Traffic Theory

TRAFFIC THEORY
INTERNATIONAL SERIES IN
OPERATIONS RESEARCH & MANAGEMENT SCIENCE
Frederick S. Hillier, Series Editor Stanford University
Fang, S.-C., Rajasekera, J.R. & Tsao, H.-S.J. / ENTROPY OPTIMIZATION
AND MATHEMATICAL PROGRAMMING
Yu, G. / OPERATIONS RESEARCH IN THE AIRLINE INDUSTRY
Ho, T.-H. & Tang, C. S. / PRODUCT VARIETY MANAGEMENT
El-Taha, M. & Stidham, S. / SAMPLE-PATH ANALYSIS OF QUEUEING SYSTEMS
Miettinen, K. M. / NONLINEAR MULTIOBJECTIVE OPTIMIZATION
Chao, H. & Huntington, H. G. / DESIGNING COMPETITIVE ELECTRICITY MARKETS
Weglarz, J. / PROJECT SCHEDULING: Recent Models, Algorithms & Applications
Sahin, I. & Polatoglu, H. / QUALITY, WARRANTY AND PREVENTIVE MAINTENANCE
Tavares, L. V. / ADVANCED MODELS FOR PROJECT MANAGEMENT
Tayur, S., Ganeshan, R. & Magazine, M. / QUANTITATIVE MODELING FOR SUPPLY
CHAIN MANAGEMENT
Weyant, J./ ENERGY AND ENVIRONMENTAL POLICY MODELING
Shanthikumar, J.G. & Sumita, U./APPLIED PROBABILITY AND STOCHASTIC PROCESSES
Liu, B. & Esogbue, A.O. / DECISION CRITERIA AND OPTIMAL INVENTORY PROCESSES
Gal, T., Stewart, T.J., Hanne, T./ MULTICRITERIA DECISION MAKING: Advances in MCDM
Models, Algorithms, Theory, and Applications
Fox, B. L./ STRATEGIES FOR QUASI-MONTE CARLO
Hall, R.W. / HANDBOOK OF TRANSPORTATION SCIENCE
Grassman, W.K./ COMPUTATIONAL PROBABILITY
Pomerol, J-C. & Barba-Romero, S./ MULTICRITERION DECISION IN MANAGEMENT
Axsäter, S. / INVENTORY CONTROL
Wolkowicz, H., Saigal, R., Vandenberghe, L./ HANDBOOK OF SEMI-DEFINITE
PROGRAMMING: Theory, Algorithms, and Applications
Hobbs, B. F. & Meier, P. / ENERGY DECISIONS AND THE ENVIRONMENT: A Guide
to the Use of Multicriteria Methods
Dar-El, E./ HUMAN LEARNING: From Learning Curves to Learning Organizations
Armstrong, J. S./ PRINCIPLES OF FORECASTING: A Handbook for Researchers and
Practitioners
Balsamo, S., Personé, V., Onvural, R./ ANALYSIS OF QUEUEING NETWORKS WITH
BLOCKING
Bouyssou, D. et al/ EVALUATION AND DECISION MODELS: A Critical Perspective
Hanne, T./ INTELLIGENT STRATEGIES FOR META MULTIPLE CRITERIA DECISION MAKING
Saaty, T. & Vargas, L./ MODELS, METHODS, CONCEPTS & APPLICATIONS OF THE
ANALYTIC HIERARCHY PROCESS
Chatterjee, K. & Samuelson, W./ GAME THEORY AND BUSINESS APPLICATIONS
Hobbs, B. et al/ THE NEXT GENERATION OF ELECTRIC POWER UNIT COMMITMENT MODELS
Vanderbei, R.J./ LINEAR PROGRAMMING: Foundations and Extensions, 2nd Ed.
Kimms, A./ MATHEMATICAL PROGRAMMING AND FINANCIAL OBJECTIVES FOR
SCHEDULING PROJECTS
Baptiste, P., Le Pape, C. & Nuijten, W./ CONSTRAINT-BASED SCHEDULING
Feinberg, E. & Shwartz, A./ HANDBOOK OF MARKOV DECISION PROCESSES: Methods
and Applications
Ramík, J. & Vlach, M. / GENERALIZED CONCAVITY IN FUZZY OPTIMIZATION
AND DECISION ANALYSIS
Song, J. & Yao, D. / SUPPLY CHAIN STRUCTURES: Coordination, Information and
Optimization
Kozan, E. & Ohuchi, A./ OPERATIONS RESEARCH/ MANAGEMENT SCIENCE AT WORK
Bouyssou et al/ AIDING DECISIONS WITH MULTIPLE CRITERIA: Essays in
Honor of Bernard Roy
Cox, Louis Anthony, Jr./ RISK ANALYSIS: Foundations, Models and Methods
Dror, M., L’Ecuyer, P. & Szidarovszky, F. / MODELING UNCERTAINTY: An Examination
of Stochastic Theory, Methods, and Applications
Dokuchaev, N./ DYNAMIC PORTFOLIO STRATEGIES: Quantitative Methods and Empirical Rules
for Incomplete Information
Sarker, R., Mohammadian, M. & Yao, X./ EVOLUTIONARY OPTIMIZATION
Demeulemeester, R. & Herroelen, W./ PROJECT SCHEDULING: A Research Handbook
TRAFFIC THEORY
by
Denos C. Gazis
KLUWER ACADEMIC PUBLISHERS

NEW YORK, BOSTON, DORDRECHT, LONDON, MOSCOW
eBook ISBN: 0-306-48217-7
Print ISBN: 1-4020-7095-0
©2002 Kluwer Academic Publishers

New York, Boston, Dordrecht, London, Moscow
Print ©2002 Kluwer Academic Publishers

Dordrecht
All rights reserved
No part of this eBook may be reproduced or transmitted in any form or by any means, electronic,
mechanical, recording, or otherwise, without written consent from the Publisher
Created in the United States of America
Visit Kluwer Online at: http://kluweronline.com

and Kluwer's eBookstore at: http://ebooks.kluweronline.com
To my wife, Jean
This page intentionally left blank
CONTENTS
CHAPTER 1. TRAFFIC FLOW THEORIES 1
Macroscopic Traffic Flow Theory – Kinematic

Waves in Traffic 1
Shock waves 6
Flow versus concentration relationships 9
The Boltzman-like Model of Traffic 17

The two-fluid model of traffic 19
Microscopic Traffic Flow Theory – Car-Following Models 21

Local stability 22
Asymptotic stability 26
Acceleration noise 31
Non-linear car-following models 33
Some Questionable Paradoxes 36

The Smeed paradox 36
The Braess paradox 38
Additional Flow Theory Work, and Future Challenges 40

viii
CHAPTER 2. QUEUEING AND DELAYS AT

ISOLATED INTERSECTIONS 45
Traffic Characteristics 46
The Gap Acceptance Function 49
The Delay to a Single Car 50
Delays to Pedestrians 63
Queueing and Delays of Vehicles 71
The Delay at a Traffic Signal 77

Evaluation of delays 79
Optimization of traffic signal settings 87
The vehicle-actuated traffic signal 89
CHAPTER 3. TRAFFIC CONTROL 101
Objectives of Traffic Control 103
Single, Isolated Intersection 105
Synchronization Schemes for Arterial Traffic 106

A mixed integer-linear programming approach 112
The effect of queues on progression 115
The TRANSYT method 118
The Combination method 121
The SIGOP program 124
Traffic Responsive Operation of Traffic Lights 128

Single intersection 128
Systems of intersections 132
Additional contributions:
The SCOOT program 135
The OPAC, PRODYN, CRONOS, and COP programs 135
The Onset of Oversaturation 136

ix
Oversaturated Systems 138

A single oversaturated intersection 138
Complex oversaturated systems 154
Oversaturated Store-and-Forward Networks 165
Freeway Control 171
Systems Affected by Geometric Details 172
CHAPTER 4. TRAFFIC GENERATION, DISTRIBUTION,

AND ASSIGNMENT 185
Network Representation of a Transportation System 186
Trip Generation and Distribution 189

Trip generation 190
Trip distribution 198
Time-Independent Traffic Assignment 206

Deterministic traffic assignment models 207
Discrete choice, stochastic, models 211
Global Network Optimization Models 219
Dynamic Traffic Assignment 225
Traffic Assignment in Congested Systems 229

Delivery of traveller information services 237
APPENDIX
Application of Kalman Filtering for Density

Estimation in Traffic Networks 245
Index 257
Preface
“ Everything should be made as simple

as possible—but not simpler”
Albert Einstein
Traffic Theory, like all other sciences, aims at understanding and

improving a physical phenomenon. The phenomenon addressed by Traffic
Theory is, of course, automobile traffic, and the problems associated with it
such as traffic congestion. But what causes congestion?
Some time in the 1970s, Doxiades coined the term "oikomenopolis" (and
"oikistics") to describe the world as man's living space. In Doxiades' terms,
persons are associated with a living space around them, which describes the
range that they can cover through personal presence. In the days of old,
when the movement of people was limited to walking, an individual
oikomenopolis did not intersect many others. The automobile changed all
that. The term "range of good" was also coined to describe the maximal
distance a person can and is willing to go in order to do something useful or
buy something. Traffic congestion is caused by the intersection of a
multitude of such "ranges of good" of many people exercising their range
utilisation at the same time. Urban structures containing desirable structures
contribute to this intersection of "ranges of good".
xii Preface
In a biblical mood, I opened a 1970 paper entitled "Traffic Control --

From Hand Signals to Computers" with the sentence: "In the beginning there
was the Ford". The Ford was indeed the beginning of man's modern age of
transportation, and the enlargement of individual oikomenopolis, and this
beginning took place shortly after the beginning of the 20th century. For a
few decades after that, people exploited the capabilities of the new medium
by trial and error, (and I am suppressing here the urge to add "mostly error").
The Ford, and its follow up companion cars, increased the mobility of people
and gave birth to urban sprawl and its concomitant traffic congestion. Traffic
signals, borrowed from railroad train management technology, were brought
into the scene to improve safety and combat congestion, with mixed results
due to inadequate understanding of the inner workings of automobile traffic.
And then Traffic Theory was born.
Around the middle of the 20th century, some notable scientists laid the
foundations of understanding of automobile traffic. You can find the names
of some of these scientists in an article I wrote for the 50th anniversary issue
of Operations Research, entitled "The Origins of Traffic Theory". Much
work followed after those pioneering contributions, some of it productive
and some readily classified as "solutions in search of a problem". Traffic
Theory followed the tradition of other physical sciences, with many
scientists entering the field in order to apply what they knew to model
traffic. Some of them did not make the necessary effort to justify their
modeling effort in terms of realistic models of human behaviour, often
ignoring the admonition of Albert Einstein shown at the beginning of this
preamble. Comments on such misguided effort will be made later in this
book. Suffice it to say here that mere good fit with data is not sufficient to
justify a theory, which must also be based of a rational consideration of
physical properties of the things that the theory purports to describe. This
book concentrates on theoretical work tested over the years, which is based
on such foundations of proper pursuit of science.
Traffic theory, after modelling of traffic movement, moved into

modelling of interactions of traffic movements aimed at improving them
through appropriate control of traffic movement through devices such as
traffic signals or other means of regulating the rate of flow of traffic past a
control point. Thus, traffic control theory for congested or uncongested
systems was developed, once more with mixed success in producing the
desired improvement of traffic flow.
If we were to evaluate the overall effect of progress in traffic theory on

traffic management, we would find "mixed success" as the universal
xiii
measure of this effect. This is so in spite of considerable effort made in

recent years to utilise advanced computer technology in order to improve the
management of transportation systems, under the label of "Intelligent
Transportation Systems (ITS)". A short explanation of why success is mixed
may be the following:
a) Some good traffic theoretical models are inadequately used.

b) Some other good theoretical models are missing, and inadequate
attention is paid toward their development.
c) Some of the suggested solutions to the traffic problem are aptly
described by the edict of Harold Mencken : “For every problem there is a
solution that is simple, neat, and wrong” .
In this book, I present key models of traffic flow and associated traffic
phenomena such as conflicts in traffic, traffic generation and assignment,
and traffic control. The book builds on my earlier work, (*Traffic Science,
edited by Denos C. Gazis, Wiley Interscience, New York, 1974), and the
work of other transportation scholars. I have focused on identifying the
appropriate use of the various models developed over the years for the
improvement of traffic systems. I discuss how proper use of some of these
models may accelerate the successful deployment of ITS, and also
enumerate opportunities for development of additional models needed for
continued improvement of ITS. The book comprises four chapters, namely:
1. Traffic Flow Theories: A description of the movement of traffic and the

interaction of its component vehicles with each other.
2. Delay Problems at Isolated Intersections: A description of queuing and
delays at isolated intersections.
3. Traffic Control: A discussion of algorithms for optimum control of
single intersections and systems of intersections, including networks.
4. Traffic Assignment: A discussion of traffic generation, distribution and
assignment in a transportation network.
I would like to express my gratitude to Walter Helly, Donald McNeil and

George Weiss, and the late Leslie Edie, whose contributions to the
aforementioned book “Traffic Science” provided the basis for the
corresponding chapters of this book.
The book is intended as a textbook for a college Transportation Science

curriculum, and as a reference book for researchers in Transportation
Science. I have concentrated in presenting the fundamental concepts and
methods in the various areas of traffic theory, but not necessarily covering
xiv Preface
all the contributions, which do not alter substantially these concepts and/or
methods. Some of these contributions are listed in the References section of
each chapter as “additional references”.
Denos Gazis
Chapter 1
Traffic Flow Theories
Traffic flow theories may be viewed as the foundation of Traffic Science.

They are intended to provide understanding of phenomena related to the
movement of individual vehicles along a highway as they interact with
neighbouring vehicles. It is the consequences of such interactions that
determine the fundamental characteristics of highways, such as their
capacity and their ability to sustain various levels of traffic flow.
Traffic flow theories were among the earliest contributions to Traffic

Science, and they have included both a Macroscopic, and a Microscopic
treatment. The macroscopic treatment views traffic as a continuum akin to a
fluid moving along a duct which is the highway. The microscopic treatment
considers the movement of individual vehicles as they interact with each
other.
1.1 MACROSCOPIC TRAFFIC FLOW THEORY –

KINEMATIC WAVES IN TRAFFIC
Lighthill and Whitham (1955) laid the foundation of macroscopic traffic

flow theory in their seminal paper describing traffic as a fluid moving along
a duct. They discussed traffic along a reasonably crowded road, with no
appreciable gaps between individual vehicles. In this case, traffic may be
viewed as a continuum, and its characteristics correspond to the physical
characteristics of the imaging fluid. At this point, some definitions are in
order:
2 Chapter 1
q = the "flow rate" of traffic along a highway lane, in vehicles per hour.
u = the (average) speed of the traffic, in miles per hour.
k = the "density" of traffic, in vehicles per mile.
The above quantities satisfy the relationship
At this point, we should discuss how these quantities are measured.

Basically, traffic is observed in two ways. First, observations may be made
at two or more fixed points along a roadway, with measurements taken at the
time of occurrence of traffic events at each of these points. Events, in this
case, are the arrivals of some parts of vehicles, e.g. the front, at each of the
observation points. Second, observations may be made at two or more
instants of time. The events observed at each instant are the positions of each
vehicle on the roadway. The first way uses a short segment of the roadway,
dx , and a large interval of time T ; the second uses a short interval of time,
dt , and a large segment of the roadway X. This is illustrated in Figure 1,
where the trajectories of some 25 vehicles are shown in the space versus
time domain, with some of these trajectories traversing the X versus T
rectangular domain.
If is the total number of vehicles traversing dx in time T,

with at speed we can compute the values of the three traffic variables
by looking at the vehicles as distributed either in time or in space. First,
1. Traffic Flow Theories 3
considering the vehicles as distributed in time, and remembering that in this

case the average speed of these vehicles must be their harmonic average, we
obtain
By substituting where is the travel time across dx , and by

multiplying the expression for q by dx in numerator and denominator, we
obtain
Now, assuming a space distribution of vehicles observed on a stretch of

length X over time dt, we let be the total number of vehicles
observed and remember that in this case the mean speed is the space mean.
We thus obtain the following values of the three variables:
4 Chapter 1
Substituting for where is the distance traveled by vehicles at

speed in the time interval dt, we obtain
A comparison of Eqs. 2 and 3 shows that they are equivalent, and lead to
the following general definitions for the three traffic variables:
The mean flow q of vehicles travelling over a section of a roadway of

length X during a time interval T is the total distance travelled on the
roadway by all vehicles which were on this section during any part of time T
divided by XT, the area of the space-time domain observed.
The mean concentration k of vehicles travelling over a roadway section
of length X during a time T is the total time spent by all the vehicles on X
during time T divided by XT.
The mean speed u of vehicles travelling over a roadway section of
length X during a time interval T is the total distance travelled on the
section by all vehicles that were on it for any part of time T divided by the
total time spent by all vehicles on the section during time T.
These definitions may be further generalised by using an area of any

shape of the space-time domain, such as a circular or elliptical, rather than
the rectangular one shown in Figure 1. In any case, these definitions make
sense if the trajectories of the vehicles do not deviate very much from a
steady state situation corresponding to a bunch of parallel trajectories.
In their discussion of traffic phenomena, Lighthill and Whitham discussed

essentially transitions between steady state situations. They further assumed
that q and k were related in a fashion described by what came to be known
as the “Fundamental Diagram”, shown in Figure 2. The rationale for the
existence of a Fundamental Diagram is simple. Clearly, at zero density
(k=0), there is no traffic and consequently q=0 . Traffic flow is also zero at
some “jam density” , , when cars are stopped bumper-to-bumper.
Assuming a continuous relationship between q and k, by Rolle’s theorem
there must be at least one maximum point between 0 and . The

assumption of one maximum point produces the Fundamental Diagram of
Figure 2, which describes the following behaviour of traffic. As the density
increases from zero, the flow also increases, starting with an almost linear
rate of increase and a lower rate as the density increases and the average
speed of the traffic, shown in Figure 3, decreases. The flow reaches a
maximum at some optimum speed and density, and then starts decreasing
with increasing density, reaching zero at jam density, when the speed is also
zero.
Finally, Lighthill and Whitham stipulated a principle of conservation of

vehicles, expressed by the relationship
6 Chapter 1
Assuming q = q (k) , we derive from Eq. 4 the following relationship
where is the slope of the tangent at the corresponding point on

the Fundamental Diagram. The solution of Eq. 5 is
where F is an arbitrary function. Equation 6 implies that inhomogeneities

such as changes in concentration of cars propagate along a stream of traffic
at constant speed which is positive or negative with respect to
a stationary observer, depending on whether the concentration is below or
above the optimum concentration corresponding to maximum q ( Figure 2).
For example, the speed V will be positive for point 1 in Figure 2, and
negative for point 2.
1.1.1 Shockwaves.
An important derivative of the above theory is the discussion of the

propagation of disturbances when traffic changes from one steady state to
another. Let us assume, for example, that traffic conditions change by a shift
from concentration to concentration (Figure 4). The transition
takes place at point S , which in general moves along the roadway, and we
are about to find out how it does so. The corresponding points on the flow
versus concentration curve are points 1 and 2 in Figure 2. We can
visualise the movement of individual vehicles corresponding to these two
points as shown in Figure 5, where the trajectory of each vehicle transitions
from a straight line with slope to one with slope The movement of
point S of Figure 4 is represented by the line separating the domains 1 and
2 in Figure 5. The speed of propagation of the “shock wave” associated with
the change from state 1 to state 2 is the slope of this line. If it is positive,
the shock propagates forward with respect to a stationary observer. If
negative, the shock propagates backwards. If we denote this speed of
propagation of the shock wave as we can compute its value as follows.
Let us use the notation
space mean speed of vehicles in region 1

space mean speed of vehicles in region 2
speed of vehicles in region 1 relative to the moving
line S (Fig. 4).
= speed of vehicles in region 2 relative to the moving
line S.
The number N of vehicles crossing the dividing line S is

8 Chapter 1
which leads to the relationship
Equation 7 is a restatement of the principle of conservation of vehicles,

and may be rewritten in the form
and using Equation 1, (q = u k ),
It may be observed that Equation 9 can be obtained directly on geometric

grounds from Figure 5, which obviously satisfies the principle of
conservation of vehicles. It is the fundamental equation for the speed of
propagation of “kinematic” waves in traffic according to Lighthill and
Whitham.
At this point, a few observations are in order regarding the underlying

assumptions in the derivation of Equation 9, and corresponding limitations in
its applicability. First of all, it is clear that a vehicle does not change its
speed abruptly from to as shown in Figure 5, but rather there is a
gradual transition from one speed to another. As long as this transition is
very similar among different vehicles, the derivation of Equation 9 is
reasonably valid in describing something akin to a shock generated by the
transition from one steady-state flow situation to another. However, one
should be careful not to use Equation 9 in describing a succession of rapid
transitions from one state of flow to another, when the vehicles do not have a
chance to relax to a steady-state flow before entering another transition.
What complicates matters in real life is the fact that the Lighthill-Whitham
theory inherently describes traffic in an infinitely long, flat, and straight
roadway. The existence of geometric variations in the roadway influences
the behaviour of traffic in all respects, including the propagation of
kinematic waves. However, Equation 9 does provide a reasonably good
description of the shock wave generated when traffic meets a bottleneck, as
indicated in Figures 2 and 5, where the flow downstream cannot
accommodate the flow upstream. Moreover, it may be observed that the
derivation of Equation 9 does not require the existence of a unique flow
versus concentration relationship. It can be derived on geometric grounds

from Figure 5, and only presupposes the existence of two possible steady-
state flow conditions, and
1.1.2 Flow versus concentration relationships
The preceding discussion applies to any flow versus concentration

relationship, and the question naturally arises which such relationship is the
appropriate one. Several such relationships have been proposed over the
years, and some have been tested against experimental evidence.
Richards (1956) proposed a hydrodynamic model of traffic essentially

identical to that of Lighthill and Whitham, independently of them. He also
proceeded to explore the properties of his model assuming a specific
relationship between speed and density, namely,
which leads to a parabola for the q versus k curve,
Richards used this flow-concentration relationship explicitly in all his

discussions of traffic phenomena, thus limiting the applicability of his
results.
The first model for the q-k curve derived entirely on the basis of a
hydrodynamic model was that proposed by Greenberg (1959), who assumed
that traffic could be approximated by a one-dimensional fluid satisfying the
continuity equation. He assumed an equation of motion for the fluid, which
10 Chapter 1
states that the acceleration of the traffic stream is inversely proportional to

the concentration k and directly proportional to the concentration gradient
with a constant of proportionality of If the speed is given as a
function of location and time, i.e. u = u (x , t ), the equation of motion
becomes
Now, the continuity equation, (Equation 4), together with the relationship
q = uk, yield
If it is now assumed that the speed of the traffic stream is only a function
of the concentration, i.e. u = u ( k ) , then
Substituting Equations 15 into Equations 13 and 14, and solving for

du/dk, we obtain
Integrating, and using the end condition u = 0 when we obtain

The maximum q is given by Equation 16 at when the

corresponding speed at maximum flow, is equal to a, yielding finally
A plot of q versus k according to Equation 17 is shown in Figure 6,

together with field measurements taken at the Lincoln Tunnel of New York
City, as discussed by Edie (1974). There is absence of data for very low and
very high densities, and a scatter of data, particularly past the point of
maximum flow, but overall the fit is reasonably satisfactory.
12 Chapter 1
There are many possible causes for the scatter of data at densities
corresponding to relatively dense traffic. Some of these causes are discussed
in what follows.
1.1.3 Additional considerations of the macroscopic traffic flow

theory
As mentioned above, the representation of the properties of a roadway by

a single flow versus concentration relationship applies, at best, on an
infinitely long, flat, straight roadway. The existence of entrance and exit
ramps, curves, and slopes introduces variations, which may be handled by
assuming concatenation of roadway sections with different flow versus
concentration relationships. This will generate the appropriate transitions
from one q,k state to another, including the emergence of bottlenecks.
In addition to geometry, some special considerations of driver behaviour

may also dictate a departure from a single q versus k relationship. Newell
(1965) was the first one to consider the effect of different driver behaviour
during acceleration and deceleration. It turns out that all drivers generally
allow greater gaps between cars during acceleration than they do during
deceleration. This is not surprising, because drivers may drive close to the
minimum distance for safe driving during deceleration, but they may relax
and let gaps increase for a while during acceleration. This introduces,
inherently, lower values for flow during acceleration than during
deceleration, for the same value of density. Newell assumed that the
properties of the roadway during acceleration were described by a different q
versus k relationship, and defined this effect as an “acceleration bottleneck”.
The net effect of this asymmetry in acceleration and deceleration behaviour
is that if a car, and its followers, undergo a brief deceleration manoeuvre and
then go back to their original speed, the speed recovery will be associated
with a decrease in density and flow. This will propagate indefinitely
upstream, causing effects similar to those of passing through a bottleneck of
decreased capacity, hence the name of “acceleration bottleneck”. The
phenomenon has been described as a “hysteresis effect”, altering the
transition from one q,k state to another, as shown in Figure 7. In this figure,
it is assumed that a car and its followers undergo a deceleration manoeuvre
from point 1 to point 2. The hysteresis effect is described by the dashed
curve, emanating from point 2. As the cars try to attain their original speed,
they end up not back to point 1 but to point 3, generating a shock wave that
affects all the cars behind them.
There is considerable experimental evidence supporting Newell’s model

of the asymmetry between acceleration and deceleration. Forbes et al (1958)
have reported slower response of drivers to acceleration than to deceleration.
Herman and Potts (1961) reported measurements indicating differences in
behaviour during acceleration ad deceleration. Daou (1964) found
significantly larger spacing between cars accelerating at a given speed than
between cars decelerating at the same speed. In addition, there have been
observations through aerial photographs showing the existence of
accelerating shock waves.
The deceleration may be due to the presence of a curve along the

roadway necessitating a temporary decrease in speed, followed by
acceleration to the original speed past the curve. This shows how geometry,
coupled with driver behaviour, may alter the behaviour of traffic beyond the
predictions of the Lighthill-Whitham treatment. An even more serious effect
of geometry is that caused by the existence of grades along the roadway. In
particular, an uphill portion may cause a large disturbance in traffic,
particularly if a heavy vehicle such as a truck must negotiate this uphill
portion together with the rest of the traffic because no “slow traffic” lane is
provided along the uphill portion of the roadway. If the truck slows down
sufficiently, it may block not only its own lane but also the neighbouring
lane, since cars trapped behind the truck try to escape into this neighbouring
lane, thus impeding the movement of cars along that lane as well. In a two-
lane roadway, the net effect is akin to that caused when two lanes merge into
14 Chapter 1
one, the only difference being that the merging point moves with the speed
of the slow truck. This produces the effect of a “moving bottleneck”
described by Gazis and Herman (1992). The moving bottleneck may be
transformed into a “phantom bottleneck” if the truck reaches a level or
downhill portion and resumes its normal speed. In that case, drivers trapped
behind the moving bottleneck see no bottleneck when they move past the
congested portion caused by the moving bottleneck and see no apparent
cause for this congestion.
A brief discussion of the study of the moving and phantom bottleneck

will be given here, to illustrate the use of kinematic models for the study of
complex traffic phenomena.
As mentioned already, it is assumed that the moving bottleneck is caused

by a vehicle moving at a speed substantially slower than the average
speed of the traffic stream, We assume that it moves along the right lane
of a two-lane stream of traffic. Because of its reduced speed, cars queue up
behind it, driving at the same slow speed and waiting for an opportunity to
pass. By manoeuvring for passing, they slow down the neighbouring lane as
well, roughly to the same slow speed, resulting in a two-lane queue, which is
discharged at the front of the left (unblocked) lane. The situation resembles
the merging of two lanes into one, only the point of convergence moves at
the speed of the slow vehicle, the moving bottleneck.
When a vehicle finds itself in the unblocked lane abreast of the slow
vehicle, it escapes at a speed which is generally different from both
and Let us assume that a vehicle escaped at time t= 0 (Figure 8). After a
time it will travel a distance while the slow vehicle travels a
distance If the spacing between escaping vehicles is , then the flow
past the slow vehicle is
According to our assumptions, the slow vehicle leads a pack alongside

and behind it moving at speed The front of this pack will travel a
distance during the time interval It follows that the slow vehicle
leads a two lane pack which supplies an additional flow rate, equal to
twice that corresponding to speed and concentration namely,
The total flow along the two-lane highway is
Equation 20 has all the desired properties in the limits and

and is based on reasonable assumptions regarding the effect of the slow
vehicle on other vehicles during periods of moderately heavy traffic. Next,
we can estimate the speed of the trapped vehicles behind ad alongside the
slow vehicle. This is not equal to the speed of the slow vehicle, because a
portion of the two queues escapes. Assuming that the percentage of escaping
vehicles is equally divided among the two lanes, we use the definition v=q/k
and the value of flow given by Equation 20, with concentration in each of
the two lanes, obtaining the speed of the trapped vehicles
16 Chapter 1
We can now determine whether or not the queue size increases or

decreases by using the Lighthill-Whitham estimation of the velocity of a
shock wave. The queue size remains constant if the speed of the front of the
queue with respect to a stationary observer is equal to that of the rear of the
queue. The speed of the front of the queue is obviously equal to that of the
slow vehicle, namely, The speed of the end of the queue is
Setting and using the value of Q given by Equation 20, we

obtain
The value of given by Equation 23 is a critical value of the speed of

the slow vehicle. Above this speed, the slow vehicle has virtually no effect
on the overall quality of traffic, because the queue length that might get
somehow started will keep decreasing until it is dissipated. Below that
critical speed, the queue will tend to increase indefinitely. It may also be
pointed out that the preceding treatment does not provide a means for
determining the stationary size of a queue, since any size will remain
constant if satisfies Equation 23. It is very possible that, if a moving
bottleneck is suddenly introduced into a stream with a speed satisfying
Equation 23, a non-zero queue size will be developed, which depends on the
exact movement of cars around the slow vehicle, after which the queue size
will remain more or less constant.
1.2 THE BOLTZMANN-LIKE MODEL OF TRAFFIC
Around 1959, Ilya Prigogine, who was later to receive a Nobel Prize for
his work on Statistical Mechanics, proposed a treatment of traffic on a
multilane highway analogous to the Boltzmann approach for deriving the
properties of gases from consideration of the movement of individual atoms
of gas. Prigogine (1961), and Anderson et al (1962) developed a theory of
traffic flow based on a description of traffic in terms of a probability density
for the speed, v , of an individual car. This density, f ( x, v, t), may vary as
a function of time, t, and a co-ordinate x along the highway. The basic
equation for the distribution is assumed to be
It is assumed that every vehicle enters the highway with a “desired

speed”, and drives at that speed as long as it is not impeded by a slower car.
If it does get impeded, it slows down until it gets a chance to pass, after
which it relaxes back to its desired speed. The first term of the right-hand
side of Eq. 24 is a consequence of the fact that f ( x, v, t) differs from some
“desired” speed distribution and drivers try to “relax” to their desired
speed as soon as they have an opportunity. The second term of the right-
hand side corresponds to the slowing down of a fast vehicle caused by a low
one. True to his tradition as a leader in statistical mechanics of fluid media,
Prigogine frequently referred to this second term as a “collision” term, a
rather unsettling choice of words when speaking about traffic.
The form of these two terms was chosen for mathematical convenience
and plausibility. In the end, Eq. 24 was assumed to have the specific form
where is the characteristic relaxation time, p is the probability that a car

passes another car, and is the average speed of the stream of traffic. The
second term of the right-hand element of Eq. 25 corresponds to the
18 Chapter 1
interaction term of Eq. 24, and tends to approach zero when the
concentration of cars is light and the probability of passing is close to unity.
In that case, the relaxation term is dominant. If. In addition, we assume a
highway with a homogeneous distribution of cars, then and the
solution of Eq. 25 is
If we seek a solution of Eq. 25 that is independent of time and space, then

the left-hand element of this equation is zero. If, in addition, we consider
small values of concentration, the solution of Eq. 25 shows an approximately
linear increase of flow with concentration, namely,
where is the average of the desired speed. As k increases, the flow q

falls below the straight line of Eq. 27 due to the increasing influence of
interactions between cars, and the resulting slowing down of the faster cars
by the slower ones. Moving to the range of high concentrations, we find11
that q is independent of and depends only on and p , according to
the relationship
Equation 28 describes what may be described as the “collective flow” of

traffic, which is characterised by high densities and very little passing.
The overall behaviour of traffic is described by the curve on the Figure 9

below.
Starting at zero concentration, the flow increases linearly with

concentration. The increase slows down as the concentration increases, but
the flow remains essentially “individual”, meaning that individual vehicles
have a chance to drive at their desired speed for a great percentage of the
time. At some point, after flow reaches a maximum, the interactions become
more and more intensive, until the flow becomes “collective”, meaning that
individual vehicles are forced to follow other vehicles most of the time. Such
behavior clearly matches personal experiences of all drivers, making the
Boltzmann-like model quite successful in representing real-life phenomena.
1.2.1 The two-fluid model of traffic
Herman and Prigogine (1979) extended the above discussion, proposing

a two-fluid model of traffic flow, which is applicable in the domain of
collective flow. According to this model, a portion of the cars are moving,
and another portion are “immobilised”. At the transition to collective flow,
the speed distribution function for cars splits into two parts, one
corresponding to the moving vehicles and the other to vehicles stopped as a
consequence of congestion, traffic signals, stop signs and barricades, but not
in the parked condition.
Extending the ideas in the Boltzmann-like model, they assumed that the
average speed of the moving cars, depended on the fraction of cars that
were moving, according to a power law, namely,
20 Chapter 1
where is the average maximum speed and n is a parameter to be

determined from observations. In addition, they stipulated that the fraction of
cars stopped, could be equated to the fraction of the time a car circulating
in the street network would be stopped, namely,
where is the stopped time per unit distance, T is the trip time per unit
distance, and Equations 29 and 30 lead to a simple
transcendental equation relating the stop time to the trip time, both per unit
distance, namely,
where
The two-fluid model has been scrutinised extensively over the years, and
found to represent remarkably well the behaviour of traffic during congested
periods in various cities. In general, several hours of data were sufficient to
determine the two-fluid parameters, namely and for a typical urban
street network of the central business district of a city. Thus, the two-fluid
model provided a means for a quantitative assessment of the quality of
traffic in a large number of cities in the United States and abroad. A further
investigation and validation of this model was provided by simulation
studies of Mahmassani et al (1987). The simulation provided an important
supporting tool for further development and extension of the theory,
especially the dependence of the proposed relations among network-wide
variables on the topological and operational features of a network.
In concluding this section, we may call attention to the following feature

of the two-fluid model. In a situation fraught with complexity, involving a
multitude of human operators in charge of a variety of vehicles, it is

remarkable that the behaviour of traffic can be described by a relatively
simple set of equations such as Eqs. 29 to 32. This is because the constraints
imposed by traffic congestion limit the number of choices for each
individual driver, and force the system into a collective behaviour that is
more regimented than one might expect. Obtaining a relatively simple
representation of traffic gives, in turn, hope that improvements in traffic
management may be implemented by influencing the operational
characteristics of a network, once the dependence of performance on these
operational characteristics is established through observations and
simulation.
1.3 MICROSCOPIC TRAFFIC FLOW THEORY – CAR-

FOLLOWING MODELS
The first microscopic models of traffic were proposed, independently, by

Reuschel (1950) and by Pipes (1953). They both used models inspired by
rules of driving analogous to the well known “California rule”, namely,
allowing a certain gap between a car and it leader proportional to its speed.
Reuschel and Pipes assumed that drivers achieved such goals essentially
instantaneously, and they derived the consequences of such behaviour. A
more discerning view of driver behaviour was the foundation of the car-
following studies carried out by Herman and his collaborators at the General
Motors research Labs, starting in 1957. They considered the devices under
the control of the driver, which are the brake and the gas pedal. Both control
the acceleration of the vehicle, and it was this variable that was the one
viewed as controlled by driver action. Second, they took into account the
fact that a human being has a time lag in reacting to any inputs. As for the
dominant inputs stimulating driver reaction, a combination of rational
thinking, mathematical analysis, and experimental observation led to the
relative speed between leading and following car as the leading stimulus.
The result was the basic car following equation
where n denotes the number corresponding to the lead car, n+1 that of the
following car, T is the reaction time lag, and is a coefficient of
proportionality of the driver reaction to the stimulus, variously referred to as
22 Chapter 1
the sensitivity coefficient or the gain factor. In the first car following studies,
this sensitivity coefficient was assumed constant. Herman and his colleagues
went on to test their model by conducting car following experiments, and
investigating the consequences of this model on the stability of flow of a
stream of traffic.
Stability refers to the ability of a system of cars to absorb a perturbation

introduced into the stream of traffic. “Local” stability refers to the relative
position of a fixed pair of cars. “Asymptotic” stability refers to the
attenuation or increase of the amplitude of perturbation as it moves down a
line of cars upstream of the perturbation.
1.3.1 Local stability.
Local stability was investigated by solving Eq. 33 by means of Laplace

transform techniques (Gazis et al, 1963). Imagine that the lead car undergoes
a manoeuvre during and then maintains a constant speed for t > T .
The following vehicle moves according to Eq. 33. If
the following equation is obtained for t > 0,
In order that this be a well posed problem, the value of y must be

assumed to be given in the interval To make this point explicit,
y (t) may be written as a sum of two functions, one of which is known and
the other one unknown, namely,
in which represents the initial data, given for and (t)

represents the unknown part of y, for t > T. The functions and satisfy
the conditions
Entering these values of and (t) in Eq. 35, we obtain
The Laplace transforms are defined as
and
The transform of Eq. 38 is

24 Chapter 1
from which we obtain
The denominator of Eq. 42 is now expanded into partial fractions

according to the relationship
in which is the denumerably infinite set of roots of the equation
These roots are assumed to be arranged in ascending order according to

the value of their real part, namely,
The inverse transform of is now found by using the convolution

theorem for the first term in Eq. 42, and inverting term by term, and by
inverting the second term using the result of Eq. 43. The result is
in which
It is seen that the long time behaviour of z(t), and hence y(t), is
approximately that of the leading term of the series in Eq. 45, namely,
in which is the root of Eq. 44 with the largest real part. The nature of this
root, depends on the value of the product
Thus, it is found that three different types of limiting behaviour of y(t)

are possible, depending on the magnitude of as follows:
If then is real and negative, and therefore the initial

fluctuation is damped out monotonically, for sufficiently large values of t.
If the quantity is complex with a negative real part.
Hence, the initial fluctuation is damped out as the product of a negative
exponent and a trigonometric function of time. When the
fluctuation remains constant in amplitude and oscillates sinusoidally in time.
If the quantity has a positive real part. Hence, the initial
fluctuation produces oscillations of increasing amplitude, and equilibrium is
never reached.
The preceding discussion is illustrated in Figs. 10 and 11 which show the

results of numerical computations of the trajectories of the lead and the
following car using Eq. 33, for various values of C, and assuming a
fluctuation in the acceleration of the lead car in the form of a negative square
pulse followed by a positive pulse of the same magnitude. Figure 10 is a
graph of the detailed motion of the two cars, showing the effect of the
acceleration pattern of the lead car, for different values of C. Remembering
that we find that for C = 0.5 and 0.8 the spacing is oscillatory
and damped, for C=1.57 it is oscillatory and undamped, and for C=1.6 it is
oscillatory with increasing amplitude.
26 Chapter 1
1.3.2 Asymptotic stability.
Although a chain of vehicles may be locally stable, it may result in an

asymptotically unstable situation, when each vehicle amplifies the signal and
passes it on to the cars upstream. This possibility is investigated by
considering a Fourier component of the fluctuation of the lead vehicle, the
in a chain. Let
Substituting into Eq. 33, we obtain
hence
and
The amplitude factor in Eq. 52 decreases with increasing values of n if

It may be seen from Eq. 53 that low frequencies cause the greatest
limitations on the sensitivity and time lag T. For in order to have
decreasing amplitude of the perturbation upstream, and T must satisfy the
relationship
Equation 54 gives the condition for asymptotic stability. Numerical

examples are given in Figs. 12 and 13.
28 Chapter 1
Equation 54 gives the condition for asymptotic stability. Numerical

examples are given in Figs. 12 and 13.
In Fig. 12, the spacing between successive pairs of cars is shown, after a
deceleration-acceleration manoeuvre of the lead car, and for three different
values of C.
In all three cases, we have local stability in the sense that the amplitude
of the perturbation in the spacing between cars eventually decreases toward
zero with increasing time. However, in the third case, C = 0.75, the
amplitude of the perturbation increases between successive pairs of cars,
causing asymptotic instability. This is shown more dramatically in Fig. 13,
where the position of nine cars is shown with respect to a system of co-
ordinates that moves with the original speed of the lead car. The original
spacing of the cars is 40 ft, and it is assumed that C =0.8 and T= 2 sec. It is
seen that the spacing between successive pairs of cars suffers a perturbation
of increasing amplitude as we move upstream, until there is a collision
between the and cars.
30 Chapter 1
It may be mentioned here that the original experiments carried out at

General Motors by Herman and his colleagues produced, for eight test
drivers, an average value of C equal to 0.56, indicating that people drive on
the verge of asymptotic instability. However, there are some additional
stabilising factors in real life driving. For example, people are generally able
to see more than just the car in front of them, and adjust their driving
accordingly. It has been demonstrated analytically (Herman et al, 1959) that
taking into account information from the next-nearest neighbour enhances
the stability of a stream of traffic. On the other hand, when such next-
nearest neighbour information is lacking, for example due to fog or extreme
congestion conditions, such stabilising influence may be lacking, which may
explain the higher incidence of rear end collisions in such circumstances.
1.3.3 Acceleration noise.
It is reasonable to assume that an average driver would like to maintain a

constant speed during a trip, avoiding accelerations and decelerations.
However, deviations from a constant speed are generated by a number of
factors including the geometry of the roadway, the influence of neighbouring
cars, and the driver’s own propensity toward aggressive rather than passive
driving behaviour. In general, high volumes of traffic tend to push every
driver toward car-following patterns of driving, generating greater
fluctuations from constant speed than those present at low traffic volumes.
A measure of the velocity fluctuations during a trip is given by the

standard deviation of the acceleration about a mean acceleration, and has
been defined as the “acceleration noise”. Assuming the mean acceleration
during a trip to be zero, the acceleration noise is defined as
where a(t) is the acceleration (positive or negative) at time t, and T is the

total time of the trip. A difference approximation to Eq. 55, in which
acceleration is assumed to be sampled at time intervals is given by
A “smooth” trip will be associated with small and a “rough” trip

with a large one.
The concept of acceleration noise was introduced in conjunction with

car-following studies (Montroll, 1961), and was further used to analyse the
quality of driving under various conditions. For example, Jones and Potts
(1962) used the concept in order to assess the safety and quality of driving
along different types of highways (four-lane versus two-lane), the
differences among drivers due to temperament and age, and the influence of
the presence of attractions along the highway on the overall quality of
driving. They reported the following conclusions:
32 Chapter 1
For two roads through hilly country, is much greater for a narrow
two-lane road than for a four-lane divided highway.
For a road in hilly country, is greater for a downhill journey than for
an uphill one.
For two drivers driving at different speeds below the design speed of a
highway, is much the same.
If one or both the drivers exceed the design speed, is greater for the
faster driver.
Increasing traffic volume increases
Increasing traffic congestion generated by parking cars, stopping buses,
cross traffic, crossing pedestrians, etc., increases
The value of may be a better measure of traffic congestion than travel
times or stopped times.
High values of on a highway indicate a potentially dangerous
situation.
Jones and Potts cautioned against the use of arbitrary interpretation of the
values of but they made the observation that is a low
value, and is a high value.
Helly and Baker (1967) proposed another parameter, related to the

acceleration noise, in order to describe the quality of driving, namely
where G is the velocity gradient, the acceleration noise, and u the mean
velocity during a trip. The reasoning behind their proposal of G as the
measure of goodness was that acceleration noise alone does not describe the
quality of the trip, because a fast trip and a slow trip could both have the
same value of but the fast trip would be more desirable. The time T
used in their calculation of was the total trip time, including stopped time.
Underwood (1968) reported results of observations in Australia which

suggest that G may be a better measure of traffic congestion than but
contradicts item 7 of the conclusions of Jones and Potts by questioning
whether the velocity gradient is any better than total travel time as a measure
of congestion.
As for the influence of car-following on acceleration noise, good

measurements were reported by Herman and Rothery (1964) from actual
experiments, shown in Fig. 14.
It is seen from this figure that the acceleration noise increases rapidly
with the position of the car up to about the position, after which it reaches
a plateau. This result was obtained for an isolated platoon. In actual traffic,
the movements of cars around the platoon will probably accentuate the
disturbances to the platoon and the resulting acceleration noise.
1.3.4 Non-linear Car-Following Models, and Tie-in to

Macroscopic Theory
The car-following model of traffic found a harmonious tie-in to

macroscopic theory in a paper by Gazis, Herman, and Potts23. The bringing
together of the two approaches proceeds as follows:
Let us assume that traffic consists of a line of cars following a leader

driving at constant speed We then assume that the leader undergoes a
manoeuvre, changing the car speed to a new speed Assuming that local
and asymptotic stability are satisfied, we can integrate the basic car-
following equation, Eq. 33, neglecting the time-lag T, since all cars will
reach a new steady state after a sufficiently large time interval. In fact, we
can re-write Eq. 33 in the form
34 Chapter 1
where u is the speed of the average vehicle and s is the spacing between
vehicles. Equation 58 states that the acceleration of the average vehicle,
du/dt, is proportional to the average relative speed, ds/dt. Integrating Eq. 58
we obtain
where is the integration constant. Its value can then be determined from
the end condition that the speed be zero at the jam density hence
Hence
and the flow versus concentration curve comes out as a straight line with a
slope The intercept of this line with the q axis is the maximum
value of flow, when This result does not agree with most
observations, which indicate a flow versus concentration relationship
represented by a convex curve, such as Greenberg’s model shown in Fig. 6.
But Gazis et al (1959) showed that the Greenberg steady-state model could
be derived from car-following theory by a nonlinear model in which the
sensitivity, or gain factor, was not a constant but was inversely proportional
to the spacing between cars. In this case, the car-following equation becomes
which, in the simplified form of Eq. 58, becomes
Integrating Eq. 59, we obtain
The speed must be zero at jam density, hence and Eq. 60

becomes
which is none other than Greenberg’s flow versus concentration

relationship. However, this model also suffers from the inappropriate limit
for the velocity at zero concentration, since as For this
reason, Edie (1961) proposed a somewhat more complicated model which
appeared to fit data for low concentrations better than the inverse spacing
model. Edie’s model, omitting time lags, is
which, when integrated, yields

36 Chapter 1
where is the “free-flow speed” at zero concentration. However, the

disadvantage of the Edie model is that the concentration tends to infinity as
the speed tends to zero.
All the above mentioned car-following models are included in a general

non-linear model proposed by Gazis, Herman, and Rothery (1961), in which
the sensitivity is a generalised function of the speed of the following car and
the spacing between lead and following cars. The general form of this model
is
which, with appropriate choices for l and m , gives any of the above
models, and more.
For data taken by car-following runs in the Lincoln tunnel of New York
City, Gazis et al found the best correlation for values of m= 1 and l= 2. It is
interesting to note that May and Keller (1969) investigated the possible
choice of fractional exponents for freeway and tunnel data, and found that
those giving the best correlations were m= 0.8, l= 2.8, for freeway data, and
m= 0.6, l= 2.1, for tunnel data.
1.4 SOME QUESTIONABLE “PARADOXES”
As in many other fields of science, there have been some misguided

contributions, some of which persisted in their popularity for a considerable
time. One is reminded of Mark Twain’s saying: “There is something
fascinating about science. One gets such wholesale returns of conjecture out
of such a trifling investment of facts”. We shall discuss here two “paradoxes”
based on traffic flow models, which are based on rather faulty reasoning,
although they appear to be mathematically correct.
1.4.1 The Smeed paradox
In 1967, Smeed (1967) published a paper in which he discussed a most

intriguing “paradox”. The paper contained a mathematical proof, based on
traffic flow models, that under certain circumstances a vehicle would reach
its destination earlier by starting later. The Smeed paradox was based on the
following argument. Suppose a car is standing at the entrance of a long
section of the roadway, considering making an entrance toward its
destination at the end of the roadway section.
The vehicle is supposed to be at the end of a platoon of n identical car,

which in effect make a collective decision to enter at some chosen headway
value. As the headway increases, the car has to leave later and later.
However, larger headways mean lower densities, and according to flow
theory higher speeds. Smeed went on to prove mathematically that under
certain circumstances choosing a larger headway, and therefore delaying the
departure, the car could arrive earlier than it would have with the unchanged
headway.
The problem with Smeed’s argument lies in the assumption that a vehicle
entering a roadway section in effect influences the steady-state speed of
vehicles in front. This amounts to an assumption that disturbances propagate
downstream. However, as every driver knows, traffic is a uni-directional
medium, in which disturbances can only propagate upstream. (Here, we can
ignore the minor effect of a “tailgater” influencing a car in front to drive
faster!) Another way of discerning the error in Smeed’s reasoning is the
following. Suppose the vehicle does not enter the roadway, but a “phantom”
vehicle does, which does not influence the roadway density but moves with
the steady-state speed of the traffic stream. After the vehicle enters, it will
never be able to catch up with the phantom vehicle and overtake it, leading
to the resounding conclusion that it must arrive later if it starts later. The
conclusion is that Smeed, an outstanding leader in traffic science, slipped
into an erroneous line of reasoning in this case.
The erroneous consideration of forward propagation of disturbances is

not limited to the Smeed paradox. In recent years, traffic scientists with an
electrical Engineering background started talking about roadways as having
impedance properties. The idea is that just as wires have their resistance
increased with increasing electric current flow, highways have their
impedance increased, slowing down the traffic movement, when the volumes
of traffic are increased, or when the density of traffic is increased. This is all
fine, except that cars do not transmit information from one to another as fast
as electrons, and the signals they transmit are unidirectional, going upstream.
This means that the concept of impedance of roadways can only be applied
for steady state conditions of traffic movement. Applying it to dynamic
38 Chapter 1
traffic phenomena runs the danger of fallacy similar to that of the Smeed
paradox.
1.4.2 The Braess paradox
The Smeed paradox never gained wide acceptance, because several

traffic theorists, including the author of this book, saw its fallacy and
communicated it widely. However, there is another paradox, the Braess
paradox (1968), which has received wide publicity and continues to be
accepted as the absolute truth. The Braess paradox states that building an
additional link in a network may not only fail to improve the performance of
the network, but it may decrease it. The general idea of the paradox
describes the performance of a network such as shown in Fig. 15. It is
assumed that cars enter at point A and exit at point B. One starts with a
network that contains only the links AB, AC, BD, and CD, and asks whether
or not it makes sense to build the link BC. Braess assumed some
“performance” value for each of the four existing links as a function of the
flow along the link, which corresponded to the delay to the traffic stream
in the link. He computed the delay for the entire network by assuming that
the stream splits among the two routes, abz and acz. Assuming congested
conditions, any increase of the flow along any link would substantially
increase the delay to drivers. Braess went on to compute the performance of
the network after addition of the link bc, which would attract some of the
traffic along abz. He found that under certain circumstances the penalty to
the cars would increase in the new network, because even an improvement
along bz would be far exceeded by the worsening along cz.
The problem with the Braess paradox is that one can question the specific
delay functions that he assumed. In the example he gave corresponding to
Fig. 15, these delay functions, if assumed as multiples of the inverse of
speed, would correspond to extraordinary, and very debatable, flow versus
concentration relationships. For example, he assumed for link 1 the delay
function 10q, where q is the volume. If we set the delay as where is
the link length, then we obtain the corresponding q versus k (density)
relationship a rather unlikely relationship for which q tends
to infinity as k goes to infinity. The distribution of traffic after the addition of
the link is obtained from an optimisation procedure, which depends on these
unrealistic delay functions. For an uncongested system, anyone familiar with
network flow theory recognises that the “minimal cut” of the network cannot
possibly get worse after addition of a link. For a congested system, the point
z could only be viewed as an exit point, which will probably be the
bottleneck, and its performance is not in any way affected by the addition of
a link. The most that can be said in this case is that the addition of the link
does not improve the performance of the system, but it does not make it
worse either. To be sure, the addition of an attractive, new link may draw
some drivers into a route choice which degrades the performance of the
system, even for realistic choices of delay functions. This does not mean that
the system has worsened, and it may very well be that trial and error, plus
some sensible instructions to drivers, can bring drivers into a better set of
route choices. In any case, the conflict between individual route choices in a
network and global optimisation objectives has been well known for many
decades, and there is nothing paradoxical about it, as discussed in Chapter 4
of this book.
There is even an element of forward propagation of traffic noise in the

Braess discussion. A diversion of traffic along the link 5 (Fig. 15) is
assumed to change the behaviour of the forward link 3 only. In real life, it is
likely to affect even more the link 2, particularly during periods of
congestion, when the node C may become a “Throughput Limiting Point
(TLP), as discussed in Chapter 3 of this book.
So, the Braess paradox is an attractive statement, which has even

generated articles in the popular press, but it not a very helpful design tool
for transportation planners. In recent years, many people have cited the
Braess paradox as correct, in the sense that construction of new roadways, or
improvement of existing ones, attracts new traffic, which results in worse
traffic conditions than before construction or improvement. This “traffic
40 Chapter 1
generation” effect may be true, but it has nothing to do with the original
statement of the Braess paradox.
1.5 ADDITIONAL FLOW THEORY WORK, AND FUTURE

CHALLENGES
Several models of traffic systems and traffic flow have been proposed in
recent years, including those based on “chaos theory”, “cellular automata”,
and “hopping models”. Some of them had the advantage of providing a
convenient simulation tool, and in this way may have been helpful to
researchers, since calibration of these models showed them to be reasonably
good in describing actual traffic conditions. However, many of these models
fail to describe the actual behaviour of traffic, and the partial fit with some
measurement does not justify their promotion as improved modelling tools.
For example, the hopping models essentially convert automobile traffic into
what might be called a “moving block” system, using the nomenclature of
railway traffic, and it is very unlikely that this is how people drive. The
chaos theories, while very attractive to journalists who love the concept of
chaotic behaviour of traffic, do not contribute any understanding of this
behaviour. Many of the seemingly chaotic features of traffic can be
explained as the effects of accidents, the geometry of the roadways, or such
phenomena as the “moving and phantom bottleneck” discussed earlier in this
chapter.
The above comments do not mean to imply that the modelling of traffic
flow is finished. There are many challenges ahead, and meeting these
challenges holds promise of additional tools for improved traffic
management. One of the needed key improvements of traffic flow theory
concerns the effect of roadway geometry.
Virtually all the traffic flow models discussed in this chapter are good
mainly for an infinitely long, flat, and straight roadway. However, every
good driver knows that the geometry of the roadway affects driving patterns.
For example, a curve may necessitate some slowing down of traffic before
the curve with some acceleration along the curve and past the curve. This
effect would alter the nature of either the macroscopic or the microscopic,
car-following models of traffic. An additional effect of geometry is that due
to inclined roadway sections. An example of the effect of such sections was
given in this chapter, in the discussion of the “moving and phantom
bottleneck”, which may result from the slowing down of a truck by an uphill
portion of the roadway. It should also be mentioned that the effect of uphill
portions has been considered in the case of tunnel traffic, discussed in

Chapter 3 of this book. In general, work is needed in augmenting traffic flow
models to include the effects of geometry. Such augmentation will be
particularly needed if automation ever becomes a standard replacement of
drivers, or if railroad train control progresses to “moving block”, automated
train control.
42 Chapter 1
References
Anderson, R. L., Herman, R. and Prigogine, I. (1962), “On the statistical distribution
function theory of traffic flow”, J. Opns. Res., 10, 180-195.
Braess, D. (1968), “Uber ein Paradoxen der Verkehrsplannung”, Unternehmenstorchung,
12, 258-268.
Chandler, R. E., Herman, R. and Montroll, E. W. (1958), “Traffic Dynamics: Studies in
Car Following”, Operations Res., 6, 165-184.
Daou, A. (1964), “On Flow in Platoons”, Bull. Operations Res. Oc. Amer., 25ty National
Meeting, 12, B40.
Edie, L. C. (1961), “Car-Following and Steady-State Theory for Non-Congested Traffic”,
Operations Res., 9, (1), 66-76.
Edie, L. C. (1974), “Flow Theories”, in Traffic Science, edited by D. C. Gazis, Wiley-
Interscience, New York, pp. 1-108.
Forbes, T. W., Zagorski, M. J., Holshouser, E. L. and Deterline, W. A. (1958),
“Measurement of Driver Reactions to Tunnel Conditions”, Proc. Highway Res. Board, 37,
345-357.
Gazis, D. C., Herman, R. and Potts, R. B. (1959), “Car-Following Theory of Steady-State
Traffic Flow”, Operations Res., 7, 499-505.
Gazis, D. C., Herman, R. and Rothery, R. W. (1961), “Nonlinear Follow-the-Leader
Models of Traffic Flow”, Operations Res., 9, 545-567.
Gazis, D. C., Herman, R. and Rothery, R. W. (1963), “Analytical Methods in
Transportation: Mathematical Car-Following Theory of Traffic Flow”, Proc. Am. Soc. Civil
Engrs, Eng. Mech. Div., 6, 29-46.
Gazis, D. C. and Herman, R. (1992), “The Moving and ‘Phantom’ Bottlenecks”, Transp.
Science, 26, No.3, 223-229.
Greenberg, H. (1959), “An Analysis of Traffic Flow”, Operations Res., 7, No.l, 79-85.
Helly, W. and Baker, P. G. (1967), “Acceleration noise in a congested signalized
environment”, in Vehicular Traffic Science, Proc. Third Intern. Symp. on Th. Of Traffic Flow,
American Elsevier, New York, pp. 56-61.
Herman, R., Montroll, E. W., Potts, R. B. and Rothery, R. W. (1959), “Traffic Dynamics:
Analysis of Stability in Car Following”, Operations Res., 7, 86-106.
Herman, R. and Potts, R. B. (1961), “Single-Lane Traffic Theory and Experiment”, in
Theory of Traffic Flow, edited by R. Herman, Amsterdam:Elsevier, 120-146.
Herman, R. and Rothery, R. W. (1962), “Microscopic and Macroscopic Aspects of Single
Lane Traffic Flow”, Operations Res. Japan, 5, 74.
Herman, R. and Prigogine, I. (1979), “A Two-Fluid Approach to Town Traffic”, Science,
204, 148-151.
Jones, T. R. and Potts, R. B. (1962), “The measurement of acceleration noise – a traffic
parameter”, Oper. Res. 10 (6), 745-763.
Lighthill, M. J. and Whitham, G. B. (1955), “On Kinematic Waves, II. A Theory of
Traffic Flow on Long Crowded Roads”, Proc. Roy. Soc. (London) 229A, 317-345.
Mahmassani, H. S., Williams, J. C. and Herman, R. (1987), “Performance of Urban

Traffic Networks”, in Transportation and Traffic Theory, edited by N. H. Gartner and N. H.
M. Wilson, Elsevier Science Publishing Company, Amsterdam, pp. 1-20,
May, A. D. and Keller, H. E. M. (1969), “Evaluation of Single and Multi-Regime Traffic
Flow Models”, in Beitrage zur Theorie des Verkehrs Flusses, Bonn: Heransgegeben vom
Bundesminister fur Verkehr, pp. 37-47.
Montroll, E. W. (1961), “Acceleration noise and clustering tendency of vehicular traffic”,
in Proc. Symp. on Th. Of Traffic Flow, Elsevier, New York, pp. 147-157.
Newell, G. F. (1965), “Instability in Dense Highway Traffic”, in Proc. 2nd Intrn. Symp. on
the Theory of Road Traffic Flow, edited by J. Almond, Paris: Office of Economic Cooperation
and Development, 73-83.
Pipes, L. A. (1953), “An Operational Analysis of Traffic Dynamics”, J. Appl. Phys., 24,
274-281.
Prigogine, I. (1961), “A Boltzmann-like approach to the statistical theory of traffic flow”,
in Theory of Traffic Flow, R. Herman, Editor, Elsevier Publishing Co., Amsterdam.
Reuschel, R. (1950), “Fahrzeughewegungen in der Kolonne bei gleichformig
beschleunigtem oder verzogertem Leitfarzeug”, Zeit. Osterr. Ing. und Arch. Ver., 95, 52-62,
73-77.
Richards, P. I. (1956), “Shock Waves on a Highway”, Operations Res., 4, 42-51.
Smeed, R. J. (1967), “Some Circumstances in Which Vehicles Will Reach Their
Destination Earlier by Starting Later”, Transp. Sci., 1, 308-317 .
Underwood, R. T. (1968), “Acceleration noise and traffic congestion”, Traffic Eng.
Control, 10 (3), 120-123, 130.
Additional References
Buckley, D. J. (1962), “Road Traffic Headway Distributions”, Proc. Aust. Road Res.
Board, 1, 153-187.
Daganzo, C. F. (1997), Fundamentals of Transportation and Traffic Operations, New
York: Elsevier.
Darroch, J. N. and Rothery, R. W. (1972), “Car Following and Spectral Analysis”, in
Traffic Flow and Transportation, edited by G. F. Newell, New York: Elsevier.
Dunne, M. C., Rothery, R. W. and Potts, R. B. (1968), “A Discrete Markov Model of
Vehicular Traffic”, Transp. Sci., 2, No. 3, 233-251.
Gazis, D. C., Herman, R. and Weiss, G. (1962), “Density Oscillations Between Lanes of a
Multi-Lane Highway”, Operations Res., 10, 658-667.
Greenberg, H. (1959), “An Analysis of Traffic Flow”, Operations Res., 7, No. 1, 79-85
(1959).
Greenshields, B. and Loutzenheiser, A. (1940), “Speed Distributions”, Proc. Highway
Res. Board, 20.
44 Chapter 1
Haight, F. A. (1963), Mathematical Theories of Traffic Flow, New York: Academic

Press.
Kometani, E. (1955), “On the Theoretical Solution of Highway Traffic Under Mixed
Traffic”, Jour. Faculty, Kyoto Univ., 17, 79-88.
Miller, A. J. (1961), “Analysis of Bunching in Rural Two-Lane Traffic”, J. Roy. Statist.
Soc., B-23, 64-75.
Montroll, E. W. and Potts, R. B. (1964), “Car-Following and Acceleration Noise”, An
Introduction to Traffic Flow Theory, Highway Research Board Special Report 79.
Newell, G. F. (1955), “Mathematical Models of Freely Flowing Highway Traffic”,
Operations Res., 3, 176-186.
Newell, G. F. (1962), “Theories of Instability in Dense Highway Traffic”, J. Oper. Res.
Soc. Japan, 5, 9-54.
Newell, G. F. (1965), “Instability in Dense Highway Traffic”, in Proc. 2nd Intern. Symp.
on the Theory of Road Traffic Flow, edited by J. Almond, Paris: Office of Economic
Cooperation and Development, pp. 73-83.
Newell, G. F. (1968), “Stochastic Properties of Peak Short-Time Traffic Counts”, Transp.
Sci., 2, No. 3, 167-183.
Oliver, R. M. (1952), “Distributon of Gaps and Blocks in a Traffic Stream”, Oper. Res.,
10, 197-217.
Prigogine, I., Herman, R. and Anderson, R. (1965), “Further Developments in the
Boltzmann-like Theory of Traffic”, in Proc. 2nd Intern. Symp. on the Theory of Road Traffic
Flow, edited by J. Almond, Paris: Office of Economic Cooperation and Development, pp.
129-138.
Prigogine, I. and Herman, R. (1971), Kinetic Theory of Vehicular Traffic, New York:
Elsevier.
Underwood, R. T. (1963), “Traffic Flow and Bunching”, J. Aust. Rd. Res., 1, 8-25.
Vaughn, R. and Hurdle, V. F. (1992), “A Theory of Traffic Flow for Congested
Conditions on Urban Arterial Streets. I: Theoretical Development”, Transp. Res., 26B, 381-
396.
Wardrop, J. G. (1952), “Some Theoretical Aspects of Road Traffic Research”, Proc. Inst.
Civil Engineers, Part II, 1, No. 2, 325-362.
Wright, C. C. (1972), “Some Properties of the Fundamental Relations of Traffic Flow”, in
Traffic Flow and Transportation, edited by G. F. Newell, New York: Elsevier, pp. 19-32.
Chapter 2
Queueing and Delays at Isolated Intersections
A driver is delayed along his journey by a combination of factors, which

includes the interactions with other cars and the effect of regulatory devices
such as stop signs and traffic signals. In order to improve transportation
systems, it is desirable to be able to estimate the effects of any regulatory
devices which are to be used. Such estimates can be achieved through the
use of the methodologies of queueing theory, which was originally
developed for the study of communication systems. To be sure, many of the
standard techniques of queueing theory, such as the assumption of specific
functional character of time intervals between events, are only
approximations to the values encountered in real systems. Nevertheless, the
study of traffic system behavior through the methods of queueing theory
gives us a handle for designing improvements in such systems.
In this chapter, we discuss the effect of isolated intersections on the delay

to drivers. In Chapter 3, we shall discuss the effect of coordination of a
system of traffic signals in reducing such delays. In this chapter, we discuss
the effect of stop or yield signs, as well as the effect of traffic signals. The
estimates are derived through the use of the theory of stochastic processes
and, in particular, on the methodologies of renewal processes and queueing
theory.
46 Chapter 2
2.1 TRAFFIC CHARACTERISTICS
Traffic moving along a highway can be described as a function of space,

and as a function of time. For the purpose of the present discussion, it is
particularly useful to consider the intervals of time between events, such as
the arrival times of individual vehicles at a particular point. Let the sequence
of arrival times, at a fixed point, of successive cars in an unimpeded stream
of traffic be The headways between cars, are defined as
and is defined to be equal to whether or not the time t = 0

corresponds to the arrival of a car at the fixed point. The assumption made in
most traffic studies is that the successive headways are independent and
identically distributed random variables. Let us denote the probability
function for the successive gaps by If the instant t = 0 at which
measurements are begun is chosen at random, uncorrelated with the arrival
of any individual car, then the probability density function for denoted
by is given by (see Adams, 1936)
where µ is the mean headway given by
It is expected that, in any realistic model of traffic, µ is finite. The most

widely used form for in traffic studies is the negative exponential
2. Queueing and Delays at Isolated Intersections 47
This density function was first observed experimentally by Adams. Later,

theoretical justification for this choice was given by Weiss and Herman
(1962), and also by Breiman (1963) and Thedeen (1964). Weiss and Herman
derive their proof in the following way.
Consider a stream of traffic on an infinitely long, homogeneous highway.

Suppose that each car is represented by a dimensionless point traveling at
constant speed, , which is sampled from a probability density function f
( ), which is not a delta function or a combination of delta functions. We
also assume that if a car reaches a slower car, it can pass immediately
without incurring any delay. Let us further assume that at time t = 0 the cars
are distributed on the highway in a way such that the probability that a car is
located between x and x + dx along the highway is where
This would be the case, for example, if
successive headways were identically distributed random variables with a
finite second moment. The preceding assumptions fully define the dynamics
of changes in the headways of the cars, and so we can calculate the
probability, that at time t, the distance between a given car and
any other car will be between x and x + dx. A detailed consideration of the
movement of the two cars, at speeds and yields
For this becomes
Thus is a constant. Translating this result into the time domain,

we conclude that the probability density for the passage of two successive
cars past a fixed point is
where q is the mean flow given by

48 Chapter 2
The preceding discussion does not take into account explicitly driver and
vehicular interactions, which would alter the resulting headway distribution.
There have been numerous contributions toward making the headway
distributions more realistic than the idealized exponential distribution.
Among other issues, the exponential distribution assigns a relatively high
weight to very small gaps, near zero, which are in effect impossible due to
the finite size of cars. One suggestion was that of Schuhl (1955) who
suggested the function
Obviously, the introduction of additional parameters increases the

likelihood of a good fit to observations. However, for the purposes of our
discussion, the exponential distribution is sufficiently successful in giving
some information regarding the queueing phenomena in traffic.
There is another important contribution to the study of successive

headways between cars, which should be mentioned here, and it is the
description of traffic as an aggregate of “traveling queues”, by Miller
(1961). He rightly observed that because of restrictions in passing
capability, cars tend to bunch into platoons which move at the speed of a
leader, leaving gaps between successive bunches. The definition of a bunch,
or queue, is arbitrary, but Miller defined it by the requirement that headways
between successive cars be no more than 8 sec., and the relative speed of
cars within a queue be within the range of -3 to 6 miles/hr. Successive
queues were assumed to be independent, and the spacing between queues
was assumed to follow a negative exponential distribution. Miller reported
good agreement with data provided by the Swedish State Roads Institute.
The distribution of the number of cars in each queue was found to be
described equally well by the distributions
where n is the number of cars in the queue, and r and m are parameters
fitted from the data.
2.2 THE GAP ACCEPTANCE FUNCTION
A driver or a pedestrian, wishing to move forward, is often impeded by

traffic which is obstructing his path. To describe this situation, the concept
of “gap acceptance” has been introduced. The idea is that a driver does not
move until the headway between successive cars blocking his path,
otherwise known as the “gap”, exceeds a certain critical value. If this gap is
denoted by G , then the behavior of the driver is summarized by a gap
acceptance function which represents the probability that the driver
will move when facing a gap G. The simplest form of the gap acceptance
function is a step function
where H(x) is the delta function; namely, H(x)=0 for x<0, and H(x)=1 for
Measurements of actual driver behavior reveal a more complicated
behavior than that described by Eq. 10. For this reason, alternative
expressions for have been suggested, in conjunction with
experimental evidence. For example, Herman and Weiss (1961) found that,
on the basis of experiments conducted under somewhat idealized conditions,
the gap acceptance function could be represented by
50 Chapter 2
where the parameters and T were found to be and T =3.3

sec.
In addition to the gap acceptance function by an individual, the

“population gap acceptance function”, has been defined as the
probability that a gap will be accepted by a randomly selected member of a
population of drivers. Assuming that for an individual driver is a step
function H(G-T), where T is a random variable with a probability density
u ( T ), the population gap acceptance function is
The population gap acceptance function is useful in interpreting field data

on gap acceptance, which by necessity involve the compilation of acceptable
gaps by a population of drivers, rather than experiments with a single driver.
At this point, some remarks are appropriate regarding the realism of the
presented models of gap acceptance. One obvious observation is that human
behavior may not be as simplistic as stipulated by the models which assume
that the size of the acceptable gap remains constant for a given individual. A
few observations will convince anyone that this may not be the case, because
the patience of a driver may be tested to the point of shrinking the size of
acceptable gaps. Another effect on gap acceptance is the succession of gaps.
A driver looking for an opening may skip a gap which appears to be
acceptable, if he detects an even more pleasing gap coming next.
Nevertheless, the methodology presented here gives a handle for estimating
the delays to traffic movement caused by traffic competing for the same
roadway space, for the purpose of planning improvements to transportation
systems.
2.3 THE DELAY TO A SINGLE CAR
Let us now assume that a single car arrives at a stop sign at time t = 0,
and faces an unobstructed stream of traffic while trying to either cross, or
merge with the main stream of traffic. We assume that the driver’s gap
acceptance function is We will calculate the probability density,
of the delay to the driver, following the work of Weiss and Maradudin
(1962). In what follows, we will talk about a “merging maneuver”, but the
results are also applicable to the case of crossing a single stream of main
traffic.
The delay time is either zero, if an acceptable gap is found at t = 0, or it

is a positive quantity if one or more cars pass the stop sign and the merging
maneuver takes place immediately after the passage of a car on the main
road. Let us assume that successive headways along the main road are
identically distributed independent random variables with probabilitydensity
The probability of an initial acceptable gap, denoted by is given
by
and therefore we can express in the form
where is the Dirac delta function, and represents the contribution

due to a delayed merge. The probability that an arbitrary gap is acceptable is
where w(t)dt is the probability that a car on the main road passes the
intersection in the time interval (t, t + dt), and no merge has been made up to
that point. We now show that w(t) is the solution to an integral equation of
convolution type. Let us define two functions
52 Chapter 2
which have the following meaning: dt is the probability that the first
gap is between t and t + dt and is unacceptable, and dt has the same
meaning for succeeding gaps. If a car in the main stream passes the
intersection at time t and no merge has been made, it is either the first car to
do so, or the last such event occurred at some time and the succeeding
gap was unacceptable. These two possibilities are summed up in the
equation
Since Eq. 17 contains a convolution integral, the use of Laplace

transforms is appropriate. Let us denote the Laplace transform of a function
of t by the same symbol of a function of s with an asterisk, so that
The Laplace transform of Eq. 17 leads to the relationship
Equations 14, 15, and 19, in concert, yield
It may be possible, in some cases, to obtain an explicit expression for

(s) by inverting Eq. 20. However, in most cases, it is sufficient to use Eq. 20
as a moment generating function, and obtain the moments of delay time. To

this end, we use the relationship for the moments
Using this formula, we find that the first two moments of the delay are
given by
When exp (– t / µ) and is the step function

the expression for becomes
If we now expand the right-hand side of Eq. 23 in series, and invert term
by term, we obtain
54 Chapter 2
The mean and variance of the delay time are readily obtained from the
expression for in Eq. 23, and they are
Some interesting variations on the preceding discussion have been

worked out to account for variations in the behavior of the driver attempting
to merge. The first one addresses the difference dictated by a yield sign as
opposed to a stop sign. In this case, a driver reaching the intersection at
some speed may be willing to accept a smaller first gap than the subsequent
gaps, which he has to face after coming to a complete stop. If we assume that
in this case is the gap acceptance function for the first gap, and
that for succeeding gaps, then in Eqs. 17 and 22, and are to be
replaced by
If we further assume
then the first and second moments of delay are found to be

Two variations on the preceding treatment have been considered by

various researchers. Weiss and Maradudin (1962) considered the case of the
“impatient driver”, whose gap acceptance behavior changes with passing
time, so that
where is the gap acceptance function for the gap preceding the arrival
of the car on the main highway. Drew et al (1967a, 1967b) also
considered this case and gave detailed results for the case in which the gap
acceptance is a step function with a threshold decreasing in some regular
manner with successive passages of cars on the main highway.
The second variation concerns a driver who considers more than the gap
he faces at any moment. Specifically, it considers a driver who also looks at
the next following gap, and waits for that one if it is larger that the one just
in front even though it is acceptable. This case was considered by Weiss
(1966), who treated the case of a negative exponential headway distribution
and a certain probability that the driver would skip an acceptable gap for the
sake of a larger following gap. The probability function assumed by Weiss
was
where T is the critical gap and is a parameter associated with the

probability that a larger following gap is preferred by the driver. It is
interesting to note that numerical results obtained on the basis of Eq. 30
show that the added delay due to such behavior on the part of the driver are
negligible compared to the basic delay due to gap acceptance limitations.
Another interesting generalization of the treatment, treated by Weiss

(1967), concerns the case in which the gap acceptance function depends not
56 Chapter 2
only on the time gap, but also on the speed of the oncoming vehicle on the
main road. In this case, let us denote the gap acceptance function by
and let be the probability that the gap is between t and t + dt ,
and the speed of the oncoming vehicle is between and The
treatment given for the standard case can be adapted by making the
substitutions for a few key variables
and analogous substitutions for the variables and Detailed

calculations were made for
which corresponds to a speed-headway distribution for light traffic. If, in

addition, can be expressed as
then the mean delay is given by
If is a linear function of
where is the mean speed, and is
with a and r representing the mean and variance of namely,
The mean delay comes out to be
where
Numerical results presented by Weiss (1967) seem to indicate that the

effect of. the dependence of gap acceptance on the speed of the oncoming
vehicle is not very pronounced, and they are unlikely to produce measurable
changes in the merging delay.
58 Chapter 2
There is an interesting departure from the treatment of the intersection

delay problem in term of finding an acceptable gap among the headways of
successive individual cars. Instead, cars are considered as arriving in
bunches, or platoons. Miller (1961), who did considerable work in modeling
platoons, proposed this “random queue” model of gap acceptance. A queue,
or platoon, was defined as a group of cars whose speed and relative spacing
did not vary more than a certain small amount, and it was assumed that no
merge was possible while a queue was passing the merge point. Let the
probability density function (p.d.f.) of the time of passage of a queue be q
(t), and let the p.d.f. of the headway between queues, i.e. the difference in
time between the end of one queue and the beginning of the next, be In
addition, let us define the mean duration of a queue as and the
mean associated with as µ . Then, by arguments based on Renewal
Theory (Cox, 1962), the probability that the merging car arrives during an
interval between two successive queues is
We now again define the p.d.f. of delay time as and its Laplace
transform as which is found to be
where is the Laplace transform of q(t). From Eq. 41, we derive the
mean delay as
and the probability of no delay as
It may be observed that the Eq. 22 for the delay caused by a sequence of
cars rather than queues, is derived from Eq. 43 by setting v = 0, which also
implies p = 1. Miller tested his model against some data on pedestrian
crossings in Sweden, finding that his “random queue” model fitted the data
slightly better than the “random car” model, and the probability of
immediate crossing was predicted definitely better by the random queue
model. An additional generalization of this model has been provided by

Weiss (1969), who considered a mixture of cars and trucks along the main
road.
Let us now consider a somewhat more complex problem. Instead of

contesting against a single stream of traffic, let us assume that a driver is
trying to cross a multilane highway. We first have to define a gap acceptance
function for crossing an n-lane highway. The first
theoretical treatment of the problem was done by Tanner (1951), who
assumed a negative exponential distribution of headways in each lane, and a
gap acceptance function
where Equation 44 assumes a certain

“symmetry” in the treatment of all lanes. However, measurements of gap
acceptance behavior for crossing two-lane highways indicate that is not a
symmetric function of the gaps (Blunden et al 1962, Wagner 1966). In fact,
it is very likely that a larger gap is required by the driver for the lanes farther
away from his position.
A model which allows for asymmetry in the treatment of the gaps at

different lanes was proposed by Gazis et al (1967). They assumed that the
gap acceptance function could be written as
and that the headway distribution in each lane was a negative exponential,
with the mean headway in lane j denoted by
A general discussion of the above model is rather difficult, but useful

results can be obtained if we impose some additional assumptions. Gazis et
al assumed that a driver waiting at the crossing point orders the gaps
according to their size, so that The gaps are then
examined in sequence starting with the shortest one and proceeding to the
longest one. Since the gap acceptance differs from lane to lane, it is possible
that a short gap may be acceptable and a slightly longer one unacceptable. If
is the shortest unacceptable gap, then the time to the next regeneration
point of the underlying renewal process is taken to be and all larger
gaps are taken out of consideration.
60 Chapter 2
We now let be the p.d.f. of the waiting time, and its Laplace
transform. First of all, the probability of zero delay is the product
where is the probability that the gap in lane r is
acceptable, namely,
If the delay is greater than zero, we define a regeneration point, as

indicated above, as the end of the smallest unacceptable gap. Let be
the p.d.f. of the time of arrival of the first unacceptable gap, and its
Laplace transform. Then we derive
and therefore
To complete the analysis, we need to calculate To this end, let us

suppose that the gap in lane r is unacceptable. Then, the first car to arrive in
any other lane either arrives after the one in lane r , or it arrives before but
the corresponding gap has been found acceptable. Therefore, we derive
If we can derive an expression for and go on to

find moments of the delay. Using the notation
we find the following expression for the expected delay

We can also derive an approximation to Eq. 51, which is quite satisfactory

for most practical applications. The approximation consists of using the
formula we derived for the delay for crossing a single lane of traffic, with
where the parameters
and are given by
A comparison of numerical results based on Eqs. 51 and 52 shows that the

approximation of Eq. 52 gives results deviating no more than 10% from
those of Eq. 51. The preceding theoretical results have been further checked
through simulation studies (Gazis et al 1967), which indicated that the
simulation and theoretical estimates of delay were within about 5% of each
other. This is not to imply that the preceding discussion provides a perfect
description of the delay in merging or crossing. There are many deviations
from the assumed behavior, which affect the delay. They concern both the
nature of the gaps, and the behavior of drivers in accepting or rejecting them.
Nevertheless, the preceding discussion gives a handle in assessing the
presence of extreme cases of crossing and merging problems.
62 Chapter 2
2.3.1 The “Flying Start”
An interesting variation of the preceding theme was also treated by Gazis

et al. It concerns a driver who approaches the intersection at some speed and
makes a decision on whether to cross or not, without stopping, (assuming, of
course, that there is no stop sign at the intersection). Since the car is in
motion, it is capable of crossing faster than it can from a complete stop.
Therefore, the gap acceptance function at time t = 0 for the first gap, to be
accepted with a “flying start”, is different from that for following gaps,
which must be accepted after a complete stop. We assume that the gap
acceptance function at time t = 0 is
where the may be different from the which are used for
subsequent gaps. Let be the Laplace transform of the p.d.f. for the
delay following the first unacceptable gap. We also define the probabilities
that the driver crosses with a flying start, or after the first unacceptable gap,
namely,
We also define the p.d.f. for the wait to the end of the first unacceptable
gap be given by the expression in Eq. 49, with replaced by
(t). Then, and satisfy the relationships
from which we obtain the solution

It may be noted that Eq. 56 reduces to Eq. 48 when If,

furthermore, we assume that the gap acceptance functions are step functions,
then the expected delay is derived in the same way as that given in Eq. 51.
Moments of the delay time are obtained by differentiating Eq. 56 and setting
s = 0, where the derivatives of and are also evaluated at s =
0. The final result is
where is given by Eq. 51 with replaced by and is given by

the same equation without any changes.
2.4 DELAYS TO PEDESTRIANS
When a pedestrian, or a group of pedestrians, arrive at a crossing point,

they may be delayed by a string of cars passing by, until they find a gap
large enough to allow them to cross the intersection. The behavior of
pedestrians can be quite complex, depending on their age, experience, and
risk aversion, but we shall use some simplifying assumptions, which have
been found to describe reasonably well the queueing of pedestrians at
crossing points.
One of the earliest contribution to the pedestrian queueing problem was

that of Tanner (1951). Tanner assumed, basically, a single lane of traffic,
although multiple lanes could be considered by using a gap acceptance
function corresponding to the minimum acceptable gap. The traffic was
assumed to have a negative exponential distribution, the pedestrians were
assumed to arrive according to a Poisson input, and all pedestrians were
assumed to have a common gap acceptance function. A more general
discussion, removing some of the restrictions of the Tanner treatment, was
provided by Weiss (1963, 1964), who also considered a single lane of traffic
but with a general stationary headway distribution, and an unrestricted gap
acceptance function. In what follows, we outline basically the Weiss
treatment of pedestrian queueing.
It is rational to assume that a pedestrian arriving at an intersection, either

crosses immediately, or joins a group of other waiting pedestrians and
crosses with them when they perceive an acceptable gap. The Weiss solution
is based on the treatment of embedded Markov chains. Transition
64 Chapter 2
probabilities, are defined for the event that “there are m people in the
queue at the regeneration point, given that n people were in the queue at the
previous regeneration point”. There are two possible cases in computing
. If n > m , the transition can only be realized in the sequence
namely, the group of n crosses as a whole, and then at least m new
pedestrians come to the crossing point, of which m remain in the queue. If,
however, the transition can either occur as or
since the initial group does not necessarily have to cross, but can be
augmented to the size m with the addition of some new arrivals. Therefore,
is decomposed as
where accounts for the indirect transition

accounts for the direct transition On the basis of our previous
discussion,
Let us consider a headway of t. On the basis of our assumption of a

Poisson arrival of pedestrians, the probability that an arriving pedestrian will
see an initial gap between x and x+dx is dx/t. Therefore, the probability
that a pedestrian arriving at some time within the interval (0,t) will cross
immediately is
Let us now calculate assuming that the pedestrian arrival rate is

If the gap is t, the pedestrian group crosses with probability or they
do not cross, and k + m pedestrians arrive with probability
Of the new arrivals, m remain in the queue and k cross the road on
arrival. These assumptions imply that is given by
where
Let us now consider which corresponds to the case when

In this case the transition can take place either in the manner
described by Eq. 60, or the original group does not cross, and k + m – n
pedestrians arrive, of whom m – n remain in addition to the original group.
On this basis, we obtain
If we now define the parameters

66 Chapter 2
then are obtainable from the transition matrix
We may now proceed to derive the equilibrium properties of the

underlying Markov chain defined by this transition matrix. Let
be the vector of steady state probabilities, meaning that is the
solution to the equation
Writing this equation in component form, we get
which can be solved by the method of generating functions. Let us define

Then Eq. 66 implies that
The moments of the queue length can now be derived from these last two
equations by differentiation, leading to the following expressions for the
mean queue length and the associated variance
Assuming and we find that

the formulas of Eq. 69 are reduced to
which are identical to those given first by Tanner (1951).
The preceding statistics are for queue lengths at a regeneration point, but
similar results can also be obtained for the statistics of queue length at a
random time. In agreement with a general result by Little (1961), it is found
that the mean number of pedestrians at a random time is
68 Chapter 2
where is given by the expression in Eq. 22. For a negative exponential

headway distribution and a step gap distribution, we obtain the following
expression for
which was also first derived by Tanner (1951).
One may observe that there are a number of deficiencies in the underlying
model of the preceding discussions. For example, the assumption that all
pedestrians have the same gap acceptance function is obviously a
questionable one. The question is, how much of an effect this assumption
has on the results. It turns out that this particular deficiency can be
circumvented quite easily. Let us suppose that the gap acceptance function
depends on the gap and on a random variable v with associated p.d.f. g(v),
and let the rate parameter also depend on v. Then, Eq. 71 changes into
It turns out that unless very heavy traffic intensities are involved, the
corrections due to a random variation in the gap acceptance function are
negligible, as was shown by Weiss (1964).
Another deficiency of the presented theory is the assumption that the gap
acceptance function is independent of the group size. If we introduce a set of
gap acceptance functions, such that is the relevant function for
a group n , the general problem cannot be solved in a closed form. However,
if we further assume that only a finite number of are different, then a
solution can be obtained. For example, if
that is, each arriving pedestrian makes a decision to cross on the basis of
(t), but thereafter makes decisions on the basis of then
where
There have been additional discussions of variations on the theme pursued

above. One of them concerns crossings in which pedestrians are given
preference, such as “zebra crossings”, so named because of the stripes
painted in the streets in the British system of this type of crossing. Smeed
(1958) presented some data regarding the delay to vehicles at such crossings,
involving two variables. The dependent variable was the journey time of a
car over a stretch of the road covering the crossing, and the independent
variable was the average number of pedestrians crossing per unit of time. A
linear relationship between the two variables was found. It turns out that, if
vehicular queueing effects are ignored and pedestrian flow is relatively
small, this linear relationship can be found on theoretical grounds.
Let us assume that the interval over which measurements are made is
large enough so that the travel time over it is a constant whether a stop is
made or not, namely, the deceleration and acceleration times due to a stop
are negligible. Then the total travel time is equal to where is the
expected delay at the zebra crossing. If the arrival of pedestrians is described
by a Poisson process with arrival rate and the driver has a step gap
acceptance function, we may use the expression given in Eq. 25 for with
If, furthermore, is small, we can expand the exponential to find
the approximation so that the total travel time comes out to be
a linear function of as found by Smeed (1958, 1968).
Incidentally, the value of T suggested by Smeed’s data lies between 5.4
70 Chapter 2
and 6.5 sec. Smeed also went on to estimate the cost of pedestrian delays,
in conjunction with the cost of traffic delays, with idea of leading to warrants
for construction of zebra crossings.
Another interesting problem concerning pedestrian crossings is one

considered by Thedeen (1969), who investigated the statistics of crossings at
intersections, where there are push-button controls. The model allows for a
Poisson arrival of pedestrians at a corner, where there is an indicator which
can register either a red (stop) or green (go) for the pedestrian. A time or
greater must elapse between the end of one green period and the beginning
of the next one. If a pedestrian arrives during a red period and pushes the
control button, a time of at least must elapse before the start of the next
green phase. This time is exactly if the first pedestrian arrives after a time
(assumed positive) measured from the end of the last green phase. If
he arrives at time the next green phase starts exactly at after
the end of the last green phase, so that the pedestrian waits a time for
the start of the next green phase. It is also assumed that a pedestrian arriving
during a green period will always be able to cross. Thedeen derived three
results, namely, the expected delay to a single pedestrian, the expected total
pedestrian delay over a long period of time, and the expected number of
pedestrians crossing during a green period. We present here the last of these
three, as an example of the mathematical techniques that can be used.
Let us assume that a green period ends at t = 0. The first pedestrian to

arrive does so either in the time interval or in the interval
In the first instance, if the pedestrian arrives at time
the time to the next green period is and the probability
that there are r arrivals in this period is In
the second instance, if the first pedestrian arrives at then the
time to the next green phase is so the probability of r arrivals in the
period Hence the probability that r pedestrians
arrive during a red phase is
Therefore, the expected number of pedestrians crossing during the

combination of a red and a green phase is
where is the duration of the green phase. Thedeen also showed that the
expected waiting time for a single pedestrian is
and the limiting value of the total waiting time, measured over a reasonably
long total period of time, is just Furthermore, the long-term portion of
time spent in the green phase, i.e. the fraction of the time during which cars
are delayed, is
where is the expected number of arrivals during the green period.
2.5 QUEUEING AND DELAYS OF VEHICLES
Let us now consider a sequence of vehicles arriving at an intersection and

attempting to merge, or cross. The difference from the previous discussion is
that the sequence of vehicles involves a succession of individual decisions
involving gap acceptance. Unlike the case of a group of pedestrians, where
all the pedestrians were assumed to cross together at an acceptable gap, this
is a case where, in some instances, the number of cars crossing is smaller
than the number of cars arriving, over a relatively large time period, so that
the size of the queue increases indefinitely. If we assume that the arrival of
cars in the feeder road is described by a Poisson process with parameter
then there exists a critical value such that
72 Chapter 2
If the queue increases without limit, but

If the queue remains finite with probability 1.
Evans, Herman, and Weiss (1964) found an expression for in the

following way. The derivation requires information about the time it takes
for a car to move from the second to the first position of the queue, as well
as information about the gap acceptance process. Let us assume that the
move-up time is which is a random variable with probability density g
and that the gap acceptance process is similar to that we assumed for the
single-vehicle merging problem. We also assume that if the first driver in
line merges during a gap G , the second driver in line sees a gap
that is, the move-up time is always smaller than an acceptable gap. The
critical rate can then be written in terms of the quantity M (t), which is the
expected number of cars from an infinite queue that merge in time t, if we
take the limit of M(t) as t goes to infinity, namely,
Evans et al showed that this limit existed, and that could be calculated
as the ratio of the expected cars to merge in a single gap over the expected
headway. If is the probability that n cars merge during a gap of t,
and E (t) is the expected number of cars that merge during that gap, then by
definition
Therefore, on the basis of its definition,
where is the p.d.f. of the headways, and µ is the mean headway.

The values of can now be computed recursively using the relations
where the second of Eqs. 84 is derived by assuming that the first car merges
with a move-up time and is followed by a merge of exactly n – 1 cars.
A single integral equation for E (t) can be derived from Eq. 84 as
Equation (85) does not have a simple solution in general, but it does in the
special case when we have a constant move-up time We then have
where is the Dirac delta function, and Eq. 84
becomes
The solution of Eq. 86 is
where is to be set equal to zero if its argument is negative. In particular,

if then
74 Chapter 2
This result implies that is given by
where In particular, when
The parameter can be found experimentally if we know the input rate,

on the feeder road, and the probability, that there are no cars in the
queue, with the additional assumption that the queue remains finite as
To this end, the following approach was suggested by Oliver
(1962). The asymptotic expected number of cars to merge in time t is
for large values of t . By equating results in the limit
we find
so that can be estimated from information available for situations with

relatively low flow volumes.
The preceding discussion is the most general one for the queueing
problem encountered in merging or crossing situations. The generality, of
course, also entails difficulty in obtaining closed form solutions for specific
situations. Several investigators have addressed such specific situations
using simplified models in which some of the features ofqueueing are either
ignored or approximated. Some of these contributions have preceded the
work described above, and others followed it. Perhaps the first treatment of
vehicular queueing was by Tanner (1953, 1962), who considered queues
generated by two conflicting traffic streams competing for the use of a
stretch of road AB wide enough for only a single vehicle. Such a situation
might arise when two lanes must merge into one. It is assumed that traffic
stream 1 arrives at point A as a Poisson process with a rate parameter
and stream 2 arrives at point A with a rate parameter A vehicle of type

j (j=1,2) can cross the segment AB at a constant time It is further
assumed that there is a minimum time gap between the passage of two
successive vehicles of the same type through AB . Tanner drew a distinction
between three cases:
(1)
(2)
(3)
The most interesting of the three, from the point of view of applications,
is the first one, and it is the only one that was treated. It corresponds to the
situation when either one of the queues has control of AB , and retains
control as long as there are still some vehicles in that queue. The other two
cases shown above allow intermittent changes of control between queues
without the intervention of any external controls.
The method of solution is a rather complicated version of the technique

of embedded Markov chains, and the results are rather complicated even for
the special cases considered. For example, when that is, when
there is no minimum gap between successive vehicles on AB , the mean
delay to a vehicle in traffic stream 1 is
with a similar result for after an interchange of the subscripts 1 and 2 .
Hawkes (1965a, 1965b) treated a slightly different version of the same

problem, in which the time needed to cross the segment AB is a random
variable, with different distributions for the two streams. The analysis leads
76 Chapter 2
to a coupled pair of integro-differential equations for the joint distributions

of virtual waiting times, and hence to an expression for the mean delay when
the system is n equilibrium. If and are the first and second moments
of delay to an isolated car crossing the intersection, and then the
generalization of Eq. 92 obtained by Hawkes is
Yeo and Weesakul (1964), and Evans, Herman, and Weiss (1964) have
treated variants of the Tanner model, in which the merging time is related to
the critical gap appearing in a step gap acceptance function. In this case, a
value of merging time is chosen from an arbitrary distribution, and this time
is the critical gap in a step gap acceptance function. Consequently, both of
these models take into account the possibility of a distribution of critical
gaps over the population of cars. The expression obtained by Evans et al is
obtained as follows. Let g (t) be the p.d.f. of the merging time, and its
Laplace transform. Let the input of cars on the feeder road be Poisson with
rate parameter and let the p.d.f. for headways be
Let us also define a dimensionless rate parameter Then, the
mean delay is obtained as
where for a stationary solution. In particular, if

that is the merging time is a constant for the entire population, then
The mean queue length can then be computed as One

interesting conclusion in the study of Yeo and Weesakul is that for low
traffic densities, the replacement of a distribution of general form for g(t) by
a delta function is a good approximation, but it shows poor

agreement with exact results at high densities.
2.6 THE DELAY AT A TRAFFIC SIGNAL
Traffic signals, and the delay they cause to individual drivers, are of
paramount importance and have received a great deal of attention. One
wants to understand the delay effect of signals, and then try to minimize it by
proper design of the signal cycle. The first problem is to find the delay to a
single traffic stream arriving at an intersection, which handles several such
streams. In order to obtain expressions for the expected delay for a single
stream, it is necessary to specify the signal settings, the arrival process of the
incoming stream vehicles, and the manner in which the vehicles pass
through the intersection.
2.6.1 The arrival process
The starting point for many studies of delays at traffic signals has been the
assumption of a simple Poisson process describing the arrival of cars. This
assumption is reasonably satisfactory for very light traffic, and the absence
of a nearby upstream signal which creates bunching of the cars into platoons.
In heavy traffic, when the interactions between cars cannot be neglected, a
generalization of the simple Poisson process is the “compound Poisson
process”, in which, if N (t) is the number of arrivals in any interval of length
t,
where I, the “index of dispersion” of the process is given by

78 Chapter 2
In this case, the index of dispersion for a compound Poisson process is

never less than unity, which is the value for a simple Poisson process. Some
experimental studies, for example that of Miller (1964), have indicated that
departures from the simple Poisson traffic are indeed in this direction, with
the dispersion index often in the range of 2. Nevertheless, it would be of
interest to obtain a more general and mathematically tractable process in
which the dispersion index can have any value, including one less than unity.
One way of doing this was suggested by Darroch34, and may be described as
follows.
In the process of binomial arrivals described in Section 2.1, let the time
be regarded as consisting of discrete, contiguous intervals, each of duration
h . Let the numbers of arrivals in the intervals h be independent, identically
distributed random variables with a probability generating function
This process specializes to the binomial process if
However, if
and if, in addition, we assume that the arrivals during a particular h interval
are randomly distributed over the interval, the process reduces to the
compound Poisson process described above. Since the index of dispersion
for the binomial and compound Poisson process are, respectively, smaller
and greater than unity, the above process generalizes the simple Poisson
process in both directions of the dispersion index. In general, it may be
shown that the dispersion index for Darroch’s process is given by
2.6.2 The Signal Settings and Departure Process
A period T of a fixed-cycle traffic light is divided into a “red phase”, a

“yellow (or amber) phase”, and a green phase. Cars are not allowed to move
when they view the combination of the red plus yellow phase, R , and they
move during the green phase, G = T – R. The simplest way in which
departures can be modeled, is by assuming that the cars move when they are
free to do so, and the departure intervals are independent and identically
distributed random variables. In the case of a single lane of traffic, it is not
unreasonable to assume that the departure headways are identical,
corresponding to some “saturation flow” rate. This assumption has been
made in most of the literature concerning fixed-cycle lights. One may expect
that the departure time of the first vehicle may be somewhat larger than that
of the rest of the vehicles, but this can be accommodated by enlarging
somewhat the value of R. It is also assumed that if the cars arrive during the
green phase, and there is no queue, they can move on, so they incur no
departure delay.
If all the vehicles do not proceed straight through the intersection, but
make a right or left turn, the department times are not necessarily the same
for all cars. Cars turning right may slow down somewhat, possibly unevenly,
and cars turning left slow down even more, unless they are given a “green
arrow” corresponding to an obstruction of the through traffic in the opposite
direction, which competes for the same road space as the left-turning
vehicles. In any case, it is reasonable to divide the cars into classes, and
assume that, for each class, the departure times are independent, identically
distributed random variables. This, of course, assumes that the classes do not
interfere with each other, as is the case if there is enough storage space to
accommodate slowed-down vehicles making a turn. If there is not enough
space, we have a “spill-back” effect, which drastically changes the
computation of delays, as discussed by Gazis (1965).
A theoretical analysis of delays due to vehicles turning contesting for

roadway space with oncoming traffic was given by Newell (1959). Darroch
(1964) also considered the interference of turning vehicles (and pedestrians),
with the straight-on movement of vehicles. Gordon and Miller (1966) also
discussed this problem, and Little (1961) evaluated the delays due to various
motorists wishing to make turning movements in Poisson traffic. However,
these last two references are concerned only with the delays to turning
vehicles, and not the delays to oncoming traffic.
2.6.3 Evaluation of Delays
In evaluating the delay to individual cars, the following parameters must

be considered. Individual cars arrive at the merging or crossing point in such
a way that the instants of arrival form a renewal process. The cars are
provided service in the order of their arrival, and the service times are
80 Chapter 2
independent, identically distributed random variables. There are three

random variables of interest: the queueing time of an individual, i.e. the time
spent waiting in the queue until service gets started, the number of individual
cars in the system at the instants of time just after the departures of
individual cars being serviced, and the busy period, i.e. the interval of time
during which there is at least one car waiting. If the distributions of either the
inter-arrival intervals or the service periods have an exponential tail, the
distributions or the their transforms, for all three random variables, are
generally obtainable, although a detailed evaluation of all these distributions
may be complicated.
The traffic signal situation is complicated, compared to classical

queueing theory, by the fact that during the red phase of the signal n service
is possible. Another difference from the classical queueing problem is that
the vehicles at the traffic signal do not necessarily depart at the order of their
arrival, except in the case of a single lane of traffic. The random variable of
most interest in the case of a traffic signal is the total delay to all vehicles
during a cycle. The reason that we do not focus on the delay to an individual
vehicle is that it is more difficult to obtain than the delay to all vehicles. This
is due to the complications introduced by the red phase. The evaluation of
the total delay is the most important one anyway, because it gives
information about the cost of system delays, and also provides inputs for
improving the settings of traffic lights. In any case, a measure of delays to
individuals is provided by the average delay, obtained by dividing the total
delay by the number of cars arriving during the cycle, starting with the red
phase. This is assuming, of course, that the intersection is undersaturated,
and all cars are services during the green phase.
Figure 1 shows the queueing process leading to the creation of a queue Q

(t) of vehicles at the traffic light, starting at the beginning of the red phase.
The delay to the vehicles during the interval (t, t + dt) is
so that the total delay to vehicles during the cycle time T is
2.6.4 Mean Delay for Poisson Input and Constant Departure

Time
Let A(t) be the number of vehicles which join the queue during th time
interval (0,t), where the origin is taken to be the start of the red phase. No
departures are possible during the red phase, (0,R), but departures proceed
unrestricted during the green phase, (R,T). We assume that A(t) is a Poisson
process with and that each vehicle takes a constant time s to
depart from the queue. Let
where
82 Chapter 2
The expected values in Eq. 105 are
To find consider first the associated variable defined by

Eq. 106 , with the upper limit of integration set to infinity. Thus, may be
regarded as the total waiting time in a busy period for a queueing process
X(t) with simple Poisson arrivals, constant service times, and
X(0) = Q(R). It can be shown, (see, for example, McNeil 1968, and Daley
and Jacobs 1969), that if
Now, to avoid negligible algebraic complications, we assume that the

green phase contains an integral number of departure times s , that is T – R
is an integer multiple of s . Since is obtained from by neglecting
that part of the queueing process after time T, we obtain, using Eq. 108,
Now let us assume that the queue is in statistical equilibrium. A necessary

and sufficient condition for this is that the average number of arrivals per
cycle be less that the number of cars that can be services during the green
phase, which is true if or
In this case E [Q(0)] = E [Q(T)] and and

Q(R) = Q(0) + A(R), therefore,
and
Using Equations 109, 110, and 112, we obtain
from which, using Eqs. 104 and 105, we obtain, for the expected total wait
per cycle,
84 Chapter 2
Dividing the right-hand side of Eq. 114 by the average number of

cars arriving in a cycle, we can find the mean delay to an individual vehicle.
Furthermore, if the traffic is relatively light, we can ignore the term
multiplied by E [Q(0)], since it is unlikely that any cars remain in the queue
at the end of the green phase. Otherwise, an expression for E [Q(0)] must be
obtained before Eq. 114 can be utilized. Before discussing this problem, we
shall discuss a generalization of Eq. 114 for more general arrival and
departure processes.
2.6.5 The Mean Delay for a More General Arrival Process
Let us assume that the arrival process is that postulated by Darroch

(1964), while departure times remain constant. Assuming that arrivals occur
at random during the h intervals, we have so Eq. 107 remains
valid. To avoid some algebraic complications in seeking an expression for E
we assume that s is a multiple of h , which is not an unreasonable
simplification. Then, it can be shown that Eq. 108 generalizes to
where I is the index of dispersion of the arrival process A(t) .

Consequently, Eq. 114 becomes
which is essentially the equation obtained by Darroch. If the arrivals are

binomial with h = s , Eq. 102 gives so Eq, 116 has the special
form
which is the formula obtained by Beckmann et al (1956). Actually, the

formula of Beckmann et al contains an s instead of 2s in the right-hand
side of Eq. 117, because in their model they assumed that the first departure
occurred at time s/2 after the start of the green phase, rather than s , as
assumed in our discussion. The discrepancy between the two formulas
disappears if the value of E [Q(0)] in the Beckmann et al discussion is
increased by which is just the average number of arrivals in half an h
interval.
Let us now consider the simplest arrival process of all, namely, the case
when the vehicles in the input stream are all separated by the constant
interval This simple model was first investigated by Clayton(1941),
who used it to obtain an optimal signal setting of a traffic signal which
minimizes the delay at an intersection. While the process Q (t) is now
completely deterministic and therefore relatively trivial, it is worth
considering here because it gives a measure of the minimum possible delay
at an intersection. The amount by which the expected delay exceeds this
minimum delay may be viewed as due to the randomness of arrivals. Clearly
in this case Q(0) 0 and I = 0, so Eq. 116 becomes
as obtained by Clayton. If we compare this formula to Eq. 116, we may

regard the second term inside the braces as arising from a possible overflow
from one cycle to the next, while the fourth term incorporates the
randomness in the arrival process.
2.6.6 Mean Delay for Random Departure Times
Random departure times may be generated by differences in behavior

among drivers, but also by the different time required for a right turn by
some vehicles, as compared with the time for straight through movement
from the same, single lane of traffic. We shall consider two important
differences in this case, the effect of randomness in departure time in the
computation of the expected delay, and the fact that the green phase may not
coincide with the end of the departure time of a vehicle.
86 Chapter 2
With respect to the effect of randomness, we assume that the departure

times are independent, identically distributed random variables, with a mean
value s and a coefficient of variation C . The arrival process is first
assumed to be a simple Poisson process.
The expected total delay during the red phase is the same as given for
constant departure times, and therefore given by Eq. 107. To find the
expected total delay during the green phase, we can use a result of
Daley and Jacobs (1969), which generalizes Eq. 108 to
The fact that the end of the green phase does not coincide with the end of
a departure time does not allow us to simply replace Q (R) by Q (T) in
Eq. 119 and subtracting. Such a replacement yields a value which is greater
than the correct value by the amount where is the expected
difference between a departure time that starts at time T and the amount by
which a departure time that ends the time period T overshoots the end of
the green phase, therefore where V is the forward recursion
time in the departure process. On the basis of renewal theory, (see, for
example, Cox 1962, p.63), we find
and then, using Eq. 119, we find

Equation 121, together with Eqs. 107, 111,112, and 120, yields
We can observe that Eq. 122 allows us to drop the requirement of

coincidence of the end of the green phase with the end of the departure time
of a vehicle for all cases, including the case of constant departure times
discussed earlier. Additional comments regarding the case of negative
exponentially distributed departure times can be found in Weiss (1974) . A
discussion of the “overflow” at the traffic signal, defined as E [Q(0)], can
also be found in Weiss.
2.6.7 Optimization of Traffic Signal Settings
The subject of optimization of traffic signal settings will be examined in

some detail in the following Chapter, on Traffic Control. An extensive
discussion of the optimization problem based on the queueing theory
approach discussed in this chapter can be found in Weiss (1974). In
particular, Weiss discusses the work of Miller (1963), who obtained
analytical expressions for the optimal signal settings of an intersection.
The general idea in Miller’s contribution is to compute the expected

delays to competing streams of traffic, and then find the signal setting which
minimizes the total delay to drivers. The problem becomes quite complex,
and various approximations are introduced to make it tractable. All these
approximations are justified on the basis of realistic expectations about the
values of the various variables involved, concerning traffic volumes, the
range of traffic signal settings, etc. One example given by Miller concerns an
intersection with four competing streams, a Northbound, a Southbound, an
Eastbound, and a Westbound. For such an intersection, one can obtain the
rate of delay, in vehicle hrs / hr, for different settings of the fixed-signal. The
settings are associated with the four values of which is the fraction of
effective green time available to traffic stream entering from direction i.
Figure 2 below shows the rate of delay obtained by Weiss for a particular
example of traffic input rates. In this example, these rates are:
Northbound 720 vehicles/hr
Southbound 1,080 vehicles/hr
Eastbound 180 vehicles/hr
88 Chapter 2
Westbound 360 vehicles/hr.

Clearly the North/South green setting of the traffic light must be larger than
the for East/West traffic which is considerably lighter than the North/South
traffic, and this setting is shown in Fig. 2 as The “saturation flow” ,
i.e. the rate of departure of vehicles from the queue, was assumed to be
1,800 vehicles/hr.
It is seen from this figure that the rate of delay increases rapidly for values
of the light cycle time smaller than the optimum, but less rapidly for values
larger than the optimum. This means that it is wise to err on the safe side by
stretching the traffic light cycle toward somewhat larger values than the
computed optimum, if such a setting is realistic.
Numerous contributions to the analysis of “bulk service” queues have

been made over the years, many of them aiming at improvements to the
theory that would make it easier to implement. We could cite the work of
Downton (1955), Gaver (1959), Miller (1959), Jaiswal (1964), Neuts (1967),
Cohen (1969), Teghem et al (1969), and Chaudry (1979). A particularly
useful contribution was that of Powell (1953), who proposed a method for
obtaining an approximate, closed form moment formula for bulk arrival,
bulk service queues. In contrast to previous work, which relied on classical
transform methods, requiring the determination of complex zeros of some

functions, or iterative techniques, Powell used transform results to obtain
approximate, closed form formulas. The approximate formulas were
obtained by solving exactly the problem for several values of the input, and
then using linear regression, in which the unknown functions were the
dependent variables, and some appropriate input functions were the
dependent variable.
2.6.8 The Vehicle-Actuated Traffic Signal
The models we have presented thus far involve traffic lights in which the red
and green phases are of fixed duration. However, there are situations, in
which one direction may start getting higher input, while the cross direction
still has portions of the green phase to spare. If we have the intersection
properly instrumented to detect such variations in real time, we can hope to
borrow some green light from the under-loaded direction and give it to the
over-loaded one. For example, we can wait until the queue in the overloaded
direction empties before we give the green to the cross traffic, thus utilizing
the available total green time more effectively, and reducing the overall
delay to users. This defines traffic actuation of signals, in which the green
and red phases of both directions vary in response to actual measurements of
demand. If the departure processes in the two competing directions are the
same, the best policy is to wait, if possible, until the queue in the over-loaded
direction empties, and then switch. The challenge, of course is to position
traffic sensors properly, so that they can determine the end of the queue. This
end is generally defined by measuring the headways of the arriving vehicles.
It should be pointed out here that in most cases there is maximum allowable
green period, intended to avoid having the vehicles in the opposing direction
wait for inordinate amounts of time. (It has been said that “if a red phase is
more than two minutes long, drivers assume that the light is broken down
and choose to ignore it”). This limitation on the duration of the green light
may negate our desire to completely exhaust the queue in the over-loaded
direction, even though this may result in an increase in the expected delay to
the users.
In the case of the fixed-cycle system, it was possible to determine the

expected delay, and go on to optimize the signal setting, by considering each
competing stream of traffic separately. This is not possible for traffic
actuated signals, because the two streams must be considered together in
seeking optimization. The analysis becomes even more complicated if we
allow turning movements, and hence contention for green time by more than
two streams. Consequently, the case of just two competing streams at an
intersection of two one-way streets has received most of the attention, since
90 Chapter 2
it is more tractable mathematically. What is lacking from most of the studies

of this problem is the application of the “California rule” of stop-and-then-
go, if you have space for the “go” part.
Even in the simple case of two intersecting on-way streets, the analysis
gets to be quite complicated, unless we remove the upper bound for a green
phase, as was done, for example, by Tanner (1953), and by Darroch et al
(1964). In both these papers, it is assumed that the arrivals are a simple
Poisson process, but Darroch et al allow for variable departure times and lost
times, whereas Tanner does not. Tanner’s model is more realistic in dealing
with the change-over of allocation of the green, at the expense of increased
mathematical complexity, which does not burden the model of Darroch et al.
In the analysis that follows, we simplify matters by assuming that all the
departure times have a fixed duration s , the lost times are of fixed length L,
and the competing streams of traffic are a simple Poisson process with rate
in the i direction. It is further assumed that the traffic signal may extend
the green in either direction until the queue is exhausted, which is
determined by the detection of a headway of at least in the arrivals near
the end of the queue. This is a special case of the Darroch et al model, but
their generalization to independent and identically distributed departure
times, possibly with different distributions for the two competing streams, is
straight-forward. The optimization problem consists of finding the values of
which minimize the total rate of delay at the intersection. Since there is
no “overflow” at the end of the green phase, with headways exceeding
the expected total waiting time per cycle is given byEq. 116, without the
term involving E [Q (0)]. The red period is now a random variable, and so
we have, in equilibrium,
In order to evaluate the first two moments of (Fig. 3), we consider

two successive light cycles and obtain
where is the time it takes to discharge the queue i, is additional time that
elapses before the beginning of a headway of at least and the
superscript j denotes the cycle number.
Suppose that there are vehicles queued in the direction i at the

beginning of its green phase. Then is the convolution of busy
periods in a queue with simple Poisson arrivals with rate and constant
service times of duration s . We can now use known results from queueing
theory (see, for example, Cox and Smith 1961, p. 55) and obtain
Now, is the number of arrivals in the red period and therefore

it is Poisson distributed with mean Therefore, taking expectations in
Eq. 125, we obtain
92 Chapter 2
Now, taking expected values in Eq. 124 and using Eq. 126, we obtain
Solving Eq. 127 for the two we obtain
and a similar expression for with the subscripts 1 and 2 interchanged

in Eq. 128, for reasons of symmetry.
Next, we obtain the average cycle time. From Fig. 3, we see that
From Eqs. 128 and 129, we obtain

The values of can be obtained by squaring both sides of Eq. 124

and taking expectations. Since and are independent, we obtain
and a similar expression for with the subscripts 1 and 2

interchanged. Explicit expressions for all the expected values of variables
can now be obtained using Eqs. 126 and 128, but they are too complicated to
show here.
To obtain the total rate of delay per unit of time, i.e.

we have to determine the first two moments of
which is just the waiting time for stream i until the sensing of the first gap
of at least duration , and therefore, using Eqs. 25 we obtain
It can now be shown that the condition for equilibrium is, as expected,
Next, we can determine the values of which minimize the total delay
per unit of time. In general, graphical or numerical methods are required for
this task, except in special cases. Interestingly enough, the derivation of the
rate of delay is simplified considerably if the intersection is operating close
to saturation, that is, if is very close to unity. In such a case, the
right-hand side of Eq. 128 is very large, since its denominator is very small.
This results in and being the dominant terms in the right-
94 Chapter 2
hand side of Eq. 131. Neglecting the other terms of the right-hand side and
using Eq. 126, we obtain
and a similar expression for with the subscripts 1 and 2

interchanged. In addition, if is close to unity, Eqs. 128 and 130
yield
We now solve Eq. 134 and its counterpart for as simultaneous

equations, and using Eq. 135 we obtain
Using Eq. 130, we can further reduce this expression to
From Eqs. 123 and 136 we now obtain a good approximation for the rate
of delay, as tends to unity, given by
The right-hand side of Eq. 136 only involves the values of through the
mean cycle time E [T ] . The values of for which E [T ] is minimum are
obtainable from Entering Eqs. 132 into Eq. 130 and
differentiating, we obtain the optimal values
subject to the additional requirement that Eq. 138 gives the values
of which minimize the total rate of delay, at high traffic volumes. It
should be pointed, however, that the usefulness of the result may be limited
by constraints on the maximum value of the red phase, in which case one
may seek settings which are as close to the optimum obtained above as
possible.
A recent review of queueing problems in transportation has been provided

by Hall (1999). In this review, Hall discusses delays not only in roadway
traffic situations, but also in railroad train movements, covering additional
aspects of queueing in transportation, with an emphasis on “bulk service
models”, which arise naturally in traffic because of the consolidation
function performed by the vehicles. A particularly interesting contribution to
the optimization of the management of bulk arrival, bulk service queues has
been given by Powell (1986). Powell bypasses the honerous task of
evaluating complex zeros of functions derived from transform methods used
in queueing studies by seeking suitable approximations to the equations
involved. He obtains closed form expressions for the mean and variance of
the length of the queue at departure instants, as well as the moments of the
length of the queue in continuous time, and of the waiting time distribution.
96 Chapter 2
References
Adams, W. F. (1936), “Road Traffic Considered as a Random Series”, J. Inst. Civil

Engrs., 4, 121-130.
Beckmann, M. J., C. B. McGuire, C. B. and C. B. Winsten, C. B. (1956), Studies in the
Economics of Transportation , New Haven: Yale University Press.
Blunden, W. R., Clissold, C. M. and Fisher, R. B. (1962), “Distribution of Acceptance
Gaps for Crossing and Turning Maneuvers”, Proc. Aust. Road Res. Board, 1, 188-205.
Breiman, L. (1963), “The Poisson Tendency in Traffic Distribution”, Ann. Math. Statist.,
34, 308-311.
Chaudry, M. L. (1979), “ The Queueing System M* / G / 1 and its Ramifications”, Naval
Res. Legist. Quart., 26, 667-674.
Clayton, A. J. H. (1941), “Road Traffic Calculations”, J. Inst. Civil Engrs., 16, No. 7, 247-
284 (1941); 16, No. 8, 588-594.
Cohen, J. W. (1969), The Single Server Queue, London: North Holland.
Cox, D. R. and Smith,, W. L. (1961), Queues, London: Methuen.
Cox, D. R. (1962), Renewal Theory, London, Methuen.
Daley, D. J. and Jacobs, D. R. (1969), “The Total Waiting Time in a Busy period of a
Stable Single-Server Queue, II), J. Appl. Prob., 6, 573-583.
Darroch, J. N. (1964), “On the Traffic-Light Queue”, Ann. Math. Statist., 35, 380-388.
Darroch, J. N., Newell, G. F. and Morris, R. W. J. (1964), “Queues for a Vehicle-Actuated
Traffic Light”, Operations Res., 12, 882-895.
Downton, F. (1955), “Waiting Time in Bulk Service Queues”, J. Roy. Stat. Soc., B 17,
256-261.
Drew, D. R., LaMotte, L. R., Buhr, J. H. and Wattleworth, J. A. (1967), “Gap Acceptance
in the Freeway Merging Process”, Report 430-2, Texas Transportation Institute.
Drew, D. R., Buhr, J. H. and Whitson, R. H. (1967), “The Determination of Merging
Capacity and its Applications to Freeway Design and Control”, Report 430-4, Texas
Transportation Institute.
Evans, D. H., Herman, R. and Weiss, G. H. (1964), “The Highway Merging and Queueing
Problem”, Operations Res., 12, 832-857.
Gaver, D. P. (1959), “Imbedded Markov Chain Analysis of a Waiting Line Process in
Continuous Time”, Ann. Math. Stat., 30, 698-720.
Gazis, D. C. (1965), “Spill-back from an Exit Ramp of an Expressway”, Highway
Research Record, 89, 39-46.
Gazis, D. C., Newell, G. F., Warren, P. and Weiss, G. H. (1967), “The Delay Problem for
Crossing an n Lane Highway”, Proc. 3rd Intern. Symp. on Traffic Flow, New York, Elsevier,
pp. 267-279.
Gordon, J. D. and Miller, A. J. (1966), “Right-Turn Movements at Signalized
Intersections”, Proc. Third Conf. Aus. Road Res. Board, 1, 446- 459.
Grace, M. G., Morris, R. W. J. and Pak-Poy, P. G. (1964), “Some Aspects of Intersection

Capacity and Traffic Signal Control by Computer Simulation”, Proc. Second Conf. Aus. Road
Res. Board, 1, 274-304.
Hall, R. W. (1999), “Transportation Queueing”, in Handbook of Transportation Science,
edited by R. W. Hall, Boston: Kluwer Academic Publishers.
Hawkes, A. G. (1965a), “Queueing at Traffic Intersections”, Proc. 2nd Intern. Symp. on the
Theory of Traffic Flow, edited by J. Almond, Paris: Office of Economic Cooperation and
Development, pp. 190-199.
Hawkes, A. G. (1965b), “Queueing for Gaps in Traffic”, Biometrica, 52, 79-85.
Herman, R. and Weiss, G. H. (1961), “Comments on the Highway Crossing Problem”,
Jaiswal, N. K. (1964), “A Bulk Service Queueing Problem with Variable Capacity”, J.
Roy. Stat. Soc., B 26, 143-148.
Little, J. D. C. (1961), “Approximate Expected Delays for Several Maneuvers by a Driver
in Poisson Traffic”, Operations Res., 9, 39-52.
McNeil, D. R. (1968), “A solution to the Fixed Cycle Traffic Light Problem for
Compound Poisson Arrivals”, J. Appl. Prob, 5, 624-635.
Miller, A. J. (1961), “A Queueing Model for Road Traffic Flow”, Proc. Roy. Statis. Soc.,
B-23, 64-90.
Miller, A. J. (1963), “Setting for Fixed-Cycle Traffic Signals”, Operations Res., 14, 373-
386.
Miller, A. J. (1964), “Settings for Fixed-Cycle Traffic Signals”, Proc. 2nd Conf. Aus. Road
Res. Board, 1, 342-365.
Miller, R. G. (1959), “A Contribution to the Theory of Bulk Queues”, J. Roy. Stat. Soc., B
21, 320-337.
Neuts, M. F. (1967), “A General Class of Bulk Queues with Poisson Input”, Ann. Math
Stat. 38, 759-770.
Newell, G. F. (1959), “The Effects of Left Turns on the Capacity of Road Intersections”,
Quart. Appl. Math., 17, 67-76.
Oliver, R. M. (1962), “Distribution of Gaps and Blocks in a Traffic Stream”, Operations
Res., 10, 197-217.
Powell, W. B. (1986), “Approximate Closed form Moment Formulas for Bulk Arrival,
Bulk Service Queues”, Transp. Sci., 20, 13-23.
Schuhl, A. (1955), “The Probability Theory Applied to Distributions of Vehicles on Two-
Lane Highways”, Poisson and Traffic, New Haven: Eno Foundation, pp. 59-72.
Smeed, R. J. (1958), “Theoretical Studies and Operational Research on Traffic and Traffic
Cogestion”, Bull. Intrn. Statist. Inst., 36, 347-375.
Smeed, R. J. (1968), “Aspects of Pedestrian Safety”, J. Transport Econ. Policy, 2, No. 3,
1-25.
Tanner, J. C. (1951), “The Delay to Pedestrians Crossing a Road”, Biometrika, 38, 383-
392.
98 Chapter 2
Tanner, J. C. (1953), “A Problem of Interference Between Two Queues”, Biometrica , 40,

58-69.
Tanner, J. C. (1962), “A Theoretical Analysis of Delays at an Uncontrolled Intersection”,
Biometrica, 49, 163-170.
Teghem, J. Loris-Teghem, J. and Lambotte, J. P. (1969), Modeles d’ attente M / G / 1 et
GI / M / 1 a Arrivees et Services en Groups, Lecture notes in Operations Research and
Mathematical Economics, 8, New York: Springer-Verlag.
Thedeen, T. (1964), “A Note on the Poisson Tendency in Traffic”, Ann. Math. Statist., 35,
1823-1824.
Thedeen, T. (1969), “Delays at Pedestrian Crossings of the Push-Button Type”, Proc. 4th
Symp. on Theory of Traffic Flow, Bonn: Bundesminister fur Verker, pp. 127-130.
Wagner, F. A. (1966), “An Evaluation of Fundamental Driver Decisions and Reactions at
an Intersection”, Highway Res. Rec., 118, 68-84.
Weiss, G. H. and Herman, R. (1962), “Statistical Properties of Low Density Traffic”,
Quart. Appl. Math., 20, 121-130.
Weiss, G. H. and Maradudin, A. A. (1962), “Some Problems in Traffic Flow”, Operations
Res., 10, 74-104.
Weiss, G. H. (1963), “ An Analysis of Pedestrian Queueing”, J. Res. Natl. Bur. Stds., 67B,
229-243.
Weiss, G. H. (1964), “Effects of a Distribution of Gap Acceptance Functions on Pedestrian
Queues”J. Res. Natl. Bur. Stds.,68B, 13-15.
Weiss, G. H. (1966), “The Intersection Delay Problem with Correlated Gap Acceptance”,
Weiss, G. H. (1967), “The Intersection Delay Problem with Gap Acceptance Function
Depending on Space and Time”, Transport. Res., 1, 367-371.
Weiss, G. H. (1969), “The Intersection Delay Problem with Mixed Cars and Trucks”,
Transport. Res., 3, 195-199.
Weiss, W. H. (1974), “Delay Problems for Isolated Intersections”, Chapter 2 in Traffic
Science, edited by D. C. Gazis, New York: John Wiley and Sons.
Yeo, G. F. and Weesakul, B. (1964), “Delays to Road Traffic at an Intersection”, J. Appl.
Prob., 2, 297-310.
Additional References
Ashworth, R. (1970), “The Analysis and Interpretation of Gap Acceptance Data”, Transp.
Sci., 4, 270-280.
Blumenfeld, D. E. and Weiss, G. H. (1970a), “On the Robustness of Certain Assumptions
in the Merging Delay Problem”, Transp. Res., 4, 125-139.
Blumenfeld, D. E. and Weiss, G. H. (1970b), “On Queue Splitting to Reduce Waiting
Time”, Transp. Res., 4, 141-144.
Blumenfeld, D. E. and Weiss, G. H. (1971), “Merging from an Acceleration Lane”,
Transp. Sci., 5, 161-168.
Buckley, D. J. (1968), “A Semi-Poisson Model of Traffic Flow”, Transp. Sci., 2, 107-133.

Burns, L. D., Hall, R. W., Blumenfeld, D. E. and Daganzo, C. F. (1965), “Distribution
Strategies that Minimize Transportation and Inventory Cost”, Operations Res., 33, 469-490.
Cheng, T. E. C. and Allan, S. (1992), “A Review of Stochastic Modeling of Delay and
Capacity at Unsignalized Intersections”, European J. of Operations Res., 68, 247-259.
Dunne, M. C. (1967), “Traffic Delay at a Signalized Intersection with Binomial Arrivals”,
Transp. Sci., 1, 24-31.
Edie, L. C. (1956), “Traffic Delays at Toll Booths”, Operations Res., 4, 107-138.
Gibbs, W. L. (1968), “Driver Gap Acceptance at Intersections”, J. Appl. Psych, 52, 200-
211.
Grace, M. M. and Potts, R. B. (1964), “A Theoty of Diffusion of Traffic Platoons”,
Grace, M. G., Morris, R. W. J. and Pak-Poy, P. G. (1964), “Some Aspects of Intersection
Capacity and Traffic Signal Control by Computer Simulation”, Proc. Second Conf. Aus. Road
Res. Board, 1, 274-304.
Haight, F. A., Bisbee, E. F. and Wojcik, C. (1962), “Some Mathematical Aspects of the
Problem of Merging”, Highway Res. Board Bull., 356.
Hall, R. W. (1991), Queueing Methods for Services and Manufacturing, Englewood Cliffs,
NJ: Prentice Hall.
Lindley, D. V. (1952), “The Theory of Queues with a Single Server”, Proc. Comb. Phil.
Soc., 48, 277-289.
May, A. D. and H. E. Keller, H. E. (1967), “A Deterministic Queueing Model”, Transp.
Res., 1, 117-128.
Miller, A. J. (1972), “Nine Estimators of Gap-Acceptance Parameters”, Proc. 5th Intern.
Symp. on the Theory of Traffic Flow ans Transp., G. F. Newell, Ed., New York: Elsevier, pp.
215-235.
Newell, G. F. (1965), “Approximation Methods for Queues with Application to the Fixed-
Cycle Traffic Light”, SIAM Rev., 7, 223-240.
Newell, G. F. (1982), Applications of Queueing Theory, London: Chapman and Hall.
Oliver, R. M. and Bisbee, E. F. (1962), “Queueing for Gaps in High Flow Traffic
Streams”, Operations Res., 105-114.
Raff, M. S. (1951), “The Distribution of Blocks in an Uncongested Stream of Automobile
Traffic”, J. Amer. Statist. Assoc., 46, 114-123.
Solberg, P. and Oppenlander, J. D. (1966), “Lag and Gap Acceptances at a Stop-
Controlled Intersection”, Highway Res. Rec., 118, 68-84.
Wardrop, J. G. (1968), “Journey peed and Flow in Central Urban Areas”, Traffic Engr.
Cont., 528-532.
Weiss, G. H. (1964), “Effects of a Distribution of Gap Acceptance Functions on Pedestrian
Queues”, J. Res. Nat. Bur. Stds., 68B, 13-15.
Chapter 3
Traffic Control
Traffic problems, and measures of traffic control intended to reduce these

problems, predate the automobile. In “Famous First Facts” by Kane (1964),
we find information related to traffic problems encountered by horse driven
carriages in 1791. A New York City ordinance at the time proclaimed:
“Ladies and Gentlemen will order their Coachmen to take up and set down
with their Horse Heads to the East River, to avoid confusion”. Thus, the first
one-way street was implemented. For many year after that, traffic problems
continued to intensify, partially reinforced by the mixture of traffic the road
facilities were expected to serve. The automobile firmly established the
existence of traffic problems, and the steady deterioration of traffic
networks, insofar as driver comfort is concerned. There have been brief
periods when, in some locations, improvements in traffic systems were
catching up with traffic, but over about one century of the existence of the
automobile, the traffic situation has gone from bad to worse.
Traffic control came about mostly for reasons of safety, to avoid

coincidence of two cars in space at the same time. The earliest devices of
traffic control were hand signals of traffic policemen, sometimes mistaken
by early motorists as friendly greetings. The first traffic light was installed in
Cleveland in 1914, (see Marsh, 1927), although Detroit has claimed being
the first, with a “first traffic light” shown at the Ford Museum in Dearborn,
Michigan. Traffic lights proliferated, and the evolution of signals operating
under advanced design concepts is shown in Fig. 1. The evolution was
intended to accomplish something beyond safety, namely, to reduce
unnecessary stops by individual cars and consequently reduce delays and the
overall travel time required for a trip.
102 Chapter 3
Coordination of a string of traffic signals was intended to allow cars to

move without stops through a sequence of intersections, preferably in both
directions. The first such “progression” system was installed in Salt Lake
City, in 1918, (see Marsh, 1927). Later on, the concept of the “multi-dial”
system came along. The idea was to adjust progression settings for different
traffic conditions, given the habits of the driving population. The multi-dial
system generally provided three settings, one for morning rush hour, one for
evening rush hour, and one for average conditions. Later, attempts were
made to make the transition from one synchronization scheme to another
responsive to actual measurements of traffic. The best known example of
such a system was the “PR System” installed in Baltimore in the early
1950s.
On the other side of the picture, “traffic actuated signals” were

introduced, whose settings were adjusted to respond to actual traffic
measurements taken by sensors. Some of these traffic- actuated signals
detect traffic only on a secondary street intersecting a main street, and keep
the main street traffic moving until some traffic appears on the secondary
street. Others measure traffic in both competing directions and allocate
“green phases” to these competing streams appropriate for their respective
volumes. All of these signals were designed for isolated intersections.
Computerized traffic control was introduced in the mid-1960s, with

leading projects in Toronto, Canada, and San Jose, California, (see Gazis and
3. Traffic Control 103
Bermant, 1966, and San Jose Project, 1966). In these computerized systems,
progression choice was based on actual measurements of traffic, and it was
coordinated with traffic-actuated control of “critical intersections”. Finally,
in recent years, the concept of “Intelligent Transportation Systems (ITS” has
been introduced. An ITS system is supposed to optimize the design of traffic
lights in a network, and also inform the drivers of events that influence the
performance of the traffic system, such as accidents and other “incidents”.
In this chapter, we start with a discussion of the objectives of traffic

control. We then discuss the elements of traffic control, including the
optimization of the operation of a single traffic light, or a group of traffic
lights, optimal synchronization schemes, and finally optimal operation of
control devices in a network. Particular attention is paid to the issue of
congestion, which causes networks to function as “store-and-forward”
systems, in which queueing is virtually inevitable, and the throughput of the
network depends on the throughput of critical points defined as “Throughput
Limiting Points (TLPs)”.
3.1 OBJECTIVES OF TRAFFIC CONTROL
There are two main objectives in traffic control. One is Safety, and the
other one is Comfort and Convenience of Drivers. The safety objective is
obvious, and it consists of separating streams of traffic to avoid collisions,
and instructing drivers on points of danger, such as sharp curves. The
objective of maximizing the comfort and convenience to drivers comprises
means for allowing drivers to complete their journey in as short time as
possible, and means for easing the driving task, for example by reducing the
number of unnecessary stops. Viewed as a control theory problem, traffic
control entails the goal of improving an objective function , subject to
constraints of the traffic system. Optimization is to be achieved by selecting
any of a number of controllable parameters of the system. Such parameters
include the split , i.e. division of the green phase among competing streams,
the offset of a sequence of lights, i.e. the time difference between the start of
the green phase of these lights, etc.
The objective function of the optimization problem can be any

combination of the following measures of performance of a network:
Average delay per vehicle,

Maximum individual delay,
Percentage of cars that are stopped as they traverse the system,
104 Chapter 3
Average number of stops for an average trip,

Throughput of intersections or the system at large, and
Travel time through the system (average or maximum).
The average delay, or equivalently the “aggregate delay” to all users, is
the most commonly used objective function, and it is related to the last item
listed above. The other criteria of performance listed above are used either as
constraints in the optimization process, or as objective functions in special
situations. For example, minimizing the average number of stops is the
objective of the design of progression, although it is not explicitly stated as
an objective function.
We shall discuss the following traffic systems, which have been

investigated over the years:
An isolated intersection,
A system of intersections amenable to signal coordination, or
coordinated control for optimizing the objective function,
Critical traffic links, such as tunnels, bridges, and special freeway
sections, and
An urban street network.
Since the onset of computerization of traffic control, some of the above

systems have been considered in combination. For example, traffic signals in
a network have been designed to accomplish two types of control,
macrocontrol, (or major loop control), and microcontrol, (or minor loop
control). The idea is to devise an overall strategy of coordination of traffic
lights in a network, which makes the largest possible contribution to the
optimization process, e.g. toward minimizing delays to drivers. Such a
global strategy constitutes the macrcontrol. Over and above that,
microcontrol is instituted, comprising the optimal management of “critical
points” in the network, generally in response to actual real-time
measurements of traffic in the vicinity of these points.
The pursuit of ITS has, in principle, integrated all the above objectives,
and added a few others, as will be discussed later. One of the distinguishing
points of ITS is the desire to communicate with drivers, in order to alert
them as to the current traffic conditions. Such communication is expected to
influence the route choices of drivers, thus affecting the inputs to the
optimization problem for the network.
3.2 SINGLE, ISOLATED INTERSECTION
Let us consider an isolated intersection serving two conflicting streams

perpendicular to each other. The intersection may actually be serving four
streams, i (= 1,2), two in the north/South and two in the East/West
directions, but it is assumed that two conflicting streams are dominant at a
particular time of the day, a normal occurrence in urban traffic. We define
the following quantities
= Traffic flow volume along the i direction

= Saturation flow along the i direction
= Duration of green phase favoring stream i.
= Common cycle of all intersection along an artery
The saturation flow is defined as the flow achieved when a platoon of

vehicles crosses the intersection, starting from a complete stop. It is tacitly
assumed that individual vehicles depart with the same headway, but a slight
variation from vehicle to vehicle would not alter the results substantially. We
define, in addition, the traffic cycle as c , and the total lost time during
acceleration and clearing of the intersection, (mostly equal to the yellow
phase), as L .
We can now use the following reasoning: During the green phase
the throughput in the i direction is It is reasonable to set the green
phases in a way that provides equal saturation levels, in both
directions, where of course it is assumed that these saturation levels, and
their sum, are smaller than unity. This yields to the empirical formula5 for
the split of the effective green
where the green phases in the two directions satisfy the relationship
The above relationships have been the bible of traffic engineers for ever.
Webster (1958) investigated the optimum setting of traffic lights under
realistic assumptions about arrival patterns and queue discharge during the
106 Chapter 3
green phase, and obtained the optimum traffic cycle, and “split” of the green
between the two directions. Webster found that the split defined by Eqs. 1,2
was pretty close to the one obtained by his optimization process, in most
practical situations.
As was mentioned already, the saturation levels in the two directions

should add up to less than unity. In fact, we must have
The condition of Eq. 3 guarantees that the intersection is

“undersaturated”. The case of the “oversaturated” intersection is discussed
later in this chapter.
The delay to drivers at a single intersection can be further reduced if one

takes into account the variation of inputs. This can be achieved in practice
with proper instrumentation measuring the traffic inputs, leading to the
“traffic actuated” operation of the traffic light. Dunn and Potts (1964) have
developed an adaptive control algorithm based on such real-time
measurements of traffic input.
3.3 SYNCHRONIZATION SCHEMES FOR ARTERIAL

TRAFFIC
The objective of properly synchronizing a string of traffic signals along an

artery is to allow as many drivers as possible to go through the sequence of
traffic lights without stopping, preferably in both directions. This provides
“comfort and convenience” to drivers, the comfort reflected by the reduced
number of stops, and the convenience by reduced travel times. It is assumed,
at this point, that the system in “undersaturated, that, is, every intersection
has adequate aggregate green time to serve the traffic in both directions.
Figure 2 shows an atypical time-space diagram. It is atypical because, as

every traffic engineers knows, life is not so kind as to provide many
opportunities of relatively regular arrangements in space and time as those
shown in Fig. 2.
On the basis of the diagram shown in Fig. 2, we can discuss some of the
basic properties of a progression design, which allow variation of a single
design in order to accommodate varying demands. Extensive discussions of
the principles of progression design can be found in Dunne and Potts (1964),
and in Matson et al (1955).
The objective of a progression design is to maximize the green through

bands in both directions of an artery. The efficiency of a progression design
is defined as the sum of the widths of the two through bands divided by the
cycle, indicating the percentage utilization of the cycle. It may be seen that
the efficiency of a design does not change if we perform a number of
transformations, such as the following:
We can stretch the time scale of the diagram in Fig. 2 . This amounts to a
change from a set of speed and cycle (v , c) to a set with
an arbitrary constant, with the same efficiency of progression.
108 Chapter 3
We can “shear” the diagram of Fig. 2 from top to bottom, obtaining a

progression scheme with different speeds in the two directions and the
same efficiency.
We can trade off some decrease of the bandwidth in one direction for an
equal increase in the opposite direction, thus accommodating increasing
traffic volumes in the direction getting the increase. Again, the
efficiency remains the same.
The stretch of the time scale may be made at sometime corresponding

to changing traffic conditions. It is well known in traffic engineering that
during light traffic conditions, the shortest possible cycle time is the best. As
the traffic volume increases, we want to be able to increase the cycle time, as
shown, in Fig. 3. The increase in cycle time corresponds to a decrease in
speed, which is anyway a natural consequence of increased traffic volumes.
On the basis of all three of the above observations we can simply seek an
optimal design for some speed and cycle, from which we may them obtain a
multitude of alternate designs accommodating different traffic demand
situations. An optimal design can be obtained from the following
observations.
From symmetry, we see that the offsets between successive intersections,

(i.e. the time differences between the middles of the green phases), must be
such that the middle of the green phase of any intersection must coincide
with the middle of either the green or the red phase of any other intersection.
If, in addition, all travel times between intersections are integer multiples
of the half-cycle, we can achieve a perfect utilization of the green phase at
all intersections, and in both directions, resulting in the highest possible
efficiency. If the distances between intersections do not allow this perfect
solution, we have some right or left interference at various intersections,

resulting in less than full utilization of the green phases. This is shown in
Fig. 4 below, where the case of full utilization of the green phase, diagram
(a), is contrasted with the case of partial utilization, diagram (b). Depending
on the alignment of mid-phases, some of these intersections produce a right
or left interference, or both, in which case they are “critical” intersections.
For example in Fig, 2, the critical intersections are Nos, 0, 5, 6, and 7. It is at
these intersections that adjustments can be made to enlarge the bandwidth of
one direction at the expense of the other. In such cases, one can try to
improve the system by finding an alignment of midphases such that the sum
of the maximum left interference and the maximum right interference is
minimized. This can be accomplished with the following finite algorithm.
110 Chapter 3
1. Let be the distances of the intersections from the one with the
minimum green phase, taken as the reference intersection, the green
phases of these intersections, and the green phase of the reference
intersection. We compute
where the symbols “mod” and “int” denote the modulo (i.e., the
residue after division), and “int” denotes the integer part of the division.
The quantities li and are the left and right interferences of the light
i on the band emanating from the reference intersection. Aligning the
mid-phases results in intersection i being associated with either a right
or left interference.
2. We then arrange the left interferences in descending order. Let be

this arrangement, and the analogous arrangement for the right
interferences. On feasible solution is to select an alignment of mid-
phases corresponding to all left interferences, in which the bandwidth is
3. We can now try to improve the solution by trading off left interferences
for right interferences. This is accomplished by shifting the offset of the
corresponding intersections by c/2 . If the number of left interferences
is k , then the optimum bandwith is obtained by maximizing over k the
quantity
It should be noted that not all values of k must be considered, but only
those corresponding to increasing values of
4. After we are through with this interim rearrangement of midphases, we

align them by computing
and setting the mid-phases of the intersection i identical to that of the

reference intersection when and opposite when
112 Chapter 3
3.3.1 A Mixed Integer Linear Programming Approach
An interesting variation of the progression design was introduced by

Little (1964), who allowed small changes of speeds between pairs of
intersections, on the theory that drivers could learn to adapt to these changes.
Allowing these changes could generally result in wider through-bands.
Little addressed his adjusted progression design problem for an entire

network using a mixed-integer linear programming (MILP) approach. The
objective function was a weighted sum of the bandwidths of all arteries in
the network. Each bandwidth depends linearly on the offsets that are the
continuous variables in the MILP formulation. The offsets must satisfy
certain constraints involving some integer variables. These constraints
express the fact that adding relative offsets between successive pairs of
intersections along a closed path in the network, one must obtain an integer
multiple of the cycle. In principle, Little’s algorithm can handle in a general
way any configuration of an arterial network, admittedly by a rather large
MILP problem.
If we limit ourselves to maximizing the through-bands along an isolated

two-way artery, we have a considerably more tractable problem. We assume
that the cycle, splits, and travel time ranges between intersections are given.
After we obtain an optimal solution, by shearing the time-space diagram
between pairs of intersections, we can obtain one corresponding to equal
travel times in both directions. Thus, the solution will be symmetric with
respect to the two directions of traffic, although we can subsequently take
some bandwidth from one and give it to the other, if the traffic volumes call
for it.
Little’s formulation starts with the following definitions, where all the
variables are expressed as multiples of the cycle c .
= red phase of intersection i ,
= travel time from i to j
time from the center of the red phase at i to the center of a red
phase at j , where the two red phases are adjacent to a through-band,
and on the same side of the band
time from the right (left) side of the red phase of

intersection i to the green band, where, from symmetry we need only
consider one direction.
From Fig. 5 we obtain the following relationships:
where
In addition, looking at Fig. 4 , we obtain the constraints

114 Chapter 3
The optimization objective can now be defined as follows: Find

b, and which maximize b subject to the constraints expressed
by Eqs. 8 to 10, plus the positivity constraint
The optimization problem is now a MILP problem, with the constraint

that be integer p-providing the “integer” part of the MILP problem.
As in all MILP problems, for any set of the integer values, the problem is
a linear Programming (LP) one. For the MILP problem, Little suggested a
branch and bound approach, which provides a “tree” one searches on the
way toward the solution, such as that shown in Fig. 5. The tree is formed by
considering intersections one at a time, and all the values of the integer
variable m associated with the pair of the intersection at hand and a
reference intersection. For each node of the tree, we solve a LP problem to
obtain a value of the through-band, assuming fixed values of for a
subset of the intersections. The branch and bound technique starts with the
node with the highest through-band and searches along the branches of the
tree, until all the values of have been selected.
The advantage of Little’s method is that it can be extended to consider

various other possibilities regarding light synchronization, such as the
following:
The change of speeds between successive pairs of intersections may be
allowed to vary between two bounds. This is accomplished by
introducing additional variables which must satisfy appropriate
constraints.
The through-bands in the two directions may be made to have any
desired ratio.
The cycle may vary between two bounds.
The method may be adapted for networks of arteries, by introducing
“loop constraints”, which must be satisfied by variables such as the
along closed loops of the network. Little’s suggestion for an objective
function of this problem was a weighted sum of the through-bands along
all the arteries of the network.
Little’s method, which also came to be known as the MAXBAND

method, has been extended by various investigators in order to consider a
variety of new aspect of the problem, such as: time of clearance of existing
queue, left-turn movements, and different bandwidths for each link
(MULTIBAND) for each link of the arterial, as discussed by Gartner (1991),
and Gartner et al ( 1991).
3.3.2 The effect of queues on progression
All the synchronization schemes discussed above have one basic

deficiency, namely, they are valid when the streets have rather light traffic.
As everyone knows, when the traffic density increases, queues of cars bock
the path of following cars through successive green lights, and the
progression scheme becomes meaningless. This situation has been
investigated by Gazis (1965) and is shown in Fig. 7. If there are no queues,
the ideal offset between intersections N and M would be such that the
green phase of intersection M starts d/v seconds after that of intersection
N , where d is the distance of the two intersections, and v the speed of a
moving platoon of cars.
116 Chapter 3
If, however, there is queue of length l stopped at intersection M , then

the onset of the green phase at M induces a “starting wave” that moves
toward the end of the queue with speed reaching the end of the queue
after time If the leader of the platoon starting from intersection N at
the beginning of its green phase is to reach the end of the queue in front just
as its last cars starts moving, then the car must start at time measured
from the start of the green phase at M , where is given by
If the queue stopped at N is released by turning the light green at time

the front of the platoon, traveling at speed v , will reach the end of the
queue at time
Setting we obtain a situation where the interference between

platoons is more or less minimized.
It is seen from Eq. 12 that is positive when l is sufficiently large,

leading to a situation which has been referred to by some as “reverse
progression”. Whatever the appellation, the point is that adjustment must be
made for the existence of a queue, if the upstream platoon is to go through
intersection M without stopping. It may very well be that in extreme
conditions, not all of the platoon cars will clear intersection, but some will
start a new queue, but this is the best that can be done under the
circumstances.
Now, one may suggest a method for finding the relative offset of the two
lights for all traffic situations, from light traffic to heavy traffic, (but not
extreme congestion). When the traffic is light, l is generally zero, and the
relative offset is
the standard progression offset. As l increases, also increases, up to the

point of becoming equal to
In this case l = d , and we have reached the level of full saturation. If,
along an artery, the density builds up gradually from A to B (Fig. 8), and
then decreases toward zero from B to C, then the offset with respect to A
may be given by a curve such as which can be obtained following the
method given above.
118 Chapter 3
It should be pointed out that such a design may only indirectly reduce
delays, by preventing or reducing the blocking of intersections, and the
infamous “gridlock”. Its main effect would be to minimize the number of
starts and stops of vehicles by starting each vehicle only when there is room
enough for it to proceed.
3.3.3 The TRANSYT method
Adjustments for the existence of queues has been done, directly of

indirectly, by many investigators. One of the most notable examples is the
TRANSYT method of Robertson (1969). TRANSYT simulates the motion
of cars through a network, evaluating an objective function which measures
the “inconvenience to drivers”, and then finds a good set of offsets by
searching over the domain of allowable offsets. TRANSYT is probably the
best known, and most frequently used road control strategy. Field
implementation of TRANSYT designs have indicated a possible reduction of
up to 16% in travel times through a network.
Figure 9 shows the basic structure of TRANSYT, as described by

Robertson. The “initial signal settings” include some pre-specified
coordination of lights, the minimum green durations for each intersection,
and the initial choice of splits, offsets, and cycle time, the latter assumed to
be the same for all network intersections. The “network data, flow data”
comprise the network geometry, the saturation flows, the link travel times,
the constant and known rates of turning movements in each intersections,
and some known, constant demands. The “traffic model” consists of nodes
(intersections) and links (connecting streets). Cars move along the links of
the system in platoons, with some platoon dispersion taking place as the
platoon moves along a link. This platoon dispersion is the change in the
spatial distribution of a group of vehicles due to the difference in their
speeds. TRANSYT also allows some deviation from fixed splits. Thus, in
addition to optimizing the system over the allowable ranges of offsets, it tries
to improve the traffic further by varying the allocation of green in each
intersection.
The TRANSYT procedure comprises a model used to calculate delays

and stops, and a “hill-climbing” technique for minimizing an objective
function, D, given by
120 Chapter 3
where is the average delay per hour on the link of a network of N

links, is the average number of stops per second on this link, and K is a
weighting factor.
The model of traffic movement consists of generating a pattern of flow

on network links by manipulating three types of flow patterns for each
direction of flow, namely,
1. The IN pattern, which is the flow of traffic past the downstream end of a
link, if traffic is not impeded by a traffic light.
2. The OUT pattern, which is the actual traffic flow past the intersection
into the next link, and
3. The GO pattern, which is the flow obtained if there is a residual queue
upstream of the intersection during the entire green phase.
The OUT pattern of one intersection generates the IN pattern of the next
downstream intersection, with some allowance of “dispersion”, the
spreading out of platoons due to varying speed between platoon cars. A
typical IN pattern is shown in Fig. 10, in the form of a histogram of flow
during each of 50 equal fractions of the cycle.
It may be seen that the position of the green phase along the time axis of
the IN pattern allows one to compute both the delay and the number of stops
at an intersection, which contribute to the objective function D. A minimum
for D is then obtained by the following “hill-climbing” procedure. The offset
and/or split of a single intersection is changed by a small amount, and a new
value of D is computed. The change of offset and split is continued until a

(local) minimum of D is obtained. The process is repeated by varying the
offset and split of all other intersections in pre-selected sequence, completing
one iteration. The procedure is stopped after several iterations, when the
marginal improvement of the objective function begins to get very small.
3.3.4 The Combination Method
Another method of signal synchronization is the Combination Method,

proposed by Hillier and Lott (1966). The Combination Method assumes a
common cycle and given split at each intersection. In addition, it is assumed
that
1. The amount of traffic at each intersection does not depend on the traffic
light settings, and
2. The delay along a given link of a network depends only on the settings
of the lights at the two ends of the link.
The method is based on the observation that, if the above assumptions are
true, links can be combined in series or in parallel, in computing and
minimizing delay. Consider a pair of intersections A and B (Fig. 11),
connected by two traffic links. Assumption 2 above states that the delay of
traffic along link 1 depends only on the offset between A and B. Let this
delay be given by the function
where is the offset of B with respect to A . The quantity is

defined within the interval (0,c), where c is the cycle, since it is a periodic
function of with period c .
122 Chapter 3
We also define a similar function for the delay along link 2 .

Since it is assumed that the two delays are independent of each other, they
can be added in computing the total delay for the network which includes A
and B . We can then replace the two links by a single one associated with a
delay function
We next consider two links in series connecting three consecutive

intersections A, B, and C. We define the delay functions
We then compute the delay of traffic moving from A to C , as a

function of the offset of the light at C with respect to the light at A ,
This delay depends, of course on the offset of the light B as well. For any
value of we pick the offset at B so as to obtain the minimum delay
from A to C . We do this by defining
where
This allows us to replace the links AB and BC by a single link AC in

computing the total delay for the network.
Combining links in series and in parallel yields a reduction of the

network to one with fewer links. Hillier observed that a network in the form
of a ladder, such as that shown in Fig. 12, could be reduced to a single link,
simplifying the optimization process. Allsop (1968a, 1968b) has shown that
any network cab be “condensed” by the above process into an irreducible
form, in which at least three links meet at each node. For example, the
network shown in Fig. 13 (a) can be condensed into that shown in Fig. 13
(b) by combining the links as follows:
AB and BC in series into AC ,

AC and CD in series into
and AD in parallel into
and DE in series into AE,
EF and FG in series into EG ,
GH and HI in series into GI ,
IJ and JA in series into IA .
In combining links in series, Hillier suggested discretizing the interval

and then searching for the minimum of Eq. 20 at discrete values
of .
124 Chapter 3
The synchronization of lights for an irreducible network such as that of

Fig. 13 (b) was also considered by Allsop (1968a, 1968b). He proposed an
iterative scheme for minimizing the delay in such a network by an
appropriate selection of offsets. The scheme starts with a single link, and
proceeds to form larger and larger sub-networks in successive stages. At
each stage, one estimates, for any combination of offsets at intersections on
the boundary of the sub-network, a set of offsets for the internal
intersections, which minimizes delay. The process proceeds until all links
are included, resulting in the solution to the synchronization problem.
3.3.5 The SIGOP Program
The SIGOP (Signal Optimization) program (SIGOP 1966) is one

developed under the auspices of the U. S. Bureau of Public Roads by the
Traffic Research Corporation. The program tries to minimize, in a least-
square sense, the deviation of the relative offsets of neighboring intersections
from a set of “ideal” offsets, which are stipulated, computed or adjusted on
the basis of real life measurements. It is implicitly assumed that the delay is
a periodic function of the relative offsets, as shown in Fig. 14 (a), with a
minimum at some value between –c / 2 and c / 2, where c is the cycle, or
this value incremented by a multiple of c . It is further assumed that this
ideal relative offset is independent of the offsets of the other intersections.
This delay function may be approximated by a set of parabolas with the
same minimum, as shown in Fig. 14 (b).
This leads to an objective function of the type
where is the ideal relative offset between intersections i and j , is

the design offset of intersection j , is a “link importance factor”,
weighing the traffic from i to j , and is an integer value, usually 0 or
which is selected so that the relationship
126 Chapter 3
is satisfied. Now, the procedure for minimizing D is a combination of a

Monte Carlo random starting point routine, and a direct search procedure. In
effect, we divide the multi-dimensional domain into different regions
corresponding to various choices of the values of which satisfy Eq. 23.

The SIGOP program searches for the global minimum of D within the
domain defined by Eq. 24, by obtaining systematically local minima through
the following sequence of steps:
1. A random choice is made of a which satisfies Eq. 24, and the
corresponding region is determined by finding all the which satisfy
Eq. 23.
2. The minimum of D corresponding to the region determined in step 1 is
found by diferentiating Eq. 22 and solving the resulting system of linear
equations in
3. Step 2 is repeated, changing each by until a local minimum is
obtained. This is so if a further change of any by would yield
a higher value of the minimum obtained through step 2.
Steps 1 to 3 constitute one “Monte Carlo game”. We then start mew

games with step 1 and obtain new values for local minima, until we are
reasonably certain that the global minimum is among all the local minima
obtained through the set of Monte Carlo games. Such a certainty that one has
not missed the global minimum is a mater of experience and judgement,
which is a requirement for proper use of the SIGOP program.
It may be pointed out that an exhaustive search of all the regions of the
domain of Eq. 24, for all possible values of runs into an extraordinary
amount of computations. For example, it has been estimated by Ross (1972)
that a square grid with 25 signalized intersections, all serving two-way
streets, would correspond to about 1028 regions of the domain defined by
Eq. 24.
3.3.6 Additional contributions to the synchronization problem
There have been numerous additional contributions to the study of the

signal synchronization problem. In fact, this subject has been one of the most
investigated topics in traffic theory.
Buckley et al (1966) have contributed methods for maximizing through-

bands for regular rectangular grids of one-way arteries. Bavarez and Newell
(1967) contributed an optimization scheme for through-bands for one-way
arteries, for a variety of objective functions involving both delay to drivers
and number of stops. In effect, Bavarez and Newell start a platoon at the
beginning of an artery, and look for a synchronization scheme that does the
best possible job of serving this platoon. According to their analysis, for any
time dependence of the traffic input at the entrance of the artery, there exists
in general a synchronization scheme better than the one corresponding to a
continuous through-band. The optimum scheme is obtained by “juggling”
the offsets of the intersections near the entrance of the artery, deviating from
the continuous through-band design. Of course, the method assumes exact
knowledge of the time dependence of the traffic input into the artery, which
is not always available. Gartner (1972a) used arguments of graph theory to
prove theorems concerning the optimal through-band design for a network.
The theorems involve relationships between design parameters such as
offsets, and are likened by Gartner to Kirchhoff’s laws for electrical
networks.
Two other contributions, by Okutani (1972) and by Gartner (1972b) use

dynamic programming for optimizing the synchronization of traffic signals,
by minimizing delay an/or number of stops.
Okutani assumes that the delay between pairs of intersections depends

only on the relative offsets of these intersections . He then uses dynamic
programming to obtain the optimum synchronization, by dividing the
network into sections, which are treated as stages in a multi-stage decision
process. He applied Bellman’s principle of optimality to obtain a recursion
relationship between the optimum design for n sections, and for n + 1
sections. Okutani also suggested an alternative approach to the problem
based on Pontryagin’s maximum principle (see Katz 1962).
Gartner considers portions of a network as stages in a multi-stage decision

process. Each stage is small enough, so that an exhaustive search on a small
number of independent variables is possible in optimizing the stage. The
global optimization consists of minimizing a suitable objective function
associated with the delay and/or the number of stops of the traffic. The
multi-stage process is optimized using dynamic programming. Gartner also
suggests a procedure, based on graph-theoretical concepts, which tends to
minimize the required computational effort.
128 Chapter 3
The preceding methods for optimizing the synchronization of lights in a

network have been tested, to some extent, in real-life situations. These tests
have shown that schemes, which take into account the formation of queues at
intersections are distinctly better than those that simply maximize a through-
band, insofar as minimizing delay is concerned. The margin of improvement
over simple through-band maximization was found to be as high as 15%, for
moderately heavy traffic, but not oversaturated conditions. In addition,
taking into account the existence of a queue at the beginning of the green
phase tends to mitigate the danger of the familiar “gridlock” produced by
cars entering into an intersection, without being able to clear the intersection
before the cross-traffic green starts. In general, designs that take into account
of the existence of queues tend to improve performance during periods of
moderate to heavy traffic, but not congestion.
3.4 TRAFFIC RESPONSIVE OPERATION OF TRAFFIC

LIGHTS
The operation of traffic lights evolved from fixed-time operation to “traffic

actuated” operation, based on actual real-time measurements of traffic along
competing directions. Typically, such actuation involved extending the green
phase, up to a maximum, to accommodate an increasing demand of a given
direction. When two competing directions showed signs of traffic demand
increasing toward congestion, both directions would get the maximum green.
This is desirable in conditions of heavy traffic, because it minimized the
effect of the lost time for clearance of the intersection before switching.
When computers were introduced for traffic control, in the 1960s, many of
these heuristic schemes found heir way into the code which generated the
computer decisions regarding the settings of the traffic lights. However, the
schemes did not necessarily involve a thorough analysis of the performance
of systems guided by them, because the state of the art of such an analysis
was not up to the task. Several contributions to the state of the art have been
made since that time, and will be discussed in the sequel.
3.4.1 Single Intersection
One of the first contributions in the area of traffic responsive signals for a
single intersection was made by Dunne and Potts (1964), who proposed their
own heuristic algorithm for traffic actuation, and also gave an interesting
method for representing and analyzing the operation of traffic lights. The
representation consists of describing the condition of the intersection in the
“state space” where and are the sizes of queues waiting to be

served, along the two competing directions of the intersection, as shown in
Fig. 15. Assuming that the variables are continuous, the condition of the
intersection will be represented by a continuous line. If, in addition, the
traffic inputs are assumed to be constant and continuous, and the flow rates
during the green phase equal to the saturation flows, then the condition of
the intersection is represented by a point moving along one of three possible
straight lines:
1. When the light favors direction 1, the state of the intersection,

moves along the direction of the vector where are the
input rates and the saturation flows along the two competing
directions.
2. When the light favors direction 2, the state of the intersection moves
along the vector
3. During the period of clearance which includes the yellow phase of the
light, the state of the intersection moves along the direction of the vector
as both queues are increasing.
In cases 1 and 2, it is assumed that there is a queue of cars waiting to be

served at the start of the green phase, for both directions. When the queue in
the favored direction is fully served, the corresponding axis acts as a
barrier for the state of the intersection, and the state of the intersection
130 Chapter 3
moves along this barrier until the light is changed, with only the not favored
queue increasing.
Dunne and Potts went on to propose and test an algorithm which treats the
lines and in Fig. 15 as “reflecting barriers”, meaning that
the light changes as soon as the state of the intersection reaches either one of
these lines attempting to cross. Incidentally when the reflecting barriers are
just the axes the Dunne-Potts algorithm treats the case known as the
Saturation Flow Algorithm, in which the light changes only when the
favored queue is exhausted, which may be impractical in many real-life
situations.
The Dunne-Potts algorithm yields a trajectory in state space such as the

line in Fig. 15. Under appropriate conditions, it may yield a
“limit cycle”, such as that represented by the line EFGH in Fig. 15. The
algorithm proceeds as follows:
The algorithm is stable, that is, it yields bounded values of when

the intersection is undersaturated, i.e. when
In that case, a limit cycle exists, with phases and given by
where
and l is the lost time for clearance at phase change-over. The duration of the
limit cycle is
The above results concerning the limit cycle are applicable, whether or
not the portions of the reflecting barriers coincide with the axes
However, Dunne and Potts point out that the delay to users is minimal when
the barriers coincide with axes which is the case of the saturation
flow algorithm.
Another contribution to the optimization of the operation of a traffic-

actuated traffic light was that of Grafton and Newell (1967). They defined
the saturation flow operation of the traffic light as the basis policy, and they
compared it with two other comparison policies. One of them allowed
switching of a traffic light before the favored queue was empty, and the
other allowed an extension of the green phase past the time when the favored
queue became empty. Comparing the basis policy with the comparison
policies led to the conclusion that there were domains in the space where
the basis policy was optimal , and other domains in which the comparison
policies were optimal. Grafton and Newell then applied Bellman’s Principle
of Optimality (Bellman and Dreyfus, 1962) to complete the solution. In this
case Bellman’s principle expressed the “recursive property” of the optimal
solution, namely, that whatever the initial state of the system and the initial
decisions up to an intermediate time the remaining decisions after time
should lead to an optimal policy with respect to the state resulting from the
first decisions.
Another interesting contribution in this area is that of van Zijverden and

Kwakernaak (1969), who gave a general approach for a good sub-optimal
control of a complex intersection serving time-varying demands. They
applied principles of optimization of stochastic control systems given by
Kwakernaak (1966), and obtained an optimal solution through an exhaustive
search from among all possible sequence of light settings over a short time
horizon, and repetition of this process thereafter.
132 Chapter 3
3.4.2 Systems of Intersections
A number of coordinated Traffic-responsive strategies for networks of

arteries have been developed, and frequently applied to various systems
around the world. One of the first schemes for a fully adaptive control of a
large number of intersections was given by Miller (1965). His algorithm
involves making a binary decision at time intervals h, for every traffic light,
whether to leave the light in its present state or change it. The decision is
based on the computation of the delay for the two options.
As an example, let us assume that at a N-S, E-W intersection the traffic

light is green in the N-S direction. If we keep the light in this state for an
extra h seconds, the reduction of delay for the N-S traffic can be found to
be
where
a is the duration of the amber (yellow) phase,

is the lost time for clearance between favoring the N-S or E-W
direction,
are the number of vehicles expected to move in the N and S
directions during h seconds,
are the arrival rates of vehicles per h seconds in the N and S
directions, and
are the corresponding saturation flows in the two directions.
If we defer the start of the E-W green phase by h seconds, we impose

some additional delay to the vehicles already there, and those that may arrive
before the queues in the E-W directions are exhausted, when we do give
them the green light. This additional delay is estimated to be
where are the vehicles already queued in the W and E directions,

are the corresponding arrival rates, and are the smaller
integers such that
where are the saturation flows in the W, E directions per h seconds,

and are the lost times in these directions. Equations 31 determine the
number of intervals of h seconds after the beginning of the green phase, during
which there is still a residual queue. Combining Eqs. 30 and 31, we find that, if
we extend the N-S green phase by h seconds, the net reduction of delay to all
vehicles around the intersection is
If it pays to leave the signal favoring the N-S traffic for h ,

repeating the process thereafter, as long as the maximum green phase allowable
is not exceeded.
3.4.3 Additional contributions to Synchronization of Intersection

Networks
A somewhat similar approach to Miller’s was used by Ross et al (1970), for a

small group of intersection around a “critical intersection”, which is most
likely to cause the highest delays. Both Miller and Ross et al tested their
strategies by simulation, and found them to be beneficial, particularly in
accommodating large fluctuations of traffic such as those caused by
temporary interruption of flow. Improvements to both approaches could be
134 Chapter 3
made by using the familiar adaptive control technique of a Rolling Horizon,

namely, predicting for the next, say, five periods and applying the solution
for one period.
A control strategy involving small adjustments about a progression

design, which do not extend from one cycle into the next, was tested in
Glasgow, as discussed by Holroyd and Hillier (1969). The strategy, known
as EQUISAT, consisted of equalizing saturation flows along the two
competing directions of an intersection, leading to the relationship
where and are the traffic input rates, saturation flows

and green phases, respectively.
Another strategy was proposed by Koshi (1972) for the city of Tokyo, and
it is one of the most adaptive ones proposed for a relatively large traffic
system. In general, for large systems, the general approach has been to
separate control of intersections into “Macrocontrol” and “Microcontrol”.
Macrocontrol comprises an overall strategy for the entire system, and
microcontrol involves small changes in this strategy to accommodate
detectable departures from expected demand at some “critical intersections”.
In effect, Koshi introduces microcontrol over the entire system, but he does
it in a way that generates a continuously changing grand strategy for the
entire system. The optimization is performed over the time period of one
cycle. Koshi looks at every intersection and tests to see whether or not a
small advance or delay of the next traffic light phase may decrease the
aggregate delay to the users. If it does, the change in the offset is
implemented, resulting in a continuous application of a “policy
improvement” scheme, which is responsive to changing traffic conditions. It
should be mentioned here that a continuous change from one grand strategy
to another has distinct advantages over an abrupt one, which may cause
problems in transition from one to another.
A number of other programs have been contributed over the years for
improving the control of traffic lights in a network. The onset of the use of
computers for real-time traffic control, starting in the 1960s, has made the
application of such contributions to real systems possible. Among these
contributions are the following:
a) The SCOOT Program
The SCOOT program was first developed by Hunt et al (1982) and has been
extended in several respects by others. It may be viewed as the traffic-
responsive version of TRANSYT, which was discussed earlier. It has been
used in many cities around the world, particularly in the United Kingdom.
SCOOT, like other systems, utilizes measurements of traffic volume and

“occupancy”, which is the percentage of the time a sensor detects a passing
vehicle, and hence is related to the density. The measurements are taken at
the entrance of a link. SCOOT runs on a central computer, and uses an
approacch similar to that of TRANSYT for minimizing delays. The main
difference is that, instead of assuming historically available fixed traffic
inputs like TRANSYT, SCOOT uses real-time measurements of traffic, and
is run repeatedly in real time to investigate the effects of incremental
changes of splits, offsets, and cycle time. After defining a “performance
index” related to delays, SCOOT selects the changes which would result in
an improvement of the performance index, and runs the traffic lights
accordingly. Results of field evaluations have been discussed by Luk (1984).
b) The OPAC, PRODYN, CRONOS, and COP Programs
Advanced traffic-responsive programs for networks include OPAC

(Gartner, 1983), PRODYN (Farges et al, (1983), CRONOS (Boillot et al,
1992) and COP (Sen and Head, 1997). These strategies do not consider
explicitly splits, offsets, or cycles of traffic lights. On the basis of some pre-
specified staging, they try to find, in real-time, the optimal specification of
the next few switching times, over a future time horizon H,
starting from the current time t and the currently applied stage. In order to
obtain the optimal switching times, these methods solve dynamic
optimization problems using a variety of realistic, dynamic traffic models,
which include some binary variables reflecting the impact of red/green
phases on traffic flow and the associated performance index. Some
constraints, such as constraints on maximum and minimum splits, are also
included. The performance index which is to be minimized is the total time
spent by all vehicles during the time horizon H.
Recognizing that it is difficult to predict traffic accurately for long

periods, these programs use the concept of a Rolling Horizon , H , but they
apply the results of optimization to a much shorter period, after which new
136 Chapter 3
data are collected and the optimization program is solved again for a time
horizon H.
One basic problem with these methods is the introduction of binary

variables, which require exponential-complexity algorithms for global
optimization. Specifically, OPAC employs complete enumeration of choices
corresponding to integer switching times, while PRODYN and COP employ
dynamic programming. Because of the exponential complexity of these
methods, the resulting solutions, although conceptually applicable to an
entire network, cannot be easily applied to more than one intersection.
Therefore, the optimization is generally limited to optimizing single
intersections, whose settings may be coordinated by a superimposed
heuristic control strategy, as discussed by Kessaki et al (1990). CRONOS
differs from the other methods, in that it employs a heuristic global
optimization method with polynomial complexity, which permits
consideration of several intersections, although again intersections are
optimized individually.
3.4.4 The Onset of Oversaturation
Oversaturated systems are discussed in detail in the following section.

Here, we shall make some comments regarding the behavior, and possible
improvements, of a single artery whose traffic movement may be controlled
by a series of traffic lights, as discussed by Gazis (1965). It is well known
that such arteries, with traffic light operation designed for progression, fail
miserably during periods of congestion. The reason is shown in Fig. 16,
where, for convenience, traffic characteristics are drawn as continuous lines
rather than ones with step function changes. If the net input at some
intersections, i.e. influx minus outflow of traffic, is positive along an
appreciable length of the artery, then the demand along the artery may
increase above capacity levels over a section of the artery. In Fig. 16, the
demand is shown as a continuous curve f(x). The demand along each point
is the integral of this curve, F(x). If
where Q is the maximum throughput of the artery, then congestion will set in
along a major portion of the artery. Progression becomes meaningless in this
case, but for lack of a better design it is not uncommon in many cities to
have the lights synchronized for progression anyway – to the frustration of
the traffic engineers and driving public.
Congestion may be avoided by controlling the net input along the artery,
and may take one of several forms:
1. A barrel-shaped green through band may be tried, with a bulge in the

potentially congested region. Such a band would favor the arterial flow
of traffic and also limit the inputs from the side streets. The bulge of the
138 Chapter 3
through band might be made traffic-responsive, changing with traffic

demands.
2. One may limit the influx past the point A, making the net input negative
past this point, as shown by the input curve and the
corresponding demand curve in Fig. 16.
3. One may scale all net inputs down to the curve with a
corresponding demand curve so that the maximum of
does not exceed the capacity of the artery.
Options 2 and 3 are akin to the “entrance ramp control” which is

increasingly applied to freeways to prevent their congestion. All of the above
alternatives assume, of course, that there is enough parking space in the side
streets, or alternate route choices for the traffic that is blocked from entering
the artery. In any case, the optimum solution involving constraining the
traffic entering the artery can be obtained on the basis of historical or real
time data.
3.5 OVERSATURATED SYSTEMS
In the preceding sections, we considered intersections, and systems of

intersections, in which traffic demands are not so high that they cause
excessive queueing at any point. However, every major city of today, and
even some minor ones, suffer from the experience of “oversaturation” of
certain intersections, i.e. the inability of these intersections to serve out all
the queues during a green phase, resulting in ever increasing sizes of queues.
Here, when we talk of queues, we understand that their position may not be
limited a single road segment, but may extend over several of them. In what
follows, we discuss systems which become oversaturated during a “rush
period”, but eventually serve out the built-up queues after the end of the rush
period. We begin with the simplest case of a single, relatively isolated,
intersection.
3.5.1 A Single Oversaturated Intersection
Let us consider an isolated intersection serving n competing traffic

demands. Vehicles arrive at rates ( j = 1,2,....,n), and some or all of
them are discharged during the green phase favoring their movement. It is
assumed that the discharge is at the maximum service rate, or “saturation
flow”, When the demand is low, the queues developed during the red
phase are completely served during the green phase, near the end of which
newly arriving vehicles may go through the intersection without stopping.
The average green time required for serving all the cars arriving during a
light cycle c is
where is the average arrival rate of vehicles during the light cycle. As
discussed in the preceding section, the standard practice accepted by traffic
engineers and managers has been to divide the “total available green time” in
proportion to the values of Eq. 35, leading to the same “degree of
saturation” along all competing directions. Accordingly, the effective green
phases satisfy the relationships
As was shown in the preceding section, in the case of an undersaturated

intersection, the rule of thumb shown in Eq. 36 comes very close to
optimizing the operation of the traffic light, in the sense of minimizing
delays to users.
Let us now consider, as an example, an intersection involving two

competing streams, and assume that the demand at the intersection increases
until, after some point, one or both of the queues cannot be completely
served during the green phases. This happens when the increase, so that
where L is the lost time for acceleration and clearance. Equation 37 defines
the case of oversaturation, discussed by Gazis and Potts (1965). The
allocation of the green may still be done in this case according to Eq. 36, but
140 Chapter 3
it is meaningless and very unlikely to minimize delays to the users. In this

case, queues build up during the rush period, and are only served completely
after the end of this period, provided that the end of one rush period does not
merge with the beginning of the next one! Optimization of the operation of
the light at an oversaturated intersection must then involve, not the
minimization of the delay per cycle, but the minimization of the “aggregate
delay to both streams, during the entire rush period.
The situation is shown in Fig. 17, where the cumulative arrival, Q, and
cumulative service, G, are plotted versus time, for one of the competing
streams. The quantities Q and G are the integrals over time of the arrival
rate and service rate, respectively, namely
where is the service rate, and the time origin is taken as the onset of
oversaturation. The service curve is actually more like the sawtooth curve of
Fig. 17, but it will e drawn as a smooth curve during our discussion, for
simplicity.
The area between the curves Q and G is a measure of the aggregate

delay to the users of the intersection during the rush period. Their vertical
distance at any time t is the effective size of the queue at that time, and their
horizontal distance at any point Q, such as at is a measure of
delay to an individual vehicle arriving at time The service rates must
cannot be altogether arbitrary, but they must satisfy certain constraints,
which depend on the fact that the green phase for each direction must lie
between a minimum and a maximum. If and are the bounds of the
green phases, then the service rates satisfy the constraints
The optimization of the operation of the traffic light may now be

formulated as an “optimal control” problem, with the split of the green phase
among the two directions being the control variables. On the basis of Eq. 39,
this split may vary between an upper and a lower bound, thus defining a
“control region”. Any acceptable service strategy must involve time-
dependent which satisfy Eq. 39 plus the equation
where the equality applies when sufficiently long queues exist in both
directions to supply saturation flow during the entire length of the green
phases.
The objective function of the control problem is the aggregate delay to the
users, and the statement of the problem is the following:
142 Chapter 3
Minimize the delay function
subject to the conditions in Eqs. 39 and 40 for the values of which

determine the The time T is defined by the equation
and it is assumed that after time T the capacity of the intersection exceeds
the demand in both directions.
The solution to the control problem, following Gazis (1964), proceeds as

follows, (see Fig. 18):
There exists a time T that is the earliest possible end of oversaturation. It

can be determined by assuming a single setting for the traffic light during the
entire period which corresponds to constant values of such
that both queues are exhausted together at time T. Denoting these constant
as we find that they can be determined, together with T, from the
system of equations
144 Chapter 3
In Eq. 43, and Fig. 16, the are known functions of t, based on
historical data of rush period traffic inputs for the intersection. The single
setting strategy is not the only strategy which exhausts both queues at the
same time, neither does it constitute the control which minimizes the
aggregate delay. It can be seen in Fig. 18 that we can construct multistage
service curves, which also exhaust both queues at the same time, by
increasing the delay in one direction and decreasing it in the other. The dash-
dot lines in Fig. 18 show such a strategy, in which we give a little more
green to direction 1 up to time and take some away from until T, so
that queue 1 is exhausted at time T. This decreases the delay to stream one,
at the expense of an increase in delay for stream 2, but it can be proved
easily that queue 2 will also be exhausted at time T.
Let us assume that at time the queue 1 is decreases, with respect to

the single-setting strategy, by the amount q. Queue 2 will be increased at
the same time by the amount Assuming, without loss of
generality, that we find that we have reduced the aggregate delay
by the amount
We can continue trading off delay, in a way that favors the stream with
the greater saturation flow, until we reach the optimum strategy, which is a
two-stage operation. During the first stage, stream 1 receives the maximum
green phase, and stream 2 receives the minimum one, and during the second
stage the allocation of the green is reversed. The switch-over point is given
by
with T and satisfying Eqs. 43. Incidentally, the switching from one
extreme of the control variable to the other is generally referred to as bang-
bang control , a term somewhat less than reassuring when used in
connection with traffic.
It can be seen from the preceding discussion that the exact shape of the
cumulative demand curves during most of the rush period does not
change the solution, or the minimization result, as long as the cumulative
demand curves do not dip below the cumulative service curves, as shown in
the top curve of Fig. 18. If the cumulative demand curves can be
approximated reasonably well by some asymptotic straight lines near the end
of the rush period, we can obtain simple expressions for T ,and for the
switch-over time For example, if
then
where
The total delay to the users may be further reduced, if we prolong a little
bit the agony of the drivers in stream 2, with The ultimately
optimum setting is obtained when the durations of the last straight line
portions of the two service curves, satisfy the relationships, (Fig. 18),
146 Chapter 3
The above relationships regarding the optimal control of the intersection

may be obtained by graphical differentiation, i.e. by proving that a very
small change, positive or negative, in the time length together with
the associated small change in produces zero gain in reducing the
total delay.
Finally, we address the problem of finding the optimum light cycle, c .

The light cycle affects the total throughput of the intersection, because of the
lost time L in each cycle, which may be assumed to be constant for any
reasonable range of c . The ratio L/c decreases with increasing c, thus
increasing the total throughput, and reducing the value of T. However, the
additional delay due to intermittent service, i.e. the saw-tooth area in Fig. 17,
increases with increasing c. The problem is to find the value of c
corresponding to a stationary value for the total delay. This can generally be
done for given functions and a given sequence of settings for the
traffic light. Figure 19 shown this derivation for a single setting of the traffic
light, and cumulative demand curves approximated by straight lines near the
end of the rush period, namely,
The appropriate single setting corresponds to green phases, and

service rates satisfying the relationships
Now, let
Solving for and T, we obtain
The saw-tooth area, per cycle, is equal to

148 Chapter 3
hence the total additional delay from 0 to T is
The total delay over the entire rush period is
or
where is the (constant) difference between the trapezoidal, cumulative

demand curve, such as OAT* of Fig. 19, and the actual demand curve,
which is not plotted in that figure. Finally,
The optimum cycle is obtained by differentiation, i.e.,
The general case involves a fair amount of algebraic manipulation. Here,

we give as an example the derivations for the simple case when
in which case, the optimum cycle turns out to be
For example, if and L, = 3 sec., we obtain
T= 80 min., and
Of course, such a choice for c is acceptable only if the range of the cycle
includes the derived optimum cycle of 2.3 min.
Let us briefly discuss the meaning of the above derivation of optimal

control of the oversaturated intersection. Our derivation shows that, during
the early part of the rush period, the direction with the higher saturation
should be favored to the highest possible degree. At some point in the midst
of the rush period, service should switch to maximum for the direction with
the lower saturation flow. If the timing is selected properly, both queues will
be served out at the same time, and with the minimum aggregate delay to
the users. Incidentally, this solution is far from obvious, or intuitive. This
author was told by a British traffic expert that “in England we actually favor
the poor devil on the side street who competes for service against a big
stream of traffic – sort of, rooting for the underdog”. It turns out that such
compassion, or even a democratization of traffic control equalizing the rate
of discharge of queues, fails to optimize the system. The reason can be seen
from the schematic diagrams of Fig. 20. The intersection, viewed as a
system, receives a traffic demand rate higher that the attainable service rate
during the rush period. This means that the cumulative service curve lies
150 Chapter 3
below the cumulative demand curve during the rush period, ad the area
between these two curves is a measure of the aggregate delay to the users.
The best way of reducing this aggregate delay is by maximizing the service
rate as much as possible during the early pert of the rush period, with proper
adjustments during the last part of the rush period for the purpose of ending
congestion simultaneously in both directions. The operative words here are
that we must get those cars out of here as soon as possible. It will be seen
later on in this chapter that this principle applied to even more complex
oversaturated systems.
A more rigorous mathematical derivation
So far, we have limited our approach to manipulating the details of a

graph, in effect performing graphical differentiations and the like. A more
complete solution, given by Gazis (1964), is cast in the framework of
Control Theory, and uses the Pontryagin “maximum principle”, (see
Pontryagin et al, 1962).
Let us refer to Fig. 21, which shows the cumulative demand and service
curves versus time, for two competing demands. The queue sizes and
are the state variables of the system, and they satisfy the differential
equations
where u is a control variable defined by
We seek to minimize the aggregate delay to the users, given by
subject to the end conditions
and the constraint that the control variable u be within the admissible
control domain, i.e.
This is a problem in linear control theory, and the linearity leads to a

bang-bang solution. Following Pontryagin, we see that Equations 62 and 64
are equivalent to
152 Chapter 3
where
We now define the adjoint variables satisfying the relationships

The system of Eqs. 68 can be solved, yielding
where and are constants. We now define the Hamiltonian of the

system, H, given by
where
According to Pontryagin’s maximum principle, the optimum control

maximizes the Hamiltonian, subject to appropriate end conditions. From Eq.
71, we see that the Hamiltonian H is maximized if u is equal to its upper
bound when z is positive, and equal to its lower bound when z is
negative. Using the signum function of z, (sg z = sign of z), we can write the
optimum u in the form
154 Chapter 3
The remainder of the computations toward the optimum solution are

straightforward, leading to the expressions already given for the switch-over
time and the duration of the rush period, T.
3.5.2 Complex Oversaturated Systems
The preceding discussion of the optimal control of a single oversaturated

intersection makes sense if this intersection is essentially alone in causing
congestion, and the rest of the traffic network is essentially undersaturated. It
could also be valid if other oversaturated points in the network are relatively
far from the intersection in question, thus being uncoupled from it. Now, let
us consider some systems, some involving neighboring intersections, and
seek their optimal control.
a) A Pair of Intersections
When two oversaturated intersections are close together, they are coupled
through the number of vehicles that pass through both intersections in
sequence. Gazis42 considered the system of two intersections serving the
traffic streams 1, 2, and 3, where the stream 1 is assumed to go through the
pair of intersections (1, 2) and (1, 3) without turning, (Fig. 22). Let us
assume that the saturation flows in the three directions satisfy the
relationships
If we optimize the two intersections independently, following the

preceding discussion of the optimization of the single intersection, we will
conclude that stream 1 must receive preferential treatment in both
intersections. However, if we take the point of view that streams 2 and 3
are like traffic streams in a single traffic stream moving along a boulevard,
then stream (2 + 3) should receive the preferential treatment, on the basis of
the last of Eqs. 73. It turns out that this second point of view is the correct
one, because in the first one we essentially count the delay to stream 1
twice, once at intersection (1, 2), and then again at intersection (2, 3). The
correct solution proceeds as follows:
Let us assume that the transit time and queueing storage between the two
intersections are negligible relative to the other retained variables of
comparable nature, and that the two intersections become oversaturated
roughly at the same time. We can again obtain an optimal policy of
allocation of green phases by applying the same policy improvement scheme
used for the single intersection, namely, trading off delays of the various
streams if they reduce the aggregate delay. First, we obtain the earliest end
of the rush period for each one of the two intersections, viewed as isolated.
Let these ends be the times and for intersections (1, 2) and (1, 3),
respectively. We also compute the switch-over times corresponding to
optimizing the operation of these two intersections as isolated ones. These
switch-over times are given by
If the maximum and minimum service rates for stream 1 are the same at
both intersections, the relative magnitude of is the same as that of
We now obtain the complete solution to the problem, shown in Fig 23,
as follows:
Since streams 2 plus 3 must receive preferential treatment at the start of

the rush period, we find the green allocation favoring these streams, for each
of the intersections treated as isolated ones. The corresponding service for
intersection (1, 2) is the optimum, because any delay trade-off reducing the
delay for stream 1 increases the delay to the combined streams 2 and 3 by
q greater amount. However, after the switch-over time a delay trade-off
156 Chapter 3
is possible between streams 1 and 3. The maximum trade-off corresponds

to a reduction of the delay to stream 1 by the area of the parallelogram
(ABCD), with a corresponding increase of the delay for stream 3 by the area
(abcd), which is smaller than (ABCD) since we have assumed that
The key to the optimal solution is that we can favor stream 1 when it
competes with only one of the streams 2 and 3, but we must favor the
combination of streams 2 and 3 if they compete with stream 1
simultaneously.
As a consequence, the optimum control of the pair of intersections

involves, in general, three switch-over times, one at intersection (1, 2), and
two at intersection (1, 3). It is tacitly assumed that there is enough parking
space along the direction 1 between the two intersections to accommodate
the queue before intersection 3. Otherwise, yet another adjustment to the
strategy is called for, which will not be discussed here. A complete
discussion of various aspects of the problem is given by Gazis (1964).
The preceding discussion in Sections 3.5.1 and 3.5.2 defines the

operation of network elements (intersections), and entire networks of
intersections, which are Store-and-Forward Networks. In such networks,
during the rush period, traffic queues up at some or all of the intersections,
which can be described as Throughput Limiting Points (TLPs). Optimizing

the operation of the traffic lights of the network comprises a selection of a
sequence of green phase allocation in each intersection, which minimizes the
aggregate delay to the users. The examples given in the preceding section
show that the optimal strategy involves transitions from one apex of the
control domain to another. We will now consider some additional systems.
b) Three Intersections Forming a Triangle
The optimal control of a system of intersections forming a triangle was

also considered by Gazis (1964) and shown in Fig. 24. Let us assume that
traffic moves along the directions 1 to 3 without turning movements, and
the cumulative demands are given by
where t is the time, given in hours. Let us assume that the saturation flows
are and vehicles per hour. We also assume
that the maximum green phase is 60 s and the minimum 20 s, with 10 s
allowed per cycle for clearing the intersection at phase changes.
158 Chapter 3
The admissible control domain is found to be one bounded by the

hexahedral surface OABCD shown in Fig. 25.
In Fig. 25, the variables are the fractions of the light cycle allocated to
the three traffic streams. It is assumed that a common cycle is used at all
intersections, hence
where is the usable green phase for the stream, and L is the 10 s time
allowed for clearance at phase change. For example, the apex A of Fig. 25
corresponds to and The usable green phase for
streams 2 and 3 is applied where these two streams intersect stream 1, and
they produce streams which do not constitute congestion at their intersection,
since the sum of the two greens is smaller that the available green at that
intersection. Thus, the specific allocation of the green at that intersection has
no influence on the performance of the intersection.
It may be easily ascertained that during the rush period at least one of the
green phases must be equal to the maximum, in order to decrease the effect
of the lost time for clearance. This means that only the apexes A, B, C, and
D need be considered in seeking an optimum solution. Table 1 shows the
service rates for the three streams, for each choice among the four apexes,
listed in order of decreasing total throughput of the system.
The traffic light settings corresponding to the four apexes are assumed to
be maintained for times and and it is required that all
three queues be served out simultaneously at time
We have seen from the preceding examples that the optimum control
policy must produce a concave curve for the cumulative service. The same
principle applies here, and so we are led to a procedure for finding the
optimum policy, which can be formulated as a linear programming (LP)
problem, the following:
Minimize
160 Chapter 3
subject to
where
and the unknowns must be non-negative. The solution to this LP problem,

for the specific values given above, is
Furthermore, to ensure concavity of the cumulative service curve, the

sequence of these time segments must be in decreasing order of the
corresponding service rates, namely
A key observation should be made at this point. At the beginning of the

rush period, favoring any one of the streams, usually the one associated with
the highest saturation flow does not optimize the operation of the system,
because it “starves” the downstream intersection of the unfavored streams. It
can be seen from the above solution that the best policy starts with equal
splits for all streams. Yet, as of the writing of this book, if we drive around
many congested areas around the world, we are likely to find that the local
traffic authorities set traffic lights at such triangle systems in proportion to
the saturation flows, being unaware of the consequence of such a policy in
the case of oversaturated systems.
It may be noted that the LP formulation can be used without knowledge of

the exact form of the admissible control domain, and hence the possible
number of distinct control periods, and corresponding service rates. The time
axis may simply be divided into equal segments, and the resulting queue
sizes on various links can be used as the unknowns of the LP problem, with
the objective of minimizing aggregate delay to the users of the system. The
solution to this problem is a suboptimal one, approaching the optimal
solution as the size of the time segments is decreased. This approach will be
used later on in this section for the optimization of a complex transportation
network.
c) Spill-back from an Exit Ramp of an Expressway
Let us consider the system shown in Fig. 26, comprising an expressway

1, an exit ramp 2, and the exit highway. The exit highway may also be an
expressway, in which case the exit ramp may lead into an acceleration lane.
Here, we are interested in exploring a situation in which the rate of discharge
from the exit ramp into the exit highway is controlled. We consider the case
discussed by Gazis (1965), in which the exit highway 3 is a major artery
but not an expressway, and the intersection of 2 and 3 is controlled by a
traffic light, a very common situation in many urban areas. If 3 is an
expressway, and input into it is controlled through “entrance ramp control”,
the discussion that follows is applicable without major changes.
If the intersection of 2 and 3 becomes oversaturated, then a queue

builds up along the exit ramp. Because of the limited parking capacity of this
ramp, the queue spills into the expressway, often with devastating results.
Not only does the spilled queue block one of the expressway lanes, but often,
drivers from the neighboring lane attempt to enter the exiting stream,
slowing down and obstructing the movement of the through traffic. It is not
unusual to have a three-lane expressway degenerate into barely a one-lane
roadway, causing inordinately large tie-ups upstream of the exit ramp. Cases
such as this one abound in many metropolitan areas as of the time of the
writing of this book. The question is, can one minimize the impact of this
phenomenon by properly operating the traffic light 4 , in a way that
minimizes the delay to the combined streams 1, 2, and 3?
162 Chapter 3
The cumulative demand curves and as well as the

maximum and minimum service rates for all streams, are shown in Fig. 27.
The service rates for streams 2 and 3 are determined directly from the
split of the traffic light, with being the maximum, and the minimum
service rate. The expressway is assumed to have a throughput equal to
when unobstructed, and when clogged due to spill-back. More often than
not, the saturation flows and are such that
Therefore, according to the discussion given earlier concerning an

isolated, oversaturated intersection, the optimum operation of the trafficlight
4 would involve service curves such as and with the highway
stream 3 receiving preferential treatment. However, in the present case the
ramp stream 2 is coupled with the stream 1 along the expressway through
the spillback phenomenon. If the storage capacity of the ramp is then
spillback will take place from time to time where these times are
determined by drawing the curve
and finding the intersection of this curve with
We can see from Fig. 27 that if we choose the service curves

and we can eliminate spillback altogether. The delay to stream
3 is increased by a quantity the delay to stream 2 is decreased by
and the delay which would have been caused by spillback is completely
164 Chapter 3
eliminated. It should be pointed out here that such a situation is not always
possible in many situations of this type, largely because of a very limited
capacity Let us continue with the assumption that it is possible. Since
the total change in the aggregate delay will be
If is positive, we have a net reduction of the aggregate delay, and

therefore it pays to adopt a strategy of keeping the queue along the ramp
below the critical value if at all possible. If is negative, then
spillback is not as damaging as it appears, at least in terms of total delay,
which is minimized by an optimum operation of the light 4, assumed
isolated. However, it may still be desirable to prevent, or at least limit the
congestion on the expressway for safety reasons, which may override delay
considerations. If this is the case, one may accept a small negative delay
trade-off
If the delay criterion is the dominant one, we find that a critical constant
rate of demand along the expressway exists, which is related to
according to a relationship obtained by setting in Eq. 84 equal to zero,
leading to the relationship
where
Solving Eq. 85 for we obtain
If the demand rate is smaller than then spillback is the lesser of two
evils, since it corresponds to minimum aggregate delay. By similar
arguments we may investigate the possibility of allowing spillback during a
portion of the interval If the rate of demand along the expressway
falls sufficiently below it may be profitable, in terms of delay, to adopt
a strategy such as that shown by the dashed line in the middle diagram of
Fig. 27, and the complementary service curves for streams 1 (not shown in
the figure), and 3. Such a policy will introduce some additional delay to the
stream 2, and because of spillback it will also delay some vehicles along the
expressway, but it will reduce the delays to stream 3. The net change can be
computed by an expression similar to Eq. 84, and the policy may be adopted
if it is found to be indeed profitable. Finally, if the demand rate along the
expressway falls below it is always profitable to allow spillback.
A few remarks are in order at this point. One key observation is that
intersection 4 should not, in principle, be operated as an isolated one. Many
such intersections are operated as isolated, resulting in unduly large delays
along the expressway. Indeed, in most cases, the demand along the
expressway is so great that it pays to try and eliminate spillback for as long
as possible. Another observation is that the storage capacity of the exit ramp
is a very crucial quantity. Improvements to the system should probably place
a high priority to increasing this capacity; for example, by building a long
extra lane leading to the exit ramp, and advising drivers that they can exit
only from that lane, thus preventing the blockage of the neighboring lanes by
exiting vehicles.
3.6 OVERSATURATED STORE-AND-FORWARD

NETWORKS
All the preceding examples of oversaturated systems may be viewed as

special cases of an oversaturated store-and-forward network, such as that
shown in Fig. 28. In such a network, traffic does not move freely along
certain paths, because its movement is obstructed by the presence of
166 Chapter 3
Throughput Limiting Points (TLPs) . The most common nature of TLPs is

that they are network nodes in which the output capability is lower than the
sum of inputs at that node. As a result, queues are formed upstream of a
TLP, which are not dissipated until the end of the rush period. In Fig. 28, the
nodes C, D, E, F, and G are all TLPs, causing the formation of upstream
queues on some or all the arcs converging at those nodes. It may be pointed
out that all the oversaturated systems discussed in the preceding sections can
be viewed as store-and-forward networks. For example, the single
intersection can be viewed as the store-and forward network shown in Fig.
29, with B being the TLP causing the queue formation along the two
competing directions.
It should be mentioned at this point that the existence of TLPs is more

common than one might think. Many studies of networks involve
assumptions of network performance based on phenomenological flow
versus concentration relationships. In such studies, it is customary to simply
compute an effective speed of traffic as a function of the density along a
network arc. The point made here is that, more often than not, this (high)
density is the result of the existence of a TLP, causing queueing. As a result,
the store-and-forward treatment of an urban network is not just one of many
choices, but it is a necessity, since most urban networks become store-and-
forward networks, even before heavy congestion sets in.
For the purpose of treating the optimization of the management of a store-

and-forward network, we assume the following:
1. There exist some time-dependent, possibly stochastic, origin-destination

requirements of traffic volumes, which cannot be fully accommodated
during the rush period, because of the existence of TLPs.
2. With each arc of a network are associated three parameters
A (fixed) travel time, (or travel “cost”),
A capacity restraint, i.e. the maximum possible volume of traffic along
the arc per unit of time, and
A storage capability allowing traffic to be stored just before a node at
the end of an arc, and be dispatched when roadway space becomes
available.
168 Chapter 3
3. Traffic moves from node to node at the constant speed corresponding to

the arc connecting the nodes, and then either moves past the node, or is
stored there.
With respect to item 3 above, it is of course true that queued traffic is

stored along the arc before the constricting node. However, for the purpose
of optimization, it can be considered as stored in a “bin” right at the node,
and then served on a FIFO (First-In-First-Out) basis.
In optimizing the operation of a store-and-forward network, the primary

objective is to minimize an aggregate delay function, possibly weighted for
priority assignments, subject to operational and capacity constraints. The
optimization will certainly comprise finding the optimum switching at the
nodes, allocating passage through the nodes to various traffic streams. For
even better results, route assignment for individual vehicles, or platoons of
vehicles, could decrease delays even further. To see the possible beneficial
effect of route assignment, let us look at Fig. 30, which shows an
particularly simple case of a network of one way streets, whose intersections
are controlled by traffic lights.
For simplicity, let us assume that all streets are essentially identical and
carry about the same amount of traffic, which is pretty close to saturation,
therefore the corresponding traffic streams require the entire length of he
green phase in order to be served without queueing extending to the next
cycle. With equal demands in both directions, the network would operate
optimally with equal green phases in all directions.
Let us now assume that at some time a certain portion of northbound

traffic along one of the streets, corresponding to a flow rate q wants to turn
right and go eastbound at some intersection. If we are to accommodate this
turning movement, we must take away a certain portion of the green phase
from (n + 1) north/south streams intersected by the right-turning stream.
This would result in a loss of overall throughput of the network equal to (n
+ 1) q. If, however, we split the stream in two, and merge it again at the
street, we gain back half of the lost capacity, or qn/2. The author has
frequently referred to this strategy as one using “The FLF (Five Loaves,
Few Fish) Algorithm”. This somewhat irreverent biblical reference to the
miracle of feeding a multitude with five loaves and a few fish is used to
dramatize the effect of stretching the capability of scarce resources, which in
this case is accomplished through route assignment. The point here is that
such route assignment increases the overall throughput of the system. If we
look at Fig. 20, which may be viewed as a generic figure for all congested
systems, we see that increasing the aggregate throughput of the system can
only decrease the aggregate delay to the users.
The solution to the complete problem, involving both switching at the

nodes and route assignment, is rather cumbersome and will not be given
here. Here, we shall assume that the route assignment for all traffic is given,
and we shall seek only optimum switching, which regulates the queue
discharge rates at various nodes. This switching problem can be cast in the
framework of control theory. The state variables are the queues before the
nodes, which have originated from upstream nodes including the input
nodes, and they must be dispatched optimally toward their destination. The
rate of change of these queues depends on the traffic input rates into the
network, and on some control variables which represent the switching at the
nodes. The objective function is the aggregate delay to the users, which is a
function of the state variables. Optimization is achieved if we minimize this
objective function, subject to appropriate constraints on the control variables.
The mathematical formulation is the following. Minimize
given that
170 Chapter 3
subject to some end conditions corresponding to some given values for

(0) and (T), and also subject to the constraints
In the above equations, are the state variables, are weighting

factors, the (given) traffic inputs, with known destinations and route
assignments, the control variables, U the control domain, and some
functionals which may possibly involve not only the instantaneous values of
and but also their history. If we now discretize the time and
replace the time derivatives by differences, and the integral of Eq. 88 by a
sum, we are led to a mathematical programming problem, the following.
Minimize
given that
and subject to the end conditions on and plus the constraints

on the control variables
It is reasonable to assume that the functionals in Eq. (92) are linear in

leading to an Linear Programming (LP) problem of optimization. The
optimal solution to this problem minimizes delays, assuming the given route
assignments to traffic. One possible approach to obtaining the global optimal
solution, including route assignments, is to make small variations to route
assignments, trying to bypass the points of maximum congestion, and solve
the LP problem repeatedly until we obtain a reasonable approximation to the
global optimum.
A number of additional contributions to the control optimization of store-

and-forward networks have been made over the years, including those of
D’ Ans and Gazis (1976), Singh and Tamura (1974), Michalopoulos and
Stephanopoulos (1977a, 1977b), Lim et al (1981), Davison and Ozguner
(1983), Park et al (1984), Rathi (1988), Kim and Bell (1992), and Diakaki et
al (1999). However, application of the approach has been limited in practice,
probably because many theoreticians and practitioners do not realize that all
traffic networks are of the store-and-forward variety, as soon as even one
TLP appears, let alone many of them. Many theoretical contributions
continue to use phenomenological relationships to describe flow behind such
TLPs, which is at best a rather pointless exercise. Links upstream of a TLP
do not miraculously set themselves to accommodate heavy densities by
obeying these phenomenological relationships. They simply pile-up traffic
into queues, which are served at the discharge rate supplied at the TLP.
Modeling of the details of flow upstream of the TLP only imposes
computational burdens, without making any contribution to the solution of
the problem. Papageorgiou (1999) cites one application of a variant of the
store-and-forward approach known as TUC, to which he and his colleagues
have contributed. The application took place in Glasgow, Scotland, with
very successful results.
3.7 FREEWAY CONTROL
Freeways frequently become congested, and even attain a store-and-

forward status. Congestion may be caused by something happening along the
freeway, it may be the result of excessive inputs from entrance ramps, or it
may be the result of spillback from an exit ramp discussed in Section 3.5.2c .
An example of something happening along the freeway that may cause
congestion is the case of the “Moving and Phantom Bottleneck” discussed in
Chapter 1 . Such a bottleneck may be caused by a truck unable to negotiate a
steep uphill portion of the freeway, which in this case does not have an extra
“slow traffic” lane. A very special example of such an effect will be
discussed in the following section, for the case of the Lincoln tunnel of New
York City.
Excessive input from an entrance ramp has been addressed by metering

the ramp input and restraining it, in order to prevent freeway congestion.
Ramp metering may be fixed-time, adjusted for expected traffic conditions at
particular times of the day, based on historical data. Obviously, fixed-time
ramp metering may miss the mark, since traffic may deviate appreciably
from historical data. However, any reduction of input of traffic into the
freeway is likely to reduce its congestion, albeit without preventing it
completely.
172 Chapter 3
A better approach to ramp metering is through closed-loop control based

on real-time measurements of traffic along the freeway. Many such schemes
have been suggested and tested, generally showing good results.
Papageorgiou (1999) gives a thorough discussion of such schemes, including
some advanced methods like the ALINEA strategy suggested by
Papageorgiou et al (1991).
Generally, all these control strategies do not account for what happens
behind the ramp entrance. If the entrance is from surface streets, it is tacitly
assumed that it does not matter what happens to them, as long as one gets
some improvements for the freeway operation. This is not the case, however,
if the ramp input comes from another freeway. In Section 3.5.2-c, we
discussed the disastrous effects on the supplying freeway, which an
obstruction of the ramp outflow such as ramp metering can cause. Clearly,
the best solution for ramp metering should consider the entire system, and
not an isolated point of the freeway. Here are some examples of what can go
wrong with consideration of isolated input ramps:
The freeway may already be obstructed downstream from the input

ramp, e.g. by another input ramp. In this case ramp metering does not
help the freeway, and only penalizes the supplying roadways.
The ramp input may be getting its supply from another freeway, which
may end up being jammed because of spillback, as discussed above.
We conclude that, more often than not, the freeway with its ramps and
possibly other freeways constitute yet another store-and-forward system, and
they should be treated as such. The optimization procedure should consider
not just a portion of the freeway in the vicinity of the entrance ramp, but the
entire freeway and its traffic-feeding partners, be they surface streets or other
freeways.
3.8 SYSTEMS AFFECTED BY GEOMETRIC DETAILS
As mentioned already, moving and phantom bottlenecks can cause

congestion at any point on a surface street or a freeway, thus becoming, in
effect, temporary TPLs, albeit moving ones. Particularly notable example
are the tunnels under the Hudson River, connecting New York City with
New Jersey, which have received a great deal of attention by traffic theorists,
from the early days of Traffic Theory. Examples are the contributions cited
in references 57-63.
We shall discuss here the Lincoln tunnel, which demonstrates well the
specific nature of the problem, and which was the subject of greater
experimentation with control measures aiming at improving the efficiency of
its operation. The key feature of this tunnel causing congestion is also found
in other roadways, making the tunnel experience relevant to congestion
control in general. Also common to all traffic systems is the challenge of
measuring traffic densities, for which key contributions originated in the
course of experiments aiming at improving the tunnel performance.
The Lincoln tunnel has a special geometry shown in Fig. 31. There are
three distinct sections of the tunnel, a downgrade section, a level section, and
an upgrade section. Generally, it is the upgrade portion that causes problems,
as soon as demand increases into levels of congestion.
As long as traffic densities inside the tunnel are low, traffic moves
relatively freely, at flow rates approaching 1400 cars per lane per hour.
However, congestion sets in as soon as demand increases and many cars pile
into the tunnel. Congestion generally starts at the foot of the upgrade portion,
point 3, as reported by Edie and Foote (1958). There, some cars, particularly
trucks, can get into trouble. If they have to slow down just at point 3, they
have difficulty accelerating, opening up a large gap in front and squandering
throughput. They, in effect, become “moving bottlenecks” of a very special
type, restricting traffic movements only in their lane, since lane changing is
forbidden in the tunnel. The phenomenon is not unlike the “acceleration
hysteresis” phenomenon described in Chapter 1, caused by the asymmetry
between acceleration and deceleration of cars. The result of the moving
bottleneck in the case of the tunnel is that its throughput is suddenly reduced
to 1200 cars per lane per hour, or even less than that.
The cause of congestion was recognized for many years, and several
attempts were made to mitigate it, including the work of Greenberg and
Daou (1960), Herman and Potts (1961), Edie et al (1963), Crowley and
174 Chapter 3
Greenberg (1965), and Foote (1968). In all these attempts, an effort was
made to prevent congestion by limiting the input rate into the tunnel, often
citing the paradoxical objective of “increasing throughput by decreasing
input”. The rule of thumb, which evolved during these attempts, was that
limiting the input rate to 22 cars per minute, or 1320 cars per hour, could
maintain relatively free flow and prevent congestion. The rate of 1320 cars
per hour was below the maximum observed at the tunnel, but it often did
work in preventing congestion for a few minutes. Unfortunately, this “open
loop” control failed if, by chance a truck found itself at the foot of the
upgrade at low speed and caused the beginning of congestion, after which
the limited input became too high to handle. Eventually, “closed loop”
control was implemented, first using some fixed-logic control hardware,
(Foote and Crowley, 1967), and then an on-line digital computer, (Gazis and
Foote, 1969). We discuss here this last experiment of computerized
surveillance and control of the tunnel, which achieved the most reliable
results, while also contributing to such areas as the measurement of traffic
densities in sections of a roadway.
3.8.1 Density Estimation
Before any traffic control is implemented, it is imperative to know where

the cars are. Thus, good surveillance is essential for the success of any traffic
control measures, particularly for congested systems. It certainly is a
prerequisite for the proper control of a store-and-forward system.
Surveillance at the Lincoln tunnel certainly qualified as excellent, even
though it used obsolescent technology by today’s standards.
As shown in Fig. 31, four observation points, also referred to as “traps”,

were installed along the tunnel, at the entrance, foot of the downgrade, foot
of the upgrade, and exit. The observation was done by photocells detecting
the passage of a car over each one of the detectors of the trap, with a time
resolution of about 10 msec. Typical profiles obtained by the photocell
detectors are shown in Fig. 32, where the “car with variable number of
holes” was generally a truck, with gaps seen by one but not both detectors,
and the “VW pulling a ladder” was a truck pulling an empty platform. These
last two signals were smoothed out logically to produce pulses similar to
those of an ordinary car. The four times shown in the figure were
sufficient to provide a determination of the length, speed and (constant)
acceleration of a vehicle passing over the trap. Speed and length could be
obtained with great accuracy, and acceleration with a lesser one, but the
acceleration was not used in the subsequent computations.
The speed of vehicles, together with the count of vehicles per unit of time,
were used to determine the flow parameters at the observation point. The
average speed is one of these parameters. Density, equal to the number of
cars in a section divided by its length, could be obtained by initializing the
count of cars somehow, and then using the formula
where is the number of cars in section ij at the end of the

“observation period”, and and are the cars passing over the
traps ii and jj during the observation period. Observation periods
were equal segments of time, small enough to allow timely updating of the
flow characteristics.
Equation 94 requires some initialization of the system with the count of

vehicles in a section at time zero. After such an initialization, the correct
count should be obtainable by this equation. Unfortunately, errors in the
counts and/or would degrade the accuracy of this count.
For this reason, correct updating of the count was necessary, periodically.
Both, initialization and updating were done by identifying individual
vehicles as they passed, first over the entrance trap of the section and then
over the exit trap.
The lengths of vehicles were used to identify individual vehicles, or

combinations of vehicles, such as a short vehicle following a long vehicle.
When these patterns were repeated at the exit trap, the count of vehicles
176 Chapter 3
entering the tunnel section between the times of entrance and exit of the
pattern used was the exact number of vehicles in the section at the time of
exit of the pattern. The reader will recognize that in present days, with the
availability of in-vehicle devices such as the RFID (Radio Frequency ID)
used for electronic toll payment, the detection of individual cars is a very
simple matter.
One may ask how densities can be estimated if identification of individual

vehicles is not available. The Lincoln tunnel experiment provided an answer
to this question. Having the accurate counts of vehicles at the Lincoln tunnel
made possible the development and testing of other advanced, density
estimation methods, such as the “Kalman Filtering” method, which is
discussed in an appendix of this book.
3.8.2 The Control Algorithm
The detectors in the tunnel yielded the speeds and throughputs over the
traps, and the counts of vehicles between the traps. On the basis of these
variables, a control function could be developed, which would in general
depend on the time series of these variables. Only one control action was
available, namely, a restriction of traffic input into the tunnel. Initially, a
binary decision was made between imposing or not imposing a restriction.
Later, a four-level control possibility was implemented, refining the response
to the measurements of speeds and densities over the entire tunnel
3.8.2.1 Binary Control, Single-Density Algorithm
The binary control decision was based only on the measurements of

speed, and traffic count, in the middle section of the tunnel, 23,
(Fig. 31). The rationale was that it was in this section that the “acceleration
hysteresis” initiated, by slowing trucks to almost a stop at the foot of the
upgrade. “Observations” of the traffic were done at 5 sec. Intervals, and a
control value was determined for the observation interval, with a
value equal to 0 or 1, the zero corresponding to absence, and the one to
imposition of control. The control consisted of restricting the input of traffic
into the tunnel in order to establish free flow, if possible. The control
algorithm was the following:
If fps, set
Otherwise, is given by
where H is the step function defined by
and A,B are constants allowing a variable threshold density for the
imposition and removal of traffic input control. The values of A and B most
frequently used were 25 and 1.
The single density, binary control algorithm worked some of the time, but
often failed to anticipate congestion in time. In addition, the binary nature
led to over-reaction or under-reaction at times. Or this reason, a multi-level
control algorithm, based on the measurements at all three sections of the
tunnel, was developed and implemented.
3.8.2.2 The Three-Density Algorithm
It is to be expected, and confirmed by observation, that congestion in any

part of the tunnel can degrade its performance sooner or later. Congestion
was generally defined as the case when speeds were lower than 20 mph, and
densities higher than 55 veh/mile. A more accurate description of the effect
of congestion in tunnel sections was developed through observation, and led
to the following algorithm.
Let the state of the tunnel be described by three state variables, the counts
and Then, the probability that congestion will set in within a
short period, say 30 sec., increases with increasing values of any or all the
The control algorithm was based on the tacit assumption that the
instantaneous values of were the primary determinants of impending
congestion. In addition, four possible levels of control were implemented,
corresponding to integer values for the control variable C ranging from 0
to 4. The value 0 corresponded to no control, and the values 1 to 4
corresponded to increasing restraints on entering traffic, with a red light
corresponding to the value 4.
178 Chapter 3
A table was compiled in state space in state space, which preceded

congestion by about 30 sec., while the tunnel was under no control, or under
single-density control. A smooth surface was fitted through these “critical
density” points, leading to the following algorithm:
If and re measured in fps and set C = 0,

Otherwise, compute
and set
where and are constants, H is the step function given in Eq. 96,
and µ is a history-dependent variable, taking the values 0 and 1, which
was used to generate the double-ladder function shown in Fig. 33. The
meaning of the double-ladder control is that used in many “governor-type”
control devices, namely, imposing or increasing control at some critical
value of the state variables, but removing or reducing control at lower values
of these variables, to avoid “chattering” of the control process. Thus, the
value of µ was taken as 0 in computing ascending values of C, and 1 in
computing descending values.
Application of the preceding control algorithm yielded a consistent

improvement of the operation of the tunnel, with the average throughput
increased by about 12%, compared to the uncontrolled state. This result

provided proof that prevention of the “acceleration hysteresis ” on roadways
can be very beneficial. It may be worth applying on highways prone to such
a hysteresis phenomenon, for example, a highway with a sudden steep uphill
portion and no extra lane for slow-moving vehicles.
180 Chapter 3
References
Allsop, R. E. (1968a), “Choice of offsets in Linking Traffic Signals”, Traffic Engr. &
Control, 10, 73-75.
Allsop, R. E. (1968b) “Selection of Offsets to Minimize Delay to Trafic in a Network
Controlled by Fixed-Time Signals”, Transport. Sci., 2, 1-13,
Bavarez, E. and Newell, G. F. (1967), “Traffic Signal Synchronization on a One-Way
Street”, Transport. Sci., 1, 55-73.
Bellman, R. and Dreyfus, S. E. (1962), Applied Dynamic Programming, Princeton, NJ:
Princeton Univercity Press.
Boillot, F., Blossville, J. M, Lesort, J. B., Motyka, V., Papageorgiou, M. and Sellam, S.
(1992), “Optimal Signal Control of Traffic Networks”, 6th Intern. Conf. On Road Traffic
Monitoring and Control, London, England, pp. 75-79.
Buckley, D. J., Hackett, L. G., Keuneman, D. J. K. and Beranek, L. A. (1966), “Optimum
Timing for Coordinated Traffic Signals”, Proc. Aust. Road Res.Board, 3, Pt. 1, 334-353.
Crowley, K. W. and Greenberg, H. (1965), “Holland Tunnel Study Aids Efficient Increase
of Tube's Use”, Traffic Eng., 35, 20.
D' Ans, G. C. and Gazis, D. C. (1976), “Optimal Control of Oversaturated Store-and-
Forward Transportation Networks”, Transportation Science, 10, 1-19.
Davison, E. J. and Ozguner , U. (1983), “Decentralized Control of Traffic Networks”,
IEEE Trans. on Automatic Control, 28, 677- 688.
Diakaki, C., Papageorgiou, M. and McLean, T. (1999), “Application and Evaluation of the
Integrated Traffic-Responsive Urban Corridor Control Strategy”, 78th Annual Meeting of
Transp. Res. Board, Washington, DC, Paper No. 990310.
Dunne, M. C. and Potts, R. B. (1964), “Algorithm for Traffic Control”, Operations Res.,
12, 870-881.
Edie, L. C. and Foote, R. S. (1958), “Traffic Flow in Tunnels”, Highway Res. Board Proc.,
37, 334-344.
Edie, L. C., Foote, R. S., Herman, R. and Rothery, R. W. (1963), “Analysis of Single-Lane
Traffic Flow”, Traffic Eng., 33, 21-27.
Evans, H. K., Editor, (1950), Traffic Engineering Handbook, Institute of Traffic
Engineers, New Haven, CT, p. 224.
Farges, J-L., Henry, J- J. and Tufal, J. (1983), “The PRODYN Real-Time Algorithm”, 4th
Intern. Symp. on Transp. Systems, Baden Baden, Germany, pp. 307-312.
Foote, R. S. and Crowley, K. W. (1967), “Developing Density Controls for improved
Traffic Operations”, Highway Res. Board Record, 154, No. 1430.
Foote, R. S. (1968), “Instruments for Road Transportation”, High Speed Ground Transp.
Journal.
Gartner, N. H. (1972a), “Constraining Relations among Offsets in Synchronized Signal
Networks”, Letter to the Editor, Transport. Sci., 6, 88-93.
Gartner, N. (1972b), “Optimal Synchronization of Traffic Signal Networks by Dynamic

Programming”, Proc. 5th Intern. Symp. on the Theory of Traffic Flow and Transportation,
edited by G. F. Newell, New York: Elsevier, pp. 281-295.
Gartner, N. H. (1983), “OPAC: A Demand-Responsive Strategy for Traffic Signal
Control”, Transp. Res. Record, 906, 75-84.
Gartner, N. H. (1991), “Road Traffic Control: Progression Methods”, in Concise
Encyclopedia of Traffic and Transportation Systems, edited by M. Papageorgiou, Pergamon
Press, Oxford, UK, 391-396.
Gartner, N. H., Assmann, S. F., Lasaga, F. and Horn, D. L. (1991), “A Multiband
Approach to Arterial Traffic Signal Synchronization”, Transp. Res. 25B, 55-74.
Gazis, D. C. (1964), “Optimum Control of a System of Oversaturated Intersections”,
Gazis, D. C. (1965a), “Traffic Control, Time-Space Diagrams and Networks”, Traffic
Control – Theory and Instrumentation, edited by T. R. Horton, New York: Plenum, pp. 47-63.
Gazis, (1965b), “Spillback from an Exit Ramp of an Expressway”, Highway Research
Record, 89, 39-46.
Gazis, D. C. and Potts, R. B. (1965), “The Over-Saturated Intersection”, Proc. 2nd Intern.
Symp. on the Theory of Road Traffic Flow, edited by J. Almond, Paris: Organization for
Economic Cooperation and Development, pp. 221-337.
Gazis, D. C. and Bermant, O. (1966), “Dynamic Traffic Control Systems and the San Jose
Experiment”, Proc. 8th Intern. Study Week in Traffic Engineering, Barcelona, Spain, V.
Gazis, D. C. and Foote, R. S. (1969), “Surveillance and Control of Tunnel Traffic by an
On-Line Digital Computer”, Transp. Science, 3, 255-275.
Grafton, R. B. and Newell, G. F. (1967), “Optimal Policies for the Control of an
Undersaturated Intersection”, Proc. 3rd Intern. Symp. on Traffic Theory, edited by L. C. Edie,
R. Herman, and R. W. Rothery, New York: Elsevier, pp.239-257.
Greenberg, H. and Daou, A. (1960), “The Control of Traffic Flow to Increase the Flow ”,
Operations Research, 8, 524-532.
Herman, R. and Potts, R. B. (1961), “Single-Lane Traffic Theory and Experiment”, Theory
of Traffic Flow, edited by R. Herman, Amsterdam: Elsevier Publishing Company, pp. 120-
146.
Hillier, J. A. and Lott, R. S. (1966), “A Method of Linking Traffic Signals to Minimize
Delay”, Proc. 8th Intern. Study Week in Traffic Engineering, Barcelona, Spain, V.
Holroyd, J. and Hillier, J. A. (1969), “Area Traffic Control in Glasgow - A summary of
Results from Four Control Schemes ”, Traffic Eng. & Control, 11, 220-223.
Hunt, P. B., Robertson, D. L. and Bretherton, R. D. (1982), “The SCOOT on-line traffic
signal optimization technique”, Traffic Eng. & Control, 23, 190-192.
Kane, J. N. (1964), Famous First Facts, New York: H. W. Wilson, 3rd ed., p. 626.
Katz, S. (1962), “A Discrete Version of Pontryagin's Maximum Principle”, J. Electron.
Control, 13, No. 2, 8.
Kessaki, A., Farges, J-L. and Henry, J-J. (1990), “Upper Level for Real Time Urban
Traffic Control Systems”, 11th IFAC World Congress, Tallinn, Estonia, Vol. 10, pp.226-229.
182 Chapter 3
Kim, K.-J., and Bell, M. G. H. (1992), “Development of an Integrated Traffic Control

Strategy for Both Urban Signalized and Motorway Networks”, Proc. 1st Meeting of the EURO
Working Group on Urban Traffic and Transportation, Landshut, Germany.
Koshi, M. (1972), “On-Line Feedback Control of Offsets for Area Control of Traffic”,
Proc. 5th Intern. Symp. on the Theory of Traffic Flow & Transportation, edited by G. F.
Newell, New York: Elsevier, pp.269-280.
Kwakernaak, H. (1966), “On line Optimization of Stochastic Control Systems”, Proc. 3rd
Intern. Federation of Automatic Control, London.
Lira, L H., Hwang, S. H., Suh, I. H. and Bien, Z. (1981), “Hierarchical Optimal Control of
Oversaturated Urban Networks, Intern. Journal of Control, 33, 727-737.
Little, J. D. C. (1964), “The synchronisation of Traffic Signals by Mixed-Integer Linear
Programming”, Operations Res., 14, 568-594.
Luk, J. Y. K. (1984), “Two Traffic-Responsive Area Traffic Control Methods: SCAT and
SCOOT”, Traffic Eng. & Control, 25, 14-22.
Marsh, B. W. (1927), “Traffic Control”, Ann. Amer. Soc. Political Social Sci., 133, 90-
113.
Matson, T. M., Smith, W. S. and Hurd, F. W. (1955), Traffic Engineering, McGraw-Hill
Book Company, Inc., New York.
Michalopoulos, P. G. and Stephanopoulos, G. (1977a), “Oversaturated Signal Systems
with Queue Length Constraints–1: Single Intersection”, Transp. Research, 11, 413-421.
Michalopoulos, P. G. and Stephanopoulos, G. (1977b), “Oversaturated Signal Systems
with Queue Length Constraints–2: System of Intersections”, Transp. Research, 11, 423-428.
Miller, A. J. (1965), “A Computer Control System for Traffic Networks”, Proc. 2nd Symp.
on Traffic Theory, edited by J. Almond, Paris: Organization for Economic Cooperation and
Development, pp.200-220.
Okutani, I. (1972), “Synchronization of Traffic Signals in a Network for Loss Minimizing
Offsets”, Proc. 5th Intern. Symp. on the Theory of Traffic Flow and Transportation, edited by
G. F. Newell, New York: Elsevier, pp. 297-312.
Papageorgiou, M., Haj-Salem, H. and Blossevil, J.-M. (1991), “ALINEA: A Local
Feedback Control Law for On-Ramp Metering”, Transp. Res. Record, 1320, 58-64.
Papageorgiou, M. (1999), “Traffic Control”, in Handbook of Transportation Science,
Boston: Kluwer Academic Publishers, pp. 233-267.
Park, E. S., Lim, J. H., Suh, I. H. and Bien, Z. (1984), “Hierarchical Optimal Control of
Urban Traffic Networks”, Intern. Journal of Control, 40, 813-829.
Pontryagin, L. S., Boltyanskii, V. G., Gamkrelidze, R. V. and Mishchenko, E. F. (1962),
The Mathematical Theory of Optimal Processes, translated by K. N. Trirogoff, New York:
Wiley-Interscience.
Rathi, A. K. (1988), “A Control Scheme for High Density Sectors”, Transp. Research,
22B, 81-101.
Robertson, D. I. (1969), “ ‘TRANSYT' Method for Area Traffic Control”, Traffic Engr. &
Control, 11, 276-281.
Ross, D. W., Sandys, R. C. and Schlaefli, J. L. (1970), A Computer Control Scheme for
Critical-Intersection Control in an Urban Network, Menlo Park, CA, Stanford Research
Institute.
Ross, D. W. (1972), “Traffic Control and Highway Networks”, Networks, 2, 97-123.
San Jose Traffic Control Project – Final Report, (1966), San Jose, CA: IBM Corp., Data
Processing report.
Sen, S. and Head, L. (1997), “Controlled Optimization of Phases at an Intersection”,
Transp. Science, 31, 5-17.
“SIGOP”, Traffic Research Corporation, New York, 1966. Distributed by Clearinghouse
for Federal Scientific and Technical Information, PB 17 37 38.
Singh, M. G. and Tamura, H. (1974), “Modelling and Hierarchical Optimisation of
Oversaturated Urban Traffic Networks”, Intern. Journal of Control, 20, 269-280.
Van Zijverden, J. D. and Kwakernaak, H. (1969), “A New Approach to Traffic-Actuated
Computer Control of Intersections”, Proc. 4th Intern. Symp. on Traffic Theory, edited by W.
Leutzbach and P. Baron, Bonn: Bundesminister fur Verkehr, pp. 113-117.
Webster, F. V. (1958), “Traffic Signal Settings”, Road Research Technical Paper No. 39,
Road Research Laboratory, England.
Additional references
Allsop, R. B. (1971), “SIGSET: A Computer Program for Calculating Traffic Capacity of

Signal-Controlled Road Junctions”, Traffic Eng. And Control, 12, 58-60.
Allsop, R. B. (1976), “SIGCAP: A Computer Program for Assessing the Traffic Capacity
of Signal-Controlled Road Junctions”, Traffic Eng. And Control, 17, 338-341.
Boillot, F., Blosseville, J. M., Lesort, J. B., Motyka, V., Papageorgiou, M. and Sellam, S.
(1992), “Optimal Signal Control of Traffic Networks”, 6th IEE Intern. Conf. On Road Traffic
Monitoring and Control, London, England, pp.75-79.
Elloumi, N., Haj-Salem, H. and Papageorgiou, M. (1996), “Integrated Control of Traffic
Corridors – Application of an LP Methodology”, 4th Meeting of the EURO Working Group on
Transportation Systems, New Castle, UK.
Farges, J-L, Henry, J-J. and J. Tufal, J. (1983), “The PRODYN Real-Time Traffic
Algorithm”, 4th IFAC Symp. on Transp. Systems, Baden Baden, W. Germany, pp. 307-312.
Gervais, E. F. (1964), “Optimization of Freeway Traffic by Ramp Control”, Highway Res.
Rec., 59, 104.
Haj-Salem, H. and Papageorgiou, M. (1995), “Ramp Metering Impact on Urban Corridor
Traffic – Field Results”, Transp. Res., 29A, 303-319.
Hillier, J. A. (1966), “Appendix to Glasgow’s Experiment in Area Traffic Control’, Traffic
Engr. & Control, 7.
May, A. D. (1964), “Experimentation with Manual and Automatic Ramp Control”,
Highway Res. Rec., 59, 9-38.
184 Chapter 3
Messner, A. and Papageorgiou, M. (1995), “Automatic Control Methods Applied to

Freeway Network Traffic”, Automatica, 30, 691-702.
Miller, A. J. (1963), “A Computer Control System for Traffic Networks”, Proc. 2nd Intern.
Symp. on Traffic Theory, London, UK, pp.200-220.
Moreno-Banos, J. C., Papageorgiou, M. and Schaeffner, C. (1993), “Integrated Optimal
Flow Control in Traffic Networks”, Europ. J. of Oper. Res., 71, 317-323.
Papageorgiou, M. (1983), Applicaton of Automatic Control Concepts to Traffic Flow
Modeling and Control, New York: Springer Verlag.
Papageorgiou, M. (1998), “Automatic Control Metods in Traffic and Transportation”, in
Operations Research and Decision Aid Methodologies in Traffic and Transportation
Management, P. Toint, M. Labbe, K. Tanczos, and G. Laporte Editors, New York: Springer
Verlag, pp. 46-83.
Wattleworth, J. A. and Berry, D. S. (1965), “Peak-Period Control of a Freeway System –
Some Theoretical Investigations”, Highway Res. Rec., 89, 1-25.
Chapter 4
Traffic Generation, Distribution, and Assignment
One of the key prerequisites for proper management of transportation

systems, including both transportation planning and traffic control, is
knowing where traffic appears, and why it appears. This leads to the need for
understanding how traffic is “generated”, how it is “distributed” from a set
of origins to a set of destinations, and how it is “assigned” along given
routes.
Traffic generation is the result of the existence of people in some parts of

a network, and the existence of attractions, such as jobs, in another part.
Knowing the numbers of people and attractions allows one to determine how
traffic is generated, and how it is distributed into an “origin-destination”
table for the network. The generation and distribution of traffic may be
estimated on the basis of historical data, or it may be based on probabilistic
estimates of certain attractions, such as special events or sales. The traffic
assignment issue concerns the way that these origin-destination needs are
met by selecting the routes to be taken by individual drivers.
The investigation of all three topics has borrowed from Graph Theory of
Flows in Networks, (see Ford and Fulkerson, 1962), which has been
developed over the years for the investigation of networks of all types, and
in particular communication networks. In what follows, we shall discuss the
network representation of a transportation system, and the methodology for
investigating the traffic generation, distribution, and assignment for such
systems.
186 Chapter 4
4.1 NETWORK REPRESENTATION OF A

TRANSPORTATION SYSTEM
An obvious example of a transportation network is an urban street grid

such as that shown in Fig. 1. In Graph Theory nomenclature, the intersections
of such a network are the nodes, and the street segments between nodes are
the links of the network.
If we wish to account for turning movements as well, the network around

the intersections will include additional links. For example, the network
around intersection 4 of Fig. 1 will be altered to one shown in Fig. 2, with
node 4 replaced by the nodes 16 to 19. The throughput of the links in
networks such as those in Figs. 1 and 2 are selected in such a way as to
include the effects of traffic lights at the various intersection nodes.
4. Traffic Generation, Distribution, and Assignment 187
Networks of intersections such as those of Figs. 1 and 2 may be

occasionally used for the study of regions of critical intersections. However,
for a more global transportation planning, it is customary to decrease the
granularity of the network representation.
An urban region is divided into a few sectors, then into smaller districts,
and finally into zones, such as the 15 zones shown in Fig. 3. Traffic is
viewed as moving from one zone to another in the network. It is customary
to view traffic as emanating from a centroid of the node, or reaching the
centroid if originating from another centroid. All discussions of traffic
generation, distribution and assignment refer to movements of traffic
between the centroids plus the entering and exiting odes of the network. The
centroids are connected either to the nodes of the network, or to some mid-
points along the links. The network itself may either be a geometric
representation of actual roadway segments and their intersections, or it may
be the result of an aggregation of roadway segments within zones into a
network representing traffic movements from zone to zone. The Spider Web
network, in which nodes are the centroids, and links connect adjacent
188 Chapter 4
centroids, is of the latter type, in which the links do not correspond to actual
road segments, but represent desired movements of traffic between zones. If
the network representation incorporates actual roadway sections, then the
properties of these sections can be taken into account explicitly in estimating
such things as travel times between nodes, or between centroids. If a spider
web network is used, then travel times along its links are estimates of actual
travel times between zones using all available roadway choices. In any case,
a complete description of a link would provide travel time and/or cost as
functions of the link occupancy, rate of flow, time, and other pertinent
factors. A node represents the terminus of one or more links, and may be
used either to join the termini of a number of links, or to represent a source
(origin) or sink (destination) of traffic. In Fig. 3, the numbered nodes are the
centroids, and the unnumbered ones are nodes of the roadway network.
Let the nodes of the network be labeled 1, 2, 3, ......,N . It is assumed

that all links are directed, meaning that they can only be traversed in one
direction. A directed link from node i to node j is represented by (i,j). If a
network segment between nodes ii and jj carries traffic in both directions, it
is represented by two links, (i,j) and (j,i), as shown in Fig. 4, preserving the
directed graph quality of the network.
The preceding labeling strategy allows one to describe the graph fully by
listing all the links of the graph. For example, the network of Fig. 4 is
described as comprising the links (1,2), (2,1), (2,3), (3,2), (2,5), (3,4), (3,6),
(6,3), (4,5), ad (5,6). A chain from node i to node j is a specified route
comprising links directed the same way, thus permitting travel from i to j. A
cycle is a chain that comes full circle and terminates at its origin node. A
path between nodes ii and jj is a sequence of links between these two nodes,
not all of which are necessarily directed the same way. In Fig. 4,
(1,2)(2,3)(3,6) , (12)(2,5)(5,6) , and (1,2)(2,3)(3,4)(4,5)(5,6) are chains from
1 to 6, and (2,5)(5,6)(6,3)(3,2) is a cycle from 2 to itself. A path from node 4
to node 1, which is not a chain, is (4,5)(2,5)(2,1), which is also a path from 1
to 4.
A node j is accessible to node i if there exists a chain from i to j. A

connected network is one in which there are paths between all pairs of nodes.
The network of Fig. 4 is such a connected one. A tree is all or part of a
network which is connected but contains no cycles. For example, the
network of Fig. 4 would become a tree if we removed the links (3,6) and
(6,3). The above definitions will suffice for the purpose of the discussion in
this book. A more detailed description of graphs can be found in Ford and
Fulkerson (1962) , for a mathematician’s view, and in Potts and Oliver
(1972), for the vehicular traffic application.
4.2 TRIP GENERATION AND DISTRIBUTION
In planning, or trying to improve the management of a traffic network,

one needs a good estimate of the traffic load carried by the current state of
the network. This means that one needs to know how traffic is generated at
any node, how it is distributed by apportioning of trips originating at that
node among all possible destinations, and how one route from origin to
190 Chapter 4
destination is assigned to any trip. The computed values and choices for
these three processes certainly change with the time of the day. They may
also change from one day to the next, in response to changing conditions,
(e.g. construction), or special attractions. They certainly change over long
periods of time, due to changes in the distribution of population, jobs, and
shopping and entertainment opportunities, to mention some of the primary
factors of influence.
Clearly, these three processes are interrelated. For example, if a trip is

generated because of a special attraction at a particular node of the network,
it will be natural to conclude that the newly generated trips will involve their
distribution into the attraction node. Some of the more recent work in this
area does consider the generation and distribution of traffic together. Let us
outline first some of the elements of these processes.
4.2.1 Trip Generation
Trip generation depends on the number of people living in a particular

zone, and their travel needs for various purposes, such as jobs, shopping, or
entertainment. For a first approximation, historical values of trip generation
may be used as a basis, to which may be added expected effects of current
attractions. For a more systematic approach to the task, particularly for the
purpose of obtaining trip generation estimates over a relatively long horizon,
it is customary to view the process in two stages:
1. The prediction of trip generation prediction variables, such as

population, car ownership, shopping opportunities, and job availability
at various nodes, and
2. The prediction of the number of trips originating from a zone, generated
by a particular trip purpose, as a function of the predictor variables.
It is generally assumed that all travel consists of a trip from a particular

origin to a destination, and a return trip on the same day. This is obviously
an approximation, since it ignores multi-destination travel, and trips not
generating a return trip on the same day, but for most cases of urban traffic
the approximation is a reasonably good one. Basic data needed for
estimating some of the key predictor variables, such as population, economy,
and car ownership, may be obtained through direct traveler surveys, and/or
publicly available census data. A full discussion of the processes involved in
trip generation and distribution can be found in Helly (1974).
4.2.1.1 Prediction of the Predictor Variables
POPULATION. The number of people living in a particular zone can be

changed by births, deaths, and net migration into the zone, i.e. the difference
between the number of people moving into a zone and the number of people
moving out of the zone. Birth and death rates may be assumed according to
historical data. Net migration, ideally, should be correlated to various driving
factors, such as quality of neighborhoods, proximity to jobs and attractions,
etc. For a first approximation, it is customary to simply extrapolate the
graphical or mathematical representation of past changes. Such
extrapolations may take into account the nature of a neighborhood in terms
of history and development. An old, established neighborhood is likely to
have a relatively constant rate of population change, resulting in an
exponential increase or decrease with time, possibly tending asymptotically
to some limit value. A newly developing neighborhood may have various
stages of development. For example, it may start with an early period of
slow population growth, followed by a middle stage of rapid population
increase, and ending with a third stage of exponential/asymptotic growth
similar to that of an established neighborhood.
In estimating population changes, it is important to take into consideration

expected unusual changes in conditions, such as a large change in
employment opportunities due to externally imposed conditions, such as
government initiatives. Also, estimates of population growth may involve a
region comprising more than one zone, in which case the estimate for a
single zone is obtained as a fraction of the regional estimate, taking into
account historical data regarding the value of such a fraction.
Sometimes it is appropriate to estimate population as a function of future

employment, particularly in the case of an area where employment is likely
to change drastically as a result of an infusion of investment. Frequently,
government funds may be the driving force for such an unusual growth of
employment. It may be noted that the labor force participation ratio, ( the
ratio of employment to the total population), may be abnormally high in an
area with very rapid employment growth. Furthermore, this ratio appears to
be rising constantly, though slowly, in the United States, and it is hovering
above 0.4 at the present.
ECONOMY. The economy of a region clearly affects travel in general,

and travel between nodea of a network, in particular. Economic forecasts
can, and often are made on the basis of an extrapolation from past historical
data. It is generally assumed that the local economy depends on some
principal economic sectors, which are those whose sales are largely exported
to points outside the study area, and whose scale of operation may be
192 Chapter 4
extrapolated on a regional or national basis. It is postulated that changes in

these basic sectors lead to changes in the rest of the local economy, affecting
employment and other factors pertinent to the traffic generation issue.
A more scientific approach to the prediction of long-term economic

development is known as the Leontief input-output model, which was used in
a new York area study (see Berman et al, 1961), and in the Northeast
Corridor Transportation Study, (see Putman, 1966). This model is based on
the fact that the output of a sector must be equal to the sum of the inputs
from this sector to all other sectors. It is specialized, for the purpose of
deriving traffic generation data, by taking into account explicitly purchasing
within an urban area of a portion of the products produced in that area. A
reasonable extrapolation of industrial activity into the future yields the levels
of activity of the various industrial sectors, which in turn drive the traffic
generation process.
LAND USE. Clearly, the nature of land use influences the source and
nature of travel, and consequently the traffic generation and distribution
processes. Several approaches to predicting land use have been contributed,
and partially tested. They all require extant land use inventories, and
independent forecasts of the overall demand for land, the latter expressed in
terms of building construction rates for the area. The models are either
simple extrapolations of past trends, or simulations of assumed human
behavior, both constrained by the available resources. The simulations may
be further subdivided into deterministic models, in which people are guided
only by accepted economic principles, or probabilistic models, in which
human behavior is only partially based on such principles.
An example of a simple extrapolation involves developing a multiple

linear regression equation of the form
where Y is the predicted change in the zone’s inventory of dwelling units,

retail space, factory activities, and other land use measures, to are
independent predictor variables, such as vacant land area, land value,
distance to employment sites or the centroid of the urban area, as well as
other variables describing zoning regulations influencing density and types
of construction, and availability of transit. The problem is to determine the
constants to on the basis of past trends. Having land use surveys for at
least two, and preferably more, points in time allows such a determination.
One of the early applications of this approach was for the Greensboro, North
Carolina area, by Swerdoff and Stowers (1966).
Another approach, which takes into account two different drivers of land
use, is known as the density-saturation gradient method. It was used in the
Chicago Area Transportation Study, (see Hansen, 1960), and it consists of
adjusting for consistency two extrapolated curves. One is for dwelling unit
density as a function of travel time to a central point, and the other is for
saturation levels of land use, also as a function of travel time to the central
point.
Two probabilistic simulation approaches, also discussed by Helly (1974),

are the following:
The first, known as the accessibility or gravity model. Is based on the

postulate that the likelihood of building on a site in a given zone depends
mainly on the zone’s distance from places of employment or commerce,
which gives rise to trips originating from the study zone. The other, known
as the intervening opportunity model, is based on the assumption that the
likelihood of building on a given site is a monotonically decreasing function
of the number of equivalent available sites which are nearer to the centroid
of the attractions listed above, namely, employment and commerce. Since
the predictors for both models change in time, time-dependent models have
been developed for both.
The accessibility model, first used in a time-independent form by

Hansen7, has the form
where is the forecast expected growth rate for zone i, an accessibility

index for zone i with an empirically determined constant exponent a, and
the vacant land area in zone i. The accessibility index has the general form
194 Chapter 4
where is a measure of activity in zone j such as total employment, and

is a measure of “friction” between nodes i and j, which may be made
proportional to the actual number of trips between the two nodes, as
surveyed or predicted exogenously. If no data exist allowing the estimation
of this friction, may be chosen as where is the travel
time between zones i and j , and b a constant exponent.
Intervening opportunity models have been proposed by Stouffer (1940),

Schneider (1960), Sherratt (1960), and Helly (1969), the last one being a
time-dependent one. The general approach to the intervening opportunity
models is the following.
Let us assume that there is just one centroid, and a builder’s site selection
proceeds from the site nearest to the one farthest from the centroid. Let M(r)
be the number of sites within a radius r from the centroid, where because of
the large number of sites in a city the integer-valued M(r) is treated as a
continuous variable. Let dM be the number of sites between M and M+dM,
and p dM the probability that a given house’s site is selected in the range
dM, given that it was not selected nearer to the centroid than M. The
parameter p is assumed constant. Then, the probability of not selecting a site
in the range (0, M+dM) is given by
which has the time-independent solution
If a total of houses are built, the number built on the M most central
sites is N(M) = [ 1 – Pr (M)], and the density of houses, n (M) is given by
The preceding time-independent formulation does not ensure, as it should,

that This deficiency can be overcome by a time-dependent
reformulation in which a site count variable m is introduced, with all
available sites at time zero described, in order of decreasing accessibility
from the centroid, by the integers 1 to m. Again, because of the large
number of sites in an urban area, the parameter m is treated as a continuous
variable. Now, we can define M (m, t) as the number of sites still vacant at
time t, which are no less desirable than m. We note here that M(m, 0) = m,
and that M (m, t) has the same meaning as the time-independent M Thus,
Eq. 5 still gives the probability Pr (M ), except that M is now time-
dependent.
Let N (t) be the total number of sites occupied at time t, and therefore M
(m,t) be equivalent to M(m,N). During a time increment dt, dN houses will
be built. Of these, a fraction [1 – Pr (M)] can be expected to be built on the
M most accessible and still vacant sites. The number builtduring dt on these
M sites equals the number that must be removed from the vacant site stock,
therefore we have the relationship
and using Eq. 5, we obtain
which has the solution
Now, let F (m,N) represent the expected fraction of sites occupied in the
range (m, m+ dm), which is found to be
196 Chapter 4
Figure 5, from Helly (1974), shows the form of Eq. 10 for various values
of pN and pm. In this figure, F is the probability that a site is occupied, m is
the number of sites in decreasing order of accessibility, N is the total
number of occupied sites, and p is the probability that a “considered” site is
accepted. Since the abscissa is in units of pm, large values of it correspond
to relatively inaccessible sites with large m. The curves in Fig. 5 show the
progression of a city from its beginning, with few houses and pN small, to
its later stages with increasing pN, when more and more desirable sites have
been built upon, and new construction takes place in the outer fringes.
Needless to say that this argument applies only to new construction and not
rebuilding of earlier structures.
The above discussion has focused on residential land use, but similar
models can be used for commercial or industrial developments. In such
cases, we need data on the need for such developments, and knowledge of
the often stringent constraints concerning such developments, either imposed
by authorities or dictated by considerations of entrepreneurial success.
4.2.1.2 Trip Generation from individual zones
Having the values of the predictor variables, one can proceed to the
prediction of trips originating from a zone. Trips are made by drivers and
passengers of private cars, by people riding buses and taxis, and by
commercial vehicles such as trucks. In many cities, private cars predominate,
although this may not be the case in very large metropolitan areas like New
York City. In any case, the trip generation picture is completed by
translating the predicted data on population, economy, and land use, into
individual trips.
It has been found that often, population density and household income are
the key factors influencing trip generation from households, including
commuting trips to work and other trips during the day. This is because
automobile ownership is highly correlated with these two variables. Another
factor that must be taken into account is the availability of modal choice, for
example, using available public transportation for commuting to work.
For modeling purposes, the number of trips per household may be

modeled by a correlation function of the predictor variables, fitted to
observed facts in the general area of a zone. It is reasonable to assume that
activity patterns do not change drastically overnight, and therefore an
extrapolation of current activity patterns into the future is a reasonable
approximation. However, as the planning horizon increases, account must be
taken of changes in behavioral patterns and other factors which cause
deviations from a simple extrapolation.
198 Chapter 4
Important contributions to the problem of trip generation have been made

by modeling activity patterns in an urban area, which generate trips. The
relevance of activity patterns was established by studies such as those of
Stopher et al (1979), Pas (1982), Damm (1983), and Jones (1983). Hirsh et al
(1986) contributed a dynamic, activity-based model of weekly activity
model. In their model, it is assumed that the week is divided into several
time periods. An individual selects an activity pattern at the beginning of the
week, and updates the pattern at the beginning of each subsequent period on
the basis of experiences during the preceding periods, and additional
information acquired during those periods. Figure 6 shows the decision
process for generating and updating a weekly activity program, following
Hirsh et al. Having the activity pattern leads to the determination of the
necessary trips for these activities.
4.2.2 Trip Distribution
Having the trip generation completed, the next step is to apportion the
trips originating from any zone to all possible destinations. Various models
used for this purpose are based on the levels of activity at the various
destinations, and the difficulty in getting to these destinations. It has been
found that the accuracy of trip distribution is improved if it is subdivided
into categories of activities, such as work, shopping, business or school. Any
model proposed must be validated and calibrated through the use of current
trip data, whose acquisition is one of the key challenges in transportation
planning.
In the models that are described below, it is assumed that an urban area is
divided into N zones, which comprise the totality of the area of interest.
Trips from outside the area are accounted for by being assigned to one of the
peripheral nodes of the area network. The following notation is used:
= number of trips from zone i to zone j during a specified period, e.g. a

weekday rush hour, as predicted by a trip distribution model for an
observable period.
= an estimate of obtained exogenously to the trip distribution model,
by survey or observation.
= number of trips originating in zone i during the specified
time period, as predicted by a model for an observable base period.
= an estimate of obtained exogenously to the trip
distribution model.
= number of trips from zone i to zone j during the specified time
interval, as predicted by the trip distribution model.
= an estimate of obtained exogenously to the trip distribution

model, by application of a trip generation analysis.
If the period of interest is one day, then, on the basis of reasonable

expectations about commuting patterns, it is assumed that
and A matrix form representation of any of
these quantities is a convenient form of their display.
Some of the models proposed for the treatment of trip distribution are
known as Growth Factor Models, which try to create a set of origin-
destination choices compatible with the growth of trips originating in various
zones, and the growth of attractions in other zones. Other model use the
gravity and intervening opportunity concepts discussed in Section 4.2.1 in
conjunction with land use. Yet another approach to trip generation is
provided by a variant of the gravity model known as the minimum entropy
model.
4.2.2.1 Growth Factor Models
Let us define a Growth Factor, according to the expression
If we assume that trip distribution is done entirely on the basis of such

growth factors for all the N zones, then a simple
approach to trip distribution is based on averaging the growth factors of
origins and destinations, yielding
200 Chapter 4
Admittedly, such a simple extrapolation ignores changes in travel

parameters such as travel times, which are induced by growth, and can be
used for relatively small time horizons. In any case, one must iterate in order
to produce internally consistent results. The value of
obtained on the basis of Eq. 11 is generally different from which is a
basic estimate of traffic volume for zone i. In order to adjust the so
that comes close to we may view the value obtained from Eq. 11 as
the first step of an iterative procedure, in which
The iteration is repeated until some step m for which
can be viewed as being sufficiently close to zero.
4.2.2.2 Gravity Model
The growth factor models do not take into account the relative travel costs
associated with alternative destinations. These cost can change in time
because of new roadway construction, or other changes in the general area,
including mass transit schedules and availability. The Gravity Model, (see
Voorhees, 1955), is one that tries to take into account the degree of attraction
between two zones. It states that the number of trips between two zones is
proportional to the total number of trips originating at the origin zone i, the
number of potential trip ends at the terminating zone j, and a cost function
also called a “friction” between the two zones. If the zones are near each
other, is large, and if they are far from each other, it is relatively small.
For some observed, base period, travel patterns, the model states that
where is a measure of attractions, and useful trip ends, characterizing

node j . Equation 13 states that trips are apportioned according to the
relative attractiveness of various possible end nodes. The sum in the
denominator guarantees that
namely, that the sum of the trips from zone i to all other zones is equal to
the exogenously obtained estimate of this sum. The measure of attraction
may be defined for a particular class of trips, e.g. travel to work, or it
may lump together all attractions. In the latter case, if then
may be an appropriate value.
The next challenge is to find an appropriate function which can be

applied to zones i and j. A good first approximation is to assume that
where is the travel time between the two zones, and
consequently an inverse measure of attractiveness. Assuming an empirical
function F (d), we can calibrate the model by minimizing the sum of the
squared deviations, between the model and actual measurements.
If such a function is found, it is generally assumed that its form will remain
202 Chapter 4
valid for at least an appreciable length of time, and thus future trip
distribution can be obtained as
where and are the best estimates of future attractions, trip

generation, and travel times. It may be noted here that the value of may
be specifically chosen in order to take into account such things as a major
barrier between zones, or a costly toll facility. For example, we may set
where is a correction factor accounting for the special
impediments to travel between i and j , and is equal to 1 if no such
impediments are present. A popular choice for is
where the exponent c is a constant fitted to available data. Observation of

Eq. 16 reveals the origin of the term “gravity model”, because of its
resemblance with the Newtonian law of gravitational attraction of two
masses, which corresponds to c = 2 .
4.2.2.3 Intervening Opportunity Model
The Intervening Opportunity Model, as discussed by Ruiter (1967), is

based on the assumption that, for every trip originating at a particular zone,
all appropriate possible destinations are ordered on the basis of the travel
time from the zone, in order of increasing travel time. It is also assumed that
there is a constant probability p that a considered site is accepted as the trip
end. The remainder of the formulation is analogous to the one used for the
intervening opportunity model for land use, Section 4.2.1.
Let M(d) be the number of eligible destinations nearer than d to the trip
origin, where d is commonly viewed as the travel time. Assuming that the
number of destinations is large, we can treat M(d) as a continuous
variable. Then, the probability of rejecting the nearest M sites, denoted by
P(M). can be deduced from
which has the solution
Accordingly, the value of is given by
where
= probability of not accepting any destination nearer than zone j,

= probability of not accepting any destination in zone j or nearer
than zone j.
Therefore,
where is the number of destinations nearer to the trip origin than the
destinations in zone j , and is the number of destinations in zone j.
Equation 20 may now be used to calibrate the model according to available
data, by finding the best possible value of p that fits the data. After
calibration, Eq. 20 may be used to estimate in terms of
204 Chapter 4
4.2.2.4 Minimum Entropy Formulation
The gravity model discussed earlier is a proportional model, in the sense

that the number of trips from zone i to zone j is proportional to the total
number of trips, from zone i, and to the number of attractions, in
zone j. The model may be simplified by retaining the proportionality feature
but ignoring “friction”, and therefore setting for all pairs of i and j,
in which case Eq. 13 becomes
Over a 24 hour period, we may assume that all trips are associated with a
return trip, therefore and = the number of trips observed
or estimated as originating from zone j . Then, the conditions satisfied by
this simplified model are
where T is the total number of trips in the network. The conditions in Eqs. 22
express the conservation laws, namely, that every trip has an origin and a
destination, and the total number of trips in the area is equal to the sum of all
zonal trips.
It may now be shown that the same Eq. 21 can be obtained by

minimizing the function
subject to the basic conditions of Eqs. 22. At this point we have an

explanation of the term entropy used in the label of this model. Because the
function S looks very much like an entropy function, as used in statistical
thermodynamics, the model was labeled as the entropy model, (see Potts and
Oliver, 1972, pp. 121-135).
The minimization of S in Eq. 23 subject to the conditions in Eqs. 22 may

be done through the use of Lagrange multipliers, involving minimizing the
composite function
where and c are undetermined multipliers. To minimize F, we take

the derivatives with respect to and c and set them equal to zero,
thus obtaining
plus the constraints expressed by Eqs. 22. The solution of Eq. 25 is
We now enter as expressed in Eq. 26, into Eq. 22, then multiply the
first two relations of Eqs. 22 and divide the result by the third expression,
obtaining
206 Chapter 4
Since the result shown in Eq. 27 is the same as in Eq. 21, the entropy
formulation is shown to be equivalent to the simplified gravity model, under
the assumptions used for the various parameters. More complex versions of
the entropy model may be obtained, for example, by taking into account
explicitly the relative travel times between zones. Extensions to the above
model have been given by Wilson (1967). Although the use of the entropy
concept is not universally endorsed, (see for example Beckman and Golob,
1972), it does present a capability of handling models of increasing
complexity in a rational manner.
4.3 TIME-INDEPENDENT TRAFFIC ASSIGNMENT
Having the origin-destination matrix for a network, the next and most
important question is how traffic proceeds from origin to destination. This is
a matter of route choice by an individual driver. Today, and in the
foreseeable future, such a choice is strictly in the hands of the driver. An
attempt to describe the driver’s choice was first made by Wardrop (1952),
who proposed the following two principles:
1. The journey times on all routes actually used are equal and less than
those which would be experienced by a single vehicle on any unused
route,
2. The average journey time (for all users) is minimum.
It is assumed that principle #1 is followed by individual users in selecting

their routes, and for this reason it is known as leading to the “user-optimal”
solution to the traffic assignment problem. Principle #2 is known as leading
to the “system-optimal” solution. It turns out that no matter how well the
user-optimal solution is implemented, it does not also guarantee a system-
optimal solution, because in general the two principles are not satisfied at the
same time. This incongruence of the two solutions, although not very
paradoxical, is a legitimate feature of the “Braess Paradox” mentioned in
Chapter 1 of this book.
An obvious application of Wardrop’s first principle is to assume that a

driver tries to follow the shortest route to a destination. Shortest route
computations are routine in graph theory, and they have been applied to
transportation networks. One of the special nuances in transportation
networks is that the travel time along a link is not constant but depends on
the volume of traffic along the link. Phenomenological relationships between
speed and density have been applied to account for this variability of transit
time along network links. Other models used to describe human behavior
and determine traffic assignment characteristics, take into account a set of
attributes of different route choices, which may include financial as well as
physical characteristics of networks.
We shall now discuss traffic assignment models proposed over the years,
including both deterministic models and stochastic models. We shall first
consider the case of time-independent traffic assignment, in which the traffic
demand is assumed to be essentially constant. We shall start with a
discussion of deterministic traffic assignment models, in which the travel
time or cost is assumed to be precisely known by the users of the system. We
shall then go on to stochastic models, in which some uncertainty is allowed
concerning the properties of the network. In particular, we shall discuss the
Discrete Choice models, which are a natural sequel to the preceding
discussion of traffic generation, and to some extent integrate traffic
generation and traffic assignment. We shall then proceed to the discussion of
both deterministic and stochastic models of the time-dependent variety,
known as Dynamic Traffic Assignment. We shall conclude with a discussion
of the effect of congestion on traffic assignment computations.
4.3.1 Deterministic Traffic Assignment Models
The development of Traffic Assignment models has been based on the

application of the two Wardrop principles, and particularly the first one. This
is natural, because by and large decision regarding travel choices are still in
the hands of individual drivers, who are likely to apply Wardrop’s first
principle, if they apply anyone at all. Wardrop’s second principle would be
implemented only through the intervention of traffic management authorities
interested in minimizing the aggregate cost of travel to all users. Such an
intervention has not been forthcoming as yet, notwithstanding the efforts at
developing Intelligent Transportation Systems (ITS), in which interaction
with drivers may ultimately influence their travel choices.
It should be pointed out at the start, that all models proposed for handling
driver choices from among alternatives implicitly assume superior wisdom
208 Chapter 4
and knowledge on the part of a driver, plus willingness to make drastic

changes in order to achieve optimality. This, however, is not the case with
most drivers. By and large, drivers may be knowledgeable about one or two
alternate routes to a destination, and may be willing to switch to one or the
other if there is convincing evidence that the alternative choice leads to a
congested domain of the network. Nevertheless, we shall discuss the
ambitious traffic assignment algorithms, with the thought that increasing
deployment of ITS services is likely to increase the knowledge, and
decision-making ability of drivers.
4.3.1.1 Cheapest Routes
There is a plethora of algorithms for determining the cheapest routes on a

network, with an extensive bibliography provided by Murchand (1969). All
of them are essentially based on the following premise: If a cheapest route
from node to node passes through then the portion of the route from
to is the cheapest route from to and similarly the portion of the
route from to is the cheapest route from to The reader will
recognize this a basic statement of a dynamic program as discussed by
Bellman (1958). Accordingly, the cheapest routes from node to node
of a network [ N ; L] with positive link costs are the unique
solutions of the functional equations
The various cheapest route algorithms are, in effect, procedures for solving
Eqs. 28. We start with the Tree-Building Algorithms.
4.3.1.2 Tree-building Algorithms
This algorithm consists of fanning out from the home node (origin) to all
other nodes in increasing order of their costs from the origin, as described by
Dijkstra (1959). The nodes are successively labeled with two numbers, one
describing the predecessor node on a cheapest route from the origin, and the
other the cost of the route up to the node in question. Initially, the origin is
labeled (0,0), and all the other nodes are temporarily labeled A
node is permanently labeled, when the cheapest route to it is determined.
The following notation is used in the sequel:
The graph of the network is a directed graph [ N ; L] , consisting of N

“nodes”, which are denoted by i or (i = 1,2,....,n), and L “links”
denoted by (i ,j) or
A partial graph of a directed graph [ N ; L] is a directed graph
with a subset of L, i.e.
A subgraph of [ N ; L] is a directed graph with, and
where is the set of all links of L which join the nodes of
A bipartite graph is one in which the set of N nodes is partitioned into
two complementary sets i.e.
The algorithm proceeds as follows. The last node permanently labeled is

considered as a predecessor node, and all nodes to which links from this
node are directed are looked at as possible new predecessor nodes. If the
sum of the cost to the current predecessor node plus the link cost to a
downstream node is lesser than the tentative cost label to that node, then this
new cost becomes a new tentative cost label for the downstream node, and
all other tentative labels are left unchanged. After all the possible
downstream nodes are investigated, the one corresponding to the minimum
new tentative cost is chosen as the new predecessor node, its tentative cost
becomes permanent, and the process is repeated until the permanent labeling
is completed.
It is convenient to use a duplicate notation for all nodes, by denoting them

as 1,2,...,i,...,n , and also as they are successively
selected at steps 1,2,... ,k,... of the algorithm. For simplicity, it will be
assumed that the network [N; L], (with N nodes and L links), is connected
and undirected and that the link costs are At the step,
nodes are tentatively labeled with two numbers, one being the
210 Chapter 4
current predecessor node, and the other the cost of the

currently known cheapest route from the origin to node When the
algorithm is concluded, all nodes have permanent labels and
representing the predecessor node and the cost of the cheapest
path, respectively. The numerical procedures for the algorithm are as
follows:
Step k = l :
Steps k= 2,3, . .., n :
a) For and
set
b) For all other define
c) Define by
and define the permanent labels
d) Define
Since the above procedure solves Eqs. 28 by

permanently labeling all nodes in order of increasing cost from the origin.
The cheapest path themselves are determined by tracing back to the origin
across successive predecessor nodes. This completes the tree-building
algorithm.
In the above discussion, we talked about cost associated with a link. This
cost is usually strongly tied to the travel time along the link, but additional
factors may be considered, such as turn penalties and prohibitions. In the
end, the decision-making individual is assumed to select the route
corresponding to the cheapest route, according to Wardrop’s first principle,
leading to the Cheapest Route Assignment for every individual, but not
necessarily to the minimum aggregate delay for all individuals combined. It
may be pointed out here that the procedure just outlined assumes a great deal
of flexibility, and superior wisdom, on the part of all users of the network, a
rather unlikely occurrence. Quite often, users limit themselves to just a few
route choices, making the route assignment much simpler , albeit somewhat
removed from the optimum. In traffic assignment practice, the most common
cheapest route assignment is the all-or-nothing assignment, in which link
costs are supposed to be constant and flow-independent. Traffic between any
origin and destination pair is all assigned to the cheapest route, and none is
assigned to any other route. Such an algorithm is justifiable for not too heavy
traffic, but it becomes increasingly inaccurate as congestion sets in.
4.3.2 Discrete Choice, Stochastic, Models
The following is a concise overview of the primary Discrete Choice

Models, (DCM), proposed over the years. Extended discussions of the topic
212 Chapter 4
may be found in Ben-Akiva and Lerman (1985) , and in Ben-Akiva and

Bierlaire(1999).
Decision Maker. In a DCM, a Decision-Maker who is effectively an

individual, but may also be a group making a single decision on choice,
effectively goes through the following steps:
Determines the options available,

Estimates the benefits and costs of these options,
Applies a decision rule to choose the “best” option.
Alternatives. By necessity, the alternatives presented to decision-makers

are limited to those of which they are aware. This limitation introduces an
uncertainty in modeling the generation process of choice sets, which has
motivated the use of probabilistic models that predict the probability of each
feasible choice set being selected from among all possible choice sets.
Attributes. Each alternative is associated with various attributes, which

may be shared by other alternatives. An attribute may not be a measurable
quantity, but may be any function of available data. Some assumptions of
human behavior may dictate the form of such functions. For example, out-of
pocket cost of travel may be considered alone as an attribute, but it may be
found that a better correlated attribute is the out-of-pocket cost divided by
the income of the decision maker.
Decision Rules. A decision-maker implicitly evaluates the attributes of

various route choices and determines a choice. Most models of decision-
making are based on utility theory, which involves the association of each
choice with a merit value called utility, after which the decision-maker
chooses the alternative with the highest utility. For example, if the decision-
maker is only interested in minimizing travel time, the utility of the various
route choices would be proportional to the inverse of the travel time.
The utility theory assumption regarding decisions by individuals is

elegant, but not altogether realistic. Human beings are not generally able to
assess attributes reliably, and they often make almost arbitrary decisions
which deviate from those that utility theory alone would dictate. For this
reason, models involving probabilistic decision rules have been proposed to
account for deviations from the deterministic application of utility theory.
Some of these models assume deterministic utilities and probabilistic
decision rules. Examples are the model of Luce (1959), and the “elimination
by aspects” model of Tversky (1972). Other models, used extensively in
econometrics as well as travel behavior analysis, and known as Random

Utility Models, are based on deterministic decision rules and utilities
represented by random variables.
4.3.2.1 Random Utility Models
Random utility models assume that the decision maker has a perfect
discrimination capability, but the analyst who tries to guess the decision
making process has incomplete information, and therefore some uncertainty
will exist regarding the decision maker’s choices. The utility is modeled as a
random variable, reflecting this uncertainty. Specifically, the utility that
individual n associates with alternative i in a choice set is assumed to
be given by
where is the deterministic part of the utility, and is a random term

accounting for the uncertainty. It is assumed that the decision maker chooses
the alternative with the highest utility, therefore, the probability that the
alternative i is chosen from the choice set is
To make the random utility model operational, we need to introduce

certain additional assumptions.
Location and scale parameters. If the values of are replaced by

where µ and are arbitrary real numbers, Eq. 36 is still valid,
as can be seen by the following derivations:
214 Chapter 4
It is seen that only the signs of the differences between utilities determine
the decision outcome, and not the utilities themselves. The constants µ and
are known as the scale and location parameters, respectively, and they
are chosen for various objectives. The scale factor is usually chosen for an
appropriate normalization of one of the variances of the random terms, while
the location parameter is usually set to zero.
Alternative Specific Constants. Without undue restriction, the mean of

the random terms can be assumed to be equal to any convenient value c . If
we denote the mean of the error term of alternative i by we can
define a new random variable such that This
leads to a model in which the deterministic parts of the utilities are
and the random terms are with mean c , satisfying the relationship
The terms are known as Alternative Specific Constants (ASC), and

they capture the means of the random errors. Inclusion of ASCs in the
deterministic part of the utility functions allows, without loss of generality,
for the error terms of random utility models to have a constant mean c. The
choice of the ASCs is generally dictated by reasons of computational
convenience, although it may affect the estimation process of the decision
making, as shown by Bierlaire et al (1997).
Utility Value. The utility value is divided into the deterministic part, and
the stochastic part.
The deterministic term for each alternative is a function of the

attributes of the alternative as well as the characteristics of the decision-
maker, given by
where is the vector of attributes for attribute i as perceived by the

individual n , and is the vector of characteristics of the individual. A
new vector of attributes is now defined from both and by using an
appropriate vector valued function h , according to
where the function h is generally assumed to be continuous and monotonic

in and its form is chosen to give the best representation in a specific
application. Equation 39 now becomes
and if the utility specification is linear in the parameters,
with the utility fully specified by the vector of parameters
The random part of the utility can have any of several form, and we shall
describe here the ones used in two leading families of utility functions, the
Logit family, and the Probit family.
4.3.2.2 The Logit Family
The Logit (Logistic Probability Unit) model was first introduced in the
context of binary choice, where the logistic distribution is used. When
generalized to more than two choices of alternatives, it is known as the
Multinomial Logit Model. It is based on the assumption that the error terms
are independent and Gumbel-distributed (Gumbel, 1958) , that is, for all i, n
it is described by
216 Chapter 4
where is a location parameter, and µ is a strictly positive scale parameter.

The mean of the distribution of Eq. 36 is where
is the Euler constant. The variance of the distribution is and the

probability that individual n chooses alternative i from the choice set
is given by
An implicit property of the Multinomial Logit Model is a basic lack of

correlation among alternatives, which sometimes leads to strange, counter-
intuitive results. An extension to the Multinomial Logit Model, designed to
capture some of these correlations, was first proposed by Ben-Akiva (1974),
and is known as the Nested Logit Model. It involves partitioning the choice
set into M “nests”, with the utility of each alternative
comprising two terms, one specific to the alternative and the other associated
with the nest. The Nested Logit Model corrects some of the deficiencies of
the Multinomial Logit Model by capturing choice problems in which
alternatives within each nest are correlated. No correlation across nests is
allowed by the Nested Logit Model, which therefore cannot be used when
alternatives cannot be partitioned in well separated nests of cross-correlated
alternatives.
An extension of the Nested Logit Model is the Cross-nested Logit

Model, in which an alternative may belong to more than one nest. It was first
proposed by McFadden (1978) as a special case of the Generalized Extreme
Value Model (GEV), which includes the Multinomial Logit and the Nested
Logit models. In the GEV model, the probability of choosing alternative i
within is
where G is a non-negative differentiable function, with the following form

for different Logit Models,
1. Logit model:
2. Nested Logit Model:
3. Cross-nested Logit Model:
4.3.2.3 Multinomial Probit Model
This model has been named after the “Probability Unit” (Probit) which
characterizes it. It is derived from the assumption that the error terms of the
utility functions are normally distributed, and captures explicitly the
correlation among all alternatives.
We define the vectors and where are the utility

functions Thevector of error terms
is multivariate, normal-distributed, with a vector of means 0 and
variance-covariance matrix The probability that the individual n
chooses alternative i from the choice set is given by
We now define the matrix according to
which turns out to satisfy the relationship

218 Chapter 4
The density function is now given by
where
and
The matrix has a form in which the column contains -1 across all
rows. If the column is removed, the remaining matrix is
the identity matrix.
It should be pointed out that the multi-fold integral of Eq. 51 becomes

intractable even for a relatively low number of alternatives, and the number
of unknown parameters in the variance-covariance matrix grows with the
square of the number of alternatives, making computations rather
cumbersome. A detailed discussion of multinomial probit models can be
found in McFadden (1989) . An additional variation on the theme of
Multinomial Probit has been the Hybrid Logit Model proposed by Ben-
Akiva and Bolduc (1996), which is intended to bridge the gap between Logit
and Probit models by combining the advantages of both.
4.3.3 Global Network Optimization Models
So far, we have considered networks subject to Wardrop’s first principle,

namely, networks in which route decisions are in the hands of individual
drivers. We shall now discuss the implications of Wardrop’s second
principle, namely, the case when intervention of a higher authority may
force individuals to select routes which optimize the operation of the
network for the aggregate of users. In increasing level of complexity, we can
define the following types of networks, which are to be optimized:
1) Networks with fixed link costs, and no link capacity restraints.

2) Capacitated networks, with fixed link costs.
3) Networks with flow-dependent costs.
4.3.3.1 Fixed Link Costs, no Capacity Restraints
Let us first dispense with the trivial case corresponding to fixed link costs,
and no capacity restraints. This is the case in which it is assumed that the
performance of a link is not degraded by the addition of traffic, and such
addition is unconstrained. If, for example, the cost is tied to travel time, the
assumption here is that this travel time does not change appreciably with
increasing traffic demand. This is clearly a case of a network with rather
light traffic, operating in the range of densities for which speed is almost
constant, and flow is a linear function of the density. In this case, the two
Wardrop principles are simultaneously satisfied. Optimization of individual
choices results in a global optimum for the operation of the network, since
the presence of any individual on a link does not affect the performance of
that link. Even discrete choice models, like the Logit and Probit models,
would be associated with a global optimum after individual optimization, as
long as the traffic characteristics affecting individual choices, including
uncertainties, are not influenced by traffic demand and traffic densities.
It will be useful for the discussion in this section, and others to follow, to
define certain variables and illustrate them by an example. Consider the
network depicted in Fig. 7. It is assumed that a number of OD pairs have
220 Chapter 4
been defined for this network, and that the link costs for the links j =
1,2,..., 10 are given. For example, let us assume that two OD pairs exist, A-
F, and B-F. For each OD pair, there exist a number of link paths. In the
example of Fig. 7, the possible paths, nine altogether, are:
For A-F : 1,4,9 2,7,10 1,5,10 2,6,9 1,4,8,10 2,3,4,9

For B-E: 4,9 5,10 4,8,10
We denote the paths by k = 1,2,... .,K, where K is the total number

of paths which can serve all OD pairs. We can define the link incidence
matrix, which associates each link with any of the above paths, by assigning
the value 1 if the link is part of the path, and 0 if it is not. For example, if
the A-F paths are numbered from 1 to 6, and the B-F paths from 7 to 9, the
link incidence matrix (j=1,2, . . . , . J ) is given by Eq.
52 . If we denote the link costs by we can compute a
vector of path costs, according to Eq. 53. It is easy then
to identify the cheapest route corresponding to any specific OD pair. This
cheapest route will be the one chosen by a knowledgeable individual
associated with the OD pair. It will also be the path chosen for the purpose
of minimizing the network cost for the aggregate of users, since the link
costs are not affected by the volume of traffic associated with them.
As a numerical example, let us assume that the link costs are
The cost vector would then be
where the first six components correspond to the OD pair A-F, and the last
three to the pair B-F. It can be seen that the optimum assignment
corresponds to the choice of path 1,4,9 for the A-F pair, and 4,9 for the B-
F pair. The solution was thus obtained by complete enumeration of all
choices, but it can also be formulated as a Linear Programming problem, as
discussed by Potts and Oliver (1972).
4.3.3.2 Capacitated Networks with Fixed Link Costs
Let us now assume that each link j is capacitated, namely, it can handle
traffic volumes only up to a maximum, Minimizing the network cost
222 Chapter 4
must now take into account explicitly the volumes

associated with the L OD pairs, where L = 2 in the above example. If these
volumes cannot be fully accommodated by allocation to the cheapest route
for each OD pair, we must partition them among all possible paths. Let us
assume that the volume of the pair is partitioned among the K paths, in
proportion to the variables where all the paths are
included in K, whether or not they correspond to the pair. If they do not,
the corresponding is automatically zero.
We now compute a flow matrix for all paths and links, according to
We have the following relationships:
Minimizing the aggregate cost to the users is now translated into the
objective
subject to the constraints expressed by Eqs. 54 and 55.
4.3.3.2.1 Capacitated Networks with Flow-dependent Costs
The optimization problem can become quite complex if the link costs are
complicated functions of the link flows. In many cases, traffic theorists have
assumed linear relationships between flows and costs, of the type
where A and B are constants, and is the flow along the link j. The cost
functions implied by Eq. 57 are rather crude, and the range of flows over
which they may give a reasonably good approximation has to be carefully
selected. However, they simplify considerably the optimization problem,
which again becomes a Linear Programming problem, whose variables are
the flow partitioning variables
As an illustration, let us consider the example of the network of Fig. 7,

and the associated link and OD parameters. We shall also assume that the
OD demands and link capacities are
expressed in units of vehicles per minute. For the purpose of obtaining the
minimum cost assignment, we shall also assume linear cost functions for
each link, of the type with
Nine weighting parameters for the nine possible paths are to be

determined, namely,
For OD pair A-F:

For OD pair B-F:
The matrix now becomes

224 Chapter 4
and the problem is to minimize the network cost function
subject to the constraints
The problem statement will be recognized as one of a Linear

Programming nature. Its solution is straightforward, and corresponds to the
following values of j = (l,2, ... ,9) :
which is translated into the following link flows for the ten links:
It may be observed here that before any attempt at optimization, it is

helpful to find out if the network is capable of handling the specified traffic.
This can be determined by resorting to minimal cut arguments, familiar in
graph theory. A minimal cut intersects links downstream from a source or a
group of sources, and the summed up capacities of these links is the smallest
among those produced by all possible cuts with the same upstream sources.
The total throughput of the minimal cut must exceed the sum of the upstream
sources if the network is to be able to handle the demands. In the numerical
example given, the minimal cut for source A alone is across links 1 and 2,
with throughput 48 > 27, (Fig. 7). The minimal cut for both sources A and
B is across links 9 and 10, with throughput 49 > 47, so the network
capacity is adequate for handling the demands.
4.4 DYNAMIC TRAFFIC ASSIGNMENT
Everything we have covered so far concerning traffic assignment deals

with steady-state traffic, unchanging with time. Given the fact that in most
urban areas traffic changes substantially with time, particularly during the
rush period, it behooves us to ask how such time-dependent traffic should be
assigned to different routes in the network. This question has led to the
development of Dynamic Traffic Assignment methodologies, in which the
time-dependent OD demands are explicitly considered as one tries to
minimize the travel cost, for a single user or for the aggregate of users. The
assignment of paths to different OD pairs will most likely be changing with
time in this case. Both, analytical techniques and simulation programs have
been used in dealing with dynamic traffic assignment.
One of the earliest contributions on dynamic traffic assignment was that

of Yagar (1970, 1971) , who used the simulation approach in order to
optimize individual route choice according to Wardrop’s first principle. A
226 Chapter 4
user was assumed to minimize travel cost for varying traffic demand
conditions and queue evolution. The simulation divided time into time slices
with constant demand, with queue demand at the end of a given slice added
to the new demand. Yagar also contributed a heuristic solution algorithm to
the problem. The first mathematical programming approach to the problem
was given by Merchant (1974), and by Merchant and Nemhauser (1978a,
1978b). A macroscopic model was formulated for minimizing total
transportation cost, in the form of a discrete-time, nonlinear, nonconvex
mathematical programming model. Ho (1980), and Carey (1986, 1987)
contributed discussions complementing the Merchant-Nemhauser
contributions, and in the case of Carey, also reformulating the problem as a
well-behaved convex nonlinear problem. Reviews of the topic of Dynamic
Traffic Assignment has been provided by Wie (1991), and by Florian and
Hearn (1999). Both deterministic and stochastic versions of the dynamic
traffic assignment problem have been considered. In what follows, we
discuss the basic formulation of a deterministic model.
In dealing with the time-dependent operation of a network, one needs

information regarding the time-dependency of the traffic demands over a
suitable time horizon. There are two sources for such information:
Historical data regarding travel demand from different origins to

different destinations.
Real-time traffic data, measuring actual travel demand, as well as
detecting traffic incidents which influence the performance of network
links.
Consider a traffic network [N ; L ] with nodes n, (n = 1,2, . . . , N), and

OD pairs l, (l = 1,2, . . . , L). There are also k paths in the network, ( k =
1.2...,K ), whose optimum use is the goal of the optimization problem. A
subset of these paths serves the OD pair. We consider a time period
[0, T ] for departures, and a termination time at which all traffic is
assumed to have been served. Assume that the demand functions for each
OD pair are time-dependent, given by and give rise to path
flows for The actual experienced travel time for
a path k which carries a flow generated at time t is We also
assume that the departure rates are continuous over time.
The feasible domain for the dynamic network equilibrium model must
satisfy certain flow conservation and non-negativity constraints, namely,
The extension of Wardrop’s first (user-optimal) principle may now be

stated as follows: A feasible flow with path flow rates h and path travel
times s(h) is a user-optimal solution to the dynamic traffic assignment
problem if, at any time t, the actual path travel time for the paths used is
the shortest possible among all paths. This means that we are trying to
minimize the expected travel time to individuals at anytime
Therefore, the dynamic traffic assignment is defined by the relationships
It can be shown that the equilibrium conditions lead to the variational

inequality
It is necessary at this point to state explicitly the order in which traffic

demands are served. It is reasonable to assume a First-In-First-Out (FIFO)
condition, meaning that a traffic stream cannot overtake another stream that
starts earlier. In order to satisfy the FIFO condition when the path costs are
additive, (i.e. path costs are the sums of the corresponding link costs), certain
conditions must be satisfied by the link cost functions. The FIFO condition
may be stated as
228 Chapter 4
where a denotes a network link from the set of links A . Equation 61 is

equivalent to stating that, for a small time increment
which, after dividing by and taking the limit as leads to
The question then arises as to which link cost functions satisfy the FIFO
condition given by Eq. 63. Friesz et al (1993) showed that this condition was
satisfied for a linear cost function, and a subsequent study by Xu et al (in
press) showed that FIFO conditions would also be satisfied for nonlinear link
cost functions which were not “too steep”.
Having settled the FIFO issue, we can now look for a solution algorithm
for the dynamic network equilibrium model. The algorithm requires a time
discretization, and the adaptation of a general method for solving variational
inequalities in order to compute a solution of Eq. 60. The solution algorithm
may be viewed as comprising two parts. The first is to determine
given The second, referred to as the network loading problem, is to
determine given the path flow rates A schematic diagram of
the necessary process is shown in Fig. 8. It should be pointed out that the
solution to the network loading problem is a rather complex task, since it
requires the computation of time-varying link flows, travel costs, and path
travel costs, by using the time-varying path flow rates A numerical
approach to this task has been given by Wu et al (1997).
It may be pointed out that the preceding solution algorithm does not take
into consideration any capacity constraints on the link flows and
occupancies. As a result, it is likely to give erroneous results in cases of
congestion, when link capacities may be the limiting parameters of any
assignment.
There have been additional contributions to the dynamic traffic

assignment problem, accounting for uncertainties in the data that influence
user decisions. Stochastic models have been proposed, which can account
for day-to-day or within-day variations from deterministic determination of
network loading, (e.g. Cascetta, 1989). In view of the difficulty in dealing
with the mathematics of the dynamic traffic assignment problem in a
thorough manner, a great deal of effort has been devoted to simulating
dynamic network loading and optimal operation.
4.5 TRAFFIC ASSIGNMENT IN CONGESTED SYSTEMS
In all the preceding sections on traffic assignment, there is a certain

element which is characteristic of relatively uncongested systems. It is the
fact that the properties and performance of a network link are assumed to
230 Chapter 4
depend solely on the traffic volume of that link. Whether one talks about
travel cost or something related to that cost, like travel time, the implicit
assumption is inherent in all the discussions that these link characteristics are
functions of the link traffic volumes alone. Furthermore, they often are
assumed to be explicitly represented by phenomenological relationships,
such as the flow versus concentration relationships discussed in Chapter 1 of
this book.
The above assumptions are increasingly violated as congestion sets in.

First of all, it must be realized that the phenomenological flow versus
concentration relationship is not like the property of a wire carrying electric
current. Traffic links are most often forced to carry higher and higher
volumes of traffic because traffic movement is impeded downstream. Insofar
as link traffic is relatively uniformly distributed along the link, and moving,
one can still use some phenomenological relationships. However, with
increasing congestion, this uniformity along the link is violated and the use
of such relationships is incorrect. The key observation here is that the link
does not stand alone in acquiring its properties, but it is influenced by the
downstream section, and similarly it influences the upstream section of a
traffic path. The network as a whole becomes a Store and Forward system,
such as those discussed in Chapter 3 of this book.
As stated already, the basic assumption in the study of traffic assignment

is the existence of a so-called "Fundamental diagram" of traffic behavior,
such as that discussed in Chapter 1 and shown in Fig. 9. The Fundamental
Diagram stipulates that traffic flow, q, is zero at zero traffic density, k, and
also at some Jam Density, and it attains a maximum at some
intermediate density. The basic premises underlying the Fundamental
Diagram are basically correct, although a one-to-one correspondence
between flow and concentration is substantially far from what is generally
observed in real life. All observations point to a very wide scatter of flow
values for any given concentration, particularly at relatively high
concentrations. This observation alone should make one cautious in the
application of the Fundamental Diagram to describe traffic in a network.
However, there is an even greater mistake in such an application. The

Fundamental Diagram essentially applies to steady state flow over an infinite
highway, and models reasonably well transitions from one steady-state to
another over such a highway.
However, an arc of a network is not infinite, and is not described well by

a behavior such as that of Fig. 9. A network arc is better described as a
traffic facility whose flow may be limited by a Throughput Limiting Point
(TLP) during periods of congestion. The TLP may be the point of an
entrance ramp, the downstream end of an arc, or it may be an intermediate
point with a reduced capacity. A collection of TLPs is shown in Figure 10.
The nature of the six TLPs shown is as follows:

232 Chapter 4
TLP1: This is the merging of traffic along two arcs into a single
downstream arc, where the throughput of the downstream arc is less than the
sum of throughputs of the two converging arcs.
TLP2: This is the effect of an entrance ramp feeding into an arc, and thus
limiting the throughput allocated to the traffic upstream from the ramp
entrance point.
TP3: This is the possible effect of an exit ramp. An exit ramp normally
reduces the number of cars in a traffic stream and thus improves the flow.
However, much too often, there is a Spillback effect caused by a limited
capacity of the exit ramp, or an obstruction at the exit ramp such as a traffic
light or stop sign at the end of the ramp. As described in a previous paper2, in
such a case a queue is formed along the exit ramp, which spills into the main
arc and reduces its throughput.
TLP4: This is a TLP caused by an accident which reduces the

throughput of the arc.
TLP5: This is the effect of Rubber-necking. Frequently, drivers

observing an unusual event, such as an accident on the opposing lanes of a
multilane highway, slow down to observe the incident. In doing so, they
reduce the throughput of their own arc, thus causing the formation of a TLP
TLP6: This is the effect of a Moving or Phantom Bottleneck, described

in a previous paper, (Gazis and Herman, 1992). Such a bottleneck may be
formed, for example, by a truck climbing along an uphill portion of a
highway lacking a slow traffic lane. It becomes a phantom bottleneck when
the truck reaches the peak of the uphill portion and resumes normal speed,
thus restoring the normal capacity of the arc. However, the effect of the
temporary TLP may last for some time after the formation of a TLP6.
TLP1, TLP2, and TLP3 are fixed TLPs, easily identified in a network.
TLP4 and TLP5 may be formed at any point in a network, and require good
surveillance for their timely discovery. Finally, TLP6 may be an identifiable
location, but with a possible intermittent TLP function.
The consequence of the existence of a TLP for a network arc is that exact
behavior of the flow versus concentration along the arc is more likely that
shown in Fig. 11 rather than Fig. 9. Flow does build up from zero to a
maximum with increasing density, but never falls down toward zero for
higher densities. Instead, it generally falls to a value lower than the
maximum flow, but still substantial, and equal to some Saturation Flow at
the TLP. The only instances in which the throughput of a network arc goes
to zero are cases of severe incidents or gridlocks at poorly managed
intersections, which shall be treated as exceptions to the rule.
Traffic movement along an arc is again frequently modeled today on the

assumption of the existence of a Fundamental Diagram. A related modeling
practice is the assumption of an Impedance Function for an arc, where
Impedance is understood as a reduction of the capacity of the arc with
increasing density. The following observations can be made regarding such
modeling practices:
1) The use of an Impedance Function must be made with due

consideration to the uni-directional nature of interactions between vehicles.
A vehicle influences only those vehicles behind it, and not the ones in front
of it. Thus, the introduction of a heavy pulse of traffic into an arc, does not
alter the impedance of the whole arc instantly. If an Impedance Function is
to be used, it should be limited to a relatively small region behind the traffic
pulse.
2) Once more, it should be remembered that it is wrong to assume that

the throughput of an arc goes to zero, even when vehicles are queued
bumper-to-bumper along the arc. There will always be a residual throughput
of an arc, equal to the throughput at its TLP.
234 Chapter 4
Improved modeling of the movement of traffic along an arc during

periods of congestion may be achieved as follows:
1) At low densities, the movement of individual vehicles is essentially

unimpeded, with vehicles moving at their desired speed, or the speed limit.
This behavior of traffic continues, up to a density approximately equal to
that corresponding to the maximum throughput at the TLP.
2) When the input into the arc starts exceeding the throughput of the
TLP, queueing ensues, and traffic begins to assume a stop-and-go behavior.
If congestion persists to the point of completely filling the link with
relatively stopped traffic, then the influence of the TLP extends to the
upstream section as well, imposing the same limitation on throughput. In that
domain of densities, we shall model the movement of individual traffic units
as follows:
An individual vehicle is assumed to be "served" on a First-In-First-Out

(FIFO) basis. This means that, on the average, vehicles do not clear the
TLP until all vehicles between the entrance to the arc and the TLP at the
time of the vehicle's entrance are served. While some aggressive drivers
may maneuver for improved service, the assumption of FIFO service
satisfies the requirement of conservation of traffic units, and yields
satisfactory results in the subsequent network optimization procedure.
We compute a Virtual Trajectory for each vehicle which assumes that
the vehicle moves unimpeded at constant speed, equal to its desired
speed or speed limit, up to the TLP. It is stored there until all the
vehicles ahead of it between the arc entrance and the TLP, at the time of
the vehicle's entrance, clear the TLP.
The analytical representation of the movement of vehicles along an arc is

as follows:
Let us denote by (ij) the arc of a network, from node i to node j. Let us
then consider a vehicle entering a link ij at time t and traveling toward j
with k as the next node of its travel path, (Fig. 5). Let us also assume that at
time t there are vehicles along the link ij with next destination k.
We estimate the departure time of the entering vehicle, from node j

toward node k, on the basis of the following assumptions:
There is a normal travel time, from node i to node j. This time

corresponds roughly to the speed limit along the link ij, and is attainable
in the absence of obstructions from traffic. There is a maximum

discharge rate, from link ij toward k, which is achieved as long as
there is a sufficient supply of vehicles in the link ij.
Vehicles proceeding along ij toward k are served on a First-In-First-Out
(FIFO) basis, and there is no appreciable obstruction of the traffic
movement by slow vehicles. (Any such obstruction is treated as an
incident).
On the basis of these assumptions, we estimate the Virtual Trajectory of

the vehicle as follows:
The vehicle proceeds from i to j in time

If
the vehicle's path is not blocked. The vehicle crosses j and proceeds
toward k without any delay.
If the inequality (1) is not satisfied, the vehicle joins a virtual queue at j
and waits for a period of time
236 Chapter 4
before being discharged toward k. The first term of the right-hand side of
Eq. 65 is the time required to discharge all the vehicles which are in the
link ij when the subject vehicle enters this link. Equation 65 simply
states that this time is exactly equal to the time during which the entering
vehicle must stay within the link.
In the preceding discussion, we have treated the node j as the TLP of the
arc (ij), a reasonable assumption in most cases. The estimation of the
vehicle's trajectory in effect ignores the detailed distribution of any delay
incurred by the vehicle along the link ij. The vehicle is assumed to proceed
from i to j at normal speed, and is stored at j in a storage "bin" of some sort
until it is free to proceed toward k on a FIFO basis. Of course, care should be
taken to ensure that the total storage capacity of the bin, for all traffic
streams queued at j, does not exceed the physical storage limit of the
roadway along ij. If it does, constraints must be imposed on the entrance of
vehicles at i and, as mentioned already, the upstream link must be viewed as
regulated by the same TLP.
The estimate of the travel time along different routes can now be
completed by combining current information about the queues and densities
at various points of the network with historical data concerning expected
demand along certain routes. This may be done by an individual optimizing
a route choice, or by a Traffic Management Center (TMC) aiming at the
global optimization of the system. In the latter case, attainment of global
optimization may be attained only if individuals follow instructions of the
TMC regarding their routes.
Optimization of the network operation by the TMC requires sufficiently

accurate estimates of the counts of vehicles within each link. Algorithms for
the estimation of such densities have been given in Chapter 3. Once the
densities along the network links are known, travel times to various
destinations can be computed, and adjustments to route assignments may be
undertaken, aimed at reducing the aggregate delay to all drivers. The
optimal control of the network is reduced to assignment of routes to different
drivers, and management of the allocation of downstream TLP throughput to
converging upstream arcs. As shown by Gazis (1974), for any given
assignment of routes, the allocation of TLP throughput is reduced to a
parametric Linear Programming formulation. By iterating with varying
choices of routes, we may then converge toward global optimization of the
network.
4.5.1 Delivery of Traveller Information Services
One of the key elements of Intelligent Transportation Systems (ITS),

which are diligently pursued around the world, is the delivery of Advanced
Traveler Information Services, (ATIS). On the basis of the preceding
discussion of Section 4.5, we can make some observations regarding the
quality and content of ATIS, which can be achieved. Gazis (1995) has
introduced the concept of Service Channels (SC) for the users of a traffic
network. A Service Channel is a time and space diagram of the movement of
a user’s vehicle through the network such as shown in Fig. 13.
A Service Channel is influenced by the contention between drivers for

roadway space, which causes congestion, queueing and delays. Ideally, a
traffic system should be operated in a way that minimizes such contention,
and optimizes the system according to objectives which should include the
cost of delays but other considerations such as the ability of the system to
deliver services according to the desires and needs of the users.
Having the capability of estimating the travel time along any given route,
and real-time adjustments to this estimate with changing traffic conditions,
we may now address the question of delivery of this information to the users.
238 Chapter 4
The early projections for ITS deployment intrinsically called for special
instrumentation of vehicles to receive traveler information from a TMC.
Early pilot projects were also associated with opinion surveys on price
acceptability for such instrumentation by the driving public. As of the time
of the writing of this book, The results have been less than stellar. The public
does not appear to be very convinced regarding the effectiveness of ITS
services in general, and does not seem willing to pay much to obtain such
services.
Gazis (1998) has suggested that we might be able to accelerate

deployment of ITS services by utilizing instrumentation already available,
namely, the personal computers owned by many drivers and the availability
of the internet. The internet has already been used to disseminate traffic
information, but the dissemination has been in the form of a general
description of traffic conditions in a network. In reference 4, we have
promulgated the use of Intelligent Agents in order to provide personalized
information regarding one’s travel. The scenario is as follows:
An Intelligent Agent residing on a user’s computer will have access to

information regarding traffic conditions, both historical and real time. Before
the start of a trip, the agent will estimate travel times along different routes,
possibly with various starting times, and give the user pre-trip planning
information on the best available option. The agent may also include modal
alternatives, such a public transportation and ride sharing, in the presentation
to the user. In doing so, it will try to improve the user’s Service Channel
without exerting any influence on the system as a whole. The activities of
the Intelligent Agent can start hours before the start of the trip and be
available to the user at home through a personal computer. The agent can
continue monitoring the traffic conditions during the trip and transmit en
route instructions directing recommended changes. The transmission can be
over a telephone line, in the form of voice instructions over a cellular phone,
unless the user carries a notebook in the car which can carry the Intelligent
Agent’s operations on board.
In the long run, Intelligent Agents such as those described above may
interact with the intelligence of a TMC in order to achieve a global
optimization of the operation of a network .
References
Beckman, M. J. and Golob, T. F. (1972), “Critique of Entropy and Gravity in

Travel Forecasting”, Traffic Flow and Transportation, edited by G. F. Newell, New York:
Elsevier, pp. 109-117.
Bellman, R.( 1958), “On a Routing Problem,”, Quart. Appl. Math., XVI, 87-90.
Ben-Akiva, M. E. (1974), “Structure of Passenger Travel Demand Models”,
Transportation Research Record, 526.
Ben-Akiva, M. E. and Lerman, S. R. (1985), Discrete Choice Analysis: Theory and
Application to Travel Demand, Cambridge: MIT Press.
Ben-Akiva, M. E. and Bolduc, D. (1996), “Multinomial Probit with a Logit Kernel
and a General Parametric Specification of the Covariance Structure”, presented at the 3rd
International Choice Symposium.
Ben-Akiva, M. E. and M. Bierlaire, M. (1999), “Discrete Choice Methods and Their
Applications to Short Term Travel Decisions”, in Handbook of Transportation Science,
Chapter 2, Boston: Kluwer Academic Publishers.
Berman, B. R., Chinitz, B. and Hoover, E. M. (1961), Technical Supplement to the New
York Metropolitan Regional Study, Cambrdge, MA: Harvard University Press.
Bierlaire, M., Lotan, T. and Toint, P. L. (1997), “On the Overspecification of
of Multinomial and Nested Logit Models due to Alternative Specific Constants”, Transp. Sci.
31,363-371.
Carey, M. (1986), “A Constraint Qualification for a Dynamic Traffic Assignment
Model”, Trans. Sci., 20, 55-58.
Carey, M. (1987), “Optimal Time-Varying Flows on Congested Networks”, Oper.
Res., 35, 58-69.
Cascetta, E. (1989). “A Stochastic Process Approach to the Analysis of Temporal
Dynamics in Transportation Networks”, Transp. Res. 23B, 1-17.
Damm, D. (1983), “Theory and Empirical Results: A Comparison of Recent Activity-
Based Research”, in Recent Advances in Travel Demand Analysis, edited by S. Carpenter and
P. Jones, London: Gower.
Dijkstra, E. W. (1959), “A Note on Two Problems in Connection with Graphs”,
Numer.Math. 1, 269-271.
Florian, M. and Hearn, D. (1999), “Network Equilibrium and Pricing”, in Handbook of
Transportation Science, edited by R. W. Hall, Boston: Kluwer Academic Publishers, pp. 361-
393.
Ford, L. R. and Fulkerson, D. R. (1962), Flow in Networks, Princeton: Princeton
University Press.
Friesz, T. L., Bernstein, D., Smith, T. E., Tobin, M. L. and Wie, B. W. (1993), “A
Variational Inequality Formulation of the Dynamic Network User Equilibrium Problem”,
Oper. Res. , 4, 179-191 .
240 Chapter 4
Gazis, D. C. (1974), “Modeling and Optimal Control of Congested Transportation

Systems”, Networks, 4,113-124.
Gazis, D. C. and Herman, R. (1992), “The Moving and 'Phantom' Bottlenecks”, Transp.
Science, 26, No.3, 223-229.
Gazis, D, C. (1995), “Congestion Abatement in ITS through Centralized Route Allocation”,
IVHS Journal, 2, 139-158.
Gazis, D. C. (1998), PASHAs: Advanced Intelligent Agents in the Service of Electronic
Commerce, in The Future of the Electronic Marketplace, edited by D. Leebaert, MIT Press,
145-171.
Gumbel, E. J. (1958), Statistics of Extremes, New York: Columbia University
Press.
Hansen, W. G. (1960), “Land Use Forecasting for Transportation Planning”, Highway
Research Board Bulletin, 253, 145-151.
Helly, W. (1969), “A Time-Dependent Intervening Opportunity Land Use Model”, Socio-
Economic Planning Sci., 3, 65-73.
Helly, W. (1974), “Traffic Generation, Distribution, and Assignment”, Traffic Science,
edited by D. C. Gazis, New York: John Wiley & Sons, Chapter 4.
Hirsh, M, Prashkea, J. N. and Ben-Akiva, M. (1986), “Dynamic Model of Weekly Activity
Pattern”, Transp. Sci., 20, 24-36.
Ho, J. K. (1980), “A Successive Linear Optimization Approach to the Dynamic
Traffic Assignment Problem”, Trans. Sci., 14, 295-305.
Jones, P. (1983), “The Practical Application of Activity Based Approaches in Transport
Planning: An Assessment”, in Recent Advances in Travel Demand Analysis, edited by S.
Carpenter and P. Jones, London: Gower.
Luce, R. (1959), Individual Choice Behavior: A Theoretical Analysis, New York:
J. Wiley and Sons.
McFadden, D. (1978), “Modeling the Choice of Residual Location”, in Spatial
Interaction Theory and Residential Location, Amsterdam: North Holland, pp. 75-96.
McFadden, D. (1989), “A Method of Simulated Moments for Estimation of
Discrete Response Models Without Numerical Integration”, Econometrica, 57, 995-1026.
Merchant, D. K. (1974), A Study of Dynamic Traffic Assignment and Control,
PhD Dissertation. Cornell Univ.
Merchant, D. K. and Nemhauser, G. L. (1978a), “A Model and an Algorithm for
the Dynamic Traffic Assignment Problem”, Trans. Sci., 12, 163-199.
Merchant, D. K. and Nemhauser, G. L. (1978b), “Optimality Conditions for a
Dynamic Traffic Assignment Problem"” Trans. Sci., 12,200-207.
Murchand, J. D. (1969), “Bibliography of the Shortest Route Problem”, London
Business Studies Report, LBS-TNT- 6.2.
Pas, E. I. (1982), “Analytically-Derived Classification of Daily Travel-Activity Behavior:
Description, Evaluation and Interpolation”, Transp. Res. Board.
Potts, R. B. and Oliver, R. M. (1972), Flows in Transportation Networks, New York:
Academic.
Putman, S. H. (1966), “Analytic Models fr Implementing the Economic Impact Studies for
the Northeast Corridor Transportation Project”, 30th National Meeting, Operations Research
Society of America.
Ruiter, E. R. (1967). “Improvements in Understanding, Calibrating, and Applying
the Opportunity Model”, Highway Res. Rec., 165, 1-21.
Schneider, M. (1960), Chicago Area Transportation Study, Vol. 2,1960.
Sherratt, G. G. (1960), “A Model for General Urban Growth”, Management Sciences –
Models and Techniques, New York: Pergamon, Vol. 2, pp.147-159.
Stopher, P. R., Meyburg, A. H. and Brog, W., editors, (1979), Proc.of the 4th Intern. Conf.
On Behavioral Travel Modeling, Grainau, Germany.
Stouffer, S. A. (1940), “Intervening Opportunities”, Amer. Sociological Rev., 5, 845-867.
Swerdloff, C. N. and Stowers, J. R. (1966), “A Test of Some First Generation Residential
Land Use Models”, 45th Highway Research Board Meeting, Washington, DC.
Tversky, A. (1972), “Elimination by Aspects: A Theory of Choice”, Psychological
Review, 79, 281-299.
Voorhees, A. M. (1955), “A General Theory of Traffic Movement”, Proc.Inst.Traffic
Engrs, 46-56.
Wardrop, J. G. (1952), “Some Theoretical Aspects of Road Traffic Research”,
Proc. Inst. Civil Engineers, Part II, 325-378.
Wilson, A. G. (1967), “A Statistical Theory of Spatial Distribution Models”,
Transport. Res., 1, 253-269.
Yagar, S. (1970), Analysis of the Peak Period Travel in a Freeway-Arterial
Network, PhD Dissertation, Univ. of California, Berkeley.
Yagar, S. (1971), “Dynamic Traffic Assignment by Individual Path Minimization
and Queueing”, Trans. Res., 5, 170-196.
Wie, B. W. (1991), “Traffic Assignment, Dynamic”, in Concise Encyclopedia of Traffic &
Transportation Systems, edited by M. Papageorgiou, Oxford: Pergamon Press, pp. 521-524.
Wu, J. H., Chen, Y. and Florian, M. (1997a), “The Continuous Dynamic Network Loading
Problem: A Mathematical Formulation and Solution Method”, Transp. Res. 32B, 173-187.
Xu, Y., Wu, J. H., Florian, M., Marcotte, P. and Zhu, D. L. (2002), “Advances in the
Continuous Dynamic Network Loading Problem”, Transp. Sci., (in press).
Additional references
Aashtiani, H. Z. and Magnanti, T. L. (1981), “Equilibria in a Congested Transportation

Network”, SIAM J. of Algebraic and Discrete Methods, 2, 213-226.
Arezki, Y. and Van Vliet, D. (1990), “A Full Analytical Implementation of the
PARTAN/Frank-Wolfe Algorithm for Equilibrium Assignment”, Transp. Sci., 24, 58-62.
242 Chapter 4
Beckmann, M. J. and Golob, T. F. (1972), “A Critique of Entropy and Gravity in Travel

Forecasting”, Traffic Flow and Transportation, edited by G. F. Newell, New York: Elsevier,
pp.109-117.
Dafermos, S. (1972), “The Traffic Assignment Model forMulti-Class Transportation
Network”, Tramp. Sci, 6, 73-87.
Dafermos, S. and A. Nagurney, A. (1984), “Sensitivity Analysis for the Asymmetric
Network Equilibrium Problem”, Math. Programming, 28,174-184.
Dow, P. and Van Vliet, D. (1979), “Capacity Restrained Road Assignment”, Traffic Eng.
and Control, 261-273.
Evans, S. P. (1976), “Derivation and Analysis of Some Models for Combining Trip
Distribution and Assignment”, Transp. Res., 10, 37-57.
Florian, M. and Nguyen, S. (1976), “An Application and Validation of Equilibrium Trip
Assignment Methods”, Transp. Sci., 10, 379-389.
Florian, M. and Spiess, H. (1983), “On Binary Mode Choice/Assignment Models”,
Transp. Sci., 17, 32-47.
Florian, M., Guelat, Y. and Spiess, H. (1987), “An Efficient Implementation of the
PARTAN variant of the Linear Approximation for the Network Equilibrium Problem”,
Networks, 17, 319-339.
Florian, M. and Hearn, D. (1995), “Network Equilibrium Models and Algorithms”,
Chapter 6 in Handbook in OR and MS, Vol. 8, M. Ball et al, editors, pp.485-550.
Fratar, T. J. (1954), “Forecasting Distribution of Interzonal Vehicular Trips by Successive
Approximations”, Proc. Highway Res. Brd., 33, 376-385.
Friedrich, P. (1956), “Die Variationsrechnung als Planungsverfahren der Stadt und
Landesplanung”, Schriften der Akademie fur Raumforschung und Landesplanung, Brenen-
Horn: Walter Dorn, Vol. 32.
Hall, M, (1978), “Properties of the Equilibrium State in Transportation Networks”,
Transp. Sci., 12, 208-216.
Hamburg, J. R. and Sharkey, R. H. (1961) “Land Use Forecast”, Chicago Area
Transportation Study, Chicago, 3.2.6.10.
Harmann, F. V. (1964), Population Forecasting Methods, Washington, DC: Bureau of
Public Roads.
Harris, B. (1962), “Linear Programming and the Projection of Land Users”, Penn–Jersey
Transportation Study, PJ paper No. 20.
Helly, W. (1965),“Efficiency in Road Traffic Flow”, Proc. 2nd Intern, Symp. on Theory of
Traffic Flow, Paris: OECD, pp.264-275.
Herbert, J. D. and B. H. Stevens, B. H. (1960), “A Model for the Distribution of
Residential Activity in Urban reas”, J. Regional Sci. Assoc.
Hoshino, T. (1972), “Theory of Traffic Assignment to a Road Network”, Traffic Flow and
Transportation, edited by G. F. Newell, New York: Elsevier, pp. 195-214.
Knight, F. H. (1981), “Some Fallacies in the Interpretation of Social Costs”, Quarterly J.
of Economics, 38, 306-312.
Lan, T. and Newell, G. F. (1967), “Flow Dependent Traffic Assignment in a Circular

City”, Transp. Sci., 1, No. 4, 318-361.
Leblanc, L. J., Morlok, E. K. and Pierskalla, W. P. (1975), “An Efficient Approach to
Solving the Road Network Equilibrium Traffic Assignment Problem”, Transp. Sci., 5, 309-
318.
Nguyen, S. (1976), “A Unified Approach to Equilibrium Methods for Traffic
Assignment”, in Traffic Equilibrium Methods, Proceedings 1974, edited by M. Florian, vol.
118. Lecture notes in Economics and Mathematical Systems, New York: Springer Verlag, pp.
148-182.
Papageorgiou, M. (1990), “Dynamic Modeling, Assignment, and Route Guidance in
Traffic Networks”, Transp. Res., 24B, 471-495.
Patriksson, P. (1983), The Traffic Assignment Problem: Models and Methods, VNU
Science Press, p. 223.
Powell, W. B. and Sheffi, Y. (1982), “The Convergence of Equilibrium Algorithms with
Predetermined Step Sizes”, Transp. Sci., 16, 45-55.
Sheffi, Y. and Powell, W. B. (1982), “ An Algorithm for the Equilibrium Assignment
Problem with Random Link Times”, Networks, 12, 191-207.
Smeed, R. J. (1961), The Traffic Problem in Towns, Manchester: Manchester Statistical
Society.
Smith, M. J. (1979), “Existence, Uniqueness and Stability of Traffic Equilibria”, Transp.
Res., 13B, 295-304.
Smith, M. J. (1993), “A New Dynamic Traffic Model and the Existence and Calculation of
Dynamic User Equilibria On Congested Capacity Constrained Road Network”, Transp. Res.,
27B, 49-63.
Swerdloff, C. N. and Stowers, J. R. (1966), “A Test of Some First Generation Residential
Land Use Models”, 5th Highway Res. Brd. Meeting, Washington, DC.
U. S. Department of Commerce, Bureau of Public Roads, Traffic Assignment Manual,
Washington, DC, 1964.
U. S. Department of Commerce, Bureau of Public Roads, Highway Capacity Manual,
Washington, DC, 1965.
Voorhees, A. M. (1955), “A General Theory of Traffic Movement”, Proc. Inst. Traffic
Engrs., 46-56.
Wie, B., Tobin, R., Bernstein, R. and Friesz, T. (1995), “Comparison of System Optimum
and User Equilibrium Dynamic Traffic Assignment with Schedule Delays”, Transp. Res., 3C,
389-411.
Appendix
Application of Kalman Filtering for Density Estimation in

Traffic Networks
We shall outline here the application of the methodology of the Discrete-

Time, Extended Kalman Filtering for estimating the densities in various links
of a traffic network. The application of this methodology is useful in systems
where surveillance is limited to detection of passage of vehicles over
detectors placed at discrete points along the roadway, for example, at the
entrance and exit points of a network link. Assuming that one can initialize
the state of the system through knowledge of the count of vehicles in each
section at some point in time, the counts may be updated subsequently by
adding the number of vehicles entering a section and subtracting the number
of vehicles exiting. However, most detectors are subject to errors in
measuring the exact number of vehicles passing over them, and the
cumulative error may cause a substantial deviation from the real count. The
Kalman Filtering methodology may be applied in order to substantially
reduce such errors. In particular, the “Discrete-Time, Extended Kalman
Filter1” has been proposed by Szeto and Gazis2 for estimating densities in a
“discrete-time control system”, such as a traffic network, which is
monitored and controlled at regular time intervals.
The Extended Kalman Filter is particularly appropriate for estimating the

state variables in a nonlinear (noisy) system with nonlinear (noisy)
observations. The algorithm re-linearizes the dynamics about each new state
variable estimate, thus not allowing estimation errors to propagate through
246 Appendix
time. It does so by combining two essentially independent estimates of the

state variables, one obtained from the “state equation”, and another one
obtained from an “observation equation”, both subject to errors which are
uncorrelated. The resulting estimates correspond to a minimum value of the
integral of the squared deviation from true values of the state variables.
Let k = 0, 1, 2, .... Denote a time index, where k = 0 corresponds to a

starting time. Consider a discrete-time, nonlinear, and possibly time-varying
system, whose state vector is an n-dimensional vector Assume
that the state changes according to the stochastic difference equation
where is the error (noise). Let us suppose that independent observations

provide a measurement vector which is nonlinearly related to
the state vector, and has the form
where is also a noise term. Let us assume that the state non-linearity,
and the output non-linearity,
are known, continuous, and sufficiently differentiable. Also assume that the
initial state vector is an n-dimensional random variable with known
mean
Appendix 247
and covariant matrix
where primes denote vector transposition. Now, assume zero-mean,

uncorrelated noise in both the state and the observation equations, leading to
for all k . Furthermore, assume that are mutually independent

for all k , j .
We shall use the following notation:
= estimate of the state vector based on the observations of the

variables
= predicted estimate of the vector based only on the

observations (i.e., before the measurement is made),
248 Appendix
If the error vectors have zero mean, then the second moment matrices,
are the covariant matrices of the state estimation error. Now, the discrete-
time extended Kalman Filter algorithm is best explained by decomposing it
into three different steps:
(a) Initialize :
Start the algorithm by setting
(b) Predict :
Generate according to
where
(c) Update :
Generate according to
Appendix 249
where is known as the “Kalman gain matrix”, and is given by
The algorithm is recursive in nature. Starting even with an initial guess at

the state variables, it generates a new state estimate each time an observation
vector becomes available, moving towards constant improvement of the
estimates.
Let us consider a network link shown in Fig. 1, with length L, and sensors
at the entrance and exit capable of recording the number of vehicles passing
over them and the speed associated with each vehicle. Knowing that the
count measurements are subject to error, we wish to use the speed and count
measurements in order to generate, at discrete intervals, improved estimates
of the number of vehicles within the link.
250 Appendix
Let denote the number of vehicles in the link at the discrete-time

interval, denote the number of vehicles entering the link between time k
and k + 1, and the number of vehicles exiting the link in the same time
interval. Assuming that the input and output sensor measurements are
subject to error, we have the following relationship:
where is the error (noise) term. Equation 13 is the “state equation” in the
Kalman Filter nomenclature. We now wish to utilize the speed information
in order to produce another equation, the “observation equation”. In order to
do this, we have to translate the speed information into one related to the
count of vehicles in the link, or the traffic density therein.
Let be the exponentially smoothed velocity at time k, at the entrance,

and the exponentially smoothed velocity at the exit of the link. Thus,
and are estimates of the average speed of vehicles entering and
exiting the link, respectively. The space mean speed of vehicles within the
section at time k, may be approximated by an expression of the form
where is a weighting factor. Equation 14 implies that vehicles within the

link attain their velocity through a combination of forward and backward
propagation of speed changes. For moderately heavy traffic, but relatively
uncongested conditions, one may expect a value of say, between
0.5 and 0.7, because of the one-sided nature of traffic flow. It is possible to
dynamically modify the value of in order to take into account the
backward and forward propagation of speed changes in a real application.
The space mean speed is a variable that measures the mean speed of
the entire stream of traffic. In Chapter 1 of this book, we discussed various
models relating such speed measurements to traffic densities. Although the
models discussed there are essentially appropriate for an infinitely long
roadway, they may be used for obtaining an approximation to the link
Appendix 251
density from speed measurements. In general, the stream velocity decreases

with increasing density. For the Lincoln tunnel experiment discussed in
Chapter 3, we found that a reasonable relationship between speed and
density was
where L is the link length and a and b are constant parameters. When the
stream velocity is plotted against the density we obtain a bell-
shaped curve such as that shown in Fig. 2. Again the deviations from the
approximation of Eq. 15 can be accommodated by introducing a noise term,
leading to the “observation rquation”
where is the noise term. The parameters a and b may be regarded as

time-invariant variables defined for each observation period, and
consequently satisfying the relationships
The parameters and may be viewed as additional state variables

of the system.
252 Appendix
We now observe that Eqs. 13 and 17 form a set of “state equations” of

the form of Eq. 1, and Eq. 16 is an “observation equation” of the form of Eq.
2. The state variables are the nuber of vehicles in the link, and the
parameters and The observed variable is the stream velocity
which is generated from the sensor measurements according to Eq. 14, and
leading to an “observed” value for the density, or count, of vehicles within
the section, which can be obtained from Eq. 16. The system is driven by the
counts of vehicles entering and exiting the link, plus the noise in both the
state and the observation equations. We can now obtain the state variable
estimates by using the extended Kalman Filter described above. The
computations are lengthy but straightforward, and they are not given here.
The application of the Kalman Filter was tested against the data obtained
from the Lincoln Tunnel experiment3 described in Chapter 3. As indicated in
that discussion, very accurate values for the densities were available for each
section of the tunnel through identification of individual vehicles or
combinations of vehicles as they passed over successive sensors. These
accurate values were used in order to test the power of the Kalman Filtering
approach, which was found to be more than adequate for most situations.
Figure 3 shows a comparison of actual vehicle count data, for a single tunnel
section, with counts estimated by Kalman Filtering.
The effectiveness of the Kalman Filter approach in estimating traffic

counts and densities depends strongly on the reliability of the “observation”
data. In the preceding discussion, the observation data were obtained on the
basis of a phenomenological relationship between speed and density, and a
simple averaging of speed data at the two ends of the link. Such a procedure
is likely to succeed as long as traffic is moving along the link, but it is likely
to fail when congestion sets in and a number of cars are “parked” in a queue
Appendix 253
and waiting for service. In such a case, neither the averaging of the speed nor
the phenomenological relationship are reliable. However, one may seek
alternative “observation” data for the Kalman Filter. As an example, let us
consider the case where there is some partial identification of some vehicles,
providing a handle for estimating the count of vehicles within the link.
The identification of some vehicles may be obtainable through the use of

RFID (Radio Frequency ID) devices, such as those used for electronic toll
payment. The RFID devices can identify individual vehicles at the entrance
and exit of a ramp if we simply install receiving antennas at those points.
These “tagged” vehicles can then be used as markers, whose passage over
the entrance and exit detectors define a time segment during which the
number of vehicles entering the link can be measured fairly accurately. To
be sure, we now in general will have to deal with multi-lane rather than
single lane traffic. However, we can use the information from the tagged
vehicles for estimating densities in the following ways.
1) At low densities, we can rely on the standard Kalman filtering

methodology, using speed-flow relationships for a second estimate of the
density. Alternatively, we can use the following computation. Let us assume
that the tagged vehicle covered the distance from the origin i to the
254 Appendix
destination j of the link in time (Fig. 4). This corresponds to an

average travel speed given by
where is the length of the link.
During the time of the passage of the tagged vehicle, we measure the
entrance speeds of all the entering vehicles and compute their average,
This defines a “spread” of the aggregate of these vehicles equal to
The quantity will be larger or smaller than depending on

whether or not the tagged vehicle is slower or faster than the platoon
entering behind it. It is plotted as larger in Fig. 6, implying that the tagged
vehicle is slower. We can then use as a second estimate for the arc density
the quantity
where is the number of vehicles in the platoon. It may be pointed out

that in the case of low densities, the moving average of the travel time across
an arc of several vehicles may be used in computing travel times along
different routes of the network.
Appendix 255
2) As the density increases and congestion sets in, one may assume that
passing becomes relatively difficult and any tagged vehicle moves
together with the rest of the vehicles. In this case, the travel time of any
tagged vehicle is an accurate measure of the travel time of most of the
vehicles entering the link. We can simply count all the vehicles entering
the link from the time a tagged vehicle enters to the time it leaves the
link, and use this count as the observation count.
The preceding discussions treated traffic links as isolated ones, and

therefore any errors in measurement of the passage of cars over entrance and
exit sensors were assumed as independently generated. In fact, links are
generally in tandem, and errors of connected links are correlated, since the
output error for an upstream link will be the input error of the downstream
link. This observation may be used to further improve the density estimation.
The Kalman filtering methodology may be revised to include the correlation
of errors for arcs in tandem, and adjustments to the final estimates of
densities can be made on arguments of correlation of errors. An improved
Kalman Filtering approach utilizing this correlation of errors has been
proposed by Gazis and Liu4. Preliminary results, based on actual and
computed values of traffic counts for Freeway 41, in Fresno, California,
indicate a substantial improvement of the accuracy of count estimates when
the correlation of errors is explicitly considered.
256 Appendix
References
1. Gazis, D. C. and Foote, R. S., 1969, “Surveillance and Control of Tunnel Traffic by an
On-Line Digital Computer”, Trans. Sci., 3, 255-275.
2. Jazwinski, A. H., 1970, Stochastic Processes and Filtering Theory, New York:
Academic Press.
3. Szeto, M. W. and Gazis, D. C., 1972, “Application of Kalman Filtering to the
Surveillance and Control of Traffic Systems” , Trans. Sci., 6, 419-439.
Index
Acceleration bottleneck 12 Delay at a traffic signal 77
Acceleration noise 31 Arrival process 77
Accessibility model 193 Departure process 78
Adaptive control Delay evaluation 79
Single intersection 106 Delay to a single car 50
System of intersections 132 Delay to pedestrians 63
Arrival process, Poisson 64 Density estimation 175
Assignment, all-or-nothing 211 Departure process 78
Time-dependent 193 Deterministic traffic assignment 207
Time-independent 193, 206 Discrete choice stochastic models 211
Domain, control 151, 157
Bandwidth of progression design 109 Dynamic Traffic Assignment 225
Bang-bang control 144
Bellman’s principle of optimality 127 Economic forecasts 191
Boltzman model 17 Entropy, minimum 199
Bottlenecks 12 EQUISAT strategy 134
Moving 14
Flow-concentration relationships 9
Car-following models 21 Flying start 62
Non-linear models 33 Forecasts, economic 191
Centroids of networks 187 Land use 192
Combination method 121 Freeway control 172
Complex oversaturated systems 154 Fundamental diagram 5
Computer control of traffic 102
Control domain, admissible 151 Gap acceptance 49
COP program 135 Gaps between queues 48
CRONOS program 135 Generation of trips 197
Crossing a multi-lane highway 59 Gravity model 200
Crossings with push-button control 70 Growth factor models 199
Cumulative arrival 140
Cumulative service 140 Heuristic control algorithms 128, 136
Cycle time variation 108 Hudson River tunnel 173
Hydrodynamic model 9
258 Index
Hysteresis effect 12 Oversaturated systems 138
Input-output model 192 Paradoxes

Instability in car-following 29 Smeed paradox 36
Interaction process in driving 18 Braess paradox 38
Intervening opportunity model 202 Pedestrian crossings 63
Phantom bottleneck 14
Kalman Filtering estimation of densities 245 Pontryagin maximum principle 150
Kinematic waves 2 Population forecasting 191
PROBIT models 217
Labeling strategy in networks 189 PRODYN program 135
Land use accounting 192 Progression design 107
Leontief input-output model 192
Limit cycle of traffic light 130 Queueing and delays of vehicles 71
Lincoln tunnel, control of 173 Qeueuing of pedestriaqns 63
LOGIT models 215 Queueing theory 45
Macrocontrol 104 Ramp, spill-bacck from an exit 161

Macroscopic models 1 Random Utility models 213
Markov chain 63 Relaxation process in driving 17
Mean delay Route algorithm, minimum cost 208
For Poisson input 81
For more general arrivals 84 Saturation flow 105
For random departure times 65 SCOOT program 135
Merging 71 Shock waves 6
Microcontrol 104 SIGOP program 124
Microscopic models 21 Single, isolated intersection 105
Minimum cost assignment 223 Single oversaturated intersection 138
Minimum entropy formulation 204 Space-time trajectories 2
Mixed Integer-Linear Programming Speed versus concentration 5
approach to progression 111 Spider web network 187
Moving bottleneck 14 Spill-back from an exit ramp 161
Stability of car-following models
Network representation of a Local stability 22
transportation system 186 Asymptotic stability 26
Networks, spider web 187 Store-and-forward networks 166
Network simulations 192 Synchronization schemes 106
Systems affected by geometry 173
Objectives of traffic control 103
OPAC program 135 Through-band maximizing design 112
Optimal traffic signal settings 87 Throughput Limiting Points (TLP) 166, 231
Oversaturation, onset of 136 Time-independent traffic assignment 206
Index 259
Traffic-actuated lights 128

Traffic assignment in congested systems 229
Traffic characteristics 46
Traffic control, evolution of 102
Traffic flow models
Microscopic 1, 2
Microscopic 1, 21
TRANSYT method 118
Travel cost minimization 225
Traveler information services 237
Trip distribution 198
Trip generation 190
Two-fluid model 19
Vehicle-actuated traffic signal 89

traffic queues at 91
Waves, propagation of 8
shock 6
Yield sign, delay at 54
Zebra crossings 69
Zones of network 187

Traffic Theory

Uploaded by

Copyright:

Available Formats

Traffic Theory

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Traffic Theory

Uploaded by

Copyright:

Available Formats

What are some of the traffic flow models discussed in the text?

What are some of the traffic flow models discussed in the text?

What are some techniques for traffic control and optimization mentioned?

What are some techniques for traffic control and optimization mentioned?

TRAFFIC THEORY

KLUWER ACADEMIC PUBLISHERS

©2002 Kluwer Academic Publishers

Print ©2002 Kluwer Academic Publishers

All rights reserved

Created in the United States of America

Visit Kluwer Online at: http://kluweronline.com

CHAPTER 1. TRAFFIC FLOW THEORIES 1

Macroscopic Traffic Flow Theory – Kinematic

The Boltzman-like Model of Traffic 17

Microscopic Traffic Flow Theory – Car-Following Models 21

Some Questionable Paradoxes 36

Additional Flow Theory Work, and Future Challenges 40

CHAPTER 2. QUEUEING AND DELAYS AT

The Gap Acceptance Function 49

The Delay to a Single Car 50

Queueing and Delays of Vehicles 71

The Delay at a Traffic Signal 77

CHAPTER 3. TRAFFIC CONTROL 101

Objectives of Traffic Control 103

Single, Isolated Intersection 105

Synchronization Schemes for Arterial Traffic 106

Traffic Responsive Operation of Traffic Lights 128

The Onset of Oversaturation 136

Oversaturated Systems 138

Oversaturated Store-and-Forward Networks 165

Freeway Control 171

Systems Affected by Geometric Details 172

CHAPTER 4. TRAFFIC GENERATION, DISTRIBUTION,

Network Representation of a Transportation System 186

Trip Generation and Distribution 189

Time-Independent Traffic Assignment 206

Dynamic Traffic Assignment 225

Traffic Assignment in Congested Systems 229

Application of Kalman Filtering for Density

“ Everything should be made as simple

Traffic Theory, like all other sciences, aims at understanding and

In a biblical mood, I opened a 1970 paper entitled "Traffic Control --

Traffic theory, after modelling of traffic movement, moved into

If we were to evaluate the overall effect of progress in traffic theory on

measure of this effect. This is so in spite of considerable effort made in

a) Some good traffic theoretical models are inadequately used.

1. Traffic Flow Theories: A description of the movement of traffic and the

I would like to express my gratitude to Walter Helly, Donald McNeil and

The book is intended as a textbook for a college Transportation Science

Traffic Flow Theories

Traffic flow theories may be viewed as the foundation of Traffic Science.

Traffic flow theories were among the earliest contributions to Traffic

1.1 MACROSCOPIC TRAFFIC FLOW THEORY –

Lighthill and Whitham (1955) laid the foundation of macroscopic traffic

The above quantities satisfy the relationship

At this point, we should discuss how these quantities are measured.

If is the total number of vehicles traversing dx in time T,

considering the vehicles as distributed in time, and remembering that in this

By substituting where is the travel time across dx , and by

Now, assuming a space distribution of vehicles observed on a stretch of