AND
TRAFFIC THEORY
Related Pergamon books
LESORT
Transportation and Traffic Theory: Proceedings of the 13th ISTT
BELL
Transportation Networks: Recent Methodological Advances
DAGANZO
Fundamentals of Transportation and Traffic Operations
GRIFFITHS
Mathematics in Transport Planning and Control
AVISHAI CEDER
Transportation Research Institute
Faculty of Civil Engineering
Technion  Israel Institute of Technology
Haifa, Israel
PERGAMON
An Imprint of Elsevier Science
Amsterdam  Lausanne  New York  Oxford  Shannon  Singapore  Tokyo
ELSEVIER SCIENCE Ltd
The Boulevard, Langford Lane
Kidlington, Oxford OX5 1GB, UK
This work and the individual contributions contained in it are protected under copyright by Elsevier Science, and the following terms and
conditions apply to its use:
Photocopying
Single photocopies of single chapters may be made for personal use as allowed by national copyright laws. Permission of the publisher and
payment of a fee is required for all other photocopying, including multiple or systematic copying, copying for advertising or promotional pur
poses, resale, and all forms of document delivery. Special rates are available for educational institutions that wish to make photocopies for
nonprofit educational classroom use.
Permissions may be sought directly from Elsevier Science Rights & Permissions Department, PO Box 800, Oxford OX5 1DX, UK; phone:
(+44) 1865 843830, fax: (+44) 1865 853333, email: permissions@elsevier.co.uk. You may also contact Rights & Permissions directly
through Elsevier's home page (http://www.elsevier.nl), selecting first 'Customer Support', then 'General Information', then 'Permissions
Query Form'.
In the USA, users may clear permissions and make payments through the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers,
MA 01923, USA; phone: (978) 7508400, fax: (978) 7504744, and in the UK through the Copyright Licensing Agency Rapid Clearance
Service (CLARCS), 90 Tottenham Court Road, London W1POLP, UK; phone: (+44) 171 631 5555; fax: (+44) 171 631 5500. Other countries
may have a local reprographic rights agency for payments.
Derivative Works
Tables of contents may be reproduced for internal circulation, but permission of Elsevier Science is required for external resale or distribution
of such material.
Permission of the publisher is required for all other derivative works, including compilations and translations.
Except as outlined above, no part of this work may be reproduced, stored in a retrieval system or transmitted in any form or by any means,
electronic, mechanical, photocopying, recording or otherwise, without prior written permission of the publisher.
Address permissions requests to: Elsevier Science Rights & Permissions Department, at the mail, fax and email addresses noted above.
Notice
No responsibility is assumed by the Publisher for any injury and/or damage to persons or property as a matter of products liability, negligence
or otherwise, or from any use or operation of any methods, products, instructions or ideas contained in the material herein. Because of rapid
advances in the medical sciences, in particular, independent verification of diagnoses and drug dosages should be made.
ISBN: 0 08 043448 7
©The paper used in this publication meets the requirements of ANSI/NISO Z39.481992 (Permanence of
Paper).
Printed in The Netherlands.
CONTENTS
Preface and Overview ix
International Advisory Committee xiii
In Memoriam xv
Contributors xix
Index 797
PREFACE AND OVERVIEW
These proceedings represent a further step forward in the understanding and solution of
transportation and traffic problems. This is the 14th publication in the series of Symposia on
Transportation and Traffic Theory. We are endeavouring to continue in the tradition of
Professor Robert Herman and a group of scientists who gathered at the General Motors
Research Laboratory in Michigan and initiated this undertaking in 1959. We have all been
inspired by their leadership.
It is with great sadness that we learned that Professor Robert Herman (Honourary Member) and
Professor Michael Cremer (International Advisory Committee) have passed away since our last
symposium. We have included a tribute to them in this book.
This publication goes a long way towards providing innovative, advanced knowledge
regarding traffic and transpiration problems and the analytical tools required to achieve their
solutions. Transportation issues about: Safety, Mobility, Efficiency, Productivity, Planning
and Environmental elements are of direct interest to a growing number of professionals in the
fields of communication, data processing, electronics, environmental quality, policy makers
and others. Every advanced transportation or traffic system, existing or under development,
absorbs its "know how" from models, methods and analyses represented by papers like the
ones in this publication. To name but a few: ATMS (Advanced Traffic Management System),
ATIS (Advanced TravellerInformation System), APTS (Advanced Public Transportation
System), ACVO (Advanced Commercial Vehicle Operations) and ETC (Electronic Toll
Collection), all of which are known as ITS (Intelligent Transportation Systems). ITS can be
seen as a link between technology and the drivervehicleroad system and models.
There is a saying: "A man's real worth is determined by what he does, when he has nothing to
do." The contributors to this publication are among a group of scientists who even when they
have nothing to do, think how to resolve and deal with transportation and traffic problems.
They are following George Bernard Shaw's remark that people look at existing things that do
not work (on transportation and traffic problems) and ask: why? But dream about things that do
not exist, but work and ask: why not?
This book has been compiled to follow the same session order as the Symposium and each of
the ten chapters has been prefaced by three sayings.
An overview of these chapters has been created in the following table. The thirtyfive papers
in this publication represent contributions from sixteen different countries.
Transportation and Traffic Theory
The review process that allowed the selection of .these 35 papers was particularly difficult,
following a twostage international review process. It is for this reason that, for the first time,
22 additional papers will also be published, in a separate bound volume which will be available
at the Symposium. Warmest thanks are due to all the reviewers who completed this arduous
task within a very limited time framework.
Thanks are also due to Mr. J.B. Lesort, who organised the last Symposium, to the
International Advisory Committee (listed separately) and to the Local Programme Committee
(Y. Berechman, Y. Gur, Y. Israeli, T. Lotan, D. Mahalel, A. Mandelbaum, M. Pollatschek,
Y. Prashker, D. Shefer, Y. Shiftan, I. Salomon).
We are indebted to the following main sponsors: the Transportation Research Institute of the
TechnionIsrael Institute of Technology, the Israel Ministry of Transport, the General Motors
Foundation, and the European Commission (DG Transport).
Avishai Ceder
April 1999
INTERNATIONAL ADVISORY COMMITTEE
Prof. E. Hauer University of Toronto, Canada (Convenor)
Prof. R.E. Allsop University College London, UK
Prof. M.G.H. Bell University of Newcastle upon Tyne, UK
Prof. P.H.L. Bovy Delft University of Technology, The Netherlands
Prof. W. Brilon Ruhr University, Bochum, Germany
Prof. A. Ceder Technion  Israel Institute of Technology, Israel
Prof. C.F. Daganzo University of California, Berkeley, USA
Prof. N. Gartner University of Massachusetts at Lowell, USA
Prof. H. Keller Technical University of Munich, Germany
Prof. M. Kuwahara University of Tokyo, Japan
Mr. J.B. Lesort INRETS/ENTPE, Lyon, France
Prof. H. Mahmassani University of Texas at Austin, USA
Prof. Y. Makigami Ritsumeikan University, Japan
Prof. V.V. Silyanov Moscow Automobile and Road Construction Institute, Russia
Prof. M.A.P. Taylor University of South Australia, Australia
Prof. M. Tracz Cracow Technical University, Poland
Prof. S.C. Wirasinghe University of Calgary, Canada
HONOURARY MEMBERS
Institute, and in 1981, they received The New York Academy of Sciences Award in Physical
and Mathematical Sciences.
Dr. Herman was a member of Phi Beta Kappa, Sigma Xi, the Washington Philosophical
Society, and the Royal Institution of Great Britain, and a Fellow of the American Physical
Society, the Washington Academy of Sciences, and The Franklin Institute. He was elected to
the National Academy of Engineering in 1978 for his contributions to the science of vehicular
traffic. In 1979, he was elected a fellow in the mathematical and physical sciences of the
American Academy of Arts and Sciences. He has been an associate editor of the Reviews of
Modern Physics; was one of the founders of the Transportation Science Section of the
Operations Research Society of America (ORS A, now INFORMS), and also became its first
chairman as well as the founding editor of its journal, Transportation Science. He has served
as the President of ORS A (19801981).
In 1959, Dr. Herman was corecipient of the Lanchester Prize in Operations Research for
pioneering research on the stability and flow of singlelane traffic. In 1963, he was awarded an
honorary medal by the Universite Libre de Bruxelles, and during that same year received the
Townsend Harris Medal from the Alumni Association of the City College of New York as a
distinguished alumnus and for his scientific contributions. Other prestigious INFORMS
awards include the George E. Kimball Medal (1976) and the John Von Neumann Theory Prize
(1993). Also in 1993, he received the Roy W. Crum Distinguished Service Award of the
Transportation Research Board for "his pioneering contributions to the field of traffic science".
In 1984, he was awarded an Honorary Doctorate in Engineering by the University of Karlsruhe
in recognition of his outstanding research in the mathematical foundations and development of
the theory of traffic flow. He received the first Lifetime Achievement Award of the Operation
Research Society's (now INFORMS') Transportation Science Section for his body of work on
vehicular traffic science in 1990. The award was subsequently renamed the Robert Herman
Lifetime Achievement Award.
In October and November 1990, The National Academy of Engineering presented the first
public showing of some of Dr. Herman's small abstract sculptures in exotic wood, which he
had produced over a period of about 30 years.
Notwithstanding all these accomplishments, two activities occupied a very special place in
Bob's heart: the ISTTT series and Transportation Science (the journal). In addition to
symbolizing the coming of age of traffic science as a legitimate field of human scientific
inquiry, and providing it with its scientific pillars, they also defined Bob's special family and
community. With us, he shared the love for the work, the ideas, the excitement and the people
engaged in pushing the intellectual frontiers of this field, through theoretical development and
painstaking measurement. The collective effects of ISTTT gatherings always ensured a far
greater impact than the sum of its individual parts. Bob's unbounded human contributions as a
colleague, friend, mentor and educator to virtually every generation that has entered the
Transportation Science domain will endure. As it showcases recent developments in traffic
science and transportation analysis methods, the 14th ISTTT stands as a testimonial to the
house that Robert Herman built, and a celebration of the life and intellectual energy that he so
freely shared with all of us.
Hani S. Mahmassani
Austin, Texas, February 1999
In Memoriam xvii
Hartmut Keller
1998
This page intentionally left blank
CONTRIBUTORS
Abstract The aim of this paper is to propose a methodology for the comparison of first and
second order macroscopic traffic flow models. First, we identify a certain number of difficulties
and investigate how various macroscopic models cope with them. Second, we propose a set of
problems or situations which could be used as a test workbench for the comparison of models.
i INTRODUCTION
Since the first papers by Lighthill and Whitham [1955] and Richards [1956] and the introduction
of higher order models by Payne [1971], a great number of macroscopic traffic flow models have
been developed. They are all based on a few similar variables and assumptions the state variables
are the flow q(x, t), the density k(x, t) and the mean flow speed u(x, t), defined as:
(1) « = kf
The variables q, k and u are considered as piecewise differentiable functions of space x and time
t, still allowing the existence of some singularities (discontinuities in spacetime corresponding
to shock waves, discontinuities in time corresponding to incidents and discontinuities in space
corresponding to variations in the road layout). The conservation equation constitutes the second
basic equation. It can be written as:
+ 0
(2) lr
ox § ot =
To realise a complete model, a third equation is necessary. It can be derived in different ways.
The first one, which results in the basic LighthillWhithamRichards (LWR) model, is to consider
an empirical relationship between speed and concentration:
These are first order models. In various extensions, the equilibrium relationship may depend on
space (variations in the road layout) and time (occurrence of incidents):
(4) u(x,
The important point is that this dependence is exogeneous to the model and not a part of it ue is
a function of the three variables k, x, t. A second way to obtain a third equation is to consider
relation (3) only as an equilibrium relationship, and to describe explicitly transitory states, using
for instance a relaxation process expressing the tendency of traffic to tend to an equilibrium.
Models of that kind may be derived from microscopic considerations, as made by Payne [1971],
resulting in a speed equation:
du du 1/ v dk
with v an anticipation coefficient and r a reaction time. These are higher order models. More
recently, many higher order models have been developed from the kinetic models generalizing
the model proposed by Prigogine and Hermann [1971] . Examples are [Phillips, 1979], [Helbing,
1997], [Kerner et al, 1996]. To develop such models, some simplifying assumptions have to be
made, such as considering a gaussian repartition of individual speeds [Helbing, 1997]. The
resulting equation are somewhat similar to Payne's.
For a long time, numerous experiments have been conducted to compare various classes of mod
els. (e.g [Michalopoulos et al., 1992] and [Papageorgiou et al, 1983, 1989]). Most of these
experiments have concluded to the superiority of higher order models. However it must be no
ticed that most of these experiments had as their first aim to validate a higher order model. For
these experiments, the LWR model discretization and implementation had not been optimized.
On the other hand, the values obtained in the calibration process for parameters as the reaction
time are often far from physical, as pointed out by Del Castillo et al. [1993]. More recently,
comparisons and analysis have been conducted on a more theoretical basis Schochet [1988] has
shown that Payne's model converges towards the simple LWR model when the reaction time
tends to zero. Daganzo [1995b] claimed that higher order models included fundamental flaws,
such as negative speeds, which made them unsuitable for practical use. In reaction, models have
been proposed [Liu et al, 1998], [Zhang, 1998] trying to avoid such flaws. Indepth experimental
analysis have been conducted [Kerner and Rehborn, 1997] to understand how theoretical models
are able to explain observations. In turn, these experimental results are contested [Daganzo et al.,
1997], various explanations being given of the same traffic phenomenons. To date, no definite
conclusion is possible on the respective advantages and drawbacks of each kind of model.
The purpose of this paper is to propose some clarifications to this debate. In order to achieve
this aim, we first identify some specific difficulties, and assess how the various macroscopic
models deal with them, thus achieving some measure of comparison. Further, we propose a set
of archetypical problems of modelling situations, for which we give some elements of compar
ison between models, and on which various models could be tested for performance or simply
modelling capability. Such a comparison procedure is of course quite theoretical, but the re
sults expected from the models ultimately rely on experience. Actually, some of the proposed
archetypes, such as traffic dispersion or congestion formation, are inspired directly by experimen
tal data, whereas others, such as intersection or lane drop modelling, result more from theoretical
or logical considerations.
It has been noticed for long that most models have difficulties in keeping some of the basic
variables describing the traffic flow dynamics within physically sound limits. This is particularly
true for speed and acceleration.
(6)
with T(k] a reaction time depending on k. The other equations of the models are the classical
equations (1) and (2). It is then clear that:
• If an initial condition is u = 0 (locally), then  is positive, which avoids negative speeds,
• On the other hand, consider the following initial conditions, which are those of a stopped
queue, with k = kjam, u = ue(kjam) = 0 for x < 0 and k = 0, u = Vf = ue(0) for x > 0,
with Vf the desired speed (i.e. q = 0 uniformly). Then ^ = 0 (because ue(k) = u for
all x), H = 0 (because initially q = 0 uniformly) and the queue will never start again the
solution is stationary.
It is to date not obvious whether higher order models presenting no risk of negative speed can be
developed.
In first order models, speeds are always physical for they are bounded by 0 and a maximum
equilibrium free speed.
2.1.2 Acceleration
Acceleration is not an intrinsic variable of any macroscopic model. However, it is interesting to
examine whether the acceleration of a vehicle moving with the flow is kept within reasonable
bounds. Indeed, models aiming for instance at the description of sound or pollutant emission by
traffic flow require reasonable acceleration estimates. If w(t) is the speed of such a vehicle, the
acceleration can be written as:
dw _ du du
dt dt dx
In the case of first order models, the possibility of the occurrence of unbounded acceleration
values will be illustrated by two examples. First, let us recall that, if the equilibrium speed
density relationship is dependent on the position x, it follows from the conservation equation and
the speeddensity relationship that:
dw , (duf\ dk ( , due\ dup
dt \ ok I dx \ * dk iI ox
*
6 Transportation and Traffic Theory
dw (due\2 dk
=
(9
) ~JT
at ~k HjT
\dk J ^~
ox
Since the value of  cannot be bounded (through a Shockwave, or when a queue discharges), it
is clear that the acceleration may take any value.
A second instance which may be mentionned is the effect of a variation in the road layout, i.e
ue is a function of x u(x, t) = ue(k(x, t ) , x ) . It is then possible to examine the acceleration of a
vehicle under steadystate flow conditions with ff = 0 and ^ = 0, since the analytical solution
is straightforward. Under these conditions the acceleration of a vehicle can be written:
dw u2 due
This formula results from (8) by noting that the steady state conditions  = 0 = f^ imply
dqe dk dqe
dk dx dx
It can be seen that:
• For a given road layout profile ( given ^jf), the acceleration sign depends on traffic con
ditions (fluid or congested)
• since the derivative ^ is bounded by the free speed, if ^ is very high, there is no physical
bound to the acceleration. What this presumably means in reality is that the actual use of
the road by traffic (actual capacity)is adapted to the acceleration/deceleration capabilities
of vehicles.
The representation of various road layout configurations is one of the problems to be dealt with
by traffic flow models. This problem will be investigated more into details in the next section, but
two questions are related to some theoretical issues: the notion of capacity, and the occurrence
of discontinuities in the road layout.
2.2.1 Capacities
Even if the capacity has a straightforward definition (maximum possible flow), its representation
is not that obvious. The main point of controversy is the concept of dynamic capacity for in
stance, does the maximum flow depend on the presence of a congestion [Papageorgiou, 1998] or
not.
In first order models, the solutions used for the model are entropy solutions which always max
imise the flow. The definition of capacity is thus obvious, since capacity is one of the parameters
of the equilibrium relationship.
On the other hand, if nonentropy solutions are used, as proposed for instance in [Lebacque,
1997] for bounded acceleration models, the notion of capacity corresponds to that of maximum
flow only under steady state traffic conditions. (Actually the flow can never exceed the capacity
value, but it is lower under transitory conditions).
In higher order models, there is a notion of maximum flow in steady state conditions, when the
flow speed is equal to the equilibrium one. On the other hand, outside of these conditions, the
Macroscopic Traffic Flow Models 1
flow may be lower but also higher than this value. Indeed, the speed u and the density k being
independent variables, there is no possibility to control the value of their product q = ku. This is
particularly apparent in Ross's model [Ross, 1988], ^ = £ (vj  u) (with v/ the desired speed),
in which the capacity constraint ku < qmax must be added to the model in order to recapture
such basic features as congestion propagation. For Payne's or similar models, Schochet's results
[Schochet, 1988] imply that such excessive values occur only in limited ranges of space and
time. This last result is actually only valid if the viscosity z/ and relaxation time T parameters
take physical values, i.e. are very small, as proposed in [Del Castillo et a/., 1993]. Otherwise,
convergence towards traffic states satisfying to capacity constraints is slow, or takes place over a
large space scale only.
2.2.2 Discontinuities
By discontinuities we mean mainly either space or time discontinuities of the equilibrium rela
tionships ue and qe. Spacewise discontinuities would be related to abrupt increases or decreases
in capacity. Timewise discontinuities represent accidents or incidents, which imply a more or
less extended capacity restriction for some finite duration. Finally, moving capacity restrictions
might also be considered, such as buses (in urban areas) [Lebacque, Lesort, Giorgi 1998] or slow
moving truck convoys on highways [Newell, 1993 and 1997].
Discontinuities constitute a normal feature of macroscopic models, a necessary consequence of
the continuum hypothesis and of the corresponding approximation. Actually, the size of any
feature smaller than say 50 to 100 meters should be neglected, a remark that applies also to the
unbounded acceleration estimate problem mentionned in subsection 2.1.2.
We shall discuss spacewise discontinuities more in detail in subsection 3.1. A few remarks
are of order here. First order models accommodate such discontinuities well, since boundary
conditions are well defined for these models, see subsection 3.7 the flow at such a discontinuity
is continuous (spacewise) and equal to the minimum between upstream demand and downstream
supply (entropy solution). Since speed is a function of density, a discontinuity of the speed is
generated, implying infinite acceleration and an unrealistic solution, which is flow maximizing,
a point often criticized in the literature [Papageorgiou, 1998]. It might be argued that speed
discontinuities constitute a normal feature of macroscopic models too, i.e. that the velocity
changes abruptly over a range of the same order as the vehicle spacing. Such a velocity gradient is
nevertheless not compatible with the finite acceleration of traffic (and the acceleration capabilities
of vehicles).
Second order models are much better behaved in theory, owing to the damping effect of the
diffusion term (term in £ in the speed equation (5)) only the acceleration would be discontin
uous. The speed should gradually relax towards the equilibrium speed. Nevertheless, it may
well be that this favorable situation actually results from the fact that variations of the equilib
rium relationships have not been considered in the derivation of higher order models. Indeed,
if such variations are considered, as in subsection 3.1.2, supplementary terms in ^ should be
introduced into the speed equation, implying a speeddiscontinuity at the discontinuity of ue.
Timewise discontinuities will not be discussed in detail here, we refer the reader to [Buisson et
al., 1996], [Mongeot, 1997], [Heydecker, 1994]. The difficulties involved are the same, with a
difference which is that the density, and not the flow, is,conserved at the discontinuity. Moving
discontinuities can be treated by considering the moving frame associated to the discontinuity,
a programme which has been carried out in the case of first order models (see the references
cited above), but not, to the authors knowledge, in the case of second order models. It is actually
questionnable whether moving capacity restrictions of small size, such as buses or convoys, can
be modelled at all with second order models, considering that the model equations may not be
able to enforce a small sized capacity restriction through the mechanism of relaxation towards
the equilibrium state.
Transportation and Traffic Theory
There are two possible ways of computing solutions to traffic models either to calculate analytical
solutions, which is not always possible, or to derive a discretized model from the continuous one
and to compute simulation solutions.
This model results from Phillips' model, in which the anticipation term of Payne's model is
replaced by a pressure term
_\&P_
k dx
(as resulting from a kinetical model) by assuming that the traffic pressure V is a linear function
of the density. This model, with addition of a viscosity term ^f^f , and a special choice of the
viscosity coefficient v = ^ and of the equilibrium relationship
became the study object of Kerner and Konhaiiser [Kerner et a/., 1996].
Macroscopic Traffic Flow Models
2.3.2 Simulation
Discretizations have been proposed for most models presented in the literature. Concerning first
order models, various schemes have been proposed, including [Lebacque 1984] (in which the
Godunov flux was introduced in a heuristic way], [Michalopoulos et al., 1984a] (based on a
shockfitting LaxWendroff scheme), [Michalopoulos et al., 1984b], [Bui et al., 1992] (using
Osher's formula for the Godunov flux function), [Leo and Pretty, 1992] (applying Roe's approx
imate formula for the Godunov flux function to several models including the LWR model). More
recently, dicretizations based on the Godunov scheme and presenting a better consistency with
the continuous model have been proposed by [Lebacque, 1996b] and [Daganzo, 1994]. However,
few systematical analyses of the discretization question have been made. An older example is
provided by [Leo and Pretty, 1992] and a more recent example is [Zhang and Wu, 1997].
On the other hand, some models exist only under a discretized form. It is the case of the model
proposed by [Hilliges, 1995], or of some models introducing bounded acceleration (the phe
nomenological model in [Lebacque, 1997]). The main drawback of this kind of model is that
the behaviour of the model is dependent on the discretization parameters (time 8t and space 5x
discretization steps), and that the convergence of the model is not guaranteed when 8x and 6t
tends towards zero. Considering the equivalent equation is usually not helpfull. Furthermore, the
convergences towards zero of 8x and 8t should not be considered independently, as the following
cursory analysis of numerical viscosity effects (and also Schochet's convergence results) show.
When the discretization scheme is consistent with the continuous model, the only effect of the
discretization is to introduce a numerical viscosity into the model behaviour. The effect of this
viscosity is visible on a simple example. Let us consider a simple first order model, under fluid
traffic conditions. Let
• ql(t) and q°(i) be the flows entering and leaving a discretization segment during timestep
[M + A],
• k(t) be the mean density at time t.
Under fluid conditions, and at first order approximation (for example a perturbation of a station
ary state) qr(t) can be considered as given, and q°(t + 6t) written as:
This is actually a smoothing formula (a fact not incompatible with the existence of shockwaves)
which bears some resemblance with the experimental TRANSYT smoothing formula [Robert
son, 1969].
Introducing the maximum speed umax, the viscosity factor J^f can be considered as the product
of two terms:
• A physical and intrinsic term ^— ^ which expresses that the viscosity is a function of
density
• A purely numerical term umaxj^ (The CourantFriedrichsLewy number)
10 Transportation and Traffic Theory
A symmetric formula relating qj(t + 6t) to ? 7 (t) and q°(t) applies in the congested case.
The CourantFriedrichsLewy (CFL) condition 6x > umax5t imposes to this numerical term to
be lesser than 1 in order to guarantee the stability of the model. On the other hand, the lesser
this term, the higher the numerical viscosity (therefore it should be as close to one as feasable in
order to optimize the discretization).
An extreme example is given by some flow models used for dynamic assignment, which can
be considered as first order models continuous in time, with an imposed space discretization
(the link). The basic example is the model proposed by [Merchant and Nemhauser, 1978], from
which a timecontinuous version has been given by [Friesz et al, 1989]. Indeed, a one segment
Godunov discretization of a link yields (assumingthat the downstream supply is sufficient):
with q1 the given inflow, x(t) the total number of vehicles at time t, A e the demand function and
x ( t ) / l the mean density, I the length of the link. Taking 6t —>• 0, the following model results:
d
ft(t}=qI(t}g(x(t}}
with
As explained by [Astarita, 1996] these models present a high degree of numerical viscosity,
resulting in unphysical behaviour (null or infinite traveltimes ...).
Numerical viscosity can be considered as a positive effect, as it looks like a platoon dispersion
effect (this will be developed further on). However, it must be kept in mind that part of this is a
purely numerical effect, with little possibility to fit it to physical observations.
Second order models in simulation have been analyzed carefully, notably by [Papageorgiou et al,
1989], [Michalopoulos et al., 1992] for instance. For lack of sufficient supportive mathematical
results (the theory is not as well developed as that of conservation equations, few analytical solu
tions or convergence results are known), analyses concentrate on stability, parameter calibration
and ability to reproduce satisfactorily observations. The idea of these simulations is to use cells,
cellaverages, and finite difference schemes, an approach similar to the one used with first order
models. The cell dimension is liable to be much greater than in discretized first order models.
Some allowance must be made for special features adding a constant K to the density in the l/k
terms for instance, in order to avoid divisionbyzero problems. The equation q = ku must be
discretized with care too; [Papageorgiou et al., 1989] proposed a linear interpolation formula for
the cell outflow of cell (i) during timestep k, qf, relating it to the mean flows of cells (i) and
(i + 1):
(13) g* = afc*ut* + (1  a)fc?+1u?+1
Flux vector splitting upwind schemes were proposed too [Michalopoulos et al, 1992], and may
yield more rigorous discretizations. At this point, parameter calibration has still an enormous
importance for second order models, which means that discretized second models should be
considered as phenomenological in that sense, and as having an existence in their own right,
irrespective of the continuous model from which they are derived. This is possibly a good thing
corrective features can be introduced into discretized second order models, correcting some flaws
of the corresponding continuous models, such as pointed out in [Daganzo, 1995c] (negative
speeds for instance), possibly introducing capacity bounds too.
The physical soundness of a model can only be checked by comparison to real world observa
tions. This means that
• The model must be able to explain observed physical phenomenons,
Macroscopic Traffic Flow Models 11
• The demand/supply mechanism may result, at a measurement point, and for a given value
of flow, in various density values: oscillation between both fluid and congested states,
mixed into a single observation period, can yield different values of the density, determin
istic scatter and hysteresis, even if the flow is constant;
• The composition of traffic, particularly with respect to the various destinations, which is
never measured, may result in the occurrence of unexplained perturbations, particularly at
diverge points. This has been observed for instance by [Daganzo et al., 1998].
It is only recently that systematic analysis of observations in relation to traffic models have been
conducted. Classical experiments have mainly focused on the calibration/global validation of the
model on a statistical basis. One of the first attempts at an indepth analysis has been made by
[Kerner and Rebhorn, 1997]. Considering their paper and the explanations given of the same and
similar data by [Daganzo et al., 1997], it is interesting to notice how multiple explanations can
be given of the same physical phenomenons, each following its own modelling rationale.
Other phenomenons are much more difficult to explain at, such as the constant size perturbations
observed by Kerner and Rebhorn, which cannot be explained through classical models. It can
be noticed that it would be possible to derive some nonentropy solutions of the LWR model
explaining these perturbations. However, these solutions would lead to instability problems.
du du 1 / v dk
^7 + C 5 =  ue(k)  u —— —
at ox r \ k + KOX
12 Transportation and Traffic Theory
and the discretized version of the corresponding model requires the parameter values of the
equilibrium relationship ue, of the physical parameters T, 5, v, and of the nonphysical parameters
K, C and a (this last one, already mentionned, is specific to the discretization equation (13)). Let
us recall that in the above, g would be the rampinflow for instance. The issue of the reaction time
T is also illuminating, since the estimated values vary from 1 to 50 seconds or more depending
on the authors and very different functional forms (for r as a function of density k) have been
proposed (see [Del Castillo et ai, 1993] and [Michalopoulos et al, 1992]).
This point has already been mentionned in the paragraph dealing with the basic variables of the
model. The basic problem is to represent a discontinuity in the road layout, i.e a sudden increase
or decrease in capacity (addition or suppression of one lane for instance).
The study of two different situations is of interest:
• Steadystate traffic conditions, either fluid or congested,
The null speed of the Shockwave implies q(x+, t) = q(x~, t) Vt. The following examples show
the solutions in different cases. Solutions for capacity increase and decrease are completely
symmetric, depending on the fluid/congested conditions.
Ex 1. Capacity restriction, demand below capacity
Macroscopic Traffic Flow Models 13
Ex 3. Capacity expansion, increase of downstream supply from below to above upstream capac
ity (This case is symetric of the previous one).
The lines in these examples are either characteristics (straight lines) or the obvious shockwaves.
What is important to notice is that:
• Accelerations are not bounded and are infinite in some cases
• The traffic flow through the discontinuity is always maximized by the entropy solutions
computed by the supply/demand process. Under stationary conditions anyway, the accel
erations would be infinite whatever the solution used since the flow is uniform and the
concentration is not.
(and v = Vf/(2kj)), with parameters Vf and kj having different values on the left and right
handside of the origin. The generic solution of (15) (i.e. excluding the particular solution
u= (jf + 1vQ\ /Vf) is given by:
du (u2  v/T] dx
(16)
+ 1vQ T
with boundary conditions given at infinity by the equilibrium conditions:
_ vf
u is continuous at the origin, but ~ is not, yielding for instance in the diverge case a solution of
the following type:
(the exact functional form depending on the exact value of Q). It should be emphasized here
that nothing prevents Q from being equal to the maximum throughflow of the system (i.e. the
smaller of the maximum equilibrium flow values on the left and righthandside of the origin).
This observation suggests that in general, outside transitory (unstationary) situations, second and
first order models should yield the same outflows of say capacity restrictions].
This last remark is consistent both with Schochet's [1988] convergence result and with Del
Castillo et a/.'s [1993] analysis, which showed that for theoretically consistent values of the
parameters v and T of Payne's model:
T(k) = 1
!/(*) = (1/2) (t)
with xn the position of the nth vehicle and A(l/fc) == u e ( k ) . Now, if ue depends on the position
x, we should rewrite (17) as:
dx
(18) ^(t + T) =A(z n _ 1 (*)z f l (t) J z n _ 1 (*))
or:
There is not really any argument enabling us to prefer one of the formulas (18), (19) over the
other. Both formulas express the fact that the driver anticipates the variation of the physical
layout of the track, they differ by the range of this perception of the driver. Let us consider (19),
which results in the simplest calculations and is consistent with Payne's original approximation.
With the usual approximations:
xn ~ x
xn K, u(x,t)
xnixn w l/k(x + l / ( 2 k ( x , t ) ) , t )
(xn.lxn)/2 w x + l/(2k(x,t))
it follows:
..... du 1 I ., . 1 due/, .dk\ I due
<20)
accounting for a specific acceleration due to the variation of the physical environment.
The same remark applies to the more recent model of Zhang [1998]. Indeed, in keeping with
the derivation of the model, it would be necessary to start from the following macromicro traffic
model:
(21) dx^ +T^Ug (k(Xn + A, *),*„ + A)
at
(replacing equation (16) of the abovementionned paper). In the above equation (21), we ex
pressed the awareness of the driver of the physical layout ahead of him (at a distance A, i.e. the
anticipatory nature of the driver's behavior, in keeping with the ideas of the paper [Zhang 1998].
Now, after setting x = xn+i, xn+i = u, and developing at first order in T and A, it follows:
.„. du 1 . ., . . A 9 u e / , . dk
(22) = (u (k x} u] H (k x)
dt ± J. OK ox
Since the characteristics speed being still
aA; \
~\7 J / OT
a/c
Finally, Zhang's model becomes:
du 1 , ,, , , (du,.,, C\ dk , dup
(23) ^ =I
16 Transportation and Traffic Theory
It is worth mentionning that this last model (23) does not admit the LWR model as its limit
ing case when ue(k, x) = u since the acceleration in this last model is given by (8) when the
equilibrium speeddensity depends on the position.
As a conclusion, some further research should go into adapting second order models to the pos
sible variability of physical characteristics of the links. Some ideas in that direction have been
developed in a purely numerical setting, for instance in [Papageorgiou et al. 1989] (the a coeffi
cients mentionned previously (13)).
Both problems, which are related, have been treated in the literature. The necessity of these
problems (and the related intersection modelling problem) stems from the fact that traffic models
should actually describe flow on networks, and not simply on the real line.
We should distinguish between multibehavior flow, in which the the equilibrium relationships
applying to different classes of users are different (for instance multilane flow), and multiclass
flow, in which users may belong to different classes, as for instance in the dynamic assignment
context, but whose equilibrium relationships are the same. Multiclass models are an inescapable
ingredient of any traffic flow model aiming at representing dynamic assignment situations.
Considering the case of first order models, multiclass flow description relies on the partial flows
densities concept. For each user class d, partial flow Qd, density Kd and speed Vd are introduced;
and the partial flow and density are related by the usual conservation equation and speed defini
tion equation Qd = KdVd. An equation for Vd yields a complete traffic flow model for class
d. The simplest possible model is the FIFO model Vd = V, Vd (all users have the same speed,
irrespective of their class). Characteristic of the FIFO partial flow model is the fundamental fact
that the composition of density and flow are the same. Such a FIFO model was introduced in a
discretized version in [Daganzo, 1994]; a nonFIFO extension was proposed (still in a discretized
setting) in the STRADA model [Buisson et a/., 1996]. A continuous version of this model is de
scribed in [Lebacque and Khoshyaran, 1998] the partial speed Vd results from the partial flow
Qd, estimated directly at each point as the minimum of the downstream partial supply and up
stream partial demand. The behavioral part of the model is contained in the derivation of partial
supplies and demands from their global counterparts.
Second order models have a much older history of multiclass models. We might cite for in
stance the Paynederived METANET [Messner and Papageorgiou, 1990] in which the partial
flow model is approximately FIFO the discretized partial flows are proportional to the compo
sition of the discretized density. The METANET and SSMT derived METACOR [Elloumi et
al., 1994] follows similar principles in its motorway part. Hilliges' model [Hilliges, 1995] is
approximately FIFO too (same speed irrespective of class d).
Basically, models differ on the issue of FIFO behavior of partial flows. FIFO behavior constrains
the model drastically, notably in respect to intersection modelling. Indeed, let [a, b] be a link,
T(t) the travel time of the link for vehicles entering it at time t, and xd(xi t) me composition
of the traffic at location x and time t, then x d ( G > i) — Xd(b, t + T ( t } ) . This last identity shows
that in FIFO models, the inflow composition of intersections may be completely determined
by past upstream conditions. Thus, the modelling of intersections in which this is not the case
(intersections with preselection lanes, leftturning stoarage capacity etc...) requires non FIFO
links. This is the purpose of multilane modelling. First order models do not accomodate these
models easily, because of the equilibrium speeddensity relationship. Two examples may be cited
here. The model of [Daganzo et al. 1997], treats the special case of two lane types and two user
classes, with the effective flow resulting from a user optimum. [Daganzo et al. 1997] derived
analytical solutions for the corresponding Riemann problem. Lebacque and Khoshyaran [1998]
present a general lane assignment model, based on the analogy with classical static assignment
problems. In this last model, the users d, of density Kd, have access to the set of lanes i 6 Id,
and the density Kd is split between lanes i as Kd = £ kf. With kd = 0 if i £ Id, and with Ki
Macroscopic Traffic Flow Models 17
the density in lane i (given by Ki = £) kf) being less than the maximum density Kjti of lane i,
d
the following constraints apply to the unknowns kf:
Kd = £ kf \/d
kf = 0 if i g J d , Vi,
kf > 0 Vi, d
de f
with 7j = Kjj/Kj the relative width of lane i (system optimal, entropy maximizinglike model),
or the following criterion
<Kihi
ue(s) ds
(useroptimal, similar to the Beckmann transform), yielding [Daganzo 1997] as a particular case.
An older model can be recalled here the model of [Michalopoulos et a/., 1984a), in which the
interaction between lanes is assumed to follow a simple linear relaxation behavior, of the form:
9kh
The dispersion as predicted by first order models occurs always at the head of the platoon and
concerns the faster vehicles (this is the consequence of the entropy condition). The situation is
slightly improved by considering discretized first order models, but as we have seen previously,
numerical viscosity is not really a reliable modelling tool. It can be noticed that, when computing
the solution numerically, the platoon aspect is more realistic, because of the numerical viscosity
introduced by discretization.
An illustration of the emulation of platoon dispersion by numerical viscosity effects can be given
by a derivation similar to the one presented above. Let the road be divided into cells of uniform
size 6x. If a concentration k0 is observed in cell i = 0 at time t = 0, with a null concentration
everywhere else, with the simplification that ^ is constant between 0 and ko, the concentration
at timestep p in the cell i is
fc i (p) = Cj^(l^ i fc 0
with Q = H^f. /? is necessarily less than 1 due to the CourantFriedrichLewy (CFL) condition,
and the distribution of concentration inside the platoon is thus binomial, which is quite similar
to the platoon dispersion model proposed by (Robertson, 1969) for the TRANSYT programm.
Again, this is a purely numerical effect depending on the discretization parameters. If  = ^,
which minimizes the numerical viscosity, then /? = 1 and there is no dispersion at all.
The presence of a traffic signal on the road can be considered in a simplified way as a boundary
condition q(xo, t) = 0 when the signal is red. The problem is to compute traffic flow evolution
when the signal turns red and when it turns green again. We can use as an archetypical initial
conditions that the signal has been green for an infinite time, and that a uniform undercritical
flow q0 is present. Boundary conditions are a constant demand q0 upstream and an infinite supply
downstream.
What is interesting to notice again is that the accelerations of vehicles are not bounded, neither
when negative (vehicles stop instantaneously when they join the queue, nor when positive The
first vehicle starting when the signal turns green reaches instantaneously its free speed, and the
next ones may present unphysical acceleration values too. For instance, using a Greenshield equi
librium relationship ue = umax (l  jr—\ it is easy to calculate the acceleration of a vehicle
moving with the flow at position x and time t in the fan area
du
1
dt 4t
The maximum value of this acceleration is obtained for x — —umaxt which corresponds to the
place where vehicles start. It is equal to:
dt 2t
It is clear that, whatever the value of a physical maximum acceleration, it is possible to find a
time t when this value is exceeded.
However, bounded acceleration solutions can be constructed, using for the solution of the above
example the trajectory of the first vehicle of the platoon as a boundary condition. These solutions
must nevertheless be defined in a proper functional setting [Lebacque, 1997] and still need to be
investigated in more detail.
3.5 Intersections.
Modelling intersections is a very difficult task in any macroscopic model, considering that the
interactions inside an intersection are very complex, generally microscopic in nature, and below
the scalelevel of the validity of the continuum hypothesis. From a geometrical point of view,
two possibilities coexist describe the intersection as a point, or as an extended object.
The latter approach has been used mainly in the discretized context and first order models, with
SSMT [Lebacque 1984], in which each movement inside the intersection is discretized as a sin
gle cell, and the intersection cells overlap, with the model proposed by [Michalopoulos, 1988],
also based on overlapping cells, but each movement being discretized into several cells (the dis
cretization was a three cell discretization), or with METACOR [Elloumi et a/., 1994], whose
urban part is partially SSMTderived. A different concept, the concept of nonoverlapping ex
change zones that accomodate several movements the interaction of which is summarized in the
zone dynamics, has been introduced with STRADA [Buisson et al, 1996].
Modelling interactions as points requires to combine quantities describing the traffic state up
stream and downstream of the intersection. Particularly illuminating in this respect, notably as
far as difficulties occuring in higher order models go, is Hilliges' model [1995], which exists
only in semidiscretized form:
dNj __
(24)
with /j the length of cell i. It is obvious that the necessary downstream information for any
upstream cell u is an equivalent speed of downstream cells d, whereas a downstream cell d
requires an equivalent speed of upstream cells: u
(formulas (6.12) to (6.18) of [Hilliges, 1995]. Furthermore, the FIFO or nonFIFO character of
the flow on the upstream links has an impact on the intersection models.
A few solutions have been proposed in the first order case. Daganzo [1994] considered merges
and diverges, and proposed a priority rule for merges, and a maximum flow rule for diverges
(with a FIFO upstream flow). Maximum entropylike models, constrained by upstream demand
and downstream supply, were proposed in [Lebacque and Khoshyaran, 1998]. The impact of the
nature of the upstream flow is particular clear in the case of a diverge if the flow is non FIFO, the
partial demands are proportional to the traffic composition coefficients upstream of the diverge,
but if the flow is FIFO, it is the actual partial flows (throughflows through the intersection) that
are proportional to the traffic composition coefficients.
To illustrate this point, let us consider for instance a pointwise diverge say a, with upstream link
u, downstream links i, upstream demand 5u(a, t), downstream suplies <7j(a, t) resulting from the
up and downstream densities and equilibrium functions (11) by:
Let qi(a, t),q(a, t) be the partial and total flow through the intersection, and 8i(a, t) and 5(a, t) the
partial and total demands upstream. Finally, let Xi(a~i t) be the composition of traffic entering
Macroscopic Traffic Flow Models 21
the intersection. Then, if the model is nonFIFO, the partial demands are proportional to the
compositions, yielding:
On the other hand, if the upstream flow is FIFO, the partial flows are proportional to the compo
sitions, yielding:
q(a,t) = Mini[ai(a,t)/Xi(a,t),6(a,t)]
qi(a,t) = Xi(o,*)g(M) 
In the case of second order models, friction coefficients in the speed equation have ben used to
account for the effect of off and onramps (i.e. diverges and merges). The basic idea is to assume
that the say outflow (on the off ramp) has a given speed related to that of traffic upstream of the
intersection (generally smaller), as in [Papageorgiou et al. 1989]. Similar models with friction
terms were proposed in [Michalopoulos et al., 1992], generalizing Papageorgiou's model:
du $ dk
 =(Uf(x)
[Kerner, 1997], exhibited a beautifull experimental example of two traffic jams of constant size,
persisting on a german highway for a very long period (of the order of an hour), surrounded by
fluid traffic. Such a structure cannot be explained by first oder models, except by recoursing to
non entropic solutions as illustrated by the following characteristics chart:
But such a solution is highly unstable in the sense that the solution is not unique, and that in
finitely small divergences from the right initial conditions may produce an (unique) solution
totally different from the above solution, but very close to the entropy solution with its accelera
tion fan. Let us recall the entropy solution satisfies to continuity relatively to initial conditions.
Some second order models of the KernerKonhaiiserHelbing type might be able to explain the
persistence of such structures.
The above example definitely constitutes a challenge for future traffic flow models.
Boundary conditions are an important feature they determine both intersection models and net
work entry and exit point models. No traffic equation should be writen only on an infinite track!
In the case of first order models, the supply and demand concepts yield the boundary conditions
for any link. For instance the link inflow and outflow Q(a, t) and Q(b, t) of the following link:
22 Transportation and Traffic Theory
A(M) =
In the case of second order models, to the authors knowledge, no simple boundary conditions
exist. Let us consider again Hilliges' model (eq. 24) to show the difficulties involved. For a
network entry point, corresponding to an entry cell say i, the inflow ki\Ui must be given, i.e.
the upstream density fcj_i must be given.
entry point
The same applies to the speed equation, which requires an upstream speed. Basically, upstream
boundary conditions will comprise both an upstream density and an upstream speed. For exit
points, only the downstream speed is required (the analogy with Papageorgiou's exit ramp model
is obvious). There is no simple way to accommodate supply and demand concepts (for example
a constraint on the demand at an exit point would imply a constraint on the downstream speed
Ui+i, dependent on the exit link density fcj).
4 CONCLUSION
The analysis presented above has obvious limits, since it is clear that the only real validation
of a traffic model is its confrontation to real data. However, it is also true that real data always
present a high proportion of noise that may mask some inconsistencies of a tested model, and
that several and possibly conflicting interpretations may often be given of the same observed
situations. The validation of a model against real data is thus not necessarily a guarantee that
model will behave properly in all situations. This is the reason why many papers presenting
the validation of some traffic models are of limited interest. It thus seems interesting to have a
basic framework to analyse the behaviour of a model on a set of welldefined cases, to point out
eventual inconsistencies or side effects and to compare a particular model (or class of models)
to the existing ones. From that point of view, the various cases presented in this paper may
contribute to constitute such a framework.
REFERENCES
Astarita,V. (1996). A continuous time link model for dynamic network loading based on travel
time function, in: Transportation and Traffic Theory, proceeding of the 13th ISTTT (J.B. Lesort
ed.). 79102, Pergamon, Oxford.
Bui D.D., P. Nelson, S.L. Narasimhan, (1992). Computational realizations of the entropy condi
tion in modelling congested traffic flow. Report 12327. Texas Transportation Institute, USA.
Macroscopic Traffic Flow Models 23
Buisson C. , J.P. Lebacque, J.B. Lesort, H. Mongeot, (1996). The STRADA model for dynamic
assignment. Proc. of the 1996 ITS Conference. Orlando, USA.
Cassidy M. J.,(1998). Bivariate relation in nearly stationary highway traffic, Trsp. Res. 32B, pp
4960.
Cremer M., M. Papageorgiou, (1981). Parameter identification for a traffic flow model, Automat
ica 17 pp 837843
Daganzo C.F., (1994). The cell transmission model 1: a dynamic representation of highway
traffic consistent with the hydrodynamic theory.
Daganzo C.F., (1995a). The cell transmission model 2: network traffic simulation.Trsp. Res.
28B. 4 269287. Transportation Research 29B. 2 7993.1995.
Daganzo C.F, (1995b). A finite difference approximation of the kinematic wave model. Trsp.
Res. 29Bpp 261276.
Daganzo C.F., (1995c). Requiem for secondorder fluid approximation of traffic flow.Trsp. Res.,
29B pp 277286.
Daganzo C.F., Cassidy M. J. and Bertini R. L., (1997).Causes and effects of phase transitions in
highway traffic, report UCBITSRR978.
Daganzo C.F., (1997) A continuum theory of traffic dynamics for freeways with special lanes.
Transportation Research 31 B. p 83102.
Del Castillo J.M., Pintado P. and Benitez F. G., (1993). A formulation for the reaction time
of traffic flow models. Twelfth International Symposium on Transportation and Traffic Theory,
Berkeley, California.
Elloumi, E.,H.Hadj Salem, M. Papageorgiou, (1994). METACOR, a macroscopic modelling tool
for urban corridors. TRISTAN II Int. Conf., Capri.
Friesz T.L., J. Luque, R.L. Tobin, B.W. Wie, (1989). Dynamic network traffic assignment con
sidered as a continuous time optimal control problem. Op. Res. 37, 6 893901.
Grace M.J., R.B. Potts, (1964). A theory of the diffusion of traffic platoons, Op. Res. 122 pp
529533.
Greenshields B.D. (1935). A study of traffic capacity, Proceedings of the Highway Research
Board, vol 14, pp 448477
Helbing D., (1997). Verkehrsdynamik, Springer, .
Heydecker B.G, (1994). Incidents and interventions on freeways. PATH Research Report UCB
ITSPRR 945.
Hilliges M., (1995). Bin phanomenologisches Modell des dynamischen Verkehrsflusses in
schnellstralknnetzen. PHD dissertation. Institut fiir theoretische Physik der Universitdt
Stuttgart. Shaker Verlag.
Kerner B. S., (1997). Experimental characteristics of traffic flow for evaluation of traffic mod
elling. IFAC/IFIP/IFORS symposium, Chania, Greece.
Kerner B. S. and Rehborn H., (1997). Experimental properties of phase transitions in traffic flow.
Phys. Rev. Let.19, pp 40304033.
24 Transportation and Traffic Theory
Kerner B. S., Konhaiiser P. and Shike M., (1996). A new approach to problems of traffic flow
theory.in: Transportation and Traffic Theory, proceeding of the 13th ISTTT (J.B. Lesort ed.).
79102, Pergamon, Oxford.
Kiihne R. D., (1984). Macroscopic freeway model for dense traffic. Stopstart waves and incident
detection. Ninth Int. Symposium on transportation and traffic theory. VNU Science Press pp 21
42.
Lebacque J.P., J.B. Lesort, F. Giorgi, (1998). Introducing buses into first order macroscopic
traffic flow models. Trsp. Res. Rec 1664, pp 7079.
Lebacque J.P., (1984). Semimacroscopic simulation of urban traffic. Int. 84 Minneapolis Sum
mer Conference. AMSE.
Lebacque J.P., (1995). Le modele de trafic de P. Ross Solutions analytiques, Actes INRETS 45,
Arcueil,France
Lebacque J.P. , (1996a). Instantaneous travel times for macroscopic traffic flow models. CER
MICS Report 5996.
Lebacque J.P., (1996b). The Godunov scheme and what it means for first order traffic flow mod
els. in:Transportation and Traffic Theory, proceeding of the 13th 75/77 (J.B. Lesort ed.),pp 647
677, Pergamon, Oxford.
Lebacque J.P, (1997). A finite acceleration scheme for first order macroscopic traffic flow
models. The 8th IFAC symposium on transportation systems, Chania, Greece.
Lebacque J.P. , M.M. Khoshyaran, (1998). First order macroscopic traffic flow models for net
works in the context of dynamic assignment. CERMICS Report. To be Published. Presented at
the 6th Meeting of the EURO Working Group On Transportation. Goteborg.
Leo C.J. , R.L. Pretty, (1992). Numerical simulation of macroscopic traffic models. Transporta
tion Research 26B. 3 207220,
Lighthill M.H., G.B. Whitham, (1955). On kinematic waves II A theory of traffic flow on long
crowded roads. Proc. Royal Soc. (Lond.) A 229 317345.
Liu G., Lyrint/is A. and Michalopoulos P., (1998). Improved high Order model for freeway
traffic flow. TRB 77th meeting, Washington, DC.
Merchant D.K., G.L. Nemhauser, (1978). A model and an algorithm for the dynamic traffic
assignment problem.Transportation Science 12 183(199 & 200207.
Messner A., M. Papageorgiou, (1990). METANET, a macroscopic modelling simulation for
motorway networks. Technische Universitat Miinchen.
Michalopoulos P.G, G. Stephanopoulos, V.B. Pisharody (1980). Modelling of traffic flow at
signalized links, Trsp. Sc. Vol 141, pp 941
Michalopoulos P.G., D.E. Beskos, Y. Yamauchi, (1984a). Multilane traffic flow dynamics some
macroscopic considerations.
Michalopoulos P.G. , D.E. Beskos, J.K. Lin, (1984b). Analysis of interrupted flow by finite
difference methods. Trsp. Res. B 18B pp 409421.
Michalopoulos P.G., (1988). Analysis of traffic flow at complex congested arterials. Trsp. Res.
Rec 1194 pp 7786. Trsp. res. 18B, 377395.
Macroscopic Traffic Flow Models 25
Serge P. Hoogendoorn and Piet H.L. Bovy, Delft University of Technology, Faculty of Civil
Engineering and Geosciences, Transportation and Traffic Engineering Section, Delft, The
Netherlands
ABSTRACT
In contrast to microscopic traffic flow models, macroscopic models describe traffic in terms
of aggregate variables such as traffic density, flowrate, and velocity. The implied mean traf
fic behavior depends on the traffic conditions in the direct environment of the vehicles in the
traffic stream. Using the analogy between the vehicular flow and flow in fluids encouraged
deriving these models (e.g. Lighthill and Whitham (1955), Payne (1979)). The advantages of
macroscopic models are among others the insight gained into traffic flow operations (e.g.
shockwave analysis), the applicability in model based control, the relatively small number of
parameters simplifying model calibration, and the applicability to large traffic networks.
Generally, macroscopic models consider the behavior of the aggregate traffic flow. That is,
neither a distinction of userclasses, such as traveler types (commuters, freight, recreational,
etc.), vehicles types (personcars, trucks, busses, vans), paying and nonpaying traffic, and
various types of guided vehicles, nor a distinction of roadway lanes is made. However, we
envisage that a generalization of macroscopic traffic flow models to both userclasses and
lanes is advantageous. On the one hand, this generalization increases the applicability of mac
roscopic models to the synthesis and analysis of multilane multiclass (MLMC) traffic flow.
As a result, more insight is gained into the responsebehavior of the heterogeneous multilane
flow, such as effective capacity, velocity distribution, and the distribution of vehicles over the
roadway lanes.
On the other hand, from the traffic control perspective, contemporary policies pursue a more
efficient use of the available infrastructure (e.g dynamic allocation of roadway lanes to
28 Transportation and Traffic Theory
The paper is organized as follows. First we present the MLMC generalization of the gas
kinetic flow equations of PaveriFontana (1975) for constrained and freeflowing vehicles.
Secondly, we present the macroscopic flow model using conservative variables, while subse
quently discussing the equilibrium speeddensity relations and the density lane distribution.
After discussing the numerical solution approach, we present results from application of the
macroscopic model to two test cases. Finally, in the closing section we summarize our re
search findings.
Macroscopic Traffic Flow Modelling 29
Gaskinetic models describe traffic using the reduced phasespace density (PSD) p(*,v,/),
where p(jc,v,/)dxdv equals the expected vehicle number in [jt,jc+djt) driving with velocity
[v,v+dv) at instant t. This concept is borrowed from statistical physics and can be considered
as a mesoscopic generalization of the traffic density r(x,t).
Equations describing dynamic changes in the implied velocity distributions are based on the
work of Prigogine and Herman (1971), who assumed that changes in the reduced PSD are
caused by acceleration, deceleration, and convection. The latter simply describes changes due
to the movement of the traffic. Their deliberations yielded the following equation:
dp dp [dpi [dpi
— + v— = — + — CD
dt dx
The acceleration term [dp/dt]^cc describes relaxation of drivers' speed towards a traffic
condition dependent velocity. Prigogine and Herman (1971) proposed:
[dp/dt]ACC =(Q°(jt,v,f)p(JC,v,f))/T (2)
where T denotes the acceleration time and Q°(x,v,0 reflects the distribution of desired veloci
ties, that is, of the expected desired velocity of vehicles driving with velocity v. Paveri
Fontana (1975) improved the relaxation process by considering the PhaseSpaceDensity
(PSD), which can be considered as a generalization of the reduced PSD extended with an in
dependent variable describing the desired velocity v°, that is p(jc,v,v°,0
The interaction term [3p/3?]iNT reflects fast vehicles catching up with slower vehicles, consid
ering the immediate overtaking probability n. The interaction term is composed of contribu
tions of active and passive interactions. An active interaction occurs when a vehicle driving
with velocity v interacts with a slower vehicle driving with velocity w<v. A passive interac
tion is defined from the viewpoint of the impeding vehicle: it occurs when a vehicle driving
with velocity v impedes a faster vehicle (w>v). The assumption of vehicular chaos (cf.
Prigogine and Herman (1971)) yields the following expressions for the contributions of active
and passive interactions on the dynamics to the reduced PSD:
and:
In this section we will establish the first step in the derivation approach which is the determi
nation of gaskinetic equations for multiclass multilane traffic flow operations. These equa
tions describe the dynamics of the multilane multiclass PhaseSpace Density (abbreviated as
MLMCPSD), which is disaggregated into contributions of platooning and freeflowing ve
hicles respectively due to differences in drivingcharacteristics. Similar to other gaskinetic
models, several classspecific and lanespecific processes govern the dynamics of the
MLMCPSD (acceleration towards the desired velocity, deceleration caused by vehicle inter
actions, immediate, postponed and spontaneous lane changing, and statetransitions). Typical
classspecific parameters are the desired velocity, the acceleration time, the reaction time, the
vehicle length, and the withinuserclass velocity variance (cf. Hoogendoorn and Bovy
(1998a,b)). The traffic conditions on the roadway lanes differ, due to among others the class
dependent overtaking behavior and lane preferences.
and:
def .  .
rJ(x,t) = pJu(x,v,v0,t)dv°dv = pJu(x,v,t)dv (6)
In the sequel we will show that the state of a driver  that is, whether he is constrained or
freely flowing  to a large extent determines both the lane changing behavior and the accel
eration behavior. A constrained or platooning driver refers to any driver who is impeded by a
slower vehicle in front without being able to immediately change to an adjacent lane. Con
versely, we will refer to any driver not impeded by any slow vehicle or who is able to imme
diately change lanes as a freeflowing driver. The MLMCPSD is then disaggregated as fol
lows:
Macroscopic Traffic Flow Modelling 31
,v,v,r) (7)
where auj(x,v,v°,t) and £,u\x,v,v°,t) respectively denote the contribution of the constrained and
the freeflowing MLMCPSD. Correspondingly to the mixedstate reduced MLMCPSD, we
define the separate reduced MLMCPSD 's, for constrained and freeflowing vehicles.
The gaskinetic equations that describe the dynamics of £,,/ are governed by both continuum
and noncontinuum processes. The continuum processes reflect the smooth changes in the
freeflowing MLMCPSD due to balancing inflow and outflow in the phasespace (x,v,v ,t).
We can show that the gaskinetic equations describing the dynamics of the freeflowing
MLMCPSD equal (cf. Hoogendoorn and Bovy (1998c)):
where TM° denotes the classspecific relaxation constant, reflecting the acceleration capabili
ties.
The continuum terms describing the changes in aj can be derived similarly. However, in op
position to freeflowing drivers, constrained drivers are by definition unable to accelerate to
wards their (own) desired velocity. Assuming a,/(v0v)/Tu°=0, the gaskinetic equations for the
constrained MLMCPSD become:
(10)
dt dx dt
where [dGuj/dt]wc reflects dynamic changes caused by noncontinuum processes. For both
freeflowing and constrained traffic, the noncontinuum processes reflect the influences of
deceleration and lane changing. Let us now discuss these processes in some detail.
Braking and immediate lanechanging. Given the assumption of vehicular chaos, the expected
number of actively interacting freeflowing vehicles of userclass u at (x,t) on laney driving
with a velocity v while having a desired velocity v° equals:
Uo 0
32 Transportation and Traffic Theory
where xF;(v)=2uxFH;(v)>0 equals the expected number of active interactions per unit time of a
vehicle driving with velocity v with slower vehicles of any userclass on laney.
where p,/^1 denotes the immediate overtaking probability to the leftlane 0+1) an
d rightlane
(/I) respectively. Similarly, actively interacting freeflowing vehicles on lane y'±l able to
immediately change lanes to lane j cause an increase in ^,/. In Europe, overtaking regulations
during noncongested traffic conditions follow the 'drive on the right  overtake on the leff
principle. That is,pJ^~l=Q, when traffic is not congested.
Postponed lane changing. Constrained vehicles previously not able to change lanes may
change lanes if the opportunity arises, after which the driver is able to accelerate towards his
desired velocity until it is again impeded. We will model the postponed lane changes using
the postponed lanechange rates X,rj±\ with Xul"°=XuM"M+l=0, and Xj= X^1 +X,Tj+l . Post
poned lane changing causes a migration of constrained vehicles a,/ to freeflowing vehicles
and
t^f / WPLC~' = ^r 7± V,(v, v°) (14)
Spontaneous lane changing. Freeflowing drivers may choose to change to either of the adja
cent lanes depending on the driver's preferences. This yields a flow from the current lane to
an adjacent lane in correspondence to the classspecific driver's preference. Since this process
is comparable to the postponed lanechanging process it yields comparable expressions. For
European legislation , this spontaneous lane changing process to a large extent results from
Macroscopic Traffic Flow Modelling 33
vehicles which have overtaken slower vehicles using the left lane returning to their origin
lane. In this case, spontaneous lane changes to the left lane are rare. Thus, the spontaneous
lane changing intensities satisfy yJ~>i'+l=Q. If we consider American legislation, the spontane
ous lane changes result from drivers having a distinct preference for a specific lane. That is, a
driver changes to the left lane if he prefers to be on any of lanes left of his current lane.
Assuming that freeflowing drivers only change lanes if they remain freeflowing, then spon
taneous lane changing yields the following contribution to the dynamics of ^,/ (cf. Hoogen
doorn and Bovy (1998c)):
[d^/dt]£J±l=yCJ±^Ju(v,vG) and [a^/a;]^=Yf^fV,v 0 ) (15)
Passive interactions. If any vehicle of class u on lane j driving at a velocity w interacts with a
vehicle driving at a velocity v<w, without having the opportunity to immediately change
lanes, it will assume the velocity v of the impeding vehicle while becoming constrained. In
this latter case, the number of constrained vehicles increases with rate:
due to passive interactions with respectively freeflowing and constrained vehicles respec
tively (cf. Hoogendoorn and Bovy (1998c)), where pj=pj(x,v,i) and quj=quj(x,v,t) denote the
immediate lanechanging probabilities of interacting freeflowing and constrained vehicles.
Let us define the passive interaction rates by:
def ™ def ~
^'(v,v°) = j(wv)^(w,v°)dw and ^'(v,v°) = j(wv)a^(w,v°)dw
Relaxation due to vanishing impeding vehicles. When the constrained vehicles on lane j are
able to changes lanes, vehicles that were previously constrained by the lanechanging vehicles
are 'freeflowing'. This event can in itself yield a relaxation of other vehicles. Since this proc
ess is very complex, we will assume that the process can be modeled by:
[a^'/a/] REL =d( v F ; (v))a^(v,v°) and [3^/3r] REL =d(^(v)X(v,v°) (19)
where $ is a monotonic increasing function of the mean number of active interactions ^(v) of
vehicles driving at a velocity v.
34 Transportation and Traffic Theory
Combining the gaskinetic equation for freeflowing traffic (8) with the specifications for the
noncontinuum processes derived in the previous paragraphs, yields:
i u / —u \ / ur \ / \\ r u / i ( \ i
a?
.
a*
\ V / \* " H ^ u V ' ^/
(21)
r jWx
(v))V ,
^ /' /
(g
J ;'— :4i ^ l U
]
This section presents the macroscopic equations describing the dynamics of the MLMC mac
roscopic traffic equations. To this end, aggregation operators are applied to the gasdynamic
equations (20) and (21). These operators aggregate the contributions of freeflowing and con
strained drivers to the respective conservative variables density, momentum, and energy.
Hoogendoorn and Bovy (1998c) show that using these conservative variables yields a simpli
fied derivation approach and improved numerical analysis of the resulting macroscopic mod
els. Moreover, the model can be easily recast in its primitive form (density, expected velocity,
velocity variance).
Aggregation operators. Let us define the operator H on any function a(v,v°) by:
2^[a(v,v°)] = Ja(w,H' 0 )^(jc,w,w 0 ,Odw 0 dw (22)
To derive the macroscopic flow equations, the gaskinetic equations (20) and (21) are multi
plied by V A , with £=0,1,2. Subsequently, the resulting equations are integrated with respect to
the velocity v and the desired velocity v°. Since in effect these aggregation operators total the
contributions of the vehicles driving at various velocities v to the respective conservative
variable. For instance, v^(v) reflects the contribution to the traffic momentum EuJ[v] of free
flowing vehicles driving at a velocity v. In multiplying the reduced PaveriFontana equations
vk for £=0,1, and 2, the dynamics of Ej[vk] and £,/[v*] are established. Hoogendoorn and Bovy
(1998c) show that these respectively equal to traffic density (£=0), the traffic momentum or
Macroscopic Traffic Flow Modelling 35
flow (&=1), and two times the traffic energy (k=2) of class u on lane j of freeflowing and con
strained vehicles respectively.
a* .
"(v))]
' * spontaneous / postponed lanechanging
dt dx (25)
and thus:
def
0
)] = P>(v,v 0 )] (27)
Statespecific MLMC momentum dynamics. Similar to the derivation of the vehicle conser
vation equations, we can establish the freeflowing momentum dynamics equation by assess
ing the aforementioned equations for k\ (cf. Hoogendoorn and Bovy (1998c)).
statechanging
dt
36 Transportation and Traffic Theory
where vu denotes the mean desired velocity of class u. Equation (28) shows that among oth
ers the mean momentum of freeflowing traffic of userclass u on lane j changes due to the
inflow and outflow of momentum, reflected by the spatial derivative of the traffic energy.
Alternatively stated, the arithmetic mean velocity vjmj/rj changes due to the balance of in
flow and outflow of vehicles with different velocities.
We can also derive the momentum dynamics equation for the constrained vehicles:
dt dx
Let us remark the subtle difference between the reduction in the traffic momentum due to ac
tive interactions for freeflowing and constrained vehicles. On the one hand, when a free
flowing driver interacts, he either changes lanes or joins the platoon. As a result, a density
flux from the freeflowing spatial density to the constrained traffic density causes the mo
mentum to decrease. However, this does not necessarily imply that the mean velocity of free
flowing vehicles decreases. On the other hand, when a constrained vehicle is impeded, it will
reduce its velocity. This causes the mean velocity of the constrained vehicles to decrease. As
a consequence, the traffic momentum of the constrained vehicles is decreased. Note that this
does not necessarily imply a reduction in the spatial density of constrained vehicles.
Statespecific MLMC energy dynamics. Finally, the multilane freeflowing energy dynamics
can be determined for k2. These equations are similar to the dynamic equations for the un
constrained and constrained momentum presented in the previous section, and will not be ex
plicitly presented in this paper. Their formulation can be found in Hoogendoorn and Bovy
(1998c).
To study the differences between aggregatelane conservative models (cf. Hoogendoorn and
Bovy (1998d)), and the presented MLMC model, it is worthwhile to determine the mixed
state traffic dynamics. To this end, we define the fraction of constrained vehicles:
a]u (x, t) = ruj (x,t)/ ru} (x, t) (30)
the mixed state immediate overtaking probability from lane j to laney'il:
MLMC conservationofvehicle equations. By adding the equations (24) and (25), the multi
class multilane conservationofvehicles equation results:
^ + ^ = £/=;±1(^/pu'^^ (33)
By stateaggregation, both the statechanging term and the withinlane influence of vehicles
interacting vanish, due to the fact that neither of these processes cause changes in the number
of vehicles of class u on laney. In the sequel of this section, we will see that these processes
do change the momentum and energy. Compared to the aggregatelane multiclass model of
Hoogendoorn and Bovy (1998d), additional density fluxes between the motorway lanes are
present due to the different types of lanechanging.
Mju t' rujv°  T " ( 1 ~^" ) ^ s (P,/ [vV/ (v)] + P/ [vO;! (v)]) (34)
\ ~ u '
we can establish the mixedstate momentum dynamics by adding equation (28) and (29):
Ml ml
• + 2
dt dx ^/(1aO (35)
/p j
 ^.._.+l (rcr ,/ [v¥ (v)]  n;f ^'P/ [v*F > (v)])  £ _ +i (A^'mj  A{*X )
Compared to the aggregatelane multiclass model of Hoogendoorn and Bovy (1998d), the
distinction of lanes introduces momentum flows between the motorway lanes caused by the
different types of lanechanging. The expression for the equilibrium momentum M,/ is equal
to the expression derived by Hoogendoorn and Bovy (1998d) for aggregatelane multiclass
traffic flow. The equilibrium momentum (34) describes changes in the traffic momentum of
class u on lane j due to vehicles interacting with other vehicles, without having the ability to
immediately overtake to either adjacent lane. Hoogendoorn and Bovy (1998d) show that this
interaction is asymmetric. That is, the presence of relatively slow vehicles, such as trucks,
have a more profound impact on the momentum of faster vehicles than vice versa. Bliemer
(1998) has used this quantification of the asymmetric interaction to determine multiclass
travel time functions for multiclass dynamic traffic assignment.
Alternatively, we can determine the equilibrium velocity V,/. This velocity equals the equilib
rium momentum MJ divided by the density rj, i.e.:
MLMC energy dynamics. Using a similar approach, we can determine the mixedstate energy
dynamics:
38 Transportation and Traffic Theory
dej r) g Fj  cj
™>L
dt +°(mJHJ+jJ,2) = 2 0 " ".
fa T°/(la^) (37)
J
where then equilibrium energy EU of class u on lane 7' is defined by:
where Hl/=3euj/rl/(muj/rl/)2 and 7^ respectively depict the traffic enthalpy (convective energy
flux) and the flux of velocity variance (nonconvective energy flux). By noticing that the traf
fic energy, the traffic velocity and the variance relate as follows:
eJu=jruJ((vJJ2+QJu) (39)
where 9M; denotes the velocity variance, we can establish (Hoogendoorn and Bovy (1998a)):
El = \ rj (vJuVuj + QJu ) (40)
where Qj denotes the equilibrium velocity variance:
u s (41)
In the sequel, we specify relations for the equilibrium velocity and velocity variance. Using
Muj=rujVuj and (40), we can determine respectively the equilibrium momentum and energy.
MLMCMODEL FORMULATIONS
We can summarize the model equations by defining the vector \vj=(rj,mj,ej). Hoogendoorn
and Bovy (1998c) show that the model equations (33), (35) and (37) can be recast as follows:
dt dx dt dx
where A.J is the conservative fluxJacobian. It describes how small spatial variations in the
conservative variables influence the other conservative variables over time. The vector x«;
summarizes the righthand sides of the equations (33), (35) and (37).
Using this formulation, the model can be recast into among others its primitive form, and its
characteristic or Riemann form. The former describes the dynamics of the MLMC density,
velocity and velocity variance and is consequently well suited for comparing the MLMC
model equations with other macroscopic flow models. The characteristic form describes the
dynamics of the characteristic variables. Although these variables lack intuitive appeal, they
are of dominant importance when mathematically analyzing the properties of the flow equa
tions. For instance, they reveal the way in which small perturbations are transported in the
flow along the socalled characteristic curves. It can be shown (cf. Hoogendoorn (1998c)) that
when the traffic conditions are freeflow, disturbances are transported downstream. In oppo
Macroscopic Traffic Flow Modelling 39
sition, when traffic conditions are congested, perturbations are transported in both upstream
and downstream directions.
Figure 1 shows the relations between the various formulations and their respective uses. For a
detailed account on the different model formulations, we refer to Hoogendoorn (1998c,1999).
conservation of characteristic
vehicles equations
Figure 1: Different forms of traffic flow models, the relevant variables, and the applicable
numerical solution methods.
In this section we will consider the MLMC equilibrium conditions for mixedstate traffic. To
this end, we propose a simple procedure to determine these equilibrium conditions. The dis
cussion focuses on both the distribution of density on the roadway lanes, and the equilibrium
velocity. We will consider two userclasses, namely trucks and personcars.
40 Transportation and Traffic Theory
Before presenting the approach to determine the equilibrium traffic conditions for the
MLMCmodel, the acceleration time iu, the desired velocities vu°, the fraction of constrained
vehicles O.J, the immediate lanechanging probabilities nuj, and the lane changing rates A«; are
specified. In the scope of this preliminary study, we have neglected the role of the flux of ve
locity variance, i.e. jJ=Q.
Desired velocities. For the desired velocities of personcars and trucks, the following values
have been respectively chosen:
v^rson.car =32m/s and vt°ruck = 24m / s (2)
These values agree with average values observed on twolane motorways in the Netherlands.
Note that in the Netherlands, the distinct speed limits on motorways for personcars and
trucks are 32m/s and 22m/s respectively.
The acceleration times. The acceleration time reflects the average acceleration capabilities of
vehicles of a specific userclass. Since personcars generally have better acceleration capa
bilities than trucks, we assume Tpersoncar<ttruck More specifically, we have chosen:
Vsoncar = *S ^ T
truck = ^ (3)
Let sj denote the space needed by a vehicle of class u driving on lane j on either of the desti
nation lanes j±\. The space needed is expressed as a function of the mean vehicle length Lu,
the average velocity vj and the reaction time Tj. For a vehicle of class u driving with veloc
ity v we assume:
^(v) = 2.5(L u +7» (6)
Let us assume that the velocities of userclass u on lane j are Gatm/andistributed random
variates with mean vuj and variance Quj. The random variate SUJ describing the distribution of
space needed by vehicles of class u on lane j is also Gaussian with mean 2.5(LM+7Yvu) and
Macroscopic Traffic Flaw Modelling 41
standard deviation (2.5Tltr)2Quj. The probability that a vehicle of class u on lane j can change
to either of the adjacent lanes equals:
X^1 = x^>±! ( V j, QJu, r ;±1 ) = Pr(S/ < GJ±1) (7)
We assume that the immediate lanechanging probabilities for constrained vehicles are negli
gible, i.e. qj=0. That is, the probability that an overtaking opportunity occurs precisely when
a constrained vehicle actively interacts equals zero. Thus nj equals:
±l
n^J = Tt^1 ( V ;, 9;, a', r&) = (i  a;jp^'xf"'11 (8)
11
where (V^ models both the preference for either of the adjacent lane, and whether the driver
may use that lane for overtaking. For instance, for noncongested traffic operations in Europe
we have (3,T;+1=1 and P/^O.
Spontaneous and postponed lane changing rates. The spontaneous lane changing rates are also
specified by considering the gap distribution on the destination lane. Considering European
traffic regulations, traffic must use the rightmost lane if possible. Thus, we assume:
Yr ; ' + '=0 and Yf^' = Xf"'"' / T ° (9)
where l/Tu° is the freeflow spontaneous lane changing rate. Tu° can be considered to be the
mean time needed for an overtaking maneuver.
Also postponed lane changing is a function of the available gap distribution on the destination
lanes. We propose:
K^j±l =$Ju*j±1ti~*j±l/w? (10)
where Wu° denotes the mean time waiting behind the leading vehicle given freeflow condi
tions.
Velocity variance and fraction of constrained vehicles. We assume that both the velocity vari
ance and the constrained vehicle fraction can be adequately expressed as functions of the
mean number of vehicles per unit unoccupied lanespace ^ r =l/E(G / ) (see Figure 2).
42 Transportation and Traffic Theory
0
0.00 0.02 0.04 0.06 0.08 0.00 0.02 0.04 0.06 0.08
effective density [veil / (in lane)] effective density [veh I (in lane)}
Figure 2: Equilibrium velocity variance and constrained vehicle fraction as functions of the
mean number of vehicles per unit unoccupied space on a lane.
We have assumed that the velocities are Gaussian distributed random variates, and can thus
be specified by the mean velocity vj and the velocity variance Qj. The assumption of Gaus
sian distributed velocities, and the specifications of the constrained vehicle fraction, and the
immediate overtaking probability enables determining the equilibrium velocity (36) given rj,
vj, and QUJ, i.e.:
Vuj=Vus(r,\,Q) (11)
where r, v, and 0 are vectors of respectively the densities, velocities, and velocity variance of
each class on each lane.
Figure 3: The equilibrium velocity on the rightlane and the leftlane of a twolane motor
way, for constant truck densities, and different personcar densities.
Figure 4: The equilibrium fraction of personcars and trucks using the left lane for fixed truck
density values and increasing personcar densities.
Given the aggregatelane densities ru for the respective userclasses, we can iteratively deter
mine the equilibrium lane distribution, velocity, and velocity variance satisfying (12) and
(13). In illustration, Figure 3 and 4 show the equilibrium velocities for fixed truck densities
(^truck=0,2 , and 10 truckslkmllane).
44 Transportation and Traffic Theory
In the past, several numerical approximation schemes have been proposed to determine solu
tions to a variety of macroscopic models (see Lyritnzis et al. (1994), Lebaque (1996), and
Hoogendoorn and Bovy (1998a)). Because of the increased complexity of the developed
higherorder traffic flow model there is a need for more efficient numerical approaches to ap
proximate solutions.
As with other traffic flow models, the numerical treatment of our multilane multiclass flow
model is quite cumbersome. Hoogendoorn (1999) describes a new approach to solve higher
order multilane flow models, based on the flow equations cast in conservative variables.
The resulting scheme is an adaptation of the Van Leer FluxVector splitting scheme (cf. Van
Leer (1982)). It considers the direction in which the perturbations are transported in the
MLMC traffic flow, while conserving the density, momentum, and energy in the distin
guished roadway segments that follow if only convective processes are considered. The
scheme adapts to the prevailing traffic conditions (freeflow/congested). Moreover, the effects
of the noncontinuum processes are quantified in a multistep approach, in order to prevent ve
hicles laterally flowing out of an empty lane. A fourthorder RungeKutta approach was cho
sen for the temporal discretization. For details, we refer to Hoogendoorn (1999).
In this section we discuss results of macroscopic simulation of two testcase examples, based
on the specifications proposed in the preceding section. These are: mixing traffic classes, and
a lane drop.
Mixing of classes
In the first example, we consider a twolane ringroad of 20km length. At time t=0, three ho
mogeneous regions are present. The first region Xo°=[2km,6km) consists of personcars only.
The second region X0l=[6km,8km) consists of both personcars and trucks, while the third
Xo2=[Skm,l2km) consists of trucks only (Figure 5). We assume that initially, traffic is in equi
librium. That is, the initial velocities, velocity variances, and the distribution of vehicles on
the lanes are determined by applying the approach described in the previous section (Figure
6).
When we consider the dynamics of the vehicles in the head of the region Xo2, trucks flow into
the empty downstream roadway section at a velocity nearly equal to the desired velocity. For
instance, considering instant t=4min, the fastest trucks are located at approximately x=iS.5km.
These vehicles have traveled at a velocity of approximately 95km/hr. By studying the char
Macroscopic Traffic Flow Modelling 45
acteristics of the flow model, we can show that these vehicles have traveled along the so
called Machline of the flow, defined by djc/d/=(v1/+(36,/)1/2) (cf. Hoogendoorn and Bovy
(1998d)).
trucks
x=8km
personcars
x=2km
x=0km
Figure 5: Schematics of twolane ringroad, and initial distribution of personcars and trucks.
Compared to the trucksonly region XQ where trucks use both lanes of the twolane roadway
comparably, in the mixed region XQI the laneuse of trucks is more confined to the rightlane.
Fast personcars flowing from this region into the truckonly region do not affect the veloci
ties of the trucks (e.g. t=4min, x=[\lkm,\lkm)). However, the truckdensity lanedistribution
is affected, in that trucks only use the right roadway lane. In the upstream region, where per
soncar densities are higher, the velocities of personcars are reduced to such an extent, that
trucks actively interact with slow personcars, and consequently need to reduce their velocity.
A jam forms between the personcar only region XQ° and the mixed region XQ{ . This area ap
pears due to fast personcars from the upstream region XQ° flowing into the lowvelocity
mixed region XQI. These fast personcars interact with trucks and slow personcars, and con
sequently need to slow down. This results in both a decreased velocity and an increased den
sity. The increase in the density results in even more interactions with slow vehicles. This
avalanchelike process causes the formation of congestion in the transitional area between the
two regions. This phenomenon is comparable to the formation of localized structures leading
to the development of phantomjams (cf. Kerner et al. (1996)).
Let us finally consider the tails of the regions XQ° and XQI. Clearly, none of the vehicles flow
back into the lowerdensity regions upstream of these tails. We may therefore conclude that
the model satisfies the socalled anisotropy condition (cf. Daganzo (1995)). That is, vehicles
mainly react to stimuli in front.
46 Transportation and Traffic Theory
personcars rightlane
personcars leftlane
trucks rightlane
trucks leftlane
5 10 15 10 15
road position [km] road position [km]
The second test case describes the traffic conditions on the 20km twolane ringroad at a lane
drop. The lane drop is located at X=[lkm,\3km). We will consider a leftlane drop (Figure 6).
Upstream of the lane drop, both personcars and trucks are present. We assume that at ?=0,
traffic conditions are in equilibrium (see Figure 8). The lanedrop is modeled by placing mo
tionless virtual vehicles on the left lane. The presence of these virtual vehicles causes an in
creased number of active interactions, yielding on the one hand an increase in the number of
lanechanges to the left lane, and on the other hand, an increase in the density on the left lane.
Macroscopic Traffic Flow Modelling 47
x=\Qkm
x=0km
Figure 7: Schematics of leftlane drop on a unidirectional twolane ringroad.
From the simulation results we see that initially, personcars and trucks flow into the left lane.
However, they change to the left lane quickly, resulting in an empty lane. Some of the person
cars change to the left roadway lane when arriving at the end of the lane drop at x\?>km.
Figure 8 shows that congestion occurs on the right roadway lane. That is, although vehicles
are able to flow into the right roadway lane without causing congestion upstream of the lane
drop, fast vehicles flowing into the right lane cause interacting with slow vehicle downstream
on the right lane need to reduce their velocity. This reduction in more profound, since no
overtaking possibilities occur. Since each interaction leads to velocity decrease, the right lane
is more susceptible to the seemingly spontaneous formation of jams due to vehicle interaction.
This paper presents a macroscopic multilane multiclass traffic flow model that is founded on
gaskinetic principles. The model dynamics are governed by processes of a convective nature,
acceleration towards the desired velocity, deceleration due to interaction, and various types of
lane changing. The model distinguishes constrained and freeflowing vehicles. It is cast using
the classspecific and lanespecific conservative variables density, momentum, and energy,
enabling a simplified model derivation and improved numerical analysis.
The model allows the specification of several parameters that define the characteristics of the
classes and the roadway lanes. Among these parameters are the vehicle length, the desired
velocity, the acceleration time, the reaction time, the constrained vehicle fraction, and the
lanechanging probability functions.
48 Transportation and Traffic Theory
The gaskinetic approach yields equilibrium expressions for the equilibrium momentum, en
ergy, and the distribution of densities over the roadway lanes. From these relations, expres
sions describing the equilibrium velocity, and velocity variance have been determined. The
equilibrium relations are functions of the userclass specific immediate overtaking probabili
ties, acceleration times, desired velocities, covariances between the velocity and the desired
velocity, and the asymmetric interaction between fast and slow vehicles of the same and dif
ferent userclasses.
densities
velocities
=5
2 so
lane drop
personcars rightlane
personcars leftlane
trucks rightlane
trucks leftlane
I 80
5 10 15 5 10 15
road position [km] road position [km]
From the testcase example of a twolane ringroad we conclude that the preliminary results of
macroscopic simulation are plausible. Trucks increasingly use the right lane when the person
car density increases, leaving the left lane for the faster personcars. The slower trucks are
Macroscopic Traffic Flow Modelling 49
virtually unaffected by faster personcars. However, when the velocity of personcars de
creases, trucks are affected by these slower personcars and consequently need to reduce their
velocity as well.
MLMC traffic operations near a lane drop were simulated. Again, the macroscopic MLMC
model was able simulate the traffic operations at the lane drop realistically. Moreover, also
incidents and onramps can be described.
Hoogendoorn (1999) identifies and remedies some of the model's current shortcomings. For
instance, he argues that the vehicular chaos assumption only holds for dilute traffic, due to the
increased correlation between vehicles for increasing traffic densities. This can be remedied
by considering the correlation of the velocities of unconstrained vehicles and platooning vehi
cles. Moreover, Hoogendoorn (1999) shows that platooning vehicles are able to accelerate
(to the desired velocity of the unconstrained platoon leader). Finally, the author incorporates
the vehicular space requirements in the MLMC dynamic equations.
REFERENCES
Bliemer, M. (1998). Multiclass travel time functions. Proceedings of the 6th meeting of the
EURO Working Group of Transportation.
Daganzo, C.F. (1995). Requiem for secondorder fluid approximations of traffic flow. Trans
portation Research B, 29, 277286.
Daganzo, C.F. (1997). A Continuum Theory of Traffic Dynamics for Freeways with Special
Lanes. Transportation Research B, 31, vol. 2, 83102.
Helbing, D. (1996). Gaskinetic derivation of NavierStokeslike traffic equations, Physical
Review E, 53, vol. 3, 22662381.
Helbing, D. (1997). Modelling multilane traffic flow with queuing effects. Physica A, 242,
175194.
Hoogendoorn, S.P. (1997). A Macroscopic Model for Multiple UserClass Traffic Flow. Pro
ceedings of the 3rd TRAIL PhD. Congress, vol. I.
Hoogendoorn, S.P. (1999). Multiclass Continuum Modelling of Multilane Traffic Flow. Dis
sertation Thesis, Delft University Press.
Hoogendoorn, S.P. and Bovy, P.H.L. (1998a). Multiple UserClass Traffic Flow Modelling 
Derivation, Analysis and Numerical Results. Research Report VK2205.328, Delft Univer
sity of Technology.
Hoogendoorn, S.P. and Bovy, P.H.L. (1998c). Continuum Modelling of Multilane Heteroge
neous Traffic Flow Operations. Research Report VK 2205.330, Delft University of Tech
nology.
Hoogendoorn, S.P. and Bovy, P.H.L. (1998d). Macroscopic Modelling of Multiple UserClass
Traffic Flow using Conservative Variables. Proceedings of the 6th meeting of the EURO
Working Group of Transportation.
50 Transportation and Traffic Theory
Hoogendoorn, S.P., and Bovy, P.H.L. (1998b). Modelling Multiple UserClass Traffic. Pre
print 980692 of the 1998 TRB annual meeting.
Kerner, B.S., Konhauser, P., and Schilke, M. (1996). A new approach to problems of traffic
flow theory, Proceedings of the 13th International Symposium of Transportation and
Traffic Theory, INRETS, Lyon, 119145.
Klar, A., R.D. Kiihne, and R. Wegener (1998). A hierarchy of models for multilane vehicular
traffic I: Modeling. To appear in SIAM J. Appl. Math.
Lebaque, J.P. (1996). The Godunov scheme and what it means for first order traffic flow
models. Proceedings of the 13th International Symposium of Transportation and Traffic
Theory, 647677.
Lyrintzis, A.D., Liu, G., Michalopoulos, P.G. (1994). Development and comparative evalua
tion of highorder traffic flow models, Transportation Research Record 1547, 174183.
PaveriFontana, S.L. (1975). On BoltzmannLike treatments for traffic flow: a critical review
of the basic model and an alternative proposal for dilute traffic analysis. Transportation
Research 5,9,225235.
Payne, HJ. (1979). FREFLO: A Macroscopic Simulation Model For Freeway Traffic. Trans
portation Research Record 772, 6875.
Prigogine, I., and Herman, R. (1971). Kinetic theory of vehicular traffic. American Elsevier
Publishing Co., New York.
Van Leer (1982). Flux vector splitting for the Euler equations. Proceedings of the 8th Interna
tional Conference in Numerical Methods in Fluid Dynamics, Berlin, SpringerVerlag.
51
Paul Nelson1
Alexandras Sopasakis
Department of Mathematics, Texas A&M University, College Station, Texas 778433368, U.S.A.
ABSTRACT
Attempts to improve on the basic continuum (hydrodynamic, macroscopic) model of traffic flow,
as developed in the seminal 1955 paper of Lighthill and Whitham, have largely followed the
1971 work of Payne in retaining the continuity equation, but replacing the classical traffic stream
model by a "dynamic traffic stream model" (or "momentum equation"). In the present work it
is suggested that, whatever may be the advantages and disadvantages of the Payne models, they
should not properly be regarded as the traffic flow analog of the NavierStokes equations of fluid
dynamics. Further, the ChapmanEnskog asymptotic expansion in a small parameter is shown
to lead to an alternate class of models that seem to have a more legitimate claim to that distinc
tion. Details of this expansion, about the stableflow equilibria of the PrigogineHerman kinetic
equation and in the case that the passing probability and relaxation time are constant, are pre
sented to orders zero and one. The zeroorder and firstorder expansions correspond respectively
'Also affiliated with the Departments of Computer Science and of Nuclear Engineering at Texas A&M Univer
sity.
52 Transportation and Traffic Theory
to the LighthillWhitham (LWR) model, and to the LighthillWhitham model with a diffusive
correction. These are suggested to be the correct trafficflow analogs of respectively the Euler
and NavierStokes equations of fluid dynamics. Results of a numerical simulation for a simple
trafficflow problem suggest that the diffusive term represents a correction to the LWR model that
captures, to some extent, effects stemming from the fact that vehicles actually travel at various
speeds. (By contrast, LighthillWhitham models proceed as if all vehicles travel at the average
speed corresponding to the density of vehicles in their immediate vicinity.) Some further related
work is suggested.
INTRODUCTION
Traffic stream models (fundamental diagrams, flow/density relations) date back to (at least) the
work of Greenshields (1934), and are often considered as the foundation of capacity analy
sis (Transportation Research Board, 1985). Lighthill and Whitham (1955), and independently
Richards (1956), observed that if a traffic stream model is supplemented by the equation of conti
nuity (conservation of vehicles), then the resulting partial differential equation presumably could
be solved, subject to suitable initial and boundary conditions, for the concentration (and hence
the mean speed and flow) as a function of location and time. This LighthillWhithamRichards
model (LWR model) is widely considered the most fundamental continuum (macroscopic, hy
drodynamic) model of traffic flow.
Unfortunately, observational data (e.g., Drake, Schofer and May, 1965) are at best ambiguous
as regards existence of traffic stream models.2 This has led some workers (e.g., Ceder (1976),
Hall (1987), Disbro and Frame (1989)) to suggest alternative macroscopic models that differ
from LWR models in a revolutionary manner. However, the usual approach to alternative con
tinuum models is more evolutionary, in following the work of Payne (1971) by supplementing
the continuity equation with a "dynamic traffic stream model" (or "momentum equation") that is
itself a differential equation, as opposed to the classical "static" traffic stream models that do not
contain derivatives. This gives rise to a system of two firstorder partial differential equations in
two unknowns, typically concentration and mean speed. Such systems are customarily known as
higherorder models.
Higherorder continuum theories of traffic flow frequently (e.g., Helbing, 1996b; Kerner and
Konhauser, 1993) are considered as analogs of the NavierStokes equations of fluid flow, and
correspondingly the LWR theory is viewed as analogous to the Euler equations of fluid flow. The
present work is intended as somewhat of a counterpoint to the first of these views. Specifically,
the primary objectives of this paper are as follows:
2
This difficulty was already noted by Lighthill and Whitham (1955, p. 344).
The ChapmanEnskog Expansion 53
1. It is suggested, in some detail, that there is a viewpoint from which the often repeated sup
posed analogy between current higherorder models of traffic flow and the NavierStokes
equations of fluid dynamics is questionable.
3. Some initial developments of this approach via the ChapmanEnskog asymptotic expan
sion are presented. These developments are based on the classical PrigogineHerman
(1971) equation of the kinetic theory of vehicular traffic,
The first two of these objectives are met in the following section, in the context of a general crit
ical review of Continuum Models of Vehicular Traffic. The remainder of this work is devoted
to the third objective. This is initiated with a section in which the salient features of Kinetic
Equations for Vehicular Traffic are collected. The focus in this section is particularly upon the
equilibrium solutions of the PrigogineHerman kinetic equation, and the necessity to limit further
considerations to the regime of stable flow, because of the form of these equilibrium solutions.
The next section is devoted to the The ChapmanEnskog Expansion, and contains the central
new results presented in this work. First the ChapmanEnskog expansion is described generally,
in the context of the PrigogineHerman kinetic equation. Then the zeroorder instance of this
expansion is shown to give rise to an LWR model, with traffic stream model completely defined
in terms of the knowns of the underlying PrigogineHerman equation. Finally, the firstorder
ChapmanEnskog expansion is shown to give rise to a continuum approximation that consists
of an LWR model with an additional "diffusive" term. The section on Computational Results
provides a comparison of numerical results, and a discussion of the interpretation of these differ
ences, for this firstorder diffusive approximation and the LWR model that is the corresponding
zeroorder approximation, in a simple instance of traffic flow. The paper closes with a section of
Conclusions, which primarily consists of suggestions for further related work.
models that potentially improve upon LWR models, but in a fundamentally different manner
from that underlying current higherorder models.
9 = Q(c), (1)
where q is flow, c is density, and Q is a known function. The equation of continuity (conservation
of vehicles) will be expressed as
^ ++ ^in
dt dx
The two can be combined to give a determined system consisting of a single (nonlinear) partial
differential equation in one unknown (the density),
dc .. . dc
ii+'(c)^ =0'
(q1 = dq/dc) that presumably can be solved, subject to suitable initial and boundary conditions,
for the density (and hence the mean speed and flow) as a function of location and time.
The latter equation is the most concise mathematical form of a LWR model. These are widely
used, at least conceptually. The possibility of further improved continuum descriptions of traffic
flow already was considered by Lighthill and Whitham (1955, p. 344), who suggested adding
"diffusion" (representing adjustments by drivers to the concentration slightly ahead) and "in
ertia" (representing the nonzero time required for accelerations or decelerations) effects to the
continuity equation. However, subsequent development of presumably better models, beginning
with the work of Payne (1971), has instead been directed toward socalled higherorder models.
In these, the "static" traffic stream model (1) is replaced by a "dynamic" counterpart that consti
tutes a second differential equation (in addition to the continuity equation). A typical instance
of such a "dynamic traffic stream model" (also known as "momentum equation") is that due to
Kiihne and Beckschulte (1993),
dv dv 1 ,Tr, . , oldc d2v
a+'S^MO^sisi. (2)
Here v — q/c is the mean speed, V(c] = Q(c)/c is the mean speed at concentration c according
to some associated static traffic stream model, and r, CQ, v are coefficients that presumably are
to be provided from observations. Such higherorder models have been rather widely studied in
recent years, especially in the physics literature. An exhaustive list of references, and an excellent
systematic overview of this subject (including references to workers who have used somewhat
different forms of a dynamic traffic stream model), appear in recent works by Helbing (1995a,
1995b, 1996a).
The precise relationship between LWR and higherorder models has been the subject of consid
erable discussion in the literature (e.g., Kiihne and Beckschulte, 1993; del Castillo, Pintado and
3
The terminology used here is that of Gerlough and Huber (1974), and of May (1990).
The ChapmanEnskog Expansion 55
Benitez, 1993; Daganzo, 1995b; Helbing, 1996a). The view of the present work is that, in regard
to either of these types (i.e., LWR or higherorder) of continuum models of traffic flow, there are
three fundamental issues that arise:
Validity: What is the basis, either empirical or theoretical, for the associated (static or dynamic)
traffic stream model?
Mathematical solvability: What additional conditions (boundary, initial, diverge) are necessary
to provide a solution of the traffic flow model that represents, to some reasonable degree
of approximation, what one actually observes in a real traffic network?
Computational solution: How does one effectively obtain these "real" solutions on a modem
highspeed digital computer?
A brief overview of the authors' view of the current and historical understanding of each of these
issues will be presented in, respectively, the following three subsections.
VALIDITY
In the trafficflow literature it usually is indicated that static traffic stream models should be de
termined from observations. However, as already mentioned in the Introduction, it is wellknown
that observations tend to show considerable scatter in observed flows (or speeds), especially in
the region of unstable flow. This phenomenon has led some workers (e.g., Ross (1988), p. 422;
Kiihne and Beckschulte (1993), p. 367) to doubt existence of traffic stream models, in the sense
of a relationship that expresses mean speed (or flow) as a singlevalued function of concentration.
Others, similarly motivated, have suggested a variety of alternatives to traditional continuum the
ories (Ceder (1976), Hall (1987), Disbro and Frame (1989)), although none of these as yet seems
to have found significant application to the quantitative modeling of traffic flow.
Higherorder models seem primarily to be validated on the basis of adjusting the variety of pa
rameters that they contain (e.g., r, CQ, v) to obtain a solution that provides a reasonable fit to
observations. The authors are unaware of any efforts to validate dynamic traffic stream mod
els themselves by direct comparison with observations. Any such program would appear to be
rather difficult to effect operationally, because (e.g.) each of the four derivatives appearing in Eq.
(2) (especially the second derivative) would be rather sensitive to uncertainty in measurements,
and then these errors could be subject to even greater magnification from combining them alge
braically as in Eq. (2). Additionally, most dynamic traffic stream models invoke an underlying
static traffic stream model, and it is unclear how such a dynamic model can be valid without
some degree of validity attaching to the associated static model. (The semiviscous and viscous
models of Lyrintzis, Liu and Michalopoulos (1994) are exceptions. These dynamic traffic stream
56 Transportation and Traffic Theory
models were developed on the basis of consideration of a number of carfollowing models, and
seem to give rather good results.)
There seems to have been relatively little work taking the form of a comparison of both LWR
and higherorder models against some independent standard (e.g., field data). It appears likely
that any early efforts (e.g., prior to the 1971 paper of Payne) to compare LWR models against
field data were significantly hampered by the primitive state, at that time, of the development of
numerical methods known to approximate the entropy solution that is thought (Ansorge, 1990)
to be relevant to traffic flow (see the following subsection on computation for more details). It
is a hypothesis that is consistent with the authors' reading of the literature that such efforts were
not undertaken in the U.S. soon after 1971, because of an "official" commitment to higherorder
models, notably as implemented in the wellknown FREFLO code (Payne, 1979). The work
of Michalopoulos and coworkers is an exception to the apparent general lack of an effort to
provide an objective comparison of LWR and higherorder models. This work is particularly
noteworthy, because it seems to represent one of the first attempts to use, for LWR models,
modern numerical methods that are known to approximate the entropy solution. Michalopoulos,
Beskos and Lin (1984, p. 418) concluded, in a comparison against field data of LWR ("simple
continuum") models as solved by three distinct numerical methods, that "despite the absence of a
momentum equation, the close agreement of the numerical results with field data . . . is notable."
Michalopoulos and Beskos (1984) carried out a comparison of LWR and higherorder models, as
solved using a variety of numerical methods, against detailed microscopic simulations, and con
cluded that "existing high order models are only slightly better than the simple continuum model
(i.e., LWR) at uncongested flows; however at congested flow conditions the simple continuum
model performs better ..." (p. 103). (More recent results by Lyrintzis, Liu and Michalopoulos
(1994) seem more favorable to highorder models.) Daganzo (1995b, p. 284) has noted, in re
gard to highorder models, that "although reasonable fits to specific datasets can be achieved
with adjusted models that include many parameters ... these limited successes do not constitute
a validation," and that "given enough degrees of freedom ... a good fit should be expected."
MATHEMATICAL SOLVABILITY
Mathematical solvability is best known (and understood) in the context of LWR models. The lat
ter are a class of nonlinear conservation laws that can have multiple solutions (even infinitely
many) that satisfy given initial conditions (cf. Ansorge, 1990; LeVeque, 1990). This non
uniqueness phenomenon requires imposition of some additional condition to select the solution
that represents actual traffic flow. This is typically done by selecting the solution that satisfies the
socalled "entropy condition." But in fact this condition originates from application of the second
law of thermodynamics in fluid flow (Ansorge, 1990), and it is unclear why one should expect
such a condition to hold sway in traffic theory (the valiant effort of Ansorge (1990) toward an
The ChapmanEnskog Expansion 57
interpretation notwithstanding). Lebacque (1997) has recently noted that there are conditions un
der which the entropy solutions imply a clearly nonphysical infinite acceleration, and suggested
that there are circumstances (e.g., entrance ramps) in which this is perhaps so badly wrong that
some other criterion should be used. Newell (1993) suggested the geometrically motivated cri
terion of selecting as the solution of interest the lower envelope of all solutions, when the LWR
model is reformulated in terms of A(x, t), the cumulative flow past x at time t, as the dependent
variable. A version of this criterion that is suitable for digital computation does not seem to have
been developed, nor does its precise relationship to the entropy condition seem to have been clar
ified. Boundary and intersection conditions for LWR models have been discussed by Lebacque
(1996).
Uniqueness issues do not seem to have been systematically studied in conjunction with higher
order models. Lebacque (1995) notes that Schochet (1988) has shown:
i) Existence and uniqueness of solutions to Payne's (1971) higherorder model, which corre
sponds to (2) with v = 0.
ii) The solutions of Payne's model converge, as T —> 0, to the solution of the continuity equation
with an added diffusionlike term,
iii) The solution of this "continuity plus diffusion" equation converges, as i>0 —> 0, to the entropy
solution of the LWR model.
Thus this "continuity plus diffusion" model plays the same role in justifying the entropy con
dition for the LWR model as the NavierStokes equations play for the Euler equations in fluid
dynamics. This is rather strongly suggestive that the "correct" higherorder generalization of
the LWR model is this "continuity plus diffusion" model, or something very similar. The first
order form of the continuum approximation corresponding to the ChapmanEnskog asymptotic
expansion has exactly that form, as described in the following section on The ChapmanEnskog
Expansion.
COMPUTATIONAL SOLUTION
Numerous discrete approximations, dating back to the 1960's, have been developed (see LeVeque
(1990); Ansorge (1990); Leo and Pretty (1992); and extensive references cited in these works)
that converge, in the finemesh limit, to the entropy solution of the Euler equations of fluid
dynamics. Although the Euler equations are the fluidflow counterpart of the LWR equation of
traffic flow, the first applications of such entropy methods to the LWR equations of fluid flow
seem to have come somewhat later (Lebacque, 1984; Michalopoulos, Beskos and Lin, 1984).
58 Transportation and Traffic Theory
Other workers (Bui, Nelson and Narasimhan, 1992; Leo and Pretty, 1992; Daganzo, 1995a) have
subsequently developed and applied such methods within the theory of traffic flow.
There is some evidence that this apparent early lack of awareness of the necessity to use, for
LWR models, discrete approximations that converge to the entropy solution (and of the exis
tence of such approximations) provided some of the disbelief in LWR models that led to the
original motivation for the development of higherorder models. For example, as recently as ten
years past (Ross, 1988, p. 422) the LWR model was criticized in terms that make it clear the
author was thinking of a discrete approximation that did not predict maximal flow between an
upstream region at jam density and a downstream region devoid of traffic, and therefore could
not correctly capture the "acceleration wave" that is the entropy solution of this problem (and the
traffic flow equivalent of a rarefaction wave in fluid flow). (See also Newell, 1989; Ross, 1989;
Nelson, 1995b.) Similarly, the seminal work on higherorder models (Payne, 1971, Eq. (7.10))
approximates the flow between two cells as the average of the two densities times the average
of the two velocities. This discrete approximation shares with the upstream method the property
that it incorrectly predicts zero flow between an upstream region at jam density and an a down
stream region devoid of traffic, and therefore it cannot converge to the desired entropy solution
of an LWR model. The discrete approximation apparently employed in the original version of
FREFLO (Payne, 1979) is the upwind method, which has this same defect. Because of this, one
would expect FREFLO to exhibit some characteristics of poor performance (e.g., slow conver
gence with mesh refinement) under conditions such that this model approximates the LWR model
(see the above discussion of the work of Schochet, 1988). It is a reasonable suspicion that this
contributes to the deficiencies of FREFLO and related codes that have been noted by numerous
workers (Rathi, Lieberman and Yedlin, 1987; Ross, 1988; Leo and Pretty, 1992; Michalopoulos,
Yi and Lyrintzis, 1993; Lyrintzis, Liu and Michalopoulos, 1994; and other works cited in these).
On the one hand it appears almost certain that there exist traffic phenomena (e.g., the "sponta
neous traffic jams" of Kerner, Konhauser and Schilke, 1995) that are unlikely to be describable
by LWR models. Therefore, one suspects it should be possible to develop macroscopic (hydrody
namic) models of traffic flow that improve upon the LWR model. On the other hand, higherorder
models of the type that have been introduced to date provide, at best, marginal improvements on
LWR models. (Even the explanations given by Kerner, Konhauser and Schilke (1995) for sponta
neous traffic jams seem more based on ad hoc physical models of traffic flow than on the specific
mathematical higherorder model that is nominally introduced to provide an explanatory frame
work.) This raises the question of what alternative type of continuum model might be developed
in an effort to improve on LWR models. The overall objective of the present work is to describe
initial results for an alternative approach that seems to lead to models more nearly consistent
The ChapmanEnskog Expansion 59
with the improvements suggested in the seminal work of Lighthill and Whitham (1955, pp. 344)
than are the higherorder models that employ a dynamic traffic stream model.
The motivation for the idea that there should be some alternative treatment arises from the obser
vation that the often repeated analogy between current higherorder models of traffic flow and the
NavierStokes equations of fluid dynamics is questionable, as follows. The compressible Euler
equations of fluid dynamics are a system of five equations in five unknowns, with each equation
representing conservation of some quantity (particle number, three components of momentum,
and energy) that is conserved in molecular interactions, and each of the five unknowns represent
ing a macroscopic analog (density, fluid velocity, temperature) of one of these microscopically
conserved quantities. The LWR model of traffic flow likewise is a single equation, in a single un
known, that represents conservation of the sole quantity (number of vehicles) that is conserved
in vehicular interactions, and the single unknown (concentration, or density) is a macroscopic
analog of that microscopically conserved quantity.
Thus, it seems reasonable to regard LWR models as the traffic flow analog of the compressible
Euler equations of fluid dynamics. But the NavierStokes equations of fluid dynamics likewise
consist of five conservation laws in these same five unknowns as the compressible Euler equa
tions. These are merely of a different form from the Euler equations, in that viscosity and dif
fusion are now represented, but by terms expressed in the same five dependent variables (and
their spatial or temporal derivatives). On the other hand, as described in the introduction, current
higherorder models of traffic flow introduce a "dynamic traffic stream model" that has the form
of an additional conservation law, and an additional independent variable (e.g., mean speed)
that is not a macroscopic analog of some quantity that is microscopically conserved in vehicular
interactions. On this basis alone it seems not quite appropriate to regard current higherorder
models as trafficflow analogs of the NavierStokes equations.4 Rather they seem to stem from
following the/orm of the development of the NavierStokes equations of fluid dynamics, as op
posed to following the spirit of such development, with due regard for the differences between
fluids and traffic streams.5
If one accepts this contention that current higherorder macroscopic models of traffic flow are
not the proper analogs of the NavierStokes equations of fluid flow, then it follows that the true
analog remains to be discovered. How should it be found? The answer proposed here is a de
velopment following the lines of the ChapmanEnskog asymptotic expansion of the Boltzmann
equation and its solutions that has been shown (e.g., Chapman and Cowling, 1952; Cercignani,
4
In itself this says nothing about the validity or invalidity of such higher order models. In fact it seems quite
reasonable to consider them as analogs of the Thirteen Moment method of Grad (1949), which is a quite respectable,
although seldom used, description of fluid dynamics.
5
Kerner and Konhauser (1993) even simply directly import the NavierStokes equations from fluid flow into
traffic flow. Daganzo (1995b) has emphasized the differences between traffic and ordinary fluids, and warned of the
dangers of pushing too far the interesting, and sometimes useful, analogy between the two.
60 Transportation and Traffic Theory
1988; Liboff, 1990) to lead to a hierarchy of macroscopic equations for rarefied gases. The
(compressible) Euler equations are the lowestorder (zeroorder) member of this hierarchy, the
NavierStokes equations are the firstorder approximation, and the seldomused Burnett equa
tions are the secondorder approximation. The proposed approach thus follows the spirit of the
standard theoretical development of macroscopic fluid dynamic equations, as opposed to simply
adopting the form of the fluiddynamic result of that development. In particular, the specific
characteristics of vehicular flow, as opposed to fluid flow, presumably will be represented in such
an approach, to the extent that these characteristics are incorporated in the underlying kinetic
equation that forms the starting point for the development.
The foundation of any such development is a suitable kinetic equation for vehicular traffic. Such
kinetic equations are the subject of the next section.
In this work the venerable kinetic equation of Prigogine and Herman (1971) will be employed.
One certainly can raise legitimate questions about the degree of validity of some aspects of the
PrigogineHerman kinetic equation, particularly the somewhat phenomenological "relaxation
term." As a result, other kinetic equations for vehicular traffic have been suggested; these are
summarized briefly in the concluding subsection of the present section, especially as regards the
state of knowledge of their equilibrium solutions and their prior use in developing continuum
models. In the first subsection the equilibrium solutions of the PrigogineHerman kinetic equa
tion are described. This description will be primarily for the modified PrigogineHerman kinetic
model recently considered by Nelson and Sopasakis (1998). Their equilibrium solution contains,
as a special case, that originally given by Prigogine, Herman and Anderson (1962).
The ChapmanEnskog Expansion 61
Here / is the density function for the distribution of vehicles in phase space, so that f ( x , v, t) dx dv
is the expected number of vehicles at time t that have position between x and x + dx and speed
between v and v + dv, c — c(x, t) is the spatial density of vehicles (vehicles per unit length), /0
is the corresponding distribution function for the desired speed of vehicles, v is mean speed, and
P, T are respectively the relaxation time and the passing probability.
The equilibrium solutions of the PH model are given by setting the righthand side of (3) equal
to zero, and solving for / = feq. Suppose all desired speeds lie between some positive lower
limit w_ and some upper limit w+. Then there is some critical density c^t,6 defined as the root
of
1
such that the equilibrium solutions of the PrigogineHerman model are as follows:
Here 8 is the delta function of Dirac, and a is not determined as a function of c, but rather
merely required only to satisfy a min (c) = max{0, 1  re2( ^_ p) /J_+ £^ dv} < a <
1
~ rc3(ip) fw ^jr^dv = amax(c). Further, £ = £(c, a) is defined implicitly, as a
function of both a and c, byTc2(}_P) /J_+ ^^ dv = 1  a.
This means that, in the regime of unstable flow, the mean speed (or flow) is a function of two
independent parameters. In the above description these parameters were taken as density and the
fraction (a) of the traffic in the collective ("highly platooned") mode. It is equally possible to
take them as the density and the speed (Q of the collective flow. Notice that c^u is infinite, and
therefore the collective flow does not exist, unless /0 is such that / *°^ dv is finite.
6
The critical density is to be interpreted as the smallest density at which equilibrium flow involves some vehicles
travelling slower than any driver wishes. This differs from the definition of the 1985 Highway Capacity Manual
(Transportation Research Board, 1985, pp. 16) as the density at which maximum (i.e., capacity) flow is reached.
However, the values of the two are quite close, under reasonable assumptions on /Q.
62 Transportation and Traffic Theory
In any case a (classical) traffic stream model is obtained in the stable regime, as
q(c) = I vfeq(v;c)dv.
Prigogine, Herman and Anderson (1962) (see also Chap. 4 of Prigogine and Herman, 1971) as
sumed that w_ = 0. In that case it turns out that a mai (c) = a m in(c) = a(c), so that again a (clas
sical) traffic stream model is obtained, as q(c) — f v f e q ( v ; c, a(c)) dv. In this case £(c, a) = 0,
so that the critical density is the smallest density at which some traffic is stopped, at equilibrium
flow, and the component (or mode) a6(v) represents a "collective flow" in which some fraction
(a) of the vehicles are stopped.7 Prigogine and Herman (1971, Chap. 4) provide graphical pre
sentations of a number of such classical traffic stream models, both with and without collective
flows.
But, the fact that equilibrium solutions lead to a traffic stream model is closely associated with
the fact that the equilibrium solutions are a oneparameter family, and that it is possible to take
the density as that parameter. This is the normal expectation. That expectation has its roots
in the Boltzmann equation of the kinetic theory of gases, where it is typically found that the
parameters required to characterize an equilibrium solution are macroscopic counterparts of the
microscopic invariants of molecular interactions, so that the number of parameters is the same as
the number of invariants.8 As there is only one invariant in vehicular interactions, the number of
vehicles, one therefore expects the equilibrium solutions of a kinetic equation for vehicular traffic
to comprise a oneparameter family. That expectation is realized for stable flow, and even for
unstable flow, provided one makes the somewhat unrealistic assumption that there exist drivers
desiring arbitrarily small speeds. It is therefore perhaps surprising9 to find that this expectation
is not met, in the general case of unstable flow.
To emphasize the radical distinction between the stable and unstable flow regimes, note that in
the unstable flow regime the classical static traffic stream model generally becomes a surface in
threedimensional c/a/v (or c/C/v) space, rather than a twodimensional curve. (See Nelson
and Sopasakis (1998) for figures illustrating this phenomenon.) The projection of this surface on
twodimensional c/v space is, of course, a region rather than a curve. This is qualitatively similar
to what is observed in attempted measurements of traffic stream models.
7
Thus, in the unstable flow regime, a is very similar to the "percent time delay" that is used as a primary level
of service indicator for twolane highways in the 1985 Highway Capacity Manual (Transportation Research Board,
1985). This quantity is defined as "the average percent of time that all vehicles are delayed while traveling in
platoons due to the inability to pass."
8
The equilibrium solutions of the Boltzmann equation depend upon five parameters, one for each of the five
invariants of molecular interactions. Specifically, these invariants are mass, three components of momentum, and
energy. The corresponding parameters can be taken as density, three components of the mean speed, and tempera
ture. These are the parameters that normally are taken as characterizing the wellknown Maxwellian distributions,
which are the equilibrium solutions of the Boltzmann equations, for some intermolecular force laws.
9
And emphasizes the danger of pushing fluid flow analogies too far in traffic flow.
The ChapmanEnskog Expansion 63
The continuum models to be developed in the following section will be based upon the equi
librium solutions described above. However, in order to obtain a LWR model as the zeroorder
approximation (i.e., to obtain a static traffic stream model from the equilibrium solution) it ap
pears to be necessary to restrict these developments to the stableflow regime (i.e., densities
below critical). That restriction will be followed in the present work. The development of suit
able continuum models for the unstableflow regime is presently an open problem. Note that the
nature of the equilibrium solution in the region of unstable flow appears to cast some doubt on
the validity of LWR models per se, in this region.
Approaches to obtaining LWR models from the PrigogineHerman kinetic theory of vehicular
travel have been discussed above, along with a theoretical basis for the limit of their range of
validity. The objective of this subsection is to briefly summarize work on the development of
continuum models (LWR or higherorder) from other microscopic models. The central focus
will be on recent work based on kinetic models other than that of Prigogine and Herman (1971).
Somewhat different approaches to this subject seem to have been initiated independently by
Nelson (1995a) and by Helbing (1995a). Nelson (1995a) developed a kinetic equation in which
the phenomenological relaxation term in the PrigogineHerman equation is replaced by a more
fundamental representation of speedingup (and passing is neglected). He and coworkers (Bui,
Nelson and Sopasakis, 1996; Nelson, Bui and Sopasakis, 1997) subsequently showed that the
equilibrium solutions of this kinetic equation lead to a traffic stream model that provides a fit to
data that is as close as that of any of several classical such models, but with use of only parameters
taken directly from a microscopic model of driver behavior (i.e., no "free" parameters available
to provide the best fit to observations). Note that the equilibrium solutions found from this
kinetic equation do not display collective flow at nonzero speeds, and thus the newer results of
Nelson and Sopasakis as described above, along with the empirical scatter of flow under unstable
conditions, bring into doubt the applicability in the unstable regime of LWR models based upon
this traffic stream model.
On the other hand, Helbing (1995a, 1995b, 1996a, 1996b) has based his approach upon the
PaveriFontana (1975) extension of the PrigogineHerman (1971) kinetic equation. This exten
sion includes the desired speed as an additional independent variable, on the same footing with
the desired speed, as opposed to being a parameter as in the original PrigogineHerman model.
(Daganzo (1995b) has emphasized the necessity of this, in order that the desired speed be a prop
erty of vehicles and drivers, not of roads.) Helbing (1996a) has developed a number of hydrody
namic models, by proceeding in the manner that frequently is used in fluid flow; i.e., by taking
64 Transportation and Traffic Theory
loworder polynomial moments of the underlying (PaveriFontana) kinetic equation, then using
ad hoc techniques to close these (i.e., reduce them to determined systems). Note also that Hel
bing (1995a, pp. 383384) was unable to obtain the equilibrium solutions of the PaveriFontana
kinetic equation, but rather assumed them to be Gaussian (Maxwellian) distributions, based upon
empirical data that was specifically indicated as being for stable traffic. This certainly brings into
question the validity for unstable flow of the continuum equations thus obtained.
Wegener and Klar (1996) built upon the approach of Nelson (1995 a) to obtain a more realistic
kinetic model that uses first principles to incorporate speedingup. They studied numerically the
equilibrium solutions, and associated traffic stream models, of their class of kinetic equations,
but no evidence of a collective flow seems to have been uncovered. These workers (Klar and
Wegener, 1997) also developed an "Enskoglike" kinetic model that differs from all previous
kinetic models of vehicular traffic in that the effect of nonzero vehicle length is taken into ac
count in the slowingdown and speeding up of vehicles. They derived from this approach a set of
macroscopic equations formally similar to the higherorder model of Payne (1979), but with co
efficients computable from microscopic considerations, and that clearly showed the importance
of including nonzero vehicle lengths. They also obtained computational comparisons between
the solution of the kinetic (Enskog) equation and their hydrodynamic model, for a rather real
istic lanedrop problem. It is not yet clear how suitable their particular kinetic equation might
be for unstable flow (e.g., whether it exhibits collective flow), but it seems likely that ultimately
the effect of nonzero vehicle length must be incorporated into any realistic treatment of traffic
flow. Finally, Klar and Wagner (to appeara, to appearb) are currently pursuing very interesting
studies of multilane kinetic models.
For completeness, two further lines of development should be mentioned. First, Phillips (1977,
1979, 1981) also has developed continuum models on the basis of an underlying kinetic equa
tion. However, the underlying equilibrium solution does not contain a collectiveflow mode, so
that it appears unlikely this continuum model will be suitable for modelling unstable flow. Sec
ond, there is an extensive older literature on obtaining traffic stream models from carfollowing
models. Most of this is based on the assumption of steady state conditions (constant headway;
see Nelson, 1995b), and thus seems unlikely to be helpful in providing insight into the collective
dynamics underlying unstable traffic flow. However, a paper of Newell (1962) that is perhaps
less wellknown than deserved is specifically directed toward adapting carfollowing models to
unstable flow.
Before turning to the development of continuum models via asymptotic expansion of the Prigogine
Herman kinetic equation, it is appropriate to note a significant difference between this approach
and that in the kinetically based developments of continuum descriptions of traffic flow in some
of the references cited above. In these previous works the approach is to take the first two poly
nomial moments of the kinetic equation, and then introduce ad hoc approximations as necessary
The ChapmanEnskog Expansion 65
to close the system (i.e., reduce to two the number of unknowns = dependent variables). By
contrast, in the approach via asymptotic expansions the number and nature of the moments of
the kinetic equation that are introduced at any order are determined automatically by the require
ment that the next higherorder approximation exist. Further, the number and nature of these
moments usually are very closely related to the number and nature of the quantities conserved
in the interactions described by the particular kinetic equation (invariants). For vehicular kinetic
equations one expects only one quantity (the number of vehicles) to be conserved, and therefore
the associated continuum models to involve only one moment (as opposed to the two moments of
current higherorder models) in any corresponding continuum equation. In this respect a kinetic
model of vehicular traffic is perhaps closer to the Lorentz model of the Boltzmann Equation, for
which number of particles is also the sole conserved quantity, than to the Bojtzmann equation of
the kinetic theory of gases. For this model, it has been shown analytically (Hauge, 1969) that the
nthorder ChapmanEnskog approximation has the form of a single partial differential equation
of order 2n. One can perhaps reasonably anticipate somewhat similar results for Chapman
Enskog approximations in the kinetic theory of vehicular traffic.
Here the arguments x, v and t have been omitted, in order to keep the notation as concise as
possible. Additionally, w+ is the maximum speed desired by any driver, as in the preceding
section, and <p0 is the reduced desired speed distribution, as defined by (pQ = /0/c, and otherwise
the notation is as explained above in conjunction with Eq. (3). For present purposes <^o is assumed
to depend only on v (in particular, to be independent of c), and further the relaxation time (T)
and passing probability (P) are assumed to be constants, independent of concentration (c).
The beginning point for the ChapmanEnskog approximation is to further rewrite the Prigogine
Herman kinetic equation (6) in the form
ef = K f . (7)
Here / and Kf are respectively the lefthand and righthand sides of (3), and e is a parameter that
is introduced as a formal device to separate (3) into (an infinite number of) more easily solved
66 Transportation and Traffic Theory
is then sought. However, the density (the sole "hydrodynamic" variable, in the case of traffic
flow) is specifically not expanded in a power series in e. That means it is required that the
densities corresponding to the higherorder corrections to the distribution function are required
to be zero, w+
f(n\x,v,t)dv = Q,n = l,2,.... (9)
/
Thus density is carried entirely by the zeroorder approximation to the distribution function,
f
f0(0)
(x,v,t)dv. (10)
Jo
In the ChapmanEnskog expansion the (partial) time derivative in the PrigogineHerman equation
(7) is further expanded as
fc=0
f\(k\
(In the last equality the expansion (8) has been used, with the ^ treated as linear operators.)
Here the —^— are to be determined so as to ensure existence of the terms of the formal solu
tion (8). That is, the —^— are to be regarded as unknowns, to be determined by substituting
the expansions (8) and (11) into the PrigogineHerman (7), in exactly the same way as the various
terms in the expansion (8) (i.e., the /^) are determined.
The result of carrying out these substitutions, and equating coefficients of like powers of e is
K/ (0) = 0, (12)
at order n = 0. Otherwise it is
Here L(^) is, for any function g, the linear operator defined by
1
w+
/
and
S
Jo
The ChapmanEnskog Expansion 67
where, as usual, the dependence on x and t is understood, but is not explicit in the notation. (The
assumption that P and T are constants, independent of c, is used in obtaining these equations.)
If / (0) , /(1), / (n ~ 1) exist, then (13) has a solution for /^^ if, and only if, the righthand side is
an element of the range of L(/^), or equivalently it is orthogonal to the null space of the adjoint
of L(/<°>), notationally L(/<°>)*. But L(/<°>)* is given by
and it is readily shown from this representation that the null space of L(/(°))* is exactly the one
dimensional space of constants. Thus a necessary and sufficient condition for the equations (13)
to have a solution for the f^ is the compatibility conditions that the integral over 0 < v < w+
of their respective righthand sides to be zero. But
where Eqs. (9) and (10) have been employed. Similarly, (9) implies that f™+ Sndv = 0, for
n > 1. Therefore, the compatibility conditions for Eqs. (13) are the (LWRlike) equations
^("l)/> /^(nl)
= 1,2,..., (15)
ot ox
where
. f +
(16)
Jo
If (15) is multiplied by e""1, and the results summed over all n = 1 , 2 . . . , then the result can be
written as
n=0
Here the definition of the expansion (11) has been employed. This equation, which arises as
a consequence of the basic assumptions of the asymptotic expansion rather than from ad hoc
assumptions, is the basis for the continuum approximations arising from the ChapmanEnskog
approximation, as follows. If the q^ are computed by solving Eqs. (13), then it turns out that
q^ depends on x and t only through c(x, t} and its spatial derivatives up to order n. If the sum
in (17) is truncated at n = N, and e set to unity, then the resulting partial differential equation of
order ./V is the continuum form of the ./Vthorder ChapmanEnskog approximation. The details
of these continuum approximations will now be developed, for ./V = 0, 1, in the following two
respective subsections.
68 Transportation and Traffic Theory
From (4) and (8) the zeroorder ChapmanEnskog solution of the PrigogineHerman kinetic
equation, for the stable flow regime (c < c^it), is given by
f (v.c] _
_ (o) _ ^ * 'AM
/f _ /f MT(1_p)u_c(c).
(19)
Note that the stable flow regime is defined quantitatively by c < c^n := rf°i°L > so that neces
sarily C(c) is negative in this regime.
Eqs. (18) and (19) comprise a LWR model. As a partial differential equation for the zeroorder
approximation to the concentration, this LWR model reads as
dQ0 _
(20)
~°'
Subject to suitable auxiliary (initial and boundary) conditions, this LWR model presumably de
termines c(°) as a function of position and time. Note that the formal asymptotic procedure gives
rise to both the conservation law and the traffic stream model, and that the latter is expressed in
terms of the given data in the underlying (PrigogineHerman) kinetic model.
As already mentioned, the starting point for the computation of /(*), and hence q^l\ is the in
stance n — 1 of (13). The compatibility condition (15) for this instance is
0(o)c ^(0)
— + ^— = 0, n = 1 , 2 , . . . .
dt dx
The righthand side of the relevant instance of (13) is then
0fO) dc d(0) f(0) f
+ v^— = —^—7— + v4r^c = —^— < —^— +v— > =
dt dt dc dt dc dx dc { dx dx
Here £' is the derivative of C, and the explicit expression (19) for gO) has been employed.
(lP)ct,C(c)'
This is a (degenerate) integral equation of the first kind for /(1). It can be solved explicitly for
/M, but it is somewhat easier to solve for q^l\ which is the quantity of actual interest. If this
integral equation is integrated over 0 < v < w+J then the result can be solved explicitly for q^
as
Therefore,
The firstorder continuum approximation associated with the ChapmanEnskog expansion then
is defined by the continuity equation (18), but now the flow is given by
It is an easy application of the CauchySchwartz inequality, along with the equality F 0 (C(c)) =
cT(l  P], which defines C(c), to show that Fi(C(c))2 < cT(l  P)F2(C(c)), so that D(c) > 0.
Therefore, the initialvalue problem for this partial differential equation is wellposed compu
tationally. Some initial computational results for this equation are presented in the following
section.
The diffusive term in (21) can be interpreted as a correction to the zeroorder LWR model, which
can be written as
l+flwol0
It is interesting to note that the possibility of further improved continuum descriptions of traf
fic flow already was considered by Lighthill and Whitham (1955, p. 344), who suggested adding
"diffusion" (representing adjustments by drivers to the concentration slightly ahead) and "inertia"
(representing the nonzero time required for accelerations or decelerations) effects to the continu
ity equation. (The present approach thus confirms the diffusion term, but not the inertia term.)
By the time of his book on nonlinear waves, Whitham (1974) mentioned both this possibility and
an approach similar to current dynamic stream models (as an effort to take account of the "time
lag in the response of the driver"), but there was no associated citation of the work of Payne
(1971). Payne (1971) and Schochet (1988) have observed that the Payne model, as its relaxation
time (which does not appear simply related to that in the PrigogineHerman model) approaches
zero, reduces to a form similar to (21). On the basis that the firstorder ChapmanEnskog approx
imation to the Boltzmann equation is the NavierStokes equations of fluid dynamics, it appears
that the advectiondiffusion equation (21) is better considered as the true analog for traffic flow
of the NavierStokes equations than are current higherorder models. This supposition seems to
be rather strongly supported by the result of Schochet (1988) described earlier, which shows that
this "continuity plus diffusion" model plays the same role in justifying the entropy condition for
The ChapmanEnskog Expansion 71
the LWR model as the NavierStokes equations play in justifying it for the Euler equations of
fluid dynamics.
COMPUTATIONAL RESULTS
For the example of this section the values P = .67 and T = .0015 hours were used. These cor
respond to 77: c/Cjam = .3333 and T = .003 hours « 11 seconds, in the notation of Chapter 4 of
Prigogine and Herman (1971). The reduced desired speed distribution used is that corresponding
to a uniform distribution of desired speeds from 40 miles per hour to 80 miles per hour. The
corresponding equilibrium solution, in the stable regime, is
Here C is given explicitly by £(c) = 40(e 02c — 2)/(e 02c — 1) mph, where c is density in vehicles
per mile per lane (vpmpl). The critical density is the root of £(c) = 0, which is C^H = 50 In 2 =
34.67 vpmpl. The corresponding traffic stream model, from the equilibrium solution, is q^ (c) =
2000 4 c£(c) vehicles per hour per lane (vphpl). A plot of this traffic stream model, for the stable
regime, is shown in Figure 1. The critical flow (i.e., the flow at critical density) is q^a =
1/T(1P)= 2000 vphpl.
The problem considered here is defined by the parameters of the preceding paragraph, and the
initial conditions
This problem was solved computationally, for both the zeroorder LWR approximation and
the firstorder "advectiondiffusion" approximation (21), by means of the generalized Godunov
method (Morton, 1996, Section 7.3) This method was selected specifically because it is a shock
capturing method, when applied to pure hyperbolic,advection problems such as the LWR model,
but it also is applicable to advectiondiffusion equations, such as the firstorder ChapmanEnskog
approximation (21) obtained above. Thus this method permits comparison of the zeroorder and
firstorder approximations, without extraneous effects stemming from the use of fundamentally
different computational methods. Some sense of the capabilities and limits of this method can
be obtained from Fig. 2. In this figure, the densities for both the exact solution of the LWR
model, as described above, and for its computational approximation by the generalized Godunov
72 Transportation and Traffic Theory
method, are displayed, for t = 1.03 minutes. Notice that the computational method captures
very well the location of the shock. The shock in the computational result displays a nonzero
thickness, that is just discernible on the scale of the motion of the front (it is three to five compu
tational spatial cells thick). This "numerical dispersion" phenomenon is practically unavoidable
for computational approximations to nonlinear conservation laws (such as LWR models). The
LWR computational results also display relatively small (about 10% of the critical density) non
physical oscillations near either edge of the computed shock. Many (but not all) computational
methods for nonlinear conservation laws display such highfrequency oscillatory errors. All in
all, these results provide some evidence of credibility for the computational methodology.
The computational result for the advectiondiffusion model (21) at the same time also are dis
played in Fig. 2. The most significant differences from the LWR results are that the shock profile
is significantly wider and smoother. (The shock is approximately .3 miles wide for the diffusive
approximation, versus .1 miles for the computed LWR results). To understand the basis for this
difference, notice that the LWR model effectively approximates the vehicular flow as if all ve
hicles travelled at the average speed. (The average speed is 57.7 miles per hour unstream of the
shock, and 59.7 mph downstream.) However, according to the underlying kinetic model there is
in fact a distribution of vehicular speeds extending from v = 40 mph to v = 80 mph. Therefore,
on either side of the shock a significant fraction of the vehicles are in fact moving faster than the
average speed. Those vehicles in fact penetrate into the lowdensity downstream region further
in a given time than would be predicted on the basis of the assumption that all vehicles move at
the average speed corresponding to the local density. It is this physical dispersive effect, arising
from the underlying statistical distribution of vehicular speeds, that seems to be represented by
the additional flow (i.e., q^) giving rise to the diffusive term in the firstorder approximation
(21).
Figure 3 is a plot of the densities, as computed from the firstorder diffusive model, from t = 0
tot = 1.03 minutes, and over a significant spatial region surrounding the initial location of the
leading edge of the platoon. The region to the left of the slowly widening shock is the relatively
crowded region initially to the left of x — 0, and that to the right is the freeflow region initially to
the right of x = 0. From the perspective of Fig. 3 the widening of the shock as time increases is
quite apparent. This is due to dispersion of vehicles at the leading edge of the platoon, because of
their different desired speed. By contrast, the shock in the comparable LWR results (not shown)
maintain a constant and narrower width.
CONCLUSIONS
It has been shown that the ChapmanEnskog expansion applied to the PrigogineHerman ki
netic equation gives rise, in the regime of stable flow, to a LWR model at order zero, and to an
The ChapmanEnskog Expansion 73
5 10 15 20 25 30
Concentration (vehicles per mile)
/\
35
+
+ \
+ t
30
1 + i
+ \
E
25 + i
i +
t20 4
I
.2 + Diffusive H
2
  Cmptd. LWR V1
a
O
— Exact LWR
1
+
Dettat = .97sec.
10 . +
Deltax = 88ft. I +
1 +
T=.0015hrs +
5 ++
P = .666 > +
w
n , i
0.6 0.7 0.8 0.9 1 1.1 1.2 1.3 1.4
Space (miles)
74 Transportation and Traffic Theory
, 1
Space (miles)
Helbing (1996, p. 2379) has earlier discussed this dispersion phenomenon, by way of answering
the objection of Daganzo (1996b) to the fact that higherorder models predict vehicular move
ment faster than the (average) vehicular speed. Dispersion of vehicles due to differing speeds
clearly is a real phenomenon that is not represented in LWR models. But this phenomenon is
captured, to some degree, by either existing higherorder models or by the firstorder Chapman
Enskog approximation presented here. The choice between these two alternatives, at least in
terms of capturing the dispersion effect, may ultimately be a matter of taste. The authors prefer
the advectiondiffusion model developed here, on the basis of the following two points:
A corrective term in the conservation equation, with no additional unknowns, seems sim
pler than an additional equation, with an additional unknown. (This point invokes the
principle of Occam's razor.)
The ChapmanEnskog Expansion 75
• The firstorder model obtained here arises from the kinetic theory via standard asymptotic
expansions, rather than by ad hoc techniques.
In any event, the results presented above clearly provide an alternative to current higherorder
models, for those seeking continuum models of traffic flow that are not subject to some of the
limitations of LWR methods.
The approach presented here also has the advantage, relative to current higherorder models, that
it can be systematically extended to give a hierarchy of arbitrarily highorder extensions of the
zeroorder LWR model. For example, in the theory of rarefied gases the secondorder Chapman
Enskog expansion gives rise to the Burnett equations of fluid flow (Cercignani, 1988). These
equations are not used in fluid flow as often as their firstorder NavierStokes counterpart, but
they have found application. It remains to be seen whether the trafficflow counterpart of the
Burnett equations has any significant application. Work related to this issue will be presented
elsewhere.
It would be of some interest to extend the ChapmanEnskog expansions presented here to the
case that P and T vary with density (c) in the manner assumed in Chap. 4 of Prigogine and
Herman (1971). However, this interest may be mostly academic in nature. This is because the
effect of this dependence upon c is most pronounced for larger values of c, which are outside
the stableflow regime for which the present theory applies. Probably a more important issue,
as regards the potential application to congested flow, is to determine a counterpart continuum
theory that is applicable to unstable flow. The results of this paper, combined with those of
Nelson and Sopasakis (1998), suggest that the applicability of LWR models may be essentially
limited to stable flow.
The asymptotic expansion of Hilbert (e.g., Cercignani, 1988) has been used in the kinetic the
ory of gases, as yet another approach to obtaining a hierarchy of continuum approximations.
When applied to the PrigogineHerman kinetic model (in the stable flow regime) it gives rise to
a hierarchy of continuum models in which the model of order n arises by adding a firstorder
partial differential equation of LWR type to the previous n firstorder partial differential equa
tions. However, the unknowns in these equation are successive corrections to the density, and the
equation added to comprise the model of order n contains not only the nth order correction, say
c^n\ but also all lowerorder contributions c^n~l\ c ^ n ~ 2 ) , . . . , c^. Thus there is a onedirectional
(triangular) coupling between the equations marking the successive orders. More details of this
approach will be presented elsewhere.
LWR models, are only compounded for succesive entries in the ChapmanEnskog hierarchy.
This is to be expected, as boundary conditions for the counterpart expansion in the kinetic theory
of gases also are a difficult issue. This issue also arises for current higherorder models, and does
not seem to have been addressed in that context.
Finally, it would be of great interest to extend the types of results presented here to other, and
more realistic, kinetic models of vehicular traffic. Perhaps the most immediate such extension
should be to a PrigogineHerman equation modified so as to incorporate the spatial extent of
vehicles in the interaction term.
REFERENCES
Ansorge R. (1990). What does the entropy condition mean in traffic flow theory? Transportation
Research B 24B, 133144.
Bui, D. D., Nelson, P. and Narasimhan, S. L. (1992). Computational realizations of the entropy
condition in modelling congested traffic flow. Texas Transportation Institute report no.
12327.
Bui, D. D., Nelson, P. and Sopasakis, A. (1996). The generalized bimodal traffic stream model
and tworegime theory. Transportation and Traffic Theory, (J.B. Lesort, Ed.), Elsevier,
Oxford, 679696.
del Castillo J. M., Pintado, P. and Benitez, F. G. (1993). A formulation for the reaction time
of traffic flow models. Transportation and Traffic Theory, (C. F. Daganzo, Ed.), Elsevier,
Amsterdam, 387406.
Ceder A. (1976). A deterministic flow model for the tworegime approach. Transp. Res. Rec.
567, 1630.
Cercignani C. (1988). The Boltzmann equation and its applications. SpringerVerlag, New York.
Chapman S. and Cowling, T. G. (1952). The mathematical theory of nonuniform gases. Cam
bridge University Press, Cambridge.
Daganzo C. F. (1995a). A finite difference approximation of the kinematic wave model of traffic
flow. Transportation Research B 29B, 261276.
Daganzo C. F. (1995b). Requiem for secondorder fluid approximations of traffic flow. Trans
portation Research B 29B, 277286.
Disbro J. E. and Frame M. (1989). Traffic flow theory and chaotic behavior. Transp. Res. Rec.
1225,109115.
Drake J. S., Schofer, J. L. and May, A. D. (1967). A statistical analysis of speed density hypoth
esis. Highway Research Record 154,5387.
The ChapmanEnskog Expansion 77
Gerlough D.L. and Huber M. J. (1975). Traffic Flow Theory. Transportation Research Board
Special Report 165, Washington, D.C.
Grad, H. (1949). On the kinetic theory of rarefied gases. Communications on Pure and Applied
Mathematics 2, 331407.
Greenshields B. D. (1934). A study of traffic capacity. Procs. Highw. Res. Board 14, 448477.
Hauge E. H. (1969). Exact and ChapmanEnskog solutions of the Boltzmann equation for the
Lrentz model. Arkivfor Det Fysiske Seminar i Trondheim, No. 5.
Hall F. L. (1987). An interpretation of speedflow concentration relationships using catastrophe
theory. Transp. Res. A 21 A, 191201.
Helbing, D. (1995a). Theoretical foundation of macroscopic traffic modejs. Physica A 219,
375390.
Helbing, D. (1995b). Highfidelity macroscopic traffic equations. Physica A 219, 391407.
Helbing D. (1996a). Gaskinetic derivation of NavierStokeslike traffic equations. Physical
Review E 53, 23662381.
Helbing, D. (1996b). Derivation and empirical validation of a refined traffic flow model. Physica
A 233, 253282.
Kerner B. S., Konhauser, P. and Schilke, M. (1995). Deterministic spontaneous appearance of
traffic jams in slightly inhomogeneous traffic flow. Physical Review E 51, 62436246.
Klar A. and Wegener, R. (1997). Enskoglike kinetic models for vehicular traffic. J. Statistical
Physics 87, 91114.
Klar A. and Wegener, R. (to appeara). A hierarchy of models for multilane vehicular traffic I:
Modeling. SIAM J. Applied Math.
Klar A. and Wegener, R. (to appearb). A hierarchy of models for multilane vehicular traffic I:
Numerical and stochastic investigations. SIAM J. Applied Math.
Kiihne R. D. and Beckschulte R. (1993). Nonlinearity stochastics of unstable traffic flow. Trans
portation and Traffic Theory, (C. F. Daganzo, Ed.), Elsevier, Amsterdam.
Lebacque, J.P. (1984). Semimacroscopic simulation of urban traffic. Procs. Int. AMSE Conf
"Modelling & Simulation," 4, 273292.
Lebacque, J.P. (1995). The Godunov scheme and what it means for first order traffic flow models.
CERMICS Report No. 9548, November.
Lebacque J.P. (1996). The Godunov scheme and what it means for first order traffic flow models.
Transportation and Traffic Theory, (J.B. Lesort, Ed.), Elsevier, Oxford, 647678.
Lebacque, J.P. (1997). A finite acceleration scheme for first order macroscopic traffic flow mod
els. Transportation Systems, Procs. 8th IFAC/IMP/IFORS Symposium, (M. Papageorgiou
and A. Pouliezos, Eds.), Tech. Univ. Crete, Chania, 2, 815820.
78 Transportation and Traffic Theory
Payne H. J. (1971). Models of freeway traffic and control. Simulation Council Proc., 1, 5161.
La Jolla, CA; Simulations Council, Inc.
Payne, H. J. (1979). FREFLO: A macroscopic simulation model of freeway traffic. Transporta
tion Research Record 722, 6877.
Phillips W. F. (1977). Kinetic model for traffic flow. Report No. DOT/RSPD/DPB/5077/17,
Utah State University for the U.S. Department of Transportation.
Phillips W. F. (1979). A kinetic model for traffic flow with continuum implications. Transporta
tion Planning and Technology 5, 131138.
Phillips, W. F. (1981). A new continuum model for traffic flow. Report No. DOT/RSPA/DPA
50/81/5, U. S. Department of Transportation, Washington D. C. (Prepared by Utah State
University, Logan, Utah).
Prigogine I. and Herman R. (1971). Kinetic Theory of Vehicular Traffic. American Elsevier Pub.
Co., New York.
Prigogine, I. Herman, R. and Anderson, R. L. (1962). On individual and collective flow. Bulletin
Acad. Royale de Belgique 48, 792804.
Rathi, A. K., Lieberman, E. B. and Yedlin, M. (1987). Enhanced FREFLO Program: Simulation
of congested environments. Transportation Research Record 1112, 6171.
Richards, P. I. (1956). Shockwaves on the highway. Operations Research, 4,4251.
Ross P. (1988). Traffic dynamics. Transp. Res. 228, 421435.
Ross, P. (1989). Response to Newell. Transportation Research B 23B, 390391.
Schochet S. (1988). The instant response limit in Whitham's nonlinear trafficflow model: Uni
form wellposedness and global existence. Asymptotic Analysis, 1, 263282.
Transportation Research Board (1985), Highway Capacity Manual. Special Report 209, National
Academy of Sciences, Washington, DC.
Wegener, R. and Klar, A. (1996). A kinetic model for vehicular traffic derived from a stochastic
microscopic model. Transport Theory and Statistical Physics 25, 785798.
Whitham G. B. (1974). Linear and Nonlinear Waves. Wiley, New York.
This page intentionally left blank
The Lagged CellTransmission Model 81
Carlos F. Daganzo, Inst. ofTransp. Studies, Univ of California, Berkeley, California, USA
ABSTRACT
Celltransmission models of highway traffic are discrete versions of the simple continuum
(kinematic wave) model of traffic flow that are convenient for computer implementation. They are
in the Godunov family of finite difference approximation methods for partial differential equations.
In a celltransmission scheme one partitions a highway into small sections (cells) and keeps track
of the cell contents (number of vehicles) as time passes. The record is updated at closely spaced
instants (clock ticks) by calculating the number of vehicles that cross the boundary separating each
pair of adjoining cells during the corresponding clock interval. This average flow is the result of
a comparison between the maximum number of vehicles that can be "sent" by the cell directly
upstream of the boundary and those that can be "received" by the downstream cell.
The sending (receiving) flow is a simple function of the current traffic density in the
upstream (downstream) cell. The particular form of the sending and receiving functions depends
on the shape of the highway's flowdensity relation, the proximity of junctions and on whether the
highway has special (e.g., turning) lanes for certain (e.g., exiting) vehicles. Although the discrete
and continuum models are equivalent in the limit of vanishingly small cells and clock ticks, the
need for practically sized cells and clock intervals generates numerical errors in actual applications.
This paper shows that the accuracy of the celltransmission approach is enhanced if the
downstream density that is used to calculate the receiving flow(s) is read { clock intervals earlier
82 Transportation and Traffic Theory
than the current time, where 0 is a nonnegative integer that should be chosen by means of a simple
formula. The rationale for the introduction of this lag is explained in the paper. The lagged cell
transmission model is related (but not equivalent) to both Godunov's first order method for general
flowdensity relations and Newell's exact method for concave flow density relations. It is easier
to apply and more general than the latter, and more accurate than the former. In fact, if the flow
density relation is triangular and the lag is chosen optimally, then the lagged celltansmission model
is a conservative, second order, finite difference scheme. As a result, very accurate results can be
obtained with relatively large cells. Accuracy formulae and sample illustrations are presented for
both the triangular and the general case.
1. INTRODUCTION
This paper describes a new finite difference approximation for the kinematic wave model
of traffic flow formulated by Lighthill and Whitham (1955) and Richards (1956), called here the
LWR model, and for the generalized continuum model that applies to freeways with special lanes
(Daganzo, 1995). The proposed scheme is conservative, in that vehicles are not created or lost
during the simulation except at the highway's entrances and exits, like the procedures described
in Bui et al (1992), Lebacque (1993) and Daganzo (1993 and 1993a). Unlike its predecessors,
however, the proposed scheme is not in Godunov's family of finite difference approximations (see
Godunov, 1961, or LeVeque, 1992), and is more accurate than they are.
In the new scheme, flows across the boundary between two cells are calculated with a
sending/receiving metaphor similar to that introduced in Lebacque (1993) and Daganzo (1993a),
using only information from the two neighboring cells.1 Such a metaphor is called here the cell
transmission (CT) model. The CT recipe has been modified to model junctions (Daganzo, 1994)
and highway links with special lanes, e.g., for turning or high occupancy vehicles (Daganzo, Lin
and DelCastillo, 1995). These modifications make it possible to model complex networks. The
similarity between the new and the old CT recipes ensures that these modifications can be extended
trivially to the new method. Thus, the improvement in accuracy obtained with the new method
should apply to network models.
The main difference between the new approach and the original CT recipe is that the density
of the downstream cell, used to calculate the "receiving" flow, is now taken from an earlier time,
with a lag of C simulation clock intervals. Newell's exact solution method for concave flow
density relations (Newell, 1993) is also based on the introduction of lags. Lags are useful for
highway traffic modeling because traffic information travels several times more slowly in the
Lebacque (1993) uses the terms "local demand" and "local supply" to express the metaphor.
The Lagged CellTransmission Model 83
upstream than in the downstream direction. In our case, we shall see that if the two wave speeds
(forward and backward) are independent of density, then the lagged celltransmission (LCT) model
turns out to be second order accurate.
Because this paper is closely related to Daganzo (1993a), the reader is referred to that
reference for more extensive introductory remarks. The remainder of this paper is organized as
follows. Section 2 presents experimental evidence pertaining to wave speeds and illustrates by
means of an example the minor difficulties one encounters with the LWR theory when the wave
speed does not decline with density monotonically, i.e., when the flowdensity relation is non
concave; in particular, it is shown that the "stable" LWR solution can have waves emanating from
a shock. The new finite difference approximation is then presented in Section 3, together with an
analysis of its accuracy and stability. The paper concludes with some examples (Sec. 4) that
illustrate the results of Sec. 3.
The LWR model is intended to describe traffic on large scales of observation, where it
makes sense to define a density function k(t, x) and a flow function q(t, x) in timespace (t, x).
It is based on the assumption that q and k are locally related by a flowdensity relation q = T(k,
t, x). When the highway is homogeneous and its features do not depend on time (e.g. no incidents
or moving bottlenecks) then the relation only includes k as an argument:
q = T(k). (1)
This is the case that will receive attention for the most part of this paper.
Flow and density are also related by the conservation equation:
k, + qx = 0, (2)
where subscripts have been used to denote partial derivatives.2 Thus, it is possible to eliminate
q from (1) and (2). The result is the following simple first order quasilinear partial differential
equation in k:
kt + T k k x = 0. (3)
2
Subscripts t, x and k will be used in this paper to denote partial derivatives of the subscripted
variable with respect to time, distance and density. All other subscripts, e.g., i, j, and {, will be
indices for the subscripted variables.
84 Transportation and Traffic Theory
In this equation kt and kx are functions of t and x, and Tk is a known function of k. This is the
conventional way of expressing the LWR model for numerical approximations.
If Tk is independent of k then the solution to (3) has the form: k(t, x) = g(xTkt) ; i.e., it
is a translationally symmetric function (wave) with wave velocity T k . The particular form of g
depends on the boundary conditions. For example, for the initial value problem where k(0, x) is
given, g(x) = k(0, x) and
The wave velocity Tk also plays an important role in the quasilinear case. In this case too,
the density is constant along wavelines (characteristics) issued from the boundary, but the wave
lines can now focus and cross. The solution can be extended into these regions of the (t, x) plane
by introducing curves, called shocks in the LWR theory, where k(t, x) is discontinuous.
Conservation of vehicles across such discontinuities results in the following equation for the shock
velocity, u :
u = Aq/Ak (5)
where Aq and Ak represent the changes in flow and density across the shock.
Equations (3) and (5), however, are not always enough to specify a solution. It turns out
that shocks can often be introduced in more than one way to form a mathematically correct solution
of a properly formulated problem; i.e., a solution that is consistent with (3) and (5) and with the
initial data. If the problem has been properly formulated, however, i.e., if it makes physical sense,
then there should be one and only one solution that makes physical sense. This solution can be
identified with a standard stability argument; i.e., by making sure that the solution does not come
undone if a small perturbation is introduced in it.3
If q = T(k) is represented by a curve as in the top part of Fig. 1, then the wave velocity is
the slope of the curve. Note that if the curve is concave, then the wave velocity declines with
density, and also with the traffic speed. That is, waves focus (into a shock) when traffic decelerates
and fan out (as an expanding wave) when it accelerates. The best evidence available and common
sense indicates that the maximum wave speed in the forward direction, obtained for low densities,
is comparable with the freeflow traffic speed, e.g., on the order of 60 mi/hr (27 m/s), and that the
backward wave speeds are several times slower. This is illustrated on the bottom part of Fig. 1,
3
The physically relevant solution must be one where the shocks are "stable"; i.e., where if (at any
given time, t 0 ) one were to replace a shock by a quick but gradual transition in density between the
values prevailing on both sides of the shock, then the shock would reform itself and the solution
would quickly approach the original (for t > t0).
The Lagged CellTransmission Model 85
0.5
0.4
Jl 0.3
 o, X
0.1
0
0.02 0.04 0.06 0.08 0.1
Density, k (veh/m)
onnn
O.1 RHO
/
c?
Omnn 
LJ_
,9
0 . I I I I I I I I I I I I I I I I I I I
C 5 10 15 20 25 30 35 40 45 5(
Occupancy (%)
2.1 An example
Let us examine how a line of cars comes to a halt according to the LWR theory, when the
flowdensity relation, T(k), is as in Fig. 2a. We assume that the corners of our curve have been
smoothed very slightly, not enough to be seen in the figure, so that the wave velocity is defined for
all densities.4
It is assumed that the leading cars in the line are in state A' (q = 0.6 veh/sec, k = 12
veh/Km) and that heavier traffic (state A with q = 1 veh/sec and k = 20 veh/Km) is found 6 Km
upstream of the leading car; see Fig. 2b. The origin of the (t, x) coordinates has been chosen so that
4
Smoothed piecewise linear qk curves such as ours, are good for illustration purposes because
they lead to solutions that are simple and easy to interpret.
The Lagged CellTransmission Model 87
Figure 2. Solutions of a leadvehicle problem with kinematic wave theory: (a) T(k) relation; (b)
a mathematically valid but physically unacceptable solution; and (c) the stable solution.
Part c is the stable solution to this problem. It includes a fan of waves (carrying state "E")
that are issued tangentially from the shock at the point where it bends. That this solution is stable
is seen by noting that if the shock is replaced by a smooth but rapid transition in density anywhere,
even at the point where the shock bends, then the wave pattern so generated will match the one in
the solution.5 This means that no new states can be introduced into the solution of Fig. 2c by an
infinitesimal perturbation and that the solution is the one that would arise physically.
That Fig. 2.c is the relevant solution can also be verified from microscopic (carfollowing)
considerations, although this is more tedious. This statement is based on the principle that any
asymptotically stable carfollowing model, in the sense of Herman et al (1959), must be consistent
with the corresponding (stable) LWR result; e.g. with the macroscopic result of Fig. 2c in our case,
provided of course that the equilibrium speedspacing relation is consistent with the T(k) of Fig.
2a and that the initial conditions are also consistent with those of Fig. 2. The development of a
coasting state, such as state "E" of Fig. 2c, is demonstrated with carfollowing theory in a longer
version of this paper (Daganzo, 1997). Figure 10 of that reference shows that the carfollowing
vehicle trajectories indeed develop a coasting state when a platoon is brought to a halt.
3.1 Background
The celltransmission model can be applied to unimodal T(k) curves with maximum flow
qmax. First one defines two monotonic curves that take values in the interval [0, qmax] , as shown
in Fig. 3: a nonincreasing receiving curve R(k) and a nondecreasing sending curve S(k). The
symbol k° is used to denote one of the densities where the maximum is achieved.
A rectangular lattice with spacings e and d is then overlaid on the (t, x)plane, as shown
on Fig. 4. It is understood that traffic flows in the direction of increasing x. The xcoordinates of
the lattice points, denoted {Xj} , represent the center of the "cells" into which the highway has been
discretized, and the tcoordinates {tj, the times at which the cell densities are evaluated. The
numbering scheme is such that xj+1 > ^ and ti+1 > tj so that traffic advances in the direction of
increasing j .
5
To check stability at the point where the shock bends, one should treat the transition from A' to
A as being rapid but smooth, remembering that the diagram of Fig. 2c is on a large scale. On a
resolution scale where the transition from A' to A can be discerned the stable shock would have
to bend gradually. Furthermore, waves would peel off from it as it curves. This detailed geometry
is consistent with the macroscopic picture painted in Fig. 2c.
The Lagged CellTransmission Model 89
If we now let K(tj, Xj) denote the average density estimated for cell j at time t ; , and we
write Q(tj + e/2, Xj + d/2) for the average flow that would advance from cell j to cell j+1 (i.e.,
crossing location Xj + d/2 ) in the time interval [t ; , tj+1] , then we require:
by virtue of conservation. The subscripts i and j have been omitted in (6) for simplicity of notation.
This will be done from now on when reference to particular cells and/or a time slices is not
necessary. In these cases it should be understood that (t, x) is a point on the lattice.
The cell transmission model is completed by a formula that gives Q in terms of the sending
and receiving functions evaluated at the upstream and downstream cells,
where Tk max is the maximum of the absolute wave speed for the given T(k) relation. For
maximum accuracy, one should choose
e=d/Tk (9)
Figure 4. Lattice and stencil for the lagged celltransmission model. Dots
are lattice points; crosses are points where the average flows are evaluated.
Equation (8) ensures that data from outside the two neighboring cells cannot influence the
calculated flow, which is a requirement for convergence.
Recall now that in Godunov's approach, Eq.(7) would be the flow at the discontinuity in
the stable solution of a Riemann problem6 for which the upstream density, ku , is that of the
upstream cell, ku = K(t, x ) , and the downstream density, k d , is that of the downstream cell, kd =
K(t, x+d); see LeVeque, 1992. The reader can verify, e.g., using the ideas in Sec. 2, that Eq. (7)
indeed yields the stable flow at the location of the discontinuity for a Riemann problem with
densities K(t, x) and K(t, x+d) — no matter how these two values are chosen  if the T(k) relation
is unimodal. This establishes that the CT model (67) is in Godunov's family of finite difference
approximations for unimodal T(k)'s.
Let Sk max and  Rk max denote the maximum (absolute) wave speeds in the forward and
reverse directions, and T k  max the maximum in any direction. We show below that whenever
6
A Riemann problem is an initial value problem where the initial density is a step function with
one step; i.e., k(0, x) = kd if x > x°, and k(0, x) = ku if x < x°.
The Lagged CellTransmission Model 91
ax,
as one would expect for most traffic streams, it is advantageous to read
the sending and receiving flows from different time slices. That is, one can define a lag, Q. = 0, 1,
2,..., and use:
instead of (7). This corresponds to the stencil depicted in Fig. 4, which predicts the flow at point
"P" when 0 = 2. The special case with { = 0 reduces to the conventional celltransmission rule.
It will be shown in Sec. 3.3 that rule (10) is most accurate when the velocity of the wave
reaching "P" happens to match the slope of one of the arrows in the figure; i.e., if the prediction
at "P" is evaluated as close as possible to the source of its wave.7 This will be the basis for our
choice for {. It also turns out, for stability reasons discussed in Sees. 3.3 and 3.5, that the backward
slope of the arrow in our diagram should not be less than the maximum backward wave speed,
R k  m a x . This means that C must satisfy:
A choice of f where (11) is as close as possible to a pure equality is recommended; i.e., where:
The accuracy of (8), (10) and (12) is evaluated below. The steps parallel those in Sec. 4 of
Daganzo(1993a).
Consider a region of the timespace plane where the waves move back (congested traffic).
Then, (6) and (10) may be rewritten as:
7
Insofar as cause and effect relations propagate as waves in the LWR model, it should not be
surprising to see that accuracy is greatest when we take our data from the lattice point closest to
the wave.
92 Transportation and Traffic Theory
In the exact theory, the solution at time t+e is related to the solution at time t by:
where Tk is evaluated for the density prevailing at (t, xTke). Thus, in a region where K > k°, it is
convenient to rewrite (13) in the following manner:
K(t+e, x) = K(t, xTke) + [K(t, x)  K(t, xTke)]  (e/d)[T(K(t0e, x+d))  T(K(We, x))]. (14)
This is useful because a second order power series expansion of the second and third terms in this
expression about point (t, xTke) yields an estimate of the error in (13) in terms of known
quantities.
The expansion of the second term is:
where a double subscript denotes a second derivative. Likewise, the expansion of the third term
can be reduced to:
and the sum of these two expressions can be further simplified. The combined result is:
The first three terms of this expression coincide with (9) in Daganzo (1993a). The last term is the
contribution to the error caused by the lag. This term becomes more meaningful if the time
derivatives Kj and Kxt are eliminated from the solution. This can be done if one notes from (3) that
kt =  T k k x , and that the xderivative of this expression (keeping time constant) is: ktx =  Tkkkx2 
Tkkxx . If we use these relationships in the last term of (15) and use p as an abbreviation for
 Tke/d , then we obtain the following expression for the error committed in time e (when the
system is congested):
The Lagged CellTransmission Model 93
The LCT method can be applied to highways and networks that have been discretized with
cells of different lengths. In this more general case the LCT lag should be celldependent so as to
ensure that the density is always read from the earliest possible time without violating the stability
condition. That is, the lag for cell j , ^ , should satisfy as tightly as possibly the inequality:
e>dj/[Rkmax(2ej+l)], (17)
e>dJ/[Sk,max(2fJ+l)], (18)
instead of (8). The forward lag can forestall the deterioration of accuracy where large cells are
being used; especially if fj is chosen so that (18) is as close to an equality as possible. The logic
for this choice is the same as that in Sec. 3.3, and the outcome is also similar; e.g., in that the
resulting method is still second order accurate in the case of a triangular T(k) relation.
The LCT rule with a variable discretization is then:
K(t+e, Xj) = K(t, Xj)  (e/dj)[ Q(t+e/2, Xj+d/2)  Q(t+e/2, Xjd/2) ] (19)
with
Godunov's procedure identifies the stable solution because, as we have seen, its flows are
always derived from the stable solution of a Riemann problem. Likewise, it will be shown here
that the flows of the LCT method (20) are based on the stable solution of a modified Riemann
problem, and therefore that the LCT solution should too approximate the stable solution.
Let us consider a modified Riemann problem (MRP) in which constant upstream and
downstream densities have been defined on a Vshaped boundary, such as the dark line in Fig. 5.
For an MRP problem to be wellposed, waves from both legs of the "V" should point into the
solution space. This will happen if the (t, x)slopes, su and sd , of these legs satisfy su > Sk max and
sd < R k max ; i.e., if the Vshaped boundary is inside the shaded wedges shown in the figure.
Consideration reveals that the stable solution of the MRP for t > 0 is independent of the
slopes of the "V" and that, as a result, said solution is also the stable solution of the conventional
Riemann problem with the same initial densities. In other words, the stable flows arising at the
discontinuity are the same for the conventional and modified Riemann problems.
Since (20) has the same form as (7), i.e., it is the minimum of a sending and a receiving
value, it expresses the stable flow of a Riemann problem with ku = K(tfje, Xj) and kd =
K(tf j+1 e, xj+1) . As we have just shown, this is also the solution of a wellposed MRP with the
same initial densities. In particular, it is the stable flow for the MRP in which the "V" passes
through the lattice points at which ku and kd have been evaluated (points "A" and "B" in Fig. 5).
This MRP will be wellposed if the "V" passing through the lattice points is inside the shaded
The Lagged CellTransmission Model 95
= 4
A"
«* max ^"^^~^
s ' f
S*, max a, II
\"B*
4. EXAMPLES
Let us start by examining the accuracy of the LCT model with both smooth and
discontinuous initial conditions.
8
The accuracy formulae of Sec. 3.3 confirm that the LCT method is most accurate when f, and
<> j+ i are chosen to be as large as possible, i.e., when the "V" forms as acute an angle as possible.
96 Transportation and Traffic Theory
Tables 1 and 2 present 17 iterations of the conventional and lagged cell transmission models
for a triangular flowdensity relation of the form:
where k is in veh/mile and q in veh/min. It is assumed that the initial density profile is quadratic
and in the congested range k e [30, 180] ; more specifically, that k(0,x) = 50 + Vix2 and x ~ 10
± 5. Because the (congested) wave velocity is constant (i.e., Rk = 0.2) the exact solution of the
problem in this range of x for small t is: k(t, x) = 50 4VXxf t/5)2. That is, in the exact solution the
density at mile x1 must equal the density that prevailed at mile x five minutes earlier.
Table 1 presents the results of the ordinary CT model when one uses e = 1 and d = 1. In
the exact solution the numbers in boldface would be equal, so that the observed discrepancy is the
CT error. The discrepancies observed in the table are consistent with what would be expected from
(16) with { = 0, and from the more detailed analysis in Daganzo (1993a).
If one uses the LCT model with e = 1, d = 1 and f = 2 , i.e., the value recommended by (12),
then Table 2 is obtained. In this case the LCT model reproduces the exact results, as one would
expect from (16) and the analysis leading to it.9
Of course, the performance of the LCT procedure deteriorates in less favorable cases, e.g.,
with nonquadratic density profiles, nonlinear T(k) functions, and nonideal lags.
This is illustrated by Fig. 6, which contains the initial and final density profiles (at t = 25
min) for a Riemann initial value problem where (21) holds and where the initial density changes
suddenly from k = 50 to k = 100 at x = 20 mile. The thin lines are the CT and LCT results
obtained with e = 1 min and d = 1 mile. Note the lesser spread of the LCT result, and that the LCT
model does not transition monotonically from one density value to the other.10 The relative
accuracy of the two models can be evaluated better with the cumulative curves of vehicle count;
see Fig. 7. Whereas in the CT model the maximum error in vehicle position is on the order of 2/3
miles (which is not surprising since d = 1 mile), with the LCT the largest error is less than 1/5 mile.
In order to illustrate the stability of the method for nonconcave T(k) relations, Figs. 8 and
9 depict the LCT solution of the lead vehicle problem in Fig. 2c using cells with d = 100 m and
time steps e = 2 sees. This is efficient because with this arrangement (8) is satisfied as an equality
— note from Fig. 2a that in our case Tk max = vf =50 m/s. Since Rk max = 350/13 m/s , the
9
Note that, for this to be the case, 0+1 time slices of internally consistent initial data had to be
specified.
10
This undesirable feature of the LCT model is typical of second order approximations (see, e.g.,
LeVeque, 1992). As a result, the LCT model may produce densities slightly greater than the
theoretical maximum from time to time. Therefore, the receiving function R(k) should be defined
for 0 < k < °° , in an implementation.
The Lagged CellTransmission Model 97
Table 1. Estimated densities at different positions with the celltransmission model. Row 3 is the
position (x) in Km. Rows 5 and 6 are the maximum flow and the jam density at the given position.
Rows 9 to 11 are the initial data, k = 50 + V^x +1/5)2, where t = 0 for row 9, t = 1 for row 10 and
t = 2 for row 11.
3 6 7 8 9 10 11
4
5 30 30 30 30 30 30
6 180 180 180 180 180 180
7
8
9 68 74.5 82 90.5 100 110.5
10 69.22 75.92 83.62 92.32 102.02 112.72
11 70.48 77.38 85.28 94.18 104.08 114.98
12 71.86 78.96 87.06 96.16 106.26 117.36
13 73.28 80.58 88.88 98.18 108.48 119.78
14 74.74 82.24 90.74 100.24 110.74 122.24
15 76.24 83.94 92.64 102.34 113.04 124.74
16 77.78 85.68 94.58 104.48 115.38 127.279
17 79.36 87.46 96.56 106.66 117.76 129.852
18 80.98 89.28 98.58 108.88 120.178 132.453
19 82.64 91.14 100.64 111.14 122.633 135.071
20 84.34 93.04 102.74 113.438 125.121 137.694
21 86.08 94.98 104.88 115.775 127.635 140.308
22 87.86 96.9599 107.059 118.147 130.17 142.897
23 89.68 98.9796 109.276 120.551 132.715 145.445
24 91.5399 101.039 111.531 122.984 135.261 147.936
25 93.4397 103.137 113.822 125.44 137.796 150.355
26 95.3793 105.274 116.145 127.911 140.308 152.691
27 97.3583 107.449 118.499 130.39 142.785 154.931
28 99.3763 109.659 120.877 132.869 145.214 157.066
29 101.433 111.902 123.275 135.338 147.584 159.09
98 Transportation and Traffic Theory
Table 2. Estimated densities at different positions with the lagged celltransmission model. Row
3 is the position (x) in Km. Rows 5 and 6 are the maximum flow and the jam density at the given
position. Rows 9 to 11 are the initial data, k = 50 + 1A(\ + t/5)2, where t = 0 for row 9, t = 1 for
row 10 and t = 2 for row 11.
3 6 7 8 9 10 11
4
5 30 30 30 30 30 30
6 180 180 180 180 180 180
7
8
9 68 74.5 82 90.5 100 110.5
10 69.22 75.92 83.62 92.32 102.02 112.72
11 70.48 77.38 85.28 94.18 104.08 114.98
12 71.78 78.88 86.98 96.08 106.18 117.28
13 73.12 80.42 88.72 98.02 108.32 119.62
14 74.5 82 90.5 100 110.5 122
15 75.92 83.62 92.32 102.02 112.72 124.42
16 77.38 85.28 94.18 1 104.08 114.98 126.88
17 78.88 86.98 96.08 106.18 117.28 129.38
18 80.42 88.72 98.02 108.32 119.62 131.92
19 82 90.5 100 110.5 122 134.5
20 83.62 92.32 102.02 112.72 124.42 137.12
21 85.28 94.18 104.08 114.98 126.88 139.78
22 86.98 96.08 106.18 117.28 129.38 142.48
23 88.72 98.02 108.32 119.62 131.92 145.22
24 90.5 100 110.5 122 134.5 148
25 92.32 102.02 112.72 124.42 137.12 150.82
26 94.18 104.08 114.98 126.88 139.78 153.677
27 96.08 106.18 117.28 129.38 142.48 156.565
28 98.02 108.32 119.62 131.92 145.22 159.47
29 100 110.5 122 134.5 147.999 162.369
The Lagged CellTransmission Model 99
12 16 20
Distance, x
optimum lag according to (12) should be f = 3/7 . Because this is not an integer, R(k) was
evaluated for a k that was an interpolation of the densities obtained with { = 0 and C = 1. The
numerical results should then still be second order accurate for densities less than or equal to that
of state "E" in Fig. 2a, and less accurate for more congested states.
The density profiles of Fig. 8 confirm this. Note how the curves have sharper steps below
the line K = 0.5 than near the top. (In the exact solution these curves would be perfect step
functions.) Note as well the good agreement between this figure and 2c; in particular, the
development of intermediate state "E", with K = 0.4, after t = 120 sees.
Figure 9 displays the Ncurves (of vehicle number) at five locations in ¥2 Km increments
upstream of the stoppage. They also agree qualitatively with the exact solution of the problem,
100 Transportation and Traffic Theory
900
750
o
700
650
which in this case is piecewise linear. Clearly, the curved arcs in the figure, which correspond to
episodes of very congested traffic, have some numerical error. However, the straight portions of
the curves match the exact solution precisely, and therefore it is easy to ascertain from the figure
the magnitude of the numerical errors in the curved sections. This error does not exceed 5 vehicles
in any of the curves.
The Lagged CellTransmission Model 101
5. CONCLUSION
The LCT procedure is also well suited for modeling intersections and inhomogeneous
highways, since the only change needed in the existing procedures is reading the traffic density for
the downstream conditions with a time lag. This modification is so minor that it can also be
applied to highways with special lanes and their junctions; e.g. by modifying the procedure
described in Daganzo et al. (1995).
It should also be said that lags impose additional memory storage requirements on an LCT
simulation, since cell densities must be stored for the past 1+fj time slices. For multidestination
networks, however, most of the storage is consumed by the cell content proportions (by destination
and entry time), which are only used as arguments as part of the "sending" functions. Thus, in an
LCT implementation the bulk of the information would only have to be kept for 1+fj time slices.
0.1
•t = 0s
1 = 40 s
t = 80 s
0.075
t=120s
t=160s
'w 0.05
c
<u
Q
0.025
0 10 20 30 40 50 60 70
Distance, x(100 m)
..
,•*
,'"
/
J
Vehicle number
/
i3
,—
///
^
B
'{/
/
,.
'^ ,—
/
V//
'/
0 40 a3 120 160 200 240 280 320 360
x=OKm
Time (sees) x = .5Km
x=1 Km
x = 1.5Km
 • • x = 2.0Km
REFERENCES
Bui, D.D., P. Nelson and S. L. Narasimhan (1992) Computational realizations of the entropy
condition in modeling congested traffic flow. FHWA Report TX92/12327.
Cassidy, M. (1998), Bivariate relations in highway traffic. Trans. Res. 32B(1), 4959.
The Lagged Cell Transmission Model 103
Daganzo, C.F. (1993) The celltransmission model: A dynamic representation of highway traffic
consistent with the hydrodynamic theory. Institute of Transportation Studies, Research
Report, UCBITSRR937, Univ. of California, Berkeley, CA. Short version in Trans.
Res., 28B(4), 269287.
Daganzo, C.F. (1993a) A finite difference approximation for the kinematic wave model. Institute
of Transportation Studies, Research Report, UCBITSRR9311, Univ. of California,
Berkeley, CA. Short version in Trans. Res. 29B(4) 261276.
Daganzo, C.F. (1994) The celltransmission model. Part II: Network traffic. PATH working paper
UCBITSPWP9412, Univ. of California, Berkeley, CA. Short version in Trans. Res.,
29B(2), 7994.
Daganzo, C.F. (1995) A continuum theory of traffic dynamics for freeways with special lanes.
PATH Technical Note UCBITSPTN9508, Univ. of California, Berkeley, CA. Short
version in Trans Res. 31B (2), 83102.
Daganzo, C.F. (1997) An enhanced celltransmission rule for traffic simulation. Institute of
Transportation Studies Research UCBITSRR9706, Univ. of California, Berkeley, CA.
Daganzo, C.F., Lin, W.H. and del Castillo, J.M. (1995) A simple physical principle for the
simulation of freeways with special lanes and priority vehicles. PATH Technical Note
UCBITSPTN9509, Univ. of California, Berkeley, CA. Short version in Trans. Res.
31B (2), 105125.
Godunov, S.K. (1961) Bounds on the discrepancy of approximate solutions constructed for the
equations of gas dynamics. J. Com. Math, and Math. Phys. 1, 623637.
Herman, R., E.W. Montroll, R.B. Potts, and R.W. Rothery (1959), "Traffic dynamics: Analysis of
stability in car following", Opns. Res. 1, 86106.
Lax, P.D. (1973) Hyperbolic systems of conservation laws and the mathematical theory of
Shockwaves. Regional Conf. Series in Applied Mathematics. SLAM, Philadelphia, PA.
Lebacque, J.P. (1993) Les modeles macroscopiques de trafic. Annales de Fonts, (3rd trim) 67, 28
45.
104 Transportation and Traffic Theory
LeVeque, R.J. (1992) Numerical methods for conservation laws, (2nd edition), BirkhauserVerlag,
Boston, MA.
Lighthill, M.J. and G.B. Whitham (1955) On kinematic waves. I flow movement in long rives.
n A theory of traffic flow on long crowded roads. Proc. Roy. Soc. A, 229, 281345.
Luke, J.C. (1972) Mathematical models for landform evolution. J.Geophys. Res., 77, 24602464.
Newell, G.F. (1961) Nonlinear effects in the dynamics of carfollowing. Opns. Res. 9(2), pp.
209229.
Newell, G.F. (1993) A simplified theory of kinematic waves in highway traffic, I general theory,
n queuing at freeway bottlenecks, HI multidestination flows. Trans. Res., 27B, 281313.
ABSTRACT
Traffic was studied upstream and downstream of a bottleneck on a freeway in Toronto, Canada
using transformed curves of cumulative vehicle count and cumulative occupancy. The
bottleneck was located more than a kilometer downstream of a busy onramp. After
diagnosing its location and the times that it remained active each day, the study focused on the
traffic patterns that arose in each travel lane. It was observed that prior to the bottleneck's
activation, the vehicles' lanechanging trends created extraordinarily high flows in the median
(i.e., leftmost) lane and that these high flows were sustained for extended durations. When a
queue eventually formed at the bottleneck, its discharge rates were considerably lower than
those flows measured prior to queueing. Within each lane, the queue discharge rates remained
nearly constant over the rush and the average rates varied only slightly across days. Finally,
vehicles arriving to the bottleneck from the nearby upstream onramp entered the freeway at
high rates, even after the bottleneck's queue propagated beyond this ramp.
1. INTRODUCTION
In an earlier study, the authors examined traffic at two bottlenecks on freeways in and near
Toronto, Canada (Cassidy and Bertini, 1999). From this study, certain reproducible patterns
were observed. For example, each bottleneck formed more than a kilometer downstream of an
onramp and their formations occurred at these same locations each day. While the bottlenecks
108 Transportation and Traffic Theory
were active,1 the vehicles discharged through them at nearly constant rates, although some time
dependencies were observed for short periods following the onset of queueing. Furthermore, a
bottleneck's average queue discharge rate did not vary much from one day to the next and
these average rates were typically 8 to 10 percent lower than the flows measured prior to
queueing upstream.
These earlier findings came by visually comparing sets of transformed cumulative curves.
Each curve was constructed from either the counts or the occupancies collected at one of
several neighboring loop detector stations. Of note, these curves described measurements that
were taken across all travel lanes, meaning that the above findings came by grouping together
the traffic streams in multiple lanes and studying them in the aggregate.
In this paper, we add to the previous findings on bottleneck flow by reporting on some
observations taken from individual lanes. At each of three neighboring detector stations
located upstream and downstream of a bottleneck, curves of cumulative vehicle count and
cumulative occupancy were separately constructed for each travel lane. Visual comparisons of
these curves revealed certain details of traffic evolution, some of which were unexpected.
It was observed, for example, that large numbers of vehicles gradually moved into the median
lane as they approached and passed through the bottleneck. This lanechanging pattern even
continued at locations more than 2 kilometers downstream of the neighboring onramp. As a
consequence, flows measured in the median lane were remarkably high; e.g., at a location well
downstream of the onramp, the median lane flows sometimes exceeded 2,600 vehicles per
hour. These high rates persisted each day for durations of up to 40 minutes before queues
formed upstream and lower discharge rates ensued.
The bottleneck's queue formed at nearly, but not exactly, the same times in each lane; i.e., this
formation occurred in the shoulder lane several minutes after it had occurred in the adjacent
lanes. Following the queue formations, the flow reductions observed downstream were most
pronounced in the median and center lanes and less so in the shoulder lane. The discharge
rates remained nearly constant so long as the bottleneck was active and free of any incidents
nearby. Although these average rates varied across lanes, each lane's average was reproduced
from day to day.
It was further observed that vehicles entered the freeway from the upstream onramp at very
high rates, even after the bottleneck's queue propagated beyond the ramp and obstructed this
flow. Thus, the onramp vehicles did not share the available capacity with vehicles in the
1
The term "active bottleneck" denotes that the discharge rates measured downstream of a queue were not
affected by traffic conditions from further downstream (Daganzo, 1997).
Observations at a Freeway Bottleneck 109
adjacent freeway lane in a strictly alternating or "onetoone" basis. Rather, motorists from the
onramp forced themselves into the queue in such a way that the gaps between the freeway
vehicles were filled by multiple onramp vehicles.
The following section provides descriptions of the bottleneck and of the loop detector data used
in this study. Section 3 describes the use of cumulative curves to identify the bottleneck's
location and the times that it remained active; uncovering these details was a requisite step for
studying the evolution of the bottleneck flows. Section 4 presents the bottleneck's traffic
patterns that were observed in each lane during a single rush. Some findings from repeating
these analyses with data from other days, and comments regarding future research directions,
are provided in the fifth and final section.
2. THE DATA
The segment of the Gardiner Expressway shown in Figure 1 was the site used in our study. As
an aside, this site has been used in several earlier studies of freeway capacity (Persaud, 1986;
Persaud and Hurdle, 1991; Persaud, et al., 1998), including the previous one by the authors
(Cassidy and Bertini, 1999). It is located in Toronto, Canada and meters are not deployed on
its onramps (although the Jameson Avenue onramp is closed during a portion of each
afternoon rush).
Loop Detector
Detector Station
The loop detector stations for measuring traffic data are labeled in Figure 1 as per the
numbering strategy adopted by the City of Toronto. These detectors record counts,
occupancies and (time) mean speeds in each lane over 20second intervals. In total,
110 Transportation and Traffic Theory
measurements were made during three weekday afternoons when the local weather bureau
reported clear skies and no measurable precipitation.
In the next section, data from one of the observation days are used to demonstrate that a
bottleneck was activated between detector stations 60 and 70. This bottleneck remained active
for more than two hours before a queue spilled over from further downstream and restricted its
discharge. We will also show that, some minutes prior to its deactivation, the bottleneck's flow
was impeded by an incident that occurred near detector 50.
In Figure 2, however, each curve, along with its corresponding time axis, was shifted to the
right by the average freeflow trip time between the respective detector and downstream
detector 80. Consequently, the vertical separations in the curves are the excess vehicle
accumulations between detectors due to vehicular delays. These shifts are advantageous
because two superimposed curves indicate that traffic in the intervening segment was freely
flowing; every feature of an upstream Wcurve is passed to its downstream neighbor a freeflow
trip time later. In addition, Figure 2 shows only the differences between each curve of
cumulative count and the line Nqo1', where go is the rate used to rescale the curves and t' is
the elapsed time from the start of each curve. This is important because reducing the curves'
cumulative count by a background flow qo magnifies details, such as the time dependencies in
the flows, without changing the excess accumulations (Cassidy and Windover, 1995).
The onramp counts at Spadina Avenue were also used to construct the curve for detector 40 so that all of the
curves in Figure 2 describe the same collection of vehicles. Conversely, a curve for detector 30 was not
included in Figure 2, since the vehicles measured at this location were not identical to those measured at the
detector stations further downstream.
Observations at a Freeway Bottleneck 111
The (nearly) superimposed curve portions in Figure 2 indicate that traffic was initially in free
flow and remained in free flow between detectors 70 and 80. The reader may use a
straightedge to confirm that curves 40, 50 and 60 exhibit increased slopes sometime shortly
after 15:40, as delimited by the large arrow in Figure 2, and this caused these three curves to
en en en
CD » eo en
01 en en en
oo co co co on en on en
c n o r i c n c n c n c n c n c n c n c n
c o c o c o c o c o c o c o c o
en on en en
e o e o e o e o e o e o e o e o
e n c n e n e n e n e n e n e n
en en en en
co co eo eo eo eo eo eo eo eo
diverge from their (two) downstream counterparts. These curve features indicate that a
bottleneck was activated between detectors 60 and 70 when increased flows arrived from
upstream. The subsequent separation between curves 60 and 50 (at about 15:51:23) reveal
when the queue arrived at detector 60. Likewise, the queue's arrival at detector 50 is made
evident by the divergence in curves 50 and 40 (at about 15:53:03).
In short, the transformed TVcurves in Figure 2 conclusively diagnose the bottleneck's location
by showing that excess vehicle accumulations occurred upstream of detector 70 while free
flow conditions prevailed immediately downstream. Especially notable are the pronounced
flow reductions (i.e., the reduced slopes of the TVcurves) that followed the queue's formation;
further details on this are provided in the next section.
Ol Ol Ol O1 Ol Ol Ol Ol Ol Ol Ol Ol Ol Ol Ol & 0) 0) 0) 05 o>
t* o o o
Ol
o o o
N A A 4x Ol Ol Ol Ol Ol Ol Ol Ol Ol 0
Ol 0) N 05 co o CO w *. 01 01 ^1 00 CD M w Ol
^
co w w w
Figure 3 reveals the approximate time that the backwardmoving queue arrived at the (freeway)
detectors at station 40. Presented in this figure is a rescaled TVcurve for the freeway lanes at
this detector station along with a rescaled curve of cumulative occupancy versus time, or a T
curve, where cumulative occupancy is the total vehicular trip time over the detectors by time t
(Lin and Daganzo, 1997). Again for the purpose of magnifying details, the Tcurve shown here
is the difference between the cumulative occupancy actually measured in the three freeway
lanes at detector 40 and the line T=b0 t', where b0 is the background occupancy rate used to
Observations at a Freeway Bottleneck 113
rescale the curve and /' is the elapsed time from the curve's start. The two curves show that a
sharp reduction in flow was followed closely by an increase in the occupancy rate at about
15:54:23. These features reveal the arrival of the queue at station 40. That this arrival
occurred some time around 15:54 will be an important part of later discussion regarding the
observed time dependencies in the onramp flows from Spadina Avenue.
Finally, Figures 4a and 4b reveal the period that the bottleneck remained active. Figure 4a
shows transformed TVcurves for detectors 60 and 80; these were constructed in the manner
previously described, but for an extended period of over 4 hours. The slope of curve 80 drops
noticeably some time around 18:00 and a similar reduction is displayed by curve 60 soon
thereafter. Notably, the queue between detectors 60 and 80 persisted, even after these flow
reductions; this is evident from the continued displacement between the two curves. These
curve features indicate that a queue from further downstream arrived at these detector stations
and thereby deactivated our bottleneck. In Figure 4b, the divergence in the rescaled curves of
TV and rreveal that the queue from downstream arrived at station 80 at about 18:12:03.
Before concluding this section, it is worth reemphasizing that the curves in Figures 2 and 4a
were instrumental in identifying the bottleneck's location and the period that it remained
active. These curves derived their value, in part, by displaying the excess vehicle
accumulations that arose between detectors and this required that the curves be constructed
from the counts taken over all travel lanes.3 Having now identified these bottleneck details (i.e.,
its location and the time it was active), traffic patterns could be studied in the individual lanes
at locations upstream and downstream of the bottleneck. Some notable findings from this
study are presented next.
Before presenting some of the findings obtained by jointly examining Figures 5, 6 and 7, a
brief explanation of their annotations is warranted. First, piecewise linear approximations
If a set of TVcurves do not describe node conservation, their vertical displacements would not be the excess
accumulations (Newell, 1982; Newell, 1993; Cassidy and Windover, 1995).
114 Transportation and Traffic Theory
Sta. 60
^ Sta. 80
10
ii
ioo 4 /
0 [/ Time, f @ 60
. 1 i : i i i i
O l U l C O O J O l C D N v l X N C O 00 CO CO to to
N '^. <jl J. NJ
fflSSsufflSNuS (Jl » *J 03 CO
Time, f @ 80
Figure 4a: Transformed Ncurves, Detectors 60 and 80.
O
c
T(80,t)b0t' O
c
o3
Q.
I
I<D
o co
o CM
00 CO
10
T~
n
c?
A N(80,t)  qf II
c?
CT §51
PO Ow o
00
Time,
0 0 0 0
u u u u
Figure 4b: Rescaled JV and Tcurves, Detector 80.
Observations at a Freeway Bottleneck 115
were superimposed on the N to highlight periods of nearly constant flow; the N usually
deviated from its corresponding linear approximation by no more than about 10 vehicles. The
start and end times for each period of nearconstant flow were selected "by eye" and rescaled
7curves (not shown here) were also used to aid in these delimitations, since (sizable) changes
in flow are accompanied by changes in occupancy. The times marking these flow changes are
labeled on each curve.4 Further, the times marking the onset and the termination of queue
discharge flows are noted in boldface type for the curves at downstream stations 70 and 80.
Also shown are the rates corresponding to each period of nearconstant flow; the numbers
shown without parentheses are in units of vehicles per hour (vph) and those in parentheses are
the corresponding average counts per minute. Finally, dotted lines are used to highlight the
average queue discharge rates measured over the rush at downstream detectors 70 and 80.
Labels specifying these average discharge flows are shown on the figures in boldface type.
We now turn our attention to the traffic patterns revealed by Figures 5 through 7. From even
cursory examination of these figures, one observes a trend in vehicle lane changing; namely,
that large numbers of vehicles changed lanes to the left while they traveled between detectors
60 and 80. The curves in Figure 5, for example, were each started with the same value of N,
but curve 60, N(60, t), lies well above curve 70 for all time t. In similar fashion, curve 70 rises
above curve 80, indicating that vehicles continued to exit the shoulder lane at locations well
downstream of the Spadina Avenue onramp.
Figure 6, on the other hand, shows no such obvious trend. Rather, the (net) flows in the center
lane remained nearly unchanged as traffic moved through the bottleneck, indicating that,
between detectors, the number that moved into the center lane nearly equaled the number that
moved out. In fact, the Figure 6 curves have been vertically displaced (by arbitrary distances)
because not separating these curves would have made it difficult to view their details.
Figure 7 shows that large numbers entered the median lane between detectors 60 and 80 and
that this trend gave rise to some extraordinarily high flows. For example, a flow in excess of
2,600 vph was measured in the median lane at detector 80 prior to the bottleneck's activation.
Remarkably, this very high rate was observed for over 40 minutes before the queue formed
upstream and a lower discharge rate ensued. (The queue's formation is signaled by the onset of
its discharge flow at detector 70).
The onset of queueing occurred at nearly, but not precisely, the same times in each lane; i.e.,
the queue appears to have formed in the shoulder lane several minutes after it formed in the
adjacent lanes. The curves at detector 60 (in all three lanes) exhibit sustained surges some
4
A few of these times are not labeled in Figure 5 so that the figure would not become cluttered.
116 Transportation and Traffic Theory
O
CO
Observations at a Freeway Bottleneck 117
ZZOSSl L£*OS:91 PSAUJV ananQ
= °b '.fb()'x)N
100 MEDIAN LANE
NOTE:
Times shown in bold mark the onsets and terminations of queue discharge.
Rates shown in bold are average queue discharge flows observed during the rush
14:55:03 15:07:03 15:19:03 15:31:03 15:43:03 15:55:03 16:07:03 16:19:03 16:31:03 16:43:03 16:55:03 17:07:03 17:19:03 17:31:03 17:43:03 17:55:03 18:07:03 18:19:03
Time, t
Figure 7: Rescaled TVcurves, Median Lane.
Observations at a Freeway Bottleneck 119
minutes prior to the queue's formation. Conversely, the onset of this queue was accompanied
by rather dramatic flow reductions. Figures 5 through 7 show that the periods immediately
following the bottleneck's activation were marked by some of the lowest discharge rates
observed during the rush, but that these relatively low flows were shortlived relative to the
rush. In the center and median lanes (Figures 6 and 7), these socalled flow collapses (Cassidy
and Windover, 1995) prevailed for just under 20 minutes before being replaced by higher
discharge flows. In the shoulder lane (Figure 5), the collapse persisted for less than 10
minutes.
Also of note, Figures 5 through 7 show that sizable reductions in the discharge flows were
measured in all lanes some minutes before the bottleneck was deactivated by the arrival of the
queue from downstream (recall that this queue arrived at detector 80 at about 18:12). Visual
inspection of N and Tcurves measured in individual lanes revealed that these flow drops were
caused by an incident (perhaps a vehicle stall or a small collision) that occurred in the shoulder
lane near detector station 50. Notably, detector 50 measured nearzero counts and occupancies
in the shoulder lane from 18:02:23 to 18:08:23. Figure 8a shows that during this period, the
shoulder lane traffic at upstream station 40 exhibited a sharp reduction in the flow coupled with
a rise in the occupancy, features which mark the passage of a queue. Conversely, Figure 8b
shows that, within this period, shoulder lane traffic at downstream station 60 exhibited sudden
reductions in both the flow and the occupancy. These features would be expected to occur
downstream of a sudden restriction (i.e., as traffic passed the obstruction and moved into the
shoulder lane). A reduction in the discharge flow also occurred in the shoulder lane at about
16:59 (see Figure 5), but this reduction was shortlived and of no real consequence.
Despite the flow variations that occurred during the bottleneck's active period, Figures 5, 6 and
7 reveal that the queue discharge rates never deviated much from the linear trends shown with
the dotted lines (the reduced discharge flows that accompanied the incident near detector 50
were excluded from consideration here). Thus, for each lane, the discharge rates can be
described as being nearly constant over the rush. By comparing the average rates that
correspond to each dotted line with the flows that prevailed prior to the bottleneck's activation,
it is clear that queueing was accompanied by longrun flow reductions, especially in the
median and center lanes. Also by examining these three figures collectively, it is evident that
the average queue discharge rates varied across lanes.
Finally, Figure 9 shows a rescaled TVcurve constructed solely from the counts on the Spadina
Avenue onramp (at detector station 40). The sustained surge in the ramp flow evident at
15:42:43 corresponds to the measured increase in upstream flow previously revealed in Figure
120 Transportation and Traffic Theory
3
o
c " O
^
f
_,^f^=~^~^ —\" "  ^^l «o
O} O
00 > ^~^^^ 00
Zo ^^^^ CO
II 0
I—
II
c? O
o x. C?
fc ., Qo \ 1 ^ ^
•** LU '• ^—i
cr K
Z ^\
10~ "^
o« \x A/(40,f)  q0t'
it) —i ^
o"
^ LJ_
K X^^ 1 ^^ o"
2f o "^\ x
fr
Time, f .^
c3 0 0 0 0 0 0 C O C 0 0 3 Co o o o o o o o o o o o o c o o o o o o o o o o o o o o o CO CD 00 CO
iD O O O O O O C3 0 0 0 0 0 0 O O O O O O O C O O O O
c. •^i g oo oo
,
o  /^^x/^\ T(60,t) b0t' ^
"'/ ss^
o
•^
~~^^ \^^^ ^•
CD >_ ^^^ •—— ^
Q.
CL
CO
o N(60,t)q0t
CD "D
O ^
\— ID
oo c
~" " ** \
5
CD Q
UJ
0
o 8
CD
CO
0
§cv \5
1
S Q.
o
2 o:
v 1015
^ "\\ •^
"o 1
II
Cr •°_
c£ 0 Q°
'\V 0
!
^ ''. \ °"
co_
s &
;
Time, t
! ' i ' 1 I ! i
' '
cO C O O O C D O O Q 3 G O Q O O O O O O O O O Q O C X ) Q 0 0 0 0 0 0 3 0 0 C O C O Q 0 0 0 0 3 C O C
o o o o o o o o o o o o o o o o o o o o o o o o o o
u c o c o u u w c o u
Figure 8b: Rescaled TV and Tcurves, Detector 60, Shoulder Lane.
Observations at a Freeway Bottleneck 121
lioo ./!*
CD
a JMf/''^ ^f
Vj
CO ,/V/<?" V.TN cjp
CD
O COi
c CN;
§ co'i
o
N(Spadina,t)  q0t'
Time, f
Figure 9: Rescaled vVcurve, Spadina Avenue onramp.
2.5 Likewise, the sharp reduction in this ramp flow at 15:50:23 corresponds closely to the
arrival of the queue from our active bottleneck; recall that this queue was shown (in Figure 3)
to have arrived to the freeway detectors at station 40 some time shortly after 15:50. Although
this queue apparently suppressed the onramp flow, vehicles continued to enter the freeway via
this ramp at a high rate; i.e., an average ramp flow of 1,650 vph persisted for nearly an hour.
This flow dropped at about 16:46:43, perhaps due to a reduction in the onramp demand
(although additional ramp detectors do not exist to confirm this). In any event, these high ramp
flows mean that, just downstream of the merge, more than half of the vehicles traveling in the
shoulder lane originated from the onramp. Thus, the merging process did not exhibit the so
called "zipper effect" (Newman, 1986) whereby freeway and ramp vehicles share the shoulder
lane in a strictly alternating fashion.
Study of the Ncurves constructed from freeway counts at detector stations 30 and 40 (not shown here)
revealed these increased flows observed in Figure 2 were part of sustained surges in the freeway flows as well
as in the flow from the Spadina Avenue onramp.
122 Transportation and Traffic Theory
Important features also not shown in Table 1 are that the bottleneck always formed at the same
location (i.e., between detectors 60 and 70) and that it was always activated by a sustained
surge in the flow from upstream. Thus, traffic transitioned from free flow to queued conditions
in a predictable way; the queues formed at an inhomogeneity, the bottleneck, due to
reproducible, exogenous reasons, i.e., the increased flows. In this instance, one might presume
that the freeway's horizontal curve is the inhomogeneity creating the bottleneck (see Figure 1).
While this may indeed be the case, it is worth noting that the same analysis methods applied to
data from another freeway location found that a bottleneck consistently formed more than a
kilometer downstream of an onramp, even though there was no obvious inhomogeneity at this
location (Cassidy and Bertini, 1999). In any event, our studies to date have revealed no
evidence suggesting that traffic can break down and form queues in a spontaneous manner.
Also of note, rows 5, 6 and 7 of Table 1 indicate that, while the bottleneck was active, the
average discharge flow (in a given lane) exhibited only small variation across days. On each
day, these rates can be described as "nearconstant" since the cumulative counts never deviated
much from a linear trend. Given these predictable features of its discharge rates, it seems
reasonable to postulate about how queues might evolve upstream of this bottleneck; e.g., by
using a continuum model of highway traffic (Lighthill and Whitham, 1955; Richards, 1956;
Newell, 1993).
Observations at a Freeway Bottleneck 123
Table 1
Some Observations Taken From Station 80
D a y l Day2 Day3
ROW
3/5/972/20/97 7/21/97
~1 Maximum Flow (vph)  Median Lane 2630 2630 2400
2 Measured Duration of Maximum Flow (min: sec)Median Lane 43:40 5:40 19:40
3 Flow Collapse (vph)  Median Lane 2280 2110 2300
4 Measured Duration of Flow Collapse (min: sec)Median Lane 19:20 7:00 5:20
5 Average Queue Discharge Rate (vph)  Median Lane 2340 2290 2330
6 Average Queue Discharge Rate (vph)  Center Lane 1920 1910 1950
7 Average Queue Discharge Rate (vph)  Shoulder Lane 1720 1690 1690
In closing, we note that the observations reported above bring to light a number of unanswered
questions. For example, the reason(s) why the observed lanechanging trends persisted well
downstream of the onramp, and the extent to which the high flows observed in the median
lane might be reproduced at other bottlenecks, are unknown. Also unknown are the causes of
the flow reductions that accompanied queueing, especially the relatively large reductions at the
onset of queueing. Finally, the potential for using control measures, such as ramp metering, to
extend the periods marked by high flows (observed prior to queueing, for example) is
uncertain.
Answers to the above will only come through additional empirical study. Since (freeway)
bottlenecks come in many forms, including merges, diverges, weaves and lane reductions, and
since the traffic patterns on each type of bottleneck may exhibit their own peculiarities, the
study of bottlenecks in each of their forms seems warranted. Cumulative curves like those
described here might be used to conduct these studies since they provide a robust way of
diagnosing the details of bottleneck traffic.
ACKNOWLEDGEMENTS
The authors are indebted to Mr. David Nesbitt, City of Toronto, for providing the data used in
this study, and to G.F. Newell for his helpful comments.
124 Transportation and Traffic Theory
REFERENCES
Cassidy, M.J. and R.L. Bertini (1999). Some traffic features at freeway bottlenecks. Transpn.
Res., 33B, 2542.
Cassidy, M.J. and J.R. Windover (1995). Methodology for assessing dynamics of freeway
traffic flow. Transpn Res. Rec., 1484, 7379.
Daganzo, C.F. (1997). Fundamentals of transportation and traffic operations. Elsevier, New
York, p. 133.
Lighthill, M.J. and G.B. Whitham (1955). On kinematic waves. 7: Flood movement in long
rivers. //: A theory of traffic flow on long crowded roads. Proc. Royal Soc., A229, 281
345.
Lin, W.H. and C.F. Daganzo (1997). A simple detection scheme for delayinducing freeway
incidents. Transpn Res., 31A, 141155.
Newell, G.F. (1982). Applications ofqueueing theory. Chapman Hall, London.
Newell, G.F. (1993). A simplified theory of kinematic waves in highway traffic 7: General
theory. 77: Queuing at freeway bottlenecks. 777: Multidestination flows. Transpn Res.,
278,281313.
Newman, L. (1986). Freeway operations analysis course notes. Institute of Transportation
Studies, University Extension, Univ. of California, Berkeley, U.S.A.
Persaud, B.N. (1986). Study of a freeway bottleneck to explore some unresolved traffic flow
issues. PhD thesis, Univ. of Toronto, Toronto, Canada.
Persaud, B.N. and V.F. Hurdle (1991). Freeway capacity: definition and measurement issues.
Proc., International Symposium of Highway Capacity, A.A. Balkema press, Germany,
289307.
Persaud, B., S. Yagar and R. Brownlee (1998). Exploration of the breakdown phenomenon in
freeway traffic. Transpn Res. Rec., 1634, 6469.
Richards, P.I. (1956). Shock waves on the highway. Opns. Res., 4, 4251.
125
Gordon F. Newell
Department of Civil and Environmental Engineering and Institute of Transportation Studies
University of California at Berkeley
ABSTRACT
Suppose that the local capacity of a highway is a smooth function of location, approximated by a
parabolic function with a minimum value at some location (the bottleneck). The flow approaching
the bottleneck increases approximately linearly with time as it exceeds the capacity of the
bottleneck. We present here an analytic solution for the resulting flow pattern upstream of the
bottleneck as predicted by the theory of Lighthill and Whitham (1955) for two different types of
analytic forms for the relation between flow and density.
Although, in each of the two cases, the formulation of the problem contains seven parameters, it is
shown that, by appropriate linear transformation of variables, the flow pattern can be described in
terms of a single dimensionless pattern. In each case, a shock first forms at some point upstream of
the bottleneck with an amplitude which increases proportional to the square root of the time from
its beginning.
126 Transportation and Traffic Theory
1. INTRODUCTION
Suppose that the capacity at each location along a section of highway is a smoothly varying
function of location, as might be the case if the highway curves or changes grade, and that the
capacity has a minimum value at some location which we arbitrarily identify as x = 0. Over some
distance in the vicinity of the bottleneck, x = 0, we will assume that the capacity Qm(x) can be
approximated by a quadratic function of the form
Qm(x) = q0 + Ax2 (1.1)
with q0 the capacity of the bottleneck and A some positive constant describing how rapidly the
capacity changes with the distance from the bottleneck.
Suppose also that at some location x =  L , well upstream of the bottleneck, there is a flow q(L, t)
approaching the bottleneck. This flow is increasing with time and at some time t = t0 becomes
equal to the capacity, q0. Over some interval of time near to we will assume that the flow is
increasing (nearly) linearly with time so we can approximate
q(L,t) = q0 + B(tto) (1.2)
for some positive constant B describing the rate of increase of the flow.
A flow larger than q0 certainly cannot pass the bottleneck nor can it pass any point upstream oi
the bottleneck where that flow exceeds the local capacity. As this flow approaches the bottleneck
vehicles accumulate behind the bottleneck. The accumulation will eventually create a shock which
propagates upstream.
Lighthill and Whitham, LW, (1955, p. 333) gave a qualitative description of how the shock forms
and propagates upstream based upon their theory of kinematic waves. In this theory it is i
postulated that there is a specified functional relation between the density, k(x, t), and the flo\\
q(x, t),
q(x,t) = Q(k(x,t),x). (1.3)
For any fixed x, Q(k, x), is a concave function of k having a maximum with respect to k, th<
capacity Qm(x) at that location. LW did not specify any specific form for Q(k, x) but it wa;
assumed to be a smooth function of k with a locally parabolic maximum with respect to k, and tha
Qm(x) had a minimum at the bottleneck.
Flows Upstream of a Highway Bottleneck 127
Our objective here is to give an analytic solution for q(x, t) based on the LW theory for two
hypothetical forms for the function Q(k, x). In the first case it is assumed that Q(k, x) for fixed x
has a triangular shape, as illustrated in figure 1 for several values of x. The slope Q(k, x)/k on the
left hand side of the triangle is the "free speed" v0 assumed to be independent of x. The slope on
the right hand side of the triangle, the (negative) wave velocity, is also assumed to be independent
of x. The maximum height of the triangle is Qm(x) as in (1. 1). Thus the family of Q(k, x) curves for
various x differ only by a scaling factor Q(x) for both q and k.
Q(x)
CT
Q(x)
ko
Density, k
Density, k
Fig. 1. Triangular qk relations
Fig. 2. Parabolic qk relations
In the second case, it is assumed that the function Q(k, x) has a local parabolic maximum with
respect to k at some density ko and some specified curvature. The ko and curvature are both
independent of x. Thus for k in some vicinity of k, we can, approximate Q(k, x) by
Q(k, x) = q0 + Ax2  C(k  ko )2 (1.4)
for some positive constant C. This form is illustrated in figure 2.
We will show for each of the two forms for Q(k, x) that, by appropriate choice of units and
coordinates, q(x, t) can be described in terms of a single dimensionless function q*(x*, t*) of a
dimensionless length x* and time t*, independent of the parameters q0, A, B, C, L etc. The
function q*(x*, t*) can also be evaluated analytically. Similar methods could be applied also to
other types of q  k relations.
128 Transportation and Traffic Theory
2. TRIANGULAR Q  K RELATION
For the triangular q  k relation illustrated in figure 1, it is advantageous first to introduce a moving
time origin traveling at velocity v0. If we measure time at each location x from the time that some
reference vehicle traveling at velocity v0 would pass (Newell 1993II)
then the relation between q' and k' would have the form illustrated in figure 3 in which the "free
speed" is, in effect, infinite in the new coordinates. The negative "wave pace" ( w0), the reciprocal
of the wave speed, is ( l/v0) plus the corresponding wave pace of figure 1, with w0 independent of
This eliminates the parameter v0, in the sense that q'(x, t') is a solution of the LW equations for
the q'  k' relation of figure 3, which is independent of v0.
According to the LW theory, the flow q(x, t') in the absence of shocks must be a constant q' along
"characteristic curves" which travel at the wave velocity associated with the flow q'. For the q'  k'
relation of figure 3 there are only two possible values for the wave pace 0 or w,. Some
characteristic curves are illustrated in figure 4. (The numerical values of the flow, distance, and time
in figure 4 are the dimensionless values q*, k*, and t* to be specified shortly.) The characteristic
curves for flow less than q0 (not shown in figure 4) are vertical lines that extend through the
bottleneck. The characteristic curve for the flow q0, labeled as 0 in figure 4, travels vertically until
it reaches the bottleneck at x = 0 then travels backward from that point at wave pace w0. A
characteristic curve for some flow larger than q0 travels vertically until it reaches the location where
the local capacity is equal to the flow, and then turns back from that point at pace w0.
If q(L, t) is linearly increasing in t, q'(L, t') is linearly increasing in f. Since the characteristic
curves (before they turn back) are vertical, a linearly increasing flow at one value of L implies a
linearly increasing flow at other values of L. Thus the value of L is irrelevant, and we can
arbitrarily choose the time origin (i.e. the t0) or the L so that the characteristic for flow qo reaches
the bottleneck at t' = 0.
Q(x)
Density, k'
The locus of points where the characteristic curves turn back are points where
q'(x, t') = q0 + Bt' = q0 + Ax2
i.e.
Bt' = Ax2 (2.4)
shown in figure 4 by a dotted curve. Obviously figure 4 depends on the values of q'(x, t')  q0, but
the value of q0 itself is irrelevant. If we increase the flows by some fixed amount and increase the
capacities by the same amount the figure would not change.
A shock will form at the earliest time when two characteristic curves of different flows intersect.
This is obviously at such time t[ , location X], and flow qi , when the slope of the curve (2.4) is
equal to the wave velocity l/w 0 , i.e.
dt' _ 2A
— ^1 — Wn .
dx B ' °
x, =  w 0 B / 2 A , t [ = (2.5)
130 Transportation and Traffic Theory
We still have the option of choosing units for x, t! and q'  q0. We could, for example, choose a
"dimensionless" distance x*
x* = x2A/Bw0 (2.6)
measured upstream of the bottleneck so that x,* = + 1, and a dimensionless time t*
t* = t'2/Bw<> (2.7)
Flows Upstream of a Highway Bottleneck 131
so that the dimensionless wave velocity or pace is +1 (in the direction of increasing x*) t,* = 1/2 .
We can also choose the units of flow so that
q*(x*, t*) = [q'(x, t')  q0] 2A/B2w02. (2.8)
The approaching flow is then q*(x*, t*) = t* and the q,* at t* = 1/2 is 1/2.
Thus, starting from a formulation which contained potentially seven parameters, q0, A, L, B, t0, v0,
and WG , the problem has been reduced to a dimensionless form with no parameters. The q*(x*, t*)
is equivalent to the q(x, t) for q0 = 0, A = 1/2, L arbitrary, B = 1, t0 = 0, v0 = °° and w0 1 (or +
1)
Finally to evaluate the q*(x*, t*) we note that, at any point (x*, t*), it can have only one of three
possible values:
q*(x*,t*) = t* (2.9)
if it is determined by the approaching flow,
q*(x*,t*) = 0, (2.10)
if it is determined by the flow which can pass the bottleneck at x = 0, or by a value determined
from the characteristic curve passing through (x*, t*) coming from the curve (2.4), t* = (l/2)x*2
where the flow is q* = t*. This last value gives
q*(x*, t*) = 1  (x* t*)  [1  2(x*  t*)]l/2. (2.11)
This last expression is a function only of x* 1* since the flow is constant along the characteristic
curves x*  t* = constant, but it applies only for 0 < x*  t* < 1/2. One can verify that (2.11) is
correct by checking that along the curve (2.4), t* = (l/2)x*2, (2.11) gives the approach flow q*(x*,
(l/2)x*2) = (l/2)x* 2 = t*.
For small values of (x* 1*) an expansion of (2.11) in powers of x* 1* gives
q*(x* , t*) = (x* 1*)2/2 + (x* 1*)312 +... (2.11 a)
thus q*(x*, t*) vanishes quadratically in (x* 1*) as x* 1* —> 0. For x* 1* close to 1/2, however,
q*(x*, t*) varies rapidly with a 1/2 power singularly. The function (2.11) is illustrated by the
curve of figure 5.
For 0 < x* < 1, it is obvious that (2.9) applies for t* < (1/2) x*2, (2.11) for (l/2)x*2 < t* < x*, and
(2.10) for x* < t*. For 1 < x*, however, the boundary separating these solutions is a shock path.
132 Transportation and Traffic Theory
LW determine the path of the shock by integrating the equations for the velocity of the shock. A
simpler and more general method (Newell 19931) is to evaluate the cumulative flow and require
that the cumulative flow be continuous across the shock.
c>
u_
0.5
r •
Time , tj *x *
if the waves at (x*, t*) are going backward. Since q*(x*, t*) = 0 along the characteristic curve t* = 0
for all x* > 0 and also at the bottleneck x* = 0 for all t* > 0, we can define A*(x*, t*) = 0 along
these lines. The A*(x*, t*) can be evaluated by integrating (2.12) along any path from a point
where A* = 0 to (x*, t*).
On the upstream side of the shock (2.9) applies. Integration of this with respect to t* for fixed x*
gives
A*(x*, t*) = (l/2)t* 2 . (2.14)
On the downstream side of the shock, for t* > x*, (2.10) applies and k*(x*, t*) = (l/2)x*2.
Integration of this with respect to x* for fixed t* gives
A*(x*, t*) = (l/6)x 3 , t*> x* . (2.15)
Equating (2.14) and (2.15) we conclude that the path of the shock is
(l/2)t*2 = (l/6)x *3 ; x* = 3(t*/3)2/3 (2.16)
if t* > x*, or, equivalently, for t* > 3, x* > 3.
The formula for the shock path for 1 < x* < 3 is much more complicated because the flow on the
downstream side satisfies (2.11). In the region where (2.11) applies, we can evaluate A*(x*, t*) by
integrating the flow (2.11) along a line of constant x* from the point (x*, x*) where A*(x*, x*) =
(l/6)x*3.
x*3 1 3Cx*t*1 2 7
= — + {l3(x*t*) + — —[l2(x*t*) }. (2.17)
6 3 2
If one integrates the expansion (2. 11 a), one can verify directly that
x *3 cx*_t*^3 ]
A*(x*,t*) =  ——  (x*t*) 4 +.... (2.17a)
6 6 8
Over most of the region where (2.17) applies, the second term of (2.17) is very small. Even at the
point where the shock first forms at x* = 1, t* = 1/2 where x*  t* is largest, the second term of
(2.17) is only (1/4) times the first term and the relative size of this term decreases very rapidly as
(x* 1*) decreases.
The path of the shock for 1 < x* < 3 is obtained by equating (2.17) and (2.14). It is quite easy to
evaluate numerically the small deviation of the shock path from the curve defined by (2.16).
134 Transportation and Traffic Theory
Figure 4 shows the path of the shock and the curves of constant flow q*. Figure 5 shows the flow
q*(x*, t*) as a function oft*  x* for various values of x*.
To give a more complete description of the solution one might wish also to draw some vehicle
trajectories. For any dimensionless solution q*(x*, t*), however, there is a whole family of
possible (dimensionless) trajectories depending on the capacity q0 of the bottleneck or its
dimensionless form
q0* = 2A qo/B2 w02.
To see this one need only observe that the "dimensionless cumulative flow" A*(x*, t*) defined by
(2.12) is the dimensionless cumulative flow for a hypothetical capacity qo  0 or, equivalently, it is
the actual dimensionless cumulative flow less a "background" cumulative flow of q0*t*. Thus, for
any q0* > 0, the dimensionless cumulative flow is actually
q0*t* + A*(x*, t*)
and the trajectories are the curves for which this is a constant. In particular, the trajectory which
passes x* = 0 at some time t0* is the curve
A*(x*, t*) = q0*(to* 1*). (2.18)
There is just one dimensionless function A*(x*, t*) as described above; the t0* labels the
trajectory, but there is still the parameter q0*. There is a different set of trajectories for each qo*.
In the region where the characteristics of figure 4 are vertical, A*(x*, t*) is given by (2.14) and the
trajectories are defined by the equation
t*2/2 = q0*(t0* 1*). (2.18a)
As expected, the trajectories are also vertical, i.e. t* is independent of x* (the vehicle speed and the
wave speed are the same), but the labels of the trajectories (vehicle number) depend on q0*.
In the region t* > x*, the trajectories are also simple. From (2.15)
x*3/6 = q0*(to* 1*) (2.18b)
which means that the time displacement of a trajectory from a vertical trajectory is proportional
to x*3 (for any qo*). The velocity in the x*, t* coordinate system is proportional to x*"2.
In the region where (2.17) applies, the trajectories are more complicated but, for any q0*, thi
trajectories are continuous across the shock path.
Flows Upstream of a Highway Bottleneck 135
3. PARABOLIC Q  K RELATION
For the parabolic q  k relation described by ( 1 .4) the formulas are somewhat more complicated
because the characteristic curves are not straight lines. In the absence of shocks the formal solution
of the LW equations is that q(x, t) is constant along characteristic curves in the x  t plane, x(t; q),
having a slope, for fixed q,
dt ok
Along the characteristic curve for flow q, the density as given by (1.4) is
k  k 0 = ± [ q 0  q + Ax2]1/2C1/2.
Thus the equation for the characteristic curve of flow q is
dx(t; q)/dt = ± 2(CA)I/2 [(q0  q)/A + x2]1/2. (3.2)
in which either the + or  sign might apply depending on whether the velocity of the wave is
positive or negative.
in which D(q) is some unspecified "integration constant" for each value of q. The D(q) is to be
determined from appropriate boundary or initial conditions. In particular, we assume here that
"sufficiently far" upstream at x = L, the flow satisfies (1.2).
Sufficiently far upstream is now interpreted to mean at a value of L such that the hyperbolic
functions in (3.3) can be approximated by positive exponentials
sinh x ~ cosh x ~ (1/2) exp (x).
Thus in (3.3), the boundary condition requires that
L = ±  q0  q(L, T)  1/2(l/2)A'/2 exp(2(AC)1/2 T + D(q(L, T))
for all T . Equivalently if we represent I as a function of q rather than q as a function of T
through (1.4), this gives
L = ±  q0  ql/2(l/2)A/2 exp( 2(AC)1/2 (t0 + B'1 (q0  q)) + D(q) ).
We would naturally specify that, for some sufficiently large L, the flow would reach q0 at some
"large" negative t0, so the argument of exp( • ) above is negative and the ± sign must be . Thus
exp(D(q)) =  q0  q  1/2(2L)'A1/2 exp(2(AC)1/210  2(AC)I/2 (q0 
136 Transportation and Traffic Theory
This relation depends on t0 and L only through a single factor L"1 exp(2(AC)1/2 t0) on the righl
hand side. Since a change in L is equivalent to a suitable change in t0, this means that if the flow is
linearly increasing with t as in (1.2) for some sufficiently large L, it is also linearly increasing with t
at other large values of L but with a displaced value of t0. Any change in this factor is equivalent to
a translation of the time coordinate and, by suitable choice of the t0, we can assign this factor any
value we wish. In particular, we will choose the time origin so that
exp(D(q)) = (AC)1/4 [2(q0  q)/B]1/2 exp(2(AC)1/2 (q0  q)/B).
The equations for the characteristics (3.3) can now be written in the simple dimensionless
form
x*(t*; q*) = + e'*+q* +q*e'*'q* (3.4)
with
x*(t*; q*) = 23/2A 3/4C1/4B/2 x(t; q)
(3.5)
I/2 1/2
t*  2(AC) t, q* = 2(AC) (q0  q)/B.
Equation (3.4) is valid either for q < q0 or q > q0 since the q* changes sign when q passes q0 .
The x*, t*, and q* are simply rescaled dimensionless versions of the x, t, and q  q0, respectively
(but not necessarily related to the corresponding symbols in section 2). Again, as in section 2, the
formulation of this problem started with seven parameters q0, A, L, B, t0, ko, and C, the first five
of which are the same as before but the parameters ko and C associated with the shape of the
parabolic q  k curve replace the parameters v0 and w0 associated with the triangular q  k curve
Again the final dimensionless form of the characteristic curves contain none of these parameters.
To complete the solution of the dimensionless flow pattern we note that (3.4) can also be writter
in the form
f21 q* l l / 2 sinh(t * q * +l/2^n I q* I) for q* < 0
x*(t*;q) = < (3.6)
[+2q *' / 2 cosh(t * q * +1 / 2£n q*) for q* > 0 .
Thus the characteristic curves for q* < 0 and those for q* > 0 all have the same shape except for i
scaling factor (q*)I/2 and a translation in time by q*  (l/2)£n  q* .
Figure 6 shows a family of characteristic curves evaluated from (3.4) or (3.6). The curves for q*< (
pass through the bottleneck at x* = 0. Note that these are drawn for a geometric sequence of q"
values 1, 1/2, 1/4, etc. These characteristic curves will cover the whole space downstream of thf
Flaws Upstream of a Highway Bottleneck 137
bottleneck, but clearly the flow at any location downstream rapidly approaches the flow q*= 0 (q
= q0) as t* increases.
Time , t '
The characteristic curve for q* = 0 is simply an exponential e"1* which is asymptotically horizontal
for t*—>°°. This characteristic cannot pass the bottleneck. The characteristic curves for q* > 0
move forward in time until the flow q* reaches the location where the local capacity of the
highway is q*. At this point the slope of the characteristic curve becomes zero. Unlike the
characteristic for q* = 0, however, the slope of the characteristic curve for q* > 0 becomes zero at
a finite time after which the characteristic proceeds to move back upstream. Note that this is
138 Transportation and Traffic Theory
similar to the pattern of figure 4 except that in figure 4 the characteristic curves change direction
abruptly.
The points at which the characteristic curves become horizontal are identified in figure 6 by the
dotted curve. They occur where the argument of the cosh (•) in (3.6) vanishes
t* = q*  1/2 In q* , x* = 2q*/2
or
t* = (x*/2)2  In (x*/2).
As a function of q* or x*, this point starts at t* = oo for q* = 0. It decreases with q* until it
reaches a minimum at q* = 1/2 and then increases again. This is the analogue of the dotted line
curve of figure 4 except that figure 4 is drawn with a moving time origin.
After a characteristic curve turns upstream it will, at some time, intersect a characteristic curve of
higher flow moving downstream so as to create a shock. In figure 6 it appears as if all the
characteristics for 1 < q* < 2 intersect at (nearly) the same point. The shock actually starts at the
first time that 9x*(t*; q*)/3q* vanishes. From (3.4) we see that
, „„ ,
(l + —  — + ) + ^4^(1  — +
2 6 3V2 4
The first term on the right hand side of (3.8) is an expansion in powers of t' of the characteristic
curve for q' = 0. The second term is a power series expansion in q' for t' = 0. Successive terms give
power series expansions in q' of the coefficients of powers oft', t'2, etc (for powers of q' of at least
1).
Flows Upstream of a Highway Bottleneck 139
Since the q]*, t)*, X* were chosen so that 3x'/3q' = 0 at t' = 0, q' = 0, we knew that the second term
of (3.8) would not contain a term proportional to q'. We would have expected this term to start
with a (positive) term proportional to q'2 , but, by some coincidence, this term is also missing. The
second term of (3.8) starts with a term proportional to q'3. This is the reason why it appears in
figure 6 as if a wide range of characteristic curves are crossing at (nearly) the same point xi*, t].
Also in (3.8) the fourth term is proportional to q' 3 .
Equations (3.4), (3.6) or (3.8) define x* or x' as a singlevalued function of q* and t*, but, to
determine the path of the shock, one must determine the flow (and corresponding densities) on
either side of the shock, i.e. q* as a function of x* and t*. From figure 6 it would appear that, in the
region where the characteristic curves intersect, there are actually three values of q* at each point
x*, t*.
In (3.8) it is not a priori obvious how many terms one needs to retain in a first approximation
because one does not know the relevant relative magnitudes of t' vs q'. If one were to keep only
terms linear or quadratic in t' and q', one might approximate (3.8) by
x' = A/2 t'(l +1')  A/2 t'q' + ...
This could then be solved for q' as a function of x' and t'
q' = x'/2 1 / 2 t' + l + t' + ..
This describes what appears to occur in figure 6. Over some range of (small) q', the characteristic
curves (nearly) pass through x' = 0, t1 = 0. The value of q' at any point (x', t') depends mostly on
the slope x'/t' of the line from x  0, t' = 0. But this also shows that the relevant magnitudes of t'
and q' are not comparable. One is interested in values of q' large compared with t' and should retain
at least the term 2l/2q'3/3 in (3.8).
The shock first forms at t' = 0 as an "infinitesimal" shock between the values q' = ±(3t)1/2 and a:
such, it travels at essentially the wave velocity for q' = 0, dx*/dt* = (2) l/2 . But even as the jump ii
flow increases with t', the shock velocity between the flows q* = q,* + (3t) l/2 and q!*  (3t)1/2 wil
stay very close to the wave velocity at the average of these two flows, namely at the wav<
velocity for q1 = 0. Thus the shock path will, in turn, stay very close to the characteristic curve fo
q' = 0 even though the amplitude of the shock is increasing rapidly with time.
Starting with (3.9) as a first approximation one can iteratively obtain higher order approximation
by including other terms from (3.8). Figure 7 shows a magnitude view (by a factor of 10) of thi
characteristics in the vicinity of X] , t ) . The cross indicates the place where the shock begins. Not
that there is a rather large curvature of the shock path near t,*; the shock velocity increases fron
(2) l/2 = 1.41 at t,* = 1.85 to about 2 fort,* = 2.0.
To determine the path of the shock over long distances it is advantageous to introduce
dimensionless cumulative flow as in (2.12), but now the dimensionless density is given by
The boundary conditions specify that, for any large fixed x*, q*(x*, t*) is linearly increasing in t*
on the upstream side of the shock, so
A*(x*, t*) = (l/4)x *2 + (l/2)(t* + In x*)2 for large x* (3.14)
From this one can show that
D*(q*) = 2q* + 2q*2
so that (3.11) determines A*(x*, t*) parametrically as a function of q* along the characteristic
curve (3.4) describing x*(t*; q*) also as a parametric function of q*. Where characteristic curves
142 Transportation and Traffic Theory
intersect each other giving multiple values of q* for the same x*, t*, there will also be multiple
values of A*(x*, t*). The "correct" A*(x*,t*) is the smallest of its multiple values and the shock
path is where the A*(x*, t*) becomes multiplevalued (Newell 19931).
To determine the path of the shock for large t*, it suffices to note that, for large positive t*, q* is
nearly zero for all x* on the downstream of the shock and A*(0, t*) is also nearly zero for large t*.
If, as in section 2, we integrate (2.12) along a line of constant t*, but now with the density given
by (3.10) for q*(x*, t*) = 0, we have
The path of the shock is now obtained by equating (3.14) and (3.15)
or
t = x*.ftix*. (3.16)
Actually the formula (3.16) is fairly accurate soon after the shock forms. One can show that to 2
second approximation
t = x*  &ix*  x*e x*
The additional term is negligible for x* > 4, only about one distance unit after the shock forms
Figure 6 shows the shock path.
As with the triangular q  k relation, one might also like to describe the vehicle trajectories. For an>
dimensionless q*(x*, t*) as illustrated in figures 6 and 7, however, there is a twoparameter family
of possible trajectories. For the triangular q  k relation there was only a oneparameter family ir
the transformed coordinate system but in the original coordinate system the trajectories would alsc
depend on the freeflow speed v0.
The dimensionless cumulative flow defined by (3.11) is a cumulative flow less a dimensionles:
background flow of q0* and density ko* so the actual dimensionless cumulative flow is
A*(x*, t*) + q0* t* + ko* x* (3.17)
with
q0*  2(AC)1/2 q0/B, k0* = 21/2 A1/4 C3/4B'1/2 k0.
Although there is only one dimensionless A*(x*, t*), there are two arbitrary parameters qo* and
ko* in (3.17) (regardless of how these may be related to the q0 and ko).
Flows Upstream of a Highway Bottleneck 143
The exact parametric representation of A*(x*, t*) along characteristic curves is not very
convenient for the evaluation of trajectories but, for most trajectories which pass through the
shock, (3.15) is valid downstream of the shock. The trajectory which passes x* = 0 at time t0*
satisfies the equation
qo*(to* 1*)  V x* = x*2/4. (3.18)
Thus the deviation of a trajectory from a straight line is proportional to x*2 (instead of x*3 as in
(2.18b)). Upstream of the shock, according to (3.14), this trajectory would satisfy
4. SOME ESTIMATES
There is no guarantee that traffic behaves in the manner described here, but, for the theory to be
plausible, the various scaling parameter should have values in a reasonable range.
In a typical rush hour we might expect the flow per lane to increase at a rate of about 1000 cars/hr.
in a period of about 1 hour. Thus a reasonable value of B would be about 103/hour2. For the
triangular q  k relation a plausible value of w0 is about 1/10 hr/km (a wave velocity of lOkm/hr). A
"typical" value of A in (1.1) is not as well defined but it might be helpful to relate the A to a
distance L* required for the capacity (1.1) to change by 100 vehicles/hr. from its value at x = 0, i.e.
let
100/hr = AL* 2 , A = 100/L*2hr.
With these values of A, B, and w0, (2.5) gives
_ L* 2 , _ L *2 hr _ _ 20L* 2
X
' ~ 2 k m ' ' ~ 40 (km) ' ' q° ~ km 2 hr
t 2 q
If the capacity of the road section did vary as in (1.1), a value of L* = 1km would certainly
described a very slowly changing capacity for typical roads. The prediction here is that, for L* = 1
km, a queue would first form about 1/2 km upstream of the bottleneck (a rather substantial
distance) at a time t] 1 of about 1 minute and qi  q0 = 25 veh./hr.
144 Transportation and Traffic Theory
The value of \\, being proportional to L*2 , decreases rapidly with decreasing L*. If L* is reduced
to 1/2 km, X] becomes only 1/8 km.
For the parabolic q  k relation a typical value of C in (1.4) would be such that the flow would
drop by about 2000 veh./hr from capacity at a density displacement of about 100 veh./km. Thus,
2000/hr = C(100)2/km2, C ~ (1/5) km2/hr.
For the same values of B and A as above, (3.5) and (3.7) predict that the shock forms at
x, ~ B(1/2) C(1/4) A(3/4) ~ (1.5)L*(3/2) (km)  (1/2 >.
This value of x, will typically be considerably larger than for the triangular q  k relation because
the parabolic q  k relation provides less room to store excess vehicles near the bottleneck (but
more space far from the bottleneck). For L* = 1 km, the value of x, is about 1.5 km. The X]
decreases with decreasing L*, proportional to L*(3/2) , but not as rapidly as for the triangular q  k
relation. For L* = (1/2) kin, x, is still about (1/2) km.
5. CONCLUSIONS
We have described here an analytic solution of the problem posed in the introduction. No claim is
made that the theory described here is consistent in detail with how traffic actually behaves bul
any attempt to verify the LW theory should recognize that a shock wave does not necessarily
start at the bottleneck itself. It may begin at some point upstream of the bottleneck at a poinl
depending on how the capacity varies with location (the A of eq. (1.1)), the rate at which the flow
increases (the B of eq. (1.2)), and the shape of the q  k curve near its maximum. Despite all the
parameters in the formulation of the problem, the final evaluation of the flow pattern is expressed
in terms of a single dimensionless form. All the parameters are absorbed in coordinate
transformations.
Installations of vehicle detectors at multiple locations along freeway sections at various places
throughout the world are presently providing very detailed information about traffic flows or
freeways. Data from these system will soon provide a means of testing the validity or deficiencies
of various models of traffic flow, including the LW theory, but to deal with masses of data ai
different locations will required special techniques of analysis.
Flows Upstream of a Highway Bottleneck 145
Many parameters are needed to characterize any particular set of observations and to compare one
set of observations with another. Even if the LW theory should fail to describe correctly some
features of the flow pattern, maybe some of the dimensional arguments will still be valid. If one
makes observations at one location having a certain set of parameters, such as the A, B, q0, ko, etc.
here, and then observations at other locations with different values of these parameters, but one
presents the data in terms of the rescaled coordinates, the A*(x*, t*) for example, then maybe one
will obtain nearly the same A*(x*, t*) for all locations. This, in itself, would mean that observation
at one location can be used to predict behavior at another location having different parameter
values.
The trick of comparing observations at two or more location by measuring time relative to a
coordinate system moving with the free speed has been exploited before. If there are no "delays",
an A'(x, t') derived from (2.2) would be independent of x. The trick of rescaling time and/or space
variables involves nothing more than relabeling the coordinate scales of x and/or t on a graph or
drawing graphs for two different locations on different physical coordinate scales.
The trick of subtracting some "background flow" from the actual flow or, equivalently, drawing
graphs of
A(x, t)  q0 t+ ko x (5.1)
for some possibly arbitrary values of q0 and ko is less "wellknown". This transformation does not
change the wave pattern or the characteristic curves and was exploited in section 3 to eliminate two
parameters from the dimensionless wave solution of section 3. This transformation, however,
does not eliminate these parameters from the trajectory pattern.
Cassidy and Windover (1995) have subtracted a background flow as in (5.1) from experimental
data, not to compare observations at different locations, but simply as a tool to "magnify"
observed flow variations in space and time. If, by appropriate choice of q0 and ko, one can greatly
reduce the magnitude of (5.1), then one can draw the graph on a magnified scale. The effect is quite
dramatic.
REFERENCES
Cassidy, M. J. and J. R. Windover (1995). A methodology, for assessing the dynamics of freeway
traffic flow. Transportation Research Record, (Washington, D.C.) (in press).
146 Transportation and Traffic Theory
Lighthill, M. J. and G. B. Whitham (1955). Kinematic waves II a theory of traffic flow on long
crowded roads. Proc. Royal Soc. (London) A 229, 317345.
Newell, G.F. (1993) A simplified theory of kinematic waves in highway traffic, I general theory, II
queueing at freeway bottlenecks, III multidestination flows. Transpn. Res. 27B, 281287,
289303,305313.
Theory of Congested Traffic Flow 147
Boris S. Kerner, DaimlerChrysler AG, FT1/V, HPC: E224, 70546 Stuttgart, Germany
ABSTRACT
Results of experimental observations of phenomena of selforganization in traffic flow on
German highways are presented. The observations allow to suggest that there are at least two
phenomena of 'selforganization without bottlenecks' in real traffic flow: (i) The spontaneous
formation of a local region of synchronized traffic flow in an initially free traffic flow and (ii)
The spontaneous formation of a traffic jam in synchronized traffic flow. A theory of
congested traffic flow which may qualitatively explain these and other diverse effects of self
organization in real traffic flow is discussed.
1. INTRODUCTION
1.1. SelfOrganization
A different scenario of selforganization for the jam's formation has been proposed in 1994 by
Kerner and Konhauser (1994), and later in other papers (Bando et al, 1995; KrauB et al.,
1997, Barlovic et al., 1998): Before the density range mentioned above is reached, there
should be a broad range of lower densities where homogeneous states of free traffic flow are
metastable states (see an explanation of 'metastable states' below in Sect. 1.4) and 'the local
cluster effect' leading to jam formation can occur (a comparison of the local cluster effect in
different traffic models can be found in (Herrmann and Kemer, 1998; Kerner, 1998b).
On the contrary, there are alternative theories which claim that there are no selforganization
processes in traffic flow and that all phenomena of the formation of spatialtemporal patterns
in traffic flow are without exception determined by an influence of on and offramps or other
freeway bottlenecks. As the consequence of these theories it is also often claimed that spatial
temporal behavior of traffic flow can be described satisfactorily by the wellknown classical
LighthillWhithamtheory of traffic flow (see Whitham, 1974), which is based on the
LighthillWhithammodel or by further developments of such 'first order' traffic flow models
where no selforganization processes are possible (e.g., Daganzo, 1997; Daganzo et al., 1998).
Only results of experimental observations of real traffic flow are able to answer the question
which traffic flow theories may actually explain related phenomena in real traffic flow. One
may mention two wellknown phenomena experimentally observed in real traffic flow:
A. The occurrence of spatialtemporal patterns consisting of different traffic jams  the 'stop
and go' phenomenon (e.g., Treiterer, 1975; Koshi et al., 1983) and
B. The breakdown phenomenon in the vicinity of a freeway bottleneck which is often
accompanied by the related 'capacity drop' (Fig. l(a)) (e.g., AgyemangDuah and Hall,
1991; Cassidy and Bertini, 1998; Persaud et al., 1998).
Are these phenomena the result of some selforganization processes or are they only the result
of initial permanent nonhomogeneity caused by freeway bottlenecks !?
Recently the nature of the breakdown phenomenon in the vicinity of a freeway bottleneck has
been disclosed (Kerner and Rehbom, 1997): This breakdown phenomenon is linked to an
occurrence of a first order local phase transition 'Free Flow => Synchronized Flow' (the
definition of synchronized traffic flow will be given in Sect. 1.6). Such local phase transitions
are one phenomenon of selforganization. Kerner (1998c) discovered the nature of the 'stop
andgo' phenomenon: It turned out that this phenomenon is linked to the double (cascade)
phase transitions 'Free Flow => Synchronized Flow => Traffic Jams and the 'pinch effect' in
synchronized flow which are both very complex phenomena of selforganization.
A freeway bottleneck usually causes a permanent nonhomogeneity in traffic flow (e.g., Koshi
et al, 1983; May, 1990; Daganzo, 1997; Cassidy and Bertini, 1998). On the one hand, it is
Theory of Congested Traffic Flow 149
wellknown from investigations of physical, biological and chemical systems (Kerner and
Osipov, 1994) that a permanent nonhomogeneity acts as 'a permanent nucleus' for phase
transitions: Spontaneous formation of a spatialtemporal pattern occurs considerably more
frequently in the vicinity of the permanent nonhomogeneity. Such phase transitions can occur
in a deterministic way, i.e., even if fluctuations are negligible. On the other hand, this
deterministic effect is also one phenomenon of selforganization (Kerner and Osipov, 1994).
Indeed, the reason of the spontaneous occurrence and the main nonlinear properties of
spatialtemporal patterns are nevertheless determined by intrinsic nonlinear properties of a
system. In other words, phenomena of selforganization can occur in such systems (although
usually considerably more seldom) even when no initial permanent nonhomogeneity exists.
The aim of this paper is to show that selforganization in experimentally observed traffic flow
occurs outside any bottlenecks. To show such 'selforganization without bottlenecks' in real
traffic flow, one needs to perform experimental observation of the behavior of traffic flow on
a long enough highway section where no on and offramps and other bottlenecks exist (Sect.
2). A qualitative theory of congested traffic flow (Kerner, 1998 a, b, 1999) which may explain
phenomena of selforganization will be considered in Sect. 3. However, before the new results
are considered (Sect. 2), some recent experimental observations which are an 'indirect' proof
of'selforganization without bottlenecks' (Sect. 1.4), some additional definitions (Sect. 1.5),
and features of synchronized flow which have been found out (Sect. 1.6) should be briefly
reviewed.
1.4. Metastable States of Free Traffic Flow and Characteristic Parameters of Wide Jams
as 'Indirect' Proof of 'SelfOrganization Without Bottlenecks'
Indeed, when wide jams propagate outside bottlenecks and free flow is formed in the outflow
of the jam, it has been found out (Kerner and Rehborn, 1996a) that there are the characteristic
(unique, or coherent) parameters of traffic flow: (i) the velocity v g of the downstream front of
a jam, (ii) the jam density p m a x , (iii) and also the flow rate q o u t , the vehicle density p m i n , the
average vehicle speed in free flow which is formed by a wide jam downstream. The
characteristic parameters do not depend on initial conditions. In particular, it has been found
out (Kerner and Rehborn, 1996b) that independently on initial conditions the characteristic
parameters were spontaneously selfformed during a development of any jam whose width
150 Transportation and Traffic Theory
Such a stationary propagation of the downstream front of a wide jam can be represented by a
line in the flowdensity plane. This characteristic line for the downstream front of the jam,
which will be called 'the line J', has the coordinates (p mjn , q out ) and (pmax, 0) in the flow
density plane; the slope of 'the line J' is equal to the mean value of the velocity of the
downstream front v g (Fig. l(b), line J). It should be noted that 'the line J' is not a part of the
fundamental diagram: It represents the characteristics of the downstream front of a wide jam.
The other result of experimental observations of wide jams found out in (Kemer and Rehborn,
1996a) is that the maximal possible flow rate in free flow qmraexe) can be considerably higher
than the flow rate out of a wide jam q out (Fig. l(b)):
q (free)
In
4 max /lout (1)
Therefore, in the range of the flow rate (and in the corresponding range of the vehicle density)
(Fig. l(b))
(2)
w
(free) A fl° rate flow rate
Qmax
(b)
Fig. 1. A possible shape of the fundamental diagram of traffic flow (a) (e.g., Brannolte, 1991;
Ceder, 1976; Hall, 1987; Hall, et al, 1986; AgyemangDuah and Hall, 1991; Koshi, et al,
1983) and (b)  the concatenation of states of free flow with the characteristic line for the
downstream front of a wide traffic jam ('the line J") (Kemer and Rehborn, 1996a; Kemer and
Konhauser, 1994).
at any chosen average vehicle density there are at least two different states of traffic: (i) free
traffic flow and (ii) wide jams. It must be noted that the existence of this range of the density
(2) is not linked to a freeway bottleneck. Therefore, a local perturbation of an initially in
average homogeneous free traffic flow can force an appearance of a jam, if the amplitude of
this local perturbation is high enough. Because there is always a finite probability of a
spontaneous occurrence of such a fluctuation in traffic flow outside any bottleneck, a traffic
jam can spontaneously occur in traffic flow without any influence from a freeway bottleneck,
i.e. 'selforganization without bottlenecks' in traffic flow can really occur.
Theory of Congested Traffic Flow 151
As well as in physical, biological and chemical systems, states of free traffic flow in the range
(2) may be called metastable states. Recall that a metastable state of a spatial system is stable
with respect to any infinitesimal perturbations. However, if the amplitude of a local
perturbation exceeds some critical amplitude, this local critical perturbation begins to grow.
The same definitions may be applied to traffic flow where local perturbations (fluctuations) of
traffic variables (vehicle speed, density and flow rate) usually occur. As it follows from (2),
the flow rate and the density q b = q o u t , pb = p min are the boundary (threshold) values
which separate stable states of free flow and metastable states of free flow with respect to the
jam formation (Kerner and Konhauser, 1994). This means that at q < q b (p < p b ) no jams can
exist for a long time or be exited in free flow (Fig. 2).
critical amplitude probability of
1
of local perturbation f first order local
phase transition /
!\ (a) 1 (b) /
^v_ ! _J
P b °r P s density p b or p density
(flow rate) (flow rate)
Fig. 2. Explanation of local first order phase transitions: (a) A qualitative shape of the
dependence of the critical amplitude of a local perturbation (see Fig. 6(a) in Kerner and
Konhauser, 1994); (b) the dependence of the probability of first order phase transitions which
is related to Fig. (a) (see Figs. 3 and 4 in Persaud et at., 1998). Threshold vehicle densities in
free flow are designated as p b for the phase transition 'Free Flow => Jam' and as ps for the
phase transition 'Free Flow => Synchronized Flow'.
The effect of the growth of a local perturbation in a metastable state of a system whose
amplitude exceeds the critical value, i.e., the growth of a critical local perturbation, is called
the nucleation effect. The critical local perturbation plays the role of a 'nucleation center' for
local phase transitions in an initial state of the system (e.g., Kerner and Osipov, 1994). A first
order local phase transition is a phase transition which occurs in a metastable state of a
distributed system and is caused by the nucleation effect. First order local phase transitions are
accompanied by a jumping (i.e., breakdown) behavior of variables of a system and hysteresis
effects. Some other general properties of first order local phase transitions are shown in Fig. 2:
The critical amplitude of the critical local perturbation is maximal at the threshold which
separates stable and metastable states of a system. This critical amplitude decreases as the
density (or the flow rate) deviates from the threshold value inside the metastable range.
Apparently, for traffic flow these properties have first been discovered by Kemer and
Konhauser (1994) from their theoretical investigation of the phase transition 'Free Flow =>
Jam' (Fig. 2(a)). Note that from statistical physics it is known that the probability of the
spontaneous occurrence of a fluctuation in a distributed system decreases with the increase in
the amplitude of this perturbation. Therefore, from Fig. 2(a) follows the other general property
152 Transportation and Traffic Theory
of first order local phase transitions: The lower the amplitude of the critical perturbation is,
the higher is the probability of the spontaneous occurrence of a first order local phase
transition (Fig. 2(b)). Note that the nature of the wellknown breakdown phenomenon in the
freeway bottleneck, as it has recently been discovered in (Kerner and Rehborn, 1997), is
linked to another first order local phase transition 'Free Flow => Synchronized Flow'.
Therefore, the latter phase transition should show qualitatively the same behavior as it is
shown in Fig. 2 (with a different threshold density p s ). The behavior of the probability of the
breakdown phenomenon in a freeway bottleneck shown in Fig. 2(b) has been discovered by
Persaud et al. (1998) that confirms the mentioned conclusion by Kerner and Rehborn (1997).
Observations (Kemer and Rehbom, 1996b) show that there are three qualitatively different
phases of traffic flow: (i) free flow, (ii) synchronized flow and (iii) wide jams. As a result,
there may be three qualitatively different types of phase transitions: 1) 'Free flow <=> Jam', 2)
'Free flow <=> Synchronized flow', and 3) 'Synchronized flow <=> Jam(s)'. A spontaneous
occurrence of these phase transitions determines the complexity of traffic flow observed in
experiments. It is also linked to the result (Kerner and Rehborn, 1997; Kerner, 1998c)) that all
these transitions are related to the same class of first order phase transitions, i.e., they are
accompanied by similar looking breakdown and hysteresis effects. Besides, for each of these
phase transitions the properties shown in Fig. 2 are valid. However, they must be
distinguished one from another because the nonlinear features of these phase transitions are
qualitatively different. This differentiation determines one of the main difficulties in traffic
flow theory.
In free traffic flow, due to the relatively low densities of vehicles, drivers on a multilane road
are able to change a lane and to pass. On the contrary, due to the higher density in 'congested'
flow, vehicles are almost not able to pass. As a result, a bunching of drivers both on each
individual lane and between different lanes of a highway can occur. Therefore, when all lanes
of a highway correspond to the same route without on and offramps or other bottlenecks,
drivers move with nearly synchronized average speed on the different lanes of the highway. A
possibility of such 'synchronized' ('collective') flow which is related to a part of the
fundamental diagram (e.g., Fig. l(a)) at higher density has been theoretically predicted by
Prigogine and Herman (1971). Synchronized flow has been observed by Koshi et al. (1983).
In congested flow a broad and complex spreading of measurement points which cover a two
dimensional region on the flowdensity plane is observed (e.g., Koshi et al, 1983). This
spreading is interpreted either as fluctuations, or as an instability, or else as a jam formation
(e.g., Hall, 1987; Helbing, 1997; Koshi et al, 1983; Hall etal., 1986).
It is wellknown that in traffic flow with high density traffic jams can appear (e.g., Treiterer,
1975). Inside a jam both the vehicle speed and the flow rate are very low or even zero. Recall
that in synchronized traffic flow, on the contrary, the vehicle speed is relatively low (but a
finite value) and the flow rate can be nearly as high as the flow rate in free flow. This
'quantitative' difference as it follows from results of experimental observations leads to
Theory of Congested Traffic Flow 153
qualitatively different nonlinear properties of traffic jams and of synchronized flow (Kerner
and Rehborn, 1996a, 1996b). Traffic flow with high density, where both traffic jams and
synchronized flow may occur, is called congested traffic flow.
Kerner and Rehborn (1996b) have found out that synchronized flow has totally different
dynamical properties in contrast to free flow. In particular, a multitude of states of free flow
may really be described by a curve on the flowdensity plane, i.e., by the fundamental diagram
(Fig. l(a), curve 'free'). On the contrary, even a multitude of homogeneous states of
synchronized flow (i.e., states which are homogeneous spatially and stationary in time; often
such states are called 'steady speed' states) cover a twodimensional region in the flow
density plane: A given vehicle speed (a steady speed) in a homogeneous state of synchronized
flow may be related to an infinity multitude of vehicle densities, and a given density may be
related to an infinity multitude of different speeds. In other words, the real dynamics of
synchronized flow cannot be described in the frame of the hypothesis about the fundamental
diagram: There is no fundamental diagram which is able to describe the properties even of the
multitude of homogeneous states of synchronized traffic flow (Fig. 3).
Q (free)
4 max
Fig. 3. Homogeneous states of traffic flow (Kerner and Rehborn, 1996b; Kerner, 1998a): (a)
Multitudes of homogeneous states of free (curve F) and of synchronized flow (hatched region)
on a multilane road, (b) A multitude of homogeneous states of flow on a onelane road.
The section of the highway has three intersections with other highways (II, "Friedberg", 12,
"Bad Homburger Kreuz" and 13, "Nordwestkreuz Frankfurt") and is equipped with 24 sets of
induction loop detectors (D1,...,D24) (Fig. 4). Each of the sets D4D6, D12D15, and D23, D24
consist of four detectors for a left (passing), a middle and a right lane, plus one for the lane
related to onramps or to offramps. The other sets of detectors are situated on the threelane
road without on and offramps, where each of them consist of three detectors only. Each
154 Transportation and Traffic Theory
traffic flow
1
o o o
—
a ro
—•
D ro
a oa o
ro to 10
. u> o en o o —' ro cx> b
] p ] 1 L JP \ 1 1 1 1 ] i ih i i i i \ i ii
] 1 hff ] ] ] ] 1 1 i 112 i i i i ] j ii
] ] ] 1 1 IQ ] 1 b i ]u ] ] ] ] i]
induction loop detector records the crossing of a vehicle and measures its crossing speed. A
local road computer calculates the flow rate and the average vehicle speed in one minute
intervals. The accuracy of the detectors according to official regulations is: better than 10% for
the flow rate higher than 600 veh/h, better than 20% for the flow rate lower than 600 veh/km;
better than 3% for the vehicle speed higher than 100 km/h; to within 3 km/h for the vehicle
speed lower than 100 km/h. If each vehicle during the interval of the averaging (one minute)
has the vehicle speed lower than 20 km/h, then the average speed is set to 10 km/h. Only if no
vehicle crosses a detector during the interval of the averaging, the average speed is set to zero.
2.1. Experimental Properties of Phase Transitions 'Free Flow => Synchronized Flow'
outside Freeway Bottlenecks
Let us first consider the dependence of the average vehicle speed (v) and flow rate (q) within
the time interval 06:20  06:40 (Fig. 5(a)) in the vicinity of the freeway bottleneck which exists
on the section of the highway due to the offramp inside the intersection 13 (D23, D22, Fig. 4)
and in the vicinity of the freeway bottleneck which exists due to the onramp inside the
intersection 12 (D15, D16, Fig. 4). It can be seen from the dependence of the average vehicle
speed (Fig. 5(a), left) and flow rate (Fig. 5(a), right) that in the whole time interval 06:20 
06:40 in the vicinity of both freeway bottlenecks free flow is realized. On the contrary, outside
from both bottlenecks some transitions between free and synchronized flow occur (D18, at
t « 06:27, t « 06:32 and t « 06:36). As it has already been mentioned in Sect. 1.6, in
synchronized flow the flow rate may be of the same order of magnitude as in free flow (Fig.
5(a), D18, right), but the vehicle speed is noticeably lower than in free flow and it is
approximately the same on different lanes of the highway (Fig. 5(a), D18, right, t>06:36).
To show that the transition from free flow to synchronized flow outside freeway bottlenecks,
which occurs at t=06:36 in the vicinity of the detectors D18 (Fig. 5(a), D18, left, up arrow), is a
local phase transition 'Free Flow => Synchronized Flow' outside freeway bottlenecks, spatial
temporal distributions of vehicle speed and the flow rate both upstream and downstream from
the detectors D18 at later times should be studied. The results of such an investigation in the
time interval 06:35  06:50 are shown in Fig. 5(b). It can be seen from Fig. 5(b) that transitions
from free to synchronized flow both upstream (D17, D16, left in Fig. 5(b), up arrows) and
downstream (D19, D20, left in Fig. 5(b), up arrows) occur later than at the detectors D18 (up
arrow at t=06:36, left in Fig. 5(b)). Besides, the greater the distance from the detectors D18, the
Theory of Congested Traffic Flow 155
later the transition from free to synchronized flow occurs. This conclusion is true both upstream
(D17, D16, left in Fig. 5(b), up arrows) and downstream (D19, D20, left in Fig. 5(b), up
arrows) of the detectors D18. This confirms that the transition from free flow to synchronized
Row first occurs only in the vicinity of the detectors D18, i.e., it is really a local phase transition
'Free Flow => Synchronized Flow' outside freeway bottlenecks.
06:20 06:25 06:30 06:35 06:40 06:20 06:25 06:30 06:35 06:40
v [km/h] 3000 7 1 [veh/h]
120 ,
1500 + :
0
06:20 06:25 06:30 06:35 06:40 06:20 06:25 06:30 06:35 06:40
v [km/h]
120 ^\
80  1500
40  D16
D16
0 o
06:20 06:25 06:30 06:35 06:40 06:20 06:25 06:30 06:35 06:40
v[km/h] 300
120 F :^x>^X^/ ° T q [vehih] D15
80 
1500
40 i
D15
0 0
06:20 06:25 06:30 06:35 06:40 06:20 06:25 06:30 06:35 06:40
Note that the transitions from free flow to synchronized flow, which occur upstream (D17,
D16, left in Fig. 5(b), up arrows) of the detectors D18 are linked to the appearance of a wave of
induced transitions from free flow to synchronized flow. Respectively, the transitions from free
to synchronized flow downstream of the detectors D18 (D19, D20, left in Fig. 5(b), up arrows)
are linked to a wave of the propagating synchronized flow. The propagation of synchronized
flow supplants free flow downstream. Indeed, as it has already been mentioned, the transitions
from free flow to synchronized flow, which occur upstream (D17, D16, left in Fig. 5(b), up
arrows) and downstream (D19, D20, left in Fig. 5(b), up arrows) of the detectors D18, occur
later than at the detectors D18. Besides, the greater the distance from the detectors D18 is, the
156 Transportation and Traffic Theory
later they occur. Therefore, the transitions from free flow to synchronized flow which occur
upstream (D17, D16, left in Fig. 5(b), up arrows) and downstream (D19, D20, left in Fig. 5(b),
up arrows) are not local phase transitions, but induced transitions. The induced transitions
upstream and the propagating synchronized flow downstream cause a widening of synchronized
flow, which has first spontaneously occurred in the vicinity of the detectors D18: The region of
localization of synchronized flow is widening both upstream and downstream over time. It
must be noted that the detectors D20D17 are situated outside bottlenecks. Therefore, the
spontaneous occurrence of the local phase transition 'Free Flow => Synchronized Flow' is
really a process of 'selforganization without bottlenecks' in traffic flow.
3000 T
1500
D20
0
06:50 06:35 06:40 06:45 06:50
3000 jq[veh/h]
1500 
D19
0
06:45 06:50 06:35 06:40 06:45 06:50
3000 T Q[veh/h]
(b) 1500 
D18
0
06:45 06:50 06:35 06:40 06:45 06:50
3000
1500 
D17
0
06:50 06:35 06:40 06:45 06:50
3000 TV
1500 "
D16
0
06:35 06:40 06:45 06:50 06:35 06:40 06:45 06:50
Fig. 5(a, b). Results of experimental observations of the phase transition 'Free Flow =>
Synchronized Flow' outside highway bottlenecks: The dependence of the vehicle speed (left)
and the flow rate (right) at the different detectors both downstream and upstream from the
location of the phase transition (D18) within time interval 06:20  06:40 (a) and 06:35 
06:50 (b) for three lanes of the highway (for the detectors D23 the vehicle speed on the off
ramp is additionally shown, D23off). Up arrows symbolically show the transitions 'Free
Flow => Synchronized Flow' at the related detectors.
Theory of Congested Traffic Flow 15 7
2.2. Experimental Properties of Phase Transitions 'Synchronized Flow => Jam' outside
Freeway Bottlenecks
The jam occurs due to the sequence of the double phase transitions 'Free Flow =>
Synchronized Flow => Jam' (Kerner, 1998c): (i) First the phase transition
'Free => Synchronized flow' occurs (D18, Figs. 5(b) and 7(a)). (ii) Then the 'pinch effect' in
synchronized flow, i.e., a selfcompression of synchronized flow is realized, where states of
synchronized flow are lying noticeably above the line J in the flowdensity plane (D20, Fig.
7(b)). (iii) In the pinch region outside bottlenecks a growing local perturbation appears (down
arrows at t = 06:50 in Fig. 6(a), D20).
In contrast to the local phase transition 'Free flow => Synchronized flow' (Fig. 5), both the
vehicle speed and the flow rate have decreased simultaneously and noticeably (down arrows
left and right at t = 06:50 in Fig. 6(a), D20). The further behavior of this local perturbation
shows the process of spontaneous selfformation of the traffic jam (down arrows in Fig. 6(a),
D19D15): The local perturbation gradually increases in the amplitude during its propagation
upstream (from the detectors D20 to D15, Fig. 6(a)), i.e., both the vehicle speed and the flow
rate in the jam decrease over time until they reach the values nearly zero (D16, Fig. 6(a)). It
should be noted that the detectors D20D17, where the processes of selfemergence and self
formation of the jam are realized, are located outside bottlenecks. Therefore, the processes of
selforganization: (i) the spontaneous occurrence of a local perturbation (D20, t= 06:50, Fig.
6(a)), (ii) the selfmaintaining and selfgrowth of this perturbation; (iii) the appearance of the
traffic jam due to the further selfgrowth of the local perturbation really occur outside
bottlenecks.
For an additional verification of the fact that the selforganization of the jam formation has
occurred outside bottlenecks, the vehicle speed and the flow rate at the bottleneck in the same
time interval as in Fig. 6(a) (from 06:45 to 07:10) have been studied (Fig. 6(b)). It can been
seen from Fig. 6(b) that no growing local perturbations have occurred downstream from the
detectors D20. Besides, both earlier and during the spontaneous occurrence of the local
perturbation (at t= 06:50, D20, Fig. 6(a)) and even later, when the process of the self
formation of the jam continues (D19  D17, Fig. 6(a)), only free flow exists in the vicinity of
the freeway bottleneck downstream (D23 and D22, Fig. 6(b)).
158 Transportation and Traffic Theory
D20
0
06:45 06:50 06:55 07:00 07:05 07:10 06:45 06:50 06:55 07:00 07:05 07:10
120 4v[km/h] 3000 Tq[veh/h]
1500
D19
0
06:45 06:50 06:55 07:00 07:05 07:10 06:45 06:50 06:55 07:00 07:05 07:10
120 v[km/h] 3000 r q[veh/h] ,
80 «
1500  
40 i
v D18
1 i 1—
/ \ 0^
(a.) 07:05 07:10 06:45 06:50 06:55 07:00 07:05 07:10
3000 T'
1500  A
D17
0.
07:00 07:05 07:10 06:45 06:50 06:55 07:00 07:05 07:10
3000 T q[veh/h]
V>
1500 "•••
06:45 06:50 06:55 07:00 07:05 07:10 06:45 06:50 06:55 07:00 07:05 07:10
After the jam has been selfformed, it propagates upstream with in average the same velocity
of its downstream front independently of the complexity of states of traffic flow and
independently of whether there is a freeway bottleneck or not (Fig. 8, down arrows).
Theory of Congested Traffic Flow 159
q[veh/h]
v [km/h] 3000
120 L D23
1500 ",".
40  D23off
D23
0
06:45 06:50 06:55 07:00 07:05 07:10 06:45 06:50 06:55 07:00 07:05 07:10
v [km/h] q[veh/h]
3000 T
120 \
80 • ~ 1500
(b) 40 "
D22
0
05:45 06:50 06:55 07:00 07:05 07:10 06:45 06:50 06:55 07:00 07:05 07:10
v[km/h] q[veh/h]
3000 T
120 T
2000
Fig. 6(a, b). Results of experimental observations of the phase transition 'Synchronized Flow
=> Jam' outside highway bottlenecks: The dependence of the vehicle speed (left) and the flow
rate (right) at the different detectors both upstream (a) and downstream (b) from the location
of the phase transition (D20). Arrows 'down' symbolically show the location of the jam at the
different detectors. Arrows 'up' symbolically show the transitions 'Free Flow =>
Synchronized Flow' at the detectors D16 and D15 in the vicinity of the freeway bottleneck in
the intersection 12 (Fig. 4).
q[veh/h] q[veh/h]
3000 i 3000 n
1500 1500 
D18
0
100 P[veh/km] 100
Fig. 7. Pinch effect in synchronized flow: The concatenation of states of free flow (black
quadrates), synchronized flow (circles) and the line J at D18 (a) and D20 (b). Synchronized
flow at D20 is related to higher densities than at D18.
160 Transportation and Traffic Theory
D15
Fig. 8. Propagation of the jam (Fig. 6) through the highway bottleneck and upstream of the
bottleneck: The dependence of the vehicle speed (left) and the flow rate (right) at the different
detectors. Down arrows symbolically show the location of the jam at the different detectors.
Up arrow at D16 symbolically show the phase transition 'Free Flow => Synchronized Flow' at
in the vicinity of the freeway bottleneck in the intersection 12 (Fig. 4).
2.3. Comparison of Phase Transitions 'Free Flow Synchronized Flow' outside and in
the vicinity of Freeway Bottlenecks
It is interesting to make a comparison of the phase transition 'Free Flow => Synchronized
Flow' which occurs outside freeway bottlenecks (Fig. 5) with the related effects where
freeway bottlenecks may play an important role.
First note that the wave of induced transitions 'Free Flow •=> Synchronized Flow' upstream of
the location of the initial local phase transition (t = 06:36, D18, Fig. 5, left) at time t ~ 06:43
reaches the detectors D17 (up arrow, Fig. 5(b), left) and at time t « 06:44 reaches the detectors
Theory of Congested Traffic Flow 161
D16 (up arrows in Figs. 5(b), left). The latter detector is already located in the vicinity of the
bottleneck (onramp, D15on). It could be expected that if the wave of induced transitions
'Free Flow => Synchronized Flow' would propagate further upstream, then synchronized flow
would be caught in the vicinity of this freeway bottleneck, i.e., it would be selfmaintained in
the vicinity of the freeway bottleneck for a long time. Such the 'catcheffect', as it follows
from experimental observations which have been made during other days, can be really
observed on a highway. However, in the case under consideration it does not occur. On the
contrary, the reverse phase transition 'Synchronized Flow => Free Flow' is realized at the
detectors D16 (up dotted arrow in Fig. 6(a), left). Therefore, the synchronized flow which has
spontaneously occurred outside bottlenecks (t = 06:36, D18, Fig. 5, left) has later been also
localized outside bottlenecks during the whole time of its existence.
The free flow which has occurred due to the mentioned reverse phase transition
'Synchronized Flow => Free Flow' has existed at the detectors D16 only during about 5 min.
After this time, the 'usual' phase transition 'Free Flow => Synchronized Flow' in the vicinity
of a freeway bottleneck in the intersection 12 at t « 06:57 occurs (up solid arrow in Fig. 6(a),
left). It should be noted that at a distance of about 12,6 km from the intersection 12 at time t «
06:37 another local phase transition 'Free Flow => Synchronized Flow' in a freeway
bottleneck which is situated in the vicinity of the latter intersection has occurred (up arrow at
the detectors D6 in Fig. 9, left). Both phase transitions 'Free Flow => Synchronized Flow' (up
solid arrow in Fig. 6(a), left and up arrow at the detectors D6 in Fig. 9, left) show qualitatively
the same peculiarities of first order phase transitions which have been considered in (Kerner
and Reborn, 1997).
'[km/h]
06:35 06:40 06:45 06:50 06:55 07:00 06:35 06:40 06:45 06:50 06:55 07:00
v [km/hi qfveh/h]
120 T 3000 T;
D6 D6
06:35 06:40 06:45 06:50 06:55 07:00 06:35 06:40 06:45 06:50 06:55 07:00
06:35 06:40 06:45 06:50 06:55 07:00 5:35 06:40 06:45 06:50 06:55 07:00
Fig. 9. The phase transition 'Free Flow => Synchronized Flow' in the bottleneck in the
intersection II: The dependence of the vehicle speed (left) and the flow rate (right) at the
different detectors both downstream and upstream from the location of the phase transition
(D6). Up arrows show the transitions to synchronized flow at the related detectors.
162 Transportation and Traffic Theory
Thus, there are at least two types of transitions 'Free Flow => Synchronized Flow' in
experimentally observed traffic flow:
(i) The local phase transition outside bottlenecks which is an example of 'self
organization without bottlenecks' in traffic flow (up arrow, D18, Fig. 5, left).
(ii) The local phase transition in the vicinity of a freeway bottleneck (the bottleneck
located in the intersection 12, up solid arrow at D16, Fig. 6, left and the bottleneck
located in the intersection II, up arrow at D6, Fig. 9, left).
The first difference between these phase transitions is the different reasons of their
occurrence: The phase transition (i) occurs without obvious reason due to the growth of a
local perturbation outside bottlenecks and the transition (ii) occurs due to the growth of a local
perturbation in the vicinity of a bottleneck. The second difference is that in the cases of the
phase transition (ii) synchronized flow can be selfmaintained in the vicinity of the bottleneck
for several hours after the transition has occurred. On the contrary, in the case (i)
synchronized flow usually exists for a relatively short time interval and it may propagate both
upstream and downstream from the location where the phase transition has initially occurred.
;h/h] D18leftlane
feh/h]J ^,, , . , q[veh/h]
D16leftlane 3^ •left lane
iVljirfJW
V^ o I i 1
0 20 P [veh/km] Q 20 P [ven/km] 0
Fig. 10. The transitions 'Free Flow => Synchronized Flow' in the flowdensity plane: The
local first order phase transition outside bottlenecks (a), the local first order phase transitions
in the freeway bottlenecks in the intersection 12 (b) and in the intersection II (c), respectively.
Free flow is shown by black quadrates and synchronized flow by circles. Arrows symbolically
show the transitions at the related detectors (Figs. 5  8).
However, these two types of transitions 'Free Flow => Synchronized Flow' have important
common properties: In the flowdensity plane all of them show qualitatively the same
breakdown effect in free flow (Fig. 10) which is wellknown from numerous observations of
freeway bottlenecks (e.g., AgyemangDuah and Hall, 1991; Brannolte, 1991; Cassidy and
Bertini, 1998; Persaud, et al, 1998). However, one could see from the consideration made
above that the breakdown effect in free traffic is not exclusively a property of freeway
bottlenecks: It can spontaneously occur outside any bottlenecks in free flow (up arrow at D18,
Fig. 5). The latter is linked to the fact that these breakdown phenomena have the same nature
 they are the local first order phase transitions 'Free Flow => Synchronized Flow' (Kerner
andRehborn, 1997).
transitions in traffic flow may really be explained in the frame of the selforganization
phenomena, in particular as considered in (Kerner and Rehborn, 1997; Kerner, 1998c).
On the one hand, the observations confirm that 'selforganization without bottlenecks' can
really occur. On the other hand, from these observations it follows that some features of this
'selforganization without bottlenecks' and of selforganization phenomena which occur in
the vicinity of a freeway bottleneck are qualitatively the same. Therefore, some intrinsic non
linear properties of ' homogeneous traffic flow, which do not depend on whether bottlenecks
exist on a highway or not, are essentially responsible for these effects of selforganization. A
theory of these intrinsic nonlinear properties of homogeneous traffic flow has recently been
developed in (Kerner, 1998a, 1998c, 1999). Some of the hypotheses of this theory which may
explain the observed phenomena of selforganizations presented in this paper and in (Kemer
and Rehborn, 1996a, 1996b, 1997; Kerner, 1998c) will be explained below shortly.
Experimental observations show that real traffic flow on a highway is usually non
homogeneous. However, to understand features of real traffic flow, the properties of
homogeneous states (steady states) of traffic flow should first be understood. The hypotheses
presented below are devoted to these steady states of flow.
To explain the hypotheses (i)(iii), note that in synchronized flow spacing between vehicles is
relatively low (i.e., the density is relatively high) in comparison with free flow at the same
flow rate. At low spacing a driver is able to recognize a change in the spacing to the vehicle in
front of him, even if the difference in speed is negligible. In other words, the driver is able to
maintain a timeindependent spacing (without taking fluctuations into account) to the vehicle
in front of him in an initially homogeneous state of synchronized flow. The ability of drivers
to maintain a timeindependent spacing should be valid for a finite range of spacing.
Therefore, a given vehicle speed may be related to an infinite multitude of homogeneous
states with different densities in a limited range (p^^ < p < p^"' in Fig. 11 (a)). For this
reason, the multitude of homogeneous states of synchronized flow covers a twodimensional
region in the flowdensity plane (hatched region in Figs. 3(a), 11 (a)). States of free flow
(curve F in Fig. 11 (a)) and states of synchronized flow (hatched region) overlap in densities.
However, in free flow on a multilane road due to the possibility to pass the average vehicle
speed can be higher than the maximum speed in synchronized flow at the same density.
Therefore, there is a gap in the flow rate between states of free flow and synchronized flow at
a given density.
Small enough perturbations in spacing which are linked to an original fluctuation in the
braking of a vehicle do not grow. Indeed, small enough changes in spacing are allowed,
therefore drivers should not immediately react on it. For this reason even after a time delay,
which is due to a finite reaction time of drivers T reac , the drivers upstream should not brake
stronger than drivers in front of them to avoid an accident. As a result, a local perturbation of
traffic variables (flow rate, density, or vehicle speed) of small enough amplitude does not
grow. An occurrence of this perturbation may cause a spatialtemporal transition to another
state of synchronized flow. In other words, local perturbations may cause continuous spatial
temporal transitions between different states of synchronized flow.
Theory of Congested Traffic Flow 165
Note that on a onelane road, independently of the vehicle density, vehicles cannot pass.
Therefore, homogeneous states of flow on the onelane road at higher density are identical to
homogeneous states of synchronized flow on a multilane road (hatched region in Fig. 3(b)).
The nucleation effect which is responsible for the phase transition 'Free
flow => Synchronized flow' is linked to a selfdecrease in the probability of
passing in traffic flow (P), i.e., in the probability that a driver is able to pass.
The selfdecrease in the probability of passing occurs if, owing to some
local perturbation of traffic variables (the vehicle speed or/and the density),
the probability of passing in the related local region of traffic flow has
decreased below some critical value of the probability of passing Pcr (dotted
curve Pcr in Fig. ll(d)). In contrast, the 'nucleation effect' which is
responsible for the jam's formation is linked to the growth of a local critical
perturbation of the traffic variables (the vehicle speed or/and the density).
The local critical perturbation occurs if the amplitude of a local perturbation
exceeds some critical amplitude (curves F(pert) and S(pert) in Fig. 1 l(b)).
Within the density range of free flow where both phase transitions can
occur, the probability of an occurrence of the phase transition 'Free flow =>
Synchronized flow' is considerably higher than of the phase transition 'Free
flow => Jam'. It is linked to the assumption that the amplitude of a local
perturbation of traffic variables (the vehicle speed or/and the density), which
is needed for the occurrence of the critical value of the probability of
passing Pcr (Fig. ll(d), curve Pcr), is considerably lower than the critical
amplitude of a local perturbation (Fig. ll(b), curve F*611*), which is needed
166 Transportation and Traffic Theory
for the jam's formation. The latter assumption is related to the result of
observations (Kerner, 1998c) where it has been found out that the 'double'
phase transitions 'Free flow => Synchronized flow => Jam' occurs
considerably more frequently than the phase transition 'Free flow => Jam'.
Flow rate, q
(c)
D(free)
Pmax nb(syn) max
P Pma*
Density, p Density, p
Fig. 11. Hypotheses about different types of phase transitions in traffic flow: (a, b) 
Explanation of the phase transitions 'Free flow => Jam' and 'Synchronized flow => Jam(s)',
and (c, d)  Explanation of the phase transition 'Free flow => Synchronized flow'. In Fig. a the
concatenation of the line J with homogeneous states of free (curve F) and synchronized flow
(hatched region) is shown. In Fig. b qualitative dependence of the critical amplitude of the
density local perturbation on the density in metastable homogeneous states of free flow (curve
F (pert) ) and of synchronized (curve S^pert\ which is related to a constant vehicle speed) flow
are shown. In Fig. d a qualitative shape of the dependence of the probability of passing P on
the density is shown. Homogeneous states of free (curve F) and synchronized flow (hatched
region) in Figs, a, c are the same as in Fig. 3(a).
(vi) A hypothesis about the nucleation effect which is responsible for the
jam's formation (Kerner, 1998a, c): The line J (line J in Fig. ll(a))
determines the threshold of the jam's existence and excitation. In other
Theory of Congested Traffic Flow 1 67
words, all (an infinite number !) homogeneous states of traffic flow which
are related to the line J in the flowdensity plane are threshold states with
respect to the jam's formation. The line J separates all homogeneous states
of both free and synchronized flow into two qualitatively different classes:
(1) In states which are related to points in the flowdensity plane lying
below (see axes in Fig. 1 l(a)) the line J no jams either can continue to exist
or can be excited, and (2) States which are related to points in the flow
density plane lying on and above the line J are 'metastable' states with
respect to the jam's formation where the related nucleation effect can be
realized. The local perturbations of traffic variables whose amplitude
exceeds some critical amplitude grow and can lead to the jam's formation
(up arrows in Fig. 1 l(b)), otherwise jams do not occur (down arrows in Fig.
ll(b)). These critical local perturbations act as 'nucleation centers' (nuclei)
for the jam's formation in traffic flow.
The critical amplitude of the local perturbations is maximal at the line J and
depends both on the density and on the flow rate above the line J. The
critical amplitude of the local perturbations is considerably lower in
synchronized flow than in free flow (Fig. ll(b)). The lower the vehicle
speed in synchronized flow is the lower the critical amplitude of
perturbations for states of synchronized flow being at the same distance
above the line J in the flowdensity plane.
To explain the hypotheses (iv), (v), note that in free flow of low enough densities a driver
is not hindered to pass, i.e., the related probability of passing is P=l (Fig. ll(d)). The
higher the density in free flow is the lower the probability of passing; however in free flow
up to the limit point p = p^f' tne probability of passing does not decrease drastically
(solid curve PF, Fig. ll(d)). On the contrast to that in synchronized flow of high enough
density drivers are not able to pass at all, i.e., P=0. The lower the density of synchronized
flow is the higher the probability of passing; however in synchronized flow up to the limit
point p = ps the probability of passing cannot increase drastically (solid curve Ps, Fig.
Synchronized and free flows overlap in the density, therefore pj^ > ps . As a result, the
dependence P(p) is Zshaped, i.e., it has a hysteresis loop, and therefore it consists of three
branches: (i) The branch PF for free flow, (ii) the branch Ps for synchronized flow, and (iii)
the branch Pcr which is related to the critical value of the probability of passing (dotted
curve Pcr, Fig. 1 l(d)). The latter means that if in a local region of free flow due to a local
perturbation of traffic variables (the speed or/and the density) the probability of passing is
decreased below the critical value Pcr, then an avalanche selfdecrease in the probability of
passing occurs leading to a selfformation of synchronized flow where P=PS (down arrow
in Fig. ll(d)). If on the contrary in a local region of synchronized flow due to a local
perturbation of traffic variables the probability of passing is increased above the critical
value Pcr, then an avalanche selfincrease in the probability of passing occurs leading to a
self formation of free flow where P=PF (up arrow in Fig. 1 l(d)).
168 Transportation and Traffic Theory
It has been mentioned in the Introduction that above a threshold of any first order phase
transition, the higher the distance from the threshold is the higher the related probability of
the occurrence of the phase transition. In other words, the probability of the phase
transition 'Free flow => Synchronized flow' should be an increasing function of the density
in free flow, and it should be one at the limit point p = p^aexe). Such a behavior of the
probability of the breakdown phenomenon in a freeway bottleneck has recently been
discovered by Persaud, et al (1998). This experimental fact confirms the conclusion made
by Kerner and Rehborn in (1997) that the latter phenomenon is caused by the first order
local phase transition 'Free flow => Synchronized flow'. An explanation of the hypothesis
(vi) has been made in (Kerner, 1998c).
(vii) A hypothesis about the resulting states of traffic where the nucleation
effect responsible for the jam's formation occurs. The resulting state of
traffic flow, where the nucleation effect responsible for the jam's formation
occurs, essentially depends even on small peculiarities of the initial state of
flow. In particular, if the initial state is homogeneous, then jams appear
caused by the nucleation effect. If, on the contrary, spacing between
vehicles in an initial state of flow is essentially nonhomogeneous, then the
hypothesis (viii) which will be considered below may be valid.
The hypotheses (iv)(vii) may explain the results of experimental observations where two
qualitatively different types of first order phase transitions in traffic flow have been
distinguished (Kerner and Rehbom, 1997). Besides, they may explain the following results
which have been discovered in (Kerner, 1998c): (1) 'Double' phase transitions 'Free flow
=> Synchronized flow => Jam(s)' occur considerably more frequently than the phase
transition 'Free flow => Jam'; (2) While 'stopandgo' traffic patterns occur in
synchronized flow, only single jams can appear in free flow.
(ix) A hypothesis about the process of jam propagation (Kerner, 1998c): The
velocity of the downstream front of a wide jam v g does not depend on
whether free flow or synchronized flow is formed in the outflow of the jam.
Note that the existence of three qualitatively different phases of traffic (Kerner and Rehborn,
1996b): (i) free flow, (ii) synchronized flow and (iii) jams indicates that highway capacity
depends on the phase (i, ii, or iii) on which traffic actually is in. The related 'maximal
capacity' for free flow is q™raexe), for synchronized flow it is qj^, and downstream of a wide
jam it is q out (Fig. 11 (a)).
4. CONCLUSIONS
1. Results of experimental observations allow to conclude the following:
• There are at least two phenomena of 'selforganization without bottlenecks' in real traffic
flow: (i) the local first order phase transition 'Free Flow => Synchronized Flow' and (ii)
the local first order phase transition 'Synchronized Flow => Jam'.
• The local phase transition 'Free Flow => Synchronized Flow', which occurs outside
freeway bottlenecks, can cause two waves: (i) a wave of induced transitions 'Free Flow
=> Synchronized Flow' upstream of the initial location of this phase transition and (ii) a
wave of the propagating synchronized flow downstream.
• These waves, in turn, can cause a widening of the region of synchronized flow (i.e., a
widening of congestion) both upstream and downstream.
• The wellknown breakdown phenomenon in a freeway bottleneck (e.g., AgyemangDuah
and Hall, 1991; Brannolte, 1991; Cassidy and Bertini, 1998; Persaud, et al, 1998), whose
nature is linked to the local first order phase transition 'Free Flow => Synchronized Flow'
in the vicinity of a freeway bottleneck (Kerner and Rehborn, 1997), has some differences
and some common features with the local phase transition 'Free Flow => Synchronized
Flow', which occurs outside freeway bottlenecks. A difference is that synchronized flow
can be selfmaintained for several hours in the case of the phase transition in a freeway
bottleneck. On the contrary, in case of the local phase transition 'Free Flow =>
Synchronized Flow', which occurs outside freeway bottlenecks, synchronized flow
usually exist for a relatively short time interval and the region of the location of the
synchronized flow may propagate both upstream and downstream from the location
where the phase transition has initially occurred.
2. A theory of congested traffic flow (Kerner, 1998a, c, 1999) may qualitatively explain the
different local first order phase transitions in traffic flow.
ACKNOWLEDGMENTS
I would like to thank H. Rehborn, S. Valkenberg and M. Aleksic' for their help, the
Autobahnamt Frankfurt for the support in the preparation of the experimental data and the
170 Transportation and Traffic Theory
German Ministry of Education and Research for the financial support within the BMBF
project ,,SANDY".
REFERENCES
AgyemangDuah, K., and F. Hall (1991) Some issues regarding the numerical value of
capacity. In: Proceedings of International Symposium of Highway Capacity (U.
Brannolte, ed.), p.115. A.A. Balkema, Rotterdam.
Bando, M., K. Hasebe, K. Nakanishy, A. Nakayama, A. Shibata, and Y. Sugiyama (1995).
Phenomenological study of dynamical model of traffic flow. Phys. I (France) 6,
13891399.
Barlovic, R., L. Santen, A. Schadschneider, and M. Schreckenberg (1998). Metastable states
in cellular automata for traffic flow. Eur. Phys. J. B, 5, 793800.
Cassidy, M.J., and R.L. Bertini. (1998). Some Traffic Features at Freeway Bottlenecks. Trans.
Res. B (in press).
Ceder, A. (1976). A deterministic flow model for tworegime approach. Trans. Res. Rec. 567,
1630.
Chandler, R. E., R. Herman, and E. W. Montroll (1958). Traffic dynamics: Studies in car
following. Oper. Res. 6, 165184.
Daganzo, C. F. (1997). Fundamentals of Transportation and Traffic Operations. Elsevier
Science Inc., New York.
Daganzo, C. F., M.J. Cassidy and R.L. Bertini (1998). Causes of Phase Transitions in
Highway Traffic. Trans. Res. B (in press).
Hall, F. L. (1987). An interpretation of speedflow concentration relationships using
catastrophe theory. Transp. Res. A, 21, 191201.
Hall, F.L., B.L. Allen and M.A. Gunter. (1986). Empirical analysis of freeway flowdensity
relationships. Transp. Res. A, 20, 197210.
Helbing, D. (1997). Verkehrsdynamik. Springer, Berlin.
Herrmann, M., and B.S. Kerner (1998). Local cluster effect in different traffic flow models.
Physica A, 255, \63\SS.
Kerner, B.S. (1998a) A Theory of Congested Traffic Flow. In: Proceedings of 3rd
International Symposium on Highway Capacity (R. Rysgaard, ed.), Vol. 2, pp. 621
642. Road Directorate, Ministry of Transport  Denmark.
Kerner, B. S. (1998b). Traffic flow: Experiment and Theory. In: Traffic and Granular Flow
97 (M. Schreckenberg and D. E. Wolf, eds.), pp. 239268. Springer, Singapore.
Kerner, B. S. (1998c). Experimental features of selforganization in traffic flow. Phys. Rev.
Letters, 81, 37973800.
Kemer, B.S. (1999). Congested traffic flow: Observations and theory. Preprint No. 990106,
TRB, 78th Annual Meeting, January 1014, Washington D.C.
Kerner, B. S., S. L. Klenov, and P. Konhauser. (1997). Asymptotic theory of traffic jams.
Phys. Rev. E, 56, 42004216.
Kerner, B. S. and P. Konhauser. (1994). Structure and parameters of clusters in traffic flow.
Phys. Rev. E, 50, 5483.
Theory of Congested Traffic Flow 171
Kerner, B. S. and V.V. Osipov. (1994). Autosolitons: A New Approach to Problems of Self
Organization and Turbulence. Kluwer Academic Publishers, Dordrecht, Boston,
London.
Kerner, B. S. and H. Rehborn. (1996a). Experimental features and characteristics of traffic
jams. Phys. Rev. E, 53, R4275R4278.
Kerner, B. S. and H. Rehborn. (1996b). Experimental properties of complexity in traffic flow.
Phys. Rev. E, 53, R1297R1300.
Kerner, B. S. and H. Rehborn. (1997). Experimental properties of phase transitions in traffic
flow. Phys. Rev. Letters, 79, 40304033.
Kerner, B. S. and H. Rehborn. (1998). Messungen des Verkehrsflusses: Charakteristische
Eigenschaften von Staus auf Autobahnen. Internationales Verkehrswesen 50, 5/98,
196203.
Komentani, E., and T. Sasaki (1958). On stability of traffic flow. J. Oper. Res. (Japan) 2, 11
26.
Koshi, M., M. Iwasaki and I. Ohkura. (1983). Overview on vehicular flow characteristics. In:
Transportation and Traffic Theory (V. F. Hurdle, E. Hauer and G. N. Stewart, eds.),
pp. 403426. Proceedings of 8th International Symposium on Transportation and
Traffic Theory, University of Toronto Press, Toronto..
KrauB, S., P. Wagner, and C. Gawron, (1997). Metastable states in a microscopic model of
traffic flow. Phys. Rev. E, 55, 5597.
Ku'hne, R. (1991). Traffic patterns in unstable traffic flow on freeway. In: Highway Capacity
and Level of Services (U. Brannolte, ed.), pp. 211223. A. A. Balkema, Rotterdam.
May, A.D. (1990). Traffic Flow Fundamentals. PrenticeHall, Englewood Cliffs, New York.
Nicolis, G., and I. Prigogine. (1977). SelfOrganization in Nonequilibrium Systems. Wiley,
New York.
Persaud, B., S. Yagar and R. Brownlee. (1998). Exploration of the Breakdown Phenomenon
in Freeway Traffic, Tranportation Research Board, Preprints of the 77th Annual
Meeting, Washington, D.C.
Prigogine, I. and R. Herman. (1971). Kinetic Theory of Vehicular Traffic. American Elsevier,
New York.
Schreckenberg, M., A. Schadschneider, K. Nagel, and N. Ito. (1995). Discrete stochastic
models for traffic flow. Phys. Rev. E, 51, 29392949.
Treiterer, J. (1975). Investigation of traffic dynamics by aerial photogrammetry techniques.
Ohio State University, Report No. PB 246 094, Columbus, Ohio.
Whitham, G. B. (1974). Linear and Nonlinear Waves. Wiley, New York.
This page intentionally left blank
173
INTRODUCTION
As seen in the car following model, traditional traffic flow theory assumes that the influence
of a car to the peripheral cars is onedirectional. More or less, driving action depends on
the actions of surrounding cars with each other. This traditional approach has given us
useful information to understand the traffic phenomena. This approach is, however, not
necessarily sufficient in describing the traffic behavior when the bidirectional influence
plays a dominant role in driving actions (e.g. Troutbeck, 1995).
Driving behavior between merging and through cars on an onramp merging section of an
expressway is a case of this sort. While merging cars usually avoid unsafe situations by
controlling the timing of their merge, the through car sometimes makes a "giveway" motion
and keeps safer passing by changing its lane to the passing lane next to the merging lane.
The behavior of the merging car under the influence of the through traffic also influences
the through traffic. That is, both merging and through cars affect each other. Since their
influences are not independent from one another, it should be jointly treated in the analysis
as an "interaction". Ignoring this interaction may cause an inaccurate description of the
traffic phenomena on a merging section.
174 Transportation and Traffic Theory
While many studies point out the need to analyze the giveway behavior, few studies have
explicitly dealt with such behavior. Troutbeck (1995) analyzes the phenomena on a round
about where the cars with the right of way running on the circle lane sometimes give their
way to the merging cars at the entrances. Nielsen and Rysgaard (1995) reported merging
and giveway behavior at motorway onramps in four EU countries and compared them with
several indices. These studies are useful data sources for understanding giveway behavior.
There exist many studies of lane changing behavior for traffic capacity analysis (e.g., TRB
(1985), Cassidy et al: (1990), Vermijs (1991)), they merely try to clarify the relationship
between the macroscopic characteristics such as traffic distribution ratio over lanes or lane
changing ratio and road & traffic characteristics. Chang and Cao (1991) model the fre
quency of lane changing from the viewpoint of microscopic behavior analysis. Their study,
however, does not explicitly handle the fundamental mechanism that the traffic behavior
as a whole is formed as the result of interactions of driving decisions among cars.
Under the recognition mentioned above, we developed a simple model for describing the
traffic behavior of a couple of merging and through cars while taking into consideration
the interaction between them explicitly (Kita, 1999). The study views the situation, in
which each of the drivers chooses their best action by considering his/her forecast of the
other driver's action, as a "game" and then, clarifies the mechanism under which the
traffic phenomena such as the merging and/or giveway ratio over lanes are determined
by the driving environments consisting of road and traffic characteristics and the driving
decisions of other surrounding cars.
However, the study does not necessarily investigate the existence of multiple equilibria and
the transition of equilibria, so that the correspondence between the driving conditions as
input and the chosen driving actions as output is not unique. This means that it is difficult
to examine the replication capability of the model by using observed data.
In this study, we show that a unique correspondence can be found between a certain driv
ing environment and a pair of the resultant driving actions of the drivers as an equilibrium
solution of the game, and the possibility to examine the replication ability of the model
based on observation data. For this purpose, we refine the earlier version model exten
sively in order to specify the timing and location of their driving actions corresponding to
the initial conditions, by analyzing the equilibria of the game especially their transition
along time. According to these results, a test of the model using observation data will be
implemented in a practice case study.
A MergingGiveway Behavior Model 175
74
Here, we shall analyze a give way motion often seen in the downflow section from the
merging gore by the through car encountering a merging car and avoiding facing conflict.
The merging section and the cars concerned in this study are depicted in Fig. 1. The
speed of the merging cars is assumed to be slower than that of the through cars. Under
this situation, if the merging car (Car [1] in Fig. 1) decides to merge behind the through
car (Car [2] in Fig. 1), the through car may not give its way. If the through car wishes to
give its way to the merging car, the merging car may merge in front of the through car.
In this way, both the merging car and the through car attempt to take the best action for
themselves by forecasting the other's action, respectively. To describe this situation as a
game, we specify the structure of the game as follows.
1. A merging car (Car [1] in Fig. 1, also named 'Merging Car' in this study) and
the closest through car approaching from the rear side in the adjacent lane to the
acceleration lane (Car [2] and named 'Through Car') are the only players of the game.
The other cars depicted in Fig. 1 such as Car [3] (which is also named 'Passing Car')
and Car [4] (named 'Following Car') constitute the driving environment of Car [1]
and [2].
2. The number of plays between two players over a merging section is one.
Merging
Car
Pass F01, Goi F
°°' G°°
U = (1)
V = (2)
Since driver behavior can be well described by the model provided in "Time to Collision"
(Kita, 1993), assume that each payoff is determined solely by the (TTC) function fa,
"Time to Collision (TTC)", t 1? £ 2 , *3, and t4. These are:
FQQ =
Gn =
Goo = 0 (3)
A MergingGiveway Behavior Model 177
where t\, t 2 , is, and £ 4 , represent TTCs between Car [1] and the end of the merging lane,
Car [1] and Car [2], Car [2] and Car [3], and Car [1] and Car [4], respectively, and are denned
asii = X/vi, t2 = y2/(v$vi), t3 = (y3 yz)/(v$  v$), and U = 2/4/^2 y i)> where u1} uf,
1)3, and u(= ^2)1 are the velocities of the car [1], [2], [3], and [4], respectively. Functions
fij and §ij determine the payoffs of the Merging Car and Through Car, respectively, when
the Merging Car and Through Car choose the strategies i(= 0 or 1) and j(= 0 or 1).
Equilibria
The game has the following eight possible Nash equilibria, (z*, y*). (For details about how
to obtain Nash equilibrium, refer, for example, to Rasmusen (1994)).
I. (z*, y*) = (1,1) when Fn  F01 > 0, Fw  F00 > 0, Gn  GIO > 0
III. (z*, y*) = (1, 0), (0,0), or (F, G) when Fn  Fol > 0, F10  F00 < 0, Gn  G10 > 0
IV. (z*, y*)  (0, 0) when Fn  F01 > 0, F10  F00 < 0, Gn  G10 < 0
V. (z*, y*) = (G, F) when Fn  F01 < 0, F10  F00 > 0, G u  G10 > 0
VI. (z*, y*} = (1,0) when Fu  F01 < 0, F10  F00 > 0, Gn  G10 < 0
VII. (z*, y*) = (0,0) when Fu  F0i < 0, F10  F00 < 0, Gn  G10 > 0
VIII. (z*, y*) = (0,0) when Fu  F01 < 0, F10  F00 < 0, G u  G10 < 0
All the cases above have unique equilibrium except for Case III which has multiple (three
possible) equilibria of (1,0), (0,0), and the mixed strategy equilibrium.
Case I and its equilibrium show that when the merging car has higher risk at the end of the
merging lane and the through car recognizes the low risk on the passing lane, then (Merge,
Go with Giveway) occurs. Here, (, •) indicates ('behavior of merging car (Merging Car)',
'behavior of the through car (Through Car)'). Case II and its equilibrium indicate that
when the situation of the merging car is same as in Case I and the through car recognizes the
higher risk on the passing lane, then (Merge, Go without Giveway) may be realized. Case
III and its equilibria mean that when the two cars are very close in TTC, either (Merge,
Go with Giveway), (Pass, Go without Giveway), or their mixedstrategy equilibrium are
178 Transportation and Traffic Theory
(vn) (0,0)
1
2 " (0,0)
rr \ .. ,
(0,0) n\H , ' (In ) n.n
o t4 (G,F)
3
3
OQ
cs 4. i
O
"•h
(E) (I)
change of \.^ or U
realized. Case IV and its equilibrium reveal that when the situation of the through car is
same as in Case II and the merging car recognizes higher risk of merging, then (Pass, Go
without Giveway) is realized. Case V and its equilibrium show that when the merging car
still has a longer merging lane remaining and feels more risk from the following car than
the through car, and the through car recognizes a low risk on the passing lane, then the
mixed strategy equilibrium (in which the merging car sometimes chooses 'Merge' and the
through car sometimes chooses 'Go with Giveway') occurs. Case VI and its equilibrium
indicate that when the situation of the merging car is same as in V while the through car
recognizes a higher risk on the passing lane, then (Merge, Go without Giveway) occurs.
Case VII and its equilibrium show that when the merging car is close to the following
car and still has a longer merging lane remaining and the through car recognizes a low
risk on the passing lane, then (Pass, Go without Giveway) is realized. Finally, Case VIII
and its equilibrium mean the situation of the merging car is the same as in Case VII, and
the through car recognizes a higher risk on the passing lane, then (Merge, Go without
Giveway) may be realized.
Analyses of Equilibrium
The transition among the equilibria according to the change of the environmental parame
ters, TTCs, are analyzed in this section. Fig. 3 shows the possible transitions of equilibria
due to the change of (decrease in) TTCs.
A MergingGiveway Behavior Model 179
At the equilibria I and II the merging car chooses Merge and therefore the merginggiveway
game ends. At IV, because the merging car does not choose Merge, the game does not
end. When the initial environmental situation is at IV, the game ends when the situation
moves to either II or I via III. The transition of the equilibrium from IV to II means the
change of inequality of FIQ — F0o < 0 into > 0. This change is brought about either by the
shortening of the merging lane length remaining for the merging car or brought about by
decreasing the TTC to that of the following car [3] for the merging car [1]. This transition
of the equilibrium shows that due to the increase in both the risk of staying in the merging
lane and the risk of merging to the next gap, the merge to the first gap by the merging car
becomes the outcome.
Next consider the case where the equilibrium moves from IV to I via III. The equilibrium
transition from IV to III occurs when the inequality GH — G\Q < 0 is reversed. The
merging car and through car become closer (t^ becomes smaller) and the through car,
recognizing the higher risk against the merging car, gives its way to it, resulting in III. In
III there exist three equilibria of (0.0), (1,1), and (G,F). We assume that the realizing
equilibrium is (0, 0) in III by assuming path dependent decisions by the players; before
the transition from IV to III the merging car and through car has chosen 'Pass' and 'Go
without Giveway', respectively, at the previous equilibrium of IV. The transition from III
to I occurs when the inequality F10 — FQQ < 0 is reversed by the change of the environment,
similar to the transition from IV to II. This change of environment is brought about either
by the shortening of the merging lane remaining for the merging car [1] or by the decrease
in the TTC to the following car (£3) and therefore it means that the merging car chooses
Merge due to the increase in risks of merging lane and also due to the next gap to merge.
Next, consider when the game starts with III where there exists three equilibria. Without
any additional information we cannot designate one equilibrium to realize out of three.
By considering the 'pregame' situation, however, the equilibrium selection can be done.
The game has been set to begin when the merging car appearing at the merging section.
This means that at 'pregame situation' the initial situation before the strategy choices
by both drivers are 'Pass' and 'Go without Giveway' and therefore (0,0). Employing the
path dependence assumption again, we can designate (0, 0) among the three equilibria as
the equilibrium to realize outcome III.
180 Transportation and Traffic Theory
Here the relationships between the payoffs and the TTCs given in (3) are specified so that
the payoffs are given by the corresponding TTCs themselves. The payoffs are specified by
TTCs as follows.
„, +  i^
VtL 
X vit
_
TOO —
Gin —
Goo = 0 (4)
where t is the length of time measured from when the merging car first appears at the
merging lane, 7/2 > 2/3, and 2/4 are the distances of the car [2], [3], and [4], respectively, from
the merging nose when t = 0, and x is the length of the merging lane.
By substituting the specification of the payoff functions in (4) into the three inequality
conditions, Fn  Fol > (<)0, FIO  FQO > (<)0, and Gn  G10 > (<)0, they can be
rewritten as follows.
< V2  Vi < Vi
Fw
Gn — GIO . 0
< V 2 — V i > V3 — V i
A MergingGiveway Behavior Model 181
Notice that in these three conditions that determine the equilibrium realized they do not
include the time variable t and therefore they are not affected by the progress of time. In
other words, when assuming payoffs consisting of TTC themselves the equilibrium transi
tions never appear. The merging in addition to the giveway times and locations realized are
therefore determined right away by the location of all cars when the merging car appears
at the merging nose.
Inequality Fn — F0i > (<)0: The sign of this inequality is determined by the TTC of
the merging car [1] and the following car [4], and the TTC of the merging car [1] and
the end of the merging lane. This corresponds to the comparison of the sizes of t\
and £3, indicating whether the merging car [1] will be passed by the following car [4]
before reaching the end of the merging lane.
Inequality Fw — F00 > (<)0: The sign of this inequality is determined by the TTC of
the merging car [1] and the end of the merging lane, and the TTC of the merging
car [1] and the through car [2]. This corresponds to the comparison of the sizes of t\
and t2, indicating whether the merging car [1] will be passed by the through car [2]
before reaching the end of the merging lane.
Inequality G\\ — G\Q > (<)0: The sign of this inequality is determined by the TTC of
the merging car [1] and the through car [2], and the TTC of the merging car [1] and
the passing car [3]. This corresponds to the comparison of the sizes of t2 and £3 + ^4,
indicating whether the through car [2] will be passed by the passing car [3] before
catching the merging car [1].
The game is constructed to capture the merginggiveway situation 'at a moment' of the
merging section. Accordingly, the equilibria include the ones in which the merging car
decides not to merge to the central lane (or chooses 'Pass'; y* = 0). We are interested in
analyzing not the drivers' behavior of the moment but the traffic phenomena that occur
at the merging section. The merging car should merge to the central lane at some point
in time; otherwise it stacks at the end of the merging lane. Using the game analyses given
above, we should examine when and where the merging car merges to the central lane and
also when and where the through car chooses to giveway.
182 Transportation and Traffic Theory
( yz N
IV (0,0) +  + (1,0)* \v%—vi ' '
V (G,F)  +  —
VI (1,0)  + + —
VII (0,0)    (i,i)* (M Jte_)
^v* — vi ' v^—vi '
VIII (0,0)   + (1,0)* (M
v^f=^7' >)
* Indicates the results different from the game equilibrium
According to the interpretation of the three equilibrium conditions of inequality given in the
previous section, the game equilibria can be interpreted as the traffic phenomena (merge
and giveway behaviors and their timings) under the payoff function specification by the
TTCs themselves given above. They are summarized in the sixth and seventh columns in
Table 1. In Table 1, (x,y) means whether the merging car [1] ever merges or not (x = I
when merging) and whether the through car [2] changes the lane and chooses giveway
behavior or not (y = 1 when giveway), respectively. Notice that "—" in Table 1 means the
corresponding behavior cannot occur because of the inconsistent conditions exist. tm and
tg indicate the time that the merging car [1] merges to the central lane and the time that
the through car [2] changes to the far right lane (chooses 'Go with Giveway'), respectively.
The traffic phenomenon to realize in the cases that correspond to equilibria I and II are
exactly same as the equilibrium strategy ((x*,y*) = (x, y)). In equilibrium I, the merging
car merges to the central lane immediately and the through car goes with giveway im
mediately (tm = tg = 0). In equilibrium II, the merging car merges to the central lane
immediately, and the passing car [3] stays at the central lane.
Other cases, Case III to VIII, have a different traffic phenomena from the game equilibria
(compare the entries in the second and sixth columns in table 1). The reason and the
interpretation of these cases are as follows.
In equilibrium III, The merging car [1], which will be passed by the through car [2] before
the end of the merging lane but not by the following car [4], merges to the central lane
right after being passed by the through car (into the gap between the through car [1] and
A MergingGiveway Behavior Model 183
the following car [4]). Also, the through car [2], faced with risk of catching the merging
car [1], chooses to give its way to the merging car and changes the lane to the passing lane
because it is never caught by the passing car [3]. Consequently, the traffic phenomenon in
this case will be (Merge, Go with Giveway), and the times of merging and giveway that
occur simultaneously are given by tm=tg = yij(v\ — t'i).
In equilibrium IV, the situation and the behavior of the merging car [1] are same as in the
equilibrium III above. The through car [2], which is under risk of catching the merging car
[1] before the end of the merging lane, cannot change to the passing lane because it will
be caught by the passing car [3]. Consequently, the corresponding traffic phenomenon is
(Merge, Go without Giveway). The merging time is given by tm = yil{v\ — v\] because
the merging car merges to the central lane just after being passed by the through car, and
the giveway time does not exist because no giveway behavior exists.
Under the specification of the payoff in this study, the two conditions, Fu — F0i < 0 and
FW — Foo > 0, constituting the equilibrium V and VI, imply that the following car [4] is
running ahead of the through car [2]. Consequently, these conditions cannot hold and the
traffic phenomenon corresponding to these two equilibria do not exist.
In equilibrium VII, the merging car [1] is passed by the through car [2] and also by the
following car [4] before reaching the end of the merging lane. It merges to the central lane
just after being passed by the following car [4]. The through car [2], which is assumed to
always run at the front of the following car [4], already passed the merging car when it
merges to that lane. Also, the through car [2] which is never being caught by the passing
car [3], changes the lane to the passing one. Consequently, the traffic phenomenon to
realize is given by (Merge, Go with Giveway). The corresponding merging and giveway
times, that occur simultaneously, are given by tm = tg = y$/(v\ — v\).
Finally, In equilibrium VIII, The situation and the behavior of the merging car [1] are
exactly same as the ones given in the equilibrium VII above, and it merges to the central
lane right after being passed by the following car [4]. The through car [2], which will
be caught by the passing car [3] before the merging car reaches the end of the merging
lane, does not change the lane to the passing one. Consequently, the realizing traffic
phenomenon is (Merge, Go without Giveway). The corresponding giveway time is given
as tm = ?/4/(f — fi) while the giveway time does not exist.
Consequently, the merging car merges to the central lane immediately when equilibrium
is I or II, merges after being passed by the through car [2] when III and IV, and merges
184 Transportation and Traffic Theory
erge i m m e d i a t e l y
erge behind Car [2]
erge behind Car [4]
0 i 2
TTC of Car [1] and Car [4] TTC of Car [1] and Car [2]
a) TTCs of cars and the merging behavior b)TTCs of cars and the giveway behavior
after being passed by the following car [4] when VII or VIII. On the other hand, the
through car changes to the far right lane when the equilibrium is I, III, or VII, and does
not change the lane for II, IV, or VIII. These relationships between the equilibria and the
resulting traffic behaviors are depicted in Fig. 4 on the TTC planes of (£2,^4) and (£2^3)5
by using the three inequalities governing the realizing equilibrium. Fig. 4 a) depicts the
merging behavior, and the horizontal and vertical lines that distinguish the equilibrium
areas indicate £2 = X/v\ and U = X/vi, respectively. Fig. 4 b) shows the equilibrium
area for the giveway one, and the line that distinguish the two equilibrium area is given
by £3 = (t>2 —
A CASE STUDY
To examine the proposed model, a case study is carried out by using a set of videorecorded
data observed at an onramp merging section of Nagoya I.C., Tohmei Expressway, in
Nagoya (data source is Research Group on Design of Intersections (1987)). The number
of observation samples of the merging behavior was 74. Among these, observations that
have complete sets of explanatory variables were used: there are 10 such complete sets of
data in the observation data set. The merging lane has 200m in length, and the average
velocities of the vehicles were assumed to be constant and estimated as v\ =82.8km/h for
the merging car, v\ = v\ =85.8km/h for the through and following cars, and v3 =95.7km/h
A MergingGiveway Behavior Model 185
Results
The data is plotted on Fig. 5 a) and Fig. 5 b). The solid dots and circles indicate data
that "Car [1] merges behind Car [2]" and "Car [1] merges immediately", respectively, in
Fig. 5 a), and "Car [3] goes with giveway" and "Car [2] goes without giveway" in Fig. 5
b), respectively.
Due to the velocities estimated, the lines, t^ — x/v\ and £4 = x/Vi, that distinguish the
equilibrium areas in Fig. 4 a) are both 8.696 sec. The plots of the observed data ('Merge
immediately' and 'Merge behind Car [2]') are consistent with the designated areas of the
game equilibrium, while no data that a merging car 'merges to the central lane after being
passed by the following car' or 'merge behind Car [4]' is available.
Next consider Fig. 5 b). According to the velocities of the car set, the slope that distin
guishes the two areas is (v\ — v\)/(v^ — v\] = 0.36. The nine data plots for the cars that did
not do giveway are consistent with the area designated by the game equilibrium analyses,
and that is also true for the one observation plot in which the car went without giveway.
Though the sample size is not large enough to fully support the validity of the model,
through this data references to the proposed model can be recognized, for instance it gives
a good description of the traffic behavior of giveway and merging in onramp merging
sections. This result verifies the ability of this new approach to understand the traffic
phenomena with interactions.
CONCLUSION
In this study, the traffic behaviors such as merging and giveway are expressed as game
strategy, and the relationships between the traffic phenomena to realize and the driving
environment surrounding the drivers are clarified. By understanding the change of driving
behaviors (and therefore the traffic phenomena) as the transition from one game equi
librium to another, the mechanism under which the drivers' behaviors and the resulting
phenomenon occur and change according to time progress and the change of driving en
vironment. Furthermore, by specifying the payoff functions in the merginggiveway game
186 Transportation and Traffic Theory
40
20
t3=0.36t2
0 8.696 20 60 20 40 t2 (sec)
t4 (sec)
a) TTCs of cars and merging behavior b) TTCs of cars and the giveway behavior
model, the conditions under which the merging and giveway behaviors emerge are explicitly
induced and the merging and the giveway timing and locations are specified. With this
model development a unique correspondence is found between a certain driving environ
ment and a pair of the resultant driving actions of the drivers as an equilibrium solution
of the game, and presents the possibility to examine the replication ability of the model
based on observation data. Finally, the validity and applicability of the proposed model
are checked by applying the observation data.
By employing the approach and models employed in this study, the occurrence and change
(location and timing) of driver's behaviors such as giveway and merging motions, and
the traffic phenomena resulting from the combination of such behaviors can be specified.
Accordingly, application of this study contributes to the construction of effective traffic
control systems and the design of safer and more comfortable intersections of expressways,
explicitly considering the car motions.
The game theoretic approach to the traffic behavior modeling has very little research
accumulation and several important aspect remain uninvestigated. Among others, most
importantly, the development of a new payoff estimation technique is necessary. With this,
other important progresses will be also possible: nonlinear payoff function applications
which enables us to explain the merging behavior at the middle of the gap, and game
modeling with incomplete information. Formulation to translate this microscopic model
into macroscopic results should be also developed.
A MergingGiveway Behavior Model 18 7
These various extensions of this study that have not been considered in this study are all
important; the existence of many possible future directions of this study indicates that the
model and approach can be easily extended to become more robust.
ACKNOWLEDGEMENT
The authors would like to express their appreciation to the anonymous referees for their
valuable comments. This study was supported financially in part by the GrantinAid for
Scientific Research, Ministry of Education, and Sumitomo Marine Welfare Foundation.
REFERENCES
Cassidy, M. J. et al. (1990). A proposed analytical technique for the design and analysis
of major freeway weaving sections, Inst. of Transp. Studies, Univ. of California at
Berkeley, UCBITSRR9016.
Chang G.L., and Y.M. Cao (1991). An empirical investigation of macroscopic lane chang
ing characteristics on uncongested multilane freeways, Trans. Res., 25A, 6, 375389.
Kita, H. (1993). Effects of merging lane length on the merging behavior at expressway
onramps. In: Transportation and Traffic Theory (C. Daganzo ed.), Elsevier, 3751.
Kita, H. (1999). A merginggiveway interaction model of cars in a merging section: a game
theoretic analysis, Transportation Research, Vol.33A, No.3/4, 305312.
Nielsen, M. A., and Rysgaard, R. (1995). Merging Contra Give Way When Entering A
Motorway. Road Directorate Report, 27, Danish Road Directorate, Copenhagen.
Rasmusen, E. (1994). Games and Information. 2nd edition, Blackwell Publisher, Mas
sachusetts, 6791.
Japan Society of Traffic Engineers (1987). Report on the Design of the Intersections,
Research Group on Design of Intersections, 2, Japan Society of Traffic Engineers (in
Japanese).
Transportation Research Board (1985). Highway Capacity Manual, Special Report, 209.
Troutbeck, R. J. (1995). The capacity of a limited priority merge. Physical Infrastructure
Centre Research Report, 958, Queensland University of Technology, Australia.
Vermijs, R. G. M. M. (1991). The use of micro simulation for the design of weaving
sections, In: Highway Capacity and Level of Service (U. Brannolte. ed.), 419227,
A. A. Balkema.
This page intentionally left blank
CHAPTER 3
ROAD SAFETY AND PEDESTRIANS
Marian Tracz, Chair of Highway & Traffic Engineering, Cracow University of Technology, ul.
Warszawska 24, 31155 Cracow, POLAND, email: mt@rodes.wil.pk.edu.pl
ABSTRACT
In this paper some measures for effectiveness evaluation of the methods of the identification of
high risk sections on rural roads have been proposed. Two categories of measures have been
taken into account. First concerns the level of accidents' concentration along hazardous road
sections as well as the level of concentration of accidents severity. The second category
concerns the problem of repeatability of the identification procedures during two consecutive
time periods in relation both to the location of dangerous sections along a road and to some
features of road accidents. The measures have been applied to three Polish methods and the real
data have been used to conduct a comparison. Some lacks which were found in the
identification methods in this study have been completed. The results varied for different
aspects and accident data from different roads.
INTRODUCTION
Accident risk at dangerous road sites can be an important factor while identifying and
evaluating them. In practice these dangerous sites can be found in parts of a road network
where accident intensity is comparatively large. So, the studies are usually focused on urban
192 Transportation and Traffic Theory
roads and on the sites that are prone to accidents of various types because of their geometrical
characteristics  as urban junctions.
In the identification and evaluation of dangerous road sections, the majority of highway
authorities in Poland use methods that are based on rather simple criteria such as numbers of
accidents or average density of accidents (ace./km/year) that happened on a road section during
a certain period of time (usually 3 years). In research it is commonly assumed that the number
of accidents at a site over a fixed time period is well modelled by a Poisson distribution. Traffic
volume is not taken into account due to the lack of current traffic data.
When investigating the safety of rural roads a problem arises how to identify dangerous
sections as they can have various lengths and they are not evenly spaced along a road. Polish
methods use different algorithms of dividing a road into sections for further investigations. In
addition, the time periods for which high risk road sites are identified also differ in various
methods. In consequence, the results of accident analysis are frequently not comparable.
In this study some measures have been proposed for evaluation of the effectiveness of the
methods that are oriented towards rural roads. For those who deal with the problem, these
measures should be helpful in comparing results. The paper does not only present a pure
methodological study of the analysed methods but also includes some practical aspects.
There are several methodologies that are used in analysis of traffic accidents and in the
identification of high risk sections on a road (Hauer, 1995). When using different methods,
different distributions and numbers of dangerous sections along a road can be a consequence of
changing a selection criterion. Then this can affect the effectiveness of road safety measures.
All these issues depend to a great extent on the reliability and validity of accident data (Hakkert
and Hauer). Consequently, the results of the identification of dangerous road sections are
influenced by accident analysis methodology, data collection and data handling. Therefore, an
important question arises: which identification procedure should be used to get the best results?
A variety of techniques have been applied to answer the question (Maher and Mountain, 1988;
Mountain at all, 1992a, 1992b). To choose a proper method reaserchers use statistical tools
and indices derived on the basis of probability theory. In the analyses they use real or artificial
(generated by a computer) accident data. In this paper some measures for evaluating and
comparing the performance of various identification methods have been introduced. The
process of comparison was carried out on the basis of real accident data.
Road Safety Evaluation Methods 193
Identification procedures applied to rural roads are usually accomplished on the basis of the
data that come from time periods of different lengths. Such data contain information of various
levels of reference. So the comparion of results of the identification of dangerous road sections
obtained from these data should be processed on the basis of indices that can be easily
examined. For road sections a method can be considered as a more accurate if there are:
• the high proportion of the total number of accidents on dangerous sections in relation to the
total number of accidents on the whole analysed road,
• the low proportion of the total lengths of these sections in relation to the length of the road.
An average accident density (ace. /km/year) is the ratio used very commonly. The greater its
value calculated for hazardous road sections the better the results of an identification
procedure. In extreme situations, the accident density can reach infinity for zero length
sections. There are, however, some disadvantages of using this measure. Comparing the
accuracy of results this measure works quite well for one road but not for several roads.
Average accident density is an absolute measure and can be as well considered as a
suplemantary measure.
In order to consider more intuitive measure, a certain relative ratio has been put forward in this
work. It has been called the Method Inaccuracy Coefficient (MIC) and is defined as follows:
The MIC coefficient expresses the proportion of the fraction of the total length of dangerous
sections in the road length to the fraction of the total number of accidents occurred on these
sections in the total number of accidents on the road. The value of the coefficient can exceed
unity. However, if the value really concerns high risk road sections, it is included in the interval
<0,1>. The closer to zero the value of the coefficient is, the more accurate or precise the
method is and reciprocally.
Coefficients describing accidents severity are defined in the same manner and are used here as
supporting measures. They are dentoed as MIC(I), MIC(F) and MIC(V) and calculated
194 Transportation and Traffic Theory
according to the formula (1) in which the number of accidents is respectively replaced by the
number of injuries (I), fatalities (F) and vehicles (V). The interpretation of the values of these
coefficients is similar to the MIC interpretation.
When considering the process of the identification of hazardous rural road sections, additional
potential impact should be taken into account. This is commonly known as "accident
migration" or "migration of accident risk" and referred to a phenomenon arising after applying
some remedial treatments on a road section or because of an increased percent of time delay
(travelling in platoon). These are, for example, accidents typical for dangerous overtaking
occurring sometimes a few kilometers from a place where the overtaking demand has appeared.
However, during this study it has been noticed that the ends of a high risk road section can
change during two consecutive time periods of the identification process, even if no treatment
had been applied in relation to the "earlier period" section. In this paper, such periodafter
period changing of locations of the hazardous sections along a road can be expressed by a
certain index. This index has been build using "positive" and "negative" differences between
the mutual position of the ends of consecutive (earlier and later) overlapping dangerous road
sections:
(2)
dr = sgn(r, (1)  r, (2)) • \ r, (1)  r (2) \ (3)
where: !,(!)  kilometreage of the left end of the z'th overlapping section obtained from an
identification procedure for the earlier time period; lt(2) is defined likewise for
the later time period,
r,(l)  kilometreage of the right end of the /th overlapping section obtained from an
identification procedure for the earlier time period; rt(2) is defined likewise for
the later time period,
dlj  difference between left ends of the two r'th overlapping road sections; the sign
of the difference informs whether the kilometreage of the end of the later time
period section is lower (,,plus" sign) or higher (,,minus" sign) than the
kilometreage of the respective end of the earlier time period section,
drt  defined similarity to the dlt difference with respect to the right ends.
Road Safety Evaluation Methods 195
The earlier time period, for which identification procedures have been processed, will also be
called the first time period in this paper, and the later time period will be called the second time
period. The scheme of overlapping of two identified dangerous road sections is presented in
Figure 1. In this scheme, the lower section s(l) represents a road section from the first time
period whereas the upper section s(2) represents a road section from the second time period.
The notations s(l) and s(2) also represent the respective sections lengths. The remaining
notations follow the explanation given for the equations (2) and (3). Subscripts were omitted
for simplification.
kilometreage
All possible alternatives of mutual position of the two overlapping dangerous road sections s(l)
and s(2), obtained for the first and the second time periods in accident analysis, are presented in
Figure 2. For each alternative a short description is given, that explains the meaning of signs of
the positive and negative differences (formulae (2) and (3)). As can be seen, there is one perfect
overlapping (alternative IX) among the nine presented cases. There are also a few other
alternatives that can be considered as representing satisfactory space compatibility of
overlapping sections  alternatives III VIII. Only in two cases I and II the space compatibility
seems doubtful. Uncertainty is especially justified when the length of the common part of the
two sections is shorter than 50% of the length of the total section <l(2), r(l)> in the case I or of
the total section <l(l), r(2)> in the case II.
In order to evaluate the degree of repeatability of an identification method  regarding the space
distribution of the identified dangerous sections, the Space Compatibility Index (SCI) was
introduced as follows:
(4)
196 Transportation and Traffic Theory
where: n/  number of dangerous sections on a road identified in the first time period,
ri2  number of dangerous sections on the same road identified in the second time
period,
n  number of dangerous sections on the road identified in the first time period that
have common parts with dangerous sections on the same road identified in the
second time period; i.e. the number of pairs of dangerous sections on the road
identified in two consecutive time periods. The mutual locations of sections
treated as a pair follow one of the schemes peresented in Figure 2. If all
dangerous sections from both time periods follow the schemes then ni=ri2=n,
x,  coefficient that specifies the way of overlapping of the pair of dangerous road
sections. The coefficient has been defined only for those pairs of sections that
have common parts (i=l,...,n) and it distinguishes satisfactory space
compatibility (alternatives IIIIX in Figure 2) and not satisfactory space
compatibility (alternatives III):
In the formula (5) ssli is the total length of the sum of the two overlapping road
sections determined by the very ends of these sections:
Sj(l) is the length of the /th overlapping road section identified in the first time
period.
The idea of the SCI index described by (4) has been taken from the concept of standard
logarithm information function (Nicholson, 1995). The SCI index is a complex measure that
takes into account the number and the mutual positions of dangerous road sections identified
in two consecutive time periods. The minimal value of SCI is equal to 0.3. This value describes
the most satisfactory space compatibility, where «/=«2 = « and xt =1.0 for each i=l,...,n. This
means that all dangerous road sections identified in the first time period were identified as
dangerous in the second time period with the satisfactory periodafterperiod shifts along a road
(schemes IIIIX), and any additional sections were not identified in the second time period.
Road Safety Evaluation Methods 197
s(2)
I) dl>0, dr>0; the s(2) section is shifted to the left
in relation to the s(l) section
s(2)
III) dl>0, dr<0; the s(2) section covers the s(l)
section
s(2)
VI) dl=0, dr<0; the s(2) section covers the s(l)
section, left ends are equal
s(2)
VII) dl>0, dr=0; the s(2) section covers the s(l)
section, right ends are equal
Figure. 2. Schemes of mutual positions of overlapping dangerous road sections obtained from
an identification procedure for two consecutive time periods.
198 Transportation and Traffic Theory
The SCI index evaluates the space repeatability of danagerous road sections in two consecutive
time periods. In order to support this index, some additional measures are suggested in this
chapter to check the repeatability of accident patterns in relation to overlapping road sections.
In this work, the accident patterns are expressed by qualitative features of accidents recorded on
dangerous road sections such as driver's behaviour and accident type. The classification
categories representing these features have to be independent of each other within each of the
features (Tracz and Nowakowska, 1998).
The measures accessing repeatability of accident patterns are build using commonly known
standard error ideas derived from the estimation theory. For each overlapping dangerous
section /, the occurence of classification categories of a chosen accident feature was checked
for two consecutive time periods t=l,2. Then, if a given accident category was recorded for at
least one time period, the percentages of this category occurence for the respective time periods
were calculated according to the formula (6):
The introduction of the percentages FCPy allowed to define three complementary measures of
accident pattern repeatability:
Road Safety Evaluation Methods 1 99
(9)
All notations on the righthand sides of the definitions (7)  (9) are the same as in the measures
described earlier (formulae (1), (4), (6)). The range of the measures of accident pattern
repeatability is the interval <0,1>. The value 0 of the measures MAAD and MAD confirms
perfect repeatability in accident patterns. The closer to unity the values of the MIAD and the
MAD are, the lower the degree of repeatability.
A few methods for the evaluation of traffic safety levels on rural roads have been implemented
in Poland. Some of them are attractive for traffic engineers because of their simplicity. These
are commonly used methods known as the Warsaw method and the Gdansk method (Datka at
all., 1989, 1997). Both are based on the knowledge of accident numbers and average accident
densities. Other methods are less popular due to their requirements regarding accident and
traffic data or due to probability aspects regarding stochastic character of road accidents. One
of such methods has been worked up recently (Tracz and Nowakowska, 1996). Is is based on
cluster analysis and Bayesian theory and,takes into account random character of accidents in
time and space. The three methods were taken as the subject of a survey in this work. For
simplicity they have been named here the WM method, the GM method and the BM method.
200 Transportation and Traffic Theory
A unit road section is used to define the method. It is the onekilometerlong section
determined by a road kilometreage. In the WM method the unit section on which not less than
four road accidents have been recorded during a year is classified as a pecularily dangerous
unit road section whereas the unit section with the number of road accidents equal to two or
three during a year is classified as a dangerous unit road section. In order to characterise the
safety of any road section, the numbers of both types of unit sections on this section are
considered.
However, there is a disadvantage in the definition given above. It has not been expressed how
many pecularily dangerous and dangerous unit road sections should be contained in the
considered road section in order to identify this section as a high risk one. To make the
definition more precise, the performace of the WM method applied to real accident data has
been studied. As a result, the following completion has been made in this work:
At least three unit road sections with the number of accidents not less than 2 and with the
distance between these sections not less than 2 km identify the location of a high risk section on
a road.
The method classifies a road section on the basis of its accident density D (acc./km/year).
A classification criterion uses the accident density of the road Drd that contains the considered
section. The Drd value depends on the administrative division of the Polish roads as it has to
be calculated for the part of the road placed within the administrative border of a province.
According to the GM method a road section can be classified into one of the three categories:
• relatively safe, if Drd <D<=2 Drd,
• being threatened, if 2 • Drd < D <= 3 Drd,
• dangerous, if D > 3 Drd .
The disadvantage of the GM method lies in lack of the information how the ends of the section
should be determined. After the analyses of some possibilities for solving this problem, the
following completion has been made in this work:
The sections where accidents cluster are determined initially using the single linkage method
in the way described in the BM method. Then each section is checked in order to decide to
which category it belongs according to the GM method classification.
Road Safety Evaluation Methods 201
In order to perform the road safety analysis, road sections where accidents cluster are
determined first. This is done by applying the single linkage method  one of the methods of
cluster analysis. To identify hazardous sections among the accident cluster sections, an
accidentproneness model has been developed (Tracz and Nowakowska, 1996). The model
describes the distributions of the following accident variables: number of injuries, number of
fatalities and number of vehicles involved in accidents. As the model is derived from the
Bayesian approach (Benjamin and Cornell, 1977) the identification method has been called a
Bayesian one.
The model is based on two sources of information. The first one h(x) is a prior distribution of
the accident variable X over the accident cluster sections:
for x =0
(10)
for x >1
The second source h(x\z,i) is a posterior distribution of the variable Shaving reference to
those accident cluster sections, where the value of z of the accident variable over the i time
units have been recorded:
for x =0
h(x\z,i) =
h(xl\z,i) for x> 1
x • (i + 1 + s)
In the above formulas s>0 is a scale parameter and k>0 is a shape parameter of the distribution.
It is obvious that the level of safety varies between different roads and different road sections
along a given road. Therefore, to define the change in the value of accident variable X for each
road section, the value of the cumulative probability of the h(x z,i) function, calculated for the
median m of the h(x) distribution, has been proposed. Thus, a single section can be
characterised according to the cumulative posterior probability h (x z, i) of the X value above
the median, i.e. by the weight W(X) expressed as follows:
(12)
202 Transportation and Traffic Theory
In order to include the influence of a road section length / in the identification process, a weight
function for this length has been defined by the following form:
u=
LW(l) max
for
In the BM method the accident cluster sections are subsequently used to find the prior
distributions for injuries, fatalities and total vehicles involved in accidents. Then posterior
distributions are detremined to calculate accident severity weights from the formula (12). These
weights and the road section length weight (13) are the elements of a set S, that characterises
the safety of a section:
The variable X} identifies the number ofyth accident variable as follows: X\ is the number of
injuries, X^ is the number fatalities, Xs is the number of vehicles involved in accidents on a
section.
The identification is processed on the basis of a classification value calculated from the
elements of the set S. This value is called a Safety Weight SW and is defined in the following
form:
A criterion for high risk road section is determined by a critical set S* obtained from the set S of
the elements equal to 0.5, 0.4, 0.6 and 0.75. These numbers reflect the rank of respective
elements of the set  the lower value, the higher rank. An accident cluster section is indicated as
a high risk road section, if its classification value is greater than the value SW for the critical set
S*.
In order to present practical application of the described measures and to compare the methods
of the identification of high risk road sections, real accident data recorded on four selected rural
roads in the period 19911996, from two neighbouring provinces of southcentral Poland were
Road Safety Evaluation Methods 203
investigated. These roads represent different categories and are characterised by various
lengths, traffic volumes and accident densities. High risk road sections were identified for
different time periods according to the identification methods. The studied road sections were
not treated during the two considered periods. The comparison of results was conducted using
calculus evaluation. This enabled to present advantages and disadvantages of the methods.
The main differences in the concentration of accidents and in the concentration of accident
consequences (injuries, fatalities and vehicles involved in accidents) obtained when using the
three methods (WM, GM, BM) are presented in Table 1. The WM method involved oneyear
time period data to process the identification procedure. The results given in Table 1 for this
method were calculated for the year 1991. The other methods (GB and BM) were used for
a threeyear time period. In this case, the results are given for the period 19911993.
Table 1. Comparison of the values of the Method Inaccuracy Coefficient and additional
measures for high risk road sections obtained for different indentification methods.
Accident Method Inaccuracy Coefficient
Time Method density Total length MIC MIC(I) MIC(F) MIC(V)
period [ace. /km/year] [km]
1 2 3 4 5 . . 6 7 8
Road No 7
1991 WM 2.0 83.0 (43%) 0,60(71%) 0.57 (76%) 0.59 (73%) 0.57 (76%)
9193 GM 5.4 17.8(9%) 0,24(38%) 0.24 (39%) 0.23(41%) 0.24 (38%)
9193 BM 3.9 40.0(21%) 0.33 (62%) 0.33 (65%) 0.30 (70%) 0.33 (63%)
Road No 44
1991 WM 2.7 6.0 (7%) 0.38(19%) 0.31(23%) 0.57 (13%) 0.39(18%)
9193 GM 3.8 9.2(11%) /0.23(48%) 0.21 (53%) 0.23 (48%) 0.23 (47%)
9193 BM 3.0 9.9 (12%) 6,30(39%) 0.27 (44%) 0.34 (35%) 0.30 (39%)
Road No 74
1991 WM 2.1 20.0 (24%) 0.51 (48%) 0.53 (46%) 0.45 (54%) 0.59 (42%)
9193 GM 4.5 3.6(4%) 0.20 (23%) 0.20 (23%) 0.20 (22%) 0.21(21%)
9193 BM 2.6 11.2(14%) 0.34 (40%) 0.33 (42%) 0.33(41%) 0.34 (40%)
Road No 728
1991 WM 1.7 6.0 (4%) 0,21(20%) 0.20 (20%) 0.15(27%) 0.22(18%)
9193 GM 2.0 11.6(8%) 017 (45%) 0.17(47%) 0.16(49%) 0.19(43%)
9193 BM 2.0 4.2 (3%) 0.17(17%) 0.14 (20%) 0.14(21%) 0.15(19%)
Accident denisties for dangerous sections are presented in the column 3 of Table 1. Some
additional supporting measures have been included in brackets:
 the percentage in a row of the column 4 represents the ratio of the total length of dangerous
road sections to the length of a road,
204 Transportation and Traffic Theory
 the percentage in a row of the column 5 represents the ratio of the total number of accidents
on dangerous road sections to the total number of accidents on a road,
 the percentages in the columns 6, 7 and 8 are ratios such as in the column 5 but calculated
for injuries, fatalities and number of vehicles involved in accidents.
For a road with a high accident density (road No 7) the WM method determines dangerous
sections on which the ratios of the numbers of accident features exceed 70%. However, the
total length of the sections can be fairly significant  even more than 40% of the road length
(see the percentage in the column 5 for the road No 7). The lower the category of a road, the
shorter the total length of dangerous sections but also the lower the percentage values of
accident features (roads: 44, 74, 728). For all four roads, accident densities on dangerous road
sections identified by the WM method are lower than identified by the GM and BM methods.
Values of the Method Inaccuracy Coefficient are the highest also for all roads. The situation
was similar when using this coefficient for subsequent years (1992, 1993). So, from the point
of view of the concentration of accident features the WM is not advisable.
The accuracy of the GM method proves to be very good for all roads. Accident densities on
high risk road sections obtained from this method have the highest values. The supporting
ratios (percentages in the columns 58) vary from road to road: they are the lowest for the roads
No 7 and No 74, whereas for the roads No 44 and No 728 they are the highest. The MIC
measures are very low: they range from 0.17 to 0.24. Taking into account the Method
Inaccuracy Coefficient, the performance of this method the best. Nevertheless, it should be
pointed out that the results of the GM method depend on the way in which road sections are
determined to be the subject of the further identification procedure. If the sections are selected
using the cluster analysis the selection criterion can be very weak and only a few accident
cluster sections are abandoned. So a sieve role of the GM method (Hauer and Persaud, 1984) is
almost none.
The values of comparative measures calculated for the BM method place the method, on
average, between the WM and the GM methods on a ranking list of the method performace
with regard to the degree of accident clustering. In the case of the roads No 7, No 44 and No 74
the MIC values are relatively low (0.200.34), but a little higher than for the GM method. In the
case of the road No 728 the BM gives the lowest (i.e. the best) values of the MIC coefficient.
Considering the values of the measures of the degree of accident clustering it can be said that
the best average results were obtained for the GM method, the BM method has occurred quite
acceptable and the WM method can be classified as the worst.
Road Safety Evaluation Methods 205
The time period for which high risk road sections are identified differs in the considered
methods. Consequently, two consecutive time periods taken for calculation of the Space
Compatibility Index differ in their lengths. In order to investigate the space repeatability of the
WM method the results of identification for two years (1991 and 1992) were taken into
account. For the two other methods (GM and BM) tradeoff analyses were carried out on the
basis of the results from two threeyear time intervals 19911993 and 19941996. The results of
analyses are presented in Table 2. Notations in columns 37 were taken form the definition of
the SCI measure.
The different numbers of high risk road sections identified by any of the three methods for two
consecutive time intervals can confirm the phenomenon of accidents migration and
consequently the migration of accident risk, even if a remedial treatment was not applied.
For the three roads (No 7, No 44, No 74) the ratios of the number of overlapping dangerous
road sections to the total number of such sections obtained in two consecutive time periods, are
for the WM method equal to, or very close to, the same ratios calculated for the BM method.
These ratios are better than ratios for the GM method  see column 6. The values of the SCI
index are very diversified and rank the WM method on a perfect repeatability list on the first
place (road No 74), but also on the second (road No 7) or on the last place (road No 44). The
extremally high value of the SCI measure for the WM method for the road No 44 (1.12) results
mainly from only one overlapping of high risk road sections, whereas there were together five
high risk sections identified by this method in two considered time periods. On the road No 728
the WM method identified one dangerous section in the first time period and also one such
section in the second period. However, they do not overlap and there is over fifteen kilometers
distance between them.
For all considered roads values of the n/(nl+n2) ratio for the GM method are lower than
respective ratios for two other methods. The SCI measures never place this method on the first
place on the perfect repeatability list. The worst results were obtained for the road No 74. There
were seven dangerous road sections identified in the first time period and five sections
identified in the second time period using the GM method. From these sections only one pair of
road sections overlaps  so the values of both measures of space repeatability (columns 6 and 7)
calculated for this method are most unsatisfactory. On average, the GM method gives the worst
results.
206 Transportation and Traffic Theory
For the roads No 7, No 44 and No 728 the BM method gives the lowest values of the SCI
measure and for the road No 74 only marginally larger than the SCI calculated for the WM
method. Consequently, in qualitative terms, the BM method can be considered as working
better than the two other methods. This can be confirmed by the n/(nl+n2) ratios, which are
comparatively best of all.
Evaluating all values of the space compatibility measures, it can be said that the best results
were obtained for the BM method. It means that the precision in the location of dangerous road
sections determined by this method was generaly greater than in the case of the two other
methods. The WM and GM methods perform more or less on the same level with regard to
space compatibility.
In Poland the description of accident details is included in a road police report known as Road
Accident Card. This accident information contains nine qualitative features describing accident
circumstances (Tracz and Nowakowska, 1997). Two of them have been chosen as subjects of
the analysis of the accident pattern repeatability on overlapping high risk road sections. These
are: accident type and driver's behaviour. Some categories of these features have been
combined in order to obtain more coherent classification (Tracz and Nowakowska, 1998). In
such way accident type was classified in eight values and driver's behaviour was classified in
Road Safety Evaluation Methods 207
eleven values. To mark the chosen feature in the measures MIAD, MAAD and MAD, the
notation FCP is replaced by ATP for accident type pattern and by DBF for driver's behaviour
pattern. The values of the pattern repeatability measures for the analysed methods and for the
four considered roads are presented in Table 3.
The WM method gives the best results for the road No 7 for both qualitative accident features.
The average differences in accident type pattern MAD(ATP) and in driver's behaviour pattern
MAD(DBP) do not exceed 10%. The method performs better than the GM method but worse
than the BM method for the other considered roads.
The results of accident pattern repeatability on overlapping dangerous road sections are the
worst for the GM method both in the case of accident type (columns 35) and in the case of
driver's behaviour (columns 68) for all roads.
The BM method performs best of all for the roads No 44, No 74 and No 728 despite the fact
that MIAD and MAAD values are not the smallest for some cases. The average differences in
pattern repeatability (the MAD measure) are smaller for the BM method than for the other two
methods. They range from 9.3% to 13.0%. This range is determined by the results obtained for
the road No 44. The MAD values for the roads No 74 and No 728 are included in this interval.
The MAD(DBP) value for the WM method is only 0.3% greater than such value for BM
method in the case of the road No 44.
Considering all values of the measures of accident pattern repeatability it can be said that the
best average results were obtained for the BM method. This confirms the general advantage of
208 Transportation and Traffic Theory
this method over the WM and BM methods with regard to accident pattern repeatability in
relation to overlapping dangerous road sections.
Accident data such as numbers of: accidents, fatalities, injuries and involved vehicles that were
recorded on dangerous road sections as well as total length of these sections determine the
degree of accident clustering. The ratios of these numbers on dangerous sections to their totals
for a considered road can indicate the degree of clustering; the closer to unity such ratio is, the
better the results of identification. However, a comparatively large value of the described
percentage can be accompanied by a comparatively large value of the total length of dangerous
sections in relation to the length of a road  see the results in Table 1. The Method Inaccuracy
Coefficient strikes a balance of these two characteristics, thus becoming a measure which, with
regard to accident clustering, plays a fundamental role in the comparison of methods of high
risk sections identification. The lower the values of the coefficients MIC, the better the
performace of the method. The mentioned earlier ratios can help in the interpretation of MIC
values.
The repeatability measures can confirm the accuracy of a method with respect to the location of
sites with high accident risk. It is obvious that accidents can, but not always, occur in the same
sites every year but if the location of sites where these occurences oscillate around two
consecutive time intervals is indicated, the method can be considered as proper. Therefore two
aspects are important:
• the number of dangerous sections that overlap,
• the way of overlapping (i.e. the range of overlapping).
The more important is the ratio of overlapping dangerous sections to the total number of
dangerous sections from two consecutive time intervals. If this ratio is the same for two
methods, the range of overlapping indicates a preferable method  see the results for the road
No 7 in the columns 6 and 7 in Table 2. Nothwithstanding all these aspects, the SCI values are
quite satisfactory to evaluate the repeatability of methods  compare, for example, the results in
columns 6 and 7 for all roads in Table 2.
Accident patterns described by accident type and driver's behaviour can help in making a
remedial treatment decision to be taken in order to improve road safety. So, in relation to
overlapping dangerous sections, the perfect repeatability of accident pattern is strongly
recommended. As measures of repeatability of accident pattern are related to overlapping
dangerous sections they should be treated as supplementary measures to the SCI indexes. The
interpretation of these measures is very easy and intuitive  the lower values the better.
Road Safety Evaluation Methods 209
The considerations presented in this paper have shown that the range of the values of MIC,
SCI, MAD(ATP) and MAD(DBP) measures depends on a method of high risk road sections
identification and also depends on a studied road. Following a suggestion of a referee, the
authors have considered the accuracy of these measures (Hauer, 1997). However, in the
presented study the sample size is too small (only four roads) to derive general conlusions
concerning this issue. Therefore, to look into some results, only the preliminary calculations
dealing with variability of the measures were done. The variance of the measures in relation to
the three methods is presented by coefficients of variations in the Table 4.
It can be seen from this table that coefficients of variance are small for three measures (SCI,
MAD(ATP), MAD(DBP)) for the BM method. This method can be expected to have the same
efficiency for other roads. For the WM and GM methods the variance is comparatively large.
Considering the repeatability of these methods, their efficiency for other roads can be either
satisfactory or not. For the MIC measure all coefficients of variance are rather large  so the
efficiensy of all considered methods, with respect to this measure, can be difficult to predict.
These results seem to be interesting but it should be pointed out that more accident data are
required in order to confirm conclusions concerning the accuracy of the measures. This is the
aim of the authors' future research in this area.
CONCLUDING REMARKS
The purpose of this study was to develope a methodology for comparison of the efficiency of
different methods of the identification of high risk road sections.
The analysis carried out in this work enabled to detect a few deficiences in the dangerous
section identification methods (WM and GM). Thus, some suggestions of their improvement
were formulated. The suggestions concern:
• the way of defining road sections for further high risk sections classification, and
• making some definitions of classification criteria more precise.
210 Transportation and Traffic Theory
In order to generalize the results and to qualify the methods, some main and supplementary
measures independent of a methodology have been defined. Two aspects were taken into
account. One concerns the level of concentration of accidents along hazardous road sections
and the level of accidents severity concentration. The other aspect concerns the problem of
repeatability of the identification procedures during two consecutive time periods in relation to
the location of dangerous sections along a road and in relation to some descriptive features of
road accidents.
It has turned out that conclusions from comparison are not explicit. Even simplified methods
can give satisfactory identification results provided that their deficiences are made up for. A
very simple method such as the GM method performs well in terms of accident feature
concentration and badly in terms of repeatability of the method. Comparison measures for the
WM method (also a simple one) are the worst for the feature concentration aspect and the most
diversified for the other aspects. Only in the case of the more sophisticated BM method the
values of all indices are the least controversial.
All measures considered in this work are based on the results of the process of the identifcation
of high risk road sections. The methods used in such processes varies in different countries
despite the common roots in several cases. However, having the results of identification, one
can use the proposed measures, which are transferable to other countries, in comparing
different methods or in estimating the efficiency of a method.
ACKNOWLEDGEMENTS
The authors wishes to acknowledge helpful comments and suggestions of anonymous referees.
REFERENCES
Datka S., W. Suchorzewski and Tracz M. (1989, 1997). Traffic engineering. WKiL Press,
Warszawa (in Polish).
Hakkert A. S. and E. Hauer. Extent and some implications of incomplete accident reporting. In:
Methods for evaluating highway improvements. Transporation Research Record
1185, TRB, Washington DC.
Hauer E., and B. N. Persaud (1984). Problem of Identifying Hazardous Locations Using
Accident Data. Transp. Res. Rec., 975, Transportation Research Board, 131140.
Hauer E. (1995). Identification of'Sites With Promise'. International Conference in Prague on
Strategic Highway Research Program and Traffic Safety on Two Continents. The
Czech Republic, 2022 September.
Road Safety Evaluation Methods 211
Hauer E. (1997). Observational beforeafter studies in road safety. Estimating the effect of
highway and traffic engineering measures on road safety. Pergamon (Elsevier),
Oxford, New York, Tokyo.
Maher M. J. and L. J. Mountain (1988). The identification of accident blackspots: a comparison
of current methods. Ace. Anal, and Prev., 20, 143151.
Mountain L., Fawaz B., Sineng L. (1992a). The assessment of changes in accident frequencies
at treated intersections: a comparison of four methods. TE&C, 2, 8587.
Mountain L., Fawaz B., Sineng L. (1992b). The assessment of changes in accident frequencies
on link segments: a comparison of four methods. TE&Cl, 78, 429431.
Nicholson A. (1995). Indices of accident clustering: a reevaluation. TE&C, 5, 291294.
Tracz M. and M. Nowakowska (1996). Bayesian Theory and Cluster Analysis in the
Identification of Road Accident Blackspots. 13th International Symposium on
Transporatation and Traffic Theory, (Editor JB Lesort), Pergamon (Elsevier),
Oxford, New York, Tokyo, 261276.
Tracz M. and M. Nowakowska (1997). Characteristics of some accident circumstances on road
blackspot sections. 20th International Conference on Theories and Concepts in
Traffic Safety, Lund, Sweden, 57 November.
Tracz M. and Nowakowska M. (1998). Using qualitative analysis in road safety research. Third
IMA International Conference on Mathematics in Transportat Planning and
Control, Cardiff, Great Britain, 13 April.
This page intentionally left blank
Behavioural Adaptation and SeatBelt Use 213
ABSTRACT
The technical performance of seatbelts is not in doubt, but their continuing value is diminished
if driving deteriorates following the switch to seatbelt use. Janssen (1994) demonstrated such
behavioral adaptation over one year after the switch, but the effects were not sufficient to
nullify seatbelt use. Longer experiments are probably impractical because of nearuniversal
seatbelt legislation. However, UK statistics suggest that driving speeds have increased over the
years since legislation to the extent of more than nullifying the effectiveness of seatbelts. In the
present paper, Fuller's learning model of road behavior is developed in conjunction with the
perceptual phenomenon of looming as an alternative to risk formulations of behavioral
adaptation. Looming acts as negative reinforcement for unbelted drivers, but not for belted
drivers. Because it represents threat to life, negative reinforcement persists in its effectiveness:
it will take some years for the loss of looming to affect fully the new seatbelt user's behavior.
In addition, behavioral adaptation in motorists inevitably militates against the encouragement of
environmentallysustainable  but vulnerable  modes of travel, such as walking and cycling. If
safetyrelated interventions are to be properly assessed, there must be adequate empirical data 
obtained over a timespan which permits a full assessment of behavioral adaptation  in
conjunction with a plausible theoretical framework in which the data can be interpreted.
214 Transportation and Traffic Theory
INTRODUCTION
For many years, driving had been regarded simply as a perceptualmotor skill subject to
straightforward causeeffect relationships between the driver's actions and the road
environment. In this context, any technical intervention to reduce the consequences of failure in
the performance of any aspect of driving skill would be deemed beneficial; the intervention may
be to the road environment  for example, modification of the roadlayout at an "accident
blackspot"  or to the vehicle  for example, equipping the vehicle with seatbelts, ABS
braking and airbags. However, in the last thirty years or so, it has become clear that such
thinking is superficial (Summala, 1996). For example, casualties may simply "migrate" from the
former accident blackspot to other parts of the roadsystem (Davis, 1993). One crucial
difficulty is that the technical intervention can be followed over time by changes in driving
behavior, a phenomenon labelled behavioral adaptation. Occasionally, behavioral adaptation is
in the direction of more cautious driving. More commonly, however, behavioral adaptation
entails less cautious driving characteristics. It seems that behavioral adaptation is an
automatized process occurring with little, if any, conscious effort on the part of the driver.
In a contrasting example, the dangers inherent in switching from left to righthand driving in
Sweden and Iceland were accepted as an unfortunate byproduct of legislation required to
achieve conformity with neighbouring jurisdictions. In fact, the switch initially elicited reduced
casualty rates, presumably because drivers were sensitive to the particular problems and made a
conscious effort to avoid them. Unfortunately, casualty rates then increased steadily over the
subsequent years to the levels applying before the switch (Naatanen and Summala, 1976;
Wilde, 1994).
An inference that might be drawn from the above two cases of behavioral adaptation is that all
safetypertinent interventions or situations become void over time, which has profound
implications for roadsafety policy: there seems little point in devising safety interventions if
inevitably there will merely be a return to previous casualty rates. It might be argued that little
is lost if the group to whom the safetypertinent intervention or situation is targeted fails to
demonstrate better casualty rates, but this is simplistic. The targeted group does not exist in
Behavioural Adaptation and SeatBelt Use 215
isolation: it is not helpful for vulnerable roadusers  pedestrians, cyclists and, arguably, motor
cyclists  if the targeted group consists of motorists who now drive more dangerously and
generate a more hostile environment for vulnerable roadusers.
In the bestknown example of risk theory, Wilde (1982, 1994) proposes that an individual's
performance is determined by a target level of risk, dependent mainly on psychological factors,
but including economic and cultural factors. Risk is linked with the individual's "need for
stimulation", so the frequentlycited link between high casualty rates and young male drivers
can be attributed to the high target risklevels among this group (Heino et al, 1996). In brief,
the target risklevel for the driver is reflected in behavioral factors such as habitual speed,
following distance and frequency of overtaking. Following a safetypertinent intervention or
situation, risk homeostasis returns risk to the target level, as reflected in adjustments to habitual
speed, for example.
Summala and Naatanen (1988) conceptualise risk differently. Their zerorisk theory argues that
drivers internalise safety margins; these lead to habitual patterns of behavior which the driver
does not consider to be risky. Only when the driver determines that safety margins have
become inappropriate  for example, unnecessarily eliciting slow speed  does perceived risk
"switch in" to affect behavior.
Taking the case of ABS brakes, the differences between Wilde's theory and Summala and
Naatanen's theory can be illustrated. For Wilde, the driver's experience after adopting ABS
brakes may suggest that the cautious behavior applying in previously hazardous conditions is
now unduly low in risk: increase of speed in such conditions will raise risk to tolerable levels. In
contrast, for Summala and Naatanen the driver's experience after adopting ABS brakes 
perhaps obtained by chance  may suggest that cautious behavior in previously hazardous
conditions is no longer necessary: there will be no risk if the driver now increases speed in such
conditions.
216 Transportation and Traffic Theory
Criticisms of risk
Risk theories have intuitive appeal but several issues prevent full acceptance. One issue
concerns definition. In most research, risk is necessarily defined in terms of objective casualty
statistics. However, risk has a major subjective component which completely eludes objective
statistics (Adams, 1988; Rumar, 1988). For example, flying for the population at large is
subjectively riskier than driving, but in terms of casualties per unit distance traveled the reverse
is emphatically the case (UK Government Statistical Service, 1997). Given such ambiguities, it
is unsurprising that Wilde's theory and the competing theory of Summala and Naatenen can
conceptualise risk in such contrasting ways.
The proposed mechanisms to explain behavioral adaptation are dubious in their relationship to
other evidence. For example, typical cases of homeostasis concerning energy, water and
temperature regulation are mediated via specific brain areas, particularly the hypothalamus
(Carlson, 1993; Kolb and Wishaw, 1985). In such cases, the timeconstant for homeostasis is of
the order of hours and a plausible physiological model can be postulated. In contrast, the time
constant for risk homeostasis is of the order of years, suggesting an entirely different and
unfamiliar mechanism.
Another case suggests little behavioral adaptation: British casualties in poor visibility such as
fog and night are invariably higher than in good visibility (Parker and Cross, 1981; White and
JefFery, 1980; Smeed, 1977). Since poor visibility is familiar in Britain, risk theories should
apply: drivers should evince slower speeds and longer following distances to compensate for
the increased incidence of casualties in poor visibility. Although some reduction of speed and
increase of following distances may occur, the changes in behavior are far from adequate.
Furthermore, the dangers of poor visibility might elicit caution amongst new drivers, given the
Behavioural Adaptation and SeatBelt Use 117
evidence from switching direction of travel in Sweden and Iceland. However, there is little
evidence that these inferences apply. The increased casualties in poor visibility are better
understood in terms of visual perception: there are several phenomena  reduced vection,
anomalous myopia, motion contrast, aerial perspective and so on  which entail serious
misperception of speed and distance in conditions of restricted visibility (Cavallo et al, 1997, in
press; Leibowitz et al, 1982; ReinhardtRutland, 1986, 1992a,b,c).
The above reasons suggest that the invoking of risk as a global factor in roadbehavior requires
caution. As a psychological concept risk is plausible and perhaps even necessary, if only as a
shorthand descriptor in regard to behaviour that seems unnecessarily dangerous. However,
current risk formulations seem unable to link risk to any established body of psychological
knowledge. For example, one might expect to link risk to theories and research concerning
motivation and personality (Maslow, 1987) and issues of social interaction (Hogg and
Vaughan, 1998; Moghaddam, 1998), but there seems little evidence that this has been or can be
realistically attempted.
As a starting point in developing theories of roadbehavior, it may be better to consider (a) each
safetypertinent intervention or situation in detail on an individual basis, while (b) seeking to
relate the intervention or situation closely to empirical research and theory in pertinent areas of
psychology.
The limitation of the above evidence is that it can tell us nothing about whether the driver's
behavior has altered as a result of wearing a seatbelt. This is an issue that is more difficult to
establish. Evans (1991) provides a useful overview of the types of nonintrusive datacollection
procedures that may be employed for drawing conclusions about the actual effectiveness of
seatbelts among vehicleoccupants. One procedure is to compare casualty rates in jurisdictions
matched as for as possible except in regard to seatbelt legislation. Seemingly, this is
unsuccessful because the necessary degree of matching is difficult to attain: other legislation
and differing patterns of roaduse are likely to confound the effects of seatbelt use.
Other procedures relate to the time around the introduction of legislation for compulsory seat
belt use in a given jurisdiction. This can be a simple beforeandafter count of casualties or a
more sophisticated timeseries analysis: a discontinuity is expected in an otherwise continuous
function for casualtyrates against time. The ideal source of data for such a study is a
jurisdiction in which seatbelt use switches from zero before legislation is introduced to 100%
after legislation is introduced. Unfortunately, in many jurisdictions seatbelt use has tended to
change only gradually over the years, whether legislation is introduced or not: any changes that
may result are thus difficult to attribute to seatbelt use.
However, the case of the United Kingdom may provide sufficient of a contrast in seatbelt use
during pre and postlegislation times: legislation in 1983 brought a change in seatbelt usage
from about 40% to over 90%. Scott and Willis (1985) reported about a 20% reduction in
fatalities, which  given the 50% change in seatbelt usage  is in line with Evans and Prick's
results from belted and unbelted occupants of the same vehicle. Evans and Prick's results were
also broadly confirmed in a more sophisticated timeseries analysis (Harvey and Durbin, 1986).
Nonetheless, caution is required in the interpretation: drinkdriving legislation was also
introduced simultaneously in 1983 (Broughton and Stark, 1986) and selective recruitment may
apply when a safetypertinent intervention is not compulsory: seatbelt users prior to legislation
were probably individuals who put a premium on safety (Evans, 1985). The 20% reduction in
fatalities therefore overestimates the effect of seatbelt use. This indicates that the effectiveness
of seatbelts suggested by Evans and Prick diminished over the duration of Scott and Willis'
study. Since the study extended one year either side of the introduction of legislation,
behavioral adaptation was possible.
Behavioural Adaptation and SeatBelt Use 219
An experimental study
Janssen (1994) confirms this view. This important study was unique in the degree of control
exercised over nuisance variables. Three groups were investigated  new seatbelt users who up
to the beginning of the study had been habitual nonusers, habitual seatbelt users and habitual
seatbelt nonusers  in various driving behaviors on Dutch freeways over one year. Janssen had
intended to include a fourth group  new seatbelt nonusers who had been habitual users  but
could not obtain willing participants. Compared with the other groups, new seatbelt users
increased their speeds by 1.6 km/hr  average speeds were 110 km/hr  and reduced their
following distances. These changes do not negate the benefit for seatbelt users. Empirical data
suggest that the ratio of road fatality rates at two speeds is determined by the fourth power of
the ratio of the speeds (Evans, 1991). Hence, the above increase in speed implies a 6% increase
in fatalities. The reduced following distances would add to fatalities, although quantification in
this case is not easy. Janssen found that behavioral changes had not stabilised: given time, they
could have seriously compromised the benefit for seatbelt users.
This assertion is consistent with trends in British roadtraffic speeds since the early 80s. Until
very recently officiallyavailable data on speeds and speed trends has been sparse. However,
Thomson et al (1985) as part of a study concerning roaduse in relation to schoolage children
reported that on urban and suburban roads for which a 30 mph (48 km/h) limit applies, speeds
in 1981 averaged 28 mph (45 km/h), with 38% of vehicles exceeding the limit. Corresponding
1996 figures are 33 mph (53 km/h) and 72% (UK Government Statistical Service, 1997).
Following Evans' relationship, the 8 km/h increase in mean speed implies a 92% increase in
fatalities  much more than enough to nullify the benefit of seatbelt for their users, given Evans
and Prick's 40% reduction in fatalities due to beltuse from their data for belted and unbelted
occupants of the same automobiles. Changes other than seatbelt legislation have occurred over
the eighties and nineties  for example, in regard to the comfort and easeofuse of automobiles
 which might also have contributed to the elevation of speed.
It is unfortunate that much less is known about other trends in driving behavior, such as close
following and violation of overtaking rules.
With regard to longterm empirical trends in fatalities for British automobile occupants, data
over a 20 year period  much longer than was available for the studies of Scott and Willis
(1985) and Harvey and Durbin (1986)  show a dip around 1983 and 1984 which was not
220 Transportation and Traffic Theory
maintained over subsequent years (UK Government Statistical Service, 1991). Adams (1994)
reports at length on this and similar longterm trends from other jurisdictions.
The problem for empirical research is that controlled seatbelt studies with the timespan of,
say, Achenbrenner and Biehl's (1994) ABS study are probably unrealistic, since most
jurisdictions now opt for compulsory seatbelt use. Indeed, this applies in the Netherlands, the
location of Janssen's study. In the absence of a longterm experimental study, one might appeal
to theoretical issues. However, the currentlyprevailing risk formulations that are postulated for
behavioral adaptation are problematic. Hence, the extent of behavioral adaptation for seatbelt
use remains something of an open question. Proponents of seatbelt use can argue that other
factors in the everchanging roadway environment account for the disappointing trends. Of
course, proponents need to identify these alternative factors  which they have yet to do in any
systematic way.
However, the theoretical underpinnings for behavioral adaptation need not be couched in terms
of the slippery concepts of risk and homeostasis. Earlier, I argued that each safetypertinent
intervention or situation should be considered individually and in the context of established
bodies of psychological research. In fact, there is an alternative model of driver behavior
offered by Fuller (1984, 1988, 1992) which can be developed to explain why behavioral
adaptation eventually becomes substantial in the case of seatbelt use.
Fuller's model refers to simple learning in humans and nonhumans, particularly classical
conditioning  associated with Ivan Pavlov  and operant conditioning  associated with B. F.
Skinner (Mazur, 1994; Leslie, 1996). The foundation for such learning resides in associations
with desired outcomes  rewarding stimuli or events  and associations with undesired
outcomes  aversive stimuli or events.
Operant conditioning concerns behavior that is under voluntary control at least in the early
stages in the acquisition of a given behavior. In a typical Skinnerian study, a fooddeprived
pigeon learns to peck a key if the pecking is followed by the presentation of a pellet of food. In
a contrasting example, an electrician always turns off the mains supply before examination of an
appliance, since this has prevented electric shock in the past. As learning becomes established,
the learnt behavior is likely to become increasingly automatized, with the learner having little
insight or understanding of the learnt behavior; the learning can then take on some of the
characteristics of classical conditioning. Indeed, the different types of conditioning can
simultaneously affect many cases of learning.
Another feature of the rarity of negative reinforcement on the road is that learning may not
generalise beyond the specific circumstances of the negative reinforcement. Thus, surviving
participants in fatal collisions drive cautiously ever afterwards when faced with the specific
circumstances of the collision, particularly the location of the collision  but their driving may
be little altered away from these specific circumstances (Foeckler et a/, 1978; Rajalin and
Summala, 1997).
222 Transportation and Traffic Theory
A major interest of Fuller concerns the difficulty in getting drivers to anticipate danger. For
example, discriminative stimuli signal potential reinforcement. A roadsign or markings before
a bend on a narrow winding road is a potential discriminative stimulus leading to slowing, since
such signs indicate hidden hazards. However, such roadsigns and markings are often
ineffective (Shinar et al, 1980): collisions are rare, so the driver fails to slow.
Fuller has less to say about behavioral adaptation as such. Indeed one area in which his model
and risk theories seem to conflict concerns seatbelts: Fuller (1984, p. 1139) believes that seat
belts remain beneficial for their users. This implies that the effectiveness of a negative reinforcer
is independent of its consequences: the reduced injuries for a seatbelt user do not alter the
effectiveness of a given collision as a negative reinforcer. Although this seems a dubious
proposition, the rarity of collisions argues that such issues would in fact have little effect.
Furthermore, the specificity of the learning among fatalcollision survivors was noted above
(Foeckler etal, 1978; Rajalin and Summala, 1997).
However, the scope of negative reinforcement can be extended in another way. For every
collision that a driver may experience, the "forgiving" nature of the roadway suggests that there
are likely to be many nearmisses. Although, not entailing tangible damage and injury, near
misses are in fact likely to be negatively reinforcing. It might be retorted that this is simply a
matter of classicallyconditioned or secondary reinforcement, by which an effect  say,
vestibular reaction to sudden deceleration  becomes negatively reinforcing because of its
association with the primary reinforcer of collision. Indeed, were this the case little might be
added by considering nearmisses, given the rarity of collisions and the limited chances of
associating deceleration and collision.
However, nearmisses provide primary negative reinforcement in their own right; the features
of nearmisses do not need to occur with collisions in order to become effective.
Behavioural Adaptation and SeatBelt Use 223
VISUAL EXPANSION
Potential collision with an object  for example, a pedestrian, another vehicle or a static item of
road "furniture"  is accompanied by visual expansion at the driver's eye elicited by the object.
Features of this expansion, such as angular velocity and acceleration, in conjunction with
knowledge of the object's physical size and level of driving experience, provide information
from which the driver can determine timetocollision and take controlled avoiding action
(Cavallo & Laurent, 1988; Lee, 1976; Stewart etal, 1993).
Nearmisses often entail sharp braking, with concomitant visual, vestibular and kinaesthetic
changes: the unbelted, unrestrained motorist will be thrown towards interior fittings  the
steeringwheel, dashboard, windscreen, windscreen pillars and so on. The visual changes
associated with collision between the motorist and the internal fittings are similar to those for
collision between the automobile and an object external to the automobile, with the crucial
exception that simple mathematical considerations show that visual expansion will be much
greater in the former case. This is because collision between motorist and internal fittings
concerns much closer distances and directly affects the self: while the motorist is separated
from collision with external objects by the automobile's superstructure, the motorist's motion
towards the internal fittings is not. Objects close to oneself subtend bigger visual angles than
distant objects. When object and self move closer together  as with a motorist being thrown
towards the vehicle's internal fittings  the rate of change of the subtended visual angles is of a
different order of magnitude, in comparison with the motorist and external objects.
The unbelted motorist thrown towards the vehicles's interior fittings provides conditions that
evoke the looming phenomenon, defensive and fear responses elicited by the threat of
immediate collision conveyed visually. For a wide range of nonhuman species, nothing more
elaborate than a rapidly expanding shadow backprojected onto a screen is sufficient to elicit
looming responses; neonates are affected similarly to adults  neonates having not encountered
collision cannot have learned to link visual expansion with collision (Schiff et al, 1962; Schiff,
1965). Human infants are affected similarly to other species (Ball and Tronic, 1971), while
human adults evince looming responses to moving objects that stop short of hitting them (King
et al, 1992). Such evidence argues that looming responses are "hardwired": they are the subject
of automatic and inborn neural processes.
224 Transportation and Traffic Theory
Further support for this assertion comes from human psychophysical evidence that centrifugal
visual motion  corresponding to visual expansion  elicits special responses in a number of
phenomena: motion aftereffects, reactiontimes and threshold responses (Ball and Sekuler,
1980; Georgeson and Harris, 1978; ReinhardtRutland, 1994). Motion aftereffects in particular
provide evidence of lowlevel neural processing (ReinhardtRutland, 1987, 1988). Looming
itself can therefore act as a primary negative reinforcer, without any requirement that it be
preceded by experience of a collision.
The negative reinforcement contingencies associated with sharp braking differ according to
whether or not a seatbelt is worn. The seatbelt user suffers much less motion within the
interior of the automobile than does the seatbelt nonuser (Evans, 1991): the former will be
protected from looming responses. The greater vestibular and kinaesthetic responses for seat
belt nonusers will no doubt also provide other primary negative reinforcers for this group,
although there seems to be less in the way of research to which this can be related, probably
because such responses are relatively inconvenient and difficult to control in the laboratory.
Research in aviation psychology may provide some clues (Hawkins, 1987).
Given that nearmisses are more frequent than collisions on the road  and therefore likely to be
experienced in a variety of situations and locations  greater generalisation of any learning
instigated by nearmisses is likely. In contrast with the case of fatalcollision survivors noted
above (Foeckler et al, 1978; Rajalin and Summala, 1997), slower driving among seatbelt non
users is not likely to be confined to specific circumstances of nearmisses, but is likely to apply
to all circumstances.
As noted earlier, behaviors learnt during negative reinforcement are not easily lost, because of
the threat to life that negative reinforcement represents; hence, the longterm fear of the
collision circumstances shown by fatalcollision participants (Foeckler et al, 1978, Rajalin and
Summala, 1997). Often the individual can avoid the negative reinforcement altogether: a
consequence is that the individual is then unlikely to find out if reinforcement contingencies
remain constant. Away from the road, a good example concerns food aversions. An individual
who consumes an unfamiliar food that is followed by sickness is likely to avoid that food
thereafter: the sickness is a negative reinforcer for avoidance of the food. The individual will
Behavioural Adaptation and SeatBelt Use 225
not find out if the case of sickness was a oneoff occurrence, or was truly a consequence of that
food (Carlson 1993).
A difference between seatbelt nonuse and food aversion is that the reinforcement
contingencies for the driver are less predictable, since other roadusers are everpresent: the
driver is likely to experience nearmisses, no matter how slowly he/she drives. When near
misses are found to be no longer associated with looming, extinction of slow driving becomes
possible, while positive reinforcement such as early arrival at destination can become more
salient. This analysis supports Janssen's assertion that behavioral adaptation requires much time
to be complete.
One factor not easily included in a controlled study of seatbelt use is the increasing proportion
of the driver population  new licence holders  who will always have worn seatbelts and will
never experience looming: a negative reinforcer leading to safer driving behaviors is
permanently unavailable. The performance of these drivers early in their driving careers will
therefore be more dangerous than the performance of their predecessors who initially lacked
seatbelts. Given that new licence holders include young males  notorious for dangerous
driving (Evans, 1991)  this factor will over time come to affect reported statistics substantially.
DISCUSSION
In this paper, it is argued that risk models have serious shortcomings in explaining behavioral
adaptation: definition of risk, substantiation of risk homeostasis and the fact that behavioral
adaptation does not always relate in obvious ways to risk however defined are problems.
Instead, given our rudimentary knowledge of psychological factors  particularly as they act
over long periods of time  behavioral adaptation should currently be considered on a caseby
case basis.
Regarding the case of seatbelts, Fuller's discussion of negative reinforcement on the road can
be developed to provide an explanation of behavioral adaptation among seatbelt users. While
collisions entail tangible damage and injury  so have obvious reinforcing properties  primary
negative reinforcement need not be so restricted. In nearmisses, not using a seatbelt leads to
vehicle occupants being thrown towards the internal fittings of the automobile and the looming
phenomenon  fear and avoidance responses conveyed by visual information for imminent
226 Transportation and Traffic Theory
collision  can be a primary negative reinforcer in the driver learning slow, careful behaviors. In
contrast, the nearmisses encountered by seatbelt users are not accompanied by substantial
movements within the automobile: looming is eliminated as a negative reinforcer, to the
detriment of driving behavior.
After switching from not using seatbelts to using seatbelts, the driver may avoid the fast
driving behaviors that lead to nearmisses, analogous to the avoidance applying in food
aversions. However, given that the driver must interact with other roadusers, he/she will
necessarily encounter nearmisses which with belt use no longer engender looming. Only then
can the behaviors learned under the influence of looming begin to be extinguished. Nonetheless,
the persistence of learning under the influence of primary negative reinforcers  with their link
with threat to life  suggests that learning faster driving behaviors will take much time. Another
factor as time proceeds after seatbelt legislation, is that there will be increasing numbers of
new licenceholders  particularly young males  who have always used seatbelts, for whom
looming will never have been a negative reinforcer.
The eventually substantial effects envisaged in the present proposal are consistent with other
evidence. Janssen's (1994) study demonstrated small behavioral changes, which nonetheless
would have been greater had the study been longer, an assertion supported by longterm
increases in UK traffic speeds and in casualty rates. In fact, such data suggest that behavioral
adaptation may have reversed the benefit for seatbelt users in Great Britain.
The urgent need for an adequate mix of empirical and theoretical knowledge
The case of seatbelts is an especially instructive one for our appreciation of interventions
intended to reduce casualties. Of all interventions, seatbelts have been regarded as particularly
efficacious by many governmental or quasigovernmental authorities. Their belief may seem to
be supported by certain shortterm empirical investigations. However, it is much more difficult
to amass evidence to support such an intervention in the long term, particularly when, as in the
case of seatbelts, legislation has made usage mandatory. Furthermore, the effects of other
interventions and developments can never be discounted: in Britain, the introduction of drink
driving legislation simultaneously with seatbelt legislation is a clear case in point.
It is not unusual to hear the offthecuff remark among the roadsafety community that present
driving behavior would be careful and sedate if the clock was turned back and the driver was
unbelted and surrounded by sharp rigid interiorfittings which would impact with him/her in the
event of sudden braking. Yet, it seems highly unlikely that such an assertion would ever be put
Behavioural Adaptation and SeatBelt Use 221
to the test. Once a safety intervention attains compulsory status, it is thereafter highly
improbable that that intervention will be the subject of rescinding legislation leading to its
abandonment. Furthermore, such an intervention would probably not have societal support:
recall Janssen's (1994) inability in recruiting habitual seatbelt users who were prepared to give
up seatbelt use for the year's duration of this study.
Nondrivers
The main concern of this paper has been with the consequences of a particular safety
intervention as it affects drivers. In that regard, it might be argued that it is unfortunate that a
safety intervention becomes ineffective for the group to whom it is targeted, but nothing is lost
by this: the targeted group at least may not be worse off than before the intervention.
However, this is to imply a proposition that the roadsystem entails moreorless separate and
noninteracting groups of roadusers. Unless groups are rigorously segregated, this is simply
untrue: how one group performs on the road inevitably affects other groups. As an increasing
number of authors have noted, some groups are going to be disadvantaged by any level of
behavioral adaptation among drivers resulting from interventions directed at driver safety: it is
of absolutely no benefit to vulnerable groups such as pedestrians and cyclists that drivers travel
faster than they did twentyfive years ago. As Davis (1993) eloquently shows, the persistent
disregard of the needs of vulnerable groups of roadusers amounts to an issue of civil rights; it
is one that governments  particularly in the UK  have consistently ignored under the pressure
of rich and vocal interests representing various strands of "the motor industry".
Now this lack of official concern with these civil rights is reaping problems that affect all road
users, including motorists themselves. Increasingly, congestion is affecting urban and suburban
roads. In governmental circles, the answer no longer lies in new road build. Beyond concerns
about quality of life, pollution and the like  even if the civil rights of vulnerable roadusers are
discounted  there is a realization that automobiles are extremely inefficient users of roadspace
(Highway Code, 1993; Marchetti, 1983): each new stretch of road only accommodates a
minimal increase in motoring. Such issues compel governments to campaign for greater
dependence on walking and cycling, whatever might be the dangers (Adams, 1988, 1994;
Davis, 1993;Hillman<?fa/, 1991).
One of the more bizarre developments in roaduse campaigning in Northern Ireland  and
perhaps elsewhere  is an emerging view that children  that most vulnerable group of road
users  should be taking a lead and travelling to school on foot. The "schoolrun" is seen as the
228 Transportation and Traffic Theory
Such observations all point to the need for coherent combinations of empirical and theoretical
understanding regarding the psychological factors at play in safetyrelated interventions. It may
be that this understanding always has to be atomistic and specific to individual safety
interventions. On the other hand, it may be that more holistic modelbuilding can be
undertaken; Fuller's learning model might well form the basis for such an undertaking.
Compared with risk formulations, the empirical and theoretical underpinnings of Fuller's
learning model are considerably more secure and there is no need to posit intervening variables
whose definition and investigation are at the least problematic.
Another and recentlyrecognized illustration of the need for adequate empirical and theoretical
understanding regarding psychological factors concerns the everexpanding propensity of
information technology to automate, direct navigation or otherwise affect the task of driving.
As Harris and Smith (1997) and Young and Stanton (1997) show, it is necessary that such
developments can be understood and evaluated within a solid framework of psychological
knowledge. Such technology will be of limited value for the motorist if it merely displaces
congestion from one location to another. More crucially, if the cognitive demands (or lack of
them) of the technology affect driverperformance in relation to other roadusers, then its value
Behavioural Adaptation and SeatBelt Use 229
really does become contentious  as already remarked, it is at last becoming unacceptable that
technical gains for one group of roadusers entail losses for vulnerable groups to whom they
are not applicable. Judging from the past, if the use of information technology on the road is
allowed to develop without restraint, it will become difficult to reverse the process if and when
detrimental effects that should have been foreseen become serious.
REFERENCES
Adams, J. G. U. (1988). Risk homeostasis and the purpose of safety regulation. Ergonomics,
31, 407428.
Adams, J. G. U. (1994). Seat belt legislation: the evidence revisited. Safety Science, 18, 135
152.
Aschenbrenner, M. and B. Biehl (1994). Improved safety through improved technical
measures? Empirical studies regarding risk compensation processes in relation to anti
lock brake systems. In: Changes in Accident Prevention: The Issue of Risk
Compensation (R. M. Trimpop and G. J. S. Wilde, eds.),. Styx, Groningen
Netherlands.
Ball, K. and R. Sekuler (1980). Human vision favors centrifugal motion. Perception, 9, 317
325.
Ball, W. and E. Tronik (1971). Infant responses to impending collision: optical and real.
Science, 111, 818820.
Broughton, J. and D. C. Stark (1986). The effect of the 1983 changes to the law relating to
drink/driving. Research Report 89. Transport and Road Research Laboratory,
Crowthorne UK.
Carlson, N. R. (1993). Psychology, the Science of Behavior. Allyn and Bacon, Boston MA.
Cavallo, V. and M. Laurent (1988). Visual information and skill level in timetocollision
estimation. Perception, 17, 623632.
Cavallo, V., J. Dore, M. Colomb, and G. Legoueix (1997). Distance perception of vehicles
rear in fog. In: Human Factors in Road Traffic II (P. Albuquerque, J. Santos, A.
Pires da Costa and C. Rodrigues, eds.). Braga, University of Minho, Braga, Portugal.
Cavallo, V., J. Dore, M. Colomb, and G. Legoueix (in press). Distance overestimation of
vehicle rear lights in fog. In: Vision in Vehicles 7 (A. G. Gale, ed.), NorthHolland,
Amsterdam.
Davis, R. (1993). Death on the Streets: Cars and the Mythology of Road Safety. Leading
Edge, Hawes UK.
Duma, S. M., T. A. Kress, D. J. Porta, C. D. Woods, J. N. Snider, P. M. Fuller, and R. J.
230 Transportation and Traffic Theory
Victor J. Blue, New York State Department of Transportation, Poughkeepsie, NY, USA
Jeffrey L. Adler, Rensselaer Polytechnic Institute, Troy, NY, USA
INTRODUCTION
Pedestrian flow is an important component in the design and analysis of transportation
facilities and in urban transportation planning. The need to assess the level of service of
pedestrian walkways motivates much of the work on modeling pedestrian flows. Statistically
derived flow models, expressing the fundamental relationships between density, flow, and
speed, are used to generate level of service criteria. The methodology of assessing level of
service for pedestrian walkways is similar to that for vehicle roadways.
Microscopic modeling of vehicle flows is well studied and most, if not all, of the resulting
models have been shown to generate fundamental relationships. In contrast, few attempts have
been made to develop microscopic models of pedestrian flows that capture the complex
behavioral movements and decisions. Historically, researchers have found the modeling task
to be daunting.
In recent years, Cellular Automata (CA) has emerged as a technique for modeling complex
behavior using a set of simple rules. Cellular Automata (CA), when emulating the complex
behaviors of living systems, is characterized as an artificial life approach to simulation
modeling (Levy 1997). CA is named after the principle of automata (entities) occupying cells
according to localized neighborhood rules of occupancy. The CA local rules prescribe the
behavior of automata creating an approximation of actual behavior. Emergent group behavior
is an outgrowth of the interaction of the microsimulation rule set. Unlike traditional simulation
models that apply equations and not behavioral rules; CA behaviorbased cellular changes of
state determine the emergent results. Furthermore, CA models function as discrete
236 Transportation and Traffic Theory
idealizations of the partial differential equations that describe fluid flows and allow simulation
of flows and interactions that are otherwise intractable (Wolfram, 1994).
The purpose of this paper is to present CA microsimulation models of uni and bidirectional
pedestrian walkways to demonstrate the ability of capturing fundamental properties of
pedestrian movements using a fairly simple rulebased approach. The paper begins by
characterizing the pedestrian flow problem. A brief literature review of previous modeling
efforts is provided. The formulations for the uni and bidirectional walkway cases are
described. This is followed by a discussion of results gathered from numerous simulation
experiments. Included is a comparison of emergent flow patterns to the fundamental
parameters described in the Highway Capacity Manual (1994).
BACKGROUND
Microscopic models of carfollowing behavior and continuum models of traffic streams are
mathematical relationships to describe the complex, and often chaotic, movement of vehicles
and the high degree of interaction between vehicles. It is well known that vehicle flow exhibits
nonlinear speed transitions and selforganized criticality. Lieberman and Rathi (1997) suggest
that, while the local behavior of individual vehicles may be reliably represented logically and
mathematically with acceptable confidence, the complex simultaneous interactions of the
system are difficult to describe and in mathematical forms. In response to the need to study
traffic networks, researchers have turned to using simulation.
Microscopic simulation attempts to model the behavior and movement of individual vehicles.
Simulations are constructed from a set of models that represent a variety of behaviors from
changing vehicle speed to lane switching and wayfinding decisions. Traditionally, microscopic
simulation of vehicle flow is constructed around carfollowing models. However, microscopic
simulations are difficult to develop and calibrate due to the complex behaviors and large
number of parameters that are needed. In addition, chaotic vehicular traffic phenomena that
occur at higher densities are difficult to capture with equationbased models and therefore are
also difficult to simulate with sufficient realism.
and Latour, 1995), and networklevel vehicle flows in the TRANSIMS model (Nagel, Barrett
and Rickert, 1996). While pedestrian flows are much more complex and chaotic than vehicular
flows, the success with which researchers were able to use CA to model traffic flow provides
the impetus to explore the use of CA for pedestrian flows.
Although fundamental properties for pedestrian flows and level of service methods for
assessing capacity are well established, it has been difficult to develop microscopic models of
pedestrian flows that generate these fundamental properties. Pedestrian flow is highly chaotic
and individualized. Pedestrian corridors may have several openings and support movement in
several directions. Pedestrian walkways are not regulated as are roadways. For the most part
pedestrian flows are not channeled; pedestrians are free to vary speed and allowed to occupy
any part of a walkway. Unlike roadways where vehicle flow is separated by direction, bi
directional walkways are the norm rather than the exception. However, since safety and crash
avoidance are less of a concern to pedestrians, slight bumping and nudging is often a part of
walking through crowded corridors. Pedestrians are capable of changing speed more quickly
when gaps arise and can accelerate to full speed from a standstill. In addition, it is not
uncommon for pairs or groups of pedestrians to walk sidebyside or in clusters.
Gipps (1986) came closest to developing a realistic microscopic model of pedestrian flow with
a cellbased approach using a hexagonal lattice. Individual pedestrians occupy hexagonal cells
and over discrete time intervals are moved in relation to one another. Movement is dictated
according to reverse gravitybased rules. Pedestrians are repelled from each other as they seek
their destinations in buildings. This work was limited in that the formulation is encumbered by
sequential floatingpoint calculations based on the number and proximity of pedestrians who
aim to avoid each other. In addition, the approach was tested only on onedirectional
uncongested flows with inconclusive, though apparently reasonable results for the uncongested
speedvolume relationship. Gipps also failed to yield a tenable method for handling bi
directional movements.
Fruin discusses several efforts from the 1960's to develop models of pedestrian flow within
different scenarios. There are few notable recent efforts by researchers to model pedestrians
using more traditional techniques. Lovas (1994) examined pedestrian traffic in building
evacuation with a discrete event queuing network. AlGadhi and Mahmassani (1991) simulated
crowd behavior using a set of conservation of mass flow simultaneous partial differential
equations solved for several classifications of pedestrians over discrete time steps.
238 Transportation and Traffic Theory
Helbing and Molnar (1995) developed a model in which a pedestrian behaves as if acted upon
by external attractive and repulsive forces, termed social fields. Their model displayed the
formation of lanes by opposing direction flows and oscillatory changes in walking direction at
a doorway. The paper demonstrated a single instance of these abilities and did not show results
over a range of densities or calibrate them against established fundamental flows. The CA
approach explored in this paper is distinctly different from the social fields model,
computationally much simpler and based on maximizing forward progress.
The attractiveness of using C A is that the models are based on behavioral rules, rather than
performance functions. This paper will demonstrate that emergent group behavior comes from
the dynamics of the pedestrians in motion across the defined walkway. The CA behaviorbased
cellular changes of state determine the emergent results, giving rise to very lifelike phenomena,
and to new possibilities of modeling based on behavioral rules. The relevant behaviors become
the focus of the modeling and that leads to an understanding of the factors that contribute to the
dynamics.
MODEL FORMULATION
The objective is to develop an intuitively and empirically appealing CA microsimulation for
modeling bidirectional flows on a pedestrian walkway. The model aims to employ the
essential minimal rule set needed and avoids the pitfall of trying to capture all possible
behaviors of pedestrians which are numerous to identify and problematic to verify. Many
pedestrian behaviors are unnecessary distractions from the essential parameters. It has been
observed in CA simulations that very simple models are capable of capturing essential system
features of extraordinary complexity (Bak 1996). Aiming at the essential and minimal rule set
by eliminating anything but critical behavioral factors permits a clear understanding of the
underlying fundamental dynamics.
There are two fundamental elements of pedestrian movement to capture: forward movement
and resolving conflicts. Forward movement refers to the velocity and acceleration of each
pedestrian. Desired walking speeds vary over the pedestrian population. Under less crowded
conditions, individual pedestrians will strive to accelerate toward their desired maximum
walking speed. In congested walkways, pedestrians will adapt their speed to the prevailing
flow rates in the immediate walkway neighborhood.
To account for differences in desired maximum walking speeds (termed vjmax), a pedestrian
population consisting of three walker classes is adopted:
Pedestrian Flows From Cellular A utomata 239
Conflict resolution refers to movements that are undertaken to avoid headon collisions or to
enable passing or bypassing. In a bidirectional walkway, pedestrians from opposing
directions can vie for the same location. The proposed model will introduce the concept of
"place exchange", whereby the positions of opposing pedestrians are swapped to emulate the
collision avoidance behavior. Passing movements occur when two pedestrians are moving one
in front of the other in the same direction and the trailing pedestrian wishes to accelerate past
the lead pedestrian. Bypassing occurs when two pedestrians approach one another from
opposite directions and circumvent one another. Termed "lanechanging" or "sidestepping"
behavior, the passing pedestrian shifts position to the right or left in order to enable a higher
walking speed. Though pedestrians do not follow lanes, lane changing is a helpful term to use
when referring to a CA grid of cells where de facto lanes exist.
Rule Set
The CA microsimulation is based on six rules applied across four parallel updating stages. In
the first parallel update stage, a set of lane changing rules is applied to each pedestrian on the
lattice to determine the next lane of each pedestrian. All pedestrians are assigned to the new
cells during the second stage. In the third stage, a set of forward movement rules is applied to
each pedestrian. The allowable speed of each pedestrian is based on the available gap ahead
and the pedestrian's desired speed with all the entities in their current positions. Finally, the
pedestrians hop forward to new cells in the fourth update.
In the first parallel update stage Rules 15 determine lane switching. Pedestrians can change
lanes only when an adjacent cell is available. Rule 1 determines if the adjacent cells to the
immediate right and left are available. Rule 2 resolves possible conflicts if an adjacent cell is
available but the cell two lanes over is occupied. A random number is drawn to designate the
lane as free to this pedestrian or to the pedestrian two cells away. Rule 3 eliminates a lane
change if both adjacent cells are unavailable. If an adjacent lane is free, then lane change is
determined by the maximum gap in Rule 4. If the maximum gap is common to two or more
lanes, Rule 5 breaks ties for making lane assignments. Rule 5a, an 80/10/10 split for all three
240 Transportation and Traffic Theory
lanes, assumes pedestrians generally stay in the current lane and drift out of it occasionally.
Rule 5b, a 50/50 split between the adjacent lanes, is the most reasonable assumption. Rule 5c,
a 50/50 split between the current lane and an adjacent lane, assumes pedestrians wish to have a
cellwidth of separation from a person in the adjacent lane as much as they want to stay with
the current course. These probabilities worked well for the simulation. The second parallel
update stage moves all the pedestrians into their new lanes.
The third parallel update stage determines the forward movement of the pedestrians. The gap
ahead is determined first. If they are opposing pedestrians, Rule 6 guards against deadlocks by
emulating what people actually do. Under constrained conditions opposing pedestrians may
exchange places. In actuality temporary standoffs may occur when people guess which way to
step past one another. Thus, the simulation contains a probability of a temporary standoff
between closely opposing walkers. With probability p_exchg closely opposing pedestrians
exchange places in the time step. The opposing entities each move the same number of cells,
which is 0, 1, or 2 cells. In the fourth parallel update, the pedestrians are moved forward based
on the gaps determined in the third parallel update stage.
The available gap ahead depends on the direction of flow of the next vehicle downstream.
From the gap calculation, if both pedestrians are going in the same direction, the new velocity
of the follower is the minimum of the desired velocity of the pedestrian (v_max) and the
available gap ahead. If the gap between opposing pedestrians is less than the total distance that
both pedestrians could move at maximum speed (i.e., 8 cells is the maximum  4 in each
direction), then the updated velocity is the minimum of v_max and moving halfway forward.
Moving halfway forward guards against collisions and hopping over other entities. Taking the
minimum gap of the same and opposite directions of movement is shown as a combined task at
the end of the gap calculation.
242 Transportation and Traffic Theory
SIMULATION EXPERIMENTS
A series of simulation experiments were conducted to evaluate the rule set over different uni
directional and bidirections scenarios. The pedestrian walkway is modeled as a circular lattice
of width W and length G. Each cell in the lattice is denoted L(i, j) where 1 < i <, W and 1 ^ j <
G. Pedestrian densities are predetermined at the start of the simulation and remain constant
throughout each run. At the start of each simulation, for a density d, the proportion of
occupied cells, where 0.05 < d < 1.0, N = INT (d*W*G) pedestrians are created and assigned
randomly to the lattice. The circular lattice enables the set of pedestrians to interact at constant
density while maintaining strict conservation of flow. Cells in the lattice are considered square
at 0.457 m per side. This cell size is scaled according to minimal requirements for personal
space as described in the Highway Capacity Manual. The scale is also used to generate the
speedflowdensity relationships that emerge. Density as the proportion of occupied cells is
considered more helpful for discussion of results than units of pedestrians/m2, which certainly
can be converted from d/(0.457m)2.
To capture variations in walking speeds across the population a distribution of 5 percent 'fast'
walkers; 90 percent 'standard' walkers, and 5 percent "slow" walkers (termed 5:90:5
distribution) was used. This distribution had the best realization of the fundamental diagram
compared with other distributions in the singledirection case (Blue and Adler 1998). For the
circular lattice, the walking speeds are 4, 3, and 2 cells per time step respectively and these
values fall within ranges of speed and standard deviations used by others researchers (Lovas
1994).
Pedestrian Flows From Cellular Automata 243
Fundamental parameters of pedestrian flows, space, flaw rate, and, average walking speed
were generated from the experiments. For the circular lattice, space is the reciprocal of density
and is constant for each run. Flow rate is computed as the total number of revolutions made by
each pedestrian over the duration of the simulation. Speed is computed as number of total
steps divided by the time duration.
Each simulation was run for 11,000 time steps of one second each. The first 1000 time steps
were used to initiate the simulation and the latter 10,000 were used to generate performance
statistics. Each set of experiments included runs at 19 densities ranging from 0.05 to 0.95
percent occupancy in increments of 0.05. For statistical accuracy, twenty repetitions at each
density level are run and the fundamental parameters are computed as the average over these
replications. The resulting emergent fundamental profile is a map of the relationships between
speed and flow over the range of densities.
UniDirectional Flow
The first set of experiments was conducted on a unidirectional pedestrian walkway. This
eliminates the need to resolve headon conflicts as all pedestrians are moving in the same
direction. Several model forms were fitted against the resulting fundamental data. Figure 1
depicts three volume vs. density curves: (1) the results of the microsimulation experiment, (2)
a fitted May's bellshaped curve, and (3) a composite tworegime model curvefitted to the
data.
May's flow equation was estimated from the simulated data. It was found to have the
following speeddensity relationship bell curve
The best fitting model is a tworegime speeddensity relationship. Under low densities, the
flows appear to follow May's bellshaped relationship; a linear relationship between flow and
density is suggested at higher densities. The best estimated model for the data had the
following speeddensity relationship:
As seen on the chart, the maximum flow of this two regime model is approximately 80
ped/min/meter of width. The HCM suggests a capacity around 25 Ped/min/ftofwidth,
equivalent to 82 Ped/min/meterofwidth.
The results compare favorably with work reported by Virkler and Elayadath (1994) who
examined unidirectional flows and fit various curves to field data. Their study found that two
functional forms fit the data best: (1) single regime model represented by May's bellshaped
function and (2) a tworegime model with separate linear regimes. A closer look at this study
reveals that the estimation of flow equations was based on a data set where density reached a
maximum value of 3.12 Ped/m2, corresponding to a density of 0.65 on the circular lattice.
Therefore, Virkler and Elayadath's estimation is limited to low and medium densities. The
results of the unidirectional CA model for low densities is consistent with these previous
findings, combining the best features of May's bellshaped curved and the two regime linear
model. The CA data set continues to high density and provides improvements to the
fundamental model. The CA model is a useful tool to extending examination and knowledge
of pedestrian behavior and aggregate movement.
•CA Simulation
•May's Bell Curve
2regime
BiDirectional Flow
Since cases of directionally separated flow are nearly identical to having two adjoining uni
directional walkways, the emerging flowdensity profile should be similar to the uni
directional case. The HCM suggests that depending on the directional split, the flow profile
will retain the same shape except for some reduction in capacity. If the directional flows are
roughly equal then little reduction in capacity is expected. However when there is a significant
imbalance, say, a 9010 split, capacity reductions of about 15 percent have been observed
(HCM 1994).
There are cases in which flows along bidirectional walkways are not segregated; opposing
traffic intermingles and directional lanes do not readily evolve. This usually rather shortlived
case of randomly interspersed pedestrians moving in opposing directions may occur at busy
crosswalks, in crowded subway stations, and in emergency situations, among others.
Historically interspersed bidirectional flow has not been well examined, documented, or
modeled. As a result this case is not thoroughly understood.
The C A microsimulation was used to investigate the case of bidirectional interspersed flow.
When opposing pedestrians converge on the same cell but cannot negotiate sidestepping, the
conflict is resolved by exchanging the places of the opposing pedestrians. This is realized
through Rule 6 with the use of a variable, p_exchg. To the authors' knowledge, the
characteristics of place exchange have not been examined in any field study of pedestrian
246 Transportation and Traffic Theory
activity, though it is well known that pedestrians will bypass, bump, slip past, and otherwise
exchange places with one another when necessary.
— 100/0
— 90/10
— 80/20
— 70/30
— 60/40
— 50/50
Exchange rates of 100, 75, 50, 25, and 0 percent were applied across the various directional
splits. 0 percent is the base case where pedestrians cannot move if an opposing pedestrian
creates a direct conflict; 100 percent indicates that all conflicts are resolved.
Figure 3 illustrates the resulting speeddensity curves for several directional splits based on
exchange probabilities of 100 percent success. The speeddensity plots of each directional split
show a pattern emerging that is orderly and understandable. At a density below 0.45, speeds
are reasonably reduced according to the directional split. Above a density of 0.45, the high
Pedestrian Flows From Cellular Automata 247
degree of exchange rate allows higher speeds, as splits become more even. This is because
pedestrians are not as often held back by opposing persons ahead and because more balanced
directional flow provides more opportunities for faceto face conflicts and smooth exchanges
of position.
100/0
90/10
80/20
70/30
60/40
Figure 4 illustrates flowdensity curves for 9010 splits over the range of exchange
probabilities. As the rate goes down, significant attenuation of flow is observed. In addition,
even for the 100 percent exchange case there is a significant (approximately 15 percent)
reduction in capacity as the HCM describes for the case where lane formation does not occur.
The 75 percent exchange case has a similar peak volume, appropriately diminished from the
unidirectional and laneseparated bidirectional case. Lower exchange rates appear to have
peak volumes out of range of what the HCM describes.
p_exchg = 0
p_exchg = 0.25
p_exchg = 0.50
p_exchg = 0.75
p_exchg = 1.0
Understandably, most of the historical effort that has gone into capturing pedestrian flow
characteristics has been directed toward their forward movement. However, an important
factor in understanding pedestrian flow is lane changing. Lane changing has been modeled in
Pedestrian Flows From Cellular A utomata 249
vehicular traffic (see Rickert, Nagel, Schreckenberg & Latour 1995; BenAkiva, Koutsopolous,
& Yang 1995 for example). Since extensive data gathering and modeling of pedestrian lane
changes has not been previously done, the effort here has been to perceive pedestrian lane
changing from experience and empirically capture the forward movement with reasonable
rules.
Attention to modeling of lane changing is also needed due to mode locking of pedestrians, or a
marching effect, which may occur in some instances. Mode locking, or synchronization of
motions, occurs in many physical processes (Schroeder 1991). In this case pedestrian self
organization effects the ability to synchronize movement. Mode locking is almost certain to
occur at about 0.25 density if all the entities are of the same walker class, having the same
maximum speed. If rigid mode locking occurs, slowdowns and small jams do not occur and
average speed will be unreasonably high. Even with substantial opportunity to pass, the
pedestrians achieve flexible selforganization. Selforganization arises naturally in the model,
but the lane changing rules should avoid rigid locking into step, or the forward flow
characteristics will not emerge correctly. While selforganization is of interest, and the model
does display it, it is not the primary focus of attention at this stage of research investigation.
Figure 5 illustrates the relationship between lattice density and rate of side stepping. The uni
directional lane change rate (100/0 split case in Figure 5) is effected by density in an
unexpected way with a local minimum at 0.25 and a local maximum at 0.35 density. The
phenomenon appears to be an outgrowth of selforganization at various stages of interaction
among the pedestrians. There are, thus, some spatial configurations of entities that promote
lane changing and others that impede it. In fact, spatial use and the unidirectional lane change
curve can be used to explain the shape of the Scurve in the speeddensity relationship. As the
density for the 100/0 split increases above 0.25, the spatial efficiency for forward movement
decreases and lane changing increases as entities tend to selforganize into the available space,
breaking up any even spacing between pedestrians and making forward movement more
difficult.
Figure 5 also illustrates the relationship between sidesteps and density at the 1.0 exchange
probability for interspersed flow over a range of directional splits. As can be seen from the
figure, the spatial advantages of mode locking are entirely lost and the opposite effect even
occurs. In fact, the spatial advantage of mode locking at 0.25 density becomes a disadvantage
with opposing flows and increases as flows become more evenly divided. Avoiding entities
coming from the opposite direction is most effective for reasons similar to why mode locking
is so effective in unidirectional flow.
250 Transportation and Traffic Theory
100/0
90/10
80/20
70/30
60/40
50/50
When exchanges occur freely, the lane changing is not as impeded. As the exchange
probability goes down, the lane changing increases (not shown). One finding from this analysis
is that lane changing is sometimes helpful and other times not very helpful in promoting flows.
A lot depends on the spatial use and density. Especially as density increases, when a
pedestrian changes lane to pass, another pedestrian can change lanes and block the first
pedestrian's movement. Pedestrians make their decisions independently and are myopic to the
movements others may take. In actual pedestrian traffic pedestrians can "read" the body
movements of others to a certain degree, but this can also be deceptive and oncoming
pedestrians may still wind up in a face off. In road traffic lane changing, a phenomenon,
referred by some as "snaking," can be observed (Resnick 1994). In congested traffic if an
adjacent lane moves faster, drivers switch lanes and congest the formerly moving lane. Then,
the formerly stopped lane begins to move. The traffic moves like a snake, alternating the lane
that is in freer flow. The fact that lane changing can be counterproductive or nonproductive
for forward flow does not stop lane changing from happening. Pedestrians can be as myopic as
Pedestrian Flows From Cellular A utomata 251
drivers, as can be observed, when at multiple queues, queue hopping occurs to the seeming
eternal frustration of the person who always seems to switch to the wrong queue.
It is noteworthy that the CA model has a significant capability in capturing lane changing to
the degree needed for apparently any application. At the very least, lane change rules in CA
models must prevent excessive mode locking from occurring while permitting reasonable
amounts of selforganization of pedestrians.
CONCLUSIONS
This paper demonstrated the use of CA microscopic simulation for modeling pedestrian
behavior. The basic pedestrian behaviors have been approximated with a minimum rule set
that effectively captures the essential pedestrian dynamics for the single and bidirectional
case. The rule set and scaling can be adjusted to situations and conditions that correspond, in
some measure, to actual behaviors. The correspondence of the rules to actual behaviors need
not be exact. The essential rules are simple enough to program and modify without imposing
unnecessary detail, and yet capture complex phenomena.
Over all the simulation tests, the region of density 0.20.4, where the maximum flows occur,
has the largest differences in speed and volume between tests. These are presumed to be the
speedflowdensity combinations that have the most volatile dynamics. This inference agrees
with Paczuski and Nagel (1995) that reveals complex dynamics at work, especially in the
maximum flow range of the NagelSchreckenberg automobile traffic model. This is evidently
due in part to spatial sensitivities of lane changing that affects forward movement in this
region. Since nonlinear effects occur around maximum flow, the region of maximum flow
would then be the area where the greatest concentration of effort in finetuning the model
would benefit. With further work it would be possible to reveal more about and to better
understand the dynamics in the maximum flow range.
252 Transportation and Traffic Theory
Lane changing effects yield some interesting insights into the speeddensity relationship. At
lower densities mode locking may occur that is efficient for forward movement, but requires
and permits few lane changes. At higher volumes, lane changing may help as well as hinder
overall flow and evidently has relatively small impact upon the emergent group flow behavior.
However, though not well studied or understood by field research to date, lane change behavior
is an important feature of pedestrian movement. Its inclusion in the model is essential and adds
realism.
Among those issues that would benefit from further examination, lane changing and place
exchange, especially, emerge as factors that are critical to the functioning of the model. Field
studies should verify the hypothetical lane change and position exchange phenomena.
Previously published data were used to model the distribution of walker speeds. However, the
use of a CA model as a design tool would require careful study of local pedestrian populations
with respect to characteristics of lane changing, position exchange, and the speed distribution
of walkers in free flow.
The more complex and hypothetical case of interspersed bidirectional flows clearly illustrates
the modeling power of the CA method. Its capabilities have been merely touched upon in this
investigation. The CA approach yields a viable tool for pedestrian modeling that has daunted
the efforts of researchers for years. It captures microlevel pedestrian dynamics and offers an
experimental platform for better grasping the important parameters of pedestrian flows. As a
new method, its possibilities have barely begun to be theoretically explored and considerable
opportunities exist for innovative applications. Among the many possibilities, the authors are
investigating a 4directional pedestrian model and complex walking environments, such as
shopping malls, street intersections, and bus/rail stations. Multiplemode models that combine
pedestrians, autos, trucks, buses, bicycles, and so forth (e.g., autorickshaws and scooters in
Asian traffic) are plausible extensions of the CA pedestrian and auto work done to date.
REFERENCES
AlGadhi, S.A.H. and H. Mahmassani (1991). Simulation of Crowd Behavior And Movement:
Fundamental Relations And Application. Transportation Research Record, 1320, 260268.
Bak, P. (1996). How Nature Works: The Science of SelfOrganized Criticality, Springer
Verlag New York, Inc.
Pedestrian Flows From Cellular Automata 253
Helbing D. and P. Molnar (1995). Social Force Model for Pedestrian Dynamics. Physical
Review E, 51 (5) 42824286.
Lieberman E. and A. Rathi (1997). Traffic Simulation, In: Revised Monograph on Traffic
Flow Theory, Federal Highway Administration, (N. Gartner, M. Messer, and A. Rathi,
eds.) WWW publication http://www.tfhrc.gov/its/tft/tft.htm.
Levy, S. (1992). Artificial Life. Vintage Books, New York.
Nagel, K. and M. Schreckenberg (1992). A Cellular Automaton Model for Freeway Traffic. J.
Physique I, 2.
Nagel, K. and S. Rasmussen. (1994). Traffic at the Edge Of Chaos. Artificial Life IV:
Proceedings of the 4th International Workshop on the Synthesis and Simulation of Living
Systems, pp. 222225.
Nagel, K. (1996). Particle Hopping Models and Traffic Flow Theory. Physical Review E, 53
(5)46554661
Nagel, K., C. Barrett, and M. Rickert (1996). Parallel Traffic Microsimulation by Cellular
Automata and Application for Large Scale Transportation Modeling. Los Alamos
Unclassified Report 96:0050, Los Alamos National Laboratory, Los Alamos New Mexico.
Paczuski, M. and K. Nagel (1995). SelfOrganized Criticality and 1/f Noise in Traffic, Los
Alamos Unclassified Report 95:4108 (published in Traffic and Granular Flow, eds. D.E.
Wolf, M. Schreckenberg, and A. Bachem, Singapore: World Scientific, 1996, p 41) Los
Alamos National Laboratory, Los Alamos New Mexico.
254 Transportation and Traffic Theory
Resnick, M. (1994). Turtles, Termites, and Traffic Jams: Explorations in Massively Parallel
Microworlds. MIT Press, Cambridge, Mass.
Rickert, M., Nagel, K., Schreckenberg, M. and Latour, A (1995). TwoLane Traffic
Simulations Using Cellular Automata. Los Alamos Unclassified Report 95:4367 (published
in Physica A, Vol. 231, pp. 534, 1996), Los Alamos National Laboratory, Los Alamos New
Mexico.
Schroeder, M. (1991). Fractals, Chaos, Power Laws: Minutes from an Infinite Paradise. W.H.
Freeman and Company, New York.
Special Report 209: Highway Capacity Manual (1994). Transportation Research Board,
National Research Council, Washington, D.C.,
Virkler, M. R. and S. Elayadath (1994). Pedestrian SpeedFlowDensity Relationships.
Transportation Research Record, 1438, 5158.
INTRODUCTION
Transport network flows are the results of the interaction between travel demand and supply
conditions. Flows in a network are not stable due to the fluctuation of travel demand or the
occurrence of failure in supply conditions. Even if links and nodes in a network are not
physically damaged and the supply condition of the network is normal, flows in the network
may not always be stable. Travel demand is variable from timetotime or daybyday, and
the resulting network flow will be fluctuating. We sometimes experience unusual network
flow conditions due to seasonal fluctuations of travel demand
On the other hand, traffic accidents, roadworks or natural disasters occasionally causes
damage on the supply conditions of a network. Some links in the network may be closed to
traffic, and flows in the network will become unstable. Almost all of the links are damaged in
an extraordinary natural disaster such as the great Hanshin Earthquake. We may suffer from
extremely heavy traffic congestion caused by the interaction between fluctuating travel
demand and degraded network capacity.
Figure 1 shows a conceptual relation of physical conditions and traffic flows in a transport
network. As mentioned above, the degree of unusuality of traffic flows in a network is
observed even in a physically sound network. The instability of network flows will be
magnified as the degree of the physical damage of the supply conditions becomes worse. In
the countries often suffering from natural disasters like Japan, it is necessary to describe
258 Transportation and Traffic Theory
network flows and estimate network performance in deteriorated conditions for the strategic
transport network planning under uncertainty. As well as those countries, the modern
society requires more reliable transport systems. A reliable network means a network which
can guarantee an acceptable level of service for traffic even if some links of the network are
physically damaged or large amount of travel demand is occasionally generated.
Degraded Network
Degree of Physical
Damage of Network
Network reliability models have been studied for evaluating transport networks in both usual
and unusual conditions. Turnquist and Bowman(1980) presented a set of simulation
experiences to study the effects of the structure of the urban network on service reliability,
lida and Wakabayashi(1989) developed an approximation method of calculating the
connectivity between a node pair in a network. These studies are concerned with the
reliability analysis of a. pure network. Flows in the network were not explicitly considered in
those studies.
Network flows in degraded conditions might be different from the ones in ordinary
conditions. As Khattak and Palma (1997) mentioned, however, few studies have been made
for describing traveller response in normal and unexpected situations. In particular, network
flow models have not been studied for unusual conditions. The uncertainty and inconvenience
of travel may bring the reduction of travel demand and it will result in network flow patterns.
Deterministic flow models have been used in the previous network reliability analysis.
However, this may cause inconsistency between flow description and reliability evaluation
process since flow is deterministic while network itself is not. Thus, it is necessary to employ
stochastic models for describing flows in a network. The objective of this paper is first to
formulate a network flow model for describing flows in degraded conditions. The second
objective is to propose some performance measures of a network when some links are closed
to traffic. The network flow model is involved in the evaluation process of the performance
measures.
In the following Chapter 2, the Stochastic User Equilibrium model with variable demand is
applied to describe flows in a network with some disconnected links. Chapter 3 shows some
performance measures considering uncertain factors of the network flow. Approximation
algorithms for estimating the expected value and the distribution functions of performance
measures are also described in this chapter. Numerical examples are calculated in Chapter 4 to
verify the convergence of the algorithm. A brief conclusion and implication of the models are
presented in Chapter 5.
260 Transportation and Traffic Theory
A degraded road network is identified by the state vector x = {x:,..., xa,... ,x J. The element
of the vector xa denotes whether link a is degraded or not, namely x a is equal to 0 if the link
is closed to traffic, or x a is equal to 1 if the link maintains its functions in the ordinary
condition. If all links are connected, the state vector is written as xg = {2, ,1} . This
state is referred as the ordinary or normal state in the following part of this paper. The worst
state vector is written as x w = {0,.... ,0},ir\ which all links are disconnected. X = {x} denotes
the set of the possible state vectors.
Since the occurrence of a failure in a link is uncertain, the state of a network is not
deterministic. We introduce a link connectivity, which is the probability of whether the link
is connected and not closed to traffic. The connectivity for link a is denoted by
pa (a = l,... ,L). We assume the value of link connectivity is exogenously estimated and fixed.
When the link connectivity pa is independent each other, the probability of the occurrence of
a state x in a network is calculated as;
P(x)=TlPaX° (IPJ1^* (^
a £A
According to the survey by authors, 15% of drivers cancelled their trips in the deteriorated
network, and 75% of diverted drivers were obliged to choose the routes with higher risks. We
Road Network Under Degraded Conditions 261
formulate a network flow model in a deteriorated network considering those travel choice
behaviour in the network. Travelling in a deteriorated network is more uncertain than in the
ordinary network. A stochastic travel choice model is appropriate to describe a driver's travel
behaviour. Since some of the drivers may cancel their trips, the travel demand is assumed
elastic for network situations.
In the previous studies (Asakura(1996), we have applied the Deterministic User Equilibrium
with Variable Demand. Although the elasticity of demand was considered in the model, the
route choice behaviour was assumed deterministic. In order to represent uncertain route choice
behaviour in a deteriorated network, the stochastic route choice modelling is desirable. This
will also resolve the inconsistency between route choice behaviour and trip making behaviour.
We will apply a Nested Logit based Stochastic User Equilibrium model with Variable Demand
(NLSUEVD) for describing flows in the transport network. Congestion effect is involved in
the route travel cost which is the function of the number of travellers. The trip making and
route choice behaviour is written by the Nested Logit model. Those choice probability
functions between an OD pair are given as follows,
exp(6,c[s)
p[k\rs] = — VkEKrs,rstQ (2)
I exp(6jc")
k€K
where p [k\rs] denotes the choice probability of the route k between origin r and destination s,
p [rs] presents the choice probability of generating a trip between OD pair rs and p [ rs } is
the choice probability of cancelling a trip between the OD pair, c" denotes the cost of
travelling the route k between OD pair rs, and S rs represents the expected minimum cost
between the OD pair, which is written as Eq.(5).
(5)
262 Transportation and Traffic Theory
These route cost and the expected minimum cost are the function of flows in a network. Crs
and Rrs are the fixed costs of travelling OD pair and of cancelling the travel, respectively.
6}, 02 are the parameters of route choice and travel choice. Q,Krs are the set of OD pairs
and the set of routes between OD pair rs.
When the number of potential trips of an OD pair is given as 7rs, the number of generated
trips qrs and the number of cancelled trips q rs of the OD pair are written as;
exp{82(Srs
\/rsE Q (6)
exp{  exp(e2RJ
(7)
An optimization model equivalent to the above conditions is formulated as the following Eqs.
(9) to (13).
[NLSUEVD]
J k
min.Z(f,q, q ) =
(9)
subject to
(10)
Road Network Under Degraded Conditions 263
VI ^.75 y , , jy —^ / 1 1 \
where ta(w) denotes the link cost function, and xa is the flow of a link. We assume the link
cost function is separable and monotonically increasing for its link flow.
Let us prove that the first order necessity conditions of the above nonlinear optimization
model [NLSUEVD] are equivalent to the Nested Logit model. Introducing the Lagrange
multipliers {A ri } and {/j.rs}, the Lagrangianof the [NLSUEVD] can be written as;
BL 8L
qrs =0 and > 0 V rs E Q (16)
The constraints of the original problem Eqs.(lO) to ( 1 3 ) are also hold. Since the route flow
and the OD flow are positive, the above conditions can be rewritten as;
264 Transportation and Traffic Theory
dL rs 1
In (18)
dL 1 1
In (19)
In +1 (20)
When this equation is put into the route flow conservation constraint Eq.(l 1), we obtain the
Logit model for route choice. Similarly, the OD flow is calculated using Eqs.(19), (20) and
(10). Then, we obtain the Binary Logit equation for trip making. These are the Nested Logit
equations which we have assumed for the traveller's choice behaviour in a network. Thus, the
optimal conditions of the [NLSUEVD] is proved equivalent to the stochastic route and trip
making choice models. The objective function of [NLSUEVD] is strictly convex for path flow
and OD flow, and feasible region is convex as well. I f w e eliminate cyclic paths, the feasible
region of path flow is bounded. Thus, the solution of [NLSUEVD] is unique for path flow
and OD flow.
The conventional SUE models assumed probability distribution of perceived travel times of
drivers in a deterministic network. This includes both the logit based SUE model by Fisk
(1980) and the probit based SUE model by Sheffi and Powell(1981). Although the probit
based SUE model overcomes the deficiency of the logit based model, it still remains
computational inefficiency for solving probit based SUE model. Mirchandani and Soroush
(1987) proposed a general SUE model in a stochastic network. Some limited cases are
approximately solved and consistent extension to variable demand seems difficult. The
formulated [NLSUEVD] is categorized into the logit based SUE models, and it is not
sufficient to capture the similarity between different routes "Independent from irrelevant
alternatives" property of the logit model may cause overloading to the overlapping paths in a
network. Nevertheless, the [NLSUEVD] seems useful since the route choice and the trip
Road Network Under Degraded Conditions 265
Applying the Simplicial Decomposition method, the formulated NLSUEVD can be solved.
Two phases are iterated; the phase of solving the Restricted Master problem for given path
set and the Column Generation phase of extending the path set. The iterative procedure for
the Restricted Master problem and the Column Generation problem can be summarized as
follows. The route flow vector and the OD flow vector are referred as /"={/"} and
q = (qrs, q rs j, respectively.
[Phase 0] Initialization. Set initial path set JA"M. Calculate initial feasible network flows
[Phase 1] Restricted Master problem. Solve the Restricted Master problem for the given path
set {.K™ ] and obtain network flows \f,q } .
[Phase 2] Column Generation problem. Examine the possibility of path set extension. If no
additional route is found, close the iteration. Otherwise, update the path set (K™*1 } and set
iteration counter m=m+l, and go back to [Phase 1]
For finding the additional route in the [Phase 2] , the shortest path is calculated for the link
cost with loading route flow vector f . If the shortest path is new and not included in the
current path set /C™ ], add the path to the path set.
The Restricted Master problem in the [Phase 1] can be calculated using partial linearization
method. The algorithm for the Restricted Master problem is as follows,
<Step.O> Initialization. Initial feasible network flows \f°, ^°}were calculated for given path
set { K r s } . Set iteration counter n=0.
<Step.l> Travel cost update. Update link flow xa= £ L o"kf^ and link travel cost
rs:Slk€K
266 Transportation and Traffic Theory
ta=ta(xa) for all links in a network. Then, calculate route cost c™ = £ S"k ta and the
a €A
<Step.2> Direction search. Calculate auxiliary OD flows v= (vrs, v \ and route flows
g = g™ j using following equations.
exp{62(S „ + €„)}
v,, = T.
exp{  82(Srs + CJ} + exp(62R J
exp(6Icrks)
8k " vrs
£ exp(8jc")
k€K
<Step.4> Flow update. Update the route flow vector and the OD flow vector using the
optimal value aopt.
<Step.5> Convergence check. If the difference of the flow vectors /'!+J /" +
is sufficiently small, then the flow vectors are converged and close the Restricted Master
problem. Otherwise, set n=n+l and return to <Step.l>.
Road Network Under Degraded Conditions 267
Solving [NLSUEVD] for the network at state x, we obtain the flow variables as well as travel
cost in the network. The equilibrium OD flow {qrs(x)} and the equilibrium route flow
\f"(x) } are the representative flow variables, and can be used for evaluating the performance
of a deteriorated network. In addition to those flow variables, the equilibrium route travel
cost {c"(x) j and the expected minimum travel cost between OD pair {Srs(x)} are also
available for evaluation.
The OD travel demand may decrease when some links in a network become deteriorated.
Thus, the simplest performance measure is the degree of reduction of OD travel demand from
the normal network state, that is, the ratio of the OD travel demand of a state x to that of the
normal state x 0. This is denoted by qrs(x) lqrs(x0). This measure is convenient to compare
the performance between different OD pairs. When it is necessary to evaluate the reduction
rate for an entire network, the reduction rate of total travel demand £ qrs(x)l £ qrs(xo)can
rs€n rsttt
The other performance measure is the travel cost of an OD pair. Using route travel cost
c"(x) and the route choice probability p [ k \ r s ] , the averaged travel cost of an OD pair is
represented as,
exp(6,c"(x))
c[s(x)p(k\rS) = £ c[s(x) (22)
exp(
The expected minimum travel cost between an OD pair Srs(x) will be the alternative
composite measure using route travel cost. As well as the performance measure using flow
variables, it is possible to define the ratio of the cost based performance measures, for
example, the increase ratio of the expected minimum cost such as S r s ( x ) / S r s ( x 0 ) .
We will define the reliability of a flow network using a performance measure. A reliable
268 Transportation and Traffic Theory
transport network generally means the network in which one can travel from his/her origin to
the destination without much uncertainty. The state of a network is probabilistic and the
performance measures are also random variables. Therefore, we define the reliability as the
probability of whether a performance measure is sustained within an acceptable level. The
probability is written as:
PM(x) denotes the value of a performance measure at a network state x . The flow based
performance measures are usually decreasing since the ratio will become smaller for worse
network state. The cost based performance measure is increasing to the contrary. Parameter c
denotes an acceptable level of the performance measure. The value of c is exogenously
determined considering the level of service which should be maintained even in deteriorated
situations.
When the ratio of OD flow, qrs(x) I qrs(x 0 ) , is used as a performance measure, Rrs(c)
becomes the OD flow reliability. This means the probability of whether the travel between
the OD pair is possible within an acceptable reduction rate c (0 < c < 1 ). In other words, the
OD pair is regarded connected in condition that (1c) of the travellers cancel their trips. When
the value of c is set equal to 1, any reduction of OD flow is not permitted and the highest
level of service is required.
The reliability is numerically evaluated using the expected value of the operated/failure
function, which determines whether the performance measure is within the given level. Taking
a decreasing performance measure like OD flow rate as for example, the operated/failure
function is written as,
(1 if PM(x)>c
Z(c,x)=
v
' ' L\0 if., PM(x)
p,,, , <c (24)
^ '
Note that the subscripts rs are dropped in order to avoid complexity. The probability R(c) is
the mathematical expectation of Z(c,x) weighted by the state probability p(x), which is
written as,
Road Network Under Degraded Conditions 269
When we evaluate the reliability measure for different criteria c's using the approach above, it
is necessary to calculate the operated/failure function for each value of c. Even if we use an
approximation algorithm, this is time consuming work. Here, we will show an alternative
approach for evaluating reliability. The occurrence of a state is stochastic and performance
measures are randomly distributed. If we could estimate the cumulative distribution function
of a performance measure Frs [PM ], the probability is easily calculated for any values of c,
such as,
Rrs(c)=Frs[PM>c] (26)
The cumulative distribution function is also approximated using similar methods for
approximating the expectation of the operated/failure function. This is explained in the next
section as well.
Approximation Algorithm
For a network with L links, the number of possible state amounts to 2L. If [NLSUEVD] is
calculated for each network state, the direct calculation of the expected value of reliability
using Eq.(25) requires huge computation cost. This is also true for estimating cumulative
distribution function. In this section, we show two algorithms for approximating those
equations. The original idea was presented by Li and Silvester (1984). The algorithm defines
the lower and upper bounds using the J most probable state vectors.
where xj denotes the jth most probable state vector, p(x.) represents the state probability
for the state x f and N is the number of all state vectors. Using the state vectors by the Jth
most probable state vector jc;, ,jc y ) and corresponding the values of the operated/failure
function{Z rs (c,jc y/ );j = 1,... , J } , the upper and lower bounds of the expected value can be
defined as follows. The upper bound is obtained through the optimistic expectation of the
270 Transportation and Traffic Theory
operated/failure function. If we assume that the states jc . for j=J+l to N are equivalent to the
normal state xa, the performance measure for those states are also equivalent to that of the
normal state, namely PMrs(Xj)=PMrs(x 0) for j=J+l to N. This means the values of the
operated/failure function Zrs(c,Xj) are equal to 1 for j=J+l to N. Thus, the upper bound of
R^s(c) is obtained as the expected value of Zrs(c,x) of these conditions. That is,
I p(Xj)Zn(c,x0)
On the other hand, the lower bound is obtained through the pessimistic expectation of the
operated/failure function. Assuming that the states X j for j=J+l to N are equivalent to the
worst state x w, the performance measure for those states are also equivalent to that of the
worst state; namely PMrs(Xj)=PMrs(x w) for j=J+l to N. For example, the OD flow of the
worst state qrs(x w) will probably be zero or extremely smaller than that of the normal state
qrs(x o ) . Since the ratio of the two qrs(x w)/q rs(x a) is also very small, the value of the
operated/failure function for the worst state Zrs(c,xw) is equal to 0 for any criterion c,
Accordingly, the values of the operated/failure function Zrs(c, Xj) are equal to 0 for j=J+l to
N. Thus, the lower bound of RLrs(c) is obtained as the expected value of Zrs(c,x) of these
conditions. That is,
(29)
The expected value of Rrs(c) stays between Rvrs(c) and RLrs(c). Comparing the values of the
upper and the lower bounds of the Jth iteration with those of the (J+l)th iteration, we
obtain the folio wings:
RLrs(c,Xj)<RLrs(c,XjJ
Road Network Under Degraded Conditions 271
This means the upper bound and the lower bound converge to the exact expected value of
Rrs(c) from the upper side and the lower side, respectively. Figure 2 shows the convergence
image of the upper and the lower bounds. We take the next most probable state vectors one
after another and update the approximated expected value of the reliability measure Rrs(c)
until the difference between the upper and lower bounds becomes small enough.
R(c)
R*(c)
Lower Bound
Approx. Value
N
1 2 Number of iteration J
<Algorithm>
Step.O Calculate the flow and cost variables by solving the [NLSUEVD] for the normal
state xo .Set iteration counter J=l.
Step. 1 Take the Jth most probable state vector x y , where the probability p(xj) is the Jth
largest. An efficient algorithm proposed by Lam and Li (1986).
Step.2 Calculate the performance measure PMrs(Xj) by solving [NLSUEVD] for state Xj.
1 if PMrs(Xj)>c
(31)
0 if PMrs(Xj)<c
272 Transportation and Traffic Theory
Step. 4 Calculate the upper and lower bounds using Eq.(28) and Eq.(29), respectively.
Step. 5 Check the convergence. If the difference of the upper and the lower bounds is small
enough, go to Step. 6. Otherwise, set J=J+1 and return to Step. 2.
c)l2 (32)
This algorithm seems efficient for evaluating the reliability between OD pairs rs (rs e £2) for
given criteria c. However, one must calculate the number of iterations again if it is necessary to
obtain the expected value o f R n ( c ' ) for different criteria c'.
where H(t) denotes the subset of [1,...,J], for which performance measure PMrs(Xj) is greater
than or equal to t. The difference between the upper and the lower bounds is:
j
F"(t) !*,(*) = lZp(xj) (34)
ji
which is kept constant for any range oft. This is convenient for examining the convergence of
the iteration.
When we take the next most probable state vector xj+1, the upper bound is lowered and the
Road Network Under Degraded Conditions 273
F",.j(t)*F?,,J+1(t)
FrsJ(t) and FrsJ(t) denote the upper and the lower bounds for the Jth iteration.
FrSiJ+](t) and FrSiJ+1(t) represent the upper and the lower bounds for the Jth iteration,
respectively. This means that the difference of the upper and the lower bounds becomes
small enough when a sufficient number of the probable state vectors are included. The
cumulative distribution function of performance measure Frs(t) is then approximated as:
F
rS (*) = {F",(t) + FLrs (t)} 12 for all t> 0 (36)
Once the distribution function is estimated, we can evaluate the value of Rrs(c) using Eq. (26)
for any criteria of c. This is quite useful for analyzing network reliability measures since it is
not necessary to consume additional computational cost for different criteria. The algorithm is
similar to the one for approximating the expected value of Rrs(c). That is summarized as
follows:
<Algorithm>
Step.O Calculate [NLSUEVD] for the normal network state. Set iteration counter J=l.
Step. 2 Calculate the performance measure by solving the NLSUE problem for the Jth most
probable state vector.
Step. 3 Evaluate the upper and the lower bounds using Eq.(33) for all ranges of the
perfomance measure.
Step. 4 Go to Step. 5 if the difference of the upper and lower bounds is small enough.
Othewise, set J=J+1 and return to Step. 2.
NUMERICAL EXAMPLES
In^ut Conditions
Figure 3 presents a small scale network with 4 nodes 5 directed links. Linear link cost
functions are also shown in the same figure. A pair of origin and destination flow is assumed
from node 1 to node 2. In the following part of the numerical examples, the subscript of
origin and destination pair is omitted since only one OD pair is concerned. The maximum
number of paths is 3 in the full network., and those paths are identified as shown in Figure 4.
(tj10+Xj
path1
path3
The upper limit of OD flow T equals to 40. The fixed travel cost C and the cost R for
cancelling the travel of the OD pair are assumed 0 and 50, respectively. The parameter values
0j, 02 are 0.1 and 0.2, respectively. Using these conditions for OD demand and route choice,
the OD flow q and the path flow/^ (k = l,2 and 3) are written as follows;
exp(02(C exp(0.2S)
40
exp(  62R) exp(0.2S) + exp(0.2x 50)
Road Network Under Degraded Conditions 275
where S denotes the expected cost between the OD pair and written as;
The values of link connectivity pa (for a= 1,2,3,4 and5) are assumed 0.9, 0.8, 0.7, 0.9 and
0.8. When the occurrance of link closure is independent each other, the probability for
X 1 X
network state is given by p(x) = Y[pa °(lpa) '• F°r example, the state probability for
Before evaluating network performance measures under degraded network conditions, we will
show the results of the network flow analysis. The number of links is 5, and the number of
the states is 25=32. These states are aggregated into 8 network patterns with the common set
of paths. Table 1 presents the 8 patterns and corresponding state probabilities. When the OD
pair 12 is not connected, the state is categorized into "disconnected". [NLSUEVD] is
calculated for each network state. The results are summarized in the Table 1.
It is possible to load 36.67 of OD flow onto the full network (PI) without any disconnected
links. This amount of flow is equal to 92% of the given upper limit of the potential OD
demand. For this normal state of the network, the largest proportion of the OD flow uses the
path3, and the second largest is the path1. The path2 will be used very few, since the cost
of the path2 is the highest among three possible paths.
Comparing the network patterns with some disconnected links, it is found that the
performance of the network becomes worse when link 5 is not available. If only link 5 is
removed (P2) from the full network, the amount of the OD flow loaded onto the network is
20.67. This corresponds to 56% of the OD flow for the full network. On the other hand, the
influence of removing link 4 is not so large. Just 1% of OD demand is reduced if link 4 is
disconnected (P3). Similarly, the closure of link 1(P5) results in the 10% reduction of OD
flow of the full network.
276 Transportation and Traffic Theory
Table 1 Network State Probability and SUE Flows for Network Patterns
Network Network State OD Flow Path Flow Path Cost
Pattern Probability (q) &«) (Ci,C2,C3)
P6 0.0101 8.33 0
8.33 59.99
0
"N/" 0

P7 0.1155 29.52 0
O 0 
Disconnected 0.1141 0 0 
When only one path is available (P4,P6 and P7), the performance of the network becomes
decreased. In particular, either path1 or path2 can carry 2040% of OD demand of the full
network. However, path3 is still capable to maintain 80% of the OD flows (P7). Although
the SUE calculation is executed just for a small scale toy network, it would be useful to find
the paths with higher performance under degraded conditions.
Road Network Under Degraded Conditions 111
Figure 5 shows the convergence of the upper and the lower bounds of the expected value of
the OD flow. From the previous Table 1, the exact value of the expected value of the OD flow
is calculated as;
Y J p ( X ) q ( X ) = 25.078,
where q(x .) denotes the OD flow from node 1 to 2 for the state vector jc y . Using the Jth
most probable state vector, the upper and lower bounds of the expected value can be
approximated. It is found that the difference of the upper and the bounds becomes smaller as
the number of the probable state vectors increases. This figure proves the convergence of
approximation algorithm for the upper and the lower bounds, Eb [q] and EL [q], given by the
following equations.
OD Flow
40
\ Lower Bound
Upper Bound
V Expected
30"
00
"ticiAr,
20"
10
11 16 21 26 31
N. of iteration
Figure 5 Convergence of the Upper & Lower Bounds of OD Flow and the Expected Value
The values of the upper and the lower bounds approximated using up to the 6th most
probable state are 31.20 and 23.14, respectively. The averaged value of the upper and the
lower bounds is 27.17, and the approximation error is 8.3%. When the approximation goes to
the 12th most probable state vector, the upper and the lower bounds become 26.92 and
278 Transportation and Traffic Theory
24.19, respectively. The averaged value of these two bounds is 25.56, and only 1.9% of the
approximation error remains at the 12th iteration.
The upper and the lower bounds of the cumulative distribution function of the OD flow,
Fu(q), and FL(q), are approximated as follows;
j
FU(q)= I pfrJ + V
jtH(q)
L
F (q)= £ p(xj)
H(q) where denotes the subset of [1,...,J], for which OD flow is less than or equal to q.
1.0
0.8
Lower Bound
Upper Bound
O.6
0.4
n—B
,S
J,
0.2
0.0°
0 10 20 30 40
Figure 6 Upper and Lower Bounds of OD Flow Distribution (upto 6th probable states)
i.o
Lower Bound
Upp er B ound
0.2:
0.0
Figure 7 Upper and Lower Bounds of OD Flow Distribution (upto 12th probable states)
Road Network Under Degraded Conditions 279
Figure 6 shows the upper and the lower bounds of the cumulative distribution function of the
OD flow at the 6th iteration. The difference between the upper and the lower bounds is
nearly equal to 0.220 and may not be converged enough. Figure 7 shows the approximated
function for the 12th iteration, in which the difference of the upper and lower bounds reaches
0.074. The difference may not be satisfactory small. However, the distribution function can
be well approximated since it is given by the average of the upper and the lower bounds.
CONCLUSION
Flows in a transport network are not stable even if all links are in service. When several links
are deteriorated by some reasons, the fluctuation of network flows will be magnified. The
uncertainty and inconvenience of travel may bring the reduction of travel demand. It is
necessary to consider those aspects of network flows when we evaluate the performance of
the transport network. This paper aims to show a network flow model in degraded
conditions, and then to propose performance measures incorporating flows in the network
with some links being closed to traffic.
Solving [NLSUEVD] for a state of the network, we obtain the flow variables as well as travel
cost in the network. Those flow and cost variables can be used for evaluating the performance
of a deteriorated network. The simplest performance measure is the degree of reduction of OD
travel demand from the normal network state, that is, the ratio of the OD travel demand of a
state to that of the normal state. When it is necessary to evaluate the reduction rate for an
entire network, the reduction rate of total travel demand can be used as a performance
measure.
Then, the reliability of a flow network was shown using a performance measure. It is defined
as the probability of whether a performance measure is sustained within an acceptable level.
For a large scale network, the number of possible state is enormous. The direct calculation of
the expected value of the reliability requires huge computation cost. We have shown the
algorithm for approximating the upper and the lower bounds of the expected value. The
algorithm is applied to approximate the probaility distribution function of a performance
measure. Through numerical examples in Chapter 4, the convergence of the algorithm was
verified.
As well as being useful for estimating the expected value, the distribution functions will be
widely available to discuss the characteristics of the performance measures. When those
performance measures are compared between different OD pairs, it is possible to find less
reliable OD pairs. The links reducing network performability will be found and the results will
be useful to improve the physical durability of those links. The formulation of the network
design problem from the view of reliability is possible. Although solving the reliability
network design model will be more difficult than usual bilevel network design problems, this
will be an attractive field of extension.
Of course, we have not yet applied the proposed flow model and the performance measures
to actual transport networks. In parallel with this study, the behavioural survey of car drivers
is now on going for studying their travel choice behabiour when some links are closed to
traffic due to heavy rains. The external validity of the methodology will be examined using
those data.
Road Network Under Degraded Conditions 281
REFERENCES
Abstract
This paper presents a methodology for evaluating the reliability of transportation networks.
While tools already exist to determine the expected benefits of travel demand management or
new infrastructure, tools have yet to be developed which take into account disbenefits arising
from randomly occurring disturbances. The paper focuses on the sensitivity of both path travel
times and expected minimum origindestination travel times to normal withinday demand and
supply variation, where demand variation takes the form of perturbations to origindestination
flows and supply variation takes the form of perturbations to link saturation flows. Two extreme
cases are distinguished; one where route choices fully respond to the perturbations,
corresponding to the more major, longerterm incident and the other where route choice does
not respond, corresponding to the more minor, shorterterm incident. A logit assignment model,
referred to as the Path Flow Estimator (PFE), is linearised with respect to the parameters
affected by withinday variation, using sensitivity expressions. As analytically derived
sensitivities are sometimes difficult to calculate for large networks, their approximation by finite
differencing is considered. Results obtained for a large network (5000 links, 8000 OD pairs) in
York are discussed, as well as results obtained for a much smaller network (100 links, 60 OD
pairs) in Leicester.
i. INTRODUCTION
1.1 Background
The potential sources of disruption to transportation networks are numerous, ranging at one
extreme from natural or manmade disasters, which tend to occur rather infrequently, to at the
other extreme events, which occur on a daily basis. The scale, impact, frequency and
predictability of such events will of course vary enormously. While little can be done about
their scale, frequency or predictability, particularly where natural disasters are concerned, it
Transport Operations Research Group, University of Newcastle, Newcastle upon Tyne, NE1 7RU
!
Department of Transportation Engineering, Kyoto University
1
Department of Civil and Structural Engineering, Hong Kong Polytechnic University
284 Transportation and Traffic Theory
This paper describes the theory behind a software tool being developed for the analysis of
network reliability. While many tools exist for studying the impact of new transport
infrastructure or travel demand management measures on traffic flows (for example, the widely
used CONTRAM and SATURN programs), there are no tools for assessing the impact of such
measures on the reliability of transportation networks.
Network reliability has two dimensions. The first relates to the connectivity reliability of a
network. When links fail in unfavourable configurations it may no longer be possible to reach a
given destination from a given origin, in which case the network becomes disconnected.
However, even a connected network may fail to provide an adequate level of service. For
example, random events may cause unacceptable variation in origintodestination travel times,
making it difficult for travellers to arrive at their destinations on schedule. The second
dimension of reliability is therefore the performance reliability of a network.
Previous work on network reliability (Du and Nicholson, 1993; lida and Wakayabashi, 1989)
has focused principally on connectivity in degradable transportation networks. Although
Asakura and Kashiwadani (1991, 1995) have looked at the problem of travel time reliability, the
field of performance reliability is distinctly underresearched.
This paper focuses on the performance reliability of transportation networks in the face of
normal withinday variation of origindestination flows and link properties. It makes use of the
Path Flow Estimator (PFE), a Stochastic User Equilibrium (SUE) assignment model, in order to
estimate the distributions of the variables of interest (flows, travel times). Reliability with
respect to given performance criteria can then be computed on the basis of those distributions.
The eventual objective is to develop a software tool that facilitates the design of robust
transportation networks.
The Path Flow Estimator (PFE) is a flexible traffic assignment tool. It was originally developed
in the DRIVE 2 project MARGOT to estimate path travel times for route guidance systems
given vehicle detector data, traffic signal times, and possibly also travel time data from probe
vehicles. The PFE is based on the notion of a Stochastic User Equilibrium (SUE) between
demand and supply. The demand for the paths is determined by a logit path choice model and
the link costs, which characterise the supply side, are determined primarily by flowdependent
delay formulae. The delay function currently used is the Kimber and Hollis (1979), time
dependent function which allows for oversaturated conditions . The PFE utilises the fact that
under a SUE, in contrast to a deterministic user equilibrium, path flows are uniquely defined at
the equilibrium.
The equilibrium path flows are found by solving an equivalent optimisation problem iteratively.
An outer loop generates paths and an inner loop assigns flows to paths according to the logit
path choice model. The inputs take the form of traffic counts and/or trip tables and the outputs
take the form of estimated path flows and travel times, as well as data which may be derived
Network Reliability Assessment 285
from these, like junction turning proportions, link flows where unmeasured and link delays.
Further information about the PFE can be found on http://www.ncl.ac.uk/~nsg5/PFE/html.
The structure of the algorithm is illustrated in Fig. 1. Interesting features of the algorithm are;
the inclusion in path generation of the "pressure" exerted by active link constraints, and the
allowance for measurement errors through userdefined confidence intervals.
Input: Static network data and dynamic flow and signal control data
Outer loop
Generate paths of "least resistance" taking delay and the pressure
exerted by active link constraints into account
Inner loop
Assign measured flows to paths according to the logit path
choice model taking flowdependent delays into account
Output: Path flows and travel times, turning movements, unmeasured link flows
Fig. 1 : The Path Flow Estimator
The approach to network reliability analysis adopted in this study is based on the sensitivity of
the PFE outputs, in particular path travel times and expected minimum origindestination travel
times, to perturbations of the inputs, in particular the origindestination demands and link
saturation flows. Tobin and Friesz (1988) obtained expressions for the sensitivity of
deterministic user equilibrium link flows to perturbations in origindestination demands and link
travel times. Bell and lida (1997) extended the method to obtain expressions for the sensitivity
of minimum origintodestination travel times to perturbations in origindestination flows and
link travel times. Using the same approach, Bell and lida (1997) have obtained comparable
sensitivity expressions for the PFE.
This paper focuses on normal withinday variations of traffic conditions, like the morning peak
hour for instance. The measures of network performance reliability considered are based on
travel times, since broadly speaking the main function of a transportation network is to carry
travellers to their destinations within acceptable times.
Two measures of potential interest to both network users and planners are:
The first measure is defined as the probability that the travel time on a given path is less than an
acceptable threshold. This is of direct interest to the users, since path travel time reliability is a
factor likely to influence their route choice. The second measure is taken at a more aggregate
level, namely the level of the OriginDestination (OD) pair. The expected minimum OD travel
time is the expected lowest perceived travel time amongst all the relevant paths, allowing for
286 Transportation and Traffic Theory
random variations across users in the perception of travel time. Thus changes of route
undertaken by users to avoid perceived increases in link travel times are taken into account. The
second measure may be more useful for planners than for the network users themselves.
Consider the logit path choice model with perceived path travel time
Ck=g?—j
where g™ is the actual travel time on a path k between an OD pair rs, £ is a random, Gumbel
distributed, perception error term, and 9 is a scaling factor, then the expected perceived
minimum travel time is given by
For both types of reliability measure, an estimation of the distributions of actual (not perceived)
path travel times in the face of normal withinday variations of exogenous parameters in the
demand or the supply side of the transportation system is required.
There are basically two types of variation which affect travel times in a transportation network
on a daily basis. One relates to the demand, characterised by average OD flows during the
period of interest, and the other relates to the supply, for example the saturation flows of the
links in the network. Both types of variation can have a range of causes and impacts in terms of
their scale and duration. Examples of sources of demand and supplyside variation are
presented in Table 1.
For both the demand and the supplyside sources of variation, the distinction between
minor/shortterm and major/longterm impacts serves to indicate whether or not there is likely
to be an induced rerouting. For more minor, shortterm incidents, there may be insufficient
time for information to disseminate to network users, and even where information has
disseminated, it may not be worthwhile for users to respond by changing route.
These sources of fluctuation can be modelled by the use of appropriate discrete or continuous
probability distributions for the OD flows and link saturation flows. To model the impacts the
fluctuations will have on the travel times, and hence estimate path travel time distributions,
Network Reliability Assessment 287
linear relationships between the PFE estimates of path travel times and the sources of random
variation (the OD flows and the link saturation flows) are established. Two approaches are
adopted, depending on the severity and timescale of the perturbations, reflecting whether or not
induced rerouting is expected.
3. PFE SENSITIVITY
3.1 Linearising the PFE
Linearising the PFE allows approximate distributions for model outputs to be obtained from
distributions of model inputs thanks to the conservation of form for certain commonly used
types of distribution. For instance, the summation of independent normal variates has a normal
distribution whose defining moments are straightforward to compute. The linearisation is valid
only for small variations about some point, in this case taken to be the SUE 'solution obtained
for average input values. /
Where the fluctuations may induce rerouting, one extreme case arises where full adjustment to
the fluctuations has occurred. SUE sensitivity analysis may be used to estimate the distributions
of equilibrium path travel times. The other extreme case arises where no rerouting at all occurs.
The fluctuations effect only the supplyside, since the routechoice proportions do not respond.
3.2 Notation
Let:
h = Vector of path flows
v = Vector of link flows
c = Vector of link travel times
g = Vector of path travel times
t = Vector of OriginDestination (OD) flows
s = Vector of link saturation flows
A = Linkpath incidence matrix
B = ODpath incidence matrix
P = Matrix of path choice proportions
a = Dispersion parameter of the logit model.
The PFE gives an estimation of average flows and travel times in a network for some period of
the day. The solution yields flows and travel times consistent with each other by virtue of the
equilibrium principle, and the assignment of flows to paths is governed by the logit model as
follows
with
/7 = exp(a.gy(h,s))/
The relation between path costs and path flows is through the fact that path travel times are the
summation of link travel times, and that link travel times depend on link flows in relation to
their saturation flows through monotonically increasing functions
where / e j indicates that the summation is over all links on path j , and the relation between
link flows and path flows is
where j e / indicates that the summation is over all paths using link i.
This SUE solution can be interpreted as the mean value of a network equilibrium where users
minimise their perceived travel times. Due to random errors in travel time perception, the model
is stochastic and therefore the estimated flows and travel times are also random variables.
However the mean values provided by the SUE are a reasonable estimator of average network
traffic conditions on a particular period of the day, like the morning peak hour. We shall
hereafter refer to the 'base solution' as being the PFE stochastic equilibrium obtained for a
given trip matrix t 0 , considered to be the average trip table, and the vector s0 of average link
saturation flows. From this base we can then look at the effects of random variations in both
demand and saturation flows on path travel times and expected minimum OD travel times
necessary for calculating reliability measures.
Assuming OD flows are normally distributed with known parameters, Asakura and
Kashiwadani (1991) ran a static User Equilibrium (UE) assignment several times, with a
demand sampled from these normal distributions, in order to estimate OD travel time
distributions. A similar approach would be possible for the PFE. However, for large networks
with many OD pairs, this method is computationally prohibitive and also simulation adds an
element of nonreproducibility.
Another approach would be to estimate variances of travel times from the variances of OD trips
by making linear approximations of the relationship between OD flows and the path travel times
obtained from the equilibrium assignment model. This can be done by applying sensitivity
analysis to the PFE model.
When f(h) is minimised subject to t = Bh and h > 0, the Kuhn Karush Tucker optimality
conditions are
In h* = ag*  BTu*
It can be shown that this implies the logit path choice model
Vf(h*) + BT u* = 0
V 2 f(h*) Ah + BT Au = aAg
r«Agl ~H B T TAhl
[At J ~ B 0 J_AuJ
TAhlJj,, J12TaAgl
Au J
L J L 2i J22iAt J
where
At optimality In h* + ag* + BTu* = 0 implying hj = exp( agj 1^). From this logit model
Hence u = In t + az, so
Var(z) = [dz/dss = s,] Var(s) [dz/dss = s,]T + [dz/dtt=t»] Var(t) [dz/dtt = t*]T
Var(h) = [dh/dss = s,] Var(s) [dh/dss = S*]T + [dh/dtt = t,] Var(t) [dh/dtt = t*]T
Gradients and variances for path travel times rather than path flows may be obtained from the
gradients of path travel times to path flows, via the link travel time Jacobian.
Calculating those expressions requires extensive matrix operations (like inversion), which for
large networks are thought to be unmanageable. Therefore it may be preferable to approximate
these gradients by finite differencing
t = t,]*[Az/Att = t*]
s = s,]«[Ah/Ass =s,]
To obtain finite differences one needs first to run the PFE program with the trip table t 0 to
obtain a base solution in all variables of interest . Then the model has to be run sequentially
after applying a small change A/y (say 1 vehicle/hr) to each OD pair j , one at a time. To reduce
the number of iterations within the PFE at each run, the initial solution is taken to be the base
solution, which is assumed to be not so distant from the perturbed solutions. Thus after each run
j , the differences with respect to the base solution in terms of path travel times can be
calculated. Those differences divided by the htj provide the finite differences approximations
to the partial derivatives needed.
If we assume that there is no rerouting induced by the fluctuations of demand and saturation
flow, then it is no longer appropriate to consider expected minimum OD travel time. Instead we
can look at distributions of path travel times around the equilibrium caused by the demand and
supplyside perturbations. The following gradients reflect the corresponding sensitivities
Network Reliability Assessment 291
4. EXAMPLES
Network data were provided by York City Council (YCC) in SATURN (a simulation and
assignment model) format, as part of a European project AIUTO, which looked at how to model
the effects of various travel demand management measures. The SATURN model was based
upon one constructed by YCC in 1992 and extensively updated by them in 1996.
The nodebased data from SATURN had to be translated into a linkbased format suitable for
the PFE. Also the morning peak hour trip matrix, obtained from a survey with real number
entries, was transformed into an integer trip matrix by a rounding operation that sought to
maintain approximately the same column and row totals as in the original real valued matrix.
The resulting PFE network thus consists of about 2000 links, 8635 OriginDestination (OD)
pairs, and includes such necessary information as link freeflow average speed, link saturation
flows, as well as the signal times. A representation of the modelled network is shown in Fig. 2.
The base SUE solution was obtained by running the PFE on the York network with the average
morning peak hour trip matrix. Fig. 3 shows the network with the congested links highlighted
(links with the ratio of flow over saturation flow greater than 0.9 were defined as congested).
292 Transportation and Traffic Theory
It was proposed in this study to obtain morning peak hour variances of both equilibrium and
nonequilibrium path travel times. The source of variation was taken to be the trip table (OD
flows). For equilibrium path travel times we used sensitivity of the PFE solution about the base
solution calculated previously. Since applying analytical sensitivity expressions was not
practically possible for such a large network, it was decided to approximate the derivatives of
equilibrium path travel times with respect to OD flow variation by finite differences. This was
carried out by simply changing each OD flow tt , one at a time, by one unit of flow, and then
calculating a new PFE solution trip matrix thus perturbed. The difference in path travel times
with respect to the base solution give the approximate derivatives
This meant running the PFE 8635 times (the number of OD pairs) with a differently perturbed
OD matrix each time. Eventually the outcome would yield an estimation of the equilibrium
path travel time variances, through the linearisation of the relationship between equilibrium path
travel times and OD flows discussed in the previous section
i<=OD pair
Network Reliability Assessment 293
What is needed as input are the variances of each OD flow. Due to lack of empirical statistical
data, some assumptions about the distributions of OD flows were required. For simplicity, we
used a normal distribution for each OD flow centred around tf , the base solution. This implied
a normal distribution as output for the path travel times.
To have a range of OD flows with negligible negative values, and given that the majority of all
OD pairs had flows of the order of 1 veh/hr, it seemed quite reasonable to take a standard
deviation a at a third of the mean value // . So for each OD pair i , the variance of flow was
thus assumed to be
The finite differences method was applied to obtain variances of SUE path travel times. Four
days of computational time were necessary in order to compute the required variances, despite
the fact that each of the 8635 PFE runs were started from the base solution, expected to be not
too distant from the perturbed solutions. We also obtained the variances for nonequilibrium
path travel times using finite differences for proper comparison, even though the variances
could have been computed analytically as shown in the previous section.
With the variances thus calculated, reliability measures could be computed for each path. The
performance criteria, or the travel time threshold set for reliability, was arbitrarily taken at 1 1 0%
of the mean path travel time.
Thus for each path j the travel time reliability Rj , defined as the probability that the travel time
will be less than 1 10% of the mean travel time g° , was calculated as follows:
/Var[g,]'
distributed variable x with mean // and variance a1. The values of O are obtained by
tabulation.
Given the huge number of paths in this network (around 23000), a thorough analysis of all the
results was not really presentable here. For illustrative purposes, we chose a particular OD pair
with two clear cut alternatives to analyse the results and illustrate the differences between the
equilibrium and nonequilibrium methodology for path travel time reliability. Fig. 4 shows the
two paths (with the arrows) on a zoomed part of the network (in the centre of town), along with
the congested links (highlighted). The table below indicates the average path travel times with
the values of the standard deviations and reliability for both equilibrium (full route adjustment)
and nonequilibrium (no route adjustment) distributions.
294 Transportation and Traffic Theory
We can see that for both paths the standard deviation increase quite significantly when
readjustment to equilibrium is not considered in the calculation. This suggests that the effects of
drivers rerouting is to lessen to impacts of variations.
We can also notice that path 2 experiences hardly any significant variations (whether in the
equilibrium case or not) compared to path 1. By looking at the map, the explanation is clear;
path 2 does not include any congested links whereas path has two of them. So fluctuations of
demand on those congested links is bound to produce greater changes in travel times than on the
links that are less congested in the base solution. This is a feature that actually appears for all
the most unreliable paths in the network; they all contain one or several congested links.
While this appears sensible, it still prompts a question about the usefulness of conducting such a
timeconsuming calculation; the most unreliable paths could indeed have been identified before
the calculations of the variances just by finding out the paths containing congested links in the
base solution.
This limitation comes from the linear approximation used in both methods, which is only valid
in the vicinity of the base solution. Therefore it would be useful to account for larger variations,
Network Reliability Assessment 295
which tend to happen in reality anyway. However to model the impacts of these larger variations
the nonlinear effects need to be included.
To illustrate the points made in the previous example further, the methodology was applied to a
much smaller network , thereby allowing a more complete overview of the results to be
presented in the scope of this paper.
The network considered here is a small part of an urban network in Leicester , England. It
consists of 103 links (including microlinks at junctions), 9 origins and 9 destinations . A
representation of the network is shown in Fig. 5. The topological data (including signal
timings), along with a trip table for a peak period were made available for a Phd project on OD
estimation based on SCOOT traffic counts. We didn't use the link detector data here.
O signalised
junction
f • —•^
.
unsignalised
junction
m
*.._. 1 1 d«J d«_.
»
iin
' !,•<: NK.222
^222. C
N00224 )22^ (* NM2« ) 22.,
4.2.2 Computation of path time variances and reliability measures with respect to OD flow
fluctuations.
We used the PFE model on the Leicester network with the trip table available, and applied the
sensitivity expressions defined in chapter 3, with and without route response. It was here
possible to compute the sensitivity expressions in the full (equilibrium) route response case,
296 Transportation and Traffic Theory
directly without using finite differences, given the small size of the network (typically the
number of paths generated was of the order of 100 veh/hr).
As in the York case we assumed normal distributions of OD flows centred around the trip table
values t°, and again chose a standard deviation of (t° 13), to be consistent with the previous
example. However since the order of magnitude of most OD flow was here much larger than in
the York case , we put a limit of 30 veh/hr on the standard deviation, which corresponds to the
largest deviation in the York case. This was to avoid very large deviations, which given the first
order approximation used in the method, would probably lead to unrealistically large variances
in the equilibrium travel times. Note that a deviation of 30 veh/hr is not strictly speaking small
enough to warrant a valid use of the first order approximation for the relation between the
equilibrium solution and the OD flows. However it is conjectured that for moderately congested
network, the secondorder derivatives of the equilibrium flows will be sufficiently small so as to
make the second order terms negligible.
The following table shows the time deviations and reliability (defined as in the York example)
results obtained for the 20 most 'unreliable' paths, in the equilibrium and nonequilibrium
variance evaluation case.
As for the two paths in the York example, we can see here that, for all the paths, the standard
deviation is greater, and thus the reliability lower, when equilibrium route response is not taken
into account. We also highlighted two links , 215L and 215K, which are included in all the most
unreliable paths, and those two links happen to be the only oversaturated links in the whole
network This confirms the relation between congestion in the average equilibrium situation and
high variations in travel times and again points to the limitation of the firstorder approximation
in gaining more useful insights.
The next table shows the standard deviation of the expected minimum OD travel time, for the
20 most 'unreliable OD pairs:
We can see that all OD pairs connecting DES_B have a substantially higher variability in travel
time compared to all other pairs. A quick look at the network shows why: To get to DES_B, it is
impossible to avoid either link 215L or link 215K. The same could be said for DES_A on the
evidence of Fig 5; however in reality link 215K is divided into 2 sublinks, and it is only the
rightturn sublink leading to DES_B that is actually oversaturated.
s. FUTURE WORK
Including nonlinear effects is a problem in the sense that the form of the input distributions
(demand, capacity) is no longer conserved in the output (travel time) distributions. We present
here briefly the problem of attempting to derive path travel times distributions knowing the
distributions of OD flows , and keeping the pathchoice proportions fixed (thus restricting the
nonlinearity to the supply side only).
We shall concentrate on the link travel times c,(v,(h),^ ; ). Suppose, without loss of generality,
the demand t to be normally distributed around t°, the base trip table.
298 Transportation and Traffic Theory
Then assuming the path choice proportion matrix fixed at the base solution, P0 , the link flow
v, (h) also becomes a normally distributed random variable, centred around v,° , since there is a
linear relationship:
v
76i W 7a
and
Var[v,]=
provided the fw are independent, (a condition that is not necessarily satisfied, but which we will
assume holds here as we did in the previous calculations).
Thus the travel time ct (v,, st) is also a random variable, being a function of a random variable,
though it is not normal, since the link cost function is nonlinear. However the expected value
and the density function can be calculated, using the density function of v^, fv
where the inverse function c~'(x) exists, since the link cost function is monotonically increasing
as a function of link flow v only.
The density function fc (x) can then be obtained by taking the derivative of F ( x ) .
For reliability studies, we are mainly interested in path travel times, which being summations of
random link travel times, are also random variables whose cumulative probability function is
just what is needed to obtain the required measures of reliability . The probability of a trip cost
on a path j being close enough to the 'normal situation' (base solution) value serves as a good
performance measure and will be given by Fg (gj + s). The problem is thus how to compute
V
Suppose , for the sake of clarity, that a path X is composed of link 1 and 2.
Then we have:
Network Reliability Assessment 299
z+y<x
If Cj and c, were independent variables, Fgx (jc) would then be quite easy compute, since in
that case:
The fc (x) , density functions for the link costs, had been calculated previously. So Fg (x) may
be calculated relatively easily ( at least in this case of two variates ).
= z+y<x
IJ
However, in the case of fluctuating demand, c, and c, will not in general be independent (they
will be affected by some common flows). Thus the joint density function fc c (z,y) will not be
easy to compute, especially for paths containing many correlated links. Some approximation
methods might be needed. Furthermore, even if we supposed that the travel times on all the
links were independent then we will be faced for large realistic networks with the nearly
impossible task of multiple integrals. Thus it appears that further research in heuristics to a)
include the correlation between the numerous random variates involved in the determination of
path travel times and b) to integrate the density functions of those variates is necessary and
justified.
6. CONCLUSIONS
We have presented a methodology for obtaining some performance reliability measures for a
transportation network. The work is essentially centred around estimating typical distributions
of path travel times resulting from random fluctuations of exogenous factors. Having identified
those factors, a method based on sensitivity (linearisation) of a SUE model, with two different
options depending on whether rerouting is considered or not. We showed some results for a
large realistic network and also a smaller one to illustrate the differences between the two
options and noticed that paths subject to bigger variations in time are those containing links that
are congested in the average situation. We concluded by indicating that more useful results
might be obtained if nonlinear effects were taken into account so as to allow a wider range of
fluctuations to be considered.
300 Transportation and Traffic Theory
REFERENCES
Asakura, Y and Kashiwadani, M(1991). Road Network Reliability Caused by Daily Fluctuation
of Traffic Flow. Proceedings of the J91 PTRC Summer Annual Meeting in Brighton, Seminar
G, 7384.
Asakura, Y and Kashiwadani, M (1992). Road Network Reliability Measures Based on Statistic
Estimation of DaytoDay Fluctuation of Link Traffic. Proceedings of the 6th World Conference
on Transport Research, Lyon, France, June.
Asakura, Y and Kashiwadani, M (1995). Traffic Assignment in a Road Network with Degraded
Links by Natural Disasters. Journal of the Eastern Asia for Transport Studies, Vol.1, No.3,
11351152.
Bell, M.G.H. , Lam, W and lida, Y (1996), A TimeDependent Multiclass Path Flow Estimator.
Proceedings of the 13th International Symposium on Transportation and Traffic Theory, Lyon,
France, July.
Bell, M.G.H. and lida, Y (1997), Transportation Network Analysis, Wiley, England.
Du, Z.P. and Nicholson, A.J. (1993). Degradable Transportation Systems Performance,
Sensitivity and Reliability Analysis. Research Report, No. 938, Department of Civil
Engineering, University of Canterbury, Christchurch, New Zealand.
Du, Z.P. and Nicholson, A.J. (1997). Degradable Transportation Systems: Sensitivity and
Reliability Analysis. Transportation Research B, 31, No 3, 225237.
Kimber, R.M. and Hollis, E.M. (1979). Traffic queues and delays at road junctions.
TRRL Laboratory Report 909, Transport and Road Research laboratory, Crowthorne, England.
Tobin, S.J. and Friesz T.L. (1988). Sensitivity Analysis for equilibrium network flow.
Transportation Science, 22, 242250.
A Paradox For A Dynamic Equilibrium Assignment 301
Takashi Akamatsu
Masao Kuwahara
ABSTRACT
This paper demonstrates that the capacity increasing paradox in a transportation networks
as in Braess(1968) does also occur under nonstationary settings, in particular, under
dynamic traffic assignment with endogenous timevarying origindestination (OD)
demands. Through the analyses, the analytical formulae for the solutions of the dynamic
equilibrium assignment are explicitly derived for two kind of networks: the networks
with a onetomany OD pattern and the reversed networks with a manytoone OD pattern;
the formulae clarify the significant difference in the properties of the two dynamic flow
patterns. This also leads us to the findings that one of the crucial conditions that
determine whether the paradox occurs or not is the OD pattern of the underlying
networks.
1. INTRODUCTION
particular condition for the network with a onetomany OD pattern, while the corresponding
paradox always arises for the reversed network with a manytoone OD pattern. We then show that
the same results also hold for the model with endogenous OD demand. Finally, the last chapter
summarizes the results and remarks on some further research topics.
2.1 Networks
assignment, which is a natural extension of the static user equilibrium assignment; the DUE is
defined as the state where no user can reduce his/her travel time by changing his/her route
unilaterally for an arbitrary time period.
Under the DUE state, the users who depart their origin at the same time, regardless of their
routes, have the same arrival time at any node that is commonly passed through on the way to their
destination. Furthermore, under the DUE state, the order of departure from the origin must be
kept at any node through destinations. From these property, we can define the unique equilibrium
arrival time at each node for each departure time from the origin.
As defined in the previous section, link travel time ctj(f) depends only on the vehicles which
arrived at the link before time /. Therefore, together with the above discussion on the order of
arrivals at a node, it is concluded that the travel time experienced by the vehicle that departs from
an origin at time s is independent of the flows of the vehicles that depart from the origin after time s.
Consequently, we can consider the assignment sequentially in the order of departure from the
single origin. That is, the assignment can be decomposed with respect to the departure time from
the single origin provided that the OD pattern is onetomany. Similarly, for a manytoone OD
pattern, we can easily conclude that the assignment can be decomposed with respect to the arrival
time at the single destination. For the detailed discussions on this property, see Kuwahara and
Akamatsu (1993) and Akamatsu and Kuwahara (1994).
We will first show the formulation for Enet and derive the solution in 3.1; and then the
formulation and the solution for Mnet will be examined in 3.2.
(1) Formulation
The DUE assignment on a network with a onetomany OD pattern can be decomposed with
respect to the origin departuretime as mentioned in chapter 2. Hence, once we know the method of
solving the equilibrium pattern for one particular departuretime, we can obtain the equilibrium
pattern for whole time periods by successively applying the same procedure at the order of the
departuretime. In the following, we consider the problem of obtaining the equilibrium pattern for
vehicles departing from origin o at time s, assuming that the solutions for vehicles departing before
time s are already given.
In the decomposed formulation with origin departure time s, two kinds of variables, ( y]} ,T' ),
play a central roll: r,s is the earliest arrival time at node i for a vehicle departing from origin o at
time s; y", is the link flow rate with respect to s, that is, y*l = dFtJ (T ) / ds , where FtJ(f) denote the
cumulative number of vehicles entered into link ij at time t. In addition, we denote the number of
vehicles with destination d departing from origin o until time s (cumulative OD demand by
departuretime) by Qoli (s) .
In the DUE state, each user choose his/her route whose travel time is (ex post) minimum over
the network. In other words, the links with positive inflows should be on the minimum path tree.
In our saturated networks, all the links have positive inflows, and therefore the minimum path
condition for users with origin departuretime s is written as c(s) + A 7 T = 0 , where c(s) is an L
dimensional column vector with elements c* =c r:/ (r; v ), r(s) is an (Nl) dimensional column
vector with elements r,v . Since the equation above should hold for any s, taking the derivative
with respect to 5, we have
dt 0 otherwise
where /i y (/) is the standard link flow rate defined as dFtJ(t)ldt. Hence, in our saturated
306 Transportation and Traffic Theory
networks where all links have queues, the rate of change in the time needed to traverse link if for
users with origin departure time s, dc^ I ds , can be represented as:
_
ds dr* ds I Jly j ds
Noticing here the definitional relationship y*'. = A, (r/ ) • dr* I ds we see that the dc^lds
reduces to a function of y*. and r/:
(3.2a)
ds fitj ds
or equivalently
= M~'y(s)A+ \fs. (3.2b)
ds ds
where M is a diagonal matrix whose oth diagonal element represents the maximum capacity of link
a, y (s) is an L dimensional column vector with elements yy.
Substituting (3.2) into (3.1), we obtain
A( ~\
= 0, Vs (3.3)
Vs. (3.4)
ds
On the other hand, in the decomposed DUE formulation, the flow constraints that consist of
the FIFO condition for each link and the flow conservation at each node over a network reduce to
the following equations (for the detail, see Kuwahara and Akamatsu (1993), Akamatsu and
Kuwahara(1994)):
(2) Solution
The rank of the matrix AMA^ generally depends on the choice of a reference node. For a
A Paradox For A Dynamic Equilibrium Assignment 307
network with a onetomany OD, the rank of AM A [can be less than Nl when we choose an
arbitrary node that is not an origin as the reference node. The rank, however, is always Nl when
an origin is employed as the reference node. Furthermore, since the value of drt (s) I ds for an
origin node is always 1 from the definition of rt ( s ) , it is natural to choose an origin as the
reference node. Thus, by setting an origin as the reference node, we obtain the equilibrium
solution, dr(s)/ ds, by the following formula:
v (37)
ds ~' ds
In addition, we can obtain the equilibrium link flow pattern, y(s), by substituting (3.7) into (3.4).
(1) Formulation
The DUE assignment on a network with a manytoone OD pattern can be decomposed with
respect to the destination arrivaltime as shown in chapter 2. In the following, we consider the
problem of obtaining the equilibrium pattern for vehicles arriving at a destination at time u,
assuming that the solutions for vehicles arriving before time u are already given.
For the networks with a manytoone OD pattern, by decomposing with respect to the arrival
time at a single destination, the discussions almost parallels to those in the previous section, hi the
decomposed formulation with destination arrival time u, two kinds of variables, (>"//» r , ), play a
central roll: T" is the latest arrival time at node / for a vehicle reaching destination d at time u;
yl is the link flow rate with respect to u, that is, y>l = dFl} (T" }l du. hi addition, we denote the
number of vehicles with origin o arriving at destination d until time u (cumulative OD demand by
arrivaltime} by Q^ (u).
The formulation almost parallels the discussions in 3.1. First, the minimum path
conditions for saturated networks reduces to the following conditions:
Vw. (3.8)
du du
Then the link travel time with a point queue for saturated networks also should satisfy
(3.9)
du du
Substituting (3 .9) into (3.8), we obtain
^M v«. (3.10)
308 Transportation and Traffic Theory
On the other hands , the link flow y should satisfy the flow constraints:
.
 du du
Thus, we see that the DUE assignment has a unique solution (dt(u)l du and y(u)) if the rank of
AMA r isN1.
(2) Solution
An arbitrary network with a manytoone OD pattern can be obtained by reversing the
direction of all links and origin/destinations of a network with a onetomany OD pattern.
Therefore, it is natural to expect that, "reversing" the result in 3.1, the rank of AMA^ become
Nl when a destination is chosen as the reference node. However, it is not the case for this
problem; the rank become less than Nl even ifwe set the destination as the reference node;
furthermore, we can prove that the rank is less than Nl for any choice of the reference node.
The reason why the rank of the matrix AMA[ becomes less than Nl is that there exist
particular origins (we call this "pure origins") that are not traversal nodes (i.e. the origin which has
no links arriving at the origin). Letting B, be the (/'/) element of A*MAl , we easily see that
'f
(3.13)
Hence, the column vectors of AMA[ corresponding to the pure origin are always zero, and the
rank of AMAT_ necessarily decreases by the number of pure origins.
To see this fact more precisely, we divide the node set N into two subsets: the set of pure
origins, N,, and the set of the other nodes, N2. Then, we divide A*, A*_, d r (u)ldu and dQ(u)/du
into the two blocks corresponding to N, and N 2 , respectively:
"<&,(")" ~dQ}(u)~
dt(u] du dQ(u) du
A = A =
A, du dr2(u) du dQ2(u)
du du
where i th element of dQ2(u) /du is defined as  ^ {dQad (u) I du} = ^ jukd if / is an orign,
dQij(u)/du if / is a destination, zero otherwise. Note that A, _, which is the first block of A_
corresponding to N,, is always 0 according to the definition of the pure origins. Rewriting (3.12)
with these partitioned variables, we have
A Paradox For A Dynamic Equilibrium Assignment 309
du .r dr(u) du
= A'MA (3.14)
dQ2(u) du dr2(u)
du du
That is,
(3.15a)
du du
T
*?2(") = A,MA'_ dr2(u)
(3.15b)
du du
This means that no condition which determines the dr} / du for the pure origins is included in the
equilibrium condition (3.12), while the dr2 / dufor the traversal nodes can be obtained by
(3.16)
du du
Thus we see that the solution of the DUE assignment with a manytoone OD pattern can not be
unique and that for the problem to have a unique solution we should add appropriate conditions to
resolve the indeterminacy of the dt\ I du .
The previous chapter analyzed the solution of the DUE assignment where only user's route choice
is endogenously described given timevarying OD demands. This chapter extends the analyses to the
case where the timedependent OD demands are endogenously determined (we call the model "DUE
assignment with Elastic demand") by incorporating the user's departure time choice. The model
employed here is the simplest one that consistently unifies the two kind of dynamic equilibrium models:
the dynamic equilibrium assignment presented in the previous chapter and the dynamic equilibrium
model of departure time choice as is well known since Vickrey (1969) or Hendrikson and Kocur (1980).
For expositional brevity, the following assumptions are made in this paper:
1) The users with the same OD pair are homogeneous, that is, their utility functions are all the same and
their desired arrival time is unique;
2) The users who arrive later than the desired arrival time do not exist [This is not a restrictive
assumption but one just to make the exposition as simple as possible; it is easy to extend to the case
where late arrival is permitted.].
3a) For the problems with onetomany OD pattern (i.e. when we consider the problem on the basis of
the origin departuretime), the disutility function for the users with destination d leaving origin at time s,
310 Transportation and Traffic Theory
VJs), is given as the linear combination of their travel time from the origin to destination d and their
"schedule delay":
Vod (s) = a} {rd (5)  5} + a2 {td  Td (5)}, (4.1)
where a,, a2 are positive parameters that satisfy al > a2, Td(s) is the destination arrivaltime for the
users who start from origin at time s, and td is the users' desired arrival time.
3b) For the problems with manytoone OD pattern (i.e. when we consider the problem on the basis of
the destination arrivaltime), the disutility function for the users with origin o arriving at the destination
at time u, Vj(u), is given as the linear combination of their travel time from origin o to the destination
and their "schedule delay":
Vod («) = a} {u  TO (u)} + a2 {td  u], (4.2)
where TO(U) is the origin departuretime for the users who arrive at destination at time u.
4) The networks can be regarded as "saturated networks" that is defined in the previous chapter.
(1) Formulation
In this section we consider the networks with a onetomany OD pattern where all nodes except the
origin are destination, i.e., there are no nodes that are neither origin nor destination. [This is simply for the
convenience of expositional brevity. The appropriate division of the node set easily extends our analyses
to the general case where there are some nodes that are neither origin nor destination. See Appendix.]
The elastic demand DUE employed in this chapter is defined as the state where no one can improve
his/her utility by changing either his/her route or their departuretime unilaterally. To formulate this,
consider users who choose time s as departure time. Since the users choose their optimal route
(conditional on the optimal departure time) in the DUE state, the equilibrium conditions for the route
choice should be represented by the following differential equations as shown in Chapter 3:
(4.3)
ds ds
where the origin node is selected as a reference node as discussed in 3.1 . Then, the condition that no user
can improve his utility by changing his/her departuretime in the DUE state can be represented by
\ f s , Md. (4.4)
ds
Substituting the definition of disutility function (4.1) into this, we obtain the equilibrium rate of change in
the destination arrivaltime as follows:
= _ Vj> Md (4.5)
ds a, flu
A Paradox For A Dynamic Equilibrium Assignment 311
[We are assuming that networks can be regarded as "saturated networks" and all OD pairs have positive
OD flows during the period of time. In general we should consider the analysis period to include the time
where some OD pairs have no generation of OD flows. By introducing appropriate classification,
however, the general case can be reduced to the combination of our basic case (the case where all OD
pairs have positive OD flows during the period for our analysis) and the case presented in Appendix.].
Thus, the elastic DUE conditions are represented as the following system of differential equations:
a
=E \ (4.6a)
ds a, a
dQ(s)_
(4.6b)
ds " ' ds
where E is an (Nl) dimensional column vector whose elements are all equal to 1. It is worthwhile to
compare the equilibrium conditions with those for the fixed demand case. In the fixed demand DUE
model, eq.(4.3) with a given constant vector dQ(syds determines dr(s)/ds . On the contrast, in the
elastic demand DUE, dr(s}/ds is first determined from the departuretime equilibrium condition, and
then eq.(4.3) with fixed dr(s)lds determines dQ(s)/ds.
(2) Solution
By setting appropriate boundary conditions, we can obtain the solution (t(s),QCs)) for the
differential equation (4.6). For the boundary conditions, we first set the initial time Ss of the time period
(measured with respect to the origin departuretime) during which eq.(4.6) holds (i.e. the networks can be
regarded as "saturated networks" and all OD pairs have positive OD flows). Then we give the value of
cumulative OD flows for the time ss and for the final time of the period:
Vd (4.7a)
Vd (4.7b)
where s(td) is an origin departuretime of the final users who arrive at destination d at time td (note that
we do not have to give the value of s(td} explicitly).
Integrating the second equation of (4.6) from time ss to s with the initial condition (4.7a), we have
*,), (48)
 a,  a2
where Q is an (Nl) dimensional vector with elements g .
We then solve (4.6) with respect tot. Integrating the first equation of (4.6) from time Ss to time
s(td } reduces to
where t, t(.sv ), and s^ are (Nl) dimensional vectors with elements td, rd (ss), and s(t^, respectively.
312 Transportation and Traffic Theory
The length of the time period that appears in the right hand side of (4.9), s(td )ss , can be obtained by
substituting (4.7b) into (4.8):
'(QQ)
~ . (4.10)
Hence, from (4. 1 0) and (4.9), we can determine the initial equilibrium arrival time corresponding to ss :
Thus, the equilibrium pattern (t(s), Q(s)) with the boundary condition (4.7) is given by
Vs (4.12)
a
Q(.?)=Q+AMA[E ' (sss)
a, a.
and the corresponding equilibrium disutility Td (s) is calculated by
(1) Formulation
In the following we consider the networks with a manytoone OD pattern where all nodes except
the destination are origins, i.e., there is no node that is neither origin nor destination. For the general case
where there are some nodes that are neither origin nor destination, see Appendix.
We divide the node set N into two sub sets: the set of origins N,, and the set of the single
destination, N2. Then, we divide A*, A*_, dr(u)/du and dQ(u)/du into the two blocks
corresponding to N, and N2 , respectively:
The elastic demand DUE employed here is defined as the state where no one can improve his/her
utility by changing either his/her route or their departure/arrivaltime unilaterally. Since the users choose
their optimal route (conditional on having chosen his/her optimal departure/arrivaltime) in the DUE state,
the equilibrium conditions for the route choice should be represented by the following differential
equations as shown in Chapter 3:
A Paradox For A Dynamic Equilibrium Assignment 313
v
du ' du
v
_ (AIMA[_) (4.16)
du ' "' du
The condition that all the users can not improve their utility by changing his/her arrivaltime (or departure
time) in the DUE state can be represented as
dV
°d\u> = Q VM, Vo. (4.17)
Substituting the definition of disutility function (4.2) into this, we obtain the equilibrium rate of change in
the destination arrivaltime as follows: 
du a,
(4.19)
du
It is worthwhile to compare the equilibrium conditions with those for the fixed demand case. In
the fixed demand DUE model, we tried to determine dr(u) I du from the eq.(4.15) with a given constant
vector dQ(u)/du. Then we encountered the indeterminacy of dr(u) I du due to the decrease in the rank
of matrix A*MA*. On the contrast, in the elastic demand DUE, the indeterminacy problem is resolved
since dt(u)l du is first determined from the departuretime equilibrium condition, and then eq.(4.16)
with fixed dt(u)l du determines dQ(u)/du.
(2) Solution
As in the case of onetomany OD pattern, we can obtain the solution (i(s), Q(s)) for the differential
equation (4.19) by giving appropriate boundary conditions. For the boundary conditions, we first set the
initial time us of the time period (measured with respect to the destination arrivaltime) during which
eq.(4.19) holds (i.e. the networks can be regarded as "saturated networks" and all OD pairs have positive
OD flows). Then, on a parallel with the discussion in 4.1, it is natural to give the value of cumulative OD
flows from us and for the final time td :
conditions are not enough to determine the value of T . Hence, instead of (4.20a), we give the time
needed to travel from origin o to the destination at the initial time us as a new boundary condition:
us  TO (us ) = rod = given Vo . (4.20c)
Integrating the second equation of (4. 19) from time u to td with the initial condition (4.20c), we have
«) VM (4.21)
We next solve (4.19) with respect tot . Integrating the first equation of (4.19) from time us to time u
with the initial condition (4.20c), we obtain
~  ^ ) VM. (4.22)
a,
and the corresponding equilibrium disutility rd (s) is calculated by
p = fl2(/rfMJ)E + fl1r (4.23)
5. PARADOXES
Having derived the formulae for the solution of the dynamic traffic equilibrium assignment
so far, now we can discuss the capacity increasing paradox. The paradox presented here is a
situation such that improving the capacity of a certain link on a network worsen the total travel cost
over the network; this is a dynamic version of Braess's paradox which is well known in the static
assignment. Using the results obtained in Chapters 3 and 4, we derive the necessary conditions
for the occurrence of the paradox for Enet and Mnet, which are shown to be significantly
different.
We consider the paradox for the network shown in Fig. 5.1, where node 1 is a unique origin;
nodes 2 and 3 are destinations; the maximum departure rate of link a (a = 1,2,3) is given by //„.
For the brevity of notation, we employ the superscript" • " as the derivative operation with respect
to origin departuretime s in this section, (e.g. t,(s) = dr^lds, Q^s) = dQod(s)lds}.
1 1 0
1 0 1  1 0 0
A = 1 0 1 A= A = (5D
0 1 1 0 1 1
0 1 1
Hence,
*> 0
(5.2)
The equilibrium pattern for the vehicles with the departure time 5 from a single origin can be
calculated using the results of Chapter 3. From (3.6), we first obtain the rate of change in
equilibrium arrival time:
Substituting these into (3.3), we have the following equilibrium link flow pattern:
(5.4)
To discuss the "capacity increasing paradox", we employ the total travel time for the users
departing from an origin from time 0 to T as an indicator for measuring the efficiency of the
network flow pattern:
s
X f y* =Z f Q (5.5)
We then refer to the situation "paradox" if increasing the capacity of a certain link, //0, causes the
increase of TC (i.e. dTC/d/na > 0 implies "paradox").
Let us examine whether the paradox arises or not for the network in Fig. 4.1. Substituting
(5.3) into (5.5), we obtain TC:
316 Transportation and Traffic Theory
(5.6)
From (5.6), we easily see that the increase of //, or ju2 always decreases TC (note that both //,
and ju2 appear in only the denominator of TC), that is, the paradox does not arise for links 1 and 2.
Increasing ju3, however, causes the paradox. The reason is that since
if the condition:
S
'° ' 2 ^ " > *0 '3 ^ " * (5.8)
(5.10)
ata2 a,0 2
where Qad = Qod  Q . Then (4. 1 3) gives the equilibrium disutility for each origin:
A Paradox For A Dynamic Equilibrium Assignment 317
1 1 "
p2 =a,(t2 5,)+(a, a2)  012, p, =a}(t3 ss) + (al a2)  013  (5.12)
/^y"3 /" 2 +/" 3
We define the sum of disutility experienced by all users over a network, TC, as an indicator for measuring
the efficiency of the network usage:
. (513)
d
Note that the capacity of link 1 should be greater than that of link 3 (i.e. //, > /J3) in order for (5.11) to
satisfy the (physically evident) condition Q}2 (s)  Qt 2 (s v ) > 0. Hence dTC/dju3 > 0 holds only if
It is noteworthy that the condition (5.1 6b) is identical in form to the condition for the fixed demand case.
We consider the paradox for the network in Fig. 5. 2, where node 1 is a unique destination;
nodes 2 and 3 are origins; the maximum departure rate of link a (a = 1 ,2,3) is given by jua.
For the brevity of notation, we employ the superscript " • " as the derivative operation with respect
to destination arrival time u in this section, (e.g. T^^dr^u}/ du ,Qod(u) = dQad(u)/du)
*•—'
Fig. 5.2. Example Network with Two Origins and Single Destination
318 Transportation and Traffic Theory
A MA  = 0 (5.18)
//3
The equilibrium pattern for the vehicles with the arrival time u at a single destination can be
calculated from the results in Chapter 3. From (3.16), we first obtain the rate of change in
equilibrium arrival time for nodes 1 and 2:
T2(u) =  (5.19)
Substituting these into (3.10) yields the link flow rates (with respect to u):
(5.20)
Note that this flow pattern is significantly different from that for the reversed network (see (5.4)).
hi order to determine the rate of change in equilibrium arrival time for node 3 (= the pure
origin), adding an appropriate condition is required. Here we assume for node 3 that the OD flow
rate measured at the origin, <?31 = dQ3l (u) I dr3 (u) = Q3} (u) I f3 (u), is given. On the other hand,
the OD flow rate measured at the destination, q3] = Q3l (u) , is determined from (3.15a):
qod(u) du dT0(u) du ^
~ f \ (5.22)
q3](u)
Defining the total travel time for the users arriving at an destination from time 0 to T as an
indicator for measuring the efficiency of the network flow pattern:
TC
* (5.23)
A Paradox For A Dynamic Equilibrium Assignment 319
let us examine whether the paradox arises or not in the network in Fig. 5.2. Substituting (5.19),
(5.21) and (5.22) into (5.23), we obtain the TC for this network:
(5.24)
du
where Q^(u}= \ qod(u)du. We see from this equation that the increase in /^, or ju2 will
always decrease TO, the paradox does not arise for links 1 and 2. However, the increase in the
capacity of link 3 always results in the occurrence of the paradox. This fact can be easily
examined as follows. Calculating the derivative of TC with respect to ju3, we have
Note that f2 (u) should be positive in the DUE state. The reason is that if f 2 («) is not positive
the users with the destination arrival time if> u must depart from their origin before the users with
arrival time u, and this contradict the assumption that the state is in the DUE. Therefore, from the
(5.25) and the fact that f 2 («)>0 for any u, the inequality dTC I dju3 > 0 always holds; we see
that the paradox for link 3 takes place without any additional conditions.
Hence, from (4.21) and (4.22), we obtain the equilibrium arrival times and OD flows:
, ,
We also get the equilibrium disutility from (4.23):
P2=a2(tus)+a}r2}(us), p3 = a2(t us)+a^(us). (5.29)
Let us define the sum of disutility experienced by all users over a network, TC, as an indicator for
measuring the efficiency of the network usage:
320 Transportation and Traffic Theory
TC^p0Qod (530)
Substituting (5.29) into the definition (5.30), we get the TC for the network in Fig.5.2:
(531)
^ = a}(a,a2](tus}{r3,(us)r2l(us)}. (5.32)
a//3
6. Concluding Remarks
This paper discussed a capacity increasing paradox under a dynamic equilibrium assignment
with elastic OD demands: the paradox is a situation such that improving the capacity of a certain
link on a network worsen the total travel cost over the network. Our analysis in a simple network
disclosed that the paradox arises only on a particular condition for a network with a onetomany
OD pattern, while the corresponding paradox always arises for the reversed network with a many
toone OD pattern. This is the asymmetrical result that can not be seen in the classical static
assignment framework; it is particular to the dynamic assignment with queue. Furthermore, we
show that this property holds not only for the assignment with fixed OD demands but also for the
assignment with elastic OD demands.
A Paradox For A Dynamic Equilibrium Assignment 3 21
In this paper, particular simple networks were employed to demonstrate the paradox. Note,
however, that the examples presented here are not the exceptional ones that can hardly be observed
in practical situations but the ones that can be seen universally if we regard the example networks
as a macroscopic representation of real road networks. Therefore, we think that the examples,
despite their simplicity, describe one of the essential points that should be considered in deciding
practical traffic management operations such as ramp metering or addition of lanes in freeways.
We recognize that there are still several relevant topics to be studied. First, we should extend
our analysis to the paradox in a more complex network by exploiting the analytical formula of the
DUE solution derived in this paper; it may be possible to obtain systematic methods for general
networks that detect (without computing the equilibrium patterns) the links where the paradox
takes place; the exploration of this possibility would be an interesting future topic. Secondly, we
should analyze more realistic case where the assumption of "saturated networks" are relaxed; the
exploration would be achieved by employing not only the analytical approach just as shown in this
paper but also the numerical approach based on the recent convergent algorithms for the DUE
assignment (see Akamatsu (1998)). Finally, we should explore the case where physical queues
are explicitly incorporated into the analysis. Though the incorporation of physical queues may
cause very complex phenomena as shown in Daganzo(1998), comprehensive studies on this topic
would be indispensable for a clear understanding of the properties of dynamic network flows.
Acknowledgements
The authors gratefully acknowledge stimulating discussions with Nozomu Takamatsu on the topic
of this paper. Thanks are also due to Benjamin Heydecker and three anonymous referees for their
helpful comments and suggestions.
REFERENCES
Akamatsu T. (1996). The Theory of Dynamic Traffic Network Flows, Infrastructure Planning
Review 13, 2348.
Akamatsu T. (1998). An Efficient Algorithm for Dynamic User Equilibrium Assignment for a One
toMany OD Pattern, submitted to Transportation Science.
Akamatsu T. and M. Kuwahara (1994). Dynamic User Equilibrium Assignment on Oversaturated
Road Networks for a OnetoMany / Manyto One OD Pattern, JSCE Journal of
Infrastructure Planning and Management IV23, 2130.
Arnott R., De Palma A., and R. Lindsey (1993). Properties of Dynamic Traffic Equilibrium
Involving Bottlenecks, Including a Paradox and Metering, Transportation Science 27,
148160.
Bernstein D., T.L. Friesz, R.L. Tobin, and B.W. Wie (1993). A Variational Control Formulation
322 Transportation and Traffic Theory
Appendix
In Chapter 4 it is assumed that all the OD pairs have always positive flows during the
period of analysis. In this appendix, we briefly demonstrate how the formulation can be
extended to the case where some OD pairs have no OD flows. The formulations for the oneto
many OD problem and the manytoone are presented in turn.
For the destinations with positive OD flows, the arrival times are governed by the departure
time equilibrium condition (4.5):
=E (A2)
ds
For the other destination nodes, the arrival times should be determined from the route choice
equilibrium condition (3.6). Rewriting the condition (3.6) with the variables defined above,
ds (A3)
dr2(s)
0
Much more than documents.
Discover everything Scribd has to offer, including books and audiobooks from major publishers.
Cancel anytime.