be
fully determined without operational experience.
Takasaki et al. (1989) evaluate reasons for oversaturation using the human
information processing approach developed by Rasmussen (1986). The authors
hypothesize that a driver’s decision strategy depends not only on the information sources
that are available in the driving environment, but also on his or her cognitive expectations
(knowledge and beliefs). Rasmussen (1986) indicates that an effective process of
information pickup is only possible if the individual possesses an internal representation
of the choice situation he is facing and if he is capable of dynamically revising this
representation. When the driver does not possess this capability he or she is unable to
use the information efficiently and oversaturation occurs.
Sheridan (1990) suggests the use of "mental workload" models to analyze how
much attention the combination of ADIS and otherwise conventional driving demands
may require. He warns that the additional tasks of getting and using information from
ADIS while driving may add sufficient workload that makes the overall workload level
unsafe.
Overreaction - Consider a situation in which a substantial fraction of drivers receive
information on traffic conditions. Overreaction occurs when drivers’ reactions to traffic
information cause congestion to transfer from one road to another. It may also generate
oscillations in road usage. Overreaction may happen if too many drivers respond to
information on traffic conditions.
Several instances of this type of overreaction have been reported. For example,
a French radio service which advises drivers going on holiday how to plan their itinerary
and optimal day for departure has experienced overreaction in past years. Another
instance is provided by Fujii (1986), discussed in Boyce (1988b): "If we use a simple
scheme such as shortest-time path guidance (when most of the vehicles are under control
of the system), it might cause oscillation of traffic flow among alternative routes. In
order to avoid this situation, an advanced scheme such as route distributing guidance
should be prepared." Using a simulation model, Kobayashi (1979) found that oscillatory
behavior may occur when road guidance by shortest path route is extended to a large
proportion of drivers (90%).
Concentration - In general, a group of drivers traveling from an origin to a destination
tend to use different routes and departure times because they have heterogeneous
preferences and diverse perceptions of network conditions. Differences between perceived
and actual network conditions imply that drivers who are not well informed may select
alternatives which are not in their best interest. Information tends to reduce the variation
among drivers because it increases uniformity of perceptions of network conditions
I;TC 33
around the true values. As a result a greater number of drivers may select the best
alternatives (from their individual point of view) and consequently drivers with similar
preferences will tend to concentrate on the same routes during the same departure times.
Thus, more information could potentially generate higher levels of traffic congestion.
A theoretical example of the potentially adverse effects of concentration is given
by Arnott et al. (1988 and 1989). They analyze a situation with one or two routes
connecting one origin to one destination. When capacity and/or demand fluctuate, they
show that for some parameter values more information increases average user cost and
perfect information may be associated with a higher average cost than no information.
The effect of concentration is also reflected in the observation made by Kawashima
(1989), while commenting on the CACS pilot experiment in Tokyo, that the efficiency
of road guidance systems providing the shortest path tends to decrease as the number of
equipped vehicles increases. Similarly, Mahmassani and Herman (1988) have shown,
using a stated preference approach, that there is an optimal fraction of individuals who
should receive route guidance information.
It should be stressed that the nature of the concentration and overreaction
processes are very different. Concentration is intrinsic to any system and holds even with
perfectly rational drivers. Overreaction is a consequence of the fact that the drivers
and/or the driver information system are unable to forecast perfectly the use of
information by all drivers.
Page 34
CHAPTER 3 INTEGRATED FRAMEWORK
FOR THE IMPLEMENTATION OF ADIS
3.1 The Proposed Framework
The Need for an ADIS Implementation Framework
Based on the discussion presented in Chapter 2, the following observations can be
made:
» The actual benefits realized from driver information systems will depend heavily
on the quality of the traffic information provided to drivers (Bonsall 1989). For
instance, the occurrence of adverse impacts due to improved information may not
be avoided if the provision of driver information is not based on a well-designed
implementation framework.
» The unavailability of accurate traffic information may result in poor guidance
directives being provided to drivers. The likely impact would be a gradual loss
of confidence in the system and its eventual abandonment. For example, a recent
survey of users of the ALI-SCOUT demonstration project (Janko, 1991) found a
very low compliance rate among commuters receiving guidance who, after some
experimentation, came to believe that they could do better than the guidance being
provided by the system. The probable reason for this poor system performance
is the fact that guidance is being based on crude predictions of traffic conditions.
Therefore, another task of the ADIS implementation framework is to propose an
operational system that is able to provide drivers with guidance that they can have
confidence in.
These observations support the need for an ADIS implementation framework that
is designed to provide guidance which is characterized by the desirable properties
discussed in Chapter 1. This chapter describes the framework being proposed in this
dissertation to achieve the above-mentioned goals.
System Structure
A dynamic network modeling approach is critical to the effectiveness of real-time
driver information systems. Such a modeling approach is needed to assess accurately
network performance as well as to forecast traffic conditions that may exist in the near
future in order to develop real-time diversion strategies to alleviate both recurring and
non-recurring congestion conditions.
A proper framework for ADIS implementation should be able to integrate the
functional needs referred to above into an operational system. Figure 3.1 illustrates the
Page 35
system structure and information flow of the framework within which real-time ADIS
should be implemented.
Stated briefly, the functions performed by each element in the system are as
follows:
The Surveillance System consists of detection equipment deployed on the various
network elements (for example, detectors in the pavement, video cameras,
possibly other optical recording equipment). The surveillance system may also
include beacons which gather information concerning the identities of vehicles
passing selected points on the network over time. In addition, equipped vehicles
themselves may act as elements of the surveillance system by providing
information on travel times on specific sections of the network. The collected
data may consist of information on flows, speeds, travel times, the numbers of
queued and moving vehicles on each link, and incident detection. The actual
information gathered will vary from system to system depending on the particular
components included in the system and the coverage of the network by the
surveillance system. This information is collected continuously, rapidly processed
and made available to other elements of the real-time system.
The Congestion Prediction (COP) element has the responsibility for providing the
Control and Routing module with the information that is needed to implement
routing and guidance strategies. Among the principles adopted in this dissertation
are the need for COP to provide CAR with projected traffic conditions and the
fact that COP should be performed by a Dynamic Traffic Assignment Model
(DTA that will take driver response into account in its projections.
The Control and Routing (CAR) element generates guidance advice in response
to information provided by the surveillance system and by the congestion
prediction model. The fact that CAR should maintain projection/guidance
consistency constitutes another principle adopted in this dissertation.
The discussion that follows in this section and in section 3.2 describes the information
flow and model interactions embodied in the proposed framework. Section 3.3 presents
the principles behind the proposed framework while sections 3.4 and 3.5 describe the
CAR and COP, respectively.
Information Flow
The flow of information from one element in the system to another is briefly
described below (refer to Figure 3.1):
Infrastructure Data: Infrastructure data is an umbrella term including all network
attributes which are generally invariant with time or which change slowly with
ag. 6
time. This includes the network topology, the geometric attributes of all network
elements, the control devices that are installed in the network, any channelization
or other type of lane control, circulation restrictions such as one-way streets and
prohibited turns, etc. This information is required by the COP and CAR.
Historical O-D Data: The historical O-D data generally consists of origin-
destination information obtained by surveys and/or inferred from traffic counts
by assignment models. This historical information is, for the most part, slowly
changing over time, but should be updated periodically to reflect changes in the
infrastructure and in the characteristics of origins, destinations and the proportion
of drivers that have access to route guidance information.
The Surveillance System: The information provided by the surveillance system
consists mainly of direct measurements of volume, speeds, occupancy, and the
presence of incidents. Eventually the information may also include travel time
data from equipped vehicles. |
Updated O-D Data: The most recent information from the surveillance system and
route information (if available) can then be combined with historical origin-
destination data to provide updated 3-dimensional O-D matrices for the
subsequent time periods, as will be described in Chapter 4.
COP: Congestion prediction is then performed and provides estimates of traffic
conditions on the network. The updated O-D data is used by only a few COP
modules including the proposed DTA. Since route choice modelling and the
provision of guidance are sensitive to the destinations associated with flows, O-D
data is required to implement the DTA associated with the proposed framework.
Link flow data is not destination-specific and hence would not be sufficient for
the purposes of the DTA but may be useful for other COP schemes, as will be
discussed later in this chapter.
CAR: The traffic conditions identifies by the by the COP are then transmitted to
the CAR. The routing strategies need this information in order to develop an
optimal response to the developing traffic environment.
Guidance: The outputs of the CAR generally take the form of route guidance
information. This route guidance information can be displayed to the driving
public either externally on variable message signs or through in-vehicle radios or
displays for those vehicles with the appropriate equipment. The subsequent
behavior of drivers given the availability of such guidance determines the actual
traffic conditions that will be in effect during upcoming time intervals. In
addition, guidance data may also be transmitted to the COP as inputs to its
model.
[
rape oo ]
3.2 System Interactions
Computational Cycles and Projection Horizon
Of major importance in the operation of the system described above is the time
synchronization among the different elements. As figure 3.2 indicates, COP processing
is conducted as a sequence of cycles with a new execution cycle being initiated
periodically, at selected points in [real] time. For each such cycle COP will project
traffic conditions for some time, T,, into the future. During the execution cycle, the COP
transmits to CAR the projected "state" of the system. The purpose of these calculations
is to provide a basis for the CAR to compute routing strategies for the projection period
that are put into effect at the end of the execution cycle.
Role of Surveillance Data
As the COP projects future conditions, its computations are actually "ahead" of real
time. Information provided by the surveillance system is only referenced when the COP
and the real-world are synchronous in time, at points tg, t;, t,, ..., shown in Figure 3.2,
which are the beginning of new computational cycles. At these points in time,
‘refreshing” is required and the real-time software must be attuned to the real-world as
perceived by the surveillance system. COP will require surveillance data relating to
volumes, travel times, and the location and expected duration of incidents. By
periodically introducing the current state of the system, at times, t, as "initial conditions”
in the COP computational process, the COP will keep track of the real-world.
Illustration and Incident Handling
To illustrate the operation of the system, consider a projection horizon, T,, of 40
minutes. Suppose further that this 40 minute period is subdivided into eight 5 minute
"cycles". During the first cycle (starting at t) the COP model will predict traffic patterns
within the network over a 40 minute horizon. These values are the same as the ones that
were reported to have been used in the LISB experiment in Berlin (see Chapter 2)
whereby T, = 40 minutes and the guidance was updated every 5 minutes. However, it
is expected that the duration of the computational cycle will become shorter as cheaper
and/or more powerful computers become available.
Now consider the onset of an incident somewhere in the network between t, and t,,
as indicated in Figure 3.2. The exact time the incident occurs is not of great significance
and there will possibly be a lag between that time and the instant at which processed data
originating from various elements of the surveillance system indicate the presence of the
incident. Of significance is the fact that the next time that the projection process is
repeated, (i.e., at time, t,) the presence of the incident and its expected time of removal,
T., are assumed to be known as shown in figure 3.2. The COP model will then operate
Page 38
on a network environment which includes the presence of this incident from time t, until
time T,. Since, in the example provided in Figure 3.2, the projected horizon extends
beyond T,, the results of the calculations will reflect both the presence of the incident and
its removal. Thus, the route guidance policy that is put into effect by CAR over time
horizon, T,, will also reflect the presence and removal of the incident. In the absence of
any substantive information regarding the severity of the incident (such as from police
reports or video monitors, for example), it may be assumed that its duration will extend
at least until the end of the projection horizon (T, = T,). This last case of T, = T; is not
shown in figure 3.2.
Moving on to t, the entire computational process is repeated and all the new traffic
control and route guidance measures are taken into account by the COP. Note that it is
possible for a new updated estimate, T,, to be computed based on new information that
was received in the time interval {t,, t,}. Thus, the new calculation procedure initiated
at t2 will reflect the new guidance strategies and the change in the status of the incident.
Factors Affecting the Computational Cycle Length and the Projection Horizon T,
The computational cycle length and the projection horizon duration will depend on
two factors:
1. Computational Resources: As mentioned above, the computational cycle involves the
COP predicting traffic conditions and the CAR computing routing strategies for the
projection horizon T,. The ratio of T,/(t,,;-t;) (the ratio of real time to execution time) is
governed primarily by the computational resources available for COP and CAR to
execute. For example, if T, is set to be 40 minutes, the associated computational cycle
length t,t, required to execute COP and CAR may be 2 minutes given one computing
system while another system may demand 5S minutes. Obviously, a longer projection
horizon (longer foresight) and a shorter computational cycle (more frequent update of
guidance) are likely to improve the effectiveness of the routing strategies being
implemented, but the associated computational costs become more extensive.
2. The Rate of Change of Traffic Conditions: Returning to Figure 3.1, the flows of
information among the system elements should support the operation of the real-time
system. These flows of information should respond more frequently as congestion
develops in the traffic environment. This is due to the fact that the onset of congestion
creates a traffic environment that depends strongly upon the COP model for properly
estimating the associated consequences. Consequently, the number of cycles within each
projected time horizon may change on-line depending upon the status of the traffic
environment. Specifically, the cycles may be shorter during periods of heavy congestion
in order to keep up with the rapidly changing conditions. These factors may also
influence the duration of the projection horizon, T,. If the required computing resources
are available, however, shorter cycles may be used throughout the COP implementation
period. Such shorter cycles would enable more frequent updating of guidance which
2a: | — 33
reduces the potential occurrence of overreaction.
3.3 Principles Underlying the Proposed Framework
The major principles underlying the proposed framework were introduced in Chapter
I and are the following:
Principle 1: COP Should Provide CAR with Projected Traffic Conditions
The basis for routing strategies associated with ADIS may consist of either historical,
current, or predicted traffic conditions. While the use of historical data may provide a
basis for static guidance, its use alone as a basis for CAR decisions is not expected to be
of any value for adaptive routing. The main reason behind this is that in order to provide
adaptive routing decisions, the guidance system needs reliable information concerning
evolving, day-specific traffic conditions. Historical traffic data is a bad indicator of such
conditions, especially in situations where traffic patterns display a significant amount of
day-to-day variability. An analysis of the performance of various real-time routing
strategies by Koutsopoulos and Xu (1992) confirmed this intuition and indicated that the
use of historical data as a basis for real-time routing advice is significantly inferior to the
use of current or predictive information, for example. Since this dissertation is concerned
with adaptive routing decisions, the use of historical data alone is not considered a viable
option and is not analyzed in any detail.
In many existing demonstration projects, guidance passed to drivers consists of
information regarding current traffic conditions (Smart Corridor, AMTICS, and RACS;
see Chapter 2). Moreover, a study conducted by Mahmassani and Jayakrishnan (1989)
to evaluate the impact of guidance on system performance assumed that drivers will have
access to information regarding current travel times to their destinations. In addition,
some researchers assert that routing strategies may be formulated based on a control-
theoretic approach that requires information on current traffic conditions only. For
example, Papageorgiou and Messmer (1991) use feedback control methods to split traffic
between an O-D pair among different routes. They claim that their methodology has low
sensitivity with respect to unknown future demand levels and compliance rates which are
assumed to be exogenous "disturbances". However, the authors do warn that their
feedback concept which is based on observations of current traffic conditions may not
achieve its goal of establishing dynamic user optimum conditions if strong oscillations in
the demand levels occur or if the network performance displays strong nonlinearities in
case of severe congestion. These remarks by the authors provide further evidence that
using current traffic conditions as a basis for guidance will not succeed if current traffic
conditions are not good predictors of future conditions.
One of the major principles embodied in the framework being proposed in this
dissertation is that such routing strategies have to be formulated based on a forecasting
Pag. 20
or projection of future traffic conditions on the network (Ben-Akiva et al., 1991) rather
than instantaneous traffic conditions. The rationale behind using predictive information
is that drivers’ travel decisions are affected by future traffic conditions expected to be in
effect when they reach downstream sections of the network on their way to their
destinations. The LISB and AUTOGUIDE systems use projected travel times in setting
the guidance based on a similar rationale. Therefore, the most useful type of guidance
that can be provided to a driver faced with travel decisions would be based on a
projection of traffic conditions. In addition, guidance based on traffic information that
is predicted using an advanced COP module (as in the framework proposed in this
dissertation and based on the availability of good predictors of time-dependent O-D
demands) is most capable of improving the travel time reliability of drivers since it can
help them avoid long future delays based on its look-ahead capability. This issue is
discussed in more detail next under the second principle.
Principle 2: A Dynamic Traffic Assignment Model (DTA) Should Be Used for COP
In all existing driver information systems and demonstration projects, the guidance
passed to drivers is either based on current traffic conditions (Smart Corridor, Travtek,
AMTICS, and RACS; see Chapter 2) or on simple predictions of future traffic conditions
(LISB and AUTOGUIDE; see Chapter 2).
The travel time prediction methodology utilized in the LISB experiment, for example,
constructs a projection ratio of the historical travel time on a specific link to the current
travel time, as reported by equipped vehicles (Hoffman and Janko, 1990). This ratio is
used to predict travel times for vehicles using that link during all future time intervals.
Since only a few vehicles are equipped in Berlin, many links are not used by equipped
vehicles during particular time intervals, and therefore no estimate of the projection ratio
would be available. Therefore, in order to improve the estimate of this ratio, it is
modified to reflect current conditions on neighboring links as well as conditions in
preceding time intervals. Koutsopoulos and Xu (1992) note that a major problem with
this methodology is the fact that the projection ratio is used to predict travel times for all
future time intervals, thus implicitly assuming that it remains constant for the entire
prediction interval. To remedy this particular problem, Koutsopoulos and Xu suggest the
use of information discounting. However, the methodology remains heuristic in nature
and suffers from other omissions which are discussed next.
In all these systems and projects, as well as in all analyses being conducted by
researchers related to ADIS (such as Mahmassani and Jayakrishnan 1989; Papageorgiou
and Messmer, 1991; Mauro, 1991), there has been no consideration whatsoever of the
response of motorists to route guidance in setting such guidance. Such an omission
entails a major shortcoming in that the potential concentration of traffic on the
recommended routes and the overreaction of drivers in their response to guidance
information is ignored. This problem is expected to become more severe as the number
of guided vehicles increases. Guidance validity, and as a result driver compliance, would
29¢2 4.
be adversely affected in such schemes.
To overcome this problem would require the guidance to be based on an advanced
COP module that makes its predictions of future congestion in the network based on:
» current traffic conditions (consideration of initial conditions)
» predicted O-D demand levels (sensitivity to future demand patterns)
» guidance being provided and anticipated driver response to guidance
» traffic control actions to be implemented
» reduction in capacity due to incidents that have been detected
We propose that such COP capabilities may be achieved through a Dynamic Traffic
Assignment module, or DTA, that would differ significantly from the methodologies
currently being used in demonstration projects. The DTA would incorporate a dynamic
driver behavior module and a dynamic network performance module. The two modules
together would predict the driver response to guidance information and take that response
into account when projecting network performance. This is expected to reduce the
adverse impacts of improved information.
In this dissertation, we will concentrate on the use of a dynamic traffic assignment
model for the prediction of congestion. In the next chapter we describe how such a
model may fit into the context of the framework described above, its constituent elements,
as well as a candidate approach for its implementation.
Principle 3: CAR Should Maintain Guidance/Prediction Consistency
Based on the discussion presented above, it becomes clear that none of the guidance
systems in existence attempt to anticipate the impact of guidance being provided. The
same holds true for analyses conducted by researchers in relation to ADIS. In this
dissertation, a major principle underlying the framework being proposed is that
consistency has to be maintained between the guidance being provided to drivers and the
predicted traffic conditions. That is, the information system has to check that the
guidance being provided will prove to be optimal to guided drivers based on a prediction
of the future traffic conditions. This would result in guidance information which is
consistent with anticipated driver behavior and network conditions. This ensures the
validity of the guidance information and would encourage its use by more drivers. In
addition, consistency has to be maintained between the traffic conditions as perceived by
the COP and the actual traffic conditions so that the COP remains attuned to the real
world.
3.4 Control and Routing (CAR) Strategies
When specifying the CAR module, three major issues have to be addressed, namely:
Vy = 7
nz , 42
(1) What logic should the CAR module adopt in order to provide drivers with guidance
advice? Should information or route directives be provided to drivers? Should route
directives be based on shortest path guidance or on route distributive guidance? How
should route distributive guidance be implemented? How can the CAR logic ensure
guidance/projection consistency?
(2) Given that guidance provided at specific locations of the network can be updated
periodically, what is the impact of temporal update frequency on the effectiveness of
routing strategies?
(3) Given that drivers may receive guidance updates at various locations as the move
through the network towards their destinations, what is the impact of spatial update
frequency on the effectiveness of routing strategies?
Each of these three issues is discussed below.
3.4.1 CAR Logic
Information vs. Route Directives
The most basic distinction between various logic types used as part of CAR relates
to whether CAR provides drivers with information or route directives.
If information is passed to drivers then CAR would constitute a simple link whereby
the traffic data output from COP is interpreted, relayed to drivers, and presented in an
understandable form. This represents the case in many demonstration projects such as
AMTICS, RACS and Smart Corridor, and has served as the basis for ADIS analyses
conducted by several researchers (see, for example, Mahmassani and Jayakrishnan, 1989).
Route Directives
Single Route Guidance vs. Route Distributive Guidance
When route directives are employed, specific routing strategies may be based on the
assignment of all traffic during a guidance interval to one path (usually the shortest path)
while a different policy may be to distribute traffic over a number of alternate routes
leading to the desired destination. Distributing traffic would entail the determination of
the optimal fraction of traffic to be guided to the different routes, a process which may
require significant computational effort in real-time.
In a paper by Jeffrey et al. (1987), the authors indicate their concern that guidance
provided by AUTOGUIDE may overload a single "best" route to which drivers are
guided. To deal with this problem they suggest that the "closeness" of journey times
along alternative routes can be used to determine what proportion of vehicles should be
af 4 9
sent along each route.
Papageorgiou and Messmer (1991) used feedback regulation in an attempt to establish
a dynamic user optimum whereby different routes used by flow departing at the same
time between an O-D pair have the same travel time and no unused path between the
same O-D pair has a shorter travel time. The feedback regulator would observe current
traffic conditions and determine the way traffic should be split between different paths
connecting O-D pairs, with the aim of establishing the dynamic user optimum conditions.
The first feedback law adopted consisted of shortest route guidance with vehicles
traveling between a specific O-D pair being guided to the route currently having the
shortest travel time between the O-D pair. However, the authors indicated that such a
logic may lead to strong perturbations of traffic flow especially if a large fraction of
vehicles are equipped. Therefore, they suggested an alternative logic based on a "smooth
regulator" using a more advanced feedback law that leads to route-distributive guidance
and the specification of optimal rates of splitting traffic among alternate routes.
Mauro (1991) used the NEMIS package for microsimulation of urban traffic to assess
the performance of a decentralized route guidance scheme. The strategy is implemented
by first performing an off-line stochastic user equilibrium assignment of traffic demand
on the network in question in order to obtain nominal values of link state variables. In
real-time, the local controllers try to keep network performance as close as possible to the
nominal values by varying the splitting rates based on the observed deviation of state
variables from nominal values. The local controllers are not attempting to optimize any
specific function but are merely trying to reduce the effects of disturbances. In case of
incident congestion, the methodology assumes the existence of a link "observer" capable
of detecting abnormal conditions and subsequently initiating an "information wave"
backward in the network. This "wave" propagates along the same paths as the
"congestion wave" would and enforces adjustments in the splitting ratios for all links that
are included in paths leading to the incident location. This scheme also represents a
distributive routing strategy but the logic used to determine the splitting ratios is different
from that used by Papageorgiou and Messmer in the paper referred to above and which
is based on feedback control concepts.
Difficulties in Implementing Route Distributive Guidance
Depending on the specific technology used for providing drivers with route directives,
it may not be technically possible to distribute traffic over a number of routes at the same
instant of time. For example, if variable message signs are used, the best that can be
accomplished is to change the sign within the guidance interval so that the time average
of the route directives would correspond to the fractions we wish to achieve for each
route. On the other hand, if in-vehicle units are used, it is technically possible to provide
different drivers with different route directives at the same time in order to split drivers
between routes according to the optimal fractions. However, this process may not be
politically or legally acceptable due to its inequity implications and a scheme similar to
Dace 44
what was suggested for variable message signs may have to be adopted.
In this dissertation we propose that shortest route guidance be used whenever possible
unless it does not succeed in maintaining projection/guidance consistency in which case
it becomes necessary to modify the CAR logic. This issue is discussed next.
Maintaining Guidancel/Projection Consistency
It is important to recognize that there is a closed loop within the control center
computations in the diagram of Figure 3.1. The loop involves the COP sending data on
projected traffic conditions to CAR which in turn sends back to the COP information
which defines the guidance environment that is in effect at specific points in time. This
guidance information should be taken into account by the COP when projecting traffic
conditions for the next forecasting period.
Specific route guidance should only be provided after going through the loop at least
once and having the CAR ensure that the guidance is consistent with traffic conditions
projected by the COP. The guidance that is provided to drivers should represent a fixed-
point solution of the COP/CAR interaction. For example, if shortest path guidance is
being used, consistency would be achieved if traffic conditions and travel times predicted
by the COP indicate that vehicles are guided to the route that is predicted to be shortest.
If consistency is not achieved, the guidance advice has to be revised. Future traffic
conditions resulting from the revised guidance have to be predicted by the COP and
consistency with the provided guidance checked again. This loop may need to be
repeated several times until consistency is achieved. It is worth noting that if shortest
path guidance is being used then the scheme described above may not converge and it
may not be possible to achieve COP/CAR consistency. This is particularly true if the
guidance is not updated frequently and therefore COP predictions take into account the
response of a large number of drivers to guidance measures, with possible oscillations in
potential messages and route usage between subsequent executions of the COP/CAR
cycle. Such oscillations and the implied lack of convergence would indicate that the
control center is not capable of managing and mitigating the effects of the potential
overreaction that will occur due to provision of guidance. This necessitates that the CAR
logic using shortest path guidance be modified to control the occurrence of overreaction
and to deal with COP/CAR inconsistency.
Consistency Check
Whenever the COP/CAR consistency check referred to above indicates a lack of
convergence using single route guidance, the travel time on the alternate routes would not
be sufficiently different to warrant sending all guided traffic to the shorter route without
causing overreaction. In such a case the CAR logic should either increase the temporal
update frequency (as discussed in subsection 3.4.2) or guided vehicles should be
Dage 45
distributed in such a way that they do not concentrate on any single one of the alternate
routes.
Route Directives with Guidance Threshold
The practice in existing route guidance experiments and studies of guidance being
provided to a specific route whenever the travel time on such a route is evaluated to be
shorter than that on alternate routes even by a very small amount often causes the
guidance/projection inconsistency. As was indicated above, if all guided vehicles are sent
to a route that is only marginally shorter than alternate routes, overreaction is likely to
occur. We suggest here the use of a guidance threshold when route directives are in
effect so that guidance is provided only if the travel time on the route to which drivers
are directed is shorter than that on alternate routes by a specific threshold. In the absence
of a route that is shorter than alternate routes by at least this threshold, guided vehicles
have to be distributed over alternate routes. Such a scheme would have a potential to
reduce overreaction associated with improved information and may be implemented based
on current or projected traffic conditions. While this scheme does not specifically identify
occurrences of inconsistency as the "consistency check” scheme would, it offers a
computational advantage over that approach since a full projection/guidance consistency
check (which requires a number of iterations and may involve checks over a long
projection horizon) need not be made.
Implementation
In the case of collective route guidance technologies such as variable message signs
the two schemes described above may implement "vehicle distribution” by simply
providing "no guidance" and leaving drivers to distribute themselves as they do under
normal traffic conditions (no incident congestion). Such "natural" distribution is expected
to yield good results since it would not impose excessive loads on any of the alternate
routes. In addition, in the cases when it is called for, it would eliminate the
computational and communication requirements associated with the determination of the
optimal splitting fractions associated with a route distributive strategy, and the actual
provision of such guidance.
In the case of in-vehicle units, on the other hand, drivers expect to receive route
guidance at all times and it may not be acceptable not to provide guidance as a way of
distributing guided vehicles. One way of distributing guided vehicles in this case is to
provide guidance in such a way that the time average of the directives relating to specific
routes are equal to the average fractions of vehicles that would take each of these routes
under normal traffic conditions. This would eliminate the computational effort required
to determine the optimal splitting fractions, but guidance still needs to be communicated
to the in-vehicle units.
3.4.2 Temporal Update Frequei:.~¥
Page 46
The frequency of guidance updates plays a major role in determining the relative
effectiveness of CAR schemes. If updating is infrequent, there is a real danger that a
CAR logic such as shortest path guidance may overload the shortest path and cause
overreaction. In this case route distributive guidance might be necessary, at a possibly
significant computational cost. On the other hand, if guidance is updated frequently,
shortest path guidance might perform satisfactorily and route distributive guidance might
not be called for. The tradeoff here is between the extra computational efforts involved
in (1) computing the splitting ratios for distributive guidance, and, (2) providing more
frequent updates in the case of single route guidance. For instance, it may require similar
computational efforts to provide distributive route guidance every 2 minutes or single
route guidance every 1 minute, for specific ADIS operating conditions. In that case, a
comparison of the effectiveness of guidance provided in each of the two cases would be
of interest. However, the specifics of this tradeoff are not likely to become clear except
after some experience with the actual ADIS operation. Obviously, these issues require
further analysis and research.
In addition, the guidance frequency may mitigate the impacts of using a less advanced
information basis for CAR. For example, use of current traffic conditions as a basis for
CAR would require very frequent updates of route guidance to avoid situations of serious
overreaction where congestion shifts from one location to another (Ben-Akiva et al,
1991).
3.4.3 Spatial Update Frequency
The temporal update frequency discussed above relates to updates at a specific
guidance location. Another consideration in the implementation of CAR strategies is the
availability of spatial updates for a specific vehicle that is en-route to its destination.
Depending on the network structure, it may be possible for such a vehicle to receive
guidance at more than one location during its trip. This would provide the CAR with
possibilities of revising and correcting guidance that was provided to vehicles at an
upstream location but which is no longer valid due to emerging traffic conditions (such
as the occurrence of accidents). Such possibilities of spatial updates are potentially
significant in improving the effectiveness of various CAR strategies. In the analysis by
Koutsopoulos and Xu (1992) referred to above, it was observed that as the spatial update
frequency increases, the adverse effects that were observed with low temporal update
frequency at high guided fractions are somewhat alleviated. Clearly, this observation is
consistent with our a-priori expectations. A high degree of spatial update frequency is
also likely to make an advanced COP module less necessary.
However, it should be noted that the benefits from a high spatial update frequency
depend to a large degree on the network structure. Specifically, the network structure has
to provide opportunities for re-guidance and route diversion if the spatial update of
guidance is to have any effect. For instance, for two routes in parallel with no crossovers,
updating guidance along the route when it is not possible to switch to an alternate route
ne 17
would be worthless.
3.5 Possible COP Schemes
3.5.1 Types of Information Provided by COP to CAR
As was discussed in section 3.1, there exist two main types of information that could
be provided by COP to CAR for use by routing strategies associated with Advanced
Driver Information Systems (ADIS). Some researchers hold the view that proper routing
strategies require information on current traffic conditions only while others assert that
such routing strategies have to be formulated based on a forecasting or projection of
future traffic conditions on the network. The rationale behind the two approaches was
presented in section 3.1.
For the approach which bases its routing strategies on projected traffic conditions,
there are the additional questions as to what projection methodology to adopt and the
required accuracy of such methodology. These questions will be discussed next.
3.5.2 COP Models
Here we focus on the COP models that are required to provide CAR with the various
types of information it may require to formulate routing strategies. The information types
that may be provided by COP are:
(1) The most recent observations from the surveillance system:
In this case COP merely provides a link between the surveillance system and CAR.
This scheme would preclude any need for projected traffic demand levels (updated O-D
data) or for projecting future congestion levels. The problems involved in having COP
provide CAR with current traffic measurements as opposed to projected traffic conditions
were discussed in section 3.3 when presenting the principles underlying the proposed
framework.
(2) Projected traffic conditions:
Conceptually, this scheme is more appealing although its implementation is more
complicated. Three possible methodologies are presented below.
(a) The first approach consists of a Dynamic Traffic Assignment model (DTA). Such
a model is described in detail in Chapter 4. Briefly, the DTA provides the ability to
project traffic conditions while taking into account potential driver response to
guidance and the predicted time-dependent origin-destination matrices. The DTA
involves a prediction of driver behavior given the availability of guidance, an
pr -— LOa7 0
assignment of the predicted time-varying path flows to the network, and the
subsequent determination of resulting flows and congestion on the various links of the
network by time of day. To provide such capabilities, the DTA requires updated
predictions of O-D demands, a dynamic driver behavior module, and a dynamic
network performance module.
It should be noted that the provision of guidance based on projected traffic conditions
is quite difficult in congested networks since such conditions are dependent on the
ways in which drivers respond to the information. In other words, the validity of
predicted network conditions depends on their consistency with current and future
drivers’ choices which depend on the drivers’ use of such information. This scheme
would also carry with it a simultaneity problem in order to ensure consistency
between provided guidance and projected traffic conditions (the guidance/projection
consistency requirements were discussed in subsection 3.4.1).
(b) Another approach to congestion prediction which involves significantly less
computational and hardware requirements is the use of statistical time-series methods.
Such an approach has been used for adaptive traffic control systems (see, for example,
Stephanedes et al., 1981 and Okutani et al., 1984) and would make use of historical
as well as recent traffic observations from the surveillance system to come up with
congestion predictions. This approach is appealing because it does not require
detailed modelling of O-D patterns, network structures, or driver response. However,
its validity may be limited to relatively short projection periods. Both schemes are
worthy of further investigation and it might well be that the best approach would be
some combination of the two.
(c) An alternate procedure may consist of tracking time-dependent splitting fractions
at junctions and time-dependent demand levels at entrances to the network, with both
being updated on-line based on real-time flow data using a statistical procedure that
relates them to historical and recent values. The proposed procedure would predict
how flows incoming to the various junctions in the network would split at these
junctions thus providing the capacity to predict congestion on the various links of the
network. If incoming traffic demands and splitting ratios at each junction are
projected over a period of 30 minutes, for example, it would be possible to detect the
occurrence of queues or congestion over that time span. Therefore, it may be argued
that the application of this procedure would in effect provide the same type of
information as a DTA model based on estimated 3-D O-D matrix.
Procedure (c), while computationally appealing, suffers from a number of
disadvantages, namely:
» There is no real look-ahead feature: The prediction of flows based on splitting ratios
at individual junctions may provide estimates of traffic conditions at links that are
directly downstream of the junction. However, the impact of these flows on junctions
Y"dorFi lp at0
that are further downstream is not identified. The DTA procedure, by assigning
traffic to the network and tracing its movement from origin to destination, provides
a real look-ahead feature that takes into account the effects of flow that is now at a
certain junction on traffic conditions at downstream locations in the future.
» Lack of behavioral rules: The prediction of splitting rates at junctions based on
historical data and updated according to real-time flow data would preclude any
behavioral assumptions as to the en-route driver decisions that may involve diverting
from pre-planned routes to alternate routes based on observations of traffic conditions
or guidance provided by an informative system.
« Unreasonable paths: In addition, the prediction of splitting rates is based on past
observations of splitting rates and does not take into account the downstream travel
times. Therefore, since flow entering the network is routed at successive junctions
to its destination based on splitting ratios predicted as such and without any regard
to the cumulative effect of such junction-by-junction routings, the path that a packet
ends up following between its origin and destination may be unreasonable in terms
of the travel time encountered.
» Applicability: Procedure (c) applies mostly to cases in which a complete count of
inflows to and outflows from the network being considered are available, as may be
the case with freeway networks. However, the unavailability of such data in urban
street networks makes procedure (c) infeasible in these types of networks.
It should be noted that the first two disadvantages related to procedure (c) also apply
to procedure (b) outlined above in that it does not incorporate any behavioral rules
and the look-ahead feature it embodies does not take into account the future impact
of flow currently at a specific network location on downstream links.
If, in procedure (c), the splitting ratios at junctions are not determined solely by a
statistical procedure but also by behavioral rules that take into consideration the
decisions that may be taken by individual packets at junctions (possibly based on
downstream travel times), then this alternate procedure becomes more plausible.
However, the absence of a look-ahead feature for en-route diversions, as noted in the
first point above, remains a weakness of the alternate procedure. Obviously, the DTA
procedure would not suffer from such a shortcoming.
However, for short term projection, where the horizon is of the order of 5 to 10
minutes, such a statistical procedure has its advantages. On the other hand, the
behavioral models embedded in the DTA are more capable of capturing changes in
traffic conditions over longer projection horizons.
a ie,I)
Other Issues Related to COP Models
Associated with each approach are the issues of traffic data collection and
computational and accuracy level requirements. These requirements have a significant
impact on the suitability of the projection method with respect to both hardware needs
and effectiveness. In question are the following issues: (1) whether the additional effort
required to base driver information on projected traffic conditions is justified or not; (2)
the level of reliability at which such projections become beneficial; and, (3) the traffic
projection methodology most likely to lead to desired results.
The IMIS Experience
A previous attempt to tie together a quasi-dynamic traffic assignment with route
guidance for future traffic levels was done in the IMIS corridor studies (Gartner and
Reiss, 1987). In this system, O-D demand levels are projected for a future time interval
using a quasi-dynamic traffic assignment. A system-optimal assignment routine is then
used to allocate flows among the alternative facilities in the corridor. Route guidance
information, via variable-message signs, is then provided to the drivers in the hope that
suitable numbers of them would be diverted to achieve the envisioned "optimal" flow
distribution. Ramp metering rates and signal control strategies are then implemented in
accord with the anticipated flows. While novel for its time (these studies were done in
the late 1970’s) this scheme also has some weaknesses. Principal among them is the fact
that the assignment is a static one and link performance is only evaluated by crude
performance functions.
3.5.3 Maintaining Consistency with Actual Traffic Conditions
Referring to Figure 3.2, at the beginning of new computational cycles "refreshing" is
required so that the COP computations taking place at the control center remain consistent
with the real-world traffic conditions as perceived by the surveillance system. COP will
require surveillance data relating to volumes, travel times, and the location and expected
duration of incidents. By periodically introducing the current state of the system, at times
t, as "initial conditions" in the COP computational process, the COP will keep track of
and remain consistent with actual traffic conditions.
3.6 The Role of COP and CAR in Reducing the Adverse Impacts of Improved
Information
3.6.1 Possible COP/CAR Scenarios
Figure 3.3 is a schematic representation of the interrelationships between traffic
conditions, guidance information, and travel decisions by informed drivers. It indicates
that a proper modelling framework would provide the DTA with information regarding
“AEC - 1
guidance that had been given to drivers. The information processing module in the DTA
models how drivers use the guidance information to formulate their updated perceptions
of travel conditions in the network. These updated perceptions form the basis of expected
travel adjustments and decisions made by drivers, and this process is modeled by the
dynamic driver behavior module. Driver decisions, in their totality, define the forecast
traffic demand on the network. This is input to the dynamic network performance module
whose job is to translate those demand patterns into projected traffic conditions. The loop
is completed when the projected traffic conditions are relayed to the CAR in order to
come up with the next round of guidance information. As mentioned above, guidance is
implemented only after ensuring consistency with predicted travel times. Guidance
information and projected traffic conditions are consistent and reliable only when the
interaction between the two is explicitly considered as depicted in Figure 3.3.
Figure 3.4 presents a case in which guidance information is not provided as an input
to the COP model. In this case the forecast traffic demand would be based on historical
data and surveillance data and would not reflect the adjustments in driver travel behavior
based on the guidance being provided. Assignment of these patterns to the network
would produce projections of network performance which, in fact, would be inconsistent
with the guidance information in the sense that they assume that drivers do not react to
it at all. This serves to illustrate the importance of the driver behavior module as a part
of a proper COP system since it would make sure that projected traffic conditions reflect
the impact of guidance on travel adjustments made by drivers.
Figure 3.5 depicts a degenerate case in which no traffic assignment is used at all. In
this case current traffic conditions form the basis of the guidance information. Moreover,
the guidance strategy is a very simple one based on indicating to all drivers the shortest
path in effect at the time the guidance is provided, ignoring all driver reactions to the
guidance information. An enhanced version of this scenario attempts to base the shortest
path guidance on projected traffic conditions. In this case, future traffic conditions are
estimated from a comparison of current conditions with historic conditions. As was
discussed in Chapter 2, one of the advanced guidance systems currently in operation (the
Berlin system) utilizes this enhanced scenario as a basis of providing guidance to drivers.
3.6.2 Dealing with Overreaction and Concentration
Next we briefly review the two behavioral phenomena (overreaction and
concentration) which may negate some of the beneficial effects of improved information
and discuss possible schemes to reduce these potential adverse impacts. These
phenomena were discussed in Chapter 2.
Overreaction - Overreaction occurs when drivers’ reactions to traffic information cause
congestion to transfer from one road to another and/or generate oscillations in road usage.
Effect of Type of Information Provided by COP to CAR and COP Model:
do €r Ey)
When current traffic information is being provided, overreaction is likely to take place
if drivers fail to consider or underestimate the potential responses of other drivers.
However, in such a case overreaction may be avoided if drivers’ decisions are based on
correct expectations of other drivers’ reactions. For example, some drivers may not shift
to a reportedly faster route because they anticipate a rush of the other drivers to that
route. Similarly, predictive traffic information that is based on the statistical approach is
not able to foresee the occurrence of overreaction since driver response to guidance does
not enter into the predictions.
In the ADIS framework being presented in this dissertation, guidance based on
predictive traffic information using a DTA is expected to be provided to drivers. In such
a case, overreaction may be avoided if the DTA accurately predicts driver behavior and
reaction to information, including the fraction of drivers who will follow the guidance
advice, and if CAR maintains prediction/guidance consistency.
Effect of CAR Logic:
Using shortest path guidance is likely to fail if CAR ignores the impact of drivers’
reactions to information. This scheme does not perform well in congested networks since
the single route to which drivers are guided will become congested quickly as drivers
switch to it and it would no longer represent the best route. This results in an
inconsistency between predicted and realized traffic conditions and leads the guidance
system to suggest (or drivers to switch to) other routes thus resulting in an oscillatory
behavior which is characteristic of overreaction cases.
The "consistency check” approach described in subsection 3.4.1 ensures that an
alternate CAR logic is resorted to whenever the occurrence of overreaction can not be
avoided using single route guidance. Similarly, using a "guidance threshold” in
association with route directives is bound to reduce the impact of overreaction since the
route to which drivers are guided would be shorter than alternate routes by some amount.
Therefore, even if a moderate quantity of traffic shifts to this route, it still might not
cause a lot of congestion on the route, and overreaction may be avoided.
The use of a guidance threshold can be thought of as equivalent to drivers behaving
according to boundedly rational rules whereby they switch to an alternate only if it offers
an advantage greater than a specific threshold. It is interesting to note that Mahmassani
and Chen (1991) observed that very little improvement in overall system performance
occurs when drivers adopt a myopic behavioral rule as opposed to having a non-zero
route switching decision band. This myopic behavior may be viewed as similar to having
a guidance threshold of zero.
Effect of Temporal Update Frequency:
More frequent guidance updates are also likely to reduce the potential occurrence of
he 33
overreaction since that would limit the number of vehicles that are exposed (and react)
to any specific guidance message.
In summary, the occurrence and severity of overreaction may be reduced either by (a)
basing the guidance on predictive traffic information that accounts for driver reaction, by
(b) providing guidance in such a way that traffic is distributed over several reasonable
paths, or by (c) performing more frequent guidance updates.
Concentration - In general, a group of drivers travelling from an origin to a destination
tend to use different routes and departure times because they have heterogeneous
preferences and diverse perceptions of network conditions. Differences between perceived
and actual network conditions imply that drivers who are not well informed may select
alternatives which are not in their best interest. Information tends to increase the
uniformity of driver perceptions of network conditions around the true values. As a
result, a greater number of drivers may select the best alternatives (from their individual
points of view) and consequently drivers with similar preferences will tend to concentrate
on the same routes during the same departure times. Thus, more information could
potentially generate higher levels of traffic congestion.
Effect of CAR Logic
When current information concerning traffic conditions is provided to drivers, it may
not be possible to reduce the impact of concentration. However, when route directives
are provided the impact of concentration may be mitigated if drivers travelling between
the same origin-destination pair are distributed over different feasible routes instead of
being uniformly provided with the same guidance.
Effect of CAR Update Frequency
More frequent guidance updates, both temporal and spatial, are likely to reduce
concentration. With more frequent temporal updates, a smaller number of drivers would
be exposed to the same information at the same location. On the other hand, with a
higher frequency of spatial updates, drivers who received information at one guidance
location may receive more updated information at downstream guidance location. Both
schemes are expected to succeed in reducing concentration since both have the ability to
reduce the uniformity of driver perceptions of network conditions that are likely to occur
with lower frequencies.
Effect of COP Model
If information or route directives provided to drivers are based on projected traffic
conditions provided by a DTA, the information acquisition and integration models within
the dynamic driver behavior module would be able to duplicate the occurrence of
concentration based on the more uniform driver perceptions resulting from information
Dage 54
provision. In a manner similar to the case of overreaction, this modeling helps identify
potential cases of concentration in order to alleviate their effects. This alleviation may
then be achieved by using proper guidance strategies (such as customized route directives)
or traffic control measures, if available.
Summary
The above discussion indicates that several COP and CAR properties may be utilized
by ADIS in order to reduce the potential of overreaction and concentration. Four such
properties are:
(1) Guidance and routing strategies have a better chance of succeeding if they are based
on projected traffic conditions and include an explicit model of dynamic driver travel
decision making in response to guidance within the COP. This ensures that the impact
of the guidance being provided on driver travel decisions are accounted for when the COP
projects traffic conditions.
(2) They should have the capability of utilizing more advanced CAR logic which may
include the use of route distributive guidance whenever required.
(3) The adverse impacts of guidance can be reduced substantially if the guidance
frequency is high.
Therefore, it may be concluded that guidance which considers potential driver
response to guidance, maintains guidance/projection consistency, resorts to an advanced
CAR logic whenever needed, and which is updated frequently is likely to constitute a
basis for more effective ADIS.
3.7 Evaluation Criteria
The importance of a clear set of evaluation criteria for the various COP/CAR schemes
can not be overemphasized. For instance, Papageorgiou and Messmer (1991) define the
following as conditions for dynamic user optimum: For traffic leaving an origin during
a specific time interval, the splitting ratios should be such that no traffic would follow the
route with the longer travel time to the destination. These conditions are thought to be
useful since they parallel the static user equilibrium conditions whereby no driver can
improve his travel time by unilaterally changing routes. However, when presenting the
results of their analysis, they fail to give a specific indication of how well their feedback
regulator works and how close it comes to satisfying these conditions which they set out
to achieve through their feedback control logic. They mention that these conditions are
met "for most" time intervals but that this might not be the case if strong oscillations of
the disturbances (compliance rates and demand levels) occur. Obviously, a more precise
Ne ~ 55
assessment of the performance of the strategy would have been helpful. Similarly, Mauro
(1991) provides only a descriptive assessment of the effectiveness of his proposed routing
strategy and asserts the need for a more extensive analysis before any conclusions may
be drawn. In addition, there should always be specific statements about the infrastructure,
hardware, and computational needs of any routing strategies so that the cost and
operational implications are well understood.
We suggest the following criteria for evaluating various routing strategies:
(a) Effectiveness in improving traffic conditions: More advanced routing strategy have to
offer tangible benefits compared to the no-guidance situation as well as compared to
simpler strategies. These benefits may be measured by the reduction in average travel
times on the network for guided, unguided, and all traffic.
(b) Potential validity of the routing strategy. Guidance provided to drivers has to be
perceived as being valid in the sense that drivers have to feel that they are being guided
to the best available routes. Reliability may be measured as the percent of guided drivers
who receive "correct" guidance leading them to the optimal routes. This criterion is
essential since the perceived validity of any guidance system has a significant impact on
driver compliance.
(c) Travel time reliability: One of the more significant advantages associated with the
adoption of a real-time route guidance system consists in its ability to provide advice that
will allow drivers to avoid long delays. The ability of any guidance system to achieve
this goal may be measured in various ways; for example, the maximum travel time
experienced by guided and unguided vehicles and the percent of guided vehicles which
encounter long travel times would constitute possible measures of travel time reliability
which may be used to compare the effectiveness of various COP/CAR and the
improvement they provide over the no-guidance case or over simpler guidance strategies.
Another measure of travel time reliability used by Koutsopoulos and Yablonski (1991)
is the standard deviation of travel time with and without guidance.
(d) Implied traffic data needs and infrastructure requirements: This criterion defines the
operating environment of the routing strategy in terms of the extent to which it requires
traffic data collection equipment such as sensors, detectors, etc. Obviously, these
requirements go a long way towards determining the cost associated with implementing
specific routing strategies.
(e) Implied computational requirements: Implementation of any routing strategy requires
a certain computational effort to analyze raw data and formulate routing advice.
However, these computational requirements vary widely depending on the volume of raw
data that needs to be analyzed, the information basis for CAR, as well as the specific
logic being used for CAR. Therefore, it is essential that the different routing strategies
Mac £6
be evaluated taking this computational intensity dimension into consideration.
Another useful measure of effectiveness that was used in the IMIS study (Gartner and
Reiss, 1987) consisted of comparing the time required to clear congestion after an incident
is removed for the guidance and no-guidance cases. The magnitude of reductions in
incident congestion clearance time that may be attained through guidance determine the
potential of guidance in mitigating the impacts of non-recurrent congestion.
ard wt ]
Figure 3.1 Proposed Framework
Surveillance System
Volumes, Speeds, Incidents
Historical "Updated
O-D Data 0-D Data
Congestion Prediction
Infrastructure
Data
Traffic | ‘Route Information!
Conditions & Guidance _
Control and Routing
Control
Center
Actual Traffic Conditions
Flow of Information nformation
EE,TE
Flow of Information
under Some Scenarios System Element
Page 58
Figure 3.2 Real-Time System Operation
Projected
Time
1
Time Interval To
over which
future conditions
are projected by
the COP module
‘Current” time at Elapsed time
which the COP needed to
and CAR are executed axecute the COP
and CAR modules
~My
Co, Real Time
J
To= Start of Incident
T = [Projected] End of Incident
D =
C2 53
Figure 3.3 Framework Implementation
Control and Routing
Projected Guidance
Traffic Conditions information
Dynamic Network information
Performance Processing
Dynamic Driver Behavior
Dynamic Traffic Assignment
Page 60
Figure 3.4 No Driver Behavior Modelling
Control and Routing
Projected Guidance
Traffic Conditions Information
Dynamic Network
Performance
Forecasted Traffic Demand
Congestion Prediction
Page 6]
Figure 3.5 Degenerate Case
Shortest Path Guidance
Current/Projected Guidance
Traffic Conditions Information
Page 62
CHAPTER 4 DYNAMIC TRAFFIC ASSIGNMENT (DTA):
BACKGROUND, LITERATURE REVIEW,
AND PROPOSED FRAMEWORK
Chapter Objectives and Layout
The objectives of this chapter are as follows:
» to justify the need for dynamic traffic assignment models and to describe how
they differ from traditional traffic assignment models (section 4.1)
» to review the state-of-the-art in dynamic traffic assignment and make conclusions
concerning the suitability of existing models for use as COP modules within the
proposed framework for real-time driver information provision (section 4.2)
» to present a proposed approach to and elements of the DTA (section 4.3)
focussing on dynamic network performance (section 4.4) and the required O-D
updating module (section 4.6)
4.1 Need for Dynamic Network Models
Chapters 4 and 5 of this thesis present the details of the proposed approach to the
development of the real-time Dynamic Traffic Assignment model (DTA) that will serve
as a COP model within the proposed framework. As a background to the proposed
approach this chapter provides an overview of dynamic network models and the need for
such models in an ADIS environment.
Traffic Assignment Models
Traffic assignment models represent the interaction between demand for travel
over a transportation network and the network performance. Traffic flows over a network
may be assigned based on a given level of travel demand, some hypotheses regarding
driver behavior, and models of link performance.
Each link in a road network is typically characterized by a level of service (or
impedance function) which is affected by the flow using it. The most common measures
of service in transportation networks are the time and cost required to traverse a particular
link. As a result of congestion, travel time on urban streets is an increasing function of
flow. From this interaction between link travel times and link flows originates the notion
of equilibrium in the analysis of flows on transportation networks (see, for example,
Manheim, 1979 and Sheffi, 1985). Therefore, one of the main objectives of transportation
analysis has been to model the interaction between congestion and trip decisions made
by individuals and subsequently to obtain the expected flow patterns throughout the urban
transportation network. This is known as the traffic assignment problem since its task is
to assign an origin-destination matrix of travel demand onto the network.
3,
. 1. 3
User Equilibrium and System Optimum
To solve the traffic assignment problem, the rule by which drivers choose a route
must be specified. This rule represents a function which splits travel demand over
alternative paths.
A reasonable behavioral assumption is that every motorist will try to minimize the
travel time from his origin to his destination. A stable condition is reached only when
no traveller can improve his travel time by unilaterally changing routes. This condition
characterizes what is referred to as the User-Equilibrium (UE) principle. It is to be noted
that, in the absence of any changes in travel demand or network capacity, the point of UE
represents a stable equilibrium since at that point no force tends to move the flows out
of equilibrium.
On the other hand, flow patterns may be derived from a System-Optimal (SO)
assumption. In this case, it is assumed that all motorists are distributed over the
alternative paths so as to minimize the total system travel time. In this case, a driver may
be able to decrease his travel time by unilaterally changing routes. Such a situation is
unlikely to sustain itself (if no controls are in place) and consequently the SO flow pattern
is not stable.
Heuristic Equilibration Techniques
Several heuristic methods for finding the user-equilibrium flow pattern exist and
have been used extensively. The all-or-nothing method has been widely applied as a
traffic assignment procedure despite the fact that it ignores the equilibrium problem
altogether. This method assigns all traffic travelling between a certain origin and
destinationpair to the shortest route connecting the pair, regardless of the impact of the
assignment on network performance. Other heuristic techniques include capacity restraint
(as in FHWA, 1977, for example) and incremental assignment (see, for example,
Manheim and Ruiter, 1970) methods. These techniques involve repetitive applications
of all-or-nothing assignments in successive iterations. Obviously, the dependence between
flows and travel times are again ignored when traffic is assigned at every iteration. Only
the travel times resulting from the previous assignment are used in the current iteration.
As such, these methods are either not guaranteed to converge or may result in non-
equilibrium flow patterns.
Mathematical Programming Formulation for UE Solutions
The shortcomings of the heuristic techniques have motivated the development of
an equivalent minimization program whose solution provides the UE traffic flow patterns.
Sheffi (1985) presents in detail the solution of the equivalent minimization program using
the convex combinations method. Speed of convergence and computational experience
are also discussed and it is concluded that this solution framework provides UE traffic
Ys “4
flow patterns at a computational effort comparable to that of the heuristic techniques.
Violations of the UE Principle
Most current traffic assignment models are based on the User Equilibrium
principle which states that at equilibrium travel times on all used routes between a
specific origin-destination (O-D) pair are equal and smaller than travel times on all
unused routes. This condition is based on the assumption that travelers choose optimal
routes and have complete knowledge of travel times on all available routes (omniscience);
that is, the underlying principle is that route choice can be modeled as a deterministic
process (the assumptions associated with this principle are discussed in footnote 1 later
in this chapter and in section 5.2 of the next chapter). In actual route choice this
condition is most likely to be violated since drivers are not sufficiently informed to
choose the optimal route. The violation of this assumption is most prevalent in the case
of dynamically changing traffic conditions whereby a driver’s knowledge of road
conditions is continually updated and revised.
Inadequacy of Static Formulations
Traditional steady-state traffic equilibrium models assume that traffic conditions
are uniform and travel demand constant over the course of the analysis period. However,
demand for travel exhibits variability over time-of-day and from day to day. Therefore,
it is clear that static frameworks fail to model (i) how congestion builds up and dissipates
in different parts of the network at different times and (ii) how the flow patterns change
from day to day adjusting to changes in network performance. This inadequacy of static
network equilibrium models in dealing with network performance under time varying
demands has been pointed out by a number of researchers. Friesz (1985) indicated that
including dynamic considerations is among the improvements needed to enhance the
predictive capability of steady-state network models. Ben-Akiva (1985) emphasized that
static network equilibrium formulations fail to capture essential features of traffic
congestion since, first, traffic is assumed to be uniformly distributed during the modelled
time interval and, second, the temporal distribution of the traffic is assumed to be constant
from day-to-day. For instance, during the rush hour observed traffic flows are clearly not
uniform and it is common knowledge that delays occur at different locations during
varying time intervals. The non-uniform distribution of peak demand can be attributed
to the fact that during the peak period most travellers wish to arrive at their destinations
within a short time window, but due to the lack of adequate capacity some travellers have
to shift their departure times so that they arrive at their destinations a little earlier or later
than desired.
Levels of Dynamic Traffic Conditions
—-
yr: Al
Ln 3
Traffic conditions in a network can be labelled "dynamic" whenever travel patterns
change over certain time horizons. Models of network traffic flows may be dynamic in
one of the following three senses:
(a) Day-to-day dynamic models which consider the evolution of traffic flows from
day to day while assuming uniform distribution of traffic within any specific day. Such
models usually simulate drivers’ daily path choices.
(b) Within-day dynamic models whereby travel demand may have different levels
over the course of the day. In this case departure time choice becomes a significant
dimension of overall travel choice. Such models usually simulate drivers’ daily route and
departure time choices.
(c) Real-time dynamic models in which adaptive henavior is considered by
allowing for en-route driver rerouting.
Need for Dynamic Traffic Assignment (DTA)
The Congestion Prediction (COP) element has the responsibility for providing the
Control and Routing module with the information that is needed to implement routing and
guidance strategies. Among the principles adopted in this dissertation arethe need for
COP to provide CAR with projected traffic conditions and the fact that COP should be
performed by a DTA that will take driver response into account in its projections. The
DTA predictions of future congestion in the network are based on:
» current traffic conditions (consideration of initial conditions)
» predicted time-dependent O-D demand levels (sensitivity to future demand
patterns)
guidance being provided and anticipated driver response to guidance
The DTA should proceed as follows. Information regarding predicted O-D
patterns as well as guidance that is given to drivers by the CAR is provided to the DTA.
The dynamic driver behavior module in the DTA models how the guidance provided to
drivers would cause them to adjust their travel choices, whether at the pre-trip or en-route
levels. The forecast driver decisions are input to the dynamic network performance
module whose job is to translate those demand patterns into projected traffic conditions
that are later relayed to the CAR in order to come up with the next round of guidance
information. The DTA procedure, by assigning traffic to the network and tracing its
movement from origin to destination, provides the required look-ahead feature that takes
into account the effects of flow that is now at a certain junction on traffic conditions at
downstream locations in the future.
Therefore, dynamic traffic assignment involves the following processes:
age £6
» predicting driver behavior given the availability of guidance
(departure time as well as pre-trip and en-route diversion choices)
» assigning time-varying path flows to the network
» determining the resulting flows on different links of the network by time of day
As such, dynamic traffic assignment has to deal with the variability of link
performances by time of day resulting from the time-varying flows on these links. As
noted above, this represents a significant departure from the static formulation where
origin-destination flows, and consequently link flows and performances, are assumed to
be constant over the analysis period.
The dynamic traffic assignment methodologies that will be reviewed in the next
section will be evaluated based on their ability to perform the functions associated with
COP, as outlined above. These functions are considered imperative if the DTA is to
function in a role that supports the operation of real-time driver information provision.
4.2 Review of Dynamic Traffic Assignment Methodologies
Whereas methods to solve for static traffic equilibrium are well established and
widely accepted, this is not true in the case of dynamic traffic assignment. Several
approaches have been suggested to perform dynamic traffic assignment, and many others
are being developed, with no single methodology being generally accepted. In fact, all
the methodologies which have been suggested so far suffer from significant shortcomings
which render their applicability within the context of ADIS questionable. In what follows
we present the state of the art in dynamic traffic assignment by outlining the approaches
suggested by several researchers and pointing out the advantages and disadvantages of
each approach.
a) Optimization Techniques:
Merchant and Nemhauser (1978 a,b) were among the first researchers to model
dynamic network flows. They provided a discrete-time, system-optimizing dynamic
traffic assignment model for a single destination. In their model, the temporal distribution
of travel demand within each time interval is exogenously given. Moreover, arc travel
costs and exit flows for each time interval are functions of the amount of traffic on the
arc at the beginning of the time interval. Reasonable requirements are imposed on the
arc travel cost and exit flow functions. However, it is assumed that the number of
vehicles that leave an arc in a given time periods depends only on the number on the arc
at the beginning of that period and that none of the vehicles that enter during a period
leave during that period. This assumption places the restriction that the assignment may
be valid only for suitably small time periods or suitably large arc travel times. The
resulting formulation is a nonlinear, nonconvex mathematical programming problem for
which Merchant and Nemhauser present a specialized solution algorithm and optimality
ral 67
conditions. The suggested algorithm considers a piecewise linear version of the model
which can be solved for a global optimum using a one-pass simplex algorithm.
Moreover, it is shown that this piecewise linear program has a "staircase" structure (only
elements i and i+l in row i are non-zero) that is amenable to decomposition techniques
or compaction methods for sparse matrices. It is to be noted, however, that the
nonconvexity still requires computationally expensive schemes.
Carey (1986, 1987) confirmed the validity of the above formulation and expanded
on it. His formulation included flow controls which can be used to keep the actual
outflow from links below the unrestricted capacity levels. He also showed that if such
flow controls are included in the constraints then the formulation becomes a convex
nonlinear mathematical program. The fact that the formulation is a convex program
reduces the analytical and computational problems encountered in the Merchant-
Nemhauser model cited above. Thus, this program can be solved using any of a variety
of well-known algorithms developed for convex programs; for example, a piecewise linear
version of the model is a standard linear program. Moreover, it is also possible to
develop faster algorithms that take advantage of the special structure of this problem,
including the network type subset of constraints and the intertemporal staircase structure
of the constraints. Finally, Carey also outlines extensions of his model which handle
multiple destinations and commodities but notes that not all these extensions yield convex
programs.
Finally, Ho (1980) developed a solution algorithm to the Merchant-Nemhauser
formulation consisting of successive linear optimizations. More recently, Ho (1990)
presented a solution of the same formulation (for a single destination) using parallel
processing on a Hypercube Multicomputer.
The above two models by Merchant-Nemhauser and by Carey share several
significant shortcomings. First, both models represent system-optimum traffic
assignments. Such assignments do not reflect actual driver behavior in which individual
drivers seek to minimize their O-D travel times over the network. The actual traffic
behavior would be better depicted by a descriptive assignment in which a generalization
of the static user equilibrium principle is developed. It should be noted, however, that
Carey establishes conditions under which no flow controls are in effect in which case the
solution to his model approximates user equilibrium. Second, in both formulations there
is no guarantee of consistency between flow entering a link, link traversal time, and exit
flows. Such consistency is trivial in the case of steady state formulations but is a central
requirement of any dynamic traffic assignment model. Finally, both models are
formulated for one destination only. Extensions to multiple destinations have not been
fully analyzed and present non-trivial conceptual and analytical complications.
b) Heuristic Techniques:
Janson (1990) presents a dynamic traffic assignment algorithm which consists of
 {a 38
once-through, incremental assignment. In this model the analysis period is divided into
discrete sub-periods with departure rates for all O-D pairs and for each sub-period being
predetermined. Janson states the optimal conditions for dynamic traffic user equilibrium
(DUE) as follows: (i) All paths used by trips departing at the same time between a given
O-D pair must have the same travel time impedance; and, (ii) No unused path between
this O-D pair for that departure time has a lower travel time impedance. Obviously, these
conditions are a generalization of the static deterministic user equilibrium conditions’.
Janson then presents an algorithm which is designed to "produce assignments that
approximate the DUE optimality conditions". The basic assumptions of the algorithm are:
(i) Route choice decisions are made at the time of trip departure based on projected link
impedances; and, (ii) Projections of future link volumes are obtained by multiplying
current link volumes by factors accounting for changing levels of travel demand in future
time intervals. Consequently, the algorithm functions as follows:
Step 1. Initialization
Step 2. Increment current time interval t
Step 3. Identify origin zones with departures in interval t
a) select an origin zone at random
b) find shortest paths to all destinations based on:
» projected link impedances; and,
 when trips are projected to traverse each link
Step 4. Incremental assignment
a) assign part of the trips departing from current origin to shortest path
found
b) update assigned link volumes and travel times for future time periods
based on new assignments from current origin
c) if all trips from origins with departures in interval t have been processed
go to step 5; else go to step 3
Step 5. If all departure time intervals in analysis period have been processed stop;
else go to step 2
The advantage of the scheme proposed by Janson is its speed for large networks.
This is due to the once-through, incremental assignment scheme he uses.
However, Janson’s method suffers from a number of significant shortcomings.
s—————
! It should be noted, however, that these conditions are associated with the assumptions that individual travellers
have the same tradeoffs between choice factors and that the behavior of an individual in a given situation can be
predicted with certainty. As a result, at equilibrium individuals necessarily have the same travel time or, in general,
the same utility level. Taste variations between individuals may be captured by introducing multiple market segments
into the model and by recognizing the fact that travelers have differing perceptions of travel time and other choice
factors (see Ben-Akiva and de Palma, 1987). The prior use of stochastic models for dynamic traffic assignment is
discussed later in this chapter.
‘age HY
First, it does not represent an attempt at obtaining equilibrium flow patterns for the
dynamic problem. The ad-hoc nature of link volume projections and their use as a basis
for assigning traffic to shortest paths does not ensure the attainment of the optimal
conditions which Janson sets out to approximate. This deficiency is fatal if such a model
is to be used as the basis for Route Guidance Systems since the look ahead feature would
be flawed due to the expected inconsistencies between projected and actual shortest paths.
Providing route guidance to drivers on the basis of this assignment would quickly lead
to drivers losing faith in such a system.
Other shortcomings of this model include the fact that it models only one driver
decision: the choice of route at the origin of the trip and does not allow for en-route
diversion behavior. This constitutes a major flaw if such a scheme is to be used in the
proposed context of ADIS since it is expected that such systems will need to provide
information to drivers which may influence their departure times and routes.
¢) Optimal Control Techniques:
A number of researchers have approached the dynamic traffic assignment problem
by formulating it as an optimal control problem. The first such effort was by Luque and
Friesz (1980) who reformulated the Merchant-Nemhauser model of dynamic system
optimal traffic assignment as a continuous time optimal control problem. This optimal
control framework was later extended to the problem of dynamic user optimal traffic
assignment. Two recent papers by Wie (1989, 1990) present the background, theory, and
formulation of the dynamic system and user optimal traffic assignment from the control
theoretic approach, and are discussed next.
The problem is formulated to predict the temporal evolution of traffic flows on a
congested multiple origin-destination network. Demand patterns are predetermined, thus
the choice of departure time is not modelled within this framework.
The user optimal formulation is claimed to be a dynamic generalization of
deterministic user equilibrium. Network users are assumed to base their continuously
revised route choice decisions on the minimization of travel times. Moreover, the users
are assumed to have complete information on the current state of the network at each
instant in time, and this information forms the basis of the selection of the revised
minimum cost route. Such information presumably could be supplied by in-vehicle
computers. As such, it is assumedthatdrivers have no perceptions of how travel times
may change by the time they traverse the downstream sections of the network. However,
a driver is free to revise his route at any intersection if his current route is no longer
optimal on the basis of updated (current) traffic information. As such, at each instant in
time, and for each O-D pair, traffic is assigned to paths such that the travel times of all
paths that will be used are identical and equal to the minimum instantaneous path travel
time. This results in a static user-optimal network equilibrium at each instant of time, if
currently intended routes were followed.
VeagN- 79
As Wie himself states, the model is not intended to predict time-varying flows
corresponding to a true dynamic generalization of deterministic user equilibrium whereby
all drivers with the same departure time and same origin-destination pair experience the
sametravel cost over all used routes with no unused routes experiencing a lower travel
cost.
The system optimal formulation is structured similarly but assumes that all
network users cooperate in minimizing the total transportation cost, or that individual
routing decisions are completely controllable by a central traffic authority.
This formulation of "dynamic" traffic assignment is not really dynamic, but rather
represents a series of static assignments based on time-varying network traffic conditions.
Inherent in the model are the assumptions that as drivers move along their routes, they
may instantly and continuously revise their route choices in response to the emerging
traffic conditions. Thus, it is assumed that instantaneous adjustments from one
equilibrium pattern to another are possible.
The formulation also suffers from several other shortcomings, including the
unrealistic user behavior assumptions whereby users are assumed to base route choice
decisions on current traffic conditions. In fact, drivers choose their shortest paths based
on travel times they expect to encounter on downstream network sections at the time they
traverse these sections. Drivers are expected to form perceptions of these future travel
times at downstream sections based on their previous experiences. This route selection
mechanism is expected to be an integral part of any model which is to be useful for real-
time route guidance, but is completely ignored in Wie’s formulation as well as in many
other formulations. In addition to the above mentioned shortcomings which make this
model unsuitable for utilization as a real-time dynamic traffic assignment model
supporting ADIS functions, no solution algorithms were suggested by Wie and no
implementation or computational experience was reported. |
In a more recent paper by Boyce et al. (1991), optimal control theory was also
used to model dynamic user-optimal traffic assignment. In this formulation both inflows
and outflows from links are treated as control variables. The authors demonstrate the
equivalence of the optimal control program with instantaneous dynamic user optimal
conditions and show how the continuous time formulation can be transformed into a
discrete time formulation that can be solved by the Franke-Wolfe technique. While this
new model does not represent a true dynamic user optimal situation (as discussed above
with reference to the model suggested by Wie), the model offers a solution algorithm and
the authors present some preliminary computational results.
d) Dynamic Traffic Assignment with Boundedly Rational User Behavior:
Chang et al. (1985) and Mahmassani and Chang (1985,1986) investigated the day-
0-day dynamics of departure-time decisions of urban commuters and introduced the
"
aye
boundedly-rational notion of an indifference band of tolerable schedule delay. This
satisficing behavior was suggested as an alternative to the utility maximizing decision
rules adopted in other studies.
The framework consisted of dynamic simulation and assignment modules and the
analysis focused on a single-destination commuting corridor with parallel routes.
Elements of the framework were (i) a "macroparticle traffic simulator" (MPSM) which
simulated vehicular movement on freeways and arterials given time-dependent demand
rates, and, (ii) a user decisions component which determines the time-dependent departure
rates resulting from daily individual departure time and route choice decisions of
commuters. The users’ decision component is responsive to experienced system
congestion and available exogenous information and includes a simple learning
mechanism.
More recently, Mahmassani and Jayakrishnan (1989) used a similar framework for
investigating the performance of congested urban traffic systems under real-time in-
vehicle information. The effect of the fraction of guided users and mean indifference
band across the population on the path selection behavior of users in response to supplied
information was investigated.
Again, the network structure was limited to a commuting corridor with routes in
parallel which are connected by links allowing for route switching. Only simplified route
switching scenarios are considered, and drivers are provided with information regarding
current traffic conditions only. As such, this model represents a rather simplistic, basic
framework providing no capability for predicting future downstream traffic conditions and
consequently no look-ahead facility. The only role of the traffic simulator is to move
individual vehicles along their routes.
e) Three-Dimensional Approach:
A 3-dimensional traffic assignment method was also developed by Hamerslag
(1988). The analysis period is divided into time bands during which the travel time on
the network links remain constant. The input origin-destination matrix specifies the
number of vehicles departing an origin to a given destination during a time band. For
every flow element in this matrix the travel times are calculated for the first and last
vehicle in a group during a time band. The method keeps track of the exact location in
the network of these vehicles. The minimum travel time paths are determined in a 3-
dimensional network where the travel time to traverse a link depends on the time band.
Initially Hamerslag employed an incremental loading approach similar to that used by
Janson. More recently, Kroes and Hamerslag (1990) implemented an iterative numerical
procedure in which updated travel times are used to reassign a fraction of the traffic until
a convergence criterion is met. The model has also been augmented to include a
capability to redistribute the departure times in response to traffic congestion. The model
was tested on a prototypical network of the major roads around and through the city of
Fd, 72
Rotterdam. A static equilibrium assignment was applied to the same network and O-D
matrix. The results of the dynamic assignment consistently replicated the observed
congestion pattern while the static assignment failed to do so. These tests, however,
encountered difficulties in converging to a stable equilibrium. The reason for this is
inconsistency between the model used to calculate link performance (i.e., a volume-delay
function) and the way in which exit flows from congested links were calculated.
f) CONTRAM
The CONTRAM system (Leonard et. al., 1989; Taylor, 1990) is an existing DTA
with deterministic route choice models designed for off-line applications. It is mainly
used for the analysis and evaluation of traffic management schemes. Inputs to the model
include time-dependent O-D levels. Network performance consists of a free flow link
travel time and queuing delays at intersections.
The model is based on grouping vehicles together into "packets" in order to reduce
computational time. Each packet is treated as a single entity and is assigned to the
minimum time path. All vehicles in a packet are assumed to experience the same traffic
conditions. Time varying flow conditions are modelled by dividing the analysis period
into subperiods (of the order of 15 minutes) and estimating the average link flows during
each subperiod.
The use of packets in this model represents an advantage since such a flow
representation allows more direct modelling of the effects of queue build up and
dissipation during the peak period. However, the behavioral models used in CONTRAM
are rather crude since a deterministic route choice process is assumed in which motorists
(packets) are assumed to possess perfect knowledge even of future downstream traffic
conditions. Moreover, as is the case with most of the models that have been reviewed
so far, the model was not designed to simulate the impact of guidance on drivers.
g) Stochastic Process Approach:
Cascetta and Cantarella (1991) adopted a different approach to the modelling of
within-day dynamic traffic conditions by simulating the system evolution as a stochastic
process. (Recall that within-day dynamic traffic conditions refer to situations where
choice dimensions consist of route and departure time decisions, but no en-route rerouting
is considered.) The stochastic process is represented by an m-dependent Markov chain
where the current system state is influenced by the last m realizations of the process.
This model postulates that users make use of information about system states in
previous days and that the fractions of drivers choosing a certain departure-time and route
combination for each time interval can not be predicted but rather form a stochastic
process. The authors outline the sufficient conditions for the existence and uniqueness
of the steady state probability distributions.
,-
‘ay 5
The model divides the day into short discrete time intervals. Moreover, the
model groups together all drivers who leave the same origin at the same time and follow
the same route. If need be, a group may be split into smaller groups. It is assumed that
all members of a group experience the same travel conditions.
The model’s basic structure is as follows:
Step 1. Initialization.
Step 2. Go to next day.
Step 3. For that day:
a) Compute time-dependent path flows for that day from the demand
model.
b) Perform a network loading to compute time-dependent arc flows.
¢) Update memory & information on network performance.
Step 4. If test of convergence to stationary state is passed then stop; else go to step 2.
The demand model used in this framework is a nested logit model which is of the
utility-maximizing type. The basic utility function includes terms representing perceived
travel times by route and departure time as well as penalties for early or late arrivals
relative to a target arrival time at the destination.
As far as network performance and representation is concerned, two types of arcs
are modelled. Running arcs model link traversing while queuing arcs model deterministic
bottlenecks and include explicit capacity constraints.
Conditions on links are assumed to be homogeneous such that all users travelling
on an arc during a specific time interval experience the same traffic conditions.
Moreover, the locations of all groups are kept track of at all times with the length of a
group being ignored such that it is assumed to be concentrated at a single point. During
each time interval, groups are either moved along running arcs (and possibly onto the next
arcs on their paths) or spend the time in queuing arcs, or a combination of the two.
Given this flow network and performance modelling, the occurrence of counter-intuitive
results (such as vehicle overtaking) is precluded.
The derivation of time-dependent link flows from time-dependent path flows
requires the solution of a fixed point problem. This is due to the fact that the flow on
any arc during a specific time period is composed of the crossing fractions of different
O-D flows that leave their origins at earlier times and cross that arc during that time
interval. However, the fraction for group (k,j) depends on the arrival time of that group
at that arc as well as on the running time on that arc during that time period. These
running times in turn depend on the arc flows during the time interval being considered.
That is, the fixed point problem can be stated as:
age 4
Va ~1fh = > walt) F,
where:
f,° = the flow on running arc a during interval h
oY, = the crossing fraction; the fraction of arc flow f,, due to path flow E;
Ey, = the flow beginning to travel along path k during time interval j
This framework represents a generally acceptable formulation of the dynamic
traffic assignment problem in the case of no en-route diversion. Some of the
shortcomings of the framework are (i) The need for a more advanced information
provision/processing capability to model the availability of ADIS features; and (ii) The
use of logit in route choice may be problematic in the case of overlapping routes.
However, this is not a serious problem in this case since the logit model is applied only
to a selected set of "reasonable paths".
h) Stochastic Equilibrium Approach with Route and Departure Time Choices:
Another framework for analyzing the dynamic traffic assignment problem was
suggested by Ben-Akiva et al. (1984a, 1986) for simple networks and has recently been
extended by Vythoulkas (1989) to the case of a small but general network. The approach
incorporates a dynamic driver behavior model and a dynamic network performance model,
both of which are briefly described below.
This framework simulates stochastic equilibrium conditions whereby the route and
departure time decisions are depicted using a probabilistic path and departure time choice
model. In such a model a driver’s choice depends on perceived travel times (assumed to
be random variables) for alternative route and departure time choices. Moreover, a
penalty is placed on late or early arrivals at the destination relative to a target arrival
time. However, in the existing models it is assumed that a driver’s perceived travel time
for any route/departure time combination (on which he bases his travel choice for the next
day) is a function of the actual travel time experienced on the previous day. A relaxation
of this assumption is required if such a framework is to be useful in the context of driver
information systems.
This equilibrium-based approach to solving within-day dynamic assignment
problems simulates flows resulting from users’ choices until convergence is obtained; i.e.
flows computed in successive days are equal within some tolerance limits. This approach
consists of determining the system state in which no user can reduce his perceived
generalized travel cost by departing at a different time or choosing a different route.
However, researchers were unable to reach any general conclusions concerning the
existence or uniqueness of equilibrium flow vectors in this context. This issue remains
an open research question.
"age 75
The basic framework, as proposed by Ben-Akiva and de Palma and later
generalized by Vythoulkas, consists of two major components: (i) A travel time model;
and, (ii) A demand adjustment mechanism. The implementation of this framework
represents groups of vehicles moving along the network and the analysis period is divided
into short subintervals of time.
The travel time model takes as input origin-destination flow rates by path and by
time of day and consists of three modules. A traffic flow module computes time-varying
link inflows, outflows, occupancies, and densities based on link volume conservation
equations and assuming homogeneous link traffic conditions. Given link densities, time-
dependent link speeds (and link travel times) are next solved numerically using a BPR-
type speed-density relation. Finally, based on the time-dependent link travel times, a set
of "reasonable paths" is obtained, and the time-dependent O-D travel times are computed.
This last step involves the solution of a fixed point problem since the time a vehicle takes
to cross a link depends on the time it entered that link, which was not kept track of
during the assignment.
The demand adjustment model pre-specifies a fraction of drivers who will review
their previous day travel decisions. Out of these drivers, the fraction who will only
review their route decisions are also specified. The travel choices of these different
groups of drivers are then predicted using demand models. The more general demand
model takes as input time-dependent origin-destination travel times and estimates the
proportion of drivers selecting a certain travel pattern consisting of a combination of route
and departure time. The travel decision process is based on a mixed discrete (route
choice) / continuous (departure time choice) nested logit model under the assumption of
utility-maximizing drivers. Another demand model provides route choice predictions for
drivers who decide to keep the same departure time decision but review their route
decision. It is possible to emulate various driver behavior types by modifying the utility
function specification. For example, it is possible to emulate habitual behavior by adding
extra utility to the repeated choice of the previous day’s travel pattern.
The major advantage of this framework is that is presents a general structure for
the demand-supply interactions in dynamic traffic models which can be extended to deal
with ADIS environments. Moreover, as opposed to all the above mentioned models
(except Cascetta’s model), this model is an operational approach which deals with general
networks (considering Vythoulkas’ extension) and which has already been applied to a
real network.
The shortcomings of this model include a simple driver learning process, the
possibly problematic use of logit for route choice, and the absence of consistency
requirements between link speeds, occupancies, and exit flows in Vythoulkas’
formulation. The inconsistency results from the following sequence of calculations:
the speed by which flow travels on a link is determined at the time it enters the
7
“a6
link
» flow is assumed to maintain the same speed for the time it spends on a link
» thus, the time taken by flow to cross a link is determined at the time it enters
that link
» on the other hand, the outflow from a link is computed independently based on
the density and speed in effect at the time the outflow is computed
In conclusion, this framework is conceptually appealing for the analysis of within-
day dynamic traffic conditions, but its use in the context of driver information systems
requires several enhancements to the driver behavior and network performance models.
Summary of the State-of-the-Art
Even though the literature on approaches to dynamic traffic assignment is rich and
diverse, none of the proposed methodologies have been widely accepted or endorsed as
totally suitable for implementation within an ADIS context. The following areas
presented the most difficulties to researchers:
Representation of Guidance
In the methodologies discussed above the representation of possible guidance that
may be available to drivers is inadequate. Most methodologies were not intended to
function in an ADIS context so the availability of guidance to drivers was completely
ignored. In the other cases simplifying assumptions were made as to the type of guidance
that would be available to drivers, in some cases presuming that drivers are completely
informed about current traffic conditions all over the network.
Driver Behavior
With driver behavior in a dynamic context and with the availability of guidance
representing a situation that is fundamentally different from traditional ones, the
behavioral models included in the methodologies reviewed above were deficient in terms
of capturing several aspects of the behavioral process. Specifically, the following
deficiencies were observed in the different methods:
» adopting a system optimal assignment
» ignoring some dimensions of the driver choice process such as the departure time
and en-route diversion decisions
» improper modelling of route choice decisions
» simplified modelling of the driver’s acquisition and processing of usage of
information or guidance in an ADIS environment
Network Performance
Lo oo
= 44C I!
The computation of link flows from path flows is essential and non-trivial in the
DTA. For the DTA to predict correctly network performance, it should be able to depict
the movement of flow along the network from node to node. This requires the
formulation of link performance and flow conservation relations that are significantly
more complicated than is the case with static models.
Some of the models reviewed above suffered from flawed network performance
relations that resulted in inconsistencies in link volume, speed, and exit flow calculations.
Look-Ahead Feature
For the DTA to function as a COP module, it has to perform a truly dynamic
assignment of traffic that takes into account the effects of flow that is now at a certain
junction on traffic conditions at downstream locations in the future. Several of the
reviewed methodologies incorporated not-truly-dynamic assignment and/or simplified
look-ahead features by:
- repeatedly assigning vehicles based on current (not projected) traffic
conditions
using crude heuristics to project link volumes
Implementation Issues
While some of the reviewed methodologies were able to deal with many of the
issues described above, the implementation and modelling experience was limited to
simple network structures. The extension to more general networks presents tremendous
difficulties, especially as far as network performance is concerned.
In addition, a number of the methodologies did not incorporate any solution
algorithm and/or no computer implementation was described.
4.3 Proposed Approach and Elements of DTA
Some of the basic ideas related to the DTA approach being proposed here were
developed in previous research (Ben-Akiva et al., 1984a, Ben-Akiva et al., 1986 and
Ben-Akiva and de Palma, 1987). This work has recently been extended by Vythoulkas
(1989) to the case of a general network and a similar approach has also been suggested
by Cascetta and Cantarella (1991). All the referenced pieces of work were reviewed in
the previous section.
Since the DTA approach will serve as a COP module that will be used to predict
the effects that travel decisions by informed drivers may have on overall traffic conditions
(as discussed in Chapter 3), the approach should explicitly treat the distribution of traffic
Nn
Ee 78
by time-of-day and the drivers’ pre-trip and en-route adjustment processes. This requires
significant enhancements to the driver behavior and network performance models that
appeared in the works referenced above. Such enhancements are discussed in the
remainder of this chapter as well as in Chapter 5.
The proposed DTA approach incorporates the following items:
» representation of CAR features
» dynamic driver behavior modelling
» dynamic network performance modelling
At this point, we briefly discuss the first two items while the next section presents
the details of dynamic network performance modelling.
Representation of the CAR Features
The DTA must be able to represent the different guidance strategies that may be
implemented by an ADIS system. For instance, a CAR logic which uses information
provision or route directives has to be accommodated within the DTA, as is the case with
different temporal update frequencies. The proposed DTA approach is able to represent
the time-varying nature of routing information and incorporates dynamic interfacing with
the CAR, as was discussed in Chapter 3.
Dynamic Driver Behavior Modeling
A detailed discussion of dynamic driver behavior modelling for the DTA is
presented in Chapter 5. At this point it is sufficient to say that the DTA approach has to
capture the potential effects of the new information services on the departure time as well
as pre-trip and en-route path choices of individual drivers. Therefore, models of the
dynamic choices open to drivers with access to guidance should reflect the fashion in
which new information concerning traffic conditions may affect driver behavior.
4.4 Dynamic Network Performance Modeling
The function of the DTA is to provide CAR with projected travel times on various
links in the network. To perform this task, the DTA has to take into account the fact that
traffic flows and network capacities vary with time. Temporal variations in traffic flows
at various points of the network are caused by:
The time-varying nature of travel demand,
The response of drivers to anticipated traffic congestion,
The delays at upstream bottlenecks which affect the arrival times
at downstream facilities.
Se d wt 7
On the other hand, temporal variations in network capacities include primarily
changes in capacities caused by:
Incidents,
Traffic control actions.
Therefore, in a dynamic traffic assignment model all the following variables are
space and time dependent:
O-D trips
Link capacities
Link volumes
Link travel times
4.4.1 Flow Computation Difficulties
The drivers’ travel choices on a particular day (i.e. departure times and routes of
the drivers) translate into specific time-dependent path flows on the network. The
dynamic network performance model of the DTA, whose elements are discussed below,
has to determine time-dependent link flows and travel times that are consistent with
predicted path flows. For simple networks (such as a number of routes in parallel
between one O-D pair) this procedure is straightforward. However, for general networks,
the correspondence between arc flows and path flows in a dynamic network is not as
trivial as in the case of "steady state" networks where the two are simply related by a
link/path incidence matrix. In dynamic networks, flow on any link during a specific
interval is composed of path flows leaving their origin in that interval and/or in previous
intervals and traversing (totally or partially) the link during the specified time interval.
Moreover, due to congestion, the time required to cross any link during a specific time
interval depends on the link flow during that interval and, for oversaturated links, on link
flows in previous intervals.
Algorithms of varying complexity and level of detail could be used to deal with
this problem. These algorithms range from micro-simulation in which individual cars are
tracked through the network, to macro-simulation algorithms which extend the familiar
static network assignment algorithms to cope with time-varying demand and travel times.
Many of the algorithms formulated to compute link flows from path flows in the case of
within-day dynamic traffic conditions require the solution of a "fixed-point problem".
One such case was discussed earlier in this chapter whereby Cascetta and Cantarella
(1991) solved the fixed-point problem in their DTA model by using meso-simulation in
which groups of vehicles are tracked through the network.
Moreover, in the previous section it was indicated that some of the dynamic traffic
assignment models which have been suggested by researchers (for example, Vythoulkas,
1989) suffer from an inconsistency in the calculation of the link travel times and link exit
",
aes 30
flows. Moreover, it was shown that a model developed by Mahmassani and Herman
(1984) to simulate dynamic user equilibrium produces counter-intuitive results at the
microscopic level in the sense that drivers may actually leave later from their origin and
end up arriving earlier at their destination while using the same route. This indicates that
using the continuous fluid approach and traditional traffic flow relations may not work
in the dynamic case. Such inconsistencies are avoided within the proposed approach by
tracing the movement of "packets" along the network. This flow representation is
discussed below.
4.4.2 Dimensions of Dynamic Network Performance Models
The specification of dynamic network performance models has direct implications
on the soundness and computational requirements of the DTA. Some of the dimensions
of dynamic network performance models which should be considered in detail are the
following:
(a) Time representation (continuous or discrete)
The complex interactions between the various ADIS elements (including the
dynamic network performance module of the DTA being analyzed here) necessitates the
use of simulation to support the required system functions. Towards that end, a discrete
time representation of network performance parameters seems to be the way to go with
the specific size of the time-step being determined based on computational and accuracy
requirements.
(b) Flow representation (continuous flow, micro-level, or macro-level)
At this time, it appears that a continuous representation of flow is inappropriate
for the problem at hand since it does not allow for distinctions between en-route travel
decisions or information levels of different categories of drivers. Therefore, it is
suggested that flow be represented by "packets" whereby each packet consists of h
seconds of flow between a certain O-D pair. This approach was used in the CONTRAM
system (Leonard et. al., 1989; Taylor, 1990). By varying h, it is possible to have the flow
representation range from micro (at the extreme, a packet may represent a single vehicle)
to macro levels. If the DTA is run frequently enough (possibly every 30 seconds or less),
it may be possible to set h = t,,, - t; (refer to figure 3.1) since the packets in that case
would be small enough to yield the required accuracy in the DTA processing. If that is
not the case, however, it will not be advisable to use the above h as a basis for generating
packets. A major consideration in packet size determination is the required computational
effort. For real time application, it may be necessary to use large packets while packets
of smaller size may be used for off-line applications.
By adopting the packet approach, the path of each packet is traced as it progresses
through the network from its origin to its destination. Moreover, each packet will consist
Daga 81
of homogeneous vehicles. All vehicles in a packet are assumed to follow the same path
through the network. However, if a packet consists of more than one vehicle, members
of the packet may face different traffic conditions and therefore may have different travel
times. Therefore, the packet may stretch or squeeze based on the travel time differentials
between the first and last drivers in the packet.
Packets are expected to possess fixed as well as time-varying characteristics.
Among the fixed characteristics are the packet’s origin and intended destination, historical
information, and decision parameters. The time-varying characteristics include location,
speed, and current information.
[n this candidate approach to the DTA, it will be necessary to maintain the correct
position of packets on links during all time intervals. As such, at the end of each DTA
projection interval the positions of vehicles which had not completed their trips yet will
be stored as initial conditions for the next projection period. Moreover, no explicit
relationship for exit flows will be needed since these exit flows would be derived based
on moving the packets along the network at the prevailing speeds. Therefore, the possible
inconsistency that was referred to above would be avoided.
(c) Travel time computation (vehicle specific or link specific, function of total density on
link or density ahead on link)
The travel time modelling approach being envisioned for the DTA consists of two
types of arcs, namely, running arcs and queuing arcs. The running arcs model travel time
needed to cross the distance separating the end nodes of a specific link under prevailing
traffic conditions. Queuing arcs model the time vehicles have to spend in queue in order
to clear the junction at the end of the current link and move on to the next link. This
queuing time may be based on deterministic queuing models with allowance for stochastic
effects. Both running arc travel times as well as queuing arc travel times on any link are
packet-specific and may be computed at the time a packet enters the link.
4.4.3 Illustration
Consider the simple network shown in Figure 4.1 in which three O-D flows (F,,
F,, and F,) enter the network at different locations but are all destined to a downstream
node. The operation of the DTA model will proceed as follows:
» Packet Generation: Incoming flows are obtained from the dynamic, real-time O-D
updating procedure as described in section 4.5 below and are subdivided over discrete
subintervals of time. The incoming O-D flow during each subinterval of time is treated
as a "packet" of vehicles.
» Packet Behavior: For the incoming packet during a specific time interval a certain level
of information is assumed; for example, incoming vehicles may be aware only of
Dag. 32
historical traffic conditions and travel times on the alternate routes to their destinations.
Alternatively, an incoming packet at a specific node may receive guidance froman
information source. The specific guidance in effect at that time is set by the CAR module
and relayed to the DTA. Based on the available information and/or guidance, the choice
of an incoming packet among alternate routes is predicted using a dynamic driver
behavior model. For instance, referring to Figure 4.1, the choice of a packet in flow
stream F, between the two outgoing links will be predicted. A similar situation applies
to flow F, while flow F, has only one outgoing link that it can use to which all of the F,
packets are assigned.
» Packet Movements Along the Network: Based on the packets’ route choices that are
determined by the dynamic driver behavior model, the dynamic network performance
model is invoked to predict the travel times and queues on various links in the network.
This is achieved by moving the new incoming packets into the network and updating the
positions of packets already on the network links. More specifically, during the
subinterval of time being considered, a packet that has just entered a link will cover a
certain length of the link’s running arc based on prevailing densities and speeds on the
running arc, as determined by the presence of other packets on the arc. As such, the
position of this packet at the end of the subinterval is recorded as is the new arc
occupancy. On the other hand, other packets may cross the whole running arc and move
into the queuing arc where they join other packets already there. During each time
subinterval, queuing arcs discharge a certain number of vehicles to downstream links with
the choice among alternate routes being determined once again by an invocation of the
dynamic driver behavior model. Finally, some packets may reach their destinations
during a time subinterval at which instance their arrival time and total O-D travel times
are recorded and they are removed from the simulation.
This process of determining incoming flows, choice of route at nodes, and
movements of packets along the network is repeated sequentially for all time intervals of
the analysis period. Statistics relating to arrival times for packets associated with
different O-D pairs and departing at various times may be obtained as a result of the
application of the above DTA procedure. Moreover, by simulating the arrival of
incoming flows over a specific projection horizon, queue lengths in the queuing arcs as
well as densities on running arcs are predicted by the DTA for that horizon thus providing
for the required congestion prediction capabilities.
4.5 Self Calibration of the DTA
Referring back to Figure 3.1, a comparison of the flow information obtained from
the surveillance system with the flows originally predicted by the DTA helps identify
prediction errors for the self-calibration of the DTA models. The prediction errors may
be attributed to errors in the O-D matrix, to the performance functions used in predicting
travel times, or to inaccuracies in the driver behavior models that predict how drivers
Page 23
behave on the network. A statistical estimation process that will identify the most
probable sources of these prediction errors may then be applied. This process of
forecasting and observing impacts through the surveillance system is one of a number of
quite useful ways of adjusting the modelling system parameters.
The travel times and speeds obtained by the surveillance system are compared
with travel times predicted by the DTA. This comparison provides a facility for re-
calibration of performance functions initially used to predict travel times. This re-
calibration is expected to improve the accuracy of models over time by "learning" from
historical prediction errors.
It is to be noted that the statistical estimation process that will identify DTA
prediction errors and the re-calibration of the performance functions will be calculated on-
line in real time. The required frequency of such updates, however, remains to be seen
and will clearly be influenced by the available computational resources.
4.6 Updating the 3-Dimensional O-D Matrix
One set of inputs required by the dynamic traffic assignment models which are to
be constructed are average (historical) O-D flows by time slice. These O-D matrices
differ significantly from the matrices available for use in transportation planning analyses,
for example. The required matrices (representing dynamic flows) entail a much more
refined description of time-varying travel demand between sufficiently small zones that
represent the sources and sinks of the network. These dynamic O-D matrices may be
obtained from historically observed time-of-day flows using an off-line DTA. It is
conceivable that matrices representing various average flow trends (by day of week,
month of year, ...) be constructed and utilized as the basis for real-time updating.
It is important to note that data describing time-of-day variability in link flows or
in O-D flows is almost nonexistent at this time. In other words, the 3-D O-D data
required to implement the DTA is mostly unavailable and it would probably take some
time before a database including the required type and level of detail of data is
established. This process, however, is likely to be expedited by the substantial
improvements in communication and computing devices, and the resulting advances in
traffic surveillance equipment.
Real-time traffic flow data obtained from the surveillance system is combined with
the historical O-D matrix referred to above. The outcome of this process is an updated
3-D O-D matrix which is to be used in the next DTA projection period. The frequency
of such updates will depend on the capabilities of the surveillance system to provide new
extensive flow data as well as on the computational requirements of the updating process.
Although the focus here has been on the use of DTA for congestion prediction,
Dase 84
updated 3-D O-D matrix may also be used by other COP scenarios. For instance,
projections of future travel times may be obtained using heuristic methods that compare,
among other items, predicted O-D levels provided by the updated O-D matrix with current
or historic levels.
4.6.1 O-D Updating Methodologies
The problem of estimating O-D flows in real time based on historical and recent
measurements of traffic flows has received increased attention in recent years in view of
its applicability for real time driver information and traffic control systems. The problem
may be applied at the level of the network or individual junctions. Real time estimation
of O-D flows would require measurements of traffic flows at the entrances and/or exits
of intersections. It is also useful to have measurements of travel times between counting
locations. These measurements may obviously be obtained by the surveillance system.
The aim is to estimate the proportion of traffic that will go from each origin to each
destination.
The DTA Approach
The dynamic O-D updating procedure that we envision would involve an iterative
application of an off-line DTA. In this process a historical matrix will be adjusted to
reflect the latest measurements from the surveillance system. Thus, the development of
a real-time DTA will also benefit the process of dynamic O-D updating. This procedure
is analogous to methods of static O-D estimation using a static equilibrium assignment.
Unfortunately, there is not yet published literature on dynamic O-D estimation
with a DTA. The existing literature on dynamic O-D estimation considers simple
situations where the distributions of travel times between counting stations are known.
As such, in addition to unavailability of historical 3-D O-D data required to implement
the DTA, the analytical tools required to provide predictions and updates of the O-D
matrix have not been fully developed yet. The development of such tools will constitute
part of the overall research effort required to implement ADIS.
Methodologies without a DTA that have been proposed to obtain the dynamic O-D
estimates fall into two major categories, namely, least squares estimation and the Kalman
filtering procedure.
Least Squares Estimation
Variations on the least squares approach (Cremer and Keller, 1987; Nihan and
Davis, 1987; Kessaci et al., 1989) include constrained estimation whereby constraints are
imposed on permissible values of O-D flows to ensure that fitted values of O-D flows are
consistent with flow conservation and non-negativity constraints; discounted estimation
whereby recent observations are given more weight, a procedure which is necessary when
‘dpc 35
the proportions of flow destined from a specific network entry point to various exits vary
over time; and recursive least squares estimation to improve the estimation efficiency.
The Kalman Filtering Procedure
The Kalman filtering formulation is essentially a method developed for signal
processing which provides optimal estimates of the current state of a dynamical system.
It consists of two equations, namely, a transition equation and a measurement equation.
The transition equation describes how the set of state variables (the parameters to be
estimated) changes over time. The measurement equation(s) provides predicted values
of some observed variables based on the (estimated) state variables. A comparison of the
predicted and measured values of the observed variables provides a basis for updating
estimates of the state variables for the next time interval.
The application of the Kalman filtering approach to the problem at hand would
have O-D flows represented as state variables that are to be estimated while the observed
variables are the traffic counts. Note that this approach may be combined with a DTA
as follows: Dynamic traffic assignment, using the estimated O-D flows and playing the
role of the measurement equation, will provide predicted values of traffic counts at
specific stations. These predicted traffic counts are then compared with the actual
observed counts at the stations. The prediction error detected from this comparison of
predicted and observed traffic counts is used in the transition equation to estimate new
values of the O-D flows for the next time interval.
4.6.2 Illustration
To illustrate many of the ideas presented above, consider the simple network
shown in Figure 4.1. Three O-D flows (F,, F,, and F;) enter the network with all of them
heading to a single destination. Counting stations (e.g. loop detectors) are assumed to
exist at several locations in the network, such as points A, B, ... H. The methodology for
estimation of the O-D flows proposed here makes use of the Kalman filtering approach
in conjunction with a DTA module.
Starting with estimates of F,, F,, and F,, the DTA is invoked to load these flows
onto the network leading to estimates of flows crossing the various counting stations.
However, it is to be noted that the network flows resulting from the O-D flows will occur
at the counting stations at different times in the future. The DTA itself will determine
these lags for each station/O-D combination. In addition, the DTA will determine the
fractions of flows at stations that are attributable to each of the 3 O-D flows.
Once these predicted lagged flows at the various stations on the network are
obtained, they are compared with the actual observed flows at the stations as provided by
the Surveillance System loops. The prediction errors (difference between predicted and
observed flows at stations) are then used in the transition equations to provide updated
~
cd Ta
» 4
estimates of the O-D flows for the next time interval. Obviously, the transition equations
would also make use of the sensitivity of each counting station observation to each O-D
flow and to the network structure, using a properly defined structure of the transition
equations.
age 87
Figure 4.1 Sample Network
F2
yD
p—
1G
 Nn
pc
o
A 45»
Fi +9,
J 7
-
rs
cz’ E
age 88
CHAPTER 5 DYNAMIC DRIVER BEHAVIOR MODELLING
The purpose of Advanced Driver Information Systems (ADIS) is to provide
information on traffic conditions that will benefit individual drivers and reduce
congestion. In chapters 3 and 4 we argued that the effectiveness of ADIS as dynamic
traffic management schemes necessitates the use of a DTA model for congestion
prediction and that such a model should entail a proper understanding of potential driver
response to information provision. As such, one component of the envisioned DTA
should be a dynamic driver behavior model which is different from behavioral models
embedded in most of the existing traffic assignment models in which the driver’s
knowledge of the attributes of his available travel options is assumed to be invariant and
complete (see section 4.1). These traditional models are not valid when driver behavior
is to be analyzed in the context of dynamic traffic conditions and driver information
systems in which perceptions of attributes of various alternatives are continually updated
and revised. Therefore, the dynamic driver behavior model has to be capable of
analyzing driver behavior in an environment that includes changing traffic conditions as
well as information provision and modelling how drivers perceive attributes and their
behavioral response to travel experiences and information provision. Such a response
involves the acquisition and processing of new information, perception revision, and
adjustments to the daily and en-route driver behavior resulting from updated perceptions.
This capability is required to analyze the potential effects of new information on
individual drivers and on overall traffic conditions.
Objectives of This Chapter
The objectives of this chapter are the following:
» to review ways of modelling imperfect driver information and the manner in which
drivers integrate new information and update their perceptions; to review existing
models of travel behavior adjustment in response to acquired information
» to formulate a framework for dynamic driver behavior which describes acquisition and
processing of various types of traffic information and adjustments to travel behavior;
» to design a data collection program which will provide the data required to estimate
many of the component models of the framework.
Outline of This Chapter
To fulfill the objectives stated above, an overview of the issues involved in
dynamic driver behavior modeling is presented in section 5.1. Section 5.2 presents
various models which have been proposed for modelling imperfect driver knowledge and,
specifically, drivers’ perceptions of travel time. Section 5.3 describes some models of
perception updating and information integration in response to additional information.
Simple models of day-to-day and en-route adjustments in travel behavior are presented
Page 39
in section 5.4. Section 5.5 discusses how models of travel adjustment in response to
information may be enhanced by incorporating psychometric data. The proposed
framework of driver behavior in an information context and relevant models are outlined
in section 5.6. Section 5.7 discusses the implications of this framework on information
provision policies. Section 5.8 describes the data collection program required to gather
information necessary for estimating the proposed models. Finally, section 5.9 presents
some concluding remarks and suggests some directions for future research.
5.1 Overview
Driver Behavior Under Time-Dependent Network Conditions:
Departure Time and Route Choice
Driver choices that are modelled in a 2-D (static) network are usually limited to
route choice, but may also include choice of alternate destinations, alternative modes, and
the no travel option. In such a context, driver behavior is thought to consist of pre-trip
and en-route components. In the basic case where no information system exists to aid
drivers in their travel decision making, the pre-trip component consists of drivers choosing
a pre-planned route from among the available alternatives they know. This process may
be modelled using a route choice model similar to the one proposed by Ben-Akivaet.
al. (1984b) whereby chcice sets of "labelled" paths are defined. A labelled path is
defined as the optimal physical path with respect to some criterion function (e.g., travel
time, distance, congestion, traffic signalling, etc.). Once en-route, a driver may decide
either to follow his pre-planned route or to divert to an alternate route. The diversion
decision is likely to be triggered by the driver’s own observation of traffic conditions on
the part of his planned route that he has already covered. That is, if the driver detects
traffic conditions on his pre-planned route that are significantly worse than the conditions
he usually experiences, he may decide to divert. The dynamic driver behavior module
of the DTA has to model these driver decisions in order to anticipate their impact on
overall traffic conditions.
In a 3-D (time-dependent) network a traveller who anticipates changing travel
conditions over time also has a choice of departure time from the origin and a variety of
en-route choices that include the possibility of switching to alternate routes, making an
unscheduled stop, and changing the intended destination.
When modelling departure time and route choice, it can be assumed that drivers
are only concerned with travel time and schedule delay. Schedule delay is defined as the
difference between the desired and the actual arrival time. The essential idea introduced
by Vickrey (1969) is that there exists a trade-off between travel time and schedule delay.
In other words, a traveller may choose to incur a larger travel time in order to reduce
schedule delay or vice versa.
Page 90
Assume that the disutilities of travel time and schedule delay are piecewise linear
functions (see, for example, Ben-Akiva et al., 1986). Let o be the value of travel time,
B the value of one minute of early arrival and y the value of one minute of late arrival.
The driver incurs no schedule delay if he arrives at his destination during the interval of
time [t" - A, t* + A]. Let tt’(t) be the travel time on route r when departing from the
origin at time t. The utility for an individual leaving at time t and using route r is
V(r,t)= - aut’(t) - PMax[0,t"-A-t-tt'(t)] - yYMax[0,t+tf'(t)-t-A] - FF (5.1)
where F* is a fixed travel cost associated with route r. The driver tries to find the
departure time t° and the route r* which minimize the cost incurred:
V(t) = Max V(r,t) 5.2)
rt
This problem is very difficult to solve in real networks even in a deterministic
environment (that is even when the parameters of the road system and the total demand
are constant and known). There are three sources of difficulty:
(i) Suppose that the decisions of all individuals are known. A marginal driver
faces time-of-day dependent travel times on a given network. If B, v, and F are
all equal to zero, the problem is to find a minimum travel time (or generalized
cost which may consist of a linear function of travel time and distance) path in a
three-dimensional network. However, in the general case the calculation of the
optimal route and departure time is an NP-hard problem which is difficult to solve
(see de Palma et al., 1990).
(ii) The drivers’ choices are made given the other drivers’ choices.
(iii) In most situations of practical interest, the travel times on the different links
of the transportation network are not known with certainty. Travel conditions
change from day to day because of unpredictable changes in demand and incidents
on the road network. This uncertainty is associated mainly with:
capacity levels;
demand characteristics; and
differences in individual anticipation of traffic conditions.
The combination of the physical uncertainty and the drivers’ uncertainty gives rise
to a driving environment which is by nature very stochastic. This implies, in particular,
that drivers may respond to changing travel conditions by adjusting their travel patterns
over time.
Dynamic Driver Choices
Page 91
The dynamic nature of travel choices made by drivers reflects the fact that drivers’
experiences, observations, and other information concerning traffic conditions affect driver
behavior. This results in drivers adjusting their travel choices by modifying their route
and/or departure times on a specific day, as well as possibly diverting from an intended
route to an alternate route while on their way to their destinations.
When a Driver Information System is available, a central requirement of the DTA
is that it take into account the impact of any such guidance information on driver
decisions when projecting future traffic conditions. In such a case, a driver’s pre-trip
choices and en-route diversion decision will not depend solely on his own experience with
and observation of traffic conditions but will also be affected by information being
provided concerning downstream traffic conditions on the pre-planned route as well as
on alternate routes. Thus, the DTA has to include models of pre-trip and en-route driver
behavior which are sensitive to the drivers’ characteristics and the motorists’ access to
traffic information and route guidance. In this dissertation we are most interested in the
impact of ADIS on traveler behavior, and in modelling this impact through the dynamic
driver behavior module of the proposed DTA.
Information, Acquisition, and Integration
The dynamic driver behavior framework should incorporate the drivers’ acquisition
and processing of traffic information since it is needed to analyze the potential efiects of
new information on individual drivers and on overall traffic conditions. A driver’s
information state at any instant is represented by knowledge of the network and
perceptions of travel times on various routes at different times of the day. A driver’s
updated travel time perceptions are based on travel time experiences accumulated from
day to day and information received from ADIS sources. These updated perceptions
constitute the basis on which drivers perform pre-trip planning or en-route revisions of
their decisions.
5.2 Modelling Imperfect Driver Information
Driver Perceptions
A driver bases his choice of travel pattern on his perceptions of the values of
choice factors. A major assumption of deterministic network equilibrium models is that
drivers have perfect information in the sense that their perceptions of attributes of travel
alternatives perfectly reflect the actual values of such attributes and, consequently, that
they are able to identify and follow their "optimal" routes. This assumption of perfect
information may be acceptable for planning studies where static user-optimal equilibrium
simulates drivers’ behavior in the long run.
However, it is well known that a driver’s perceptions represent his imprecise view
of the actual attributes of travel alternatives. A driver’s perceptions are formulated on the
Dace |2
basis of a number of stimuli:
» his own experience
» driver information sources
» word of mouth
A number of factors result in perceptions being an imperfect reflection of actual
attribute values (Stern and Bovy, 1989):
» imperfect sensory pickup and/or mental processing of experienced travel
conditions
» lack of information concerning attribute changes
» stochastic nature of attributes of travel patterns
(due to incidents and interactions with other users of the system)
Therefore, dynamic models of driver behavior should take into account the fact
that drivers possess imperfect information especially when driver behavior is to be
analyzed in the context of ADIS.
A Taxonomy of Attribute Values
A better understanding of the problem at hand may be gained by taking a more
comprehensive look at the different values which may be associated with attributes (Ben-
Akiva and Kaysi, 1991). The overall relationships between the various values of travel
choice attributes are presented in Figure 5.1.
It is to be emphasized that while perceived attribute values constitute the basis for
travel decisions, they are latent, unobservable variables. This observation underscores the
difficulties that are encountered in modelling adjustments in driver travel behavior and
necessitates a closer examination of the relationship between perceived attribute values
and the two attribute values most closely related to it, namely, actual and reported
attribute values.
Relation Between Perceived and Actual Attribute Values
It is important to investigate the relationship between perceptions and actual values
for the following cases:
» travel pattern attributes
» differences in attribute values among travel alternatives
» changes in attribute values for a specific travel alternative
Research into this problem indicates that the relationship between perceived and
actual values of travel pattern attributes is not a simple one. A basic difficulty stems
ae 33
from the fact that perceived values are unobserved. Studies have also indicated that there
are threshold values for perceiving differences in attribute values. There is evidence that
individuals can only perceive differences when they are beyond a value referred to as the
"Just-Noticeable-Difference" level. This observation should have implications for drivers
choosing among alternatives with similar attributes.
Models of Stochastic User Perceptions
One model of the relationship between perceived and actual attribute values which
has been used widely in analyzing the performance of transportation networks is the
stochastic driver perceptions model. This model recognizes the fact that driver decisions
are based on perceived values of attributes (such as travel time) and assumes that drivers
possess imperfect information and thus imperfect perceptions of actual travel times.
Perceived travel times are modelled as random variables with t = t + € where t is the
actual travel time (on a specific link, for example), t* the perceived travel time, and € the
perception error. To operationalize this model, a specific probability distribution has to
be selected for the perception errors.
A recent study by Koutsopoulos and Lotan (1990) utilized this model of stochastic
travel time perceptions to investigate the effectiveness of motorist information systems
in reducing recurrent traffic congestion. In this study, perception errors are assumed to
be normally distributed and the distribution itself is assumed to be influenced by the
availability of information on traffic conditions. The exact relation is formalized through
the coefficient of variation B with B = 0 signifying perfect information and B = 0.5 (high
variance) a complete lack of information. The study simulated various scenarios by
varying the level of information (f), the % of informed drivers, and the spatial extent of
information. Conclusions reached included the need for better modelling of travel time
perceptions and the effects of information.
Even though this stochastic model explicitly recognizes the fact that perceived
values differ from actual values, it still assumes that the average perceived travel times
and the perception error distributions are the same for all drivers and that perceived travel
times are centered around the true values of travel time. Moreover, it is to be noted that
actual values of attributes are also unobservable, with measured values providing their
best available indicators.
Relation Between Reported and Actual Attribute Values
Some experimental studies have focussed instead on the relation between objective
(actual) time and estimated (reported) time. For short durations of time (less than 30
seconds) the relationship appears to be linear (Allen, 1979). For longer durations of time,
the "Power Law" (y=a.x") developed by Stevens (1967) has been used as a model of this
Page Ja
relationship but its fit is only marginally better than that of a linear relationship, if at all
(Allen, 1979; Leiser and Stern, 1988).
A different approach to modelling subjective time estimation is based on the
"Attention Allocation Theory" (Brown, 1985). The theory assumes that individuals
possess internal clocks which advance at a speed dependent on the amount of time
allocated to them. With attention being a finite resource, the amount of attention devoted
to concurrent tasks determines how fast the internal clock will advance. If tasks
concurrent with driving are demanding, the clock will advance slowly leading to an
underestimation of time.
Leiser and Stern (1988) developed a model which relates subjective estimates of
travel time while driving to three mediating factors, namely, the en-route obstacles (turns
and traffic lights), driving distance, and average driving speed. The study indicated the
existence of two kinds of obstacles: turns and traffic lights. Turns require increased
attention on the part of the driver thus resulting in underestimation of travel time. Traffic
signals, on the other hand, are "empty pauses” which allow the driver to devote more
attention to advancing his internal clock, and an overestimation of travel time ensues.
Driving distance was found to be directly related to subjective time since, for a given time
duration, a longer distance would involve a larger number of events which in turn results
in a higher estimate of travel time. Finally, the impact of driving speed on subjective
time estimates was found to be mixed since a higher speed requires more attention
(leading to underestimation) while involving more events (which result in overestimation).
Relation Between Perceived and Reported Attribute Values
One source of difficulty in analyzing the relationship between perceived and
reported attribute values is the existence of cognitive dissonance effects (Stern and Bovy,
1989). These effects reflect the observation that the traveller usually reports the attributes
of chosen and non-chosen travel patterns so as to rationalize his choice. Reported values
of choice factors are usually biased indicators of the perceptions. Therefore, any model
of this relationship should explicitly recognize the bias resulting from these cognitive
dissonance effects.
5.3 Perception Revision and Information Integration
As drivers go through travel experiences and gather information from various
driver information sources as well as other informal sources (such as word of mouth),
they tend to revise their perceptions by integrating the different pieces of information with
their previous perceptions. The process of perception revision is a complex one. In this
section we first present a simple model of perception revision and then evaluate the use
Deca 13
of Bayesian updating as a method of combining various sources of information and
analyze its suitability as a model of updating driver perceptions of travel times.
5.3.1 A Simple Information Integration Model
The following simple procedure represents a model of the way drivers process
information. In general, drivers have access to three generic sources of information on
any specific day w:
(a) Historical perceptions on day w of travel conditions on various routes I"(w). This
information may be updated every day.
(b) Exogenous information I¥(w) concerning events which are specific to day w and are
independent of drivers’ decisions (such as weather conditions or the occurrence of
accidents). The exogenous information consists of media reports and DIS inquiries which
are used to forecast travel conditions for that specific day and to construct the driver’s
current perceptions of travel conditions on the network I°(w) for that day.
(c) Personal experience IX (w) acquired by drivers choosing travel pattern p on day w.
Updating Historic Perceptions
The exogenous information in (b) relating to day w-1 as well as the experience
acquired on day w-1 in (c) will be used to update the historical travel time perceptions
at the end of day w-1 as follows:
[ I(w-1), BB(w-1), TX (w-1) |] => I'(w)
At the beginning of day w the driver constructs his updated historical perceptions
of travel timestt",(w) which constitute the individual’s best estimates of travel times on
a usual day for various travel patterns (r,t) consisting of route r and departure time t. His
update is a function of the previous historical perceptions tt",(w-1) and the information
acquired by the personal experience tt*.. .(w-1) associated with his chosen travel pattern
on day w-1 (r",t") or the acquired exogenous informationtt%(w-1) on day w-1. More
precisely, consider the following updating mechanisms:
te (W) = (1- @) thu w(w-1) + @* 5. u(w-1)
th (w) = (1- @©) tf, (w-1) + ¢° tf (w-1); (rp) = (1)  (5.3b)
D<g<l (5.3¢)
where g* and qF are positive parameters expressing, respectively, the impact of personal
experience and exogenous information on the individual’s travel time updating
Page 96
mechanism. The limiting case of g* =0 (or g© =0) corresponds to a situation where the
driver does not update his information from one day to another. Note that it is possible
that a driver gives different weights to the information received through various route
guidance and driver information systems; in such a case the value of the weight q° has
to be specific to the source of information.
Formulation of Current Perceptions for Day w
In a similar fashion, the driver uses the exogenous information to construct his
best forecast of the travel time on day w I(w). Therefore, [I*(w), IF(w)] => I(w).
At the beginning of day w the individual constructs a travel forecast representing
his current perceptions tt°, (w) for different travel patterns (r,t) which is a function of his
most recent historical perceptions tt", (w) and of the exogenous information tt", (Ww)
received at the beginning of day w (from media reports and D.L.S.), as follows:
tt (Ww) = (1- ¢) tt, (Ww) + q" tt® (w) (5.4)
where qF is a positive parameter which measures the impact of exogenous information
provided to the road users on the travel forecast made by the individuals. The currently
perceived travel times may be updated later as the driver moves towards his destination
based on other exogenous information that becomes available and based on the driver’s
own observations of traffic conditions.
Driver Inference
In route guidance systems where predictive information is not provided, drivers
have to project future traffic conditions based on information concerning current traffic
conditions. As such, the individual may infer the traffic conditions at time t (when he will
use the route) from the knowledge of the traffic conditions at time t’ (t'< t) when he
receives the information. Similarly, the individual who does not receive information on
a specific route has to infer this information from similar (e.g. spatially close) routes.
The specific way individuals perceive these temporal and spatial correlations is a topic
which remains to be explored.
More Realistic Models
It should be noted that the travel time updating mechanism presented here is a
simple one that assumes that travel time information is perceived by drivers through one
parameter such as the mean. A more realistic representation of travel time perceptions
may be through a range of values associated with a certain level of confidence on the part
of the driver. An advantage of this representation is that the range and confidence level
are equivalent to the specification of a mean and variance for a distribution of the
perceived travel times in which case the coefficient of variation may be used to compare
—-
 = 7
TART
the quality of information relating to different routes. In the next subsection we evaluate
the use of Bayesian updating as applied to travel time perceptions having a specific
probability distribution.
5.3.2 The Use of Bayesian Updating
One possible model of information integration is based on Bayesian Updating (see
the section on "Behavioral Decision Theory" in Rasmussen, 1986). This model starts off
by assuming a certain initial level of knowledge (prior information) and formulates the
final level of knowledge (posterior information) given the availability of new information.
The model may also be extended to the case where the levels of knowledge and the
available information are described by probability distribution functions.
Assuming that the prior probability distribution of perceptions concerning a certain
attribute of a travel pattern is normally distributed, the posterior distribution would also
be normally distributed in the case of Bayesian Updating if the probability distribution
of the new signal or new information is similarly distributed. The formulae which present
the new parameters are:
i 1 1
—— De —— — 15 4)
2 2 2
Oposterior Cprior info
1 = 1 =
_ Oo, 2 . prior 2 "ni
goer Ono (8.6)
1 1
ee
2 2
Oprior info
Equation (5.5) presents the relation between the variances of the average travel
times associated with the posterior perceptions, prior perceptions, and new information.
Equation (5.6) expresses the mean value of posterior travel time perceptions as a function
of the means and variances associated with the prior perceptions and new information.
As is apparent from the above formulae, the mean of the posterior perceptions is
affected by the means of the prior perceptions as well as the new information. However,
more weight is given to the source having a mean with a smaller variance since it would
indicate a higher level of confidence. One useful aspect of Bayesian updating is that it
can handle the specification of a level of confidence for each information source. This
would relate to the variance perceived to be associated with information obtained from
different sources.
Another characteristic of the Bayesian updating model is that
Page 98
2 < Z 5,73)
Oposterior O prior
Zz < - (£.7D)
OUposterior Oinfo
That is, new information always reduces in the variance of the mean travel time
from the prior to the posterior distributions, reflecting less uncertainty (and therefore more
confidence) regarding the value of mean travel time. This is so because it is assumed that
the mean travel time has a single, static value in which case additional observations will
provide a better basis for estimating this mean. Bayesian updating does not handle the
situation where the above may not be true, in which case new observations are related to
a new mean that old observations were not associated with. As will be discussed below,
day-to-day learning and day-specific learning may present such situations.
It should be noted that while new information always increases the confidence in
the average travel time estimate with Bayesian updating (as in equations 5.7a and 5.7b),
a traveler’s estimate of the average variance of the distribution of travel time is described
by a formula similar to equation 5.6 and therefore may actually increase or decrease. In
such a case the traveler may have more confidence in his estimate of the mean travel time
while fecling that the travel time is more variable.
Day-to-Day Learning
Day-to-day learning occurs by a process through which the driver updates his
historical (average) perceptions of attributes of various alternative travel patterns based
on his daily experience with his chosen alternatives. The update may also take into
account other information received daily concerning non-chosen alternatives.
The Bayesian updating model is applicable in the case of day-to-day learning if
the day-to-day average travel time has a "true", static value. In such a case, a driver’s
prior confidence about his average travel time perceptions is strengthened by repeated
observations which are consistent with the prior perceptions. This translates into a
posterior distribution of the mean perceived value that has a lower variance.
However, the above situation will not hold if significant structural changes occur
in network conditions, such as an expansion of capacity or an increase in demand
affecting specific sections of the network. In such a case, the new traffic conditions may
be sufficiently different that the traveler’s previous perceptions of travel times on these
sections of the network are no longer valid and have to be completely revised.
Day-Specific Learning
Na
ee Y“3
Day-specific perceptions are first formulated by combining historical perceptions
of average-day conditions with day-specific information which may initially be obtained
while the driver is still at home. Beyond that point, the driver receives en-route
information which is used to update his day-specific perceptions. This learning process
may be modelled using Bayesian updating since, on any specific day, traffic conditions
have one "true" actual value which the driver is being informed of. It is expected and
logical that as the driver progresses along his route, the variance in his average travel time
perceptions will decrease because of two factors: 1) more precise information will be
available concerning traffic patterns that are expected when he crosses various sections
of the network; and 2) the remaining portion of the trip is continually decreasing thus
reducing elements of uncertainty.
However, in assuming the existence of a true value which the driver is being
informed of, one would not be considering the possibility of certain occurrences or
incidents downstream which drastically change expected travel patterns thus rendering
some earlier information outdated and worthless. It may even be desirable to "subtract"
the effect of earlier information or formulate a totally new distribution of perceptions.
5.4 Existing Models of Travel Behavior Adjustment
in Response to Acquired Information
As drivers acquire and integrate new information, they may adjust their travel
behavior. Depending on the source and timing of newly acquired information, the
adjustment may be in either the day-to-day or en-route travel behavior. In what follows
we present examples of previous efforts at modelling these two types of adjustment. We
also review a recent study of drivers’ diversion decisions in response to delay due to its
relevancy to the topic being pursued.
Day-to-Day Adjustment
A simplified model of day-to-day adjustment was suggested by Ben-Akiva et al.
(1984) with later extensions by Vythoulkas (1989). This model assumes that a constant
fraction of drivers revise their daily travel decisions. Moreover, the model assumes that
individuals are informed about the distribution of actual travel times on day t-1 and that
they use this information in making their travel decisions on day t.
Another study by Horowitz (1984) assumed that drivers acquire their information
on travel costs from previous commuting experiences. The formulation of this
experience-gathering mechanism for a simple network with two links in parallel and
stochastic travel time perceptions was set up as follows:
where:
Page 100
1
Cp iY 8,=0 or t=1
k=1
t-1
(5.8)
on > 8, w,(DIC, +e,
otherwise
) 80
| tel
15.01
5 {4 iftraveller used link i on day t
it | 0 otherwise
w,(t): the weights given to previous experiences (k = 1,2....,t-1) when the driver makes
his route choice for day t,
measured cost of travel on link i for day t,
random variable representing the perception error,
weighted average perceived cost on link i at the time the driver has to make his
route choice for day t, and
the initial (day 0) perceived cost of travel on link i.
Since the travel decision for time period t depends on the average perceived costs
C, this formulation results in the probability of a traveler choosing a given link in a
particular time period being dependent on the traveller’s previous route choices. Due to
the complexity of this formulation even for a two-link case, simulation was resorted to
in operationalizing this experience-gathering mechanism.
En-Route Adjustment
Mahmassani and Jayakrishnan (1989) present a model of the path selection
decisions of individual motorists along their journey in response to supplied information.
The model is implemented in the context of a simplified corridor-like network structure.
In this model it is assumed that drivers with equipped vehicles know the travel time from
the present location to their destination on their current path as well as on the best path
based on current traffic conditions. It is assumed that an on-board computer receives and
analyzes information on prevailing traffic conditions and provides the relevant data to the
driver. The commuter route choice is assumed to follow boundedly-rational behavior
whereby a motorist switches from his current path only if the improvement in remaining
travel time exceeds some threshold level (or indifference band). Let TTC(k) and TTB (k)
be the travel times from node k to motorist j’s destination along the current and best
routes, respectively. The driver behavior can then be operationalized using the following
satisficing decision rule:
where:
Page 101
ITCf k) —- TTB ) > max(n. TTC1k»T) (5PNqr
3,(k) = {o Gr (k-
d,(k) is a binary indicator variable equal to 1 when user j switches from the current path
to the best alternate and 0 if the current path is maintained;
n; is the threshold level for user j, as a fraction of the remaining trip time on the current
path with 1; 2 0 Vj; and,
7, is an absolute minimum travel time improvement below which user j will not switch
routes.
In this model the indifference band is expressed as a percent improvement in
remaining trip time over the current path, while an absolute minimum level is maintained
to avoid switching when TTC;(k) becomes small as the driver nears the destination. In
the simulation study reported in this paper behavioral rules were applied at the individual
vehicle level.
Modelling Drivers’ Diversion Decisions in Response to Delay
Khattak et al. (1991) investigated short term driver diversion response to incident-
induced congestion delay and evaluated the ways in which drivers use real-time traffic
information. Work trip drivers destined to the central business district in the Chicago area
were intercepted at downtown parking garages during the AM rush hours. The survey
was conducted in April 1990 and a total of 660 questionnaires (8 pages, 112 questions)
were received. The questionnaire asked respondents several questions including:
» the number of times they travelled to downtown Chicago during the last month;
» their work start time and the start time flexibility;
e at what time was the selected route chosen;
» the average travel time to work by the best and alternate routes;
» the number of years the usual route has been used;
» the number of alternate routes used;
» whether they experienced en-route delays;
» the length of the experienced delay;
» the source of information on the delay;
» whether they diverted because of the delay;
» the savings/loss in travel time because of the diversion; and,
» the loss/savings in travel time had they diverted.
Survey results indicated that 42% of the respondents had diverted to an alternate
route in response to en-route delays. Moreover, 76% of those who diverted thought they
gained some time while about 50% of the respondents felt that they could have saved
time by diverting but did not do so. The reason for this latter figure lies in the fact that
drivers observe delay incrementally and may have realized its full extent at a time when
Page 102
diversion was no longer possible.
The survey data on respondents’ reported experience in the face of a recent delay
was used to estimate a multinomial logit diversion choice model. The alternatives in this
diversion decision were either to stay on the usual route or to divert to an alternative
route. Several models were estimated and they included the following variables:
» Characteristics of the delay experience (weather, trip direction, length of delay,
information source on delay)
» Attributes of "usual" and alternate routes (travel time, congestion, scenery,
reliability, neighborhood safety, stress experienced while driving, traffic stops,
overall rating)
» Trip characteristics (reported travel time on usual and alternate route, length of
time usual route has been used)
» Socioeconomic attributes (age, gender, income, location of residence, personality
factors)
Several models were tested with one being finally selected. The final model
indicated that there exists a preference for staying on the usual route and that information
has a greater impact on the decision to divert than direct observation . The reason for this
could be the fact that drivers have more options to divert at the time they get traffic
information on incident-induced congestion than when they observe congestion
themselves. Moreover, delay is observed in increments whereas traffic reports provide
a more comprehensive view of congestion. Mahmassani et. al. (1990) also found that
drivers who listen to radio traffic reports are more likely to divert.
The model also indicated a tendency to divert as length of delay and travel time
increase. As was expected, a strong perception of congestion on the alternate routes
reduces the diversion tendency while the number of alternative routes used has the
opposite effect. A Stated Preference Index which indicates a driver’s inherent tendency
to divert as well as an "adventure and discovery" personality index were found to be
significant in modelling drivers’ diversion decisions. Finally, suburban residents were
found to be less likely to divert.
Discussion
The models discussed above are helpful starting points in addressing the problem
at hand but suffer from a number of shortcomings. Specifically, the day-to-day
adjustment model suggested by Horowitz treats all individuals as equal in the weights
they attach to different past experiences and fails to specify a basis for determining these
weights. Moreover, it only deals with learning from experience and concentrates
specifically on day-to day adjustment. On the other hand, the adjustment model
suggested by Mahmassani and Jayakrishnan deals only with a very specific scenario of
en-route driver information provision and does not consider individual differences which
lage 103
may affect the diversion decision or the effect of the provided information and driver
experience on day-to-day travel adjustments. Finally, the model by Khattak et al.
represents an initial attempt at identifying how driver characteristics and attitudes may
influence their diversion decisions in response to delay when driver information systems
are available. However, the incorporation of the driver attitudes into the final "decision
to divert" model through variables in the utility function representing a "stated preference
index" and an "adventure and discovery index" could definitely be improved upon.
Next Steps
None of the models reviewed above purport to be a comprehensive model of the
impact of information availability on driver adjustment decisions. Need therefore exists
for a modelling effort which:
» takes into account the fact that drivers possess imperfect knowledge of
network attributes and traffic condition (section 5.2);
» presents reasonable models of information integration by drivers (section 5.3);
» properly incorporates driver perceptions, attitudes, and preferences (section 5.5);
» presents a unified framework of dynamic driver behavior in response to
information (section 5.6); and,
» considers the implications of the driver behavior framework on information
provision policies (section 5.7).
In addition, a data collection program which would provide the data necessary to
estimate the various models constituting the proposed framework has to be designed
(section 5.8).
5.5 Incorporating Psychometric Data
The incorporation of psychometric data in models of dynamic driver behavior and
adjustment in response to information may significantly improve the predictive power of
such models. The required psychometric data may include indicators of driver
perceptions and attitudes as well as stated preference data, as described below.
Perceptions
As discussed above, driver decisions in the context of information availability are
influenced by driver perceptions of various choice factors. At the most basic level, driver
travel decisions are based to a large degree on perceived travel times (as opposed to
actual travel times). As mentioned earlier, the modelling framework should treat these
perceived travel times as latent, unobservable variables and use reported travel times as
their indicators. Another example of the importance of perceptions is the impact of
perceived reliability of different information sources on the drivers’ decisions to consult
Page 104
these sources.
Attitudes
Driver attitudes also influence their travel adjustment decisions. For instance,
"habit following" and "adventure and discovery" are examples of attitudes which are
expected to impact the drivers’ propensity to adjust their travel behavior. These attitudes
are likely to be related to some socioeconomic characteristics of drivers, and through a
proper data collection program design one can obtain several indicators of such attitudes
which can help in estimating the required models.
Stated Preferences
Stated preference data which indicates the drivers’ weighing of various choice
factors, when coupled with revealed preference data, should provide a superior basis for
understanding, modelling, and estimating models of driver behavior. The power of stated
preference data lies in its ability to provide insights into some yet unimplemented options
as well as potential driver choices in other cases for which revealed preference data is
limited. Of course, the documented difficulties and biases with stated preference data all
have to be properly taken into account.
5.6 Integrated Framework for Modelling Dynamic Driver Behavior
In this section we present a proposed framework for dynamic driver behavior
modelling as well as outlines of the component models. The basic characteristics of this
framework are:
» the choice of travel pattern by drivers consists of a process rather than a simple act;
» drivers are active seekers and users of internal as well as external information; and,
» feedback from previous travel patterns affects later choices.
5.6.1 Dynamic Driver Behavior: The Basic Concepts
In this subsection, we present the basic concepts underlying the dynamic driver
behavior model, with special emphasis being placed on the way individuals process and
use different sources of information.
The framework describing the relationship between elements relevant to dynamic
driver behavior is illustrated in Figure 5.2. It takes into account the fact that generally an
individual has access to limited information (Newell and Simon, 1972), has a limited
capacity to process the information (Slovic, 1972), and attempts to find the best
alternative within time and effort constraints (Bettman, 1979; Heiner, 1983). A similar
framework has been discussed and applied by de Palma and Papageorgiou (1989) in the
Page 105
context of consumer behavior.
The approach adopted here for the analysis of dynamic driver behavior
characterizes the driver as interacting with his choice environment, seeking and acquiring
information from various sources, processing this information and then selecting from
among available alternative travel patterns (see Bettman, 1979). Since drivers’
acquisition, interpretation, and integration of information from various sources have
significant impacts on their travel choices, decisions regarding type, quantity, and methods
of information provision require an understanding of these cognitive processes affecting
driver behavior. Their implications for information provision policies are discussed in
section 5.7.
The main components of the drivers’ decision process are described below.
Goals: The driver’s objective is to arrive at his destination within a given period of time
while incurring the lowest possible cost. Costs include monetary costs (fuel, wear and
tear, tolls, etc.), travel time costs, and psychological costs (e.g., anxiety and late arrival
stress). A driver, based on his preferred weighing of the different cost parameters, aims
to minimize the overall sum of all such costs while best satisfying his objectives.
Information acquisition: Since drivers make travel choices to accomplish specific goals,
the associated motivation leads drivers to devote attention to acquiring available
information relevant to their goals. Drivers can acquire historical information either by
retrieving it from their own memory or by retrieving it from some database (collective
memory). Drivers can acquire current information either by direct observation or by
having access to information systems. The acquired information is evaluated given its
source and how it compares with previous knowledge. Predictive information is either
inferred by drivers based on historical and current information or is provided by an
information system.
Processing capacity: Drivers have differing abilities to combine and process a variety of
information concerning road conditions.
Computational ability: Drivers also have differing abilities to use available information,
perform travel forecasts, and develop heuristic decision procedures.
Decision rules: Driver behavior covers a wide range of possibilities. At one end of the
range lies habitual behavior whereby the driver does not seek new information. At the
next level are drivers who take into account some or all of the information available to
them when making their travel decisions. Finally, at the far end of the range of possible
behaviors are those drivers whose travel decisions are "strategic" in that they are based
on available information as well as the anticipation of how other drivers might react to
this information.
Page 106
The decision rules adopted by drivers are concerned with (1) information
acquisition and (2) the heuristics used by drivers to select their best alternative. For the
second type of decision rules, one possible class of rules is associated with drivers who
follow bounded rational decision rules and search among different alternatives until a
desired level of satisfaction is reached. The choice dynamics become more complicated
if the desired level of satisfaction changes over time and reflects personal experience.
Another class of rules may depict drivers as utility maximizers.
Reviewing: From one day to the next, some drivers are inflexible and do not change
their choices. However, given that the traffic conditions vary from day-to-day, other
drivers may consider other alternatives and modify their choices. They may also decide
to review their current decision rules.
Actual decision: The decisions made by the drivers are: choice of travel mode, departure
time, intended route and actual route.
After acquiring and processing information, drivers have to compare alternatives.
The limited processing capacity of drivers leads some to use simple heuristics such as
"Use the same route as yesterday”, "Use the route with the least number of traffic lights”,
or "Use the most reliable route". Other drivers may be able to weigh several choice
attributes and make a more balanced decision. The specific choice rules utilized depend
to a great degree on:
» individual differences in driver processing capacities;
» driver familiarity with choice situation; and,
» information availability.
Finally, the outcome of the actual travel choice provides additional information to
the driver and is expected to have an impact on future travel choices and, possibly,
decision processes (choice heuristics). The impact is likely to depend on how the choice
outcome is interpreted and the inference made concerning why it occurred. If the travel
experience was as expected, simplifications may take place in the travel choice process
(such as repeating the same choice if he feels satisfied with it or choosing the fastest
among a set of alternatives which are perceived to be similar). Another type of change
may be an elaboration in the travel decision process whereby the traveller learns from
outcomes to discriminate among alternatives by adding more criteria. These changes in
decision mechanics may also occur as a result of external information or interpersonal
contacts.
It is worth noting, however, that outcomes of travel choices are likely to be
weighed more heavily if they depart from expectations or if there was an a-priori goal of
observing the outcome.
5.6.2 The Hierarchy of Dynamic Driver Choices
Page 107
In the context of repetitive travel (commuting or repetitive shopping, for example),
the drivers’ dynamic behavior can be represented by a hierarchy of pre-trip and en-route
choices.
Pre-Trip Decisions
Every driver is assumed to possess historical perceptions of travel times on various
routes leading to his destination which are based on his day-to-day experience. When
making pre-trip decisions, drivers utilize these historical perceptions as well as any new
day-specific traffic information they acquire from various sources including media reports.
Pre-trip information may lead drivers either to continue following habitual travel choices
or to review their route and/or departure time choices.
The various steps in the individual’s pre-trip choice procedure as shown in figure
5.3 are described below.
(a) An individual may integrate the previous day experience and information with his
historical perceptions of travel time to form updated historic perceptions.
(b) Before leaving the origin of a trip, a driver may acquire and process new day-specific
traffic information based on various media reports. Individuals may have different ways
of processing and integrating information obtained in steps (a) and (b) to form a set of
current perceptions.
(c) Based on the new set of current perceptions, a driver may or may not decide to review
his former choice. A traveler who does not review his previous choice continues to use
the same route and departure time. For example, a driver may decide to review his
previous travel choice if the value of the utility expected from a new travel choice
exceeds the utility derived from the previous travel by a specific threshold.
(d) If the driver decided to review his travel pattern, he may choose to acquire and
process further information.
(e) A driver who chose to review his travel pattern decides upon an intended route and
a departure time given his new level of information. The decision on day w+1 may or
may not depend on the choice made on day w. In the latter case, habit formation and
transaction costs are ruled out. Optimal route and departure time choice is a complex task
in a congested transportation system and most probably drivers do resort to simplifying
procedures (see Ben-Akiva et al. 1984b).
En-Route Behavior
The next stage in the proposed framework is the en-route driver behavior. Once
a trip begins, a driver may receive new information (without actively searching) about the
Pace 108
quality of his travel choice and other choices. Based on this information and his
historical perceptions, a driver forms current perceptions regarding expected travel times
on possible routes to his destination and may decide to review his chosen route. The
driver may then actively decide to acquire and process new information and then make
a new route choice decision based on this information. This procedure is shown in figure
5.4 and is repeated until the destination is reached.
5.6.3 Components of the Modelling Framework
Figure 5.5 is a schematic representation of the elements involved in driver travel
choices in an information context. The "black box" in the middle represents the latent
processes which occur in a driver’s mind. The indicators shown in Figure 5.6 are needed
to gain a better understanding of the unobservable personal factors which affect the
various stages in driver behavior. These indicators relate directly to the psychometric data
discussed in the previous section.
Habit Following
This model classifies drivers as habit followers (who always follow their daily
habits without learning from previous experiences or consulting driver information
sources) and non-habit followers. The classification is expected to be based on some
socioeconomic factors (age, gender, level of education, ...) as well as driver attitudes
towards change.
Pre-Trip Decision to Revise Travel Pattern
Drivers who do not automatically follow their habits may frequently revise their
recent travel pattern decisions. A decision to revise the travel pattern does not commit
the driver to actually modifying his travel pattern: it merely relates to whether the driver
is considering such a possibility. Simon (1967) proposed that individuals possess a
scanner, or a mechanism for continually monitoring events occurring in their choice
environment and identifying instances which may require changes in current choices. In
addition, an interrupt mechanism stops implementation of current choices and initiates
responses to new conditions. That is, the scanner and interrupt mechanisms allow for
adaptation to changing conditions.
In the case of dynamic driver behavior, a basic force behind interrupts is a
departure of travel conditions from expectations. In other words, a driver may consider
revising his travel choices if the travel conditions he is experiencing differ significantly
from what he expected or is accustomed to. An important issue here is how much
deviation from expectations is necessary to cause an interrupt. Other causes for interrupts
may be what is referred to as "incidental learning" or the existence of conflicting pieces
of information. In such cases, drivers have to decide whether current travel choices still
make sense or must be revised.
Page 09
One way of modelling the driver’s pre-trip decision of whether to revise his travel
pattern may be by a system of equations with latent variables. The pre-trip decision to
revise the travel pattern is likely to depend on:
» the discrepancy between the driver’s current perceptions of the travel time on the
chosen travel pattern and his perceptions of an average day for that travel pattern
» the driver’s attitude towards trying new alternatives.
Moreover, it is assumed that some indicators of this "adventure" attitude would
be available. A simple, prototypical model may consist of the following set of equations:
Structural Equations
u*fw+l) = a, + €.A" + a,[tt*(w)-tt *(w) (5.11)
Dv (5372)A
Measurement Equations
1 ifu’ 20
d a (5.13)
otherwise
A = ANA" + eg (5.14)
tw) = Att (w) + €, (5.15)
(tH (ow ) —= A,tt"(w) + € (£79)4
where:
u*(w+1) = the perceived benefit from a pre-trip revision of the travel pattern on day w+1;
A’ = driverattitude towards trying new alternatives;
A = indicator(s) of A" (such as the inclination to try new alternatives, take risks to
avoid delays, and the fear of getting lost as in Khattak et al., 1991)
tt"(w)= the perceived travel time for the alternative chosen on day w based on the
travel experience and information associated with day w;
tt""(w)=the perceived average (historical) travel time for the alternative chosen on day w;
tt(w)= indicators of the travel time perceived on day w for the alternative chosen on
that day, including measured and/or reported values of travel time;
it(w)= indicators of the perceived average (historical) travel time for the alternative
dag, 0
chosen on day w;
v = vector of socioeconomic variables;
d = 1 if the driver revises his pre-trip travel pattern and 0 otherwise;
Ol, Ol, 0, B, and A,, A,, A, are arrays of the unknown parameters and e, &, €,, €,, and
g, are disturbance terms and vectors.
Equation 5.11 indicates that the benefit that a driver derives from a pre-trip
revision of his travel pattern on day w+1 depends on the discrepancy he perceives
between his usual (historic) travel time and his travel time on the previous day for the
alternative he took on that day. It is expected that if the travel time experienced on the
previous day had been significantly higher than the average experienced travel time for
that alternative, the driver will be unsatisfied with his previous day choice and will
perceive a potential for high benefits if he revises that choice. The equation indicates that
the driver’s attitude towards trying new alternatives will also figure into the decision to
revise the travel pattern. For instance, if a driver’s attitude inclines him strongly against
trying new alternatives and towards adhering to his habitual travel choices, this factor in
the equation will always involve negative perceived benefits from revising the travel
pattern. Equation 5.12 relates the latent attitude variable to measured socioeconomic
variables. The attitude towards trying new alternatives is likely to be affected by
socioeconomic characteristics such as age and gender. Equations 5.13 - 5.16 express the
measured indicators as functions of the latent variables.
Decision to Acquire New Day-Specific Information
In making travel choices, and especially after deciding to revise their travel
patterns, drivers examine relevant information in memory. The degree of this "internal
search" depends on (Bettman, 1979):
» amount of stored information
(prior experience, "spectator" learning, and individual differences);
» relevance and suitability of stored information
(inter-choice time & satisfaction with previous choices);
» degree of perceived conflict; and,
» stability of choice situation.
If the internal search is found to be insufficient or if several pieces of information
conflict drivers may acquire additional information from external sources. Drivers make
conscious decisions concerning whether to seek additional day-specific information from
various sources before they start their trip as well as during the trip itself. Models of
these decisions as well as the degree to which additional information is sought should
take into consideration a number of factors, namely:
» individual differences:
» use of pre-trip vs. en-route processing;
Page 111
» processing capabilities;
» concern with optimality of choice;
» choice environment factors:
» information availability and accessibility;
» time pressure;
current traffic conditions;
difficulty of choice situation
(information load and ease of processing/information provision format);
» perceived reliability of the different information sources.
degree of conflict between information sources;
perceived "costs" of obtaining information
(time, effort, money, decision delay, and psychological costs);
» perceived value of information in helping to make travel choice
(increased satisfaction with choice and psychological benefits);
The specific source of information that will be consulted in an external search is
expected to depend on:
» individual differences; and,
stage in choice process.
Information Iitegration
As information is acquired, it is actively processed and evaluated. Drivers
perceive the information they acquire as being characterized by varying levels of
reliability. The acquired information may be placed in memory, distorted, or even
ignored. The various processes affecting information processing and memory are quite
complicated.
A model of perception revisions and information integration by drivers exposed
0 various information sources has to:
» allow for increases and decreases in the variance of the perception distribution
as new information is accessed; and,
» take into account the discrepancy between prior perceptions and new information
when calculating the variance of the posterior perceptions.
In formulating a proper information integration model, it is probably useful to
obtain indicators of driver perceptions regarding the different weights and reliabilities they
attach to different information sources and relate these perceptions to specific
characteristics of the sources themselves.
Choice of Travel Pattern
Page 112
One model of the drivers’ pre-trip choice of travel pattern appears in Ben-Akiva
et al. (1986). This model, presented in section 5.1, indicates that the utility of route and
departure time choice depends on travel time as well as schedule delay. In the context
of dynamic driver behavior, such a model would be applied based on the drivers’ updated
perceptions.
En-Route Decision to Divert to Alternate Route
Different scenarios of information provision lead to different models of the en-
route diversion decision. At this point we will adopt a scenario similar to that suggested
by Mahmassani and Jayakrishnan as an example, and propose extensions to the model to
take into account individual decision factors.
The scenario being considered here relates to drivers receiving information, while
en-route to their destinations, regarding the remaining travel time on their current route
(TTC) as well as on their best alternative route (TTB). Let TTI = TTC - TTB denote the
travel time improvement offered by the alternate route. An improved model would
consist of a system of equations incorporating latent variables and would relate the
perceived utility of the decision to divert to:
» improvement in travel time due to diversion relative to remaining travel time on
current route;
» perceived reliability of information system providing guidance; and,
» driver’s attitude towards trying new alternatives.
While such a model may not be the final word on modelling en-route decisions
to divert, it serves the purpose of demonstrating that better explanatory power may be
gained by incorporating data on attitudes and perceptions in such models. A simple,
prototypical model may consist of the following system of equations with latent variables:
Structural Equations
4 3o + B(TTITIC) + B,R* + BA” + & (5.17)
X ] + & (5.18)
”~
A k - (5.19)
“7?
Measurement Equations
Me-3 13
g-Ql fu 20 [VeycM
otherwise
A mA") (5.21)
R Mmi(R™) (5.22)
where:
ut = the perceived benefit from en-route diversion to an alternate route;
R" = perceived reliability of information system providing the guidance;
R = indicators of perceived reliability of information system (e.g. degree of use of
system);
Z vector of observed characteristics of information system.
All other variables are similar to the ones that were previously defined in relation
to equations 5.11 - 5.16.
Equation 5.17 demonstrates that a driver’s perceived benefit from an en-route
decision to divert to an alternate route depends on the relative improvement in travel time
over his current route that the information system indicates will occur if he diverts
(TTI/TTC). That is, if the relative benefit is high, the driver will perceive that significant
benefits are likely to materialize if he diverts. The perceived benefits are also affected
by the driver’s perceived reliability of the information system: if he distrusts the system,
the factor in equation 5.15 related to the information system reliability will have the effect
of substantially reducing the perceived benefits from a diversion. In such a case, the
driver may not divert unless the reported benefits are quite substantial. The factor related
to the attitude towards change works in a manner similar to what was described for
equation 5.11. Equations 5.18 and 5.19 relate the latent perception and attitude variables
to measured causes. For instance, the perceived reliability of the information system may
be a function of the frequency of guidance update and the degree of sophistication of the
surveillance system used to support information provision. Equation 5.20 - 5.22 express
the measured indicators as functions of the latent variables.
5.7 Implications for Information Provision Policies
Based on the discussion presented in the previous sections, it is evident that driver
Information provision strategies have to be formulated recognizing the following factors:
Page 114
a) Drivers are characterized by having a limited processing capacity which leads many
of them to use heuristics and other techniques to simplify their travel decision making.
Therefore, the amount and complexity of information processing which drivers might be
able to perform should not be overestimated.
b) Individual differences in processing abilities as well as prior experiences play an
important role in determining driver response to information. For example, some drivers
may not perform en-route processing. As such, it seems that several methods of
information provision may need to be utilized to reach different types of drivers which
find specific sources of information to be easier and more effective.
c) It should not be assumed that more information is always better for drivers. In fact,
if too much information is available or if the information requires a lot of processing,
much of that information may be ignored. Therefore, the goal of information provision
should not merely be to make the information available, but rather to make sure that it
is usable.
d) The format in which driver information is presented influences the information
acquisition and processing tasks performed by drivers. Therefore, information should be
presented in formats which enable more effective processing by drivers.
It should be noted that the discussion presented in this section is not meant to be
comprehensive in its treatment of the human factors aspects of ADIS. Rather, it is
limited to issues that are directly related to the basic concepts of dynamic driver travel
behavior which were presented above.
5.8 Design of a Data Collection Program
In order to estimate models dealing with dynamic driver behavior in the context
of ADIS such as the ones described above, an extensive data collection program is
required. Moreover, the data collection effort will differ significantly from traditional
efforts for two reasons:
the focus is on dynamic driver behavior which dictates multiple observations
over time;
the necessity of real-time observation of driver behavior.
5.8.1 Potential Sources of Data on Dynamic Driver Behavior
The potential sources of data that may be useful for understanding and estimating
models of dynamic driver behavior fall into three categories:
i) Data from driver surveys: Surveys represent the traditional method of obtaining data
Jase 115
regarding driver behavior. With respect to dynamic driver behavior, traditional surveys
are useful in providing data on driver attitudes, perceptions and stated preferences as well
as information regarding the driver’s habitual travel patterns and reaction to information
provision (Khattak et al., 1991; Kitamura and Jovanis, 1991; Ygnace, 1991). If surveys
are to be used for analyzing dynamic driver behavior, they need to be augmented to
include questions that relate to a driver’s travel behavior over a number of days. Such
data will help us understand how drivers modify their travel behavior with time in
response to specific experiences or information sources. A survey that was conducted at
MIT in May 1991 illustrates the types of questions that may be asked in such surveys
(see Appendix A).
(ii) Data based on driving simulators: A number of recent studies have focussed on the
use of driving simulators to analyze driver behavior under various scenarios of
information provision (Allen et al., 1991; Bonsall and Parry, 1991). These studies put
drivers in simulated driving circumstances and observe their reactions and travel behavior.
Drivers would typically have in front of them a screen that presents a simplified map of
a real or imaginary road network with an assigned trip origin and destination appearing
on the map. Drivers have to "travel" between these two points, receiving guidance at
various points in their trip, and making route diversion decisions as appropriate. The
simulator would "observe" driver decisions and store such information for later analysis.
(iii) Data from demonstration projects: This last source of data is potentially the most
useful for analyzing driver behavior in the context of information provision. It would
involve documenting the behavior of drivers participating in demonstration projects, and
relating that to prevailing traffic conditions and guidance being provided. This would be
of great help in clarifying such issues as the relationships between driver compliance and
guidance validity and reliability. However, the difficulty lies in the fact that such projects
have as their primary aim the provision of guidance to drivers, and that requiring them
to observe, document, and store driver behavior associated with specific travel conditions
and guidance would impose significant additional data processing and storage needs.
5.8.2 Types of Data to be Collected
The basic focus of the data collection effort will be on the observation of dynamic
driver behavior under a variety of possible information transactions. Therefore, four basic
types of data need to be collected, namely:
» Information transactions performed by drivers;
» Driver decisions (including timing and location);
» Actual travel conditions experienced by drivers; and,
» Indicators of drivers’ perceptions, attitudes, and preferences.
Information Transactions
Page 116
The various sources of information which will be considered to be available to the
driver are the following:
» day-specific, pre-trip information;
» day-specific, en-route information; and,
» previous day experience (and, possibly, information).
If a survey is to be conducted to obtain such information, it should include explicit
questions concerning:
» whether the above-mentioned sources were consulted (i.e. information acquisition
decisions); and,
» the specific information that was relayed by these sources.
Under some information provision technologies, it may be possible to keep track
of the information that was relayed to drivers in which case it would not be necessary to
question drivers concerning information received. This would be desirable since a
possible source of error due to incorrect reporting by drivers of information received
would be eliminated. Such data could be obtained from demonstration projects or driving
simulators.
Concerning the drivers’ experience as a possible souice of information, it is
necessary to question drivers about their perceived experienced travel time on their chosen
travel pattern (as well as on non-chosen alternatives) at the end of each day of
observation. This would provide values of reported travel times which serve as indicators
of perceived values. Moreover, if data is available concerning the actual travel times for
the reported travel patterns (see discussion below), it would be possible to estimate a
model of the relations between perceived values on the one hand and each of the actual
and reported values on the other hand.
Driver Decisions
Driver decisions include initial pre-trip choices of route and departure time as well
as en-route diversion decisions. This information is to be obtained directly from drivers
through the questionnaire. It is conceivable that demonstration projects and future IVHS
systems will be able to provide this information automatically through two-way
communication between vehicles and the control center.
Actual Traffic Conditions
The collection of information regarding actual traffic conditions experienced by
drivers during their daily commute is advantageous since it provides a means of
understanding the relation between actual, perceived, and reported traffic conditions.
However, this data may be difficult to get hold of at the present time since the technology
of  EZ 117
is not in place to monitor traffic conditions continuously on a large scale basis. This
would leave one with the inferior option of relying solely on travel times as reported in
surveys by drivers themselves, and the analyst would have to take into account the
potential problems with these values (bias, cognitive dissonance). Obviously, driving
simulators are not of much help here since they can not reproduce the true distinction
between actual and perceived travel times.
Indicators of Driver Perceptions, Attitudes, and Preferences
As previously mentioned, reported attributes of alternatives (such as reported travel
times) need to be collected as part of a survey since they serve as indicators of perceived
attributes (latent variables). Moreover, indicators of relevant driver attitudes also need
to be collected. Relevant attitudes include:
» consultation of driver information sources;
» reliability of various information sources;
» learning from experience; and,
» changes in habits and trial of new alternatives.
Finally, stated preference questions should also be part of the survey as they would
indicate the drivers’ relative weighing of various choice situations. Stated preference
questions could relate to pre-trip decisions to revise the travel pattern, acquire
information, and the actual choice of travel pattern as well as the en-route decisions to
review the current route, acquire new information, divert, and the actual choice of
alternate route.
5.9 Concluding Remarks and Future Research Directions
This chapter has presented an overview of the issues involved in modelling
dynamic driver behavior in the context of information systems. It presented some existing
ways of modelling imperfect driver information and, specifically, travel time perceptions
and their relation to actual and reported travel times. The shortcomings of existing
information integration models and travel adjustment models were also discussed.
Requirements were suggested for a comprehensive modelling framework for dynamic
driver behavior, the need for incorporating psychometric data was clarified, and outlines
of the component models were presented. Finally, a general design of a data collection
program was suggested.
This analysis can be seen as a first attempt at analyzing the problem at hand.
Extensive research still needs to be performed as related to the following issues:
» modelling imperfect driver knowledge of traffic conditions
» modelling information integration and perceptions revision
Page 118
» specifying many of the component models of the framework
(especially the important information integration model)
specifying what types of data may be obtained from the various sources
identified above
identified how data obtained from the different sources can be combined in a
useful and sound manner
In addition, extensive work has to be done on uncovering the human factors
aspects of the various driver information technologies and how the workload associated
with each may affect driving habits and safety aspects.
As far as implementation is concerned, data has to be collected, and models need
to be estimated before the proposed framework can become operational.
>
2 119
Figure 5.1
Taxonomy of Attribute Values
Measured
values
 measurement
error
Actual
Values
perception
Historical
error
Choices;
Information o/ Perceived
Values
seme mamas pe"
bias;
cognitive
dissonance
Choice
"Reported
values
120
Figure 5.2 Drivers’ Decision Process
Goals
(from diff. sources
including memory)
(review —— b
| i Lo 1c 1 (acquire Info i
Reviewing | decision | Decision 1(acq
I rules) Rules | info) | Acquisition
~ Processing
Capacity
(review (combining
(choose info)
decision)
travel
pattern) Computational
Ability
(using )
Actual info /
prediction)
Decisions
FD
Figure 5.3 Pre-Trip Driver Behaviour
[ Historic
Perceptions Previous
of Network
(a) Day
Conditions exper /info
Current
Updated Historic| Media
Perceptions Reports
Current Perceptions
orion ©)
Previous
lravel No
Travel
Pattern Pattern
ir Yes
iin
iin
Co (d)
Acquird No Select
_ New DIS New
Travel
Nf (e
Yes Pattern
po
Figure 5.4 En-Route Driver Behaviour
Go to Next Intersection le—
Yes Reached ~~
Destination?.~
or No
| Current Perceptions |
Y eevvii ewNo 1
“ | + . 4
Route
Choice?"
w Yes
~~ .
BIE Yes
New D.I.S.
info ?
«
Process
\
3
| New Info
Updated Route Choice
i 29
Figure 5.5 Modelling Framework
-.
- Socio- :
Past | Actual
Economic Choices Conditions lL
Charact.|
- Info
sources
info
(Tearning)e Acquisition
(Attitudes) Perceptions}
(Preferences)
| Situational
Constraints,
Choice
7
Figure 5.6 Role of Psychometric Data
A ‘PerceptualILnLd itcuadtionrgs Indicators
Tearning)
 ~~
(Attitudes) (Perceptions
(Preferences)
Wr ap ap am gm
Situational
otated Constraints
Preferences
Choice
98
CHAPTER 6 CASE STUDY EXPERIMENTAL DESIGN
6.1 Objectives and Outline of Case Study
Chapters 6 and 7 focus on achieving the final objective of the dissertation, namely,
to conduct prototypical evaluation analyses of the proposed framework vis-a-vis other
information provision scenarios and under various conditions. For each of the scenarios,
an investigation of the interactions outlined in Chapter 3 between the system elements
(the surveillance system, COP, and CAR) is carried out.
Specifically, a case study is presented that
» evaluates the relative benefits that may be achieved by adopting the principles
underlying the proposed ADIS framework (refer to Chapter 3)
» analyzes the impact of other factors (described below) that are related to the
CAR and the DTA on the effectiveness of ADIS; in addition the sensitivity of
the results to demand and supply characteristics is evaluated
This chapter presents the details of the case study experimental design while
Chapter 7 analyzes the simulation results.
Outline of Experimental Factors
The case study will involve an analysis of the impact of the following
experimental factors:
Guided Probability: The relation between guidance effectiveness and the
probability of vehicles being guided is investigated in the case study to determine
whether adverse impacts of information materialize as the guided probability
increases.
CAR Factors: The case study will evaluate the impact of the following CAR
factors:
» CAR logic: approaches to maintaining consistency COP/CAR (Principle 3);
» Spatial update frequency: availability of re-guidance possibilities
» Temporal update frequency: frequency of guidance update
COP Factors: The impacts of three basic issues related to the COP module are
evaluated in the case study, namely:
Type of information provided by COP to CAR: the use of predicted traffic
conditions as opposed to other types of traffic information such as last
Pag_26
reported travel times or current measurements of queue lengths (Principle 1)
» COP methodology: the use of the DTA for congestion prediction with the ability
it incorporates to take into account potential driver response to guidance
(Principle 2)
DTA-specific issues: It was asserted in Chapters 3 and 4 that in order to
implement the DTA as a COP scheme, an O-D prediction module is required.
The case study analysis evaluates the impact of the quality of predictions
provided by the updating module on the guidance effectiveness. In addition, the
impact of the surveillance system’s ability to detect the occurrence of accidents
is studied as it relates to the performance of the DTA. This last issue is relevant
both under COP factors as well as under the role of the surveillance system.
Role of Surveillance System: The surveillance system acts primarily to provide
traffic information that will support the various COP schemes. The analysis will
identify the surveillance system capabilities that are required by each of these
schemes.
Demand and Supply Characteristics: Several factors extraneous to the
CAR/COP/Surveillance modules comprising the ADIS have an impact on the
effectiveness of the guidance being provided. Specifically, the role of the
following factors will be analyzed in the case study:
» O-D demand patterns including peaking extent and time
» day-to-day demand variability as related to the performance of the O-D updating
module required by the DTA
» demand level and the congestion intensity it imposes
» stochasticity of network capacity (e.g. accident frequency)
Tools for Case Study
Since the complexity of the problem precludes an analytical solution, the tool we
will use for this investigation is simulation of two small, prototypical networks relying
on an experimental design that addresses the relevant issues.
6.2 Elements of Case Study
Network Structure
Before proceeding to discuss the experimental factors associated with the case
study, we introduce the networks that will be used in the simulation. The two networks
are presented in figure 6.1 and have the following properties:
hoth networks consist of 4 real links (labelled 1 to 4);
age 127
each of these four links has a potential bottleneck at its end
» links DE and DF of network II are dummy links; the dummy links are included
to force flows p, and p, to use links 2 and 4, respectively
The performance functions associated with the links are described later in this
section. The implications of using these two networks are discussed in the next section
which describes the experimental factors.
O-D Flows
Associated with the networks are 3 O-D flows all of which are destined to node
G, as follows:
» ®(t) entering at node C
» p,(t) entering link 2 at node E
» 0,(t) entering link 4 at node F
It should be noted that flow ®(t) in network II can change route at point D which
provides it with 4 possible paths to its destination. The same is not true in the case of
network I where flow ®(t) has only two possible routes to the destination. Moreover,
flows p,(t) and p,(t) have no route choice as they have to follow links EG and FG,
respectively. These flows represent "external demands" that are loaded ontc the network
and affect its traffic conditions but are outside the scope of the guidance system.
Information Provision Scenario
The information provision scenario that will be used in the simulation model
consists of drivers receiving route guidance at specific points in the network. For network
[, guidance will be provided to vehicles at point C concerning which of the two routes
they should follow to their destination, point G. As far as network II is concerned, it is
assumed that guidance will be provided at both points C and D. The message at point
C will guide vehicles to the route they should follow between points C and D while the
message at D will be related to that portion of the trip between points D and G.
Driver Response to Guidance
The behavior of guided and unguided drivers has to be accommodated within the
case study since both categories of drivers affect the overall traffic conditions on the
network. The behavior of each of these categories is described below.
Guided Vehicles
It is assumed that drivers of guided vehicles will have a specific probability of
complying with the guidance. This probability of compliance, together with the percent
Page 128
of equipped vehicles, will determine the probability that any vehicle will be considered
to be "guided". This "guided probability" constitutes an important parameter of the case
study and can be interpreted as follows:
» probability vehicle is equipped * probability of compliance vehicle is equipped
All other vehicles are considered to be "unguided".
The guided probability may also be related to specific guidance technologies such
as Variable Message Signs (VMS) or in-vehicle units. ‘In these cases the guided
probability can be interpreted as follows:
» VMS: 1.0 (probability of receiving guidance) * probability of compliance
» in-vehicle units: probability that vehicle is equipped * probability of compliance
While these two technologies are treated as equivalent within the simulation, they
represent different information provision scenarios with potentially varying driver
responses. Therefore, in the simulation we will be mostly concerned with the generic
alternative and will not try to distinguish between the impacts of different technologies
on driver compliance.
Unguided Vehicles
In order to model the behavior of unguided drivers, a stochastic user equilibrium
solution was obtained for each of the two networks. Such a solution assumes that the
drivers’ perceptions of travel times on the different links in the network are random (see
Sheffi, 1985) and assigns the O-D demand among alternative routes in such a way that
the perceived travel time on the various routes between the O-D pair is equal.
The stochastic equilibrium conditions were obtained using average O-D demand
patterns. Route choices made by drivers are described by a logit model where the path
travel time represents the only factor in the driver’s disutility function. The scale
parameter which measures the variability in choice behavior among individuals was set
to 0.5, a value that reflects a relatively small perception variability. The simulation
algorithm used to obtain the stochastic equilibrium pattern is based on the method of
successive averages described in Sheffi (1985), as extended for dynamic traffic conditions.
The time-of-day-dependent patterns of path flows that correspond to stochastic
user equilibrium were obtained for each network. Based on these patterns, the fractions
of vehicles that follow each of the two routes leading out of node C (and node D in the
case of network II) for different times of the day were obtained. These "stochastic
equilibrium fractions" simulate the average behavior of drivers for a no-information,
average-conditions case.
Page 129
In the case study, drivers who are unguided at point C (or point D in network II)
are assumed to follow each of the two routes with a probability equal to the stochastic
equilibrium fractions. In other words, the unguided vehicles are assumed to distribute
themselves, on average, according to the same fractions as the ones that occur in the
stochastic equilibrium case. The rationale behind this assumption is that unguided drivers
are expected to follow their "normal behavior" which responds to average traffic
conditions and, therefore, can not adjust to day-specific random effects such as accidents
or demand variations. Moreover, such drivers are not expected to be able to discriminate
fully between traffic conditions that vary by time of day. In the case study it is assumed
that unguided drivers are only able to discriminate amongst three periods of the day
which experience significantly different traffic volumes (uncongested, congested, and
moderately congested). Therefore, the stochastic equilibrium probabilities are averaged
over these three periods of the day. These averages will be referred to as "stochastic
probabilities”.
Aggregate Driver Behavior
For traffic leaving C during a specific time interval t, and assuming the guidance
message indicates that drivers should follow route R1 (i.e. link 1), the probability that any
individual vehicle will follow route R1 is:
P:,(t) = Probability that vehicle is guided *
Probability of compliance (R1)|vehicle is guided
+ Probability that vehicle is unguided * |
Probability it follows R1|vehicle is unguided
— Guided Probability * 1.0 + (1 - Guided Probability) * SPg,(t)
where SPg,(t) is the stochastic probability that a driver will follow route R1 during time
interval t. The two terms in the equation refer to guided and unguided vehicles,
respectively. A "guided" vehicle will follow the guidance with probability 1 since the
compliance rate is already factored into the "guided probability”. For an unguided
vehicle, the probability that it follows route R1 is equal to the stochastic probability of
following R1 during that time interval.
Given that each vehicle departing C during time interval t has a probability Pg,(t)
of following route R1, the total number of vehicles following route R1 during that interval
of time may be represented by a binomial distribution. The reasonably large number of
cars leaving during each interval permitted the use of the normal approximation to the
binomial in the case study to determine the number of vehicles following each route
during that interval.
Page 130
Dynamic Network Performance
To implement the dynamic network performance model, flows on the network are
subdivided into packets with the size of each packet determined by the incoming flows
during each interval of time.
The movement of the first and last vehicles in each packet is simulated. The link
performance functions used to determine the travel times on links, or the time a vehicle
spends on a specific link, assume that each link consists of a running section that has a
constant free flow travel time and a bottleneck at the end (refer to figure 6.1). The travel
time on the running sections of the four links is assumed to be zero. As such, we will
consider delays at bottlenecks as the primary measure of travel disutility faced by traffic.
The implications of this assumption on the relative effectiveness of guidance scenarios
is discussed in Chapter 7. Moreover, each bottleneck is modelled as a deterministic
queue which possesses a specific discharge capacity. Queues in the bottlenecks build up
or dissipate based on the level of the incoming flows relative to the bottleneck discharge
capacity.
6.3 Experimental Factors
6.3.1 Guided Probability
The "guided probability” was defined in section 6.2. As the probability of a
vehicle being guided increases, the effect of the behavior of guided vehicles on overall
traffic conditions becomes more pronounced. Moreover, it was noted in Chapter 2 that
the adverse effects associated with information provision (concentration and overreaction)
occur as the guided probability increases. Therefore, it was essential that the impact of
this important factor be taken into consideration within the case study.
6.3.2 CAR Factors
Spatial Update Frequency
The network structures described in section 6.2 and shown in figure 6.1 were
designed so that the analysis can evaluate the impact of spatial update frequency on
guidance effectiveness. The availability of re-guidance possibilities along the route should
improve guidance quality since it allows revision of guidance advice at upstream locations
on the route. The downstream guidance locations may advise drivers to switch from a
route they were following to an alternate route leading to their destination. This
switching usually involves some excess time that the driver has to experience in order to
change routes. We refer to this excess time due to switching as the "re-guidance penalty".
The two networks being used in the case study embody two extremes of network
Nage [31
topology as it relates to spatial update frequency in that:
» Network I represents the case in which traffic that is guided at point C cannot
switch from the route selected at C. As such, no possibilities exist for adapting to
evolving traffic conditions or for recovering from mistakes that were made when guiding
vehicles at C. This is equivalent to having a "re-guidance penalty” which is very large.
 Network II depicts the case in which traffic that is guided at node C can be re-
guided at point D without any extra cost (such as that related to a cross over to an
alternate route). This situation is equivalent to having a "re-guidance penalty" of zero.
Temporal Update Frequency
As discussed in Chapter 3, the frequency of change of guidance message is an
important factor that the CAR may resort to in reducing the potential adverse effects of
information provision, thus potentially making the guidance more effective. It was also
pointed out that as the update frequency increases, shorter computational cycles are
required. This places heavier computational requirements on the system. The tradeoffs
between guidance effectiveness and computational requirements that come into play when
the update frequency is modified will be evaluated in the case study.
CAR Logic
The CAR logic that is used in the case study is based on providing drivers with
route directives. Moreover, it relies basically on shortest path guidance whereby all
traffic is guided to the route determined to be shortest according to the information
provided to the CAR in each of the scenarios that will be introduced below.
Referring to the discussion presented in Chapter 3, the following 3 approaches to
CAR logic are evaluated in the case study:
(i) Basic: The basic CAR logic provides guidance to a specific route whenever it
is determined that the travel time on that route is shorter than that on the alternate
route by even a very small amount. In the projective scenario, the COP/CAR
computational cycle to ensure consistency is terminated after a certain number of
iterations even if no convergence is achieved. In that case, shortest path guidance
is still provided knowing that it may not yield COP/CAR consistency.
(ii) Consistency Check: This approach can only be implemented with the
projective guidance scenario which attempts to reach COP/CAR consistency
through an iterative process (see section 6.5). If such consistency is not arrived
at after a maximum number of iterations, the CAR logic switches from single
route guidance to route distributive guidance or "no guidance", as was described
in Chapter 3.
Page 132
(iii) Guidance Threshold: This approach would compare travel times on the
alternate routes based on the information provided by the COP module. Guidance
is provided only if the travel time on one route is shorter than the alternate route
by more than a specific threshold. If the difference in travel time between the two
routes is smaller than the threshold, guided vehicles have to be distributed or no
guidance is provided and all vehicles are considered to be unguided. This
approach may be implemented with all the information provision scenarios used
in the case study whereby the difference in last reported, current, or projected
travel times on alternate routes are evaluated to determine whether or not it is
larger than the threshold.
Issues related to the implementation of distributive route guidance, consistency
check, and guidance threshold and ways of dealing with the "no guidance" case were
discussed in Chapter 3.
6.3.3 COP Factors
Types of Information Provided by COP to CAR
The four basic scenarios used in the case study to assess the impact of the
information provided by COP to CAR on guidance effectiveness are described below.
Scenario 0: No Guidance
This scenario serves as a baseline for measuring the benefits that may be achieved
by the three guidance scenarios. In this scenario all vehicles are treated as being
unguided.
Scenario 1: Guidance Based on Last Reported Travel Times
In this scenario it is assumed that the basis for guidance consists of what vehicles
themselves report to the control center concerning the travel times they experienced while
crossing specific links of the network. In this case, the surveillance system would consist
of the vehicles themselves as well as beacons installed at the entrance and exit to each
link. In this way, the beacon detects when a vehicle exits a specific link and relates that
to the time it entered that link to come up with estimates of the link travel time. In this
scenario, there is no real COP module since the surveillance data provided by vehicles
and beacons (the last reported travel time) are used directly to set guidance.
Scenario 2: Guidance Based on Instantaneous Information
In this case guidance is based on instantaneous information that is provided by the
surveillance system. The information needed to come up with instantaneous estimates of
travel times on the various links will vary based on the link performance model being
Page 133
used. For the case study, we require instantaneous measurements of queue lengths at
bottlenecks associated with all the links in the network. This may be provided by
detector loops or by an advanced video monitoring system. Again, no real congestion
prediction (COP) module is required here since surveillance data is used directly to set
guidance.
Scenario 3: Guidance Based on Projected Traffic Conditions
This scenario represents the strategy which requires explicit projection of traffic
conditions as a basis for setting guidance messages, in accordance with Principle 1.
Moreover, based on Principle 2, a full DTA is implemented in order to come up with the
necessary projections. The experimental design issues specific to the DTA are discussed
next.
DTA-Specific Issues
The quality of the travel time predictions provided by the DTA associated with
scenario 3 is affected by the availability of timely information regarding:
> current queue lengths on the various links in the network
(discussed in the next subsection)
It should be noted that while current queue lengths may be simulated using
the DTA, in practice this may not yield acceptable results due to the
possibility of cumulative errors in simulating such queue lengths. That is,
if the DTA is overestimating the incoming O-D flows over a period of
time, this causes a cumulative overestimation of the queue lengths on the
links affected by the incoming flows. If the DTA does not frequently self-
calibrate (see Chapter 3), the errors in the simulated queue lengths may
grow. Therefore, it is assumed here that the DTA will have access to
actual measurements of queue lengths instead of simulating such queues
itself.
» predictions of the evolving, time-dependent O-D flows ®(t), p,(t), and p,(t)
(obtained from an O-D updating module as described in Chapter 4)
reductions in capacity that result from accidents
(obtained from the surveillance system and based on an incident detection
capability)
The impact of each of the last two issues on the performance of the DTA is
discussed next.
Page 134
O-D Predictions
It was pointed out in Chapters 3 and 4 that in order to implement the DTA as a
COP scheme, an O-D updating module that provides predictions of demand levels should
be available. The quality of these O-D predictions is an experimental factor of
significance.
The O-D demands, while having average levels and a general pattern, could on
any day be significantly different from their average levels. The role of the O-D updating
module is to predict the demand levels that are expected to materialize at different times
on a specific day, given average flow trends and levels as well as flows that have already
been observed at specific locations on that day. This information is then provided to the
DTA to be used in projections of travel times. The case study evaluates the impact of
the quality of O-D predictions on guidance effectiveness by assuming that information
available to the DTA regarding future O-D demand ranges between two extremes as far
as real-time O-D updating capabilities are concerned, namely:
» the DTA is aware only of the average daily demand patterns;
i.e., no real-time O-D updating is involved
the DTA is informed of the actual realizations that will occur on a specific day;
i.e., a perfect real-time O-D updating mechanism is available
There will be more discussion about these O-D prediction cases in Chapter 7.
Incident Detection
The surveillance system will probably not be able to detect instantly the
occurrence of accidents. It is assumed that there exists a certain "delay" between the time
an accident occurs and the time the surveillance system and, subsequently, the DTA
become aware of the associated capacity reduction. The impact of such incident detection
delay as it relates to the performance of the DTA is evaluated in the case study.
Note that it is assumed that the system is always aware of the current queue
lengths at the different bottlenecks in the network. While current queue lengths may
indicate the occurrence of an accident, the incident detection capability referred to above
reflects the ability of the system to predict further buildup of queues due to reduced
capacity.
Predicted Driver Behavior
The DTA used for projecting traffic conditions requires a "prediction" of driver
behavior. As discussed above, the actual number of drivers who will use each route
during a specific time period is modelled using the normal approximation to the binomial
based on a random draw from this distribution. In the case study it is assumed that the
Page 135
DTA can not project the actual number of vehicles following each route. Instead, the
DTA assumes that the fractions that will use each route are identical with the probability
of an individual driver choosing such route (Pg,(t) for route R1). That is, in the DTA
prediction, the number of vehicles using each route is equal to the average value
associated with the normal distribution.
6.3.4 Role of the Surveillance System
The quality of guidance provided by each of the three guidance scenarios depends
on the availability of the following types of information from the surveillance system:
Scenario 1: This scenario requires information on the most recently experienced travel
time on each link of the network. Such information would in reality be obtained by
having the control center maintain a record of travel times that were reported to have
been most recently experienced on links in the network. Whenever a new packet exits
a link, the control center would update the "last reported travel time" entry for that
specific link and subsequently use the new entry in setting guidance. Since what is
required is simply the last reported travel times, it is expected that such information will
be readily available from the surveillance system without any significant measurement
errors. Therefore, the case study assumes that the CAR associated with this scenario will
have access to recent, accurate travel times on all links.
In the real world, a fast car may report a travel time that is shorter than was
typically experienced by other drivers, thus resulting in measurement errors. This
difficulty may be partially overcome if instead of using a single value for the reported
travel times, a number of recent values are weighted and utilized by the CAR. If
measurement errors persist, then the benefits of scenario 1 that are obtained from the case
study will be higher than what might be achievable in the real world.
Scenarios 2 and 3: These two scenarios require information concerning current queue
lengths on the various links in the network. While detector loops do not provide very
good estimates of queue lengths, more advanced surveillance devices (such as video
monitoring systems) are expected to perform significantly better. It is assumed that
scenarios 2 and 3 have access to accurate measurements of queue lengths and, therefore,
measurement errors are not taken into account in the case study. Since perfectly accurate
measurements may not be achieved in the real world, the benefits of these two scenarios
may be overstated in the case study results. However, the overstatement in the case of
scenario 3 is expected to be less significant since current queue lengths are only one input
used by COP and CAR while in scenario 2 they constitute the only input to CAR. Thus,
if there are measurement errors in the queue lengths, scenario 3 would not be affected as
much as scenario 2.
It should be noted that the surveillance system provides other inputs to scenario
3, as was discussed in subsection 6.3.3.
Mac 136
6.3.5 Demand and Supply Characteristics
There are several factors extraneous to the CAR/COP/Surveillance modules
comprising the ADIS which are expected to have an impact on the effectiveness of the
guidance being provided. In what follows we briefly discuss the role of three such
factors:
Demand Pattern
The case study will investigate whether different demand patterns will have an
impact on guidance effectiveness in general or on the comparative performance of the
various guidance scenarios. Specifically, the effect of peaking extent and time will be
evaluated.
Demand Variability
The performance of the O-D updating module required by the DTA is expected
to be influenced by the day-to-day variability of demand. Specifically, if the demand
does not show any variability from day to day, then a simple O-D updating module may
perform quite well. On the other hand, if demand is quite variable from day to day, a
simple module is expected to perform poorly, and it would be necessary to resort to a
more advanced O-D updating scheme. These issues are evaluated within the case study.
Demand Level
The impact of the level of O-D demands will be investigated in order to identify
whether congestion intensity results in differential performance by the various guidance
scenarios.
Stochasticity of Network Capacity
It has been argued that the effectiveness of real-time driver information systems
will be most pronounced in the case of highly stochastic network capacities. In the case
study, the impact of accident frequency on guidance effectiveness will be analyzed.
6.4 Measures of Effectiveness
In order to evaluate the effectiveness of guidance scenarios under various
operating conditions and guidance parameters, measures of effectiveness are required.
The following measures were selected:
Page 137
Average Travel Times
The average travel time of all drivers who travel between C and G during the
analysis period is used as one measure of effectiveness of the guidance scenarios
compared to each other and to the no guidance case. This constitutes a measure of the
overall impacts of guidance on the average travel time of all traffic departing from C,
both guided and unguided. In addition, the average travel times experienced by guided
and unguided vehicles separately will be evaluated.
The average travel time for the external demands p2 and p4 is also provided for
the different guidance scenarios. This will identify the impact of information provision
on travel times experienced by traffic that is not subject to guidance.
Guidance Validity
The percent of drivers who are "correctly guided" conveys a measure of how valid
the guidance was. Specifically, this measure indicates the percent of guided drivers who
experienced travel times that were shorter than the ones experienced by drivers who left
during the same time interval but followed the other route to the destination. The
importance of this measure lies in the fact that driver compliance with guidance is
expected to depend to a large degree on whether the guidance they receive is valid or not,
in the sense indicated above.
In the case where the CAR logic includes a consistency check or the use of a
guidance threshold, all vehicles that leave during a "distributive guidance" or "no
guidance" time interval are considered to be unguided. As such, the travel times
experienced by these vehicles do not enter into the calculation of guidance validity.
Travel Time Reliability
The ability of ADIS to provide guidance that will allow guided vehicles to avoid
long delays is expected to be one of the major advantages of such systems. The case
study evaluates this by using the following two measures:
(a) the maximum travel time experienced by guided and unguided vehicles
(b) the % of guided vehicles experiencing travel times that are significantly
longer than the average
These two measures for guided vehicles are compared across guidance scenarios.
In addition, a comparison of these measures for guided vehicles with similar measures in
the case of no guidance is provided as part of the case study results.
Page 138
6.5 Structure of the Simulation
General Simulation Structure
The simulation model consists of three major modules:
» data input
» simulation of guidance, driver behavior, and network performance over the
time period of interest
» computing performance measures
The simulation structure is presented in figure 6.2 and proceeds as follows (the
descriptions provided relate to network I; however, the simulation for network II proceeds
in a similar manner):
Step 0. Data input
Step 1. Set t=0; initialization
Step 2. Invoke the COP/CAR and set route directives based on data available from
the surveillance system; this step is detailed below under "COP/CAR"
schemes
Step 3. Invoke the driver behavior model to determine @,(t) and @,(t) (the flows
from C using links 1 and 3) based on ®(t), the route directive that was
set in step 2, and Pg,(t) as defined in section 6.2 above; the flows on each
link will consist of two "packets", one for guided and one for unguided
vehicles
Step 4. Invoke the network performance model to determine the actual traffic
conditions (including queue lengths at bottlenecks and travel times
associated with the new packets) that will be in effect based on @(t)
and ®,(t) as well as p, and p,;
Step 5. If the current time interval belongs to analysis period, collect performance
statistics (including the travel time experienced by flow constituting ®,(t)
and @,(t))
Step 6. If the current interval is the last one in simulation period stop and compute
summary statistics;
else set t=t+1
Check the temporal update frequency:
If guidance has to be updated go to step 2;
else let new route directive be the same as the old one and go to step 3
Step 0 of the simulation model involves input of data related to the various
processes and parameters that come into play. A detailed description of the data which
is input is presented in the next chapter.
The driver behavior and network performance models used in steps 3 and 4,
Daca -30
respectively, were described in section 6.2. It should be noted that the network
performance model used to simulate the movement of packets associated with actual
traffic conditions is the same as the one used in the DTA.
For each time interval within the analysis period, performance measures are
collected in step 5. Once the simulation for the relevant period is completed, summary
statistics are computed in step 6. These measures are used to evaluate the performance
of the guidance strategy in effect given specific operating conditions and guidance
parameters.
COPI/CAR Schemes
The generic function performed by COP/CAR (referred to as step 2 in the
simulation structure outlined above) for a specific time interval in the simulation is
presented in figure 6.3. It indicates that the COP/CAR schemes in general accept some
data from the surveillance system which is processed inside the COP/CAR box. The
output from the box consists of route directives.
Figure 6.4 presents the COP/CAR functions that are performed with the DTA as
a COP module. For a specific time interval, the surveillance system provides the
COP/CAR box with information concerning current queue lengths which is used to set
the initial route directives. Next, the DTA’s driver behavior module receives this route
directive as input. It also obtains predictions of O-D demand ®(t) from the O-D updating
module. These inputs are used to predict driver behavior and the resulting values of ®,(t)
and ®,(t). Next, the network performance model uses the flows predicted by the behavior
model (P,(t) and D,(t)), the current queue lengths, and the flows predicted by the O-D
updating module (p, and p,) to project the travel time that will be experienced by each
packet that leaves C at time t. As indicated in Chapter 4, this will be computed as the
average of the travel time of the first and last vehicles in the packet. CAR uses these
projected travel times and sets the route directive to the shortest path.
The system then checks for consistency: is the new route directive the same as the
old one? That is, are the packets’ projected travel times (based on which the new route
directive was set) consistent with the old route directive? If so, the route directive is
provided to drivers and the COP/CAR cycle is terminated for this time interval.
Otherwise, a new COP/CAR cycle is initiated based on the new route directive.
Conceptually, the process has to be repeated until consistency is achieved in which case
single route guidance would be provided.
However, as noted in Chapter 3, convergence may not be achieved especially since
single route guidance is being used. In this case it becomes necessary to interrupt the
process after a certain number of iterations. For the networks used in this case study, and
since shortest route guidance can only be set to two possible routes, three iterations are
sufficient. Convergence is assured whenever two consecutive iterations yield the same
Page 140
guidance advice. The lack of convergence would imply that the control center is not
capable of managing and mitigating the effects of the potential overreaction that will
occur due to provision of guidance. After the process is interrupted, the outcome depends
on whether the "basic" or "consistency check” CAR logic is being used.
The "basic" logic would provide single route guidance (based on the directive
reached during the last iteration) despite the fact that consistency cannot be achieved.
The "consistency check" logic, detecting the projection/guidance inconsistency after the
process is interrupted, would make sure that guided vehicles are distributed over the two
alternate routes. The implementation of this guidance logic was discussed in Chapter 3.
It is worth noting that in the specific implementation used in the case study, the
COP/CAR cycle takes into account only the first packets that leave following the
provision of guidance and use each of the alternate routes. That is, the projections which
take into consideration potential driver response to guidance and the process by which
consistency is ensured relate only to packets leaving during the first time interval. This
is not an issue if guidance is updated every time interval in which case every packet
would be taken into account. However, if guidance is updated every 10 intervals, the
behavior of the last nine packets in each guidance interval is not considered by
COP/CAR. Instead, the guidance for all ten intervals is set by projecting travel times
(and ensuring consistency) for the first packet only. In other words, depending on the
temporal update frequency, the projection horizon may not include all the flow that is
affected by the guidance to be set. Similarly, the "guidance threshold" approach provides
single route guidance if the predicted difference in travel times for the first packets
leaving following the provision of guidance and taking each of the alternate routes is
larger than the threshold value. Otherwise, guided vehicles would have to be distributed
over the alternate routes. This approach is meant to guard against the occurrence of
inconsistency for the packets that leave later during the guidance interval.
The advantage of this short-horizon approach is the reduced computational effort
required to implement projective route guidance since only the movement of the first
packet has to be modelled, but it is relatively myopic if the guidance interval is long.
Obviously, an implementation with a longer projection horizon is possible and is likely
to yield improved results for the projective guidance scenario, but would require a
significant increase in computational effort. For instance, one possibility with a long
guidance interval would be to project the travel times experienced by the first vehicle in
each packet that leaves during the guidance interval, based on a certain guidance message.
The first packet for which an inconsistency is predicted to occur would trigger the CAR
logic to switch to another message for the rest of the guidance interval.
6.6 The Efficiency of Scenarios 2 and 3: First Insights
Before presenting the simulation results in the next chapter, it is interesting to
Page 141
compare the guidance bases described in scenarios 2 and 3. To start with, the guidance
in scenario 2 is based solely on observation of current traffic conditions. This approach
has several advantages. First, it requires no projections of traffic conditions, thus no
uncertainty enters into the generation of guidance messages. In addition, the
computational effort required to implement this scenario is not substantial. However,
current traffic conditions may not be good predictors of future conditions. This renders
the guidance basis in this scenario suspect and crude which is its major shortcoming.
On the other hand, the predictive approach has the advantage of basing the
guidance messages on expected traffic conditions, which should provide a more plausible
and more reliable basis for guidance. However, this method has several disadvantages
including:
» potentially significant surveillance requirements to provide a basis for
O-D updating: traffic counts on many links throughout the network are
needed (see section 4.6)
potentially significant computational requirements due to the need for:
calculating the updated O-D demand levels, given the input from
surveillance
travel time projections using a potentially complicated DTA
ensuring consistency between guidance and projected traffic conditions
» reduced guidance effectiveness due to inaccuracies in predicting O-D flows
In summary, it seems that the relative value of using predictive information as
opposed to instantaneous information as the guidance basis may depend on the quality of
the available predictors of O-D levels. Even if this proves to be a non-issue, a
determination still has to be made as to whether the expected superiority of scenario 3 is
worth the extra computational effort and hardware required to implement it.
Page 142
Fig. 6.1 The Networks
~ho?2
5LT
f-
A,
~hi ( .
AN
a
all
f F
Network |
rhn4d4
rho2
a
—
» i
Tyr
Ao
wi C P
ol
a RS
+—
phi
"1
et,
—
Lo
1
Network II
rho4
| : Bottleneck
A
g. 5.2 Simulation Structure
Input Data
Step |
Yes A—
CCP/CAR : Surv. Data
Update Guidance? _
. Step 7 Step 2
. eee
Route Directives
Dr ver Behavior
Cs i pht
Ste
re oeii—————
phi!, pm3
Network Performance
1 rho2, rho4
Step 4
Step
+ — In Analysis |
Actual Period ?
Traffic Concitions
Yes al
NO Ena of
mutation Pertod? Step 6
Corrpute Perf. Stats.
Summary Statistics for Time Perand
ly
ig. 6.3 Generic COP/CAR Functions
Surveillance Data
COP/CAR
Route
Directives
 a
Fig. 6.4 COP/CAR with a DTA
Surveillance Data:
Current Traffic Cond.
 initialization
C AK 1
i
No New Route
0-D or Driver YDirective =Prediction
Behavior Old ?
phit
phi3
rho4| rho2 Network
Serformance
Proj. Traffic
Conditions
CAR
| 4G
CHAPTER 7 ANALYSIS OF SIMULATION RESULTS
7.1 Base Case Input Data
Simulation and Analysis Periods
A model run includes a simulation period over which guidance and traffic
conditions are simulated and an analysis period for which performance measures are
collected and eventually aggregated. The analysis period extends from 7 to 10 AM, the
period of time during which typical AM peak period traffic conditions are experienced.
The simulation period extends between 5 and 10 AM with the initial two hours
representing non-congested traffic conditions that are used to load the network.
Simulation Interval and Packet Size
As pointed out in Chapter 4, flow in the dynamic network performance model is
represented by "packets" whereby each packet consists of h seconds of flow following a
given path between a certain O-D pair. One such packet can be thought of as consisting
of two subpackets of guided and unguided vehicles. By varying h, it is possible to have
the flow representation range from micro (at the extreme, a packet may represent a single
vehicle) to macro levels. A major consideration in packet size determination is the
required computational effort. In the case study a simulation interval of 30 seconds was
used so that a packet represents 30 seconds worth of incoming flows. The use of 30
seconds provided a reasonable execution time while maintaining a reasonable level of
detail in flow representation.
Network Capacities
The network structures used for simulation were introduced in Chapter 6 with each
link having a bottleneck at its end. The discharge capacities associated with each of the
four bottlenecks are as follows:
s, = 2000 vehicles/hour
s, = 4000 vehicles/hour
5, = 4000 vehicles/hour
5, = 6000 vehicles/hour
O-D Demands
Figures 7.1a and 7.1b present the two demand patterns that were used in the case
study. The average demand levels are shown for ®, p,, and p, for each 5 minute interval.
The O-D demand patterns associated with p,(t), and p,(t) shown in figure 7.1a are
relatively FLAT and spread out over time, with their peaks occurring at roughly the same
Tage 147
time, which is about 15 minutes later than @’s peak. The alternate p,(t) and p,(t) demand
patterns shown in figure 7.1b differ in that the peaking patterns are SHARP and more
pronounced and occur at different times.
While figures 7.1a and 7.1b represent average demand levels, specific realizations
of the O-D demand levels for each time interval for each day being simulated are
obtained by assuming the following:
» for the whole day, there exists a "daily factor" which represents a systematic
variation around the mean for all three O-D flows
» for each time period, an actual realization of the O-D demand for that day is
obtained by assuming a normal distribution around the mean (as modified
by the daily factor)
The actual daily realizations described above are generated based on the following
DIroCess:
p y= tr & I  ) of 3 dD
where:
Pq = the incoming flow during time interval i on day d
Pp, = the average incoming flow at point C during time interval i
f, = the daily factor
Ee. = the time interval factor
The daily factor f; is assumed to be a random variable uniformly distributed
between 0.9 and 1.1). As far as g is concerned, it is assumed to follow a normal
distribution with a mean of zero. Two values of the standard deviation for 30 second
! Based on traffic counts that were collected in Toronto, Canada, the Highway Capacity Manual (1985)
reports that relatively narrow and parallel fluctuations were observed over the 77 days for which counts
were available, indicating the repeatability of the basic pattern. The range for the daily factor used in the
case study reflects this observation.
Page 148
intervals are used: 0.95 (high variability or HIVAR) and 0.30 (low variability or
LOVAR)®. The LOVAR case represents more typical variability in 30 second flows
while the HIVAR case represents a situation whereby the variability in 30 second flows
is at the higher end of possible values.
O-D Predictions Required by DTA
The DTA that is implemented in scenario 3 requires predictions of O-D demand
levels from the O-D updating module. In the case study three cases of the quality of the
information provided by the O-D updating module are considered, namely:
"Average" Case: In this case the DTA uses average historical O-D demand levels
only. Therefore, no O-D updating is required and p, is used each day. This level of
knowledge assumes that the DTA has an accurate unbiased estimate of the historical
dynamic O-D demands, including the seasonal and day-of-week patterns. While actual
estimates of the historical dynamic O-D matrix may not be perfectly accurate, the
significance of this case is the fact that it represents a scenario whereby no real-time O-D
updating is required at all. The implications of such a scenario include the elimination
of the surveillance and computational requirements associated with real-time O-D
updating.
"Day" Case: This case represents the situation whereby the DTA is assumed to be
aware of the daily factor f, associated with the O-D levels only. This means that the O-D
apdating module captures the systematic daily variation in O-D levels but cannot predict
variations about that day’s average flow values for specific time intervals.
"Perfect" Case: This case assumes that the DTA is aware of exact O-D realizations
for each time period. Specifically, f; and ; are assumed to be known. This corresponds
to an O-D updating module that can predict perfectly future demand levels.
While in the average case no real-time O-D updating is undertaken, the perfect
case requires an O-D updating module that is capable of exactly predicting future levels
2 The data for time-of-day variability in demand levels comes from a study conducted by the Minnesota
DOT (1983). This study indicated that the coefficient of variation for flows during 15 minute time intervals
was 0.181 and for 5 minute flows was 0.300. No information was available for 30 second flows, which
is the simulation interval being used in the case study. However, the coefficient of variation for 30 second
flows may be calculated from the values for the 5 and 15 minute intervals by assuming no correlation
between flows in separate 30 second intervals. These calculations indicate a coefficient of variation for 30
second intervals of 0.949 and 0.991 based on the 5 and 15 minute intervals, respectively. In case correlation
exists between flows in separate 30 second intervals, the actual coefficient of variation may be somewhat
smaller. Therefore, a coefficient of variation of 0.950 was used for 30 second flows since it lies at the
lower end of the calculated range and would provide for the possibility of some correlation. When similar
calculations were performed while completely controlling for the variability in overall demand levels due
to peaking, a coefficient of variation of 0.30 is obtained for 30 second flows.
Page 149
of O-D demands in real-time. These two cases represent extreme situations in the sense
that the actual DTA is likely to incorporate some real-time O-D updating capability, but
the resulting predictions will probably not be perfect. The O-D predictions that are likely
to be available will be in between these two cases as far as real-time O-D updating
capabilities are concerned. The day case represents one such possibility.
Accidents’
Variations in link capacities are assumed to occur as a result of accidents. For
each simulation run, it is assumed that accidents are described by the following factors:
Occurrences: The literature review indicated that incidents occur on roadways at
the rate of 0.025 incidents/hour/lane-mile. In this case study we concentrate on the
occurrence of major accidents which involve lane closures. Such accidents constitute
20% of all incidents. These values translate into an average of one major accident for the
5 hour simulation period for the networks under consideration. As a result, accident
occurrence is modeled based on an exponential inter-arrival time with a mean inter-arrival
time of 5 hours. Such a model would allow for multiple accident occurrences within the
simulation period.
Links Affected: Given that an accident occurs, it is assumed that it will affect one
of the four links, with equal probability.
Impact: Since the case study is concerned with major accidents, an accident is
considered to reduce capacity on the affected link by a factor of one half.
Duration: Based on the literature review, the duration of capacity reduction due
to an accident was modeled as a normally distributed random variable with a mean of 25
minutes and a coefficient of variation of 0.7.
Temporal Update Frequency
Three values of guidance update frequency were tested in the case study, 5
minutes, 2.5 minutes, and 30 seconds. The higher end of the range (5 minutes) was
selected since it is the frequency that is used in the LISB experiment in Berlin, as
indicated in Chapter 2. The highest update frequency (30 seconds) was selected since it
represented a value that will possibly be achievable, but only by dedicating significant
computational resources to the guidance system. Finally, an update frequency of 2.5
minutes was used as the base value. This frequency is higher than that of existing
systems but not as resource-intensive as the highest frequency case.
3 The parameter values used here for accident frequency and duration are based on a literature review
reported in Yablonski (1991).
Page (50
Guidance Threshold
In the case study, the "guidance threshold" CAR logic adopts a threshold of 1
minute. This value represents a guidance logic that uses shortest path guidance only if
the shorter route offers at least a 3.6% edge in average travel time (the average travel
time for the base case is approximately 27.5 minutes, as will become apparent later in this
chapter).
Free Flow Travel Time
As mentioned in Chapter 6, the free flow travel time on the running section of
network links is assumed to be zero. This leaves queuing delays as the only measure of
travel time. The two terms (delays and travel times) are used interchangeably in this
chapter.
The free flow travel time is likely to have some implications on the relative
effectiveness of the guidance scenarios. For instance, scenario 2 sets guidance based on
"current" queue lengths which are likely to change more significantly by the time vehicles
get to the bottlenecks if the free flow travel times were nonzero. In this case "current"
traffic conditions will be less reflective of the actual conditions that vehicles will
experience when they reach the downstream bottlenecks in the network.
In the case study, with a zero free flow travel time, the first vehicle in a packet
that leaves C during time interval t will experience queue lengths on links 1 and 3 that
are exactly equal to the values that were used in setting the guidance. The implication
is that the effectiveness of guidance based on current travel times is likely to be
overstated in the case study. Projective guidance, which makes use of current queue
lengths reported by the surveillance system, is also likely to be affected but to a lesser
degree since the queue lengths constitute just one input to the projections. In addition,
projective guidance would have to predict O-D levels further ahead into the future if free
flow travel times were not zero. This may result in a reduced O-D prediction quality.
This effect is not captured in the case study since it is assumed the quality of O-D
predictions for each of the average, day, and perfect cases does not depend on the
projection horizon.
Number of Replications
Each data point that is reported in the simulation results consists of simulating
system behavior under certain conditions for a number of days. The number of
replications should be selected in a way that offers a balance between statistical validity
and computational requirements. As the number of replications increases, the statistical
validity of the results increases but so does the simulation execution time. In this case
study, 200 days were simulated for each data point. A statistical analysis of the
simulation results is presented later in this chapter.
Page i151
7.2 Simulation Results
Unless otherwise noted, results pertain to network I, the FLAT demand pattern,
and a guidance update frequency of 2.5 minutes. The presentation of the results is
structured to reflect the impact of experimental factors presented in Chapter 6.
7.2.1 Guided Probability
In all the figures that are presented, specific performance measures for information
provision scenarios are plotted against the guided fraction. This provides us with the
facility to observe the variation of guidance performance as the situation changes from
only a few drivers being guided to fully guided traffic.
7.2.2 COP Factors
Information Provided by COP to CAR
Figures 7.2a - 7.2f compare guidance effectiveness achieved by the four
information provision scenarios using the 4 measures of effectiveness for the LOVAR
case and "basic" CAR logic. Specifically, the measures presented in each of the four
figures are as follows:
figure 7.2a: average delay for flow incoming at C
» ratio of average delay using projective guidance (perfect)
to that resulting from no guidance
» figure 7.2b: « guidance validity: percent of drivers who are "correctly guided”
figure 7.2c: + maximum delay experienced by guided vehicles
» ratio of maximum delay using projective guidance (perfect)
to that resulting from no guidance
» figure 7.2d: « fraction of guided vehicles with delays greater than 60 minutes
» figure 7.2e: + maximum delay of unguided vehicles
» figure 7.2f: + average delay for external flows
Average delay: Based on figure 7.2a, one can make the following observations:
» the greatest benefits for flow entering at C for all scenarios occur at guided
fractions of 0.2-0.5; the benefits decrease at higher guided fractions as the
adverse effects of information come into play
» the maximum benefit occurs for the projective (perfect) guidance at a guided
probability of 0.2 and 0.3; this maximum benefit amounts to a 4.6%
reduction in average delays over the no-guidance case (for flow entering
at C)
Nrnr
Li 32
» except for guidance based on last reported travel times, all other scenarios
outperform the no-guidance case up to a guided fraction of 0.7
» the delays for the projective (perfect) and projective (day) guidance are superior
to those corresponding to guidance based on instantaneous traffic conditions
up to a guided probability of 0.8 beyond which the average delays for the
three scenarios become almost equal
Guidance Validity: Based on figure 7.2b, the following observations can be made:
» the guidance validity decreases as the guided fraction increases; at high guidance
fractions the validity could be as low as 50%
» the projective (perfect) and projective (day) provide guidance with consistently
superior validity
» the guidance validity for the other scenarios is rather similar
It should be noted that even though perfect predictions of O-D demands are
assumed to be available in the case of projective (perfect) guidance, the relatively
low validity that is achieved at high guided fractions is probably due to the
relatively low temporal update frequency of 2.5 minutes. Under the "basic" CAR
logic, and with most vehicles being guided and responding to the same message
over a period of 2.5 minutes, overreaction is likely to occur causing the route to
which drivers are guided to become congested. This phenomenon results in a
reduction in guidance validity.
The validity at a guided fraction of 1.0 is obtained by comparing the delay
experienced by a packet of guided vehicles to what would have been experienced
by a packet that leaves from C at the same time but uses the alternate route.
Maximum delays for guided vehicles: Referring to figure 7.2c, the following
observations can be made:
» except for the case where guidance is based on last reported travel times,
guidance results in significant reductions in maximum delays;
reductions for the projective (perfect) scenario range from about
26.2% at a guided fraction of 0.1 to 17.7% when all vehicles are guided
guidance based on all projection scenarios consistently outperforms guidance
based on instantaneous traffic conditions; that is, even guidance using
projections based on average demand levels reduces the maximum travel
times more than guidance using instantaneous information; the foresight
provided by projective guidance seems to be successful in identifying and
avoiding situations where drivers may experience long delays
Fraction of guided vehicles with high delays: The following observations can be
made based on figure 7.2d:
Jage 33
» except for the case where guidance is based on last reported travel times,
guidance results in moderate reductions in the fraction of drivers
experiencing long delays; while 12.5% of drivers suffer from long
delays if no guidance is provided, the for guided vehicles fraction drops
in the case of projective (perfect) guidance to about 9% when the guided
probability is 0.1 and to 10.5% when all vehicles receive guidance
» all the projective scenarios outperform the scenario based on instantaneous
guidance
Figures 7.2e and 7.2f present the variation in maximum delays of unguided
vehicles and in the average delays of the external flows (p, and p,) for the different
scenarios and for the range of possible values of guided probability. These figures
indicate reductions in both measures over the whole range of guided probability for all
scenarios of information provision. In addition, all projective scenarios consistently
outperform instantaneous guidance for these two measures.
These figures provide us with evidence that the performance of projective (day and
perfect) guidance is superior to that of instantaneous guidance in terms of all performance
measures (except for average delays at a guided probability of 1.0). It is also interesting
to note that even the projective (average) case outperforms the instantaneous case in all
measures except average delay.
These result are significant in that they point to the fact that if demand is of
typical variability, using projective guidance without an on-line prediction of O-D demand
levels (average historical levels) or with simple predictors that capture only daily variation
(day) produces results that are superior to those afforded by instantaneous guidance.
Demand Variability
The LOVAR case (cv=0.30) represents O-D demands that display a realistic,
typical level of variability. The impact of more variable O-D demands on guidance
effectiveness is investigated using the HIVAR case.
The results shown in figures 7.3a-7.3d indicate that, under conditions of high
demand variability, all projective guidance scenarios are successful in reducing maximum
delays of guided vehicles to a level that is consistently lower than that achieved by the
other scenarios. However, the advantage of projective guidance, if any, is less consistent
or significant based on the other performance measures.
As noted in section 7.1, the HIVAR case uses a coefficient of variation (cv) of
0.95 for 30 second O-D flows. This represents demand levels that are highly variable and
unlikely to materialize with actual traffic conditions but are considered to provide an
upper bound on the impact of demand variability.
Page 154
Demand Pattern
The performance of the various guidance strategies for the SHARP demand pattern
was investigated using the simulation model for the HIVAR and basic” CAR logic cases.
Results are shown in figures 7.4a and 7.4b.
It can be observed that:
» the average delay in the no guidance case associated with this SHARP
demand pattern is higher than the average delay in the base case of the
FLAT demand pattern; this reflects the higher overall level of congestion in this
case which also causes overreaction even at a guided probability of 0.7 for
projective (perfect) guidance
the reductions in average and maximum delays based on guidance are
smaller for the SHARP demand pattern than for the base case FLAT demand
pattern due to the higher congestion levels which reduce the opportunities for
guidance to identify uncongested routes to which drivers should be guided
 Wi the reductions in maximum delay by instantaneous guidance are small,
particularly at high guided fractions; this is due to the more pronounced peaking
(and therefore the more quickly varying demand levels) which instantaneous
information has no way of anticipating since in this case current information is
not a good predictor of future traffic conditions
the projective (average) scenario performs better than guidance based on
instantaneous information according to both performance measures
The fact that the projective (average) guidance outperforms instantaneous guidance
is significant since it means that an accurate knowledge of the historic O-D demand levels
without any real-time O-D updating would be enough for scenario 3 to provide guidance
that is better than that provided by scenario 2. In such a case, there would be no need
for an O-D updating module, thus reducing the computational and hardware (surveillance)
costs associated with scenario 3.
7.2.3 CAR Factors
In analyzing the impact of factors that do not involve a comparison of the
information provision scenarios, we use the projective (day) scenario to illustrate the
results. A projective scenario was chosen since this is the focus of the dissertation. The
‘day" O-D prediction case was selected since it represents a middle ground between the
‘average" and "perfect" cases. Simulation results confirmed that its performance lies
between those of the other two cases.
CAR Logic
Figures 7.5a-7.5¢ evaluate the impact of CAR logic for the LOVAR case. The
Page 155
scenarios that are included are no guidance as well as instantaneous and projective (day)
guidance, with each of the latter two scenarios involving the "basic" and "guidance
threshold” approaches to CAR logic. The purpose of these figures is to illustrate the
relative performance of instantaneous and projective guidance when both are based on an
advanced CAR logic. The figures indicate that both scenarios benefit from the advanced
logic, but projective (day) consistently outperforms instantaneous for both basic and
guidance threshold cases, for all 3 performance measures being considered here. (The
only exception occurs for the average delay at a guided probability of 1.0 with the "basic"
CAR logic).
Based on these figures it can also be observed that guidance based on an advanced
CAR logic (with a threshold) is superior to the "basic" CAR logic in three respects,
namely:
 it results in significantly lower average delays at higher guided probabilities
(fig. 7.52)
the resulting maximum delays for guided vehicles are consistently lower
(fig. 7.5b)
» it provides for higher guidance validity at all guided probabilities; however,
the improvement in validity is more significant at low guided probabilities
(fig. 7.5¢)
These results indicate that using an advanced CAR logic is effective in reducing
the potential adverse effects from overreaction which occur due to a lot of drivers reacting
to guidance advice. When the guidance is based on a the "basic" approach, the route to
which drivers are guided can easily become congested since it may have been just
marginally shorter than alternate routes. On the other hand, if an advanced CAR logic
is used, the route to which drivers are guided does not become congested as easily as it
does in the case of single route guidance since that route would have been shorter than
the other routes by some quantity. Moreover, when the difference in travel times on
alternate routes is smaller than the threshold and guided vehicles are subsequently
distributed over alternate routes (using "no guidance" or route distributive guidance), no
single route becomes overly congested.
It should be noted that the advantage of an advanced CAR logic is expected to be
most prevalent when guidance is not updated very frequently. When guidance updates
are provided frequently, overreaction is reduced so the effect of the advanced CAR logic
is expected to become less important.
Finally, it should be pointed out that the drivers who are considered to be guided
for the purposes of computing guidance validity and maximum delays are those who
receive single route guidance. It is for those drivers that high levels of guidance validity
are observed since the guidance with which they are provided is based on a threshold or
after ensuring consistency based on predicted demand levels and driver behavior. The
Page 156
benefits to other driver (considered to be "unguided") are demonstrated indirectly through
the overall improvement in average delays.
Spatial Update Frequency
The advantage of projective (perfect) guidance in network I vis-a-vis other
guidance strategies lies in its ability to predict what will happen on the downstream links
and set the guidance at C on that basis. Therefore, the long foresight sets projective
guidance apart from other guidance strategies for that network. Obviously, if the demand
patterns were uniform, then the foresight would not provide any advantage and current
information would be as good as predictive information.
Figures 7.6a and 7.6b present the relative effectiveness of the various guidance
strategies when applied in network II (see figure 6.1) using HIVAR and the "basic" CAR
logic. The striking feature about the results lies in the observation that instantaneous
guidance and projective guidance (both average and perfect scenarios) generate almost
identical results as far as average delay and maximum delay of guided vehicles are
concerned (the instantaneous case does not even appear very clearly in the figure since
its data points lie beneath those of the other scenarios). The reason behind these results
is the fact that with network II guidance is provided at both points C and D, thus allowing
for revisions in the original guidance provided at C. The guidance at D is set after a
closer look is taken (by all guidance strategies) at the traffic conditions on the reraaining
portion of the trip, including the current queue lengths. The revision of guidance at D
in conjunction with the zero re-guidance penalty eliminates the advantage of the foresight
afforded by projective guidance.
Temporal Update Frequency
Figure 7.7a presents a comparison of the average delays that occur based on
projective (day) guidance for guidance update frequencies of 5 minutes, 2.5 minutes, and
0.5 minutes using the LOVAR and "consistency check" CAR logic. In addition, the
performance of the projective guidance for 30 second guidance updates and with the
"consistency check" as CAR logic is also plotted since it represents the best projective
guidance strategy that may be applied within the scope of the case study.
It is apparent that more frequent guidance updates result in lower average delays
especially when the guidance probability approaches 1.0. For a guidance update of 30
seconds, both projective (day) and the "best projective” case yield average delays that are
significantly lower. Of special importance is the fact that average delays almost do not
increase at high guidance probabilities, indicating a total elimination of overreaction.
Figure 7.7b presents a comparison of maximum delays experienced by guided
vehicles for the same range of guidance update frequencies. It indicates that at low
guided probabilities, there is only a small difference between the three frequency levels
Doge 157
being simulated. However, when the guidance fraction is high (>0.5), the differences
become more significant. Again, an update frequency of 30 seconds with the projective
(day) scenario and the "best projective" case provides further improvements in this
performance measure, and succeeds in keeping the maximum delays at almost a flat level.
Figure 7.7c indicates that more frequent guidance updates improve guidance
validity, especially when moving from a 2.5 minute to a 30 second guidance update
interval. The performance of the "best projective" case signifies that guidance validity
may be kept above 95% over all guided probabilities if such a strategy can be
implemented.
Finally, it should be emphasized that the outstanding performance of the "best
projective” case in all three performance measures is based on the three factors it
incorporates:
(i) perfect knowledge of O-D demands
(ii) frequent guidance updates
(iii) the consistency check which may only be applied if the guidance is based on
a projection of traffic conditions that takes potential driver response into account;
that is, this represents the culmination of the 3 principles embodied in the
proposed framework.
While item (i) above may not be achievable in reality, the last two items are
within reach given the availability of the required computational and surveillance
resources.
7.2.4 DTA-Specific Issues
O-D Predictions
The impact of the quality of the O-D predictions available to the projective
guidance of scenario 3 can be analyzed by comparing the results obtained from the
"average", "day", and "perfect" cases. The relative performance of these cases (in terms
of the measures of effectiveness that are used in the case study) when compared with
other information provision scenarios was discussed in section 7.2.2.
Incident Detection
The base case projective scenarios assumed that incidents are detected instantly
and that travel time projections based on the DTA take into account the occurrence and
duration of such incidents and the associated reductions in capacity. Figure 7.8a
compares the performance of the projective (day) guidance for the base scenario (referred
to as "del=0" since it is assumed that incident detection delay is zero) as well a scenario
Page 158
in which the DTA has no information at all concerning the occurrence or impact of
incidents (referred to as "no info"). The LOVAR and "basic" CAR logic are used here.
The figure indicates that somewhat higher average delays are experienced in the case of
no information regarding incidents. Moreover, the guidance validity of the no incident
information case remains consistently and significantly superior to that of the
instantaneous guidance case, as is clear from Figure 7.8b. Figure 7.8c indicates that when
no information concerning accidents is available to the projective (day) guidance scenario,
the maximum delays that result are less than but close to those with guidance based on
instantaneous travel times produces, and only slight benefits would be experienced by
adopting a projective (day) guidance scenario under these conditions.
Since the no information scenario represents an extreme situation in which no
incident detection capability is available at all, what is achievable in reality is likely to
be better than this, but not as efficient as the no delay scenario. This is likely to leave
the projective (day) scenario with tangible benefits over instantaneous guidance, especially
as far as guidance validity goes.
In both "no info" and "del=0" cases it is assumed that the DTA receives
information from the surveillance system on the current queue lengths at bottlenecks.
Instantaneous guidance receives the same information and is "aware" of incidents in as
much as these queue lengths reflect their occurrence.
7.2.5 Demand and Supply Characteristics
The impact of demand variability and demand pattern were discussed in section
1.2.2
Demand Level
Figures 7.9a-7.9¢ present simulation results for a lower congestion case whereby
demand levels are reduced by 10%. The LOVAR and "basic" CAR logic are used.
The figures indicate that with such lower congestion the advantage of projective
(day) guidance over instantaneous guidance is reduced or eliminated for all performance
measures being presented. This may be due to the fact that with lower demand levels
recurrent congestion becomes less pronounced and there may be less need for an
advanced COP module.
Stochasticity of Network Capacity
It has been argued that the greatest benefits from route guidance systems will
occur when capacities are highly variable, in which case drivers who follow their habitual
travel patterns may be in for long delays while those who follow guidance are warned and
may be spared such delays. Here we deal with the impact of accident probability (and,
Dage .59
therefore, stochasticity in capacities) on guidance effectiveness. Figure 7.10 presents
results from cases where the average accident inter-arrival time was varied from 2.5 hours
to 5 hours (base case) and to 25 hours (very low accident occurrence probability). The
figure indicates that delay savings increase with accident probability which confirms the
argument cited above. It also indicates that the more significant benefits from guidance
will occur in cases of incident congestion as opposed to recurrent congestion. Moreover,
the figure indicates that when the accident probability is low disbenefits may occur as a
result of guidance at high guided fractions. Again, this is due to overreaction which
results from the relatively low guidance update frequency (every 2.5 minutes).
7.3 Delays of Guided and Unguided Vehicles
Figure 7.11a presents the average delays of guided, unguided, and all vehicles for
the base projective (day) scenario for the HIVAR and "basic" CAR logic. It indicates that
benefits to guided vehicles in terms of reductions in average delays range from about 10%
at a guided probability of 0.1 to zero when all vehicles are guided. Moreover, unguided
vehicles also receive some benefits which are, however, smaller than those experienced
by guided ones. The maximum delays of guided and unguided vehicles plotted in figure
7.11b for the projective (perfect) scenario with LOVAR and "basic" CAR logic indicate
a similar relative pattern.
7.4 Delays by Departure Time
Figure 7.12 plots values of delays encountered by vehicles leaving C at times
between 7 AM and 10 AM and following route R1 from C to G, for one specific
replication in a simulation run consisting of 200 replications. This figure displays the
expected delay pattern whereby delays increase up to a peak and then decrease again.
However, the figure indicates that delays reach their highest values for flows departing
C between 8:45 AM and 9:30 AM. A driver behavior model that also includes departure
time choice in conjunction with a desired time of arrival at the destination is likely to
yield a delay pattern that is somewhat shifted to the left; i.e., with a desired arrival time
between 8:30 and 9:00 AM, it is expected that the peaking in the delay pattern would
occur earlier than what is shown in the figure.
7.5 Statistical Analysis of Simulation Results
The simulation runs associated with the case study are characterized by a high
level of stochasticity. Specifically, the following processes are based on probabilistic
measures:
» O-D demands: stochasticity in daily factor and time interval factor
» accidents: stochasticity in occurrence, link affected, start time, and duration
» driver behavior: stochasticity of route choice (see section 6.2)
Page 160
The service rate at each bottleneck is basically deterministic. However, the
specific value of the bottleneck discharge capacity is reduced whenever an accident that
affects that link occurs.
The probabilistic processes involved in the simulation result in performance
statistics which display a significant amount of variability. For instance, the following
distribution of occurrences of average delays over the 200 replications was obtained for
the projective (perfect) case at a guided probability of 0.50 and a guidance update
frequency of 2.5 minutes:
0-15 minutes
15-20 minutes
20-25 minutes 4.
25-30 minutes 40
30-35 minutes 32
35-40 minutes 25
40+ minutes 0
As can be observed from this distribution, the average delay values are widely
spread around the mean value of 26.63 minutes.
To obtain a sense of the statistical validity of the results, we perform the following
analysis based on simulation runs plotted in figure 7.2a (LOVAR) at a guided probability
of 0.5. For a sample of size 100 out of the 200 replications constituting each simulation
run, we compute the differences in average delays for the three of the scenarios, as
indicated below, with each replication involving the same accident situation for each of
the scenarios. The results were as follows:
Sample
Sample Mean Standard Deviation
No Guidance - Projected (Perfect) 0.807 0.124
Instantaneous - Projected (Perfect) 0.182 0.062
Using these figures, The following conclusions may be drawn:
» a 90% confidence interval of the average difference between the no guidance and
projective (perfect) case is [0.787,0.827]
» a 90% confidence interval of the average difference between the instantaneous
and projective (perfect) case is [0.172,0.192]
Finally, it should be noted that the variability in performance measures decreases
Page 161
when the level of stochasticity of some of the input processes decreases. For instance,
the following values were obtained for the average delay in the no guidance case under
different conditions:
Sample
Sample Mean Standard Deviation
LOVAR 27.54 8.54
HIVAR 30.47 9.56
Higher Accident Probability/HIVAR 32.76 10.22
Lower Accident Probability/HIVAR 28.76 8.78
The last case represents the LOVAR scenario which, it was noted, is more typical
of the level of variability that may realistically exist in traffic flows.
7.6 General Conclusions
Based on the above results, the following general conclusions can be drawn:
Issues Related to Proposed Framework
Benefits of Projective Guidance
Guidance based on projected traffic conditions offers advantages over other
guidance scenarios under the following conditions:
» Demand Characteristics: when O-D demand levels are not highly variable form
day-to-day
» COP/Surveillance: when good predictors of daily O-D demand levels are
available
» Surveillance: with the existence of fast incident detection capabilities
Network Structure: when network topology does not provide "cheap"
opportunities for re-routing
The degree to which these conditions are likely to be met in typical situations
where ADIS are to be implemented is discussed at the end of this chapter.
Even with no real-time updating of the O-D demands, and relying solely on the
average historical levels, projective guidance is still to be preferred over other information
provision scenarios if O-D demand patterns display significant peaking occurring at
different times of the analysis period, an effect which may not be captured by other
guidance scenarios. Moreover, even with such absence of real-time O-D updating
capabilities, projective guidance offers better results in terms of reductions in maximum
Page 162
delay experienced by guided and unguided vehicles as well as in the delays associated
with external flows.
Guidance based on projected traffic conditions with good predictors of daily O-D
demand levels provides improved levels of guidance validity. Although not tested in this
case study, it is expected that improved guidance validity will potentially lead to higher
levels of driver compliance with guidance and an improved perception of guidance
reliability.
The Value of the 3 Principles Embodied in the Proposed Framework
The occurrence of adverse effects from guidance at high guided fractions due to
overreaction were eliminated when the following actions were adopted:
» guidance is based on projected traffic conditions (Principle 1)
» a DTA is used for congestion prediction (Principle 2)
» an advanced CAR logic (consistency check) is resorted to (Principle 3)
» the CAR’s temporal update frequency is increased ensuring that only a small
number of drivers receive the same guidance advice
These actions resulted in average delays that did not increase at high guided
probabilities and in significant improvements in guidance validity. However, guidance
validity levels higher than 95% for all guided probabilities were only possible if "perfect"
predictions of O-D demand levels were available.
General Guidance Issues
Situations Where Guidance Offers Highest Benefits
Guidance (in general) is needed most when the traffic conditions are highly
stochastic, as when accidents occur frequently. In such cases, guidance offers a more
significant advantage by guiding drivers away from locations that are highly congested
as a result of accidents. This also means that guidance is more effective in dealing with
incident (non-recurrent) congestion than with recurrent congestion when the network is
rather congested.
Size of Benefits from Guidance
The maximum reductions in average delay that were obtained by using guidance
were modest and occurred at market penetration rates that are less than 50%. However,
some of the other performance measures indicated that guidance does offer other more
significant advantages. Specifically, the reductions in maximum delays were found to be
more significant and occurred for the full range of guided probabilities.
Page 163
Impact on Unguided Vehicles
Unguided vehicles were also found to receive some benefits from guidance. Such
benefits, however, were smaller than those experienced by guided vehicles under the same
conditions.
Generalization of Results
In this case study we were able to analyze some of the conditions under which
projective guidance may be valuable and to evaluate the impact of several factors on
guidance effectiveness. The conclusions that were reached are based on two small,
prototypical networks under a wide range of possible scenarios that represent a set of
possible effects that may take place in reality. These results provide some insights that
are likely to be useful when real-time driver information systems are to be planned and
designed. Some of these insights are discussed next.
Guidance Validity and Controlling the Adverse Impacts of Improved Information
[t was observed that the overall validity levels were low (of the order of 50-60%
at high guided probabilities) when guidance was updated infrequently and based on a
basic CAR logic or elementary COP capabilities. With the detrimental effects this may
have on driver compliance, it is not expected that such a situation would be stable or that
it is sustainable for long periods of time. As such, it is concluded that the achievement
of high levels of validity is of great significance for the operation of such driver
information systems.
The role of the three principles that were advanced in this dissertation in
improving guidance validity and in controlling the adverse impacts of improved
information was demonstrated. This should have significant implications for the design
of future ADIS projects, especially as the desired market penetration of such projects
approaches levels at which the adverse impacts become most prevalent.
System Design Tradeoffs
The case study uncovered some interesting tradeoffs between several elements of
the overall framework for implementation of real-time driver information systems. For
instance, a clear tradeoff emerged between the spatial and temporal update frequencies
associated with the CAR element on the one hand and the need for an advanced COP
module on the other hand. Moreover, another tradeoff exists between the use of an
advanced CAR logic and an increased temporal update frequency for the purpose of
controlling overreaction. These issues are to be considered carefully, together with their
implementation implications (cost, hardware, and computational requirements), when
deciding on the ADIS structure most suitable for specific situations.
lage 164
Implementation of Projective Guidance
It was indicated at the beginning of this section that 4 conditions have to be met
for projective guidance to provide benefits over other information provision scenarios.
The condition on variability of O-D demand levels is not restrictive and the level of
variability that is commonly exhibited by O-D flows satisfies this requirement. The
availability of good predictors of O-D demands necessitates that such predictors be able
to capture at least the overall variations in demands for a specific day from average
levels, as well as the existence of a very good estimate of these average levels. It is
expected that such O-D predictors, while not available at this time, are likely to be
developed in time for the operation of ADIS. Incident detection capabilities should be
able to provide the DTA with capacity reduction information within a short period of time
after the accident occurs. The technology for such a capability is being developed and
is likely to be available for the support functions required here. Finally, the network
topology determines whether or not an advanced COP capability is required based on the
availability of inexpensive en-route re-guidance possibilities. It is expected that urban
street networks may provide such opportunities while this may not be the case for
freeway networks in general.
Page 165
Fig. 7.1a FLAT Demand Pattern
“>
[3
=
be
4 :
3
a —
> 1
-
a
a
~
© eanag
a  LL LLY yr CerebrrpLdft
a
Time of Day (AM)
igi— phi —— rho2 —r— rho4
166
Fig. 7.1b SHARP Demand Pattern
3
«3
Y
fq
2
—
smd
~ &
a Q
oy
o  Trpt
Nr 7
™
4
nt
—y
3
OJ
pgm 3
wnlnad
Laan
TC Al lA a
a pty briny
3 prepregpvveeydgLI
Time of Day (AM)
sninln, phi —— rho2 —— rho4
lf
Fig. 7.2a Impact of COP Information on Average Delays
Guid Freq = 2.5 min / LOVAR / Basic
~¢
‘
x
1s) A
27.5 FR
R ON)
1
\ J) 00?
27 J a4
PP
af 978
ry
- 26.5
4
~
{» 2.960
0
J J4 0.954
26
oT
00 ©01 02 03 04 05 06 07 08 09 1.0
Guided Probability
—a— No Guidance —— Last Reported —a— Instant. Guid.
—o-Proj. Guid. (ave) —— Proj. Guid. (day) —— Proj. Guid. (perf)
HA
Fig. 7.2b Impact of COP Information on Guidance Validity
Guid Freq = 2.5 min / LOVAR / Basic
1.0
g 0.8 |
<
0.7
1 0.6
”
-
>
— mp
0.5
D.£
00 02 03 04 05 06 07 08 09 10
Guided Probability
-=- Last Reported —o Instant Guidance —a— Proj. Guid. (ave)
—o Proj. Guid. (day) —o—Proj. Guid. (perf)
69
Fig. 7.2¢ Impact of COP Information on Maximum Delays
Guid Freq = 2.5 min / LOVAR / Basic
wr 0
15
70
© 65
——
60
55 +
Ee oro 0778
ooo 0750 O70
Ld
50 SONEer
50 01 02 03 04 05 06 07 08 09
Guided Probability
—a— No Guidance —o— Last Reported —— Instant. Guid.
—o~Proj. Guid. (ave) —o Proj. Guid. (day) — Proj. Guid. (perf)
70
Fig. 7.2d Impact of COP Information on Fraction with Long Delays
Guid Freq = 2.5 min / LOVAR / Basic
0.17
= 0.16
*-
D 0.15
So
0.14
0.13
0.12
—_
0.11
0.1 —  3wl
-,
2
pe
4 0.09
i.
D.08
7%0 01 02 03 04 05 06 07 08 09  0
Guided Probability
—a— No Guidance —o— Last Reported —a— Instant. Guid.
—o-Proj. Guid. (ave) ——Proj. Guid. (day) —— Proj. Guid. (perf)
| 7
Fig. 7.2e Maximum Delays of Unguided Vehicles
Guid Freq = 2.5 min / LOVAR / Basic
he!
80
!
15
opm
70
~~
ed 65
al
60
’
55
53
=
=
01 02 03 04 05 06 07 08 09
Guided Probability
—=—- No Guidance —— Last Reported —— Instant. Guid.
—o-Proj. Guid. (ave) ——Proj. Guid. (day) -a— Proj. Guid. (perf)
1 79
Fig. 7.2f Average Delays for External Flows
Guid Freq = 2.5 min / LOVAR / Basic
14 pu
3 |
= 13.9
 Xx
< 13.8
£
ee 13.7
13.6
13.5
13.4
» 133
o 132
3} 13.1 |
-
 PD oa3
a.
 129
12.8
0.1 02 03 04 05 06 07 08 09 1.0
Guided Probability
—a— No Guidance —a— Instant. Guid. —o— Proj. Guid. (ave)
——Proj. Guid. (day) ——Proj. Guid. (perf)
1 73
Fig. 7.3a Impact of Demand Variability on Average Delays
Guid. Freq. = 2.5 min / HIVAR / Basic
"
— 31
=
“0
30.5
v
30
= J QR4
1.
75
ne
i 20.5
J
-
rr: 1 _— G.»”
rr
1 20
oss i 0953 7%
0.949
208
04 0.5 0.6 0.7 0.8 09 1.0
Guided Probability
-a— No Guid. —o— Last Reported —a— Instan. Guid.
_o Proj. Guid. (ave) —o Proj. Guid. (day) —— Proj. Guid. (perf)
Numbers below data points indicate ratio of average delay for proj. (perf) over no guidance
1 7
Fig. 7.3b Impact of Demand Variability on Guidance Validity
Guid. Freq. = 2.5 min / HIVAR / Basic
 J
: 0.8
5 07
0.6
"
a
D.5
D.4
20 01 02 03 04 O05 06 07 08 09 10
Guided Probability
—o— Last Reported —— Instan. Guid. —o Proj. Guid. (ave)
—o— Proj. Guid. (day) —a— Proj. Guid. (perf)
175
Fig. 7.3c Impact of Demand Variability on Maximum Delays
Guid. Freq. = 2.5 min / HIVAR / Basic
 Nn
90
0
a
70 ppm
et 0.813
. — 0.801
ote 0778 0783 0.788
© yj 0746 O70 ©
80
2725
0
| 0.4 0.5 0.6 0.7 0.8 0.9 10
Guided Probability
—=— No Guid. —o— Last Reported —a— Instan. Guid.
—o Proj. Guid. (ave) —o Proj. Guid. (day) -a— Proj. Guid. (perf)
Numbers below data points indicate ratio of maximum delay for proj. (perf) over no guidance
76
Fig. 7.3d Impact of Demand Variability on Fraction with Long Delays
Guid. Freq. = 2.5 min / HIVAR / Basic
ay
- 0.16
=
-
»~
L
A 0.15
~
al 0.14
~
r
e
kN 0.13
(
1
1’
"
rr
i
— 0.12
0.1!
04 05 06 07 08 09 10
Guided Probability
—e— Last Reported —a— Instan. Guid.
_o Proj. Guid. (ave) ——Proj. Guid. (day) -a— Proj. Guid. (perf)
{77
Fig. 7.4a Impact of Demand Pattern on Average Delays
SHARP / Guid. Freq. = 2.5 min / HIVAR / Basic
yr
—
3}
33.5
33 - Ao
pia .
1.012
32.5 1.007
=
~~ Nn 1.001
iy 4g! “
32 —— )“0.001
0.95
’ 0.985 0.983 0.957
2 31.5
dy
vo
50 J.1 .2 03 0.4 0.5 0.6 0.7 nN hp 1.0
Guided Probability
—a— No Guid. —a— Instan. Guid.
—o— Proj. Guid. (ave) —— Proj. Guid. (perf)
Numbers below data points indicate ratio of average delay for proj. (perf) to no guidance
178
Fig. 7.4b Impact of Demand Pattern on Maximum Delay
SHARP / Guid. Freq. = 2.5 min / HIVAR / Basic
rl
L<.
80
»”
 i—
—ail-
70
wd
x a) ss ne
= - 2500 0 0.829 0.839 0.851 0.859 0.871 0.879
60
.
rr
&
"0
0 0.1 0.4 0.5 0.6 0.7 Of "3 Nb
Guided Probability
—a— No Guid. —a— Instan. Guid.
—o Proj. Guid. (ave) —a— Proj. Guid. (perf)
Numbers below data points indicate ratio of maximum delay of proj. (perf) to no guidance
{ 70
Fig. 7.5a Impact of CAR Logic on Average Delays
Guid Freq = 2.5 min / LOVAR
 oO 28
wr!
%
af
27.5
=
hd
A \
7
=
~
“57
A . J
4 26.5
"3 0.965
»
0.yf —— 0.956
0953 gs 0952 09%
. 2F
1.0  D0 02 03 04 05 06 07 08 0.9 10
Guided Probability
-a— No Guidance —a— Instant. (basic) —o— Instant. (thresh.)
—o— Proj. (day, basic) a Proj. (day, thresh)
180
Fig. 7.5b Impact of CAR Logic on Maximum Delays
Guid Freq = 2.5 min /LOVAR
>
= 80
=
»y
75
70
lL
55
 A
50
55
oa 0751 0764
0733 0732 0734 07136 074
: nr————————————
oh
2 00 01 02 03 04 05 06 07 08 09
Guided Probability
—a— No Guidance —a— Instant. (basic) —o Instant. (thresh.)
—o—- Proj. (day, basic) —— Proj. (day, thresh)
81
Fig. 7.5¢ Impact of CAR Logic on Guidance Validity
Guid Freq = 2.5 min / LOVAR
1.9
at 0.8
 el {
eee,
nd
1) 0.7 i
v 0.6
po
po
rm ~~ r—
‘u
———
0.5 -— ef
I.
10 0.1 02 03 0.4 0.5 0.6 0.7 0.8 0.0 1.0
Guided Probability
—a— Instant. (basic) —o Instant. (thresh.)
—— Proj. (day, basic) —— Proj. (day, thresh)
89
Fig. 7.6a Impact of Spatial Update Frequency on Average Delay
Guid. Freq. = 2.5 min / HIVAR / Basic
31
”
-
bres
3
)
30.5 n
»-
2
Lg 30 ~~
rd
S$
ip
3 209.5
—ay
<
5
3
--
Ly) ~ A
L 20
SD 0.561
2.955
28.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 3 0
Guided Probability
-a— No Guid. —e— Last Reported —a— Instan. Guid.
_— Proj. Guid. (ave) —a— Proj. Guid. (perf)
R 3
Fig. 7.6b Impact of Spatial Update Frequency on Maximum Delay
Guid. Freq. = 2.5 min / HIVAR / Basic
20
<
4
80
T0
oo — ad
oes 0910 09%
< 0.888
0.864 0.877
| fag 0.854 :
50 0.822 C.tet
-
-
3G
1.0 J N.2 03 0.4 0.5 0.6 0.7 0.8 {
Guided Probability
—a— No Guid. —e— Last Reported —a— Instan. Guid.
_— Proj. Guid. (ave) —a— Proj. Guid. (perf)
R/
Fig. 7.7a Impact of Temporal Update Frequency on Average Delays
Proj. Guid. (day) / LOVAR / Consist. Check
0 29.5
“0 29
28.5
28
27.5
st
~ 27
J———
26.5
¢
f ——
m——
 ee——-
26
(2)
=)
25.5
2.
D.0 J i 02 03 04 05 06 07 0%
n
Guided Probability
—— 5 min — 2.5 min — 30 sec
—o no guid —— "Best Proj."
8 ~
Fig. 7.7b Impact of Temporal Update Frequency on Max. Delay
Proj. Guid. (day) / LOVAR / Consist. Check
A 80
\
t
$a
“ 15
oi
ur
T0
.
+
8S
60
-
 rarerr——
<g
= of— 0
=
 rT I
na D1 0.2 03 0.4 0.5 0.6 0.7 0.8 No 1.0
Guided Probability
—— 5 min —— 2.5 min — 30 sec
ono guid —— "Best Proj."
RE
Fig. 7.7¢ Impact of Temporal Update Frequency on Guidance Validity
Proj. Guid. (day) / LOVAR / Consist. Check
—
y= no
3
—
ty
iad
) 0.8 |
r 0.7
n
, =
re
N.6
 te
" —
)
D.0 0.1 07 nN 0.4 0.5 06 07 0.8 0.9 1.0
Guided Probability
min —t— 2.) min —a— 30 sec — "BestProj."
| Q 7
Fig. 7.8a Impact of Incident Detection Delay on Average Delays
Guid Freq = 2.5 min /LOVAR / Basic
 QO 28
eo
«nn
27.5
-
Ty 27
Tt
&
a
-
-
¢.
A 26.5
¢ )
20
%
0.2 03 0.4 0.5 0.6 0.7 0.8 0.9 0
Guided Probability
—=— No Guidance —a Instant. Guid.
—o— Proj. (day,del=0) Proj. (day,no info)
88
Fig. 7.8b Impact of Incident Detection Delay on Guidance Validity
Guid Freq = 2.5 min / LOVAR / Basic
).0
).85
1
J.8
5 0.75
~
3 0.7
’
*
.
-
- 0.65
»
r 0.6
—
4.
 -—
1.5% Te T—a-
-
rt
1.6
N.45
1.0 3 0.2 03 04 05 06 07 08 09 1.0
Guided Probability
_ instant Guidance — Proj. (day,del=0) — Proj. (day,no info)
8
Fig. 7.8¢c Impact of Incident Detection Delay on Max. Delays
Guid Freq = 2.5 min / LOVAR / Basic
 Awn  80
ow
iy
J 75
a“
T0
r
*
55
wil
erm’
—
50
 ht
wil
55
V0)
3.0 J.1 02 03 0a 05 06 07 0.8 nx
Guided Probability
= No Guidance —a— Instant. Guid.
—o— Proj. (day,del=0) —a Proj. (day,no info)
90
Fig. 7.9a Average Delays for Low Congestion Case
Guid Freq = 2.5 min / LOVAR / Basic
~ 17
=
0D
0
16.5
3
168
-
+
J
\
15.5
J
—
n.0 03 02 03 04 05 06 07 08 0-
Guided Probability
—=— No Guidance —- Last Reported
-« Instant. Guid. —o— Proj. Guid. (day)
9
> Fig. 7.9b Maximum Delays for Low Congestion Case
2 Guid Freq = 2.5 min / LOVAR / Basic
~ 55
50
un
45
40 mann
. a— 7
=
osor O81¢
%
0.789
0.775
35 = ” 0.752 0.764
~ 128 :
UU.O
30 lu
D0 0.1 02 03 04 05 06 07 08 0.9 0
C
_—-
= Guided Probability
—a— No Guidance —e— Last Reported
—« Instant. Guid. —— Proj. Guid. (day)
Q
Fig. 7.9¢ Guidance Validity for Low Congestion Case
Guid Freq = 2.5 min / LOVAR / Basic
).8¢
N.8
2 0.75
-
 py
>
 yt
0.7
 -—
H 0.65
0.6
C3 0.55
0.5
N.4°
.
~~) J.1 03 04 05 06 07 08 ay J
Guided Probability
-=— No Guidance —o— Last Reported
—o Instant. Guid. —— Proj. Guid. (day)
LE
Fig. 7.10 Impact of Accident ProbabilityonAverageDelay
Proj. Guid. (day) / Freq = 2.5 min. / HIVAR / Basic
“0 1.04
{
0 103
1.02
1.01
1
—
0.99
0.98
0.97
0.96
0.95
¢ 0.94
a
1.0 p01 02 03 04 05 06 07 0° ny 1.0
Guided Probability
_a lambda = 25 hrs lambda =5 hrs
i lambda =2.5 hrs —_ No guid.
_ambda is the average accident inter-arrival time
9.
Fig. 7.11a Average Delay of Guided and Unguided Vehicles
Proj. Guid. (day) / Freq = 2.5 min / HIVAR / Basic
C315
~
0 31
30.5
30
of
- 29.5
I~
Q 29
> 28.5
f
28
27.5
27
1.0 § 0. 03 04 05 06 07 08 nN< 0
Guided Probability
= Guided Vehicles  —- Unguided Vehicles
—— No Guidance —o All Vehicles
1Q¢
Fig. 7.11b Maximum Delays of Guided and Unguided Vehicles
Proj. Guid. (perf) / Guid Freq = 2.5 min / LOVAR / Basic
3
 x 4
~ 70
Co
—
“
~
= 65
Cy 60
=
3
50
0.0 0.1 0.2 0.3 04 05 06 07 08 09 1.0
Guided Probability
cm. _ No Guidance —— Guided Vehicles Unguided Vehicles
ig
JA
Fig. 7.12 Delays by Departure Time for Route R1
gr
30 "Me,
5
3
~
—
nid
Time of Day (AM)
3°
CHAPTER 8 CONCLUDING REMARKS AND
DIRECTIONS FOR FURTHER RESEARCH
8.1 Contributions
The major contributions of this dissertation consist of the following:
Proposed Framework
A general framework for the provision of real-time driver information was
formulated. An in-depth investigation of the framework produced the following results:
» identification of the major components
» definition of the functions to be performed by each component
» recognition of the options available for the operationalization of each component
The analysis also classified existing ADIS projects according to the framework
components they embody. Most importantly, principles that ensure that information
provision will characterized by a set of desirable properties were advanced.
Modelling Needs
The second major contribution of the dissertation lies in the analysis it presents
of the modelling requirements of the proposed framework. The analysis focussed on
implementing a DTA for congestion prediction and achieved the following tasks:
developed a modelling framework for the analysis of driver behavior in
the context of information systems; analyzed the role of psychometric
data concerning driver attitudes, perceptions, and stated preferences;
designed a data collection program relevant to dynamic diver behavior
in the ADIS context
reviewed various dynamic network performance modelling methods
identified elements of such models & difficulties encountered
Simulation & Evaluation
Finally, the last part of the dissertation developed an evaluation tool of guidance
strategies. This was achieved by formulating a simulation model that ties together the
zlements of guidance, driver behavior, and network performance.
A case study was conducted that provided insights into:
2age 198
» the role of the different framework components in achieving efficient guidance
» the impact of the different options available for implementing these components
» the value of the suggested principles
» the impact of various parameters on guidance effectiveness
» the impact of various demand/supply characteristics on guidance
8.2 Major Findings
This section describes the major findings of the research reported in this
dissertation. It is subdivided into three major subsections that relate to the framework,
component models, and simulation results.
8.2.1 ADIS Implementation Framework
It was concluded that the actual benefits that may be realized from driver
information systems depend heavily on the quality of guidance that is being provided.
The possible occurrence of adverse impacts from improved information was demonstrated.
This suggests the need for an ADIS implementation framework that is designed to provide
drivers with guidance they can have confidence in and to reduce the potential adverse
impacts of improved information.
In order to insure the effectiveness of real-time driver information systems, the
ADIS implementation framework for has to be based on a dynamic network modeling
approach. Such a modeling approach is needed to assess accurately network performance
as well as to forecast traffic conditions that may exist in the near future in order to
develop real-time diversion strategies to alleviate both recurring and non-recurring
congestion conditions. A framework that would satisfy these requirements is adopted in
this dissertation and has the following major components:
» a surveillance system which collects relevant traffic data
» an O-D updating module which predicts future O-D demand levels
» a Control And Routing module (CAR) which determines the guidance advice
» a COngestion Prediction module (COP) which provide CAR with congestion
information
8.2.2 Modelling Needs of Framework Components
Chapters 4 and 5 of this dissertation analyzed the modelling requirements
associated with the implementation of a DTA as the COP module. The analysis focussed
on the dynamic network performance and dynamic driver behavior elements of the DTA.
The basic conclusion that was reached is that the state of the art in these two fields is not
yet prepared to support ADIS functions and that significant research and development is
still required. Specifically, it was found that existing dynamic traffic assignment models
Page 00
suffer from several shortcomings, including not-truly dynamic assignments or the
existence of inconsistencies in network performance representation. Moreover, there is
a dearth of literature on dynamic O-D estimation using accepted methodologies that do
not require limiting assumptions.
However, much of the difficulty lies in the dynamic driver behavior modelling
aspect of the problem. In this respect, models in the literature fail to consider the fact
that in an ADIS context drivers continually update their perceptions of travel
characteristics and revise their travel choices. When a Driver Information System is
available, a central requirement of the DTA is that it take into account the impact of any
such guidance information on driver decisions when projecting future traffic conditions.
In such a case, a driver’s pre-trip choices and en-route diversion decision will not depend
solely on his own experience with and observation of traffic conditions but will also be
affected by information being provided concerning downstream traffic conditions on the
pre-planned route as well as on alternate routes. Thus, the DTA has to include models
of pre-trip and en-route driver behavior which are sensitive to the drivers’ characteristics
and the motorists’ access to traffic information and route guidance.
Modelling driver behavior in such a dynamic context requires significant
enhancements to existing models. It was indicated that psychometric data may be of help
in explaining some latent factors that enter into such dynamic behavioral decisions.
Finally, the data required to estimate such models has to be collected in a manner that
differs from traditional data collection efforts.
8.2.3 Case Study Results
Based on the case study, the following conclusions were reached:
Issues Related to Proposed Framework
Benefits of Projective Guidance
Guidance based on projected traffic conditions offers advantages over other
guidance scenarios under the following conditions:
» when O-D demand levels are not highly variable form day-to-day
» when good predictors of daily O-D demand levels are available
» with the existence of fast incident detection capabilities
» when network topology does not provide "cheap" opportunities for re-routing
Even with no real-time updating of the O-D demands, and relying solely on the
average historical levels, projective guidance is still to be preferred over other information
provision scenarios if O-D demand patterns display significant peaking occurring at
different times of the analysis period, an effect which may not be captured by other
Page 200
guidance scenarios. Moreover, even with such absence of real-time O-D updating
capabilities, projective guidance offers better results in terms of reductions in maximum
delay experience by guided vehicles.
The Value of the 3 Principles Embodied in the Proposed Framework
The occurrence of adverse effects from guidance at high guided fractions due to
overreaction were eliminated when the following actions were adopted:
» guidance is based on projected traffic conditions (Principle 1)
» a DTA is used for congestion prediction (Principle 2)
» an advanced CAR logic (consistency check) is resorted to (Principle 3)
» the CAR’s temporal update frequency is increased ensuring that only a small
number of drivers receive the same guidance advice
These actions resulted in average delays that did not increase at high guided
probabilities and in significant improvements in guidance validity.
General Guidance Issues
Guidance (in general) is needed most when the traffic conditions are highly
stochastic, as when accidents occur frequently. In such cases, guidance offers a more
significant advantage by guiding drivers away from locations that are highly congested
as a result of accidents. This also means that guidance is more effective in dealing with
incident (non-recurrent) congestion than with recurrent congestion when the network is
rather congested.
The maximum reductions in average delay that were obtained by using guidance
were modest and occurred at market penetration rates that are less than 50%. However,
some of the other performance measures indicated that guidance does offer other more
significant advantages. Specifically, the reductions in maximum delays were found to be
more significant and occurred for the full range of guided probabilities.
Unguided vehicles were also found to receive some benefits from guidance. Such
benefits, however, were smaller than those experienced by guided vehicles under the same
conditions.
Generalization of Results
In the case study it was observed that the overall validity levels were low when
guidance was updated infrequently and based on a basic CAR logic or elementary COP
capabilities. With the detrimental effects this may have on driver compliance, it is not
expected that such a situation would be stable or that it is sustainable for long periods of
time. As such, it is concluded that improvements in validity are of great significance for
Page 201
‘he operation of such driver information systems.
The role of the three principles that were advanced in this dissertation in
improving guidance validity and in controlling the adverse impacts of improved
information was demonstrated. This should have significant implications for the design
of future ADIS projects, especially as the desired market penetration of such projects
approaches levels at which the adverse impacts become most prevalent.
The case study uncovered some interesting tradeoffs between several elements of
the overall framework for implementation of real-time driver information systems. For
instance, a clear tradeoff emerged between the spatial and temporal update frequencies
associated with the CAR element on the one hand and the need for an advanced COP
module on the other hand. Moreover, there was another tradeoff between the use of an
advanced CAR logic or an increased temporal update frequency for the purpose of
controlling overreaction. These issues are to be considered carefully, together with their
implementation implications (cost, hardware, and computational requirements) when
deciding on the implementation most suitable for specific situations.
8.3 Directions for Future Research
8.3.1 Possible Extensions to Case Study
However, several extensions to the case study that would provide further insights
into the design of effective guidance systems are possible, a number of which are
suggested here:
1. Incorporating the drivers’ departure time decision process in addition to route choice
2. Consideration of travel behavior adjustment from day to day
3. The extension of the driver behavior module in the simulation model to include route
selection based on factors other than guidance itself such as:
» previous travel experiences including the perception of travel time reliability
personal differences in attitudes towards route diversion, for example
J compliance as a function of perceived guidance validity
4. The analysis of the use of projective guidance that is based on a statistical prediction
of travel time rather than a DTA
5. The consideration of the case whereby coordination between a number of guidance
points is required and the difficulties associated with the use of projective
guidance in such a case since there would be a need to take into account future
guidance at different locations when setting guidance at one location. The possible
Page 202
use of dynamic programming and/or heuristics to handle such difficulties has to
be investigated.
8.3.2 Possible Directions for Future Research
In addition to the extensions to the case study mentioned above, this dissertation
has uncovered the need for further research on a number of elements related to the
provision of real-time driver information. Some of the possible directions are:
CAR
There exists a need for further research aimed at analyzing the potential benefit
of implementing route-distributive guidance. A related problem involves the formulation
of algorithms to come up with the optimal fractions that can be directed to the alternate
routes between a specific O-D pair.
COP
The possible use of a combination of statistical methods and DTA to predict
congestion is an interesting and important problem. In addition, the use of an off-line
DTA for dynamic O-D estimation is worthy of further research.
DTA
A lot of research items related to the DTA were identified in chapters 4 and 35.
In short, further research on dynamic network performance models is required to come
up with consistent formulations that can be applied for general networks. In addition,
work has to be done on designing the implementation of DTA models on high-power
computers since it is likely that a lot of computational power has to be used if a DTA is
to be used for real-time congestion prediction for real-world size problems.
As far as dynamic driver behavior is concerned, it is apparent that we have barely
scratched the surface. Extensive research still needs to be performed as related to the
following issues:
» modelling imperfect driver knowledge of traffic conditions
modelling information integration and perceptions revision
specifying several models that are required to describe travel adjustment
decisions
specifying what types of data may be obtained from the various sources
identified above
identified how data obtained from the different sources can be combined in a
useful and sound manner
Page 203
Finally, data has to be collected, and models need to be estimated before any
framework for dynamic driver behavior modelling in an ADIS context can become
operational.
In addition to all of the above areas for research there remain two significant
topics that, although not directly relevant to this dissertation, are of great importance to
the implementation of ADIS systems. The legal, institutional and political aspects of
implementation represent one significant difficulty. The other aspect relates to the human
factors considerations involved. Such considerations limit the scope of technologies and
methodologies that may be used in conjunction with ADIS and need to be considered
carefully.
Page 204
REFERENCES
Allen, L.G. (1979). "The Perception of Time", Perception and Psychophysics, Vol. 26,
pp. 340-354.
Allen, R.W., D. Ziedman, T. Rosenthal, A. Stein, J. Torres, and A. Halati (1991).
"Laboratory Assessment of Driver Route Diversion in Response to In-Vehicle
Navigation and Motorist Information Systems", paper presented at the 70th Annual
Meeting of the Transportation Research Board, Washington, D.C., January 1991.
Arnott R., A. de Palma and R. Lindsey (1988). "Information and Time-of-Use Decisions
in Stochastically Congestable Facilities", Northwestern University, Center for
Mathematical Studies in Economics and Management Science, WP No. 778.
Arnott, R., A. de Palma and R. Lindsey (1989). "Does Providing Information to Drivers
Reduce Traffic Congestion?", Transportation Research, Vol. 25B, No. 5
(forthcoming).
Ben-Akiva, M., M. Cyna and A. de Palma (1984a). "Dynamic Model of Peak Period
Congestion", Transportation Research, Vol. 18B, No. 4/5, pp. 339-355.
Ben-Akiva, M., M. J. Bergman, A.J. Daly and R. Ramaswamy (1984b). "Modelling
Interurban Route Choice Behavior", Proceedings of the Ninth International
Symposium on Transportation and Traffic Theory, VNU Science Press, Utrecht,
The Netherlands.
Ben-Akiva, M. (1985). "Dynamic Network Equilibrium Research". Transportation
Research, Vol. 19A, No. 5/6, pp. 429-431.
Ben-Akiva, M., A. de Palma and P. Kanaroglou (1986). "Dynamic Model of Peak Period
Traffic Congestion with Elastic Arrival Rates", Transportation Science, Vol. 20,
No. 2, pp. 164-181.
Ben-Akiva, M. and A. de Palma (1987). "Dynamical Models of Transportation
Networks", Proceedings of PTRC Transport and Planning Summer Annual
Meeting, University of Bath, England.
Ben-Akiva, M., A. de Palma, and I. Kaysi (1991). "Dynamic Network Models and Driver
Information Systems", Transportation Research, Vol. 25A, No. 5 (forthcoming).
Ben-Akiva, M. and I. Kaysi (1991). "Dynamic Driver Behavior Modelling for Advanced
Driver Information Systems (ADIS)", paper presented at the 6th International
Conference on Travel Behavior, Quebec City, Canada, May 1991.
Page 205
Bettman, J. (1979). An Information Processing Theory of Consumer Choice. Addison
Wesley, Reading, Massachusetts.
Bonsall, P.W. and A.D. May (1989). "Evaluation of In-Vehicle Route Guidance Systems",
paper presented at the Sth WCTR, Yokohama, Japan.
Bonsall, P.W. and T. Parry (1991). "Using an Interactive Route-Choice Simulator to
Investigate Drivers’ Compliance with Route Guidance Advice", paper presented
at the 70th Annual Meeting of the Transportation Research Board, Washington,
D.C., January 1991.
Boyce, D. (1988a). "Route Guidance Systems for Improving Urban Travel and Location
Choices", Transportation Research, Vol. 22A, No.4, pp. 275-282.
Boyce, D. (1988b). "Contributions of Transportation Network Modelling to the
Development of a Real-Time Route Guidance System", paper presented at the first
International Conference on Transportation for the Future, Sweden, May 1989.
Boyce, D., B. Ran, and L. LeBlanc (1991). "Dynamic User-Optimal Traffic Assignment:
A New Model and Solution Technique", paper presented at the First Triennial
Symposium on Transportation Analysis, Montreal, Canada.
Brown, S.W. (1985). "Time Perception and attention: The Effect of Prospective and
Retrospective Paradigms and Task Demands on Perceived Duration", Perception
and Psychophysics, Vol. 31, pp. 115-124.
Carey, M. (1986). "A Constraint Qualification for a Dynamic Assignment Model",
Transportation Science, Vol. 20, No. 1, February 1986.
Carey, M. (1987). "Optimal Time-Varying Flows on Congested Networks", Operations
Research, Vol. 35, No. 1, pp. 58-69.
Cascetta, E. and G. Cantarella (1991). "A Day-To-Day and Within-Day Dynamic
Stochastic Assignment Model", Transportation Research, Vol. 25A, No. 5
(forthcoming).
Catling, I. and B. McQueen (1991). "Road Transport Informatics in Europe - Major
Programs and Demonstrations”, IEEE Transactions on Vehicular Technology, Vol.
40. No. 1, pp. 132-140.
Chang, G.L., H.S. Mahmassani and R. Herman (1985). "A Macroparticle Traffic
Simulation Model to Investigate Peak-Period Commuter Decision Dynamics”,
Transportation Research Record, 1005, pp. 107-120.
Page 206
Cremer, M. and H. Keller (1987). "A New Class of Dynamic Methods for the
Identification of Origin-Destination Flows", Transportation Research, Vol. 21B,
pp. 117-132.
de Palma A. and Y. Papageorgiou (1989). "Heuristic Model of Choice", Northwestern
University, Department of Civil Engineering, mimeo.
de Palma, A., P. Hansen and M. Labbe (1990). "Commuters’ Paths with Penalties for
Early or Late Arrival Time", Transportation Science, forthcoming.
Federal Highway Administration (1977). Computer Programs for Urban Transportation
Planning, U.S. Department of Transportation, Washington, D.C.
Federal Highway Administration (1987). "Urban and Suburban Highway Congestion",
Working Paper No. 10, Washington, D.C.
French, R.L. (1990). "Intelligent Vehicle/Highway Systems in Action", ITE Journal,
November 1990, pp. 23-31.
Fricker, J.D. and H.S. Tsay (1984), "Airborne Traffic Advisories: Their Impact and
Value", Transportation Research Record, 996, pp. 20-24.
Friesz, T. L. (1985). "Transportation Network Equilibrium, Design and Aggregation: Key
Developments and Research Opportunities”. Transportation Research, Vol. 19A,
No. 5/6, pp. 413-427.
Fujii, M. (1986). "The CACS Project”, paper presented at the 65th Annual Meeting of the
Transportation Research Board, Washington, DC.
Gartner, N.H., S.B. Gershwin, J.D. Little, and P. Ross (1980). "Pilot Study of Computer
Based Urban Traffic Management", Transportation Research, Vol. 14B,
pp. 203-217.
Gartner, N.H., and R.A. Riess (1987). "Congestion Control in Freeway Corridors: The
IMIS System", In Flow Control of Congested Networks (Edited by Odoni et al.),
Springer-Verlag.
Hamerslag, R. (1988). "A Three Dimensional Assignment in the Time-Space". Paper
presented at the UTSG Annual Conference, 1988, London.
Heiner, R.A. (1983). "The Origin of Predictable Behavior". American Economic Review
73, pp. 560-595.
Pac; 207
Ho, J. (1980). "A Successive Linear Optimization Approach to the Dynamic Traffic
Assignment Problem", Transportation Science, Vol. 14, November, 1980.
Ho, J. (1990). "Solving the Dynamic Traffic Assignment Problem on a Hypercube
Multicomputer”, Transportation Research, Vol. 24B, No. 6.
Hoffman, G. (1989). "Application of Route Guidance in Berlin", paper presented at the
CCCT 89 Symposium, Paris, France, September 1989.
Hoffman, G. and J. Janko (1990). "Travel Times as a Basic Part of the LISB Guidance
Strategy", paper presented at the IEE Road Traffic Control Conference, London.
Horowitz, J.L. (1984). "The Stability of Stochastic Equilibrium in a Two-Link
Transportation Network". Transportation Research, Vol. 18B, No. 1, pp. 13-28.
ITE (1989). A Toolbox for Alleviating Traffic Congestion, Washington, D.C.
Janko, J. (1991). "LISB - The Berlin Route Guidance System", paper presented at the
70th Annual Meeting of the Transportation Research Board, Washington, D.C.,
January 1991.
Janson, B. (1990). "Dynamic Traffic Assignment for Urban Road Networks", paper
presented at the 69th Annual Meeting of the Transportation Research Board,
Washington, D.C., January 1990.
Jeffrey, D.J., K. Russam and D.I. Robertson (1987). "Electronic Route Guidance by
AUTOGUIDE: The Research Background", Traffic Engineering and Control,
October 1987.
Jeffery, D.J. (1988). "Driver Route Guidance Systems: State of the Art", Telematics -
Transportation and Spatial Development Int’l Symposium, The Hague, April 1988.
Kawashima, M. (1989). "Discrete Minimal Radio Zone Communication System and Its
Application to Traffic Management and Control", paper presented at the Seminar
on Urban Traffic Networks, Japan, June 1989.
Kessaci, A., J. Farges, and J. Henry (1989). "On-line Estimation of Turning Movements
and Saturation Flows in PRODYN", Proceedings of the 6th IFAC/IFIP/IFORS
Symposium on Transportation, Paris.
Khattak, A.J., J.L. Schofer, and F.S. Koppelman (1991). "Effect of Real-Time Information
on Commuter Response to Delay", paper presented at the 70th Annual Meeting
of the Transportation Research Board, Washington D.C., January 1991.
Page 208
King, G.F. and T.M. Mast (1987). "Excess Travel: Causes, Extent, and Consequences",
Transportation Research Record, 1111, pp. 126-134.
Kitamura, R. and P. Jovanis (1991). "Driver Response to In-Vehicle Route Guidance
Systems", paper presented at the 70th Annual Meeting of the Transportation
Research Board, Washington, D.C., January 1991.
Kobayashi, F. (1979). "Feasibility Study of Route Guidance System", Transportation
Research Record, 737, pp.107-112.
Koutsopoulos, H. and T. Lotan (1990). "Effectiveness of Motorist Information Systems
in Reducing Traffic Congestion". Forthcoming, Transportation Research Record.
Koutsopoulos, H. and A. Yablonski (1991). "Design Parameters of Advanced Driver
Information Systems: The Case of Incident Congestion and Small Market
Penetration”, paper presented at the VNIS ’91 Conference, Michigan.
Koutsopoulos, H. and H. Xu (1992). "An Information Discounting Routing Strategy for
Advanced Driver Information Systems (ADIS)", paper prepared for presentation
at the 71st Annual Meeting of the Transportation Research Board, Washington,
D.C., January 1992.
Kroes, E. and R. Hamerslag (1990). "A Practical Dynamic Network Equilibrium with
Elastic Demand", paper presented at the 69th Annual Meeting of the
Transportation Research Board, Washington, D.C., January 1990.
Leiser, D. and E. Stern (1988). "Determinants of Subjective Time Estimates in Simulated
Urban Driving". Transportation Research, Vol. 22A, pp. 175-182.
Leonard, D., P. Gower, and N. Taylor (1989). CONTRAM: Structure of the Model. TRRL
Research Report # 178.
Lunenfeld, H. (1989). "Human Factor Considerations of Motorist Navigation and
Information Systems", paper presented at the IEEE International Conference on
Vehicle Navigation & Information Systems, Toronto, Canada, September 1989.
Luque, F.J. and T.L. Friesz (1980). "Dynamic Traffic Assignment Considered as a
Continuous Time Optimal Control Problem", paper presented at the TIMS/ORSA
Joint National Meeting, Washington, D.C., May 1980.
Mahmassani, H. and R. Herman (1984). "Dynamic User Equilibrium Departure Time and
Route Choice on Idealized Traffic Arterials", Transportation Science,
Vol. 18, No. 4.
Page 209
Mahmassani, H. and G.L. Chang (1985). "Dynamic Aspects of Departure Time Choice
Behavior in a Commuting System: Theoretical Framework and Experimental
Analysis", Transportation Research Record, 1037, pp. 88-101.
Mahmassani, H. and G.L. Chang (1986). "Experiments with Departure Times Dynamics
of Urban Commuters", Transportation Research, Vol. 20B, No. 4, pp. 297-320.
Mahmassani, H., G. Chang, and R. Herman (1986). "Individual Decisions and Collective
Effects in a Simulated Traffic System", Transportation Science, Vol. 20, No. 4.
Mahmassani, H., and R. Herman (1988). "Interactive Experiments for the Study of
Tripmaker Behavior Dynamics in Congested Commuting Systems", Jones P. (ed.),
New Developments in Dynamic and Activity-Based Approaches, Grover Press.
Mahmassani, H. and R. Jayakrishnan (1989). "System Performance and User Response
Under Real-Time Information in a Congested Traffic Corridor", paper presented
at the Seminar on Urban Traffic Networks, Capri, Italy, June 1989.
Mahmassani, H., C. Caplice, and C. Walton (1990). "Characteristics of Urban Commuter
Behavior: Switching Propensity and Use of Information", paper presented at the
69th Annual Meeting of the Transportation Research Board, Washington D.C.
Mahmassani, H. and P.S. Chen (1991). "Comparative Assessment of Origin-Based and
En-Route Real-Time Information Under Alternative User Behavior Rules", paper
Presented at 70th Annual Meeting of the Transportation Research Board,
Washington, D.C., January 1991.
Manheim, M.L. and E.R. Ruiter (1970). "DODOTRANS I: A Decision Oriented Computer
Language for Analysis of Multimodal Transportation Systems”, Highway Research
Record 314, pp. 135-163.
Manheim, M.L. (1979). Fundamentals of Transportation Systems Analysis. MIT Press,
Cambridge, Mass.
Mauro, V. (1991). "Evaluation of Dynamic Network Control: Simulation Results Using
NEMIS Urban Microsimulator”, paper Presented at 70th Annual Meeting of the
Transportation Research Board, Washington, D.C., January 1991.
Merchant, D.K. and G.L. Nemhauser (1978a). "A Model and an Algorithm for the
Dynamic Traffic Assignment Problem", Transportation Science, Vol. 12, No. 3,
pp. 183-199
Merchant, D.K. and G.L. Nemhauser (1978b). "Optimality Conditions for a Dynamic
Traffic Assignment Model", Transportation Science, Vol. 12, No. 3, pp. 201-207.
Page 2.0
Newell, A. and H. A. Simon (1972). Human Problem Solving. Prentice-Hall, Englewood
Cliffs, New Jersey.
Nihan, N. and G. Davis (1987). "Recursive Estimation of Origin-Destination Matrices
from Input/Output Counts", Transportation Research, Vol. 21B, pp. 149-163.
OECD (1987). "Dynamic Traffic Management in Urban and Suburban Road Systems",
Organization for Economic Co-Operation and Development, Paris.
Okutani, I. and Y.J. Stephanedes (1984). "Dynamic Prediction of Traffic Volume Through
Kalman Filtering Theory", Transportation Research, Vol. 18B, No. 1, pp. 1-11.
Papageorgiou, M. and A. Messmer (1991). "Dynamic Network Traffic Assignment and
Route Guidance Via Feedback Regulation”, paper presented at the 70th Annual
Meeting of the Transportation Research Board, Washington, D.C., January 1991.
Pisarski, A. (1987). Commuting in America, Eno Foundation.
Rasmussen, J. (1986). Information Processing and Human-Machine Interaction, An
Approach to Cognitive Engineering. North-Holland, New York.
Rillings, J.H. and R.J. Betsold (1991). "Advanced Driver Information Systems",
IEEE Transactions on Vehicular Technology, Vol. 40, No. 1, pp. 31-40.
Sheffi, Y. (1985). Urban Transportation Networks, Prentice-Hall, Englewood Cliffs, N.J.
Sheridan, T.B. (1990). "Human Factors of Driver-Vehicle Interaction in the IVHS
Environment", report to the National Highway Traffic Safety Administration,
Center for Transportation Studies, MIT.
Simon, H. (1967). "Motivational and Emotional Controls of Cognition". Psychological
Review, Vol. 74, pp. 29-39.
Slovic, P. (1972). From Shakespeare to Simon: Speculations - and Some Evidence -
About Man’s Ability to Process Information. Research Monograph 12, Oregon
Research Institute.
Smith, J.C. and K. Russam (1989). "Some Possible Effects of AUTOGUIDE on Traffic
in London", paper presented at the IEEE International Conference on Vehicle
Navigation & Information Systems, Toronto, Canada, September 1989.
Stephanedes, Y.J., P.G. Michalopoulos, and R.A. Plum (1981). "Improved Estimation of
Traffic Flow for Real-Time Control", Transportation Research Record, 795,
pp. 28-39.
Page 211i
Stern, E. and P.H.L. Bovy (1989). Theory and Models of Route Choice Behavior, OSPA.
Stevens, S.S. (1967). "Intensity Functions in Sensory Systems", International Journal of
Neurology, Vol. 6, pp. 202-209.
Takasaki, G. M. and P. F. Wasielewski (1989). "Research in Vehicle Information
Systems at General Motors", paper presented at the IEEE International Conference
on Vehicle Navigation & Information Systems, Toronto, Canada, September 1989.
Taylor, N. (1990). CONTRAM 35: An Enhanced Traffic Assignment Model. TRRL
Research Report # 249.
Transportation Research Board (1985). Highway Capacity Manual, Special Report 209,
Washington, D.C.
Tsuji, H. et al. (1985). "A Stochastic Approach for Estimating the Effectiveness of a
Route Guidance System and its Related Parameters", Transportation Science, Vol.
19, No. 4, pp. 333-351.
U.S. Department of Transportation (1989). Discussion Paper on Intelligent
Vehicle-Highway Systems.
Vickrey, W. (1969). "Congestion Theory and Transport Investment”, American Economic
Review, 59, pp. 251-261.
Vythoulkas, P. (1990). "A Dynamic Stochastic Assignment Model for the Analysis of
General Networks", Transportation Research, Vol. 24B, No. 6.
Wie, B. (1989). "Dynamic System Optimal Traffic Assignment on Congested
Multidestination Networks", paper presented at the Sth World Conference on
Transport research, Yokahoma, Japan, July 1989.
Wie, B. (1990). "Dynamic User Optimal Traffic Assignment: A Control Theoretic
Formulation", Transportation Research, Vol. 24B, No. 6.
Yablonski, A. (1991). "An Investigation of Design Parameters of Advanced Driver
Information Systems". Masters Thesis, Massachusetts Institute of Technology.
Ygnace, J.L. (1991). "An Example of Consumer Acceptance of Route Guidance
Technologies", paper presented at the 70th Annual Meeting of the Transportation
Research Board, Washington, D.C., January 1991.
Page 212
Appendix A
A Survey of
Your Home to M.I.T Commute
Center for Transportation Studies
Massachusetts Institute of Technology
May 1991
Attached is a questionnaire prepared by a research team in the Center for Transportation Studies
at MIT, in cooperation with the MIT Planning Office. This questionnaire is part of an on going
research at MIT in the area of Intelligent Vehicle Highway Systems (IVHS), which aim at
reducing congestion by providing on-line and user specific information to commuters using a
variety of technologies. It is designed to provide transportation planners with a better
understanding of the routes you follow on your daily commute, your preferences when it comes
to choosing those routes, your reliance on traffic reports, and your parking needs at MIT
You are kindly requested to fill out this questionnaire. By filling out this questionnaire you will
help us design systems which will provide drivers with relevant, useful, and reliable information.
It will also allow the MIT Planning Office to be more responsive to your preferences and needs
in the future. All responses are strictly confidential.
How to fill out this questionnaire:
The questionnaire consists of two parts; Part I asks you about your usual commute to MIT, and
Part II about your specific commute during the week of May 6 to May 10. Please fill out both
parts and return them in the envelope provided by May 16, 1991.
Please feel free to write comments in the margins wherever appropriate.
Thank you in advance for your time and effort.
PART I: Your Usual Commute to MIT
l. In a typical § day work week, how many times do you use each of the following modes to commute to work:
_ Drive Alone __Carpool Driver __Carpool Passenger
__ Public Transportation __other:
2. In a typical 5 day work week, how many days do you come to MIT? days
3. Do you come to MIT on weekends? Ooften [OJoccasionally never
FOR THE REST OF THIS SURVEY WE ARE INTERESTED IN DRIVERS BEHAVIOR AND
CHOICES. IF YOU NEVER DRIVE TO WORK, PLEASE DO NOT FILL THE REST OF THE
SURVEY. THANK YOU FOR YOUR WILLINGNESS TO PARTICIPATE.
’
4. When driving to MIT:
a. What time do you usually leave home? hour ___ min Oam O p.m.
b. What time do you usually arrive at MIT? hour ___ min ___ Oam O p.m.
5. How much flexibility do you have in choosing the time you arrive at work on a daily basis?
[J none O up to 15 minutes [J 16- 30 minutes
7] 31 - 60 minutes [J more than an hour
6. Think about a typical car trip from your home to work. Assume “regular traffic conditions, i.e. no extreme traffic
delays, no major incidents and no weather related problems. Under these conditions, how long does it usually take you
'0 drive from your home to work? Please specify a range (e.g. from 40 to 55 minutes): from to minutes.
7. How often does your driving time to work exceed the range you specified in question 6?
O very often (more than once a week) O often (approx. once a week)
O occasionally (approx. twice a month) O rarely (approx. once a month)
J very rarely (less than once a month)
8. What is the shortest driving time you have ever experienced during your home to work commute? __ minutes
9. What is the longest driving time you have ever experienced during your home to work commute? minutes
10. In a typical 5 day work week, how many significantly different routes from home to MIT do you use? routes
(by “significantly different" we mean routes which almost do not overlap, for example: Mass. Pike and Route 9, or 93
and Morrissey Blvd.)
11. Please describe below your most frequently used route to MIT by indicating the major streets, highways, and
bridges that compose the route:__
12. Do you usually make stops on your way to MIT? Ono» proceed to question 14
(1 yes, total duration of stops is approx. minutes
13. What is the purpose of your stops? O drop a passenger O pick a passenger
(Jeat [run errands OJ fin gas  J other:
14. Where do you usually park your car?
[Jat an MIT parking lot [OJ on street
[J on street at a meter [OJ other:
5. To which MIT parking facility do you have a sticker?
6. How long does it usually take you to get from your parked car to your MIT destination? __ minutes
17. What time do you usually leave MIT? hour min O.m O p.m.
)
Your Attitudes and Preferences
18. On a scale of 1 to 5, where 1 indicates "strongly disagree” and 5 indicates “strongly agree”, indicate your level of
agreement with the following statements by checking the appropriate box:
strongly strongly not
disagree agree relevant
 2 3 4 53
I am very familiar with at least 2 significantly¢differentroutesto work
I often change my planned route while driving
1
I like discovering new routes nif, a - » vr wih —
[ am willing to try new routes to avoid traffic delays
I always listen to radio traffic reports
I usually follow the recommendations of radio traffic reports
——
+
Radio traffic reports are usually reliable
When traffic reports are different from my own observation, I ignore them
I often change my route after listening to radio traffic reports:
I trust my own judgement more than traffic reports
Traffic reports do not provide relevant information
I am willing to pay in order to get more useful traffic information
19. On a scale of 1 to 5 where 1 indicates "not important at all" and § indicates "very important”, indicate the importance
of the following factors in choosing your route to work:
not important very
at all important
/, 3
Time of day |
Commute time
Habit
 rr———
Time spent stopped in traffic
Number of traffic lights
Traffic reports
Risk of delay
Weather
dig
ABOUT YOURSELF
The information requested in this section relates to your personal and household data. We need
this information to better understand how personal and family characteristics affect commuting
choices. All information collected will remain strictly confidential
20. Sex: O Male
21. Marital status: [J Married
22. What is your age group?
[O Less than 20 years 0320-29years O30 - 39 years
[J40 - 49 years 50 - 64 years 0 65 or older
23. What is the highest level of education you have completed?
0 High school or less 0 Sone College
OO Graduated College [J Post graduate work
24. What is your home Zip Code? ____
25. How long have you lived at your present home address? years
26. Do you own or rent your dwelling unit?
O own O rent
27. How many persons including yourself live in your household? persons
28. What is the total number of automobiles owned by your household? ____ automobiles
29. What is your household’s approximate yearly income from all sources (before taxes)?
0 Less than $20,000 0d $20,000 - $40,000 0 $40,000 - $60,000
Od $60,000 - $80,000 [$80,000 - $100,000 J More than $100,000
30. How long have you worked at your present job location? _ years
31. Which of the following categories best describes your position?
0 Undergraduate Student 0 Graduate Student O Academic Staff
0 Tenured Faculty (J Non-Tenured Faculty [J Administrative Staff
O Service O Support Staff [0 Research Staff
7] Other:
General
32. Write any other comments related to your daily commute to work, the way you choose and follow routes, your attitude
towards traffic reports, and your parking needs (optional):
) 16h
PART II: YOUR COMMUTE TODAY
THIS PART OF THE QUESTIONNAIRE RELATES TO YOUR DRIVING BEHAVIOR
DURING A SPECIFIC WEEK. IT CONTAINS 5 IDENTICAL SECTIONS, ONE FOR EACH
DAY FROM MAY 6 TO MAY 10. PLEASE FILL OUT EACH SECTION AFTER YOU
HAVE COMPLETED YOUR COMMUTE TO MIT FOR THAT DAY.
EVEN IF YOU MISS FILLING OUT THE SURVEY FOR ONE DAY DUE TO ANY
REASON, PLEASE FILL OUT THE INFORMATION FOR THE SUBSEQUENT DAYS.
Would you be willing to respond to a follow-up questionnaire
about your driving behavior during the next year?
If so, please fill out the following:
Name:
MIT address:
Participants will also receive a summary of our findings.
[f you wish to remain anonymous, please take the time to fill out PART II anyway.
All responses are strictly confidential.
This part is designed to monitor your daily commute from home to MIT. We need to know
whether the route you followed to work everyday was influenced by traffic information you
listened to, by unusual traffic conditions you encountered on your way to work, or by commuting
experience on the previous day. We are also interested in diversion decisions you made while
on your way to MIT.
) 1
YOUR COMMUTE FOR MONDAY, MAY 6, 1991
(1) How did you get to MIT today? [J Drive Alone dd Carpool Driver
O Carpool Passenger [J pubic Transportation U] other:
IF YOU DID NOT DRIVE YOURSELF TO WORK TODAY, PLEASE IGNORE THE
QUESTIONS FOR TODAY. CONTINUE TOMORROW WITH THE NEXT SECTION.
(2) Did you receive traffic information before you left home today?
C] yes, from radio O yes, from TV
Ono — proceed to question 5
(3) Did the information that you received before leaving home influence your route choice for today?
(alot [J somewhat O very little OJ not at an
(4) What did the information you received indicate about traffic conditions on the route you decided to take?
(1 much worse than usual [J worse than usual [J usual traffic conditions
{J better than usual [J much better than usual [J 1 don't remember
(J no information O other:
(5) Once you started your trip, what were the traffic conditions you observed at the beginning of your trip?
[J much worse than usual [J worse than usual [J usual wraffic conditions
[J better than usual [J much better than usual O other:
(6) While driving, did you receive any information about the route you were following?
J yes, which radio station? ______ Ono » proceed to question 8
(7) What did the information indicate about traffic conditions on the route you were following?
[J much worse than usual . [J worse than usual [1 usual wraffic conditions
[J better than usual U] much better than usual [J 1 don’t remember
[J no information [J other:_
(8) After you started your trip to work, was there a way to switch to another route that will take you to your destination?
O yes Ono - proceed to question 14
(9) While driving, did you switch from the route you were following ?
O yes One > proceed to question 14
(10) Before you switched routes, did you get any radio information about the route you switched to?
O yes Ono » proceed to question 12
(11) What did the information indicate about traffic conditions on the route you switched to?
[J much worse than usual [J worse than usual [3 usual traffic conditions
[7 better than usual [J much better than usual [J 1 don’t remember
[no information 1 other:_____
)]8