Lecture Notes in Economics and Mathematical Systems
(Vol. 1-15: Lecture Notes in Operations Research and Mathematical Economics, Vol. 16-59: Lecl
Notes in Operations Research and Mathematical Systems)
Vol. 1: H. BOhlmann, H. Loeffel, E. Nievergelt, EinfUhrung in die
Theorie und Praxis der Entscheidung bei Unsicherheit. 2. Auflage,
IV, 125 Seiten. 1969.
Vol. 2: U. N. Bhat, A Study of the Queueing Systems M/G/1 and
Gil MIl. VIII, 78 pages. 1968.
Vol. 3: A Strauss, An Introduction to Optimal Control Theory.
Out of print
Vol. 4: Branch and Bound: Eine EinfUnrung. 2., geiinderte Auflage.
Herausgegeben von F. Weinberg. VII, 174 Seiten. 1973.
Vol. 5: L. P. Hyviirinen, Information Theory for Systems Engineers.
VII, 205 pages. 1968.
Vol. 6: H. P. KOnzi, O. MOller, E. Nievergelt, EinfUhrungskursus in
die dynamische Programmierung. IV, 103 Seiten. 1968.
Vol. 7: W. Popp, EinfUhrung in die Theorie der Lagerhaltung. VI,
173 Seiten. 1 968.
Vol. 8: J. Teghem, J. Loris-Teghem, J. P. Lambotte, Modeles
d'Attente M/GI1 et GI/MI1 a Arrivees et Services en Groupes. III,
53 pages. 1969.
Vol. g: E. Schultze, EinfUhrung in die mathematischen Grundlagen
der Informationstheorie. VI, 116 Seiten. 1 969.
Vol. 10: D. Hochstiidter, Stochastische Lagerhaltungsmodelle. VI,
269 Seiten. 1969.
Vol. 11/12: Mathematical Systems Theory and Economics. Edited
by H. W. Kuhn and G. P. Szeg/). VIII, III, 486 pages. 1969.
Vol. 13: Heuristische Planungsmethoden. Herausgegeben von
F. Weinberg und C. A. Zehnder. II, 93 Seiten. 1969.
Vol. 14: Computing Methods in Optimization Problems. V, 191 pages.
1969.
Vol. 15: Economic Models, Estimation and Risk Programming:
Essays in Honor of Gerhard Tintner. Edited by K. A. Fox, G. V. L.
Narasimham and J. K. Sengupta. VIII, 461 pages. 1969.
Vol. 16: H. P. KOnzi und W. Oettli, Nichtlineare Optimierung:
Neuere Verfahren, Bibliographie. IV, 180 Seiten. 1969.
Vol. 17: H. Bauer und K. Neumann, Berechnung optimaler Steue-
rungen, Maximumprinzip und dynamische Optimierung. VIII, 188
Seiten. 1969.
Vol. 18: M. Wolff, Optimale Inslandhaltungspolitiken in einfachen
Systemen. V, 143 Seiten. 1970.
Vol. 19: L. P. Hyviirinen, Mathematical Modeling for Industrial Pro-
cesses. VI, 122 pages. 1970.
Vol. 20: G. Uebe, Optimale Fahrpliine. IX, 161 Seiten. 1970.
Vol. 21: Th. M. Liebling, Graphentheo;ie in Planungs- und Touren-
problemen am Beispiel des stiidtischen StraBendienstes. IX,
118 Seiten. 1970.
Vol. 22: W. Eichhorn, Theorie der homogenen Produktionsfunk-
tion. VIII, 119 Seiten. 1970.
Vol. 23: A Ghosal, Some Aspects of Queueing and Storage
Systems. IV, 93 pages. 1970.
Vol. 24: G. Feichtinger, Lernprozesse in stochastischen Automaten.
V, 66 Seiten. 1970.
Vol. 25: R. Henn und O. Opitz, Konsum- und Produktionstheorie I.
. 11,124 Seiten. 1970.
Vol. 26: D. Hochstiidter und G. Uebe, 6konometrische Methoden.
XII, 250 Seiten. 1970.
Vol. 27: I. H. Mufti, Computational Methods in Optimal Control
Problems. IV, 45 pages. 1970.
Vol. 28: Theoretical Approaches to Non-Numerical Problem Sol-
ving. Edited by R. B. Banerji and M. D. Mesarovic. VI, 466 pages.
1970.
Vol. 29: S. E. Elmaghraby, Some Network Models in Management
Science. III, 176 pages. 1970.
Vol. 30: H. Noltemeier, Sensitivitiitsanalyse bei diskreten lin.
Optimierungsproblemen. VI, 102 Seiten. 1970.
Vol. 31: M. KOhlmeyer, Die nichtzentrale t-Verteilung. II, 106
ten. 1970.
Vol. 32: F. Bartholomes und G. Hotz, Homomorphismen un<
duktionen linearer Sprachen. XII, 143 Seiten. 1970. DM 18,-
Vol. 33: K. Hinderer, Foundations of Non-stationary Dynamic
gramming with Discrete Time Parameter. VI, 160 pages. 1
Vol. 34: H. Sttirmer, Semi-Markoff-Prozesse mit endlich v
Zustiinden. Theorie und Anwendungen. VII, 128 Seiten. 1
Vol. 35: F. Ferschl, Markovketten. VI, 168 Seiten. 1970.
Vol. 36: M. J. P. Magill, On a General Economic Theory of Mo
VI, 95 pages. 1970.
Vol. 37: H. MOlier-Merbach, On Round-Off Errors in Linear
gramming. V, 48 pages. 1970.
Vol. 38: Statistische Methoden I. Herausgegeben von E. W,
VIII, 338 Seiten. 1970.
Vol. 39: Statistische Methoden II. Herausgegeben von E. W,
IV, 157 Seiten.1970.
Vol. 40: H. Drygas, The Coordinate-Free Approach to Ga
Markov Estimation. VIII, 113 pages. 1970.
Vol. 41 : U. Ueing, Zwei Ltisungsmethoden fUr nichtkonvexe
grammierungsprobleme. IV, 92 Seiten. 1971.
Vol. 42: A V. Balakrishnan, Introduction to Optimization Thec
a Hilbert Space. IV, 153 pages. 1971.
Vol. 43: J. A. Morales, Bayesian Full Information Structural AI
sis. VI, 154 pages. 1971.
Vol. 44:- G. Feichtinger, Stochastische Modelle demographis
Prozesse. IX, 404 Seiten. 1971.
Vol. 45: K. Wendler, Hauptaustauschschritte (Principal Pivot
II, 64 Seiten. 1971.
Vol. 46: C. Boucher, Leyons sur la theorie des automates
thematiques. VIII, 193 pages. 1971.
Vol. 47: H. A Nour Eldin, Optimierung linearer Regeisysl
mit quadratischer Zielfunktion. VIII, 163 Seiten. 1971.
Vol. 48: M. Constam, FORTRAN fUr Anfiinger. 2. AuflagE
148 Seiten. 1973.
Vol. 49: Ch. SchneeweiB, Regelungstechnische stochasti,
Optimierungsverfahren. XI, 254 Seiten. 1971.
Vol. 50: Unternehmensforschung Heute - Obersichtsvortriig.
ZOricher Tagung von SVOR und DGU, September 1970. Her
gegeben von M. Beckmann. IV, 133 Seiten. 1971.
Vol. 51: Digitale Simulation. Herausgegeben von K. Baukr
und W. Nef. IV, 207 Seiten. 1971.
Vol. 52: Invariant Imbedding. Proceedings 1970. Edited by F
Bellman and E. D. Denman. IV, 148 pages. 1971.
Vol. 53: J. RosenmOller, Kooperative Spiele und Miirkte. III,
Seiten. 1971.
Vol. 54: C. C. von Weizsiicker, Steady State Capital Theor
102 pages. 1971.
Vol. 55: P. A. V. B. Swamy, Statistical Inference iQ Random (
ficient Regression Models. VIII, 209 pages. 1971.
Vol. 56: Mohamed A EI-Hodiri, Constrained Extrema.lntrodu
to the Differentiable Case with Economic Applications. III,
pages. 1971.
Vol. 57: E. Freund, ZeitvariableMehrgrtiBensysteme. VIII,160
ten. 1971.
Vol. 58: P. B. Hagelschuer, Theorie der linearen Dekomposi
VII, 191 Seiten. 1971.
continuation on page
Lectu re Notes
in Economics and
Mathematical Systems
Managing Editors: M. Beckmann and H. P. KOnzi

Operations Research

118

Traffic Equilibrium Methods

Proceedings of the International Symposium
Held at the Universite de Montreal,
November 21-23, 1974

Edited by M. A. Florian

Springer-Verlag
Berlin· Heidelberg· New York 1976
Editorial Board
H. Albach ' A V. Balakrishnan' M. Beckmann (Managing Editor)
P. Dhrymes . J. Green' W. Hildenbrand · W. Krel1e
H. P. Kunzi ( Managing Editor) . K.Ritter' R. Sato . H. Schelbert
P. SchOnfeld

Managing Editors
Prof. Dr. M. Beckmann Prof. Dr. H. P. KOnzi
Brown University UniversiUit Zuric h
Providence, RI 02912/USA 8090 ZOrich/Schweiz

Editor
Dr. Michael A Florian
Universite de Montreal
P.O. Box 6128
Montreal 101
Quebec H3C 3J7/Canada

1IIt.eznat1CO>Al. SppodUIII .... TnN1c: Equ.U1b1"11m. Methods,

Moouu1, 19'710.
'!'rat1"lc eql11l.1bdUli ..tboda ; !proceecUna.l,

,"'.
IntenlatiOD&l. SyJI;p:IdUII, Mootreal , Quebec , IIC1Ve111beI" 21--23,

(Lecture note. 111 ec()l>all1c. "",,- ...thematieal

. y. teIU ; l.l8) (OpuatiOlLl ""uareh)
Blbllogr.p>;:r: p.
:t=l.u&e. 1Qdu .
1. '!'raN"1e eaa1D.evUl6- -M&th...t1elLl lI"dd .. _ ~.. _
u. 2 . 1'rat1"le .... ~ --M&tmm&t1c&l. 1IOd4lI --C<lflgJ""~
.... I . Fl.ol"hll, lUeb&el, 1939- II . Hili .
III. Seri n. IV . Sed u : OperaUOIUI reteareh {Ber.l.1.!:>j .
H!.332 . 165 1974 388. 3'1 76-3760

AMS Subject Classifications (1970): 90A15, 90Bl 0, 90B20, 90C25,

90C50, 90010

ISBN 978+540-07620-9 ISBN 978-3-642-48123-9 (eBook)

DOl 10.1007/978-3-642-48123-9

This work is sUbject to copyright All rights are reserved, whether the whole
or part of the material is concerned, specifically those of translation, re-
printing, re'Use of illustrations, broadcasting, reproduction by photocopying
machine or simitar means, and storage in data banks.
Under § 54 of the German Copyright Law where copies are made for other
than private use, a fee is payable 10 the p4Jblisher, the amount of the fee to
be determ ined by agreement with the publisher.
C by Springer- Verlag Berlin' Heidelberg 1976
\oJe dedicate this volume to the memory of Pierre Robillard
 AC KN0\4L EDGMENTS

The programme committee consisted of Professor Martin Beckmann of Brown University

&Technische UniversitSt MUnchen, Dr. Robert B. Dial of the Urban Mass Transporta-
tion Agency, Dr. John Murchland of PTRC London, Mr. Earl Ruiter of Massachusetts
Institute of Technology and Cambridge Systematics Inc., Professor Pierre Robillard
and Professor Michael Florian of the Universite de Montreal. Sadly, Professor
Robillard died in a car accident in January 1975. We would like also to express
a speci~l appreciation to Dr. Robert B. Dial, who not only contributed to the sym-
posium actively, but provided incentive and encouragement for its organization.

Also, we are gratef~l to the Transportation Development Agency of Transport Canada

and to the Urban Mass Transportation Agency of the U.S. Department of Transporta-
tion for their indirect contribution to the financing of the symposium.

Special thanks are due to Mrs. Clairette Masse for the administrative organization
of the symposium.
 PREFACE

The interest in the mathematical modeling of transportation systems stems from the
need to predict how people might make use of new or improved transport infrastruc-
ture in order to evaluate the benefit of the required investments. To this end it
is necessary to build models of the demand for transportation and models that de-
termine the way in which people who travel use the transportation network. If such
models may be constructed and their validity reasonably assured, then the predic-
tion of the traffic flows on future and present transportation links may be carried
out by simulating future situations and then evaluating the potential benefits of
alternative improvement projects.

In the attempts that were made to construct mathematical models of transportation

networks, the notion of equilibrium plays a central role. Suppose that the demand
for transportation, that is, the number of trips that occur between the va-
rious origins and destinations is known. Then it is necessary to determine how
these trips are attracted to the alternative routes available between origins and
destinations. Knight (1924), gave a simple and intuitively clear description of
the behaviour of road traffic under conditions of congestion. We quote

"Suppose that between two points there are two hiqhways, one of Hhich is broad
enough to accommodate without cro~lding all the traffic which may care to use it,
but is poorly graded and surfaced, while the other is a much better road, but
narrow and quite limited in capacity. If a large number of trucks operate between
the two termini and are free to choose either of the two routes, they will tend to
distribute themselves between the roads in such proportions that the cost per unit
of transportation, or effective return per unit of investment, will be the same for
every truck on both routes. As more trucks use the narrower and better road, con-
gestion develops, until at a certain point it becomes equally profitable to use the
broader but poorer highway."
 VIII

This simple example illustrates the principle of equilibrium among alternative rou-
tes. Some 28 years later, Wardror (1952) stated two principles that formalize this
notion of equi1ibriu~.

The first principle states that "the journey times on all routes actually used are
equal, and less than those which would be experienced by a single vehicle on any
unused route". The implication of this principle is that the routes actually used
are shortest routes under the prevailing traffic corditirnso This characterization
of equil ibrium, identical to the notion described by Knight, became accepted over
the past 20 years as a sound behavioural princir1e to describe the subdivision of
traffic over alternative routes at equilibrium. The traffic flows that result un-
der this principle are usually referred to as "user optimized" flows, in order to
distinguish them from "system optimized" flo!s lI.'hich are characterized by Wardrop's
second principle which states that "the average journey time is a minimum".
lIardrop compared these two principles and commented that the first criterion is quite
a likely one in practice, since it implies an equilibrium situation in which no dri-
ver can reduce his journey time by choosing a new route; the second criterion is the
most efficient since it minimizes the total travel time in the network studied.
Wardrop discussed the equilibrium conditions for a general network but did not
propose a method to conpute the resulting flows.

Beckmann, McGuire and Winsten (1956) were the first to develop a mathematical model
of traffic equilibrium on a network with variable demand o They showed that, for
the "user optimized" traffic equil ibrium, there exists an equivalent minimization
problem; the solution of tliis minimization problem produces traffic flows that sa-
tisfy ~iardrop' s first principle. We sketch very briefly a proof of this equiva-
lence when all traffic is assumed to be homogeneous, that is no distinction is made
between cars, buses, trucks, etc., as was done by Beckmann et al.

Consider a transportation network of nodes N and oriented links A. A node re-

presents an origin, a destination or an intersection of streets. A two way street
is represented by directed links of opposite orientation. The number of trips that
 IX

occur from origins p to destinations q are given by a function gpq (u pq ) ,

when u is the cost of the trip from p to q. On each link of the network,
pq
the congestion effects are represented by link volume delay functions, s (v )
a a
a £ A , which are convex increasing functions of the link flows v On the net-
a
work gpq (u pq ) subdivides into route flo~ls h On the network all flows must
k,pq
be conserved

~ h - g (u ) all (p,q) ( 1)
~ k,pq - pq pq

and the link flows ar~ related to the route flows as

va = LLL 6
ak,pq hk,pq all a (2 )
pq k
where

6ak,pq -1 0
if

otherwise
a belongs to path k for pair (p,q)

and v > 0 hk ,pq -> 0 (3)

a

Wardrop's first principle may now be stated as

L6 sa (v a ) = u if hk,pq > 0 (4)

a ak,pq pq

L6 sa(v a ) -> upq if hk,pq = 0 (5 )

a ak,pq

The implicit assumption made is that the travel time on a route is the sum of the
link times that form the route. Equations (4) and (5) may be interpreted as the
Kuhn-Tucker conditions of an equivalent minimization problem which is
va 9 pq
Min L J s a (x) dx - I J Wpq (y) dy (6)
a 0 (p,q)o

where Wpq (gpq) is the inverse of the demand function gpq (u pq ), subject to

constraints (1), (2) and (3). Hhen the demand function gpq (u pq ) is a decreasing

function in upq , the objective function (6) is convex, which has the implication
that the solution of this program is unique and stable.
 x

Thus, with respect to the above mathematical model of traffic equilibrium on a net-
work, equilibrium means a steady state that is reached when the various origin-
destination demands for transportation give rise to traffic flows and conditions
that maintain that demand. The actual computation of this equilibrium state is a
difficult task: the collection of data and the estimation of demand ~unctions

is an exceedingly difficult problem in practice; also the solution of the problem

(6), (1), (2) for a large size network is not a trivial task. Thus, the interest
of practitioners in the general equilibrium problem was low and advances in the
computation of equilibrium traffic flows progressed slowly. The thrust of trans-
portation planning in practice was directed to development of the urban transporta-
tion planning system, commonly referred to as UTPS.

During the 1960's, numerous transportation studies were carried out for major cities
throughout the world. In most of these studies the estimation of demand for trans-
portation was achieved by use of the sequence of trip generation, trip distribution
and modal split models. The function of these models is roughly as follows:

Trip generation - Mathematical models that relate the number of trips originating
a p , and terminating bq in a traffic zone to land use factors, employment and
socio-economic characteristics.

Trip distribution - Mathematical model that relates the origin-destination flows

b
q
,the trips originating and terminating in the traffic zones.

tft,odal spl it - Mathematical model that relates the choices among modes of transpor-
tation to service characteristics, fares, costs and socio-economic characteristics
of the travellers. Thus the origin-destination trips by all modes, 9pq are
"split" as it were, between private car trips, transit trips and others.
(-gpqcar ,g- pqtransit , etc. )

There are many reasons for the development of the transportation modelling for urban
areas in this way and a thorough discussion of the pros and cons of the urban trans-
portation planning process is beyond the scope of this preface.
 XI

It is worthwhile however to mention that these models are relatively easy to

"calibrate" against traffic patterns rceasured for some base ye~r and the goodness
of fit of the calibration has often served as a substitute for demonstrated proof of
the predictive ability of these models. The output of the first three stages of the
transportation planning process wasa fixed origin-destination matrix that served
as input to the trip assignment stage. This consists of describing the way in which
the fixed origin-destination demands give rise to traffic flows on the links of the
network o The relation of trip assignment to the traffic equilibrium problem is very
close. Rather than being given by demand functions gpq(u pq ) the demand ~s fixed,
sav -g ,and the traffic assignment problem may be stated as:
" pq

v
Min I fa sa(x) dx (7)
a 0

subject to ~ hk,pq = gpq all (p,q) (8)

and constraints (2) and (3).

Because of the similarity of this problem vlith the general equilibrium problem it
became known as the traffic equil ibrium problem with "fixed demand" and the' general
problem is usually referred to as the equilibrium problem with variable or "elastic
demand".

In most of the large scale transportation studies that were carried out, the traffic
assignment problem was solved by a variety of heuristic methods which attempted to
produce "user optimi zed" flows. The success of these methods was mixed and progress
on the computation of equilibrium flows vias achieved by academic researchers, ra-
ther than those engaged in practical transportation planning.

It was not until the mid sixties that rigorous solution methods were proposed to solve
the traffic equilibrium problem with fixed or elastic demand. An attempt was made
at the Massachusetts Institute of Technology by Martin and Manheim (1963) to solve the
problem with a method that amounts to incrementally loading the network until the
equilibrium conditions are approximately satisfied, This method, although heuristic,
 XII

renewed the interest in the problem. The short lived :Transport Network Theory Unit
of the London School of Economics and Political Science, made significant contribu-
tions. In particular Gibert (1968) and later Murchland (1969) suggested algorithms
for the solution of the traffic equilibrium problem. Bruynooghe et al (1968) of the
Institut de recherche sur les transports, Arceuil, also developed an algorithm for
the problem. At about the same time, Dafermos and Sparrow (1969) studied the pro-
blem and suggested algorithms for the fixed demand problem. Later Dafermos (1971,
1972, 1973) presented an improved algorithm, developed a generalization of the
traffic equilibrium model to multiclass-user networks and derived the pattern of
tolls that must be imposed in order to render user-optimized flows system optimal.

The revived interest in the traffic equilibrium problem was further stimulated by
the possibility of applying these methods to larger scale problems by refining the
computational methods, applying recent advances in mathematical programming techniques
and developing efficient computer codes for digital computers. This was a result of
development of algorithms for the problem with fixed demands by Leventhal et al (1973).
LeBlanc (1973) and Nguyen (1973).

Thus, the time seemed appropriate to organize a symposium on the topic and bring to-
gether the geographically scattered researchers and practitioners who had developed
the conceptual, theoretical, computational and practical aspects of traffic equili-
brium analysis. The aim was to disseminate recent progress, assess the current state
of the art and discuss fruitful avenues for future research.

BIBLIOGRAPHY

1. Beckmann, M.J., McGuire, C.B. and Winsten, C.B., (1956) Studies in the Economics
of Transportation, Yale University Press, New Haven, Conn.
2. Burynooghe, t'1., Gibert, A. and Sakarovitch, ~1., (1968) "Une methode d'affectation
du trafic", Fourth Symposium on the theory of Traffic Flow, Karlsruhe.
3. Dafermos, S.C. and Sparrow, F,T., (1969) "The Traffic Assignment Problem for a
General Network", Journal of Research, National Bureau of Standards, 73B, No.2.
 XIII

4. Dafermos, S.C. (1971) "An Extended Traffic Assignment Model with Application
to Two ~lay Tra ffi c" , Transportation Science, 5, 366-389.
5. Dafermos, S.C" (1972) "The Traffic Assignment Problem for Multiclass-User
Transportation Network", Transportation Science, 6,73-87.
6. Dafermos, S.C., (1973) "Toll Patterns for Multiclass-User Transportation
Networks", Transportation Science, 7, 211-223.
7. Gibert, A., (1968) "A Method for the Traffic Assignment when Demand is
Elastic", LBS-TNT-85, Transport Network Theory Unit, London Business School.

8. Knight, F.H., (1924) "Some Fallacies in the Interpretation of Social Costs",

Quarterly Journal of Economics, 38, 582-606.
9. Leblanc, L.J" (1973) "Mathematical Programming Algorithms for Large Scale
Network Equilibrium and Network Design Problems", Ph.D. dissertation, Depart-
ment of Industrial Engineering and Management Science, Northwestern University.
10. Leventhal, L.U., Nemhauser, G.L. and Trotter, L.E., (1973) "A Column Genera-
tion Algorithm for Optimal Traffic Assignment", Transportation Science, 7.
11. Martin, B.V. and Manheim, M.L., (1965) "A Research Program for Comparison of
Traffic Assignment Techniques", Highway Research Record, 88.
12. Murchland, J.D., (1969) "Road Network Traffic Distribution in Equilibrium",
paper presented at the conference "Mathematical Methods in the Economic
Sciences", ~1athematisches Forschunginstitut, ObenlOlfach, 1969.
13. Nguyen, S., (1973) "A Mathematical Programming Approach to Equilibrium Methods
of Traffic Assignment with Fixed Demands", publication # 138, Departement
d'informatique, Universite de Montreal.
14. Wardrop, J.G., (1952) "Some Theoretical Aspects of Road Traffic Research",
Proc. Inst. Civil Engineers, Part II (1), 325-378.
 OVERVIEW OF THE PROCEEDINGS

There were twenty-four papers given at the Symposium, addressing various aspects
of the theory, computation and application of traffic equilibrium methods. A
classification of the papers by area of interest and a brief review of each of the
papers are as follows:

(i) General Modelling Considerations

Dial* (Urban Mass Transportation Administration) described the development of UTPS

software systems for transportation planning under his direction. He emphasized
the need to bridge the gap between research workers and practicing engineers by
presenting research results in easily usable forms. Many planning questions related
to multi-mode short-run policy making must consider social and environmental im-
pacts, and traffic equilibrium models may only indirectly serve to answer such
questions.

'1urchland (PTRC, London) presented a conceptual framework for structuring theoreti-

cal research in order to produce results of greatest practical value. He stressed
the need to test the validity of planning models and measure the component errors
due to the model itself, the data sample and the forecasting process.

Wigar. (Greater London Council) repeated some of the above-mentioned themes and des-
cribed a range of theoretical problems raised by practical experiences gained with
application of equilibrium type transportation models. The need to develop models
that identify different types of users was highlighted.

(ii) Theoretical Studies

Hall and Peterson (Bell Laboratories, Northwestern University) presented a "gene-

ralized geometric programming" approach to the classical formulation of the equili-
brium problem with elastic demand.
* Unfortunately this paper was not available for the proceedings.
 XVI

Dafermos (Brown University) showed how different and complex demand generation me-
chanisms may be represented without radical departures from a tractable network
structure.

Beckmann (Brown University and Technische Universit~t MUnchen) presented an equili-

brium model for public transport use. This represents an important first cut at
this difficult problem. He showed that fares may be chosen to obtain a system op-
timum when all travellers are homogeneous.

Jensen (Technical University of Denmark) presented a different equilibrium problem

in traffic models. He presented models of statistical equilibrium on multi-lane
traffi c fl ow.

Hamerslag (D\~ar5 Hepderik en Verhey, BV, Netherlands) presented a model that achie-
ves equilibrium between land use and development of transportation facilities.

(iii) Computational ~ethods

Nguyen (Universite de Montreal) gave an extremely interesting paper on the available

algorithms for computing traffic equilibrium flows. He showed that most of these
methods are special cases of methods of feasible directions in mathematical pro-
gramming. He also reported extensive computational tests with some of these me-
thods.

Ruiter (Massachusetts Institute of Technology and Cambridge Systematics Inc.) gave

a detailed description of a particular equilibrium method, that was designed for in-
corporation into the UTPS system. (The method is similar to algorithm Al described
by Nguyen) 0

Nestle (Cambridge Systematics, Inc.) considered the error or convergence measures

that should be used in practice and gave some computational results obtained with
the use of DODOTRANS.

Evans (University of Bristol) treated a particular family of combined trip distri-

bution and assignment problems, giving proof for the convergence of an algorithm for
 XVII

the problem.

Burrell (Burrell and Associates, U.K.) compared two standard models of vehicle
driver route choice which are not equilibrium type models - Dial's STOCH model and
his own probabilistic method.

(iv) Dynamic Traffic Assignment Methods

Yagar (University of Waterloo) gave a time-dependent traffic assignment method,

which has been applied to the testing of traffic control schemes in Ottawa.

Merchant and Nemhauser (University of Rochester, Cornell University) presented a

model of a restricted dynamic assignment problem. In spite of the fact that only
one origin is considered, the corresponding mathematical programming model is
non convex and rather difficult to solve.

(v) Traffic Equilibrium and Traffic Control

Gartner (Massachusetts Institute of Technology) gave a presentation of the area

traffic control problem and presented computational methods for it. He remarked
that traffic control models and trip assignment methods had not been jointly used
in practice.

Orthlieb and May (University of California, Berkeley) described a sequence of models

that optimize signal settings and ramp metering schemes for freeway operations.

(vi) Model components

Morlok (University of Pennsylvania) presented a pioneer attempt to construct supply

functions that describe the dependence of level of service in public transit sys-
tems to the demand for those services. He illustrated his presentation with data
obtained from the city of Chicago.
 XVIII

Gaudry (Universite de Montreal) questioned the term "supply curve" that is often
used to denote a link volume delay function. He then characterized these func-
tions as production functions.

(vii) Applications of Traffic Equilibrium Methods

Florian and Nguyen (Universite de Montreal) tested the equilibrium traffic assign-
ment algorithm of Nguyen on the network of the city of Winnipeg. Agreement bet-
ween predi cted and observed f1 OvlS was very good.

Van Vliet (University of Leeds) described the design of an incremental type assign-
ment method for the Greater London area. He discussed special problems asso-
ciated with large networks.

Leblanc and ~10rlok (Southern Methodist University, University of Pennsylvania)

compared "user optimized" and "system optimized" flows for the network of the city
of Sioux Falls. At least for that city the differences were minor.

Judge (University of Leeds) reviewed route choice models for roads and showed a
method that may be used to compare the results of the model with available data.
He also gave an illustration of the method.

Genest, Burden, Muguet and Kardicali (Quebec Department of Transport) reported on

the application of DODOTRANS in the study of the feasibility of a high-speed trans-
port link to the new international airport of Montreal.

Michael Florian
Montreal, Quebec, Canada
 CONTENTS

i. GENERAL MODELLING CONSIDERATIONS

MURCHLAND, John D., "Traffic Equilibrium Methods:
The Structure of Model Research" . . . • • .
IHGAN. Mark R., "Equil ibrium Models in Use: Practical
Problems and Proposals for Transport Planning" 11

ii. THEORETICAL STUDIES

HALL, ~lichael A. and PETERSON, Elmor L., "Highway-
Traffic Equilibria Analysed Via Geometric Programming" ........... 53
DAFERMOS, Stella C., "Integrated Equilibrium Flow Models
for Transportation Planning" • . . . . . . . . . . . • . '106
BECKI-1ANN, Martin J., "Equilibrium Versus Optimum in
Public Transportation Systems" .. ·119
JENSEN, Arne, "Statistical Equilibrium" . . . . . .132
HAMERSLAG, R., "Transportation Model with Elastic
Constraints" . . . . . . . . . . . . . . . . .147

iii. COMPUTATIONAL METHODS

NGUYEN, Sang, "A Unified Approach to Equilibrium
Methods for Traffic Assignment" . . . . . . .
RUITER, Earl R., "Network Equilibrium Capabilities for
the UMTA Transportation Planning System" . . . . .
NESTLE, Richard E., "Application of Error Measures to
Traffic Assignment Results" . . . . . . . . . . . . 200

EVANS, Suzanne P., "Some Models for Combining the Trip

Distribution and Traffic Assignment Stages in the
Transport Pl anning Process" . . . . . . . . . . . 201

BURRELL, J.E., "Multiple Route Assignment:

A Comparison of Two Methods" . . . . . 229

iv. DYNAMIC TRAFFIC ASSIGNMENT METHODS

YAGAR, Sam, "Emulation of Dynamic Equil ibrium
in Traffic Networks . . . . . . . . . . . . . . . . . 240
MERCHANT, Deepak K. and NEMHAUSER, George L., "A Model
and an Algorithm for the Dynamic Traffic Assignment
Problem" . . . . . . . . . . . . . . . . . . . . . .
 xx

v. TRAFFIC EQUILIBRIUM AND TRAFFIC CONTROL

GARTNER, Nathan, "Area Traffic Control and
Network Equilibrium" . . . . . . . . . .274
ORTHLIEB, Maxence P. and MAY, Adolf D., "Evaluation
of Freeway Corridor Assignment Equil ibria" . 298

vi. MODEL COMPONENTS

NORLOK, Edward K., "Supply Functions for Publ ic
Transport: Initial Concepts and Models" . ·322
GAUDRY, Marc, "A Note on the Economic Interpretation
of Delay Functions in Assignment Problems" .368

vii. APPLICATIONS OF TRAFFIC EQUILIBRIUM METHODS

FLORIAN, Michael and NGUYEN, Sang, "Recent Experience
with Equilibrium Methods for the Study of a
Congested Urban Area" . . . . . . . . . . . . . .
VAN VLIET, Dirck, "The Choice of Assignment Techniques
for Large Networks" . . . . . . . . . . . . . . . 396
LEBLANC, Larry J. and MORLOK, Edward K., "An Analysis
and Comparison of Behavioral Assumptions in
Traffic Assignment" . . . . . . . . . . . . 413
JUDGE, E.J., "An Investigation of Some Problems
in the Calibration of Traffic Assignments" . . . . . 426
GENEST, Bernard-Andre, BURDEN, David, MUGUET, Jacques
and KARDICALI, Esat, "The Use of Traffi c Equil i bri um
Based Problem-Oriented Language for a Governmental
Feasibil ity Study" . . . . . . . . . . . . . . . . . .427
 CONFERENCE PARTICIPANTS

CLAUDE ACHIM, Departement d'informatique,

Universite de Montreal
SYLVAIN ASSER, Departement de mathematiques,
Universite de Laval
~1ARTIN J. BECKMANN, Institut fur Angewandte Mathematik,
Technische UniversitMt MOnchen and Brown University
JON BOTTOM, ~assachusetts Institute of Technology,
Cambridge
DAVID BRANSTON, Centre de recherche sur les transports,
Universite de Montreal
DAVID C. BURDEN, Ministere des Transports,
Gouvernement du Quebec
J.E. BURRELL, Burrell ~ Associates,
London, U. K.
ROBERT CHAPLEAU, Ecole Poly technique, Departement de Genie Civil,
Universite de Montreal
CLAUDE CHRIQUI, Ecole des Hautes Etudes Commerciales,
Universite de Montreal
STELLA DAFERMOS, Brown University, Division of Engineering,
Providence
ROBERT B. DIAL, UMTA, U.S., Department of Transportation,
Washi ngton
SUZANNE EVANS, University Walk, Computer Centre,
Bristol, U.K.
JACQUES FERLAND, Departement d'informatique,
Universite de Montreal
MICHAEL FLORIAN, Departement d'informatique et Centre de recherche
sur les transports, Universite de Montreal
NATHAN GARTNER, Operations Research Center,
Massachusetts Institute of Technology, Cambridge
MARC GAUDRY, Departement de sciences economiques et Centre de
recherche sur les transports, Universite de Montreal
HOWARD GAUTHIER, Department of Civil Engineering,
Ohio State University, Columbus, Ohio
BERNARD-ANDRE GENEST, Ministere des Transports,
Gouvernement du Quebec
 XXII

tHCHAEL GODFREY, iJepartment of Civil Engineering,

Ohio State University, Columbus, Ohio
JEAN GRANGER, D~partement de G~nie Civil,
Ecole Poly technique, Universit~ de Montr~al
R. HAMERSLAG, Dwars Heederi k en Verhey, BV,
Netherlands
BRITTON HARRIS, Department of City and Regional
Planning, University of Pennsylvania
LARRY JENKINS, Department of Industrial Engineering,
University of Toronto
ARNE JENSEN, Institute of Mathematical Statistics,
Technical University of Denmark, Lyngby
BILL JOHNSON, Transportation Development Agency,
t10ntrea 1
E.J. JUDGE, Institute for Transport Studies,
Leeds Polytechnic and Leeds University, U.K.
GEORGE KOCUR, Cambridge Systematics Inc.,
Cambri dge
HAROLD KUHN, Department of Mathematics,
Princeton University
MAURICE L'ABBE, Vice-recteur a la recherche,
Universit~ de Montr~al

LARRY LeBLANC, Department of Computer Science and

Operations Research, Southern Methodist University,
Da 11 as

PAUL MICHEL LION, Department of Civil and Geological Engineering,

Princeton University
WIRVIN MANHEIM, Massachusetts Institute of Technology,
Camb ri dge
DEEPAK K. MERCHANT, Graduate School of Management,
University of Rochester
JACK MORGAN, Transportation Development Agency,
Montreal
EDWARD K. MORLOK, Department of Civil &Urban Engineering,
University of Pennsylvania
JACQUES MUGUET, Ministere des Transports,
Gouvernement du Qu~bec
JOHN D, MURCHLAND, Planning and Transport Research and
Computation (International) Co. Ltd" London
SANG NGUYEN, D~parter.lent d'informatique et Centre de recherche sur
les transports, Universit~ de Montr~al
 XXIII

GEORGE L. NEMHAUSER, Department of Operations Research,

Cornell University
RICHARD NESTLE, Department of Civil Engineering,
Massachusetts Institute of Technology, Cambridge
RICHARD L. oRAM , UMTA, U.S., Department of Transportation,
Hashington
MAXENCE P. ORTHLIEB, Institute of Transportation and Traffic
Engineering, University of California, Berkeley
JEFFREY OSLEEB, Department of Civil Engineering,
Ohio State University
ELMOR L. PETERSON, Department of Industrial Engineering and
Management Sciences, Technological Institute, Northwestern University
EDHIN J. ROBERTS, Transportation Systems Center,
Cambri dge
PIERRE ROBILLARD, Departement d'informatique,
Universite de Montreal
H.J. ROODENBURG, Netherlands Institute of Transport,
Rijswijk, The Netherlands
EARL R. RUITER, Cambridge Systematics Inc. and Massachusetts
Institute of Technology, Cambridge
SIV SCHEELE, Mathematical Statistics,
Stockholm University, Sweden
DIRCK VAN VLIET, Institute for Transport Studies,
The University of Leeds, U.K,
M.R. WIGAN, Greater London Council,
U.Ko

CHRIS WITZGALL, National Bureau of Statistics,

Washi ngton
SAM YAGAR, Department of Civil Engineering,
University of Waterloo
 TRAFFIC EQUILIBRIUM ~'ETHODS:

The Structure of Model Research

John D. tlurchland
PTRC (UK) Li~ited

London, England

Introduction

This is not a paper giving the results of traffic model research, but an expres-
sion of opinion as to how such research should be structured.

Scientific development usually shows a pattern of breaking larqe problems into

parts, analysis of the parts, and then re-assembly to see ~lat should be done next.
Often a correct understanding of the parts is the only way of finding out that the
first conceptualization was wrong, and that another must be tried.

Because transportation modelling must include so many variables and relationships

it certainly counts as a "large problem". The main plea of what follows is that we
should be more self-conscious about the conduct of this research, and the goals and
sub-goals which we are currently pursuing.

The writer's experience includes academic research and also practice. Practice
can fairly be described as a bed of nails. The analyst squirms to accept so many
painful contradictions - between the legitimate demands of the decision-makers and
the applicable scope of the modelling and evaluative procedures, between the time-
liness of the results and their degree of credibility, and the many inconsistencies
within the model - that we become either fakirs or fakers.
 2

The purpose of traffic models

The purpose can be put briefly:

The aim of analysis is to assist the selection
of practicable and beneficial transportation
actions,
though each of the words, "actions", practicable", beneficial",
"selection" and "transportation" deserves further discussion.
Passing on however, the distinction between decision-making and
scientific study needs emphasis.
Decisions are the product of decision-making.
The product of scientific study is objective
understanding.
Set out like this the distinction is evident, but in actuality
it has to be struggled for. Often results are both indecisive and
unscientific.

Taking assistance to decision-making as the primary goal, it could be

that scientific understanding is not really needed. This is surely not
the case in so complex a field as transportation. However, scientific
study is not conducted in the same way as analysis for decision-making -
it requires, to mention several pertinent aspects, adequate time to carry
out, open publication to permit feedback and critical discussion,
independent confirmation of key results in varied circumstances, and
accumulation of results over time. In additon, to be useful for
decision-making, scientific studies need to be directed to the right
problems, and the results communicated back to those who should use them.

Evaluation and Forecasting

These remarks aim to show the contribution that mathematics should

make to traffic modelling - and the emphasis on the future and
implication that in the past the contribution has been small is intended.
"Mathematics" is used for want of a better term here, to include
mathematical analysis in the narrow sense, economics, statistics and
computational analysis.
 3

The general aim of selection between alternative transportation

actions implies that the primary output of the transportation model
must be an evaluation. Except for scholarly interest, models without
evaluation are valueless.

Some transportation decisions concern the near future, say the

next year or two. This applies to road traffic management measures -
one-,vay streets, traffic signals and the like, to parking control
and to many aspects of public transport by bus. For such short-run
problems it would appear that marginal change models would be most
appropriate, though these have not yet been intensively studied.
But since all significant construction takes a number of years, and
land-use development or adaption from one to several decades, the
usual demand on the model is that it produces evaluations for one
or more future years. (Quite a few 'nails' rise from the inability
to meet the obvious need for treatment of a number of future epochs,
but I pass over this.) Thus the primary model output required is a
forecast evaluation. It may seem surprising to an economist reading
this to find an emphasis on the elementary point that a model
is irrelevant to decision-making unless it provides an evaluation
of the future situation concerned. Unfortunately, the emphasis is
needed.

Because of the economy of data collection, and I suppose for

managerial reasons also, transportation models in practice are based
on data collected in a only single year (and, to be precise, from
a few days in that year). Parameters in the model are chosen to
give a best fit to this single year data. Forecasting is made
by accepting independent variable forecasts from land-use planners,
and by keeping the fitted parameters constant. Serious consideration
of the forecasting problems and forecasting errors has been absent.
The work of the econometrists who would, I imagine, deplore this cross-
section model, is unknown. Instead, the naive view that the goal of
accurate future evaluations is equivalent to an accurate fit to
base-year behaviour, holds the field.
 4

To be quite particular about this lack of attention to

forecast evaluation, let Cle mention recent work on developing what
are called disaggregate demand models. These models imply many
kinds of elasticity, and have many parameters. Evaluation hasn't
been mentioned in what I've seen or heard. Estimation is made by
fitting observations at one time point (in quite a refined statistical
manner). Forecasting is referred to, but is expected to be made by
keeping the fitted parameters fixed. Even from the behavioural point
of view, it's curious that people's actions are supposed sensitive
to everything except the laps~ of time.

The three kinds of error

In what follows the goal of forecast evaluations will be taken

as implicit, and the structure of model research considered more
abstractly. However, it is not enough merely to produce a forecast
evaluation for each future transportation alternative. We need to
know the accuracy of these evaluations - are the worths of the
alternatives significantly different or are they not ?

Errors may be systematic (e.g. a bias in the evaluation in favour

of one mode) or random. The former include some of the sharpest
of the nails. The non-systematic error will be a compound of three
separate errors:
inherent model error
data sampling error
computational error.

These three types are quite distinct conceptually. It's often

hard to separate the first two numerically (that is, to distinguish
the inability of the model to completely represent the situation
from data sampling effects). The third is within our control, and
with sufficient mathematical analysis should be eliminated.
 5

Little serious attention has so far been paid to estimating

these errors in transportation modelling. Indeed, the computational
errors remain unassessed even in the iterative calculations of distribution
and congested (capacity-restrained) assignment which we currently
use. Fox and Hayers (1968) could easily
illustrate their critical remarks from our field. "The 'scientific
method' proceeds by building up thepries from observations and
testing the theories on additional observations. In this respect
errors of measurement have to be separated from errors of
computation, and Lf the latter are too big, or cannot reasonably
be estimated, the 'scientific method' becomes a meaningless
shambles."

A paradigm for model research

The following seems to the writer a reasonable way of dividing

the development of models into distinct tasks. This division should
make it much easier. apart from what we can do ourselves, to pose
our problemsto experts on particular techniques - mathematical
programming, algorithm convergence, statistical estimation -
who do exist but have seldom been called on to help. These
five parts are as follows.

1. Explicit formulation of postulates, for travel demand and

supply.

2. Logical analysis of the postulates, individually and

together, especially to determine whether there is one
and only one solution.

3. Invention and study of solution algorithms,proving convergence,

finding the rate of convergence and measuring the final
accuracy.

4. Development of statistical estimation methods for the

parameters in the postulates, from sample survey data.

5. Evaluation.
 6

I hope these five tasks for mathematicians are clear, and the
distinctions persuasive. Because of the complexity of
transportation modelling it is evident that none of these tasks are
trivial.

The postulate formulation of 1. will depend substantially on

empirical knowledge and preconceptions of model structure, though
there are also mathematical considerations. At the least
we want to have sufficient tractability that the analysis of
implications of 2. and development of numerical solutions of 3.
may be made. 2., 3. and 4. are technical exercises of various
kinds. Evaluation, 5., is of a different nature. It is very
closely allied to 1. aIld perhaps should not be separated from it.
It requires the logical ~pplication of economic principles,
and welfare theory in particular. I suspect we have also something
to gain here from measurement theory ( and would very much like
e~perts in this to explain their achievements in simple terms).

Several brief comments on the different tasks. The postulate

analysis of 2. may show that postulates work well when taken
separately, but fail to do so when taken together - for instance
that there is no joint solution for the supply model and the demand
model. Even though the postulates seem quite simple, this analysis
may be a hard mathematical problem. Computationally, we ~ require
that the problem addressed has a unique solution, so the establishment
of this in 2. logically precedes a demonstration that a particular
algorithm converges in 3. However, a numerical demonstration might be
the easiest way of showing the possibility of multiple solutions.
It's a pity that every individual algorithm seems to need its own,
difficult, proof of convergence.

Investigating the rate of convergence can be done either

theoretically or by testing with problems of known sOlution. (The
latter can usually be constructed by working backwards.) Testing
with known problems can be easier to do. We should bear the following
in mind, quoted from Scarf and Hansen, 1973. "General argumepts
give little indication of whether the algorithm will be useful in any
specific numerical problem. For example, the simplex method of linear
programming derives its importance as a practical computational
 7

device not from any theoretical bounds on the maximum number of

possible pivot steps required to solve a problem of given size,
but rather from the empirical observation that the method generally
terminates in a surprisingly small number of iterations". :;earer home
the shortest distance algorithms used in traffic assignment are an
order of magnitude better than those discussed in operations research
journals despite the upper bounds presented for the latter. All the
same, it is helpful when the convergence rate is amenable theoretically,
To refer to the distribution model, whose computational errors were
said to be unknown above, it is not too hard to show that
convergence is ul timately linear. (Specifically, the relative
errors in the column balancing factors at the end of an iteration
become a strict weighted average of the relative errors at the
beginning).

Estimating the accuracy of the solution at termination is a

significant problem. Presumably the widespread illusion, that a
small change in the solution over an iteration is a sufficient
indication, can be eliminated by experience with known test problems.
(Fox and Mayers call this false dawn 'consistent inaccuracy'.)

Equi li brium

So far these remarks have managed to skirt the notion of equilibrium

which is the chief theme of this Symposium. The book by Scarf and
Hansen The Computation of Economic Equilibria of last year, is
intended by them to explain the theory simply and to exhibit the
successful algorithm now available for direct solution of general
problems.

From their account it appears that the term "equilibrium"

is not used in any restrictive technical sense in economics - merely
that values for the variables exist which satisfy all the necessary
relations, in problems which have an economic, supply-deI:!and, flavour.

Looking at this from the viewpoint of model analysis and

calculation, a demand for an equilibrium treatment is a demand' for
complication. Equilibrium implies action and reaction, and the
coupling together of aspects which could otherwise be treated
separately. How much simpler is theory and computation if total
 8

travel depends in no way on the transportation system provided!

The theoretical tools for estabiishing that an equilibrium exists

are the fixed-point theorems of Bromver and Kaku tani. The
practicality of a good theorem like these can be illustrated by a
problem from traffic assignment. The "stochastic" model of route
choice due to R.b. Dial (1971) was originally put forward for the
uncongested situation. In extending it, the first question of
whether there is still a solution in the presence of congestion,
granted that the set of admissable paths is fixed, is readily
answered by a didactic application of Brouwer's theorem. Indeed,
the demonstration applies to ~ principle of route choice which is
continuous in all the variables involved. (When the paths are
selected in the "stochastic" [ilodel dependent on the state of
congestion, for computational reasons, the answer is unfortunately
negative. Obviously, the selection introduces discontinuities.)

Uniqueness of the equilibrium is quite another matter. Here

we must hope that the problem is simpler than the general one - for
instance, can be expressed as a convex mathematical programming
problem, as Beckmann, Naguire and Wins ten long ago showed the
traffic assignment problem to be (1956). There are many important
and challenging problems here, both in f-ormulation and in finding
effective algorithms. They have the flavour of 'multi-agent'
problems, i. e. where several independent agents - travellers, the
road system, the community, transport operators - each have variables
they manipulate to achieve their own goals. We would rather deal
with a structure much simpler than the mathematical game of the
evident formulation, if it all possible.

Conclusion

After a series of remarks, concise in some parts and rambling

in others, let me be clear in the conclusion.

For the contribution which mathematicians can make to

the development of transportation modelling, a fivefold paradigm
for the structure of such research has been put forward. The five
tasks are postulate formulation, consistency analysis, algorithm
 9

study, statistical estimation and evaluation. This paradigm is

proposed within an emphasis that the real goal is forecast evaluation,
and that the errors due to the model, to data sampling and to
computation are to be taken as errors of the forecast evaluation.
A much clearer division of the problem into parts, each with its own
criteria for success, will help a sharper and more intense attack,
and make it easier to apply known analysis from outside this field.

It will not be sufficient to solve the research problems

indicated in a purely mathematical way. Research is not fruitful
until the results are communicated. This communication needs care
and attention. It will be helped by a general acceptance of this
division of the whole problem into logically distinct smaller
parts, and also by simple numerical illustrations.

What has been said above has contained some severe criticisms
of the current state of transportation modelling. They were meant
constructively, and I am optimistic that much progress is readily
possible. Referring to the bed of nails metaphor - of which I
was at pains to point out only a few - I do believe that many can
be drawn, and some of the rest at least blunted.

J.D. Murchland
 10

References

Dial, R.B.
A probabilistic multipath traffic assignment model
\.hich eliminates path enumeration, Transportation
Research 5 (1971) 83-111

Beckmann, M.J., McGuire, C.B. and Winsten, C.B.

Studies in the Economics of Transportation, Yale
University Press (1956). Now available through
University Hicrofilms, OP 40965.

Fox, L. and Mayers, D.F.

Computing Methods for Scientists and Engineers,
Clarendon Press Oxford (1948) vii

Scarf, H. and Hansen, T.

The Computation of Economic Equilibria,
Yale University Press (1973) 60
 EQUILIBRIU~l ~10DELS IN USE; PRACTICAL PROBLEMS

AND PROPOSALS FOR TRANSPORT PLANNING

~1ark R. Wigan
Greater London Council
London, England

ABSTRACT

The main efforts in the field of equilibrium analysis have been directed towards
the production of adequate theoretical tools. Such work is intended in the field
ana lys is to be used in app 1i cat i on to real problems. l~e cons i der here the types of
problem for which equilibrium analysis is appropriate or necessary. For some pro-
blems its use is essential; traffic restraint by fiscal means is one of these, and
is used to illustrate the practical problems and issues that can arise when equili-
brium methods are applied. The use of such techniques stimulates new evaluation,
analytical, and assessment viewpoints, and the practical and theoretical responses
required to serve these perspectives are indicated with the help of a range of stu-
dies where equilibrium methods have been usefully employed.

INTRODUCTION

The basis of forecasting traffic flow and travel demand over a network is the
balance of the efforts of travellers to minimise the effort and expenditure that they
must make to travel, and the ability of the physical system to provide an adequate
service level at the cost and demand levels that the travelling public cause to
appear. Problems arise as soon as hypotheses are introduced to allow this competi-
tion to be simulated. Any model is open to criticism that it ignores or does not
pay adequate attention to one consideration or another; far fewer ought to be open
 12

to the criticism that they do not represent a coherent and internally consistent
representation of the system in question. Even when such questions are asked and
answered, the implicit assumptions that accrete around useful but acknowledged imper-
fect models tend to become blurred and no longer get the attention that they need and
deserve when the use of such models is considered in a novel or especially demanding
context. The question of possible sources of systematic error is almost invariably
overlooked.

The simplest model of network traffic flow is the all or nothing traffic assign-
ment. This is so well known that it is an ideal illustration of this thesis.

The representation of the network is-necessarily-incomplete, and the size of the

zones is usually matched-roughly-to the degree of detail left in the network of roads.
The size of the zones, if either too large or too small for the scale of the road net-
work left in the simulation, gives rise to a disproportionate fraction on the mileage
of the network as a result of the zonal connectors, which are neither road links nor
not road links. Even before we consider the trip matrix and calibration processes,
it is not obvious what this model is simulating. Is it simulating all the trips over
a given length on a subset of the roads? Is it simulating some of the trips on all
the roads included? If there is a great mismatch between traffic flows on given roads
and the observed flows, is this due to the choice of zone size? The degree of network
detail. An error in the traveller routing assumption? An error in the travel times
used or the road characteristics as represented in the model? As most assignments
can be tuned by adjusting any or all of these factors, it would be wise to pause
before judging an assignment technique as adequate or inadequate simply because the
fit to observed flows was good. The trip matrix is the next source of problems. If
it has been constructed on a 24 hour basis, and been factored down to give the matrix
of movements for the representative hour under study then the basis on which the
matrix has been constructed is pertinent: it now represents the mean ~xpectation of
the averaged flow sampled over the whole day and factored on an average basis to give
a representative flow for a specified hour. It is pretty difficult to go out and
check this matrix against measurements of flow when this litany of qualifications has
 13

been rehearsed. The travel speeds on the links are next to draw fire: what are
these speeds? As we are discussing only an all or nothing assignment at present,
there is no question of speed flow or other capacity variation functions, merely the
simple question - what is the speed on this link? It mustbe a space mean speed, as
this is what the traffic model is actually working with. The specification of links
as directed arcs makes very sure of that. Each link has only a topological meaning
in the network, and the travel time between the end points is therefore a space mean
speed measurement over the link as a whole. This too is not quite as easy to match
with observation as might seem possible. The time mean speed is very much easier to
measure. This is not enough. The average representative hour with which we are
working - q.v. the tri"p matrix - demands that we take the hourly average mean expec-
tation of the space mean speed over the whole link in order to confirm its value.
The routing assumptions are reached at this point, and the simplistic assumption
that all travel takes place on a single shortest route can give very good results on
uncongested networks, and indeed on congested networks as long as an artful selecti-
vity is exercised on the zone sizes, network detail, and positioning of zonal connec-
tors. The reason for this ability to provide acceptable assignments using all or
nothing assumptions is twofold. First, that the matching of data to assigned flows
is problematical to say the least, and second, the degree of flexibility inherent in
the network building and zone size selection stages allows the calibration phase to
match observation without any great appeal to the routing assumptions or effects.

For congested assignments all the above qualms apply, but in addition the speci-
fication of the speed flow or other capacity variation function lends further com-
plexities to the specification process. Paradoxically a full congested assignment
is easier to handle than an all or nothing calibration, as there is no freedom to
alter the speeds on each link or the routes taken without going through a full equi-
librium process once more. This means that the dependence on the rapid dying away·
of local tuning variations that occurs in all or nothing calibrations simply is not
available under a full congested assignment. Consequently the speeds that are finally
predicted for 'now' by such a model may well not agree with the observed speeds
(assuming that the correct form of measurement was made of these speeds). In order
 14

to tune a model of this kind there are two different courses of action. The first
is to alter the level of detail on the network, and thus the capacity made available
to the trip matrix. The second course is to vary the parameters of the generalised
cost function that controls the routing balance, or to adjust the parameters of cer-
tain speed flow functions. The first course of action should be unnecessary, if any
reasonable care has been taken to match the size of the zones to the level of detail
on the network and the percentage of intrazonal trips retained for loading. The
second course of action bears close examination. The variation of the balance bet-
ween time distance and out of pocket money costs in the routing cost function leads
to an alteration in the whole flow pattern on the network as certain parts with spare
capacity become attractive at the reduced cost of longer journeys, or the reverse.
Such a tuning process should be able to achieve a good balance between congested cen-
tres and the surrounding areas in the simulated network, but this does not have any
great effect on the variation of speed flow functions. Any such variation calls into
question the basis of a speed flow (or other capacity) function on a partial network
of the kind used for transportation studies. If a capacity variation function is
specified on a link by link basis, the effects of junctions must also be defined.
The performance function of a network is an amalgam of link, junction, capacity func-
tions, and the routing patterns through the area concerned. Any network speed flow
function is therefore extremely difficult to define, and harder still to measure in
a manner which will allow major variations in flow patterns in the forecast stage to
be simulated consistently. It is not clear whether area speed flow relationships can
be defined unambigously for such major fluctuations in travel pattern, but in experi-
ments on a 4000 link network it proved to be possible to reproduce the average speed
of flow over the four areas for which speed flow functions had been defined. The
same speed flow function was assigned to all roads in the area, and no account was
taken of the known variations in capacity of the individual roads. The pattern of
traffic flow, the network topology, and the routing cost function act~d together to
define a unique equilibrium flow pattern over the network which could be matched to
the observed average flow speed over the same broad areas over which a single speed
flow function had been fixed for every road. By varying the parameters of these
 15

speed flow functions it proved to be possible to fund a unique set of parameters

under which the average speeds in all areas matched the observed averages. This was
quite difficult to achieve, as any change upset the entire pattern of flow over the
network, and equilibrium had to be reattained once more. In spite of this simplistic
manner of defining 'area speed flow', the flows on individual roads were in good qua-
litative agreement with the measured flows. If any account at all had been paid to
the fact that some roads were dual lane, segregated highways, and others back streets
with many junctions, it would have been quite easy to have obtained a good overall
assignment in quantitative terms. The size of the network made little or no differ-
ence to this experiment, and the only visible effect of its size was a slowing down
in convergence rate as equilibrium was neared as a direct result of the accumulated
effect of small variations on many links as trees were built.

In this section I have tried to illustrate how elusive the idea of 'validating
an assignment against observation' can be, without going into the more subtle aspects
of route choice, stability, uniqueness, or the link between analytical checks on con-
vergence and simple comparisons between measured and simulated link flows. These
more theoretical problems are equally real, but gain more attention than they deserve
in comparison with the more pragmatic calibration and interpretation difficulties
covered here.

WHERE EQUILIBRIUM MODELS ARE NEEDED

The use of models to simulate or forecast traffic flows and travel patterns has
two distinct aims: (1) to reproduce measured circumstances in such a way that alter-
native ways of dealing with present conditions may be assessed: and (2) to forecast
future conditions for planning or decision purposes. While the technical requirements
for a good match between supply, demand and user reactions is needed in both cases,
the utility of refined techniques may not outweigh the cost of their application
under some circumstances. Equilibrium analysis is not necessarily required for some
forecasting applications, but may be a prerequisite for any short run assessments.
The scale of the problem is not necessarily a good guide to the choice or need for an
 16

equilibrium method. Table 1 indicates some of these scale factors.

Table 1

Scale Object Elements Tools nvironment Equilibrium

Regional Strategy Modal balance Coarse LUTS * model s Nil Yes

Landuse plans Lowry Models
Regional Spatial interactior
economics

Subregional Rapid survey Direct data Special languages Nil No

of pol icy access: Interactive front
options Interactive ends, data base
models and access adaptive
displays models display
chained tools
regional and
local models

Urban Strategy Behavioural LUTS * SAA(l} Yes

tactical models PANIC (2)
design network
geography
modal split
cost models

Local Strategy of ? Elements of Nil Yes

urban minor ele- both LUTS and
ments tacti- microanalysis
cal design

Local Tactical Queues DHTM(3) None No

micro design lanes TRANSvr(4)
local traffi c control
policy parking etc
behaviour

Working down the scale, regional models have been set up to employ equilibrium ana-
lyses: DODOTRANS (5) is one example although the equilibrium process built in does
not-and-cannot-always produce good results, the objective of the equilibrium process

* Land Use and Transportation Studies

 17

is to achieve a reasonable spread of effects on supply and demand over the whole
regional spectrum of networks, which are held in a skeleton or coarsely agregated
form. This is a perfectly reasonable and useful level of precision at this scale,
and it is doubtful if any better methods are really worth using for these applica-
tions, as the many assumptions on aggregation and representation rehearsed in the
first section of this paper are most likely to outweigh any increase in apparent pre-
cision produced by more theoretically densible methods. If however such a model is
used for the setting of finely balanced competitive fare and service frequency stra-
tegies, rather than for broader questions, then the relative consistency and internal
precision with which the model and the built-in equilibrium search is carried out
becomes of prime importance. Ironically, for practical purposes, such fine tuning
searches are probably best carried through without any use of equilibrium methods,
simply because the internal consistency of the model and its realisation of the con-
sequences of different options is maintained by the use of a simple all or nothing
approach. Clearly, if one situation is severely congested and the other is not, then
it would be tempting to use equilibrium methods in both cases, but unless the con-
gested results reach a high order of stability such a course of action could disable
the model from producing useful differential comparisons.

At the subregional level sketch planning comes into its own. The indicative
nature of the process and the necessary coarseness of all types of representation
used, make it unreasonable to use an equilibrium approach. For certain special pur-
poses a variant on the simple all or nothing view can be used. This is one than can
be, and often is, confused with a true congestion responsive equilibrium; an all or
nothing view is taken, the results of taking this view traced out, and the supply
functions adjusted in consequence of the flow and demand results obtained. The all
or nothing process is then repeated until no adjustment is needed. This provides a
clear implicit statement of the level of internal precision reached, but the price
is not so often recognised; it is that a new assumption has been imposed on the
model, i.e., that!ll traffic on a given route uses the same route. This is frequent-
ly directly at variance with the motive for using an equilibrium method at all, when
a degree of congestion responsive route spreading is what is required. Although in
 18

many cases the impression is left the authors feel that they have achieved such a
state, clearly this cannot be so by the nature of the process description imposed
by the process itself.

At the urban scale the need to represent congestion in a future situation where
little or no congestion may exist in a base year, ensures that equilibrium techniques
are in demand. Once again, the use of repeated all or nothing single assignment of
traffic flows (on any number of modes) is frequently confused with the congestion
spreading of traffic over many different routes. The need for effective traffic
equilibrium models is very limited when major schemes of construction are the main
focus of attention. This is no longer true when only minor schemes are under study.
For convenience this type of problem has been labelled 'Local Urban', and comprises
a major gap in our strategic armoury of analytical tools. The need to assess a large
number of minor measures applied simultaneously to many pOints over a broad network
to achieve a policy objective places heavy demands on effective and consistent con-
gestion spreading equilibrium techniques. It would not be overstating the case to
say that without such techniques it is unlikely that the results of such an analysis
would be of any practical value. This question has been explored using a set of
different models of a large town by means of several different techniques. Further
details are given later.

At the level below 'Local Urban', we have the 'Local Micro' analyses: these
are typified by the Stanford DHTM (3), by Giannopoulos' (8) discrete model, and to
a lesser extent by TRANSYT (and its kin) signal setting models of the micro behaviour
of local traffic. Equilibrium analysis at this scale is rarely of any practical
value, as the detailed simulations carried out either include time as an explicit
dimension within the model itself, or it is equally explicitly excluded in order to
reach a time invariant simulation.

The most interesting area for practical analysis at present is the Local Urban
scale, where transport network representations are used to marshal many short run
tactical options into a single overall policy. With the move against road and faci-
lities construction in both US and UK, the short run low capital cost options of
 19

this kind are critically important tools for achieving transport and planning goals.

Hand in hand with this increase in importance of systematic traffic management

has been an increasing need to deduce the environmental consequences of any trans-
port measures. Indeed, the motive for many such schemes for systematic traffic mana-
gement is frequently solely the achievement of environmental gains. The system of
lorry routes which may now be set up in the UK as a result of the Head Commercial
Vehicles Act (1973) - popularly known as the Dykes Act ~ is a good example of the
class of proposal that must now come into the analytical framework.

Some experiments have been carried out using a full equilibrium model of a lar-
ge city, and the lorry and non-lorry traffic treated separately. The lorry traffic
is inelastic in the short run, but the effect of lorry concentration on the private
cars produces restraint through congestion, and the spreading of traffic over the
network as a result of lorry routes has two different types of effect. The restraint
on private cars through congestion is the first, and the environmental improvement
is the second. The environmental improvements come from the shift of sma11 numbers
of lorries from lightly trafficked roads onto already busy streets. The congestion
spreading effects on private cars dilute these beneficial influences, but there is
a net gain in noise, and indeed in air pollution terms. The technical requirements
for such models are difficult to meet, as the effects are all made up of tiny chan-
ges, all in substantially the same direction. To distinguish consistently between
different policies that differ marginally in this general manner, places heavy
demands on the quality of the equilibria achieved, and the need to include many
minor roads expands the problem size substantially. The need to handle two types
of traffic at once, and with elastic traffic demand levels, presents fewer theore-
tical problems than practical difficulties. The sheer time required to run through
such multiclass models on real networks cannot be discounted. The results are most
encouraging, and the variation in both network detail and overall trip matrices
that has also been introduced into this study has allowed us to confirm that the
lorry routes results are robust and repeatable even when brand new base equilibria
must be generated and the results compared of what is in effect three separate and
 20

independent equilibrium models, differenced to give comparative estimates of the

effects of the diffuse policy of lorry routing.

This example has been pursued for several reasons:

(1) It is a case where a real practical short term need has genuinely required
the full panoply of equilibrium analysis.

(2) It has been actually applied in a realistic scenario with credible indica-
tive results of excellent robustness.

(3) It raises the whole class of equilibrium problems called into question by
the need to develop environmental impact and evaluation statements within
the framework of transportation planning.

Submitted for publication early in 1973 was a systems analysis and practical
modelling technique designed to bring the estimation of environmental impacts and
their variations with different policies into the transportation planning program
format (3). This has since been put into practice and matched up to specially deve-
loped and otherwise available forecasting equations for noise levels, many different
forms of air pollution, and lead concentration. An essential need for such analyses
to be effective in practice is that congestion variations be adequately represented
in the host transportation model, and that the minor roads (where on most of the
affected population resides) be represented explicity.

The needs of environmental impact estimation cut right across the underlying
assumptions of much of present day transportation analysis. The level of network to
be included must now be set to specifically include those minor roads with little
traffic flow, and little expectation of any large flows in future: no longer can
these roads be efficiently excised at an early stage in network design, on the basis
of cordon counts and other checks designed to concentrate attention on roads that
actually have traffic on them in fair quantity.

The changes on these minor roads are now of greater importance than the changes
on the major routes, as the large %variations arising from small numbers of vehicles
 21

being encouraged to vary their routes by congestion or cost are precisely the points
at which the main environmental changes occur, On major routes the saturating effect
of extra vehicles on noise levels acts to reduce the environmental interest of these
routes, but for the needs of compensation payments. For legal purposes in the UK the
18 hour L10 is a key index. If this varies by 2dBA or over as a result of road alte-
rations or construction, then compensation is payable if the final level is over 68
dBA. Consequently we are in a position to assess the compensatory payments that a
new road may induce with some precision, but as almost all of the benefits will accrue
on the minor and lightly trafficked roads which are so underrepresented in our current
models - and so badly modelled when they are included - we are in no good position to
assess the benefits from any scheme on the environment.

Environmental impact also raises another awkward conventional assumption into a point
of some concern: impact is a product - in general terms - of a concentration of level
of a pollutant or irritant at a point and the number of people exposed to this level.
Up to this point the specification is consistent with transportation planning models
of conventional form (e.g., PANIC (3); but once the next element of this product is
added it is no longer true: exposure is the time for which a group of people are
exposed to a given level in given numbers, and the exposure of people in a shopping
street has a quite different time profile to the exposure of households in residen-
tial streets. The first solution to this difficulty is to apply simple scaling fac-
tors to streets of different kinds to take account of the fact that the effective
population present at a given time may be very dffferent to the resident population.
Similar artifices can be applied to pedestrian activity in shopping areas. As soon
as the evident inadequacies of this device are appreciated, the pressure becomes overt
in the attempt to obtain a time varying profile of traffic flows in both peak and off-
peak hours. If this need is to be satisfied then very careful thought must be given
to the manner in which the data must be collected to calibrate such a class of
analysis.

Although this discussion has been focussed on the basic problems raised by the
inclusion of environmental factors in analysis at the conventional urban planning
 22

scale, the real problems emerge when this line of approach is added to the conver-
gent demands of area traffic control. systematic local traffic management, physical
closure policies, and other very local measures which may not be used in concert to
achieve a strategic objective by a co~rdinated application over a wide area simul-
taneously. Area traffic control adds yet another element to this mix, as such con-
trol systems can be used to serve several ends by the choice of different objective
functions used to control their operation. Such small scale controls work on indi-
vidual junctions, affect many different minor roads, and produce an accumulation (of
tiny alterations to the behaviour of traffic over the network, shifting congestion
from one point to another, with small gains - or even losses - at each point. This
requires a full equilibrium model to even hope to reach a strategic assessment. It
is interesting to note that TRANSYT accepts implicity this type of overall objective
logic in its hillclimbing strategy for setting up traffic light systems in an area.
The elements at this level of analysis are queues, numbers of stops by vehicles,
turning movements, and a multitude of very detailed issues which affect traffic at
this level. This is the micro scale shown at the end of Table 1. No really effec-
tive models yet exist at this level, but when they are set up they must be of equili-
brium form if strategic goals are to be assessed with their aid. The present stress
on small scale local measures - often restrictive in their nature - makes it impor-
tant that some progress be made in this area. The environmental objectives of many
such local schemes require estimating frameworks to forecast pollution of different
kinds in terms of queues and similar elements. The data needs of systems of this
kind are daunting, but it is hard to see how these demands can be avoided in the pre-
sent climate - and equally hard to see how they could have been ignored for so long
and without such stimulation.

One compromise strategy is to develop special techniques for cordoning equili-

brium models so that differential scales of analysis can be followed through in dif-
ferent parts of the modelled area. Later in this paper a systems design is described
that has real prospects of making advances in these difficult areas practical for
those who are the final customers for such work. It is not unduly difficult to
design theoretical problems and solutions, but quite another matter to develop a
 23

real system with adequate flexibility, power, and speed to allow applications to be
made of the theories developed.

HOW EASY IS EQUILIBRIUM ANALYSIS?

The theoretical advances in the last decade are not as widely known as they
should be amongst transportation planners. The real question is not 'is equilibrium
analysis possible? 'but, 'is there a usable, economic and adequately robust techni-
que available?' The most popular device has been to use a mixture of heuristic
methods linked to a soundly based theoretical measure of convergence quality and
uniqueness. This was the method used at TRRL in the late sixties (7) and which has
continued to provide an effective platform for several programmes of equilibrium
application studies for various purposes.

The technique is disconcertingly simple, and achieves its end by picking out the
key problem of all or nothing assignment methods and treating it in a suitable manner.
The problem is caused by the use of a constant fraction of the total trip matrix for
a given loading or iteration. While this fraction might be precisely the size needed
to give a correct, or improved, flow loading on one piece of road, it is equally
likely to push the flows on other roads further from the equilibrium solution than it
was before the loading. Even if the fraction is a mere 1% on some roads this will
add up to 10, 20% or even more of the flow on that road, due to the use of a single
set of trees.

It is tempting to try to avoid this problem by introducing some multiple tree

device, such as those used by Burrell (8) and von Falkenhausen which randomise the
lengths of the arcs for the tree building process. This can on occasion produce a
slight improvement in the convergence of an iterative assignment, but at the last
analysis does not make any advance on the problem enunciated above. At the last
stage every flow is reassigned at the same fractional level as all the other flows
from origin to destination. If the process is halted at an arbitrary point when using
these randomising techniques, then it might have been supposed that the assignment
was 'better' in some way than one which seeks for a specified equilibrium.
 24

The behavioural assumptions which can be used to describe the randomised or multi-
routing methods have an appeal that can often blind the analyst to the fact that
there is still only one unique solution, and the heuristic used to reach it has no
great significance.

There are two essentially similar ways of breaking this coherence between origin-
destination flows: -

to carry out an !unloading! process as the last iteration;

2 to introduce a device to allow differential scaling.

The latter approach may either be employed with a direct estimation of a distribution
or generation function (such as the course followed by IMPACT (9)) or by means of a
null technique, in the best traditions of experimental measurement in the physical
field, where a device is introduced solely to allow differential scaling and is later
removed so that the original situation can be compared with the new version produced
(10) .

The first approach is an essential component of Wigan (10) and of Van Vliet's
(11) methods, and a small table of the options is given: -

Table 2

Elements of three equilibrium techniques in use

Method Unloading interations Differential scaling Null tests

Van Vliet (11 ) Yes No No

Hawkins ( 9) No Yes No

Wigan (10 ) Yes Yes Yes

 o is the point at which tolls have been removed.
L is the last point generated by the equilibrium model.
R is the point at which the last trips are added in again.
A is the 'Forced Adjustment' point for position R
0:::
0
~
en
0 £400
c.

en~
en
f §. £300
o~
~ 10

s-.~
",-Z £200
Ol 0:::
E'~
::I
eno
0:::-1-'
0 1\:1
u "0 £100 C11
Ol
s-
Ol
'+-
Ol
cr:
,f) 5 oL.RA

en 0::: ,
s-.~
OlIO Sta:@! " Stag e 2
E(!l .
::I Elasticity 0.1 Elasticity 2.0
S!~~,
1.0 ~
en
0:::
o
U
Tolls of 4d/mile in Central Area Tolls of ld/mile in Central Area Tolls of ld/mile in Central Area

Figure 1. REFINEMENT OF THE FLOW ASSIGNMENTS ON A CONGESTED NETWORK

AN ILLUSTRATION OF A RECOMMENDED STRATEGY
 26

Van Vliet developed his approach for large networks, where it has proved to be possi-
ble to produce acceptable equilibrium assignments using only five stages: four of
loading and one of selective unloading, The very size and potential variety of large
networks materially assists this approach, but for smaller networks it is unlikely
that the results would be anything like as good, judging from the use of the same
general type of method in experiments made as part of the development of null seek-
ing models, The huge number of small changes that arise when carrying out equilibrium
assignments and tree builds on large networks produces a level of noise as a result
of rounding and significance losses in the large number of elements accumulated along
the way, The null seeking methods are at their worst on very large networks, as will
be indicated later,

Hawkins (9) breaks the coherence of the assignment factors by computing the re-
quired level of journeys to be added from a distribution model, and working cyclicly
through the list of origins, adjusting the supply parameters as the traffic is pro-
gressively loaded onto the network, This technique will evidently give a fairly
effective and stable flow solution, but it is far less likely that it will actually
be the correct and unique solution of the joint assignment and distribution problem
posed, The cyclic nature of the assembly of the final flows effectively puts IMPACT
into the class of methods which obtain flow equilibria for fixed trip ends and then
recalculates the trip matrix before repeating the exercise, This approach has some
real difficulties when analysed from a tbeoretical standpoint, as has been suggested
by Florian (12),

Wigan's perturbation technique (10) has proved to be most effective on small

and medium sized networks of up to several thousand links, The basic innovation is
that a trip end model (which may have either or both of the characters of distribu-
tion or generation models) is used as a reference point when computing the trip
matrix to be used for the next iteration at any stage, It is used only after a full
assignment loading, so as to avoid the cyclic loading characteristics of IMPACT,
which may appear to give good convergence but cannot readily be demonstrated to do
so, By comparing the current trip costs with the level of travel demand implied by
 27

the trip end model for these costs, a matrix of both positive and negative elements
can be constructed to correct the flow and demand patterns represented by the last
iteration of the assignment, As elements may - and often are - negative, the un-
loading feature is automatically included by this device. Although this description
correctly represents the elements of the methods, the other half of the technique
lies in the way in which it is used.

An initial all or nothing assignment is done, and the chosen trip end model
calibrated to this level of cost and demand. An arbitrary disturbance is now made
to the network to start the equilibrium seeking process on its oscillatory path:
this disturbance may be a set of charges on roads, a variation in the speed flow
functions used at some point in the network, or any other similar measure. Use of
the automatic scaling method will immediately produce a change in level of demand
and of flow pattern, and once a couple of iterations have been carried out THE ARBI-
TRARY DISTURBANCE IS REMOVED. The system then continues to operate, but is not seek-
ing to reproduce the original situation. The precision with which it does so provi-
des the null test required, and the mechanism used to return to the original state
has been improving the congestion equilibrium of flow and demand throughout the pro-
cess. Once an adequate equilibrium has been found, the same automatic scaling
device is used to simulate the forecast state, once the trip end model actually
needed for forecasting has been brought into play. This model need bear no relation
to the one used for perturbations, although there is a clear practical advantage in
using the same model for both purposes. Figure 1 shows how consumers surplus
(defined by reference to the arbitrary start state for the perturbation) varies.
The agreement between the last two versions of the initial equilibrium is clearly
adequate, as the uncertainty in the forecast benefits cannot be greater than £30 in
any future situation as a result of systematic errors introduced by an inadequate
startpoint equilibrium for policy analyses.

Figure 2 shows how important such a refined base position can be in practice.
The policy under study is that of road pricing on a 400 link network. The net bene-
fits to the community are given by the difference between the toll revenue curve (+)
 28

13001)

::;--_.-

Consumer loss (,ft,r for,,]' ,dj"trn"t)

12r)OO.

Consumer loss (before forced adjus{ment)

III III
+'>Q)
U)0_ ~
Q)~ s::
IIlQ)Q)
illS::>

-
OQ)Q)
--' a:> 0::

10000

I I I
o 2 3 4 :;
Index number of perturbation solution used

Figure 2. THE ELIMINATION OF SYSTEMATIC ERRORS DUE TO POOR BASE

ASSIGNMENTS ON AN EXTREMELY HEAVY CONGESTED NETWORK
 29

and the consumers loss curve (~). The forecast of road pricing effects has been
carried out using a trip end model with elastic travel demand. The results are given
after one, two, and three perturbation stages, and it is very clear that the final
net benefit of over £1,000 had been lost in the systematic errors inherent in the
base equilibrium even after the first application of the perturbation method. This
particular example is of a network with extreme congestion, and a number of other
problems: the level of demand assumed was close to the absolute capacity of the net~

work, thereby showing how this method can achieve practically useful results under
pathological conditions.

When attention was shifted to a 4000 link model of London, the 'noise' referred
to previously did not allow the perturbation methods to show clearly the quality of
the base equilibrium. The quality was actually very good, but it was impossible to
establish the fact without repeating a set of simulated 'policy' runs on the lines
discussed for the 400 link network. It is at this point that the cruder and faster
methods of Van Vliet become the most attractive, and certainly the cheapest.

The uniqueness of the solution obtained is a point of some concern in all three
of these methods: the perturbation method has been used to examine the difference
between the two approaches to joint trip end and flow equilibria. The open circles
of Figure 3 represent recalculations of the tripend matrix after a series of seven
iterative assignments. The whole cycle is then expensively repeated to reach the
next open circle. This is the strategy of obtaining a good congested assignment, and
then recalculating the tripend demands and patterns, and repeating this sequence
until the consumer surplus (net) figures stabilise. Once again road pricing has been
used as a self~generating policy which amplifies any congestion variations further.
At stage 5 what appears to be an excellent equilibrium has been reached, one that is
precisely comparable to the results of IMPACT. The costs on each link are then for~

ced into precise agreement with the flows, and the road pricing charg~s either left
alone, or recalculated to amplify any small congestion variations that might appear.
The difference between these two shifted open circles is the significant measure:
the gap is far bigger than the apparent shift between successive tripend reassessments
 30

2000:-'
(. ...

Forced adjustment
(Tolls changed)

~~~cremental loading ?~
/ '-, ...,.~., I

•I 'e-' """"~r --_~

1500 -
.I
.I
II
Forced adjustment
(Incremental loading)

Forced/1'""-6~""-0
adjustment I
Successive loadln!J ('i' I): I s left)!

.s::;
'-
~ 1000
II)
II)
0

II)
~
OJ
E
::I
II)
s::
0
u

500

100

o 2 3 4 5 6 7

Successive loading stage

Figure 3. CONSUMERS LOSS FOR SUCCESSIVE LOADING ~10DEL FOR OPTIMUM

TOLL SOLUTION VERSUS INCREMENTAL LOADING
 31

~t stages 3, 4, 5. Thi-s process has then been pushed on two further stages to show
that the internal precision leaves much still to be desired, although the detection
of the imprecision is clearly not easy, The filled circles form a single dotted
line, and the whole line is equivalent to less than one step between open circles
in terms of cost. The dotted line was obtained using the automatic scaling device
alone, and the final convergence check shows that greater reliance can be placed on
this result than the other. Figure 4 shows how three quite different ways of using
automatic scaling still gives the identical result for tripend elasticity.

These are merely two of the many experimental checks carried out to ensure that
the results of these methods are both consistent and unique.

These methods have been applied to study traffic restraint policies on several
networks, a line of work which requires the model to consistently forecast and
delimit the effects of a 1p (3 cents) change on a single link in a 2000 link network.
Results of such studies have been published elsewhere (13).

The importance of systematic error reduction in short run analyses is illustra-

ted by figure 5, where two lines show the results of an analysis of variation of
perceived cost on the benefits from road pricing. The dotted line links sets of
results all based on the same base equilibrium and the smooth line the results fore-
cast from a different base equilibrium for each point on the line. As the level of
perceived cost varies, the equilibrium balance of flow in the initial unaffected
situation is altered and if this is not carefully corrected for by means of a
further set of perturbation corrections, although the size of the effect is small
the internal consistency of the differenced results shown is sufficient to demons-
trate the systematic distortion of the forecasts. The size of the distortion intro-
duced is not large, and unless small scale policies such as those of traffic res-
traint are under study, it is evident that for all practical purposes this source of
systematic error is insignificant: however, as an illustration of the care with
which the meaning of simulations must be examined in order to avoid other, poten-
tially more serious, sources of systematic error it is graphic. An example of such
 32

500,000

400,000

300,000

.....
.s:::
.......

~
OJ
E
:::l
III
s::
0
u
200,000 Undamped
....,
0
o
III
III
- ---0--- Damping policy 1
0
...J - - -0- - - - Damping policy 2

100,000

o 2 3 4 5 6
Adjustment stage number

Figure 4. LOSS TO CONSUMER AT EACH ADJUSTMENT STAGE. SA~IE SOLUTION

OBTAINED FROM DIFFERENT ITERATIVE STRATEGIES
 33

"Unrefined solutions" refer to

2000 benefit predictions based on
\
a pre-restraint assignment in
which the level of cost perception
is 0.865 (point %)

:t
\
\

1500
\
\ ,
\
\ Refined solutions
\
\
\
\ 1:
\
\
\
..c Unrefined solutions \

I,,
'-
.::!!.
1000
I::
::s
~
,,
o "-
u
o
...., '"
')'
01
...., ~

,
;;: \
QJ \
I::
QJ
.Q \
...., \
~ 500 \
\
\
\
\
\
\
'l\
\
\
\
\
____ I \ !
o
o 0.5 0.6 0.7 0.8 x 0.9 1.0 \1-1\ \'2
User perceived cost as a fraction ,~
of the full average cost at a speed '~I
of 32 km/h ~"
Figure 5. BENEFIT TO COMMUNITY OF A MARGINAL COST PRINCING SCHEME FOR
DIFFERENT LEVELS OF USER PERCEIVED COST: THE EFFECT OF
INCORRECTLY DEFINED BASE EQUILIBRIUM
 Model stage number
4 5 6 7 8
0·0

'" " J)~-~

Backtrack to the
-1000 startpoint

..r:::: -2000
.......
~

~ _~---CJ w
..., I \ """
0
...,
Vl
Vl
0
-3000
s-
a.>
E
Vl
'"
s::
0
u
k \ I > -~
Generation of Optimal Changes
-4000

-5000
Figure 6. EQUILIBRIA FOR TWO ELASTIC USER CLASSES IN COMPETITION UNDER ROAD PRICE
 35

a source of systematic error is the rounding effect introduced by having a large

range of sizes of link length and arc cost; this inevitably causes a degree of
rounding, and the meaning of this rounding is that the simulation is solving the new
problem posed to it; namely, what is the equilibrium, given that certain classes of
variation are insignificant? This is some way from the problem that is actually
being attacked, and like the uncritical employment of multiroute assignment, can
lead to apparently different results than an 'ideal I method would provide. The
reason of course being that they are not solving the same problem statements.

The null seeking methodology discussed here as a basis for modelling can also
be extended to multiple user classes. This is an area fraught with both theoretical
and practical difficulties. The theoretical issues have been discussed by Netter (14)
and Dafermos (15) amongst others, and the details of a practical examination of a
multiple user class formulation of traffic restraint is covered by Wigan (16). In
the current context the interest is in the use of a null seeking technique to check
on the internal consistency of the model used to represent two different groups of
people responding eleastically in both routing and trip demand to charges for using
different links of the network. This is a multiple equilibrium problem, as each
class sees distance, time and money costs with a different relative and absolute
weighting. We therefore have two complete elastic travel demand systems in competi-
tion for the same road system, in addition to these two equilibrium searches we are
also seeking the optimum price structure to cover all links of the network to optimise
the community surplus. The model is run first to reach this equilibrium,and then in
reverse (by removing the road pricing subsystem) to re-establish the startpoint
equilibrium. The precision with which this is reached indicates that there are few,
if any, sources of systematic error within the representation of travel demand and
traffic flow that has been used, and also that the models themselves have not intro-
duced any fresh sources of variation.

WHAT HAPPENS WHEN THESE METHODS ARE APPLIED?

This range of techniques and tools for equilibrium analysis has been applied to
 36

a large town to study the feasibility and possible returns from traffic restraint
of different kinds. The results of these studies are not likely to be published for
some time, but there are a number of items of general interest which can be
abstracted for comment. These fall into three categories: ~

Comparisons between previous, equilibrium, and observed assigned flows.

2 Comparisons between the behaviour of this network and the idealised test
networks used for the diagrams of this paper.
3 Robustness of the results to matrix and network variations.

The consultant assignment was carried out by an iterative methods, where the
entire matrix was assigned each time and the speeds of each link adjusted after each
assignment until fair agreement was reached. On a subjective basis, the consultants
assignment was stable and adequate for strategy forecasting: however when the speed
flow relationships used for the assignment were used to calculate the speeds on the
links at the end of the model, there were substantial differences. These are shown
in figures 7, 8. Figure 7 shows the detailed results for each type of road for both
consultants' model and equilibrium assignment. The mismatch between printout average
speed and average speed including overloaded links is the item referred to here. A
sample of 10% of all links was used to compare in detail the two model results. It
was realised almost immediately that in both assignments the 'minimum allowed flow'
had led to many overloaded links, thereby producing a distortion in the computation
of the average speed. As the consultants had not reduced their final speeds to
match their final flows, there was a marked drop in average speed when the overloaded
links were excluded from the comparison. The clustering in both vertical and hori-
zontal planes of the equilibrium model results shows how the congestion has been
efficiently spread, and all available capacity called into effective use. The con-
sultants model had far fewer overloaded links in this far~future situation (this was
a model of unrestrained demand late in the century), and consequently had vastly
overestimated the average speeds available. The equilibrium model has brought a
total of 27% more vehicle miles onto the network, and gives a very good agreement
between final assigned speeds and the speeds implied by speed flow relationships and
 37

() Overloaded links included

.. Overloaded links excluded
18 ,-----------------------------------------------------

Cap I 10
16 ~--------------.

CI 9
14 0----------- ------.

Consultants'
mode 1
cr- -- - - - - - - - Cap
- - I-11- - - - - -- ....
Vl

"
Q)
Q)
CI 8
0--- - - - - TI 1"2- - - - -
~ 10 0---------- --.
Q)
0-- _ _ _ _ _ _ C.! 7_ _ _ _ ______ ...
0>
ro
s-
Q)
CI 13
>
ro 8
0------------ -----~

ro
s-
cv
>
a
6

rL. _ _
CI 9
____ -/t - .
.
Cl 11

CI 8

2
\..F C17
(J>:; - ----- :=-- - ---e
o----1..I-.--.::-_--==_~
CI 12
Equilibrium
model
--cTTY --. CI 10

Average speeds (mile/h) for 10% sample

Figure 7. BASE ASSIGNMENTS: EQUILIBRIUM & ITERATIVE t1ETHODS

 38

the final assigned flows. Not unexpectedly the average vehicle flows in the equili-
brium assgnment were up by 22% over the consultants figures, in close correspondence
with the increased mileage run on the network.

In a similar comparison on the uncongested base year there was little to

choose between the observed flows, the consultants adjusted assignments, and the
equilibrium model results. This is remarkable only in that the ability to adjust an
equilibrium assignment does not exist, and thus suggests that in un congested condi-
tions there is (a) little reason to use either method (b) that the speed flow rela-
tionships were more consistent than one would ever have expected, and (c) that the
comparison of either assignment with observed flows raised all the difficulties
listed earlier in this paper. One useful point was that the average speed flow along
a complete major route was the best and most consistent index to use for this purpose,
partly because the variations on individual roads were averaged out, and partly
because the minor switches between reasonable routes that commonly differ between
observations and assignment were also averaged out.

These two diagrams have been included to show how the pedantic approach to esta-
blishing the underlying statement of the problem as seen by the model in use, is both
necessary and illuminating when equilibrium models are employed. The conceptual luci-
dity of equilibrium models has been found to be invaluable in practice, as every
anomaly or surprising results that appears can be speedily traced either to the data
or to a specific element of the equilibrium postulates. It must be added that this
only applies if a great deal of attention is paid to obtaining good base equilibria,
and tracking down possible sources of systematic errors of comparison within the sys-
tem in use.

The details of the results of simulating different policies on both base year
and future year networks are not available for publication as yet, but the general
behaviour shows a clear relationship to the results obtained on the test networks, and
while the policies vary in their relative success in different conditions, and the
network effects are substantial, it will be possible to put forward conclusions of a
 39

general nature in the future. The entire equilibrium analysis, modelling of policies,
and comparative analysis of results have been completed for several variations on
network capacity (i.e., details of network) and overall levels of demand: the results
for all these sets of analyses are consistent, which indicates that the general over-
all pattern of travel demand and the broad characteristics of the network of roads
are sufficient to define short run option rankings, without great precision being
necessary for the network and demand variations that might prove necessary. This is
a most promising outcome, and suggests that equilibrium analysis of short run options
is likely to become much more widely used to support the short run options that have
now become so important on the ground.

ANALYTICAL STRUCTURES FOR FUTURE RESEARCH AND APPLICATIONS

The basic emerging conflict in transportation planning models is the need to

handle large scale problems extending over large areas, but which have a fairly minor
impact at any given point, and require extreme detail in certain areas in order to
pick out the environmental effects, in addition to synthesising the broader influence
of local traffic management measures being used as part of a co-ordinated policy over
a ~Ii de area.

A two stage approach to modelling analysis might well serve these needs: the
first stage being a selection from a data base, and the second, the modelling process
proper. Our attention should then be directed towards the key problem of generating
a set of decision rules to allow us to cut a general problem down to size at critical
points in our analysis. The opportunity to return to the data base and revise the
selection process is likely to be a basic requirement for such a metamodel framework,
but we have already defined that it ~ be accessible.

The aggregation of different elements of the model at different stages should be

the initial objective. If we can determine the minimum levels of dis-aggregation suf-
ficient to allow us to input local measures and gain the necessary behavioural and
system responses, then we can do much of our task by means of a set of variable aggre-
 40

17

16

15

14 I
~- .. :
.'. .. ... -;..
.~." "

13
r 'I
12
r . ,[
l
11 ~:..--l
r r

....... 10
..c::
OJ
r--'
.
.~

E 9 I I
r--~
r
"0
OJ 8
,I,
OJ
c..
(/) I
~
7

n
7 8 9 10 11 12 13 7 8 9 10 11 12 13
Capacity indices
EQUILIBRIUM assignment speeds ITERATIVE assignment speeds
(before and after removal of (before and after removal of
overloaded links) overloaded links)

~ Including overloaded links

Excluding overloaded links
Printout average speed

Figure 8. THE EFFECT OF OVERLOADED LINKS

 41

gation and disaggregation algorithms. The interfaces between each such process being
the subject for initial decisions as the design stage.

The presupposition of this line of argument is that the huge amount of data
needed to back it up can be set up readily. This is a reasonable assumption, as
much of the data will comprise the fixed physical specification of the system as it
now stands.

If we set this up as a distinct entity, then we may use synthesis methods for
filling in blank entries, or modifying the data structures for a specified problem.
The analogy most appropriate is that of a report generator in an MP * package, where
this class of task is handled, as a ~ requirement.

Thus far the only dimension of generalisation considered is that of the level
of detailed geographical representation, with some asides to hint at the associated
data structures. We must also consider the disaggregation of users of the sytem.
Lorries, cars, buses all have their own characteristics, and so too do the travellers
who commonly use these modes of transport. Each of these classes of vehicles, to-
gether with others such as motorcycles and bicycles comprise a different and policy
sensitive element of the traffic mix which makes a different demand on the services
available and which will respond differently to each of the mechanisms that we might
use to modify the structure of either the system or the manner in which it is used.

The most intractable and awkward element that we have yet to introduce is the
time factor. The representation of an equilibrium state is far easier than the
accurate representation of a system within which transient or time structured events
occur and which interact. This is unfortunately exactly what happens within a trans-
port network: journeys that start a long way away from a given point interact with
shorter and later starting trips at points all along the progress of the longer jour-
ney through the system. As the operating body we frequently assume that we may valid-
ly take the standpoint of a person with full information: i.e., we know where everyone

* Mathematical Programming
 42

is gOing so that we can then optimise the system so that they service their needs
(as we see them) at some minimum cost to society as a whole. It we do take this
standpoint then we find that there is a difference between defining a scheme which
will fulfill a given policy and a mechanism which will act through people's behaviour
to achieve these policy objectives. This is a common distinction to have to draw in
any system where physical compulsion, regulation, and fiscal persuasion are the only
tools that can be used to influence travellers whose actions are more readily con-
trolled by self interest, information, and a healthy appreciation of the likelihood
of being caught when violating a regulation in a particular location, manner, or
time.

It would therefore be desirable to add an explicit time structure to the frame-

work. This is not strictly necessary, and for many purposes can easily be avoided by
drawing on our knowledge of the actual behaviour of people when they are affected by
fairly conventional policies. If however one of the possible methods of application
is a real time updating of the control instruments behaviour in response to a changing
situation then time must be an explicit dimension. Area traffic control is one such
policy, if an adaptive control scheme is required.

When the size of the problem area can be restricted to one part of the physical
network, then the addition of a time structure is fairly straightforward. A number
of effective approaches to this restricted problem can be found in the literature,
and essentially the addition of several replicates of the physical system - each one
covering a different time slice - is all that is required. Even discrete silumation
methods fall into this category, as the use of event- or time-stages control routines
depend on either a substantial set of time history files associated with each entity
or activity represented within the model, or alternatively a multicopy network is
used to take a snapshot of each overlaping time period for the network as a whole.
Both are simple variants of the basic idea of extending the physical representation
into another dimension associated with each action or entity.
 43

If detailed knowledge of the behaviour of the vehicles or travellers over the

system is not required for every unit of time, then we can use multicopy representa-
tions of the system to keep track of the effects as they change time. Continuous
models of processes can then be used to move between successive snapshots of the sys-
tem. An example of this approach is embedded in TRANSYT (4) where the platoon struc-
ture of the vehicles let loose by a traffic light green phase are of no further
interest to the system until they reappear as a smeared~out platoon of arrivals at
the next junction. This is a particular application of this principle, and by care-
ful selection of the levels of representation that we might use for our basic unitary
elements, a considerable amount of simplification might be achieved.

One critical decision is whether or not it is necessary to retain many different

snapshots of the system at several successive points in time: this would be needed
for many of the dynamic programming systems of optimising the operation of a control
system, and unless a data gather mechanism was available within the model to generate
such snapshots at any given point in time from the normal operation of the system,
then the control aspects of an investigation would be severely hampered.

The precedj,ng discussion was based on the restricted problem, where the extent
of the physical system surrounding the problem locality was comparatively small. This
is certainly not true of the policy actions which might be implemented by a set of
small but co-ordinated changes to each of many different localities, with the overall
objective of altering the overall use of an extended network of roads and services.

Giannopoulos (6) attempted this problem with a segmented model of Oxford Street.
The area of interest was fairly small, about half a mile square, and by representing
the whole of the street system within about 3/4 mile of Oxford St. as a central core
in a full strategic representation of the road network of London, he was able to break
down the successive arrival patterns of journeys from a multitude of d1fferent origins
that pass through the Oxford Street area at staggered intervals depending on the length
of time that journey had been on the road. This produced an effective and realistic
time dependent profile of arriving journeys, and was the first model to successfully
 44

mix the strategic and micro levels of representation in a single framework, and
although the model itself was full of detailed assumptions of dubious verity, the
overall result was satisfactory in that it demonstrated the practicality of the
approval.

A better designed and more practically oriented verslon of this model could well
serve many purposes. It would certainly be effective for the modelling and analysis
of local variations of traffic control policies over a small area and the consequen-
tial environmental effects would be describable in terms both of trips, routes and
flows and queues, rush hour peaking effects, and noise peaks due to traffic light
changes.

Unless a more flexible framework can be constructed round the Giannopoulos idea
as it now stands, then its application is severely limited. The proposition put for-
ward in the earlier part of this note is of this kind. The necessary step would be
to have a problem generator driving the Giannopoulos model framework. The whole sys-
tem of roads, services, trips etc. would be held in the form that it is usually store
for various engineering purposes, and the essential "new" process model would be the
automatic process of constructing a Giannopoulos type of nested model from the data
base to apply to the specific locality in question. The 'problem generator' facility
would then enable us to analyse a number of different localities in groups, and then
revise the external conditions if deemed worthwhile and re-run to resolve the inter-
actions.

Obviously this general framework would require a substantial effort to create,

and indeed a number of subsidiary research projects would be needed before the design
could be finalised. There are two special techniques from transport network analysis
which could be examined to help with the disaggregation process. The direct estima-
tion process of Schneider, and the cordon models of the GLC and others. Some work on
differential levels of aggregation is being done by MIT and the GLC.
 45

The general features of an analysis framework suitable for linking traffic

management transport poltcy emerge from this discussion as being:

(i) dependent on a data base accessible to a model generator

(ii) requiring specialised models or other estimation processes to fill in
the gaps tn the data base (1) that would be left in any collection of
real data.
(iii) requiring an aggregation and disaggregation process of some sophistica-
tion
(iv) requiring a model generator macromodel to carve out appropriate micro/
macro boundaries and to access the appropriate data sets
(v) the actual models for local and macro analysis need not be linked to-
gether directly, but can communicate via the data base at level (1) if
need be.

The key elements in a startegic system design of this type are:

(a) data base-real data

(b) data base-for model-generator access
(c) data generator-link between base a, base b
(d) data selection processor to generator micro/macro data sets for models
(e) micro models
(f) macro models
(g) ability for elf to revise data sets a.b.
(h) differential aggregation algorithms for elf.

The general system format is now laid out: a data base of measurable physical
information is the essential starting point.

The next level is a transformed data set that is based on the root set (a) but
has many estimation processes applied to fill it out in appropriate directions and
remove any gaps, perhaps only by interpolation, to produce data set (b).
 46

The novel element in the design is a problem oriented model generator. This
process is the only one that we shall have to design from scratch. The research
required is to establish appropriate breakpoints between macro and micro. It can
reasonably be expected that with proper design the system set up to carry out the
disaggregation and selection process could be at a sufficiently generalised level to
allow the research to be done using the system itself. This strategy has proved
successful for several projects in the transport modelling and evaluation field
carried out by the author.

The design of the intermediate (or smoothed) data bank will naturally be carried
out in conjunction with the selective problem builder for obvious reasons.

The individual models at macro and micro stages will necessarily have different
data requirements, and as these are to be the output of the selective problem builder
there is no reason to take special steps to interface macro and micro models unless
experience shows that the price of such extra complexity is worth the effort.

If an approach of this type is adopted then several of the available models

already in use may serve as the initial macro or micro procedures; and thereby
reduce the considerable effort otherwise likely to be needed to interface micro and
macro models containing very different view the situation under study. The commu-
nication through the data framework will adequately buffer both the differences in
concept and the differences in application.

The most difficult design decision is the choice of a recording mechanism for
the time stages results. The storing of replicates is not only data and space inten-
sive: we must reconcile the requirement that the form of the representation of the
snapshots that we shall need for practical use of the transforms of the data base
that comprise the models prime output must the need to represent the behaviour of the
individual models in a manner appropriate to their internal working. This is likely
to demand a lot of storage space, and consequently substantial computer resources.
 47

One way of avoiding these problems would be to build a specialised data acqui-
sition system which would allow generalised file access to the structured data
defined in data bases 1 and 2, and display on a Tectronix screen for live interactive
editing, command menu processing, or simply as a photographic record of the progress
of the problem analysis process as it proceeds, in place of weighty and intractable
printouts. This would allow the construction of a snapshot from disparate sets of
data without the need to actually compute and store it unless it was specifically
required.

The diversity of models and analysis procedures that are undoubtedly needed to
bridge the micro-macro gap is such that much time and effort could be wasted in
trying to build a tool for every job at the expense of its utility for any particu-
lar case. The database approach put forward allows us to construct only those parts
that we need at any given time, and would give us a specific interface requirement
with which to co-ordinate the different types of data and process.

A useful analogue of this system design is that of a computer CPU operated under
a hardware stack with the devices and memory banks acting as autonomous devices
requiring servicing. Many fire control systems are of this kind. The CPU is able to
execute specific jobs but as each data device interrupts it through the hardware stack,
the action is stored away with the current states of the system, and the interrupt
process handled before returning to the prime computation, which may now be operating
on modified data. Access between models is through the analogue to ported store,
namely the level 2 database, and the basic data bank remains untouched throughout.
In principle the level 2 data set would also be protected and a level 3 copy presented
to the model in use, the level 3 data set being generated by the selective problem
generator software.

To marshall the considerable amounts of information needed to analyse situations

at both macro and micro levels simultaneously we should concentrate on defining a data
set that will service the needs of already existing models at micro and macro level,
and spend our initial efforts on designing a data structure that will allow us to
 48

store and update the types of information that we seek at both levels and the kinds
of data that we might actually be able to get primary access to. If it proves worth-
while, we could distinguish between the primary data sets and the secondary data sets
produced by interpolation and deduction from the primary data and ad hoc estimation
processes that we might define for the purpose,

The prime requirement for bridging the micro~acro gap is a problem synthesis
generator which will select the relevant subset of detailed and less precise data
from the data base. If this can be achieved, then the actual modelling process will
be greatly simplified. The analysis of peaking effects, transients, and time varying
processes will force special attention to be paid to the form in which the data pro-
duced by such models will be stored and accessed.

The objectives of each subsystem or' model should be defined solely in terms of
the target level of detail and the specific policy issues to which it is to be
addressed: the link between each subsystem should not be direct, but only through a
levell, 2 or 3 dataset.

The general structure is laid out in Figure 9b, and an illustration of how this
approach is already in use in an elementary form is given in Figure 9a. The basic
data base for RRLTAP and the PANIC environmental system is held offline on tape.
The transformation routines in RRLTAP select and manipulate this information in the
form defined by the interface conventions. The results of each individual transfor-
mation are held on line within the working database, and as better or fuller infor-
mation is synthesised or forecast it is used to overwrite the relevant parts of the
working database. It has been necessary to push the simple modular operational
structure of RRTAP/PANIC into this more pretentious form for this illustration, but
the essential design elements of specified interfaces, working data structures of
the same structural form as the basic data, and the modular 'beads' for operations
on both data bases have required no stretching.

The general approach of the UTPS sketch planning system and its relationship to
the rest of UTPS shows the influence of these general perspectives, but there is a
 49

Figure 9 (a) A SPECIFIC DATA HANDLING SYSTEM

Overall networks RRLTAP Specific regional networks

Patterns of Movement Models flows, networks, etc.
Geographical
characteristics I
I

I
Household and zonal INTERFACE I
characteristics
RRLTAP conventions for I
networks, flows, matrices~
data files co-ordinates,'·~
]. Pollution data
etc. -----, (Forecast)
~------------------~~
,,
2. Impact results
PANIC
I Data Estimation ~--------------------.-------
Street by street detail Conversion of detailed : local details in a usable
on certain areas I Routines impacts of 1 for!)1at
e.g., population-·-·- ~ po 11 uti on 1/4
\- - - - - - -
BASIC DATA SET TRANSFORMS WORKING DATA SET
(Normally Offline) CONLINE, I NCLUDES REPLICATES
AND UPDATES OF PARTS OF
BASIC DATA

Figure 9 (b) GENERALISED DATA HANDLING SYSTEM

Processing
Beads
BASIC data set (which DATA DATA WORKING data set a
may be pa rt i a1) INTER INTER replica of BASIC data set,
FACE FACE with missing dates filled
, in, and new forecasts are
written, by processing
,, beads.
,
___ . __ .x. >( XX X
 50

great deal of careful work to be done if research and application systems outside
the UTPS framework are to be set up efficiently. Were the source code for UTPS more
widely available it might be possible to experiment on these lines. The investment
of resources in inflexible systems is inhibiting much needed new work on the grey
area between data base access and systems models for transport and land use planning.

THEORETICAL STRUCTURES NEEDED FOR NEH PROBLEMS

Recent work on traffic restraint, multiple user classes, and environmental pro-
blems have shown that further theoretical support is now needed for the next round
of problems to be faced with equilibrium analyses.

Traffic restraint has been included in a theoretical analysis by Murchland (17),

but only under the condition that a price may be set on any road if need to be opti-
mise the community surplus function. Practical simulation studies have already been
carried out (13), on the more usual case where only certain types of charging struc-
ture are admissable, and only over restricted sections of road or types of road.
This form of constraint should be brought within the general framework, and prefer-
ably in a form which will allow an estimate of the benefits lost by these restrictions
(compared to the unrestricted case dealt with by Murchland). This has been posed as
a singlz user class problem, as this has been the main requirement of the transport
analyst: it is now becoming important to be able to differentiate between the
impacts and benefits borne by different groups of people in planning studies. An
example of the simulation of a two-user class situation is given in figure 7, but
there is no theoretical framework for computable algorithms in the case of many user
classes, or indeed of a distribution of user classes. The introduction of environ-
mental constraints also requires mathematical assistance: environmental constraints
may appear in many forms, as levels of pollutants that must not be exceeded in a
given place, as the requirement to close a street or streets to certain classes for
vehicle, or indeed to all vehicles. With the aid of emmission concentration fore-
casting equations (some of which are now available in a form appropriate to optimi-
sing models - see (18)), it should be possible to adduce shadow costs for environ-
 51

mental standards, and resource costs for road closures, The addition of competing
objectives in the optimising function for problems of this kind will allow questions
such as the effect of influencing the degree of perception of costs or controls on
the system to be pursued with a better understanding than that provided by simple
simulation.

An extension of the tools for optimising traffic light systems is also required,
so that the optimisation of overall transport system use can properly reflect the
fact that a great many more devices now exist in a form that can influence traffic
flow and travel demand over the system as a whole.

Few of these requirements are new, but the first few applications of equilibrium
analysis to practical problems have underlined the fact that they are real - and the
tools are now to hand.

This paper is presented with the permission of the Greater London Council, but
the views expressed are solely those of the author, and do not necessarily reflect
those of the Council. Much of the experience on which this paper draws was gained
while with the Transport and Road Research Laboratory, but are the authors own views,
and not necessarily those of the TRRL.

REFERENCES

(1) us Department of Transportation, Special Area Analysis: Final Manual

Federal Highway Administration, Washington, D.C. (1973).

(2) Wigan, M.R., "Pollution Pedestrian and Noise Impacts as Part of the
Transportation Planning Process" (In course of publication).

(3) Sandys, R.C, "The Dynamic Highway Transportation Model" (DHIM) Users Manual.
Stanford Research Institute, Menlo Park, California (1971).

(4) Robertson, D.l. "TRANSYT: A Network Study Tool" LR253. Transport and Road
Research Laboratory. Crowthorne. Berks, (1969).

(5) Manheim, M. and Ruiter E., "DODOTRANS I: A Decision Oriented Computer

Language for the Analysis of Multimode Systems". Highway Research Record
314 (1970).
 52

(6) Giannopo10s, G., "Pedestrianisation: London's Oxford Street Experiment.

Transportation 3, 97716 (1974).

(7) Wigan, M., "Benefit Assessment for Network Traffic Models and Application to
Road Pricing" LR417, Transport and Road Research Laboratory, Crowthrone, Berks,
(1971) •

(8) Burrell, J., "Multiple Route Assignment and its Application to Capacity
Restraint", Strassenbau and Strassenverkersteknik Heft, 86 pp.21 O~2l9,
(1969).

(9) Hawkins, A., "IMPACT" Mathematical Advisory Unit, Ministry of Transport,

London (1972).

(lO) Wigan, M. and Banford, T., "A Perturbation Model for Congested and Overloaded
Transportation Networks", LR411, Transport and Road Research Laboratory, (197l).

(l1) Van Vliet, D., "The Choice of Assignment Techniques for Large Networks" Inter-
national Symposium on Traffic Equilibrium Methods, Montreal, (1974).

(12) Florian, M. and Nguyen, S., "A New Look at Some Old Problems in Transportation
Planning", Centre de recherche sur 1es transports, Universite de Montreal,
(1974).

(13) Wigan, M. and Banford, T., "A Comparative Network Simulation of Different
Methods of Traffic Restraint", LR566, Transport and Road Research Laboratory,
(l973).

(14) Netter, M., "Affectation de trafic et tarification au coOt marginal social

cri ti que de quelques idees admi ses" Trans. Res. (6), 411-429, (1972).

(15) Dafermos, S., "Traffic Assignment Problem for Multic1ass User Transportation
Networks", Trans. Sci. 6 (I), 73-88, (1972).

(16) DoE Transport and Road Research 1972 pp33. HMSO, London (1973).
(also Wigan, in press).

(17) Murch1and, J., "Road Network Traffic in Equilibrium", proceedings, Mathematical

models in the social sciences. Oberwolfach 1969. Pub. Anton Verlag Vau.

(l8) Wigan, M., Ing1eton, 1. and Souter, H., "Noise and Airbourne Pollution" (to be
pub1 i shed).
 ' Equ~'1'~br~a
Tra ff ~c ' Ana I d
yse .
v~a . ProgrammingI , 2, 3
Geometr~c

by

Michael A. Hal1 4 and Elmor L. Peterson s

Abstract. The "traffic-assignment problem" consists of predicting "Wardrop-

equilibrium flows" on a roadway network when origin-to-destination "input flows"

are specified. The "demand-equilibrium problem" consists of predicting those

input flows that place the network in a state of "economic equilibrium" when

the input flows are related via given travel-demand (feedback) curves to the

resulting Wardrop-equilibrium origin-to-destination "travel costs".

The traffic-assignment problem is treated as a special case of the demand-

equilibrium problem (the case in which the travel-demand curves are graphs of

constant functions); and the demand-equilibrium problem is formulated and

studied in the context of (generalized) "geometric programming". In particular,

existence, uniqueness and characterization theorems are obtained via the "duality

theory" of geometric programming by introducing appropriate "extremality con-

1. Prepared for the "International Symposium on Traffic Equilibrium Methods",

held at the Centre de Recherche sur les Transports of the Universite'de
Montreal, November 21 through November 23, 1974.

2. The authors are indebted to Professor E.K. Morlok of the University of

Pennsylvania for directing their attention to the traffic equilibrium
problems analysed herein.

3. Research partially sponsored by the National Science Foundation under

Grant ifFGU 3851 to the Urban Systems Engineering Center at Northwestern Uni-
versity, and also by the Air Force Office of Scientific Research, Air
Force Systems Command, USAF, under Grant #AFOSR-73-2sl6. The United States
Government is authorized to reproduce and distribute reprints of this paper
for Goverruaental purposes notwithstanding any copyright notation hereon.

4. Bell Laboratories, P.O. Box 2020, New Brunswick, New Jersey 08903.

5. Department of Industrial Engineering/Management Sciences and Department of

Mathematics, Northwestern University, Evanston, Illinois 60201.
 54

ditions" and their corresponding "dual variational principles" (som~times

called "complementary variational principlcs"). These dual variational prin~

ciples and extremality conditions also lead to new computational algorithms that

show promise in the analysis of relatively large networks (such as those in

relatively large urban or metropolitan areas).

All of this is done for a relatively flexible model in which (1) each

"conunodity" is permitted to preselect the only feasible "paths" (from its givcn

origin to its given destination) on which some of its input flow can be

assigned, (2) each link travel cost need only be nondecreasing aud unbounded

from above as a function of the total traffic flow on the link, and (3) each

commodity input flow, as specified by the given con~odityls travel-demand

(feedback) curve, need only be noninoreasing and not approach ~ without its cor-

responding Wardrop-equilibrium origin-to-destination travel cost approaching -=.

Moreover, analogies are drawn with "monotone network problems" (which arise in

various physical contexts, such as the analysis of electric and hyraulic net-

works, but involve only a single commodity).

No prior knowledge of traffic equilibria, geometric programming, or monotone

network theory is required.

 55

1. Introduction. In studying the traffic-assign~ent problem we consider

''multiconunodity networks" (i .e. roadway networks) on which each conunodity

(i.e. each traffic type as distinguished by origin, destination, and per-

haps other criteria, such as vehicle class) distributes its own input flow

(specified by its own travel demand) over certain preselected feasible net-

work paths (subject of course to any capacity constraints) in such a way

that a Wardrop-equilibrium [41] is achieved. A Wardrop-equilibrium occurs

when the total origin-to-destination travel cost per unit of flow (e.g.

the total origin-to-destinatioq travel time [per unit of flow]) is the same

on each path used by a given conunodity and is not greater than that on each

path feasible for but unused by the given conunodity. In particular, we

suppose that each network link contributes to each such cost per unit of

flow a term that (due to traffic congestion) is actually a nondecreasing

function of the total flow on that link and approaches plus infinity at

the link's total capacity flow (which may itself be plus infinity).

This·multiconunodity problem of predicting Wardrop-equilibrium flows

(i.e. the traffic-assignment problem) is somewhat similar to, but more

complicated than, the single-conunodity problem of predicting "equilibrium

flows on monotone networks" (e.g. "Kirchoff-Ohm equilibrium flows" on non-

linear electric or hydraulic networks). In fact, Charnes and Cooper [6]

have attempted to make appropriate extensions of the classical variational prin-

ciples [~t,lO,3,21,2,34] that reduce such single-conunodity equilibrium pro-

blems to equivalent optimization problems. However, their extensions are at

best imprecise (e.g. vague assertions about how Wardrop-equilibria can be

treated as a special case of "Nash-equilibria" are made without a definition

of the "game" under consideration). Moreover, the models treated are extremely

unrealistic (e.g. on€~way streets are nOt included, and traffic on a given

two-way street must always experience the same flow "resistance" in both

directions). Although Dafermos and Sparrow [7,8] have made such extensions
 56

preci.se and more realistic (without relying on game·theoretic concepts), a

more thorough study of the Hardrop·equilibrium problem is possible. However,

to carry out such a study with maximal simplicity and ease, the theory of

monotone netlvorks should be generalized by replacing certain vector spaces

with (more general) polyhedral cones. Actually, the required generalization

is already available in the form of (generalized) geometric programming (25,26,

28,29,32] •

In addition to relaxing some of the assumptions made by previous workers,

our approach to the Hardrop-equilibrium problem provides extra insight and

additional algorithms via the modern duality theory [14,25,26,28,29,32,35,36]

of mathematical programming. In doing so, it also subsumes both the work of

Murchland (22] and the very recent work of Evans [12] by providing the first

reasonably definitive treatment of the most fundamental questions having to do

with the existence, uniqueness and properties of Wardrop-equilibria. Finally,

it should be noted that our approach is fundamentally different frum that

taken by Rosenthal (37], which emphasizes the game·theoretic aspects of discrete

multicommodity networks on which each player controls an indivisible unit of flow.

In discussing the demand-equilibrium problem we assume in the context of

the preceding Wardrop-equilibrium problem that the input flow of a given com-

modity (given by its own travel demand) is actually related via a given nonincreasing

travel-demand (feedback) curve to the resulting origin-to-destination travel

cost per unit of flow experienced by the given commodity when the network is in

a state of Wardrop-equilibrium. In particular then, a demand-equilibrium occurs

when the network is in a state of Wardrop·equilibrium and the resulting origin-

to-destination travel cost per unit of flow experienced by a given commodity

produces via its own travel-demand curve the same input flow already specified

for the given commodity. It is, of course, clear that this demand· equilibrium

problem is far more complicated than the corresponding Wardrop-equilibrium pro-

blem; in fact, it is obvious that this demand-equilibrium problem reduces to

just the corresponding Wardrop-equilibrium problem when the travel-demand curves

 57

are the graphs of constant functions that are identical to the specified input

flows. Nevertheless, the mathematics required to treat the demand-equilibrium

problem is no more sophisticated than that required to treat the Wardrop-

equilibrium problem. Geometric programming does both jobs with equal ease.

It seems that the demand-equilibrium problem was first identified as

essentially a mathematical programming problem by Beckmann [1]. Although

Beckmann's work has spawned numerous papers on the algorithmic treatment of

more realistic models (as reviewed by Nguyen [23]), it seems that there has been

relatively little progress of a more fundamental nature. In fact, it seems

that this paper provides the first reasonably definitive treatment of the more

fundamental questions having to do with the existence, uniqueness and properties

of demand-equilibria -- a task that is carried out for even more realistic and

flexible models. In doing so. this paper also subsumes both the work of

Murchland [22] and the very recent work of Evans (12] by providing a more direct

and simple mechanism through which modern duality theory can be applied to the

study of demand-equilibria. In addition to providing considerably more insight,

this mechanism also serves as the basis for a variety of new algorithms, some

of which may be especially suited to the analysis of relatively large networks

(such as those in relattvely large urban or metropolitan areas) when used in

conjunction with appropriate geometric programming "decomposition principles"

[27,28,31]. Finally, it should be noted that there are alternative (nonprogramming)

approaches to the study of demand-equilibria -- the earliest ones having been

summarized by Potts and Oliver [33], with the more recent ones having been sur-

veyed by Ruiter [38].

2. The Model. A roadway network can be conveniently represented by a "directed

graph" (consisting of "nodes" and "directed links") on which there is multi-

commodity flow.

Each commodity (i.e. each traffic type) is associated with a (fictitious)

"return link" over which only that particular commodity flows. Such a return
 58

flow takes place only in the prescribed return link direction, ~hich is of
course from the commodity·s given destination (node) back to its given origin

(node). For the example network shown in Figure 1, there are a total of three

commodities: a single commodity with origin-to-destination pair (1,5) has

return link 1, and two commodities with the same origin-to-destination pair

(2,5) have return links 2 and 3 •

1
,,- ..- - - 'itf-- "
5 .....
"- ....

, " .I
8 :..&..-"'" ,
--...-3 .......
\ ..... ---:::L~ "
'... - ....

Figure 1.

We show the return links of the network with broken lines and the ureal links"

with unbroken lines. Each real link represents a collection of unidirectional

lanes over which any of the given commodities might flow from a certain inter-

section (node) to an adjacent intersection (node). Such a (multicommodity)

flow can take place only in the prescribed real link direction, which is of

course dictated by local traffic regulations. For the example network shown

in Figure 1, flow can take place on real link 5 only from node 1 to node 3,

etc. Thus, two-way streets are represented by at least two real links.

In the most general model to be studied here, we consider a directed graph

with a total of n links. In particular,

the return links are numbered 1,2, ••• ,r,

and

the real links are numbered r+l, r+2, ... ,n.

Hence, there are a total of r commodities and a total of (n - r) real links

 59

over which the r commodities can flow. For convenience we number the r com-

modities so that

commodity i has return link i for i = 1,2, •.. ,r.

With each commodity i is associated the family of all possible "paths" over

which that particular commodity might conceivably flow (over real links from

its given -origin to its given destination). For the example network shown in

Figure 1, the family of all possible paths over which commodity 2 might con-

ceivably flow consists of: the path over links 4,5, and 10; the path over links 4,

7, and 11; the path over links 4,7,9, and 10; the path over links 8,9, and 10; and

the path over links 8 and 11. Note, however, that links 8,9,6,8, and 11 do not

constitute such a path; the reason is that a path (by definition) can not "eros:!'

itself. This rejection of such possible "routes" is realistic because a commodity

would obviously be foolhardy to follow one of them. Moreover, it is clear that a

given commodity's possible route family is finite if and only if it includes

only paths, in which event it is of course identical to the given commo'dity' s

possible path family. Since a possible path family is obviously always finite,

and since a possible route family is clearly almost always infinite, this

rejection of possible routes that are not paths is also a mathematical nicety.

Needless to say, we assume that the network is sufficiently "connected" to

guarantee that the possible path family for each commodity i is not empty.

It is obvious that two or more different commodities exhibit the same

possible path family if and only if they have the same origin and destination.

For the example network shown in Figure 1, commodities 2 and 3 (and only com-

modities 2 and 3) exhibit the same possible path family (the one enumerated in

the preceding paragraph). To eliminate this alnbiguity, each possible path is

extended into a possible "circuit" by appending to each such path the given

commodity's return link. For the example network shown in Figure 1, the
 60

possible path over links 8 and 11 is extended into the possible circuiJ: around

links 8, 11, and 2 when it is to be associated with commodity 2, but is extended

into the possible circuit around links 8, 11, and 3 when it is to be associated

with commodity 3. Naturally, the family of all circuits obtained by appending

a given commodity's return link to the end of each of its possible paths is

termed the commodity's possible circuit family.

There are numerous reasons why a given commodity usually eliminates from

consideration some of the paths (circuits) in its possible path (circuit)

family. For example, a complete scanning of them all might be unrealistic for

even moderate sized networks (as found in moderate sized urban or metropolitan

areas); in which event only those that seem to show the most promise of being

reasonably "short" can actually be considered. Moreover, local traffic reg-

ulations may not permit a given commodity (i.e. a given traffic type, such as

trucks) on certain real links; in which event only those of its possible paths

that do not include such real links can actually be considered. Finally, other

factors (such as personal safety, driving pleasure, etc.) may cause a given

commodity to eliminate some of its possible paths from further con~ideration.

In any event, we assume that each commodity i selects at least one of its pos-

sible paths for further consideration; and we term the resulting (not necessarily

proper) subfamily of its possible path family its feasible path family. Need-

less to say, we also term the corresponding subfamily of its possible circuit

family its feasible circuit family.

For our purposes, the most convenient way to represent a path (circuit)

is to first associate each network link k with the k'th component of the vectors

in n-dimensional Euclidean space En' Then, a given path (circuit) can be

represented by the vector whose k'th component is either I or 0, depending

respectively on whether link k is or is not part of the given path (circuit).

For the example network shown in Figure 1, the possible path over links 8 and
 61

11 is represented by the vector (0,0,0,0,0,0,0,1,0,0,1) in Ello Moreover,

its extension to the corresponding possible circuit around links 8, 11, and 2

for commodity 2 is represented by the vector (0,1,0,0,0,0,0,1,0,0,1) in Ell'

while its extension to the corresponding possible circuit around links 8, 11,

and 3 for commodity 3 is represented by the vector (0,0,1,0,0,0,0,1,0,0,1) in

In the most general model to be studied here, we suppose that commodity

i has a nonempty feasible circuit (path) family that is enumerated by the integer

index set

[i] ~[m.,m. +l, ... ,n.} for i = 1,2, •• • ,r,

l. l. l.

where

... ,
Thus, there ±s a total of m feasible circuits over which traffic can flow; and

given a feasible circuit j (namely, an integer in the circuit index set

[1,2, ••• ,m}) there is a unique commodity i (in the commodity index set [1,2, ••• ,r})

such that j E [i], which means that commodity i (and only commodity i) flows

over the feasible circuit j. Moreover, the vector oj representing a given

feasible circuit j has components

1 when k = that i for which j E [i]

0 when k F that i for which j E [i] but l:s;k:S;r

oj
k
1 when r+l:;;;k:S;n and real link k is part of circuit

0 when r+l:s;k:;;;n and real link k is not part of circuit j

Now, a possible circuit flow is just a vector

zEE
m

whose j'th component Zj is simply the input flow on circuit j of that commodity

i for which j E [i] 0 Since each commodity can flow only in the given link
 62

directions, it is obvious that each possible circuit flow z must also satisfy

the (vector) inequality

z ~ O. (1)

Of course, each such z generates a possible total flow

(2)

whose k'th component xk is clearly the resulting total flow of all commodities

on link k for k=1,2, ... ,n.

Given that

d i is the total (non-negative) input flow of commodity i

and that

b k is the total capacity of real link k (which may be + co),

a fea'sible circuit flow is just a possible circuit flow z that generates a feasible

total flow x, namely, a possible total flow x such that

for k=1,2, ... ,r (3)

while

for k=r+l,r+2, ••. ,n. (4)

(Even though conditions (1) and (2) clearly imply that the inequality 0 ,;;xk in

condition (4) is redundant, it is explicitly included to indicate the interval

over which certain cost functions c k of xk are to be defined.) Although a

given feasible circuit flow z generates a unique feasible total flow x, it is

obvious that a given feasible total flow x can generally be generated by more

than a unique feasible circuit flow z; in fact, the set of all such feasible

circuit flows z is clearly identical to the set of all solutions z to (the linear)

conditions (1-2).

It is worth noting that the preceding definition of feasible flows (as

 63

well as all definitions and theorems to follow) does not invoke "nodal con-

servation laws" -- though it is easy to see that such laws are in fact im-

plicitly satisfied. Actually, all roadway network models that are explicitly

based on nodal conservation laws (in fact, many previous roadway net-

work models) are erroneous in a very fundamental way. For example, although

a flow (0,0,0,1,1,1,0,0,0,0,0) of commodity 1 on the network shown in Figure 1

obviously satisfies the nodal conservation laws, it is clearly in no real

sense a feasible commodity flow. It is fortunate though that such extraneous

flows are automatically eliminated by certain Wardrop-equilibrium solution

techniques when the travel cost per unit of flow on each real link is strictly

positive -- as demonstrated in [18].

On each real link k we suppose that the travel cost per unit of flow

(e.g. the link's node-to-node travel time [per unit of flow]) perceived by each

commodity i is a positive nondecreasing function c k (of only the total flow ~)

that is continuous at 0 and approaches +00 at bk ; that is,

(HI) c k (xk ) ~ c k (x~) when 0 ~ ~ ~x~ <b k for k = r+l,r+Z, ••• ,n,

(Hz) O<ck(O) = lim+ck(xk ) for k = r+l,r+Z, ••. ,n,

~ -+0

(H3 ) lim c k (xk ) =+00 for k = r+l,r+Z, ••• ,n.

xk -+b~

Needless to say, none of these hypotheses (H) are very restrictive in the con-

text of roadway networks.

Z.l. Wardrop-equilibria. We assume that the (real link) capacities b k are

sufficiently large to handle the commodity input flows dk ; that is, we assume

that there exists at least one feasible circuit flow z. Actually, for non-

trivial roadway networks of practical interest, there are usually infinitely

many feasible circuit flows z.

By definition, the only feasible circuit flows z that place a givel' roadway

network in a state of Wardrop-equilibrium are those z that generate a feasible

 64

total flow x for which there ar~ both (fictitious) return link travel revenueD

per unit of flow

for k=1.2 ..... r (Sa)

and real link travel costs per unit of flow

for k =r+1. r+2, •••• n (5b)

such that the resulting revenue-cost per unit of flow vector

satisfies the following "inner product" conditions

for j =~.2, ••• ,m. (6)

i f z. =0
J

To properly interpret the preceding definition. note from the defining

formula for the geasib1e circuit vectors aj that condition (6) simply asserts

the following traffic situation: for each of the feasible circuits actually

used by a given conunodity i (namely. each circuit j E[i] for which Zj>O) the

total origin-to-destination travel cost per unit of flow is the same, in fact.

just -Yi (the negative of the corresponding "return link travel revenue per

unit of flow" Yi)' which in turn does not exceed the total origin-to-destination

travel cost per unit of flow for each of the feasible circuits not used by the

given conunodity i (namely. each circuit j E [i] for wich Zj = 0).

Each vector Z that satisfies conditions (1-6) is termed a Wardrop-

equilibrium circuit flow, and each such flow Z generates a Hardrop-eQuilihrium

total flow x via equation (2). Although a given Wardrop-equilibrium circuit

flow Z generates a unique llardrop·equilibliium total flow x it is obvious

that'a given Wardrop-equilibrium total flow x can generally be generated

by more than a uniquQ Wardrop-equilibrium circuit flow z; in fact, the

set of all such Wardrop-equilibrium circuit flows Z is clearly identical to

 65

the set of all solutions z to both (the linear) conditions (.1-2) and the

"complementary slackness" conditions

eith er <~ j , y ) = 0 or Zj = 0 for j=1,2, ... ,m. (7)

There are roadway networks that satisfy our basic hypotheses (H) and for

which there are feasible circuit flows z but for which there are no Wardrop-

equilibria. A simple example is described by Figure 2,

t
for 0" x 2 < 1
for 0" x3 < 5
dl =4 c 2 (x 2) for 1" x 2 < 5 c 3 (x 3) = {6.5
1.5 +x3 for 5 ".x3
+ x 2 for 5 :;;x2

Figure 2,

Note that the lack of continuity of c 2 at x 2 = 1 seems to cause the lack

of Wardrop-equilibria for the preceding example. Actually, continuity of all

cost per unit of flow functions c k is, in general, sufficient to guarantee the

existence of Wardrop-equilibria a fact that turns out to be a corollary to

a more general existence theorem.

The more general existence theorem is concerned with (Wardrop) "quasi-

equilibria", which eeem relevant to the prediction of traffic on roadway networks

with discontinuous cost per unit of flow functions c k ' and which coincide with

Wardrop-equilibria in the case of roadway networks with continuous ck ' Although

all models known (by us) to be in u~e by traffic scientists and engineers

involve continuous c k ' we HIlspect that in many cases more accuracy could be

obtained by modeling with discontinuous c k ' For instance, link 2 of the net-

work described by Figure 2 may model a roadway artery with two traffic control
 66

lights that are synchronized and timed relative to the speed limi.t in such a

way that: when x2 < 1 each vehicle moves at the speed limit and is stopped by

only one of the lights; when x 2 >1 each vehicle moves at less than the speed

limit (due to congestion) and is stopped by both lights; when x2 = I some

vehicles are stopped by only one of the lights while other vehicles are stopped

by both lights.

2.2. Quasi-equilibria. To define quasi-equilibria, we first imitate the

theory of monotone networks [21,2,34] by embedding the graph of each cost per

unit of flow function c k in a (continuous) "complete nondecreasing curve"

Each curve 1"k is "nondecreasing" in the sense that, for each (xk'Yk) Efk , and

each (xk'Yk)Efk , either Xk$;xk and Yk$;Y k, or xk$;~andYk$;Yk' Each

nondecreasing curve fk is "complete" in the sense that it cannot be embedded in

a (properly) larger nondecreasing curve. (In general, a complete nondecreasing

curve f can be described geometrically as an infinite continuous curve that crosses

each of the lines with slope -1 exactly once.)

It is natural to treat each fk as a multivalued or point-to-set function

Yk on [O,b k ) with functional values

for k=r+I,r+2, ••• ,n,

each of which is a nonempty closed interval. For a given xk E (0, bk ) the set

Yk(xk ) is a singleton [Y k } if and only if the cost per unit of flow function c k

is continuous at xk ' in which case Yk = c k (x k). Thus, if c k is continuous on [0, b k)

 67

then c k differs from Yk only in that Yk (0) = [Yk ,;; c k (O)}, a mathematical dif-

ference that turns out to be inconsequential in the determination of Wardrop-

equilibria from quasi-equilibria.

By definition, the only feasible circuit flows z that place a given roadway

network in a state of quasi-equilibrium are those z that generate a feasible

total flow x for which there are both (fictitious) return link quasi-revenues

per unit of flow

for k=1,2, ••• ,r (Sa' )

and real link quasi-costs per unit of flow

for k = r+l,r+2, ••. ,n (Sb' )

such that the resulting quasi-revenue-cost per unit of flow vector

satisfies the following inner product conditions

=0 ifz.>O
• J
<o\y) [ for j=1,2, ••• ,m (6' )
:20 i f z.=O
J

To properly interpret the preceding definition, note from the defining for-

mula for the feasible circuit vectors oj that condition (6') simply asserts the

following traffic situation: for each of the feasible circuits actually used by

a given connnodity i (namely, each circuit j E [i] for which z. > 0) the total
J
origin-to-destination quasi-cost per unit of flow is the same, in fact just -yo
~

(the negative of the corresponding "return link quasi-revenue per unit of flow"

Yi)' which in turn does not exceed the total origin-to-destination quasi-cost

per unit of flow for each of the feasible circuits not used by the given com-

modity i (namely, each circuit j E [i] for which z. =0).

J
 68

Each vector z that satisfies conditions (1-4,5 1 ,6 1 ) is termed

a quasi-equilibrium circuit flow, and each such flow z generates a quasi-equilibrium

total flow x via equation (2). Although a given quasi-equilibrium circuit flow

z generates a unique quasi-equilibrium total flow x, it is obvious that a given qu~si-

equilibrium total flow x along with the corresponding quasi-revenue-cost per unit

of flow vector y can generally be generated by more than a unique quasi-equilibrium

circuit flow z; in fact, the set of all such quasi-equilibrium circuit flows z

is clearly identical to the set of all solutions z to both (the linear) conditions

(1-2) and the complementary slackness conditions

either (6 j ,y) =0 or z. =0 for j =1,2, ... ,m. (7 I)

J

It is easy to show that each Wardrop-equilibrium flow is a quasi-equilibrium

flow and that, when all cost per unit of flow functions ck are continuous, each

quasi-equilibrium flow is a Wardrop-equilibrium flow. However, if at least one

c k is discontinuous, there can be quasi-equilibrium flows that are not Wardrop-

equilibrium flows. For instance, the roadway network described oy Figure 2

clearly has a (unique) quasi-equilibrium flow x* = z* = (4,1,3) even though we

have already noted that it has no Wardrop-equilibrium flow. Actually, in view

of our previous interpretation of the discontinuity of c 2 in tenus of traffic-

control lights, it certainly seems reasonable to expect that (4,1,3) constitutes

a (stable) traffic equilibrium. In any event, we shall henceforth be concerned

(without any loss of generality) with the existence and properties of quasi-

equilibria.

2.3. Demand-equilibria. We now suppose that the input flow d i of a given

commodity i is not necessarily fixed, but is instead related via a given complete

nonnegative nonincreasing (feedback) curve to the origin-to-destination quasi-

cost per unit of flow -Yi experienced by the given commodity i when the net-

work is in a state of quasi-equilibrium. In essence then, we suppose that

 69

each dO i is related via a complete nonnegative nondecreasing curve to its cor-

responding return link quasi-revenue per unit of flow Yi' Since it is not

difficult to see that each complete nondecreasing curve has an "inverse"

that is also a complete nondecreasing curve, we can equivalently suppose that

each Yi is related via a complete nondecreasing curve

fi to its corresponding di ,

where

the "domain" of each such fi is a (not necessarily proper) sub-

interval Ii of the interval of all nonnegative real numbers.

Moreover, we also suppose that

the "range" of each such fi consists of all real numbers

(equivalently, a given di does not approach +00 without its

corresponding Yi approaching +00)

Needless to say, none of these hypotheses (H) are very restrictive in the con-

text of roadway networks,

As with each real link curve f i , it is natural to treat each return link

curve r i as a multivalued or point-to-set function Yi on Ii with function-

values

for i = 1,2" ,.,r,

each of which is a nonempty closed interval. For a given d i E IOi the set Yi (d i )

is a singleton {y i} i f and only if the "travel demand" from conmodity i results

in the given input flow d i for only a single origin-to-destination quasi-

cost per unit of flow -Yi'

By definition, the only feasible circuit flows z that place a given road-

way network in a state of (quasi")demanpweguilibrium are those z tlNtt generate a

feasible total flow x with corresponding feasible input flows

for k=1,2,.",r (3")

 70

for which there are both (fictitious) return link guasi-revenues per unit of

for k=l,2, ... ,r (Sa")

and real link guasi-costs per unit of flow

for k = r+l, r+2, .•• , n (5bl!)

such that the resulting quasi-revenue-cos~ per unit of flow vector

satisfies the following inner product conditions

for j=I,2, ••. ,m. (6")

To properly interpret the preceding definition, note that it differs from

the definition of quasi-equilibria only in that the input flows d i can now vary

over Ii~ and the symbol El in condition (Sa') has been replaced by the symbol

Yk (d t ) in condition (Sail). Consequently, a demand-equilibrium occurs when the

network is in a state of quasi-equilibrium and the input flow d i of a given com-

modity i is related via its own travel-demand curve fi to the resulting origin-

to-destination quasi-cost per unit of flow -yo experienced by the given commodity
1.

Each vector z that satisfies conditions (1,2,3",4,5"',6") is termed

a demand-eguilibrium circuit flO~ol, and each such flow z generates a eemand-i'quili·brium

~,. flow x via equation (2). Although a given demand-equilibrium circuit flow

z generates' a unique Memand-equilibrium total flow x, it is obvious that a given

demand~equUihril1m total flow x along with t:hp. corresponding quad-reve"'te-

cost per unit of flow vector yean generally be generate:o 0)1 mOlt! than Il unique

demand-equilibrium circuit flow z; in fact, the set of all such demand-equilibrium

circuit flows z is clearly identical to the set of all solutions z to both (the linear)

conditions (1-2) and the complementary slackness conditions

 71

for j =1,2, ••• ,m. (7")

It is, of course, clear that the problem of predicting those quasi-equilibrium

flows that result from fixed input flows dk is a special case of the problem of

predicting demand-equilibrium flows -- the case in which each return link function

Yk simply has a singleton domain {dk } and a single function value Yk(dk ) = El

!2.!. k = 1,2, ••• , r. Consequently, we shall henceforth concentrate (without any

loss of generality) on the existence and properties of demand-equilibria.

3. The Key Characterization. First, we consider the irreducible commodity

I ~ rill ~ i ~ rand \ contains at least one number d; > O}

and the (resulting) irreducible circuit (path) family

6.
J=U [i].
I

If I (and hence J) is empty, it is clear that the only feasible circuit flow

z = 0 places the given network in a state of demand-equilibrium (with a feasible

total flow x = 0 and a [rather meaningless nonunique] corresponding quasi-revenue-

cost per unit of flow vector y with components Yk =1 for k=1,2, ••• ,r, and

Yk=ck{O) for k=r+l, r+2, ••• ,n). Consequently, we shall henceforth assume

(without any real loss of generality) that I (and hence J) is not empty.

Now, we incorporate the feasibility conditions (1,2) into a set

X~ (x=~ Zj 6j I Zj ~O for j O}, (8)

J

which is obviously a "convex polyhedral cone generated by" the irreducible circuit

(path) vectors aj , j E J. We also introduce its "dual cone"

Y~ [y EEn I 0 ~<;K,y) for each x EX}, (9a)

which clearly has the representation formula

Y=[yEE IO~(oj,Y)for each jEJ}. (9b)

n
 72

This representation formula (9b) shows that Y is also a convex polyhedral cone

(whose generators can be computed in any given situation with the aid of a

linear-algebraic algorithm due to Uzawa [401).

By definition, the network-equi libriu ••l conditions (for the general
demand-equilibrium problem) are:

(1) x EX and

(II) 0= (x.y)

(III) for k=1.2, .... n.

Each vector (x.y) that satisfies these conditions is termed a network-equilibrium

~.

The following theorem is the key to almost everything that follows.

Theorem 1. Each demand-equilibrium total flow x along with the corresponding

quasi-revenue-cost per unit of flow vector y produces a network-equilibrium vector

(x,y). Conversely. each network-equilibrium vector (x,y) produces a demand-

equilibrium total flow x along with a corresponding quasi-revenue-cost per unit

of flow vector y.

l'roofA First, note from the defining equation (8) for X that conditions (1,2,3",4.5"

are equivalent to the firs.t part of condition (I) and all of condition (III).

Consequently, we can complete our proof by showing that condition (6") is equivalent

to the second part of condition (I) and condition (II).

Toward that end. simply note from the defining equation (8) for X that

(x,y);: L; z (oj .y) for x ~X. (10)

J j

It is then an immediate consequence of this identity (10) that condition (6")

imp1~oB the second part of condition (I) and condition (II). On the other hand,
it is an immediate consequence of the representation formula (9b) for Y and this

identity (10) that condition (6") is- also iUlplied by the second part of condition

(I) and condition (II). q.e.d.

 73

It is, of course, a consequence of (the key) Theorem 1 that we can hence-

forth concentrate (without any loss of generality) on the existence and properties

of network-equilibrium vectors (x,y). There are at least two reasons why such

a concentration pays extremely high dividends: 1. the network-equilibrium

conditions (I-III) are mathematically simpler than the demand-equilibrium con-

ditions (1,2,3",4,5",6"), in that the latter also explicitly involve demand-

equilibrium circuit flows z; and 2. the network-equilibrium conditions (I-III)

turn out to be essentially the "extremality conditions" for (generalized) geometric

programming and hence provide a mechanism through which the powerful theory of

(generalized) geometric programming can be applied to the study of traffic

equilibria.

The next section is devoted to actually constructing those geometric pro-

gramming "dual problems" whose corresponding extremality conditions coincide

with the network-equilibrium conditions (I-III).

4. Dual Variational Principles. Imitating the theory of monotone networks, we

first "integrate" each complete nondecreasing point-to-set function Yk' We

do so by computing the (Riemann) integral of any real-valued (point-to-point)

function Ok whose domain coincides with the domain of Yk and whose function

values ck(~) EYk(~)' Of course, for k=r+l,r+2, ••• ,n, the given real link

travel cost per unit of flow function c k is a legitimate choice for Ok' In any

event, our monotonicity hypotheses (H l ) and (H4 ) along with monotone function

theory guarantee that such integrals exist and do not depend on the particular

function-value choices 0k(~) ~Yk(xk)' Consequently, we denote such integrals

by the SymbOl~Yk(S)dS even though Yk is generally not a real-valued function.

Our monotonicity hypotheses (Hl ) and (H4 ) along with monotone function

. theory also guarantee that each such integral [kyk(S) ds gives a "convex function"

gk whose domain
 74

Ck ~ the (interval) domain of vk

along with any endpoint ek not in

the domain of Yk but for which the (11a)

e
improper integralS kVk(s)dS converges,

and whose function-values

(Ub)

where

O'k is any convenient point in the domain of vk

for which 0 ~Vk(O'k) (the existence of such an

O'k being guaranteed by hypotheses (H2 ) and (H6 ), (llc)

with hypothesis (~) actually implying that

O'k=O for k=r+l,r+2, ••• ,n).

Each such convex function gk is known to be "closed" in that its "epigraph"

is a (topologically) closed subset of E2 •

On the other hand, each clos!!d convex function gk with domain Ck ~ El has

a (point-to-set) "subderivative" function Ogk with function-values

at least one of which is nonempty. Needless to say, subderivatives are actually

generalized derivatives -- in that a convex function gk with domain Ck ~ El is

differentiable at a point ~ EC k if and only if agk(xk ) is a singleton [Yk J,

in which case the derivative gk(xk ) =Yk' However, for our purposes, the key

fact is that the subderivative function agk for an arbitrary closed convex

function gk with domain Ck I; El is a complete nondecreasing point-to-set function

 75

integra l is just
(corresp onding to a complet e nondec reasing curve) -- whose
In particu lar
gk (plus, of course, an arbitra ry "consta nt of integra tion").
that
then, the definin g equatio ns (11) for our functio ns gk imply

the domain of Yk is identic al to the set of all

(12a)
points ~ for which Ogk(xk ) is not empty,

and
(12b)

relatio n
The reader is probabl y already familia r with the more special
functio ns Yk and their
between (comple te) nondec reasing continu ous real-va lued

(closed ) convex differe ntiable integra ls gk. Althoug h a proof for the more

it is omitted because
general relatio n just describ ed herein is not difficu lt,
[34]. However, as a
of its lengthi ness and the fact that it can be found in
n, note that when
simple (but importa nt) example of the more general relatio
Yk (d k ) = El , the integra l
Yk has a singlet on domain [~} and a single functio n-value
n-value gk(dk ) =0;
gk also has the singlet on domain [~} but the single functio

from which it is obvious that Ogk (d k ) = El •

of our dual
The optimiz ation problem that serves as the basis for one
guise of a functio n g
variati onal princip les involve s the functio ns gk in the

whose domain
(l3a)

and whose functio n-value s

n
g(x) ~ L: gk (~). (l3b)
1

ble" into the functio ns

For obvious reasons this functio n g is said to be "separa
s the convex ity and
gk. Moreov er, it is not difficu lt to show that g inherit

closedn ess of the gk.

followi ng
From the cone X and the functio n g we now constru ct the
 76

optimization problem.

Problem A. Using the "feasible solution" set

calculate both the "problem infimum"

~=infg(x)
xES
and the "optimal solution" set

S*~{xES I ~=g(x)J.

The fact that X is a cone means that problem A is a (generalized) "geometric

proEramming prohl em". (Justification for this terminology is given in [28,29, 32]) •

Moreover, the fact that X is actually a convex polyhedral cone and the fact that

g is a closed convex separable function imply that problem A is a closed convex

separable prograruning £roblem with (essentially) linear constraints.

It, is, of course, obvious that each feasible total flow x is also a feasible

solution to problem A. On the other hand, it is clear that problem A can have

a feasible solution x that is not a feasible total flow only if C k is actually

larger than the domain of Yk for at least one k.

Imitating the theory of monotone networks, we also integrate each complete

-1
nondecreasing point-to-set "inverse function" Yk ' whose function-values

for k=1,2,oo.,n,

each of which is a nonempty closed interval. In particular, each integral

{ k Yk-1 (t) dt gives a closed convex function hk whose domain

(14a)
 77

and whose function-values

(14b)

Naturally, each such hk has the property that

(15)

As a simple (but important) example, note that when Yk has a singleton domain
-1
[~} and a single function-value Yk (d k ) = El , its inverse Yk has domain El and
-1
a constant function-value Yk (Yk) '" [~}, so Dk = El and hk (Y k ) '" dky k ; from which it

k
is obvious that ohk (Y k) = [h (Y k) } '" [d k }·

The optimization problem that serves as the basis for our other dual

variational principle involves the functions hk in the guise of a function h

whose domain

(16a)

and whose function-values

(16b)

Of course, this function h is separable, and h inherits the convexity and

closedness of the hk •

From the cone Y and the function h we now construct the following optimization

problem.

Problem B. Using the feasible solution set

t:.
T = Y n D,

calculate both the problem infimum

'±'~ infh(y)
yET
 78

and the optimal solution set

T'~ ~ ry n I l' = h (y) }.

The fact that Y is a cone means that problem B is a (generalized) geometric

prograrrnning problem. Moreover, the fact that Y is actually a convex polyhedral

cone and the fact that h is a closed convex separable function imply that problem

B is a closed convex s~parableprogramming problem with (essentiRlly) linear constraints

Problem B has important features not possEssed by problem A. In particular,

note that relations (14a) and (16a) imply that

(17)

which in turn clearly implies that

T =Y.

Since each cone Y contains at least the zero vector, we infer that problem B has

at least one feasible solution (even when problem A has no feasible solutions),

We shall eventually see that "geometric prograrrnning duality theory" re-

duces the study of traffic equilibria to a study of either problem A or problem

B. Actually, we shall then see that problems A and B should be studied simultaneously

in order to provide the most insight into traffic equilibria. In essence,

problem A describes traffic equilibria entirely in terms of total flows x, while

problem B describes traffic equilibria entirely in terms of quasi-revenue-cost

per unit of flow vectors y.

An important ingredient in our simultaneous study of problems A and B is

the fact that the cones X and Y defined by equations (8) and (9), respectively.

constitute a pair of (convex polyhedral) "dual cones"; that is

Y= (y EE
n
I 0 ~(x,y) for each x EX}, (18)
 79

while

X= (x EE
n
I 0 ~(x,y) for each y 01. (19)

Of course, equation (18) is just a repetition of the defining equation (9a) for

Y; and it is obvious from equation (18) that X is a subset of the right-sand

side of equation (19). To show in turn that the right-hand side of equation (19)

is a subset of X, and hence establish equation (19), simply use the representation

formula (9b) for Y in conjunction with both the "Farkas lemma" [11] and the

defining equation (8) for X.

Another important ingredient in our simultaneous study of problems A and

B is the fact that the functions gk !n4 hk defined by equations (11) and (14),

respectively, constitute a pair of (Closed convex) "conjugate functions"; that is,

(20a)

and

(ZOb)

while

(2la)

and

(2lb)

It is obvious from the defining equation (14a) for Dk that equation (20a) can be

established simply by showing that Dk(i.e. El ) is a subset of the right-hand

side of equation (20a). To do so, first observe from our hypotheses (H3) and (H6 )

along with the defining formulas for Yk that for a given Yk in Dk (i.e. &1) there
 80

exists at least one xk in the domain of Yk such that Yk is in Yk(x k), Now,

observe from equation (12b) that such a Yk must also be in ;;gk(Xk ), which in

turn implies via an elementary manipulation of the defining inequality for

In addition to showing that Yk is in the right-hand side of equation (20a) and

hence establishing equation (20a), the preceding displayed equation also

establishes equation (20b). The reason is that the defining equations (11) and

(i.e. the [signed] area ~Yk of the rectangle with sides [O,x k ] and [O,y k ] is

the sum of the [sLgned] areas gk(xk ) and hk(y k ) when Yk EYk(xk», Similar,

though somewhat more complicated, steps can be used to establish equations (21a)

and (21b). However, for complete pro~fs of both equations (20) and equations

(21) in increasing degrees of generality see ~4,9,34].

It is also important to know that the functions g and h defined by equations

(13) and (16), respectively, inherit the conjugacy of the gk and hk ; tKat is,

D= {y EE I sup [(y,x) -g(x)l is finite} (22a)

n x EC

and

hey) = sup [(y,x) - g(x)l, (22b)

x EC

while

C=[xEE I sup [(x,y)-h(y>1isfinite} (23a)

n y ED

and
 81

g(x) = sup [(x,y) - h(y)]. (23b)

Y fD

In fact, it is easy to see that equations (13), (16), (20) and (21) imply

equations (22) and (23).

The most fundamental properties of problems A and B are not induced by

the special nature of rmadway networks -- only by the (conical) duality of X

and Y and the (functional) conjugacy of g and h. We henceforth concentrate on

such properties (28,29], some of which also depend on the separability of g and h.

Notice how problem B can be obtained directly from problem A: simply

replace the convex polyhedral cone X with its (convex polyhedral) "dual" Y,

and replace the closed convex separable function g with its (closed convex

separable) "conjugate transform" h. lhe symmetry of (conical) duality demon-

strated by equations (18) and (19) together with the symmetry of (functional)

conjugacy demonstrated by equations (22) and (23) imply that the problem

obtained by applying the same transformation to B is again A. lhis symmetry

justifies the terminology "(g:eometric) dual problems" for problems A and B· Need-

less to say, it also induces a symmetry on the theory that relates A to B -- in

that each statement about A ~ B (whose proof uses only the duality of X and

Y togehher with the conjugacy of g and h) automatically produces an equally

valid "dual statement" about B ~ A (obtained by interchanging the symbols X

and Y, the symbols C ~ D, and the symbols g and h). However, to be concise,

each dual statement will be left to the reader's imagination.

Unlike the usual min-max formulations of duality in mathematical pro-

gramming (e.g. in linear programming), both problem A and its dual problem Bare

minimization problems. lhe relative simplicity of this min-min formulation

will soon become clear, but the reader who is accustomed to the usual min-max

formulation must bear in mind that a given duality theorem generally has slightly

different statements depending on the formulation in use. In particular, a

 82

theorem that asserts the equality of the min and max in the usual formulation

asserts that the sum of the mins is zero (Le, ~ +'1'=0) in the present for-

mulation. (The reader interested in the prec~se connections between the various

formulations of duality in mathematical programming should consult [26] and

the references cited therein).

By definition, the "extremality conditions" (for separable geometric pro-

grannning) are:

(I) xEX and yEY

(II) 0= (x,y)

(III) for k=1,2, ••• ,n.

Each vector (x,y) that satisfies these conditions is termed an extremal. vector.

From equations (12) we immediately see that each network-eguilibrium vector

(x,y) is an extremal vector, and each extremal vector (x,y) is a network-

eguilibrium vector. Consequently, we can henceforth concentrate (without any

loss of generality) on the existence and properties of extremal. vectors (x,y).

The main reason for doing so is that the following powerful theory of (generalized)

geometric programming is particularly suited to the study of extremal vectors.

5. Duality Theory. Proofs for the most fundamental duality theorems actually

exploit the conjugacy of g and h indirectly via the following (Young-Fenchel)

"conjugate inequality"

(x,y) s: g (x) + h (y) for each x E C and each y E D.

This inequality is an immediate consequence of either equations (22) or equations

(23); and its equality state can be conveniently characterized in terms of the

"subgradient" set

Og(x)~{YEE Ig(x)+(y,xl-X)s:g(x l ) for each x'Ecl.

n
 83

In particular, elementary (algebraic) manipulations of the defining inequality

for og (x) show that y E D and (x,y) = g (x) + h (y) when y E og (x). On the other

hand, equally elementary (algebraic) manipulations show that y E og (x) when

(x,y) = g (x) + h (y). Hence,

(x,y) =g(x)+h(y) if and only if yEdg(x);

from which it follows via the symmetry of (functional) conjugacy that

y E og (x) if and only if x E oh (y).

Finally, in the event that g and h are separable (e.g. in the present ~oad­

way network case), it is clear that

yEog(x) i f and only i f YkEogk(xk) for each k=1,2, ••• ,n;

and it is, of course, equally clear that

for k = 1, 2, ••• , n •

The following duality theorem relates the dual problems A and B directly

to the extrema1ity conditions (I-III) -- and hence to traffic equilibria.

This theorem is also basic to all succeeding duality theorems.

Theorem 2. If x and y are feasible solutions to the dual problems A and B

respectively (in which case the extrema1ity conditions (I) are satisfied), then

o !i:g(x) +h(y),

with equality holding if and only if abe extrema1ity conditions (II) ~

(III) are satisfied; in which case x and y are optimal solutions to problems

A and B respectively.
 84

Proof. The defining inequality for Y and the conjugate inequality for h show

that 0,;; (x,y) :s;; g(x) +h(y), with equality holding in both of these inequalities

if and only if the extrema1ity conditions (II) and (III) are satisfied; in

which case a subtraction of the resulting equality 0 = g (x) + h (y) from each

of the inequalities O!:g(x')+h(y) and O!:g(x)+h(y') (which are valid for

arbitrary feasible solutions x' and y') shows that x and yare actually

optimal solutions. q.e.d.

The fundamental inequality given by Theorem 2 implies important properties

of the dual infima f and'l' .

Corollary 2A. If the dual prob1 ems A and B both have feasible solutions. then

(i) the infimum ~ for problem A is finite, and

O!:f+h(y) for each yET,

(ii) the infimum 'l' for problem B is finite, and

Od+'l'

The proof of this corollary is, of course, a trivial application of Theorem 2.

The strictness of the inequality 0 s ~+ 'l' in conclusion (ii) plays a crucial

role in almost all that follows. In fact, dual problems A and B that have

feasible solutions and for which 0 <f + '1' are said to have a "duality gap" of

~+'l'. Although duality gaps do occur in (generalized) geometric programming,

we shall eventually see that they do not occur in the present context of traffic

equilibria. This lack of duality gaps is extremely fortunate because of their

highly undesirable properties. In particular, we shall soon see that they weaken

the bond between the dual problems A and B and imply that the extremality con-
 85

ditions have no solutions (Le. there are no "extremal vectors" even though

problems A and B may have optimal solutions). They also tend to destroy the

possibility of using the inequality 0 :s: g(x) + h (y) to provide an "algorithmic

stopp;i:ilg criterion".

Such a criterion results from specifying a positive tolerance e so that

the numerical algorithms being used to minimize both g(x) and h(y) are terminated

when they produce a pair of feasible solutions xt and yt for which

t t
g(x )+h(y ):s:e.

Because conclusion (i) to Corollary 2A along with the definition of ~ shows

that
t t
- h (y ) :s: ~ :s: g (x ),

we conclude from the preceding tolerance inequality that

t
~:s:g(x):s:~+e.

Hence, W can be approximated by g(xt) with an error no greater than e (and,

t
dually, Y can be approximated by hey ) with an error no greater than e). This

stopping criterion can be useful even though it does riot estimate the proximity

of xt to an optimal solution x* (and, dually, even though it does not estimate

t
the proximity of y to an optimal solution y*). In any event, it is clear

that the algorithms being used need not terminate unless W+Y < e, a situation

that is guaranteed for each tolerance e only if 0 = W+ Y (Le. there is no

duality gap).

The equality characterization of the fundamental inequality given by Theorem

2 implies the key relations between extremal vectors (x,y) and optimal solutions

x E S* and y E T*.

Corollary 2B. If (x+,y+) is an extremal vector (i.e. if (x+,y+) is a solution

to the extremality conditions (I-III», then

 86

+
(1) and y E T*.

(ii)
+
S*'"[xEXlo=(x.y). and ~Eohk(Yk> for k=1.2 ••••• n}
+

+ +
(iii) T*=(yEY IO=(x .y). and YkEogk(xk> for k=1.2 ••••• n}

(iv) o = ~ +'!'.

On the other hand. if the dual Broblems A and B both have feasible solutions

and if O=~+'!', then (x,y) is an extremal vector if and only i f xES* and yET*.

The proof of this corollary come~ from a trivial application of Theorem 2 along

with the (previously mentioned and easily established) conjugate transform

Corollary 2B readily yields certain basic (nontrivial) information about

traffic equilibria. In particular, the first part of Corollary 2B asserts

that if there is at least one network-equilibrium vector (x*.y+), then there is

no duality gap. On the other hand, the se~or~ part of Corollary 2B clearly

implies that if there is no duality ga~. then the set E of all network equilibrium

vectors (x.y) is iust the cartesian product S* X T* of the optimal solution sets

S* and T*. Consequently. when network equilibria exist, each demand-equilibrium

total flow x* (namely, each x* E S*) can be paired with each equilibrium quasi-

revenue-cost per unit of flow vector y* (namely, each y*E T*) to produce a

network equilibrium vector (x*,y*) E E. Moreover. network equUlbria exist if

and only if hypotheses (H) are sufficiently strong to guarantee the absence

of a duality gaB along l~ith the existence of optimal solutions x* E S* and

y* E T*, Actually, the existence of network equilibria is established in

section 9 by using the netwo.-k classification and reduction given in section

7 to show that 0 = ~ + 'I' and neither S* nor T* is empty.

Computationally, it is important to note that conclusions (i-iii) of

Corollary 2B provide methods for calculating all netl~ork-equilibrium

vectors (x,y) E E from the knowledge of a single net\~ork-eC1uilibrium vector

 87

(x+,y+) E E (or even from just the knowledge of a single optimal solution

x* E S*, or a single optimal solution y* E T*). Of course, the (nonempty) set Z(x,y)

of all demand-equilibrium circuit flows z that generate a given network-equilibrium

vector (x,y) E E can be calculated simply by calculating the set of all

solutions z to the feasibility conditions (1,2) and the complementary slackness

conditions (7") as explained in the next-to-1ast paragraph of subsection 2.3.

It is also important to note that conclusions (ii-iii) of Corollary 2B

imply that both the set S* (of all demand-equilibrium total flows x*) and

the set T* (of all equilibrium guasi~revenue-cost per unit of flow vectors

y*) are convex pOLYhedral sets. and hence so is the set

E = S* X T*

(of all network-equilibrium vectors (x*,y*».

+
Ogk(xk ) are closed intervals (as previously observed); and convex polyhedralness

is preserved under cartesian p~oducts. Of course, the set Z gf all demand-

e~uilibrium circuit flows z has ~he re~resentation formula

Z= u Z(x,y)
(x,y) E E

where each set

Z(x,y) ~ [z E Em I conditions (1,2) and (7") are satisfied}

is clearly a bounded convex polyhedral set,

6. Uni~ueness Theorem~. It is worth noting that various uniqueness theorems result

from the representation formula E = S* XT* by imposing strict monotonicity and/or

continuity hypotheses on the travel cost per unit of flow functions c k ' In partic-

ulJar, if there is a demand-equilibrium total flow x* such that a given c k is strictly

increasing on a (relative) neighborhood of x~, then every other demand-equilibrium

total flow x** exhibits the same unique total flow x*~ = x~ on link k. The reason

is that strict monotonicity of c k on a (relative) neighbothood of x~ implies

 88

that Yk (x~) and Yk (x*~) do not intersect when x~ f x*~; so a given y~ can not be

in both Yk(x~) and Yk(x~) unless x~=x*~. On the other hand, i f there is a

demand-equilibrium total flow x* for which x~>O and such that the given c k is

continuous at x~, then each e9-uilibrium 9-uasi-revenue-cost J?er unit of flow

vector y* exhibits the same unique quasi-cost per unit of flow Y~ = ck(x~) for

link k. The reason is that continuity of c k at x~ implies that Yk(x~) contains

a unique element c k (x~) when x~> O. Finally, if there is a unique quasi-cost

per unit of flow y~ for each (real) link k for which there is at least one

demand-equilibrium total flow x* such that x~>O (e.g. if each c k is continuous

on its domain [O,b k », then each equilibrium quasi-revenue-cost per unit of flow

vector y* exhibits the same unique quasi-revenue per unit of flow yt for a given

(return) link i if xr-> 0 for each demand-equilibrium total flow x* (e.g. i f 0

~

does not belong to the domain I. of y., as is the case when I. = &d.} and di>O).
~ ~ ~ ~

The reason is that the preceding' observation along with conditions (6") and the

finiteness of the feasible circuit family [i] clearly imply that yt can have at

most a finite number of values; while the convexity of T* clearly implies that

yt'can have infinitely many values if it has more than one value.

Other uniqueness theorems resul t from the representation formula E = S* XT*

by imposing comparable hypotheses on the travel demand curves fi' In particular,

if there is a demand-equilibrium total flow x* and a (returh) link i such that

Yi (xt) and y i (x*t) do not intersect when xt f x*t, then every other demand-

e~uilibrium total flow x** clearly exhibits the same unique total (input) flow

x*t =xt on (return) link i. On the other hand, if there is a demand-equilibrium

total flow x* and a (return) link i such that y.(x~) contains a unique element
~ ~

yr, then each equilibrium quasi-revenue-cost per unit of flow vector y* clearly

exhibits the same unique quasi-revenue per unit of flow yt for (return) link 1.

Even when there is a unique network-equilibrium vector (x*,y*), there can

 89

obviously be infinitely many demand-equilibrium circuit flows z*. In fact,

uniqueness of z* in that case clearly requires a positive independence of those cir-

cuit vectors oj for which <oj ,y*) = 0 and z'!'> O. It is clear that such circuit vectors
J
oj are less likely to be positively independent, the more the corresponding cir-

cuits "overlap" (Leo the more possibility there is for congestion). In any event,

it is easy to construct simple examples where neither x*, y* nor z* have any unique

components. Such cons,ructions are, however, left to the reader's imagination.

7. Network Classification and Reduction. In this section the family of all (roadway)

networks is partitioned into a family of "canonical networks" and a family of "degen-

erate networks". Each degenerate network has a "reduced form" that is canonical,

and each canonical network is its own reduced form.

This mapping of all networks onto canonical networks induces another partitioning

of the family of all networks, namely, a partitioning into an infinite collection

of "equiValence classes". This collection of equivalence classes is in one-to-one

correspondence with the family of all canonical networks. Corresponding to a

given canonical network is the equivalence class that consists of the given canonical

network and all degenerate networks that have the given canonical network as a

reduced form.

Networks can be classified as canonical or degenerate simply by inspection.

If a network is degenerate, the inspection yields its reduced form, which is just

a "subnetwork" of the given degenerate network. Moreover, the geometric programming

problems A and 0 associated with a network N and its reduced form n, respectively,

have the same infima ~ and ~. Likewise, the geometric programming problems B

and B associated with a network N and its reduced form n, respectively, have the

same infima '1' and 1jI. Furthermore, the optimal solution set S*(T*) for problem A

(B) can be generated from the optimal solution set "J*(:J*) for problem 0(6);

and hence the extremal vector set E = s* )( T* for network N can be generated from the
 90

extremal vector set a =,,/* X:r* for network '1/. Thus, the family of all reduced

forms (canonical networks) is the most important family from a computational point

of view. It is also the most important family from a theoretical point of view,

because the absence of duality gaps and the existence of dual optimal solutions can

readily be proved for all networks N once they are established for all canonical

networks '1/ •
Roughly speaking, a network N is canonical if each commodity i has

at least one feasible input flow d~ > 0 and if each link k in the network N is part
~

of at least one circuit j (for some commodity i for which j E [i]). In essence then,

a network N is canonical if it is impossible to infer from the feasible input

flows and the feasible circuit families alone that some link k (possibly a return link)

is unusable (i.e. x k =0 for each feasible total flow x). Naturally, a degenerate

network N is then a network that contains unusable links k (some of which may be

return links), each of which is to be pruned away (along with the corresponding com-

modity k in the case of an unusable return link k) when constructing its reduced

form n.
To give a precise classification of the most general network N (corresponding to

the demand-equilibrium problem described in section 2), use the "irreducible circuit

fafnily" J introduced at the beginning of section 3 to define the irreducible link set

K ~ [k 11:S" k,;; n and there exists a j E J for which o~ = l}.

Then, network N is said to be canonical if K = [1,2, ••• ,n}, and degenerate if

KC(l,2, ••• ,n}.

Since we have already assumed (at the beginning of section 3) that the

"irreducible commodity set" I (and hence J) is not empty, we immediately see

that K is not empty. Consequently, deleting those links k~ K from network N

produces a (nonempty) subnetwork '1/ of N, which is termed the reduced form of N.

Naturally, the only commodities that flow on '1/ (over the links k E K) are those
commodities i E I; but the feasible circuit family for such a commodity i is
 91

clearly still [i]. Needless to say, each link k E K retains the same "characteristic

functions" Yk' and hence 71 inherits from N all of the hypotheses (H) along with

all of the properties previously derived therefrom. In addition, 71 has

properties not generally possessed by N -- the most important ones stemming

from the obvious fact that 71 is canonical and its own reduced form.

Two simple observations are crucial to ultimately showing that essentially

nothing is lost in replacing network N with its reduced form 71. The first

observation is that the defining equation (8) for X along with the defining

equation for K clearly implies that

each vector x E X has components x k = 0 for k ~ K. (24)

The second observation is that identity (10) for X along with the defining

equation for K clearly implies that

the vector y- in En with components

_ {6for,kEK
Yk ~ (25)
-1 for k~K

is orthogonal to each vector x E X,

and hence y - E Y.

In replacing network N with its reduced form 71 it is helpful to employ the

linear transformation

defined so that the image v =£(v) is obtained from v simply by deleting the k'th

component of v for each k ~ K. Thus, n is equal to the number of links in 71; and,

unless o~herw1se stated, a vector represented by a script symbol V resides in the

image space En of £. However, it is convenient not to relabel the indices on the

components of V; consequently, the j'th component of v may be vk = vk where j <k.

It is clear that the geometric programming problem a associated with the re-

duced network 71 is constructed from the cone

 92

Z=L(X) (26)

and the function () whose domain

C- = X C (27a)
Ii:. k

and whose function-values

(27b)

ProblemO. Using the feasible solution set

calculate both the problem infimum

({l fl inf () (x)

=xERI
and the optimal solution set

It is equally clear that the geometric programming problem B associated

with the reduced network n is constructed from the cone

~! =L(Y) (28)

and the function h whose domain

./)= X D (29a)
Kk
and whose function-values

(29b)

Problem B. Using the feasible solution set

calculate both the problem infimum

Ijt fl inf h(y)

=yE:r
and the optimal solution set

:r* ~ [yU-/ljt =h(y)].

 93

We have already observed that the reduced form n inherits from network N all

of the hypotheses (H) along with all of the properties previously derived therefrom.

Consequently, the "reduced form" a inherits from problem A all of the previously

derived properties of A; and the "reduced form" B iDherits from PI!Oblem B all

of the previously derived properties of B. In addition, problems a and B have

important properties not generally possessed by problems A and E, respectively.

However, such properties can be fully exploited only after problems A and Bare

more completely related to their reduced forms a and R, respectively.

The following theorem brings to light the most important relations between

problem A and its reduced forma.

Theorem 3. Problem A has feasible solutions if and only if its reduced form

a has feasible solutions, in which case

(i) qJ=t

(ii) Jk =;l(S*)

(iii) S* =[1 ("tk) nS

Proof. If x E X, then X =:l(x) fZ by virtue of equation (26). If, in addition,

x f C, then x EC- and (J~) = g (x) by virtue of equations (27) and relation (24), along

with both our choice of ~ in equations (11) and the definition of K in terms of

J and I. These two facts establish the following lemma.

~ a. If xES, then x =:l(x) EJ, and (J~) = g (x).

Now, i f X ez, then x =ol(x) for some x fX by virtue of equation (26). More-

over, relation(24) asserts that xk = 0 for each k i K; so the definition of :l

shows that x is uniquely determined by x. If, in addition, xE(!', then xEC and

g(x) =(J~) by virtue of equations (27) and relation (24), along with both our

choice of ~ in equations (11) and the definition of K in terms of J and I. These

two fact-s establish the following lemma.
 94

Lemma b. If xLI, then there is a unique xE S such that x=£(x), and g(x) =g'x-).

To prove the first assertion of Theorem 3, first observe that Lemma a

implies that the reduced forma has feasible solutions when problem A has feasible

solutions, and then observe that Lemma b implies the converse.

To prove conclusion (i) of Theorem 3, use Lemma a to infer that (jl';; ~, and

then use Lennna b to infer that ~ ';;(jl.

To establish conclusion (ii) of Theorem 3, first let xE£(S*). Then

x =£(x) for some xES such that g (x) = iji. Thus, from Lennna a and conclusion (i)

we deduce that xE.,II<, which shows that .,II<:1£(S*). Now, let xE.,II<, which

means that X E,.P and g Q:) = (jl. Then, from Lennna b and conclusion (i), we deduce

the existence of an xES for which x=£(x) and g(x) =~. Thus, xE S* and

hence xE£(S*), which establishes the relation ,.P* I;£(S*) and consequently com-

p1etes our proof of conclusion (ii).

-1
To prove conclusion (iii), first observe that the relation S*!;£ (,.1*) nS
follows innnediately from conclusion (ii) because S* C S. Now, suppose that
-1
xE£ o(,.P*) ns, which means there is an xE,.P* such that £"(x) =x. Then PQ:) =(jl,

and g (x) = g Q:) because of Lemma b. These two equations and the assumption that
-1
xES imply that x E s* by virtue of conclusion (i). Hence, S* =£ (.,11<) n s,
and thus our proof of Theorem 3 is complete.

The following theorem brings to light the most important relations be-

tween problem B and its reduced form B.

Theorem 4. Both problem B and its reduced form B have feasible solutions, and

(i) Ijr = 'i'

(ii) :1* =£(T*)

(iii) T* =£-1 (:11,) n T.

Proof. As previously noted, the existence of feasible solutions to problem B

 95

is a direct consequence of equation (17) and the fact that each cone T = Y contains

at least the zero vector. Since the reduced form B inherits all of the pre-
viously derived properties of B, we infer that problem B also has a nonempty
feasible solution set :r ='11. This proves the first assertion of Theorem 40
If yET (Le. yEY), then y=£(y) q, (Le. yE:r) by virtue of equation (28).

Moreover, h(y),;h(y) by virtue of equations (29), along with the first part of
-1
hypothesis (H2 ) , the construction of Yk from c k ' and both equation (14b) and

the definition of K in terms of J and I. These two facts establish the following

lemma.

~ c• ..!E yET, then y=£(y) E:r, and hey) ,;h(y).

Now, i f yE:r (Le. yE~I), then y=£(y) for some yEY (Le. yET) by virtue

of equation (28). Moreover, relation (25) clearly implies that such a y is not

uniquely determined by y and can in fact be chosen so that Yk ,; 0 for k f:. K. If

such a y is chosen, then hey) =h(y) by virtue of equations (29), along with the
-1
first part of hypothesis (H 2), the construction of Yk from c k ' and both

equation (14b) and the definition of K in terms of J and I. These facts establish

the following lemma.

Lemma d. If yE:r, then there is at least one yET such that y=,i(y),

and hey) =h(y).

We can now prove conclusions (i), (ii), and (iii) of Theorem 4 by using

Lemmas c and d in the same manner that Lemmas a and b were used to establish

conclusions (i), (ii), and (iii) of Theorem 3 .de leave the details to the

reader and hence consider the proof of Theorem 4 to be complete.

Needless to say, Theorem 3 shows that problema is equivalent to problem

 96

A, and Theorem 4 shows that problem B is equivalent to problem B. Consequently,

net\~ork 7i is equivalent to network N.

8. Roundedness Theorems. It is worth noting that boundedness of the ex-

trema1 vector set E (consisting of all network-equilibrium vectors (x,y» is

intimately related to boundedness of the dual optimal solution sets S* and T'~.

In particular, the representation formula E = S* X T* clearly implies that E is

bounded if and only if S* and T* are both bounded. Moreover, we shall soon

see that hypotheses (H) guarantee that S'~ is all"ays bounded, while T'" is

bounded only if network N is canonical. Needless to say, an immediate corollary

to these facts is that E is bounded only if N is canonical. Moreover, it is

worth noting that the boundedness of s* implies via conditions (1,2) and the

non-negativity of the feasible circuit vectors 6 j that the set Z (consisting

of all demand-equilibrium circuit flows z) is always bounded.

To show that S* is bounded, it is of course sufficient to show that g(x)

approaches +'" when some component Xq of x E C becomes unbounded (because g(x''') '= ~

for each x* E S*). Now, using the defining equation (13a) for C along with the

defining equation (lla) for Ck ' we infer from the defining equations for (the

domains of) Yi and Yk that each such component Xq is bounded from below by O.

Moreover, our hypotheses (H) and the defining equations (lIb) and (llc) for gk(xk )

clearly imply that gk (xk ) ;0, 0 for each x k E C k while gq (x q ) approaches +00 when

x approaches +00. The desired conclusion is now a direct consequence of the

q
defining equation (13b) for g(x).

To show that T* is unbounded if network N is degenerate (and T* is not

-1
empty), simply note that the non-negativity of the inverse functions Yk (as

implied by hypotheses (1\ -H3 ) and hypothesis (liS» along with both the defining

equations (14) and (16) for h and equation (17) for D imply that any positive
 97

multiple of the vector y described by relation (25) produces a vector in T*

when added to a vector in T*. Consequently, canonicality of N is a necessary

condition for boundedness of T* (and hence E) -- though we conjecture that

boundedness of T* (and hence E) is r.ot characterized by canonicality of N.

However, characterizations of the boundedness of T* (and hence E) can probably

be obtained via refinements of the general results given in [30].

9. The Main Existence Theorem. The following fundamental theorem shows that

our hypotheses (H) are sufficiently strong to guarantee the existence of traffic

equilibria.

Theorem 5. If there is at least one feasible circuit flow z (i.e. if there is

at least one vector z that satisfies conditions (1,2,3",4», then there is at

least one extremal vector (x+,y+).

Proof. As previously observed, the existence of a feasible circuit flow z

implies the existence of a feasible solution x to problem A. Since we have

also previously observed that problem B has at least the feasible solution 0, we

deduce from Corollary 2A that

the infima ~ and ~ are both finite.

Now, the fact that Y is nonempty and convex (as inferred from equation (9»

and the fact that D = En (as asserted by equation (17» clearly imply vie Theorem

6.2 on page 45 of [36] that problem B has a feasible solution yO E (ri Y) n (ri D),
where (ri W) denotes the "relative interior" of the given set W. From the existence

of such a feasible solution yO and from the preceding displayed statement, we

deduce via (the geometric programming version of Fenche1's) Theorem 5 in sub-

section 3.1.4 of [28] that 0 =~ + ~ and S* is not empty.

 98

In view of the final assertion of Corollary 28, we now need only show that

T* is not empty to complete our proof. To do so, we first note from conclusion

(ii) to Theorem 4 that it is sufficient to show that :r* is not empty. Toward

that end, observe that Theorems 3 and 4 along with the preceding displayed

statement imply that

the infima ~ and V are both finite.

Now, the defining equations for I and J along with the feasibility conditions

(2,3") obviously imply that the feasible circuit flow z has components z. = 0
J
for each j il. J. Consequently, the definition of network '1/ clearly implies that
the vector 2 obtained from z by deleting component z. for each j il. J is a
J
feasible circuit flOlv for '1/. Moreover, it is clear that X =;£(x) where x and x

are the total flows on '1/ and N reApectively resulting from 2 and z respectively.

From the defining equation for I and the feasibility conditions (2,3",4)

it is obvious that a slight (possibly zero) perturbation of 2 can be used to

produce a feasible 2' with strictly positive components. Moreover, it is

equally obvious that a slight (possibly zero) perturbation of 2' can then be

used to produce a feasible 20 with strictly positive components such that the

resulting total flow X o has components x~ E (riC-k ) for each k E K. Now, it h

a consequence of the strict positivity of 2 0 along with equations (8) and (26)

for X and X respectively that XO E (ri X), by virtue of the linearity of;£ and

Theorem 6.6 on page 48 in [36]. Moreover, it is a consequence of the relations

x~ E (riC\) along with equation (27a) for C- that XO E (riC,), by virtue of the

first formula at the top of page 49 in [36].

From the existence of this feasible solution X O E (ri Z) n (riC,) and from the

preceding displayed satement, we deduce via the (unstated) dual of (the

geometric programming version of Fenchel's) nleorem 5 in subsection 3.1.4 of

[28] that 0 = ~ + ~ and :r,~ is not empty. q.e.d.

 99

10. Computational Considerations. There are at least four different approaches

to computing traffic equilibria (x,y), but only (very limited versions of) two

of them have formed the basis for all computer algorithms [5,13,15,20] known

to us.

The first approach [5,13,20] seems to be the only one that is now widely

used in practice. In essence, it "simulates" the behavior of drivers by using

"shortest path algorithms" (as described, for example, in [16,19]) to

iteratively solve in heuristic ways the defining equations (1-6) for Wardrop

equilibria. Even though such algorithms have (to the best of our knowledge) been

used only when all cost per unit of flow functions ck are continuous and strictly

increasing, it is widely known that they frequently fail to converge. Of

course, such numerical difficulties are likely to be compounded when discon-

tinuous ck are present and demand equilibria are being sought.

The second approach is (to the best of our knowledge) now being set forth

for the first time (even in the special context of computing Wardrop equilibria

when all ck are continuous and strictly increasing). It uses any algotithm

(possibly those in [24]) that can solve the network-equilibrium conditions (I-III).

This approach is analogous to a well-known (rather successful) approach used

in analysing electric and hydraulic networks. In the context of such networks, con-

dition (I) constitutes the "Kirchoff current and potential (conservation) laws", con-

dition (II) is implied by (I) and hence is redundant, and condition (III) is just

"Ohm's law".

The third approach is essentially a simplification, amplification, and

generalization of the variational approach ~onsidered (in the context of Wardrop

equilibria) by Dafermos and Sparrow [7.8] and computerized by many workers (e.g.,

see [15]). It uses any algorIthm that can solve the closed convex separable
optimization problem A. Actually, any such algorithm that exploits the absence

of nonlinear constraints (in particular, any of the most powerful convex programming
algorithms [42] presently available for solving linearly-constrained problems)
 100

can be used. To do so, simply replace the only (conceivably) nonlinear con-

straints xk E Ck with the (generally weaker) linear constraints

for k=1,2, ... ,r (3' , ')

and

for k=r+l,r+2, ••• ,n. (4' , ')

This reformulation does not alter the infimum ~ and the optimal solution set Si(

for problem A, because the defining equations (lla) and (lIb) for Ck and gk(x k )

respectively clearly imply via the defining equation (13b) for g(x) that g(x)

approaches +00 when at least one component x approaches some point not in C •
q q
Of course, once a single optimal solution x* E S* has been computed, the set

E of all traffic equilibria (x*,y*) can be computed via the technique outlined

after Corollary 2B. This approach is analogous to a classical variational

approach [39,10,3,2l,2,34J used in analysing ele~tric and hydraulic networks.

In the context of such networks, problem A consists of minimizing the "content"

of the network (which is frequently just the power dissipated in the network),

subject to only the Kirchoff current (conservation) law.

The fourth approach is essentially a simplification, amplification, and

generalization of the variational approach considered by Murchland [22] and

more recently by Evans [12]. It uses any algorithm that can solve the closed

convex separable linearly-constrained optimization problem B (in particular,

any of the most powerful convex programming algorithms [42] presently available

for solving linearly-constrained problems). Needless to say, once a single

optimal solution y* E T* has been computed, the set E of all traffic equilibria

(X*,y*) can be computed via the technique outlined after Corollary 2B. This

approach is also analogous to a classical variational approach [39,10,3,21,2,34]

used in analysing electric and hydraulic networks. In the context of such

networks, problem B consists of minimizing the "co-content ll of the network

 101

(which is frequently just the power dissipated in the network), subject to only

the Kirchoff potential (conservation) law.

Unlike the first and second approaches, the only nonlin~ar functions to

be reckoned with in the third and fourth approaches (the functions gk and hk

respectively) are both contiIll' .IUS and convex. In fact, a multivalued Yk (e.g.

a discontinuous c k ) induc~5 only a lack of differentiability in an otherwise

-1
differentiable gk; and a multivalued Yk (e.g. a c k that is constant over at

least one nontrivial interval) induces only a lack of differentiability in an

otherwise differentiable h k • Naturally, the third and fourth approaches can

be carried out in unison to prov.ide an algorithmic stopping criterion, as

explained after Corollary 2A. However, that criterion may not be of much use

because it provides direct information only about the degree of convergence of

g(x) and h(y) -- not the desired degree of convergence of x and y. Nevertheless,

the degree of convergence of x and y can ;requently be deduced from the degree

of convergence of g(x) and h(y) through an analysis of the rates of change of

the Yk and the v~l (e.g. the second derivatives of the gk and the hk when they
exist). In any event, the fact that D = En (as asserted by equation (17))

obviously endows the fourth approach with a numerical advantage over the third

approach (because C'f En). Moreover, the immediate availability of the data

required in the representation formula (9b) for Y (i.e., the generators 6 j for

X) endows the fourth approach with still another numerical advantage over the
third approach (because comparable data required in the corresponding (unstated)

representation formula for X can be obtained only from the generators for ¥,

which, to the best of our knowledge, can be computed only with the aid of a

rather intricate linear algebraic algorithm due to Uzawa [40])0

Actually, network planners need only know T* to uncover the possible

"bottlenecks" in a given network. In fact, it is clear from linear programming

theory that they need only know the "extreme points" and "recession directions"

of the convex polyhedral set T* to uncover··-euch bottlenecks 0 Of course, the

 102

same information can also be extracted from the extreme points of the convex

polyhedral set S* (which has no recession directions because of its boundedness)0

Moreover, it is worth noting that either set of extreme points (and recession

directions) can be computed without computing the set Z (of all demand-equilibrium

circuit flows) -- simply by using either the second, third, or fourth approaches.

Consequently, the first approach is, no doubt, the least attractive approach

for netlyork planning, even though i t is the only approach that is now widely

used.

Finally, each of the preceding approaches should clearly be used on the

reduced network ~ when network N is not canonical. Moreover, the de-

composition principles given in (27,28,31,32] should probably be used in con-

junction with the preceding approaches when ~ is large but "sparse" (e.go when

~ results from modeling a metropolitan roadlYay network in which various

identifiable subnetworks are only IYeak1y linked to one another).

 103

References

1. Beckmann, M.J., McGuire, C.B., 'Jinsten, C.B., Studies in the Economics

of Transportation, Yale Univ. Press, New Haven, Conn., (1965).

2. Berge, C., Ghouila-Houri, A., Programming, Games, and Transportation Networks,

~iiley, New York, (1963).

3. Birkhoff, G., Diaz, J.B., "Nonlinear Network Problems", Quart. App1. Math., 13
(1956), 431-443.

~. Boas, R.P., Marcus, M., "Inverse Functions and Integration by Parts", Amer.
Math. Month., 81(1974), 760-761

5. Bureau of Public Roads, Traffic Assignment Manual, U.S. Gov't. Printing Office,
Washington, (1964).

6. Charnes, Ao, Cooper, W.W., "Multi copy Traffic Network Models", Proco Sym. on
Theory of Traffic Flow, held at General Motors Research Lab., Harren,
Michigan, (1961).

7. Dafermos, S. C., "An. Extended, Traffic Assignment Model with Application to

Two-IJay Traffic", Trans. Science, 2., (1971), 367-389.

8. Dafermos, S.C., Sparrow, F.T., "The Traffic Assignment Problem for a General
Network", Jour. Res. Nat. Bur, Standards, 73B, (1969), No.2, 91-1170

9. Diaz, JoB., Metcalf, F.T., "An Analytic Proof of Young's Inequality", Amer.
Math. Month., 77 (1970), 603-609.

10. Duffin, R.J., "Nonlinear Networks IIa", Bull. Amer. Math. Soc., 53, (1947),
963-971,

11. Duffin, R.J., Peterson, E.L., Zener, C., Geometric Programming, John \-liley
and Sons, New York, New York, (1967).

12. Evans, S.P., "Some Models for Combining the Trip Distribution and Traffic
Assignment Stages in the Transport Planning Process", paper presented at
the Intern. Sym. on Traffic Equilibrium Methods, Centre de Recherche sur les
Transports, Univ. de Montreal, (1974).

13. Federal Highway Administration, Urban Transportation Planning, ',lashington,

(1972), 111-1 to 111-20.

14. Fenchel, -,'1., "Convex Cones, Sets and Functions", Mathematics Dept. Mimeographed
Lecture Notes, PrincetoG University, (1951).

15, Florian, M" Nguyen, S., "Recent Experience with Equilibrium Methods for Studying
Congested Urban Road Networks", paper presented at the Inter, Sym. on Traffic
Equilibrium Methods, Centre de Recherche sur les Transports, Univo de Montr~al,
(1974) •
 104

16. Ford, L.K., Fulkerson, D.R., Flows in Networks, Princeton University Press,
Princeton, New Jersey, (1962).

17. Hall, M,A., "Civil Engi.neering Applications of (Generalized) Geometric Progratmn

Jour. of Eng. Optimi.zation, to appear in Vol. 1.

18. "On the Definition of Feasible Flows on Roadway Networks", in

preparation,

19. Hu, T.C., Integer Programming and Network Flows, Addison-\~esley, Reading,
Mass" (1969).

20. Huber; M., et al., "Comparative Analysis of Traffic Assignment Techniques with
Actual Highway Use", N.C.H.R.P. report No. 58.

21. Minty, G.J., "Monotone Networks", Proc. of Rcyal Soc., London, Ser. A 257,
(1960), 194-212.

22. Murchland, J.D., "Road Network Traffic Distribution in Equilibrium", paper

presented at the Conf. on Math. Meths. in the Econ. Sciences, Mathematisches
Forschungsinstitute, Oberwolfach, (1969).

23. Nguyen, S., "A Unified Approach to Equilibrium Methods for Traffic Aflsignment",
paper presented at the Inter. Sym, on Traffic Equilibrium Methods, Centre
de Recherche sur les Transports, Univ. de Montreal, November, 1974.

24,. Ostrowski, A.M., Solution of Equations and Systems of Equations, 2nd Ed.,
Academic Press, New York, New York, (1966).

25. Peterson, E. L'~, "Symmetric Duality for Generalized Unconstrained Geometric

Programming", SIAM Jour. Appl. Math., 19, (1970), 487-526.

26. '/' "Mathematical Foundations of Geometric Programming", Appendix

to "Geometric Programming and Some of Its Extensions", Optimization and Design,
(Avriel, Rijckaert, llilde, eds.), Prentice-Hall, Englewood Cliffs, N.J., (1973),
244-289.

27. , "The Decomposition of Large (Generalized) Geometric Pro-

gramming Problems by Tearing", Proc. NATO Inst. Large Scale Systems,
(Himmelblau, ed.), North Holland, Amsterdam, (1973), 525-545.

28. "Geometric Programming - A Survey", SIAM Review, to appear.

29. , "Geometric Programming - A Handbook", Chapter V of Handbook

of Operations Research, (Elmaghraby, Moder, eds.), Van Nostrand-Reinhold,
Princeton, N.J., to appear.

30. , "On the Existence of Recession Directions in Geometric Pro-

gramming", in preparation.
31. , "The Decomposition of Large (Generalized) Geometric Programming
Problems by Geometric Lagrangians", in preparation.

32. _____________ , The Mathematical Foundations of Convex and Nonconvex Programmin!

in preparation.

33. Potts, R.B., Oliver, R.M., Flows in Transportation Networks, Academic Press,
New York, N.Y., (1972).
 105

34. Rockafeller, R. T., "Convex Programs and Systems of Elementary Monotonic

Relations", Jour. Math. Anal. Appls., 12., (1967), 543-564.

35. , "Duality in Nonlinear Programming", Amer. Math. Soc. Lects. in

App1. Math!, 11, (G.B. Dantzig, A.F. Veinott, eds.), Amer. Math. Soc., Providence
R.I., (1968), 401-422. '

36. ~~~________ , Convex Analysis, Princeton University Press, Princeton, N.J.,

(1970).

37. Rosenthal, R.W., "The Network Equilibrium Problem in Integers", Networks, 1,

(1973), 53-61.

38. RId ter, E.R., "The Prediction of Network 'Equilibrium - The State of the Art",
paper presented at the Intern. Conf. on Transpor. Res., Transpor. Res. Forum,
Bruges, Belgium, (1973).

39. Thompson, H, (Lord Kelvin), Cambridge and Dublin Math. Jour., (1848),84-87.

40. Uzawa, H., "A Theorem on Convex Polyhedral Cones", Studies in Linear and
Nonlinear Programming, (K.J. Arrow, L. Hurwicz and H, Uzawa, eds.), Stanford
University Press, Stanford, (1958), 3.

41. Wardrop, J.G., "Some Theoretical Aspects of Road Traffic Research", Proc 0

Inst. Civ, Eng, Part II, 1, (1952), 325-378,

42, Zangwill, W.I" Nonlinear Programming: A Unified Approach, Prentice-Hall,

Englewood Cliffs, New Jersey, (1969).
 INTEGRATED EQUILIBRIUM FLOH MODELS
for TRANSPORTATION PLANNING*

Stella C. Dafermos
Brown University
Providence, Rhode Island, USA

1. Introduction

The conventional transport model (recursive in character) is a combination of

three submodels, trip generation, trip distribution and assignment. As a rule, these
stages are treated separately as if the travel choices involved in each one were made
independently of the other. A careful examination of this procedure, however, reveals
some significant weaknesses. Although quite sophisticated models based on behavioral
assumptions have been developed for the assignment stage, the models available for
the distribution and generation stages are less refined. As a reSUlt, serious inter-
nal inconsistencies arise. For example, the travel costs that are used as inputs for
the distribution and generation stage are not the same as those which are outputs
from the assignment stage. As a partial remedy, practitioners have devised in recent
years an iterative scheme, in which the above model is imbedded, in the hope that the
interrelationship between demand, supply and cost will be taken into account. Unfor-
tunately, the above scheme does not produce in general equilibrium flows (see e.g.
[6]). For these reasons we strongly believe that the substitution of the "first
generation" recursive transport model by a "second generation" integrated transport
model based on behavioral assumptions and the equilibrium concept will be an important
step towards better planning of transportation systems.

* This research was supported by ftS.F. grant No. GP28931.

 107

Recently some models have been proposed (e.g. [9]) which attempt to combine the
functions of generation, distribution and, potentially, assignment into a single pro-
cess. However, the special assumptions involved make these approaches not generally
applicable. Also, the computational methods associated with the above models do not
produce in general equilibrium flows (see [6]). In this article we introduce a novel
approach for modelling the transport problem based on a behavioral assumption. The
success of the behavioral models developed for the assignment problem (e.g. [8] and
the references cited therein, [2], [5], [7]) indicates that such an approach might
prove fruitful.

We adopt here the riatural behavioral assumption that each user chooses his
origin, his destination as well as his path so as to minimize his "travel cost". Of
course in the above statement "travel cost" should be interpreted in a very liberal
fashion. In reality, additional factors such as "attractiveness" of the origins
(residential areas) and destinations (places of work) have to be taken into account
but these can be incorporated into the model as "travel costs" by a straightforward
modification of the network.

Actually we propose here three variants of an integrated transport model, each

based on different reasonable circumstances. Interestingly, we establish a mathema-
tical equivalence which reduces these integrated transport problems for a network
into assignment problems for a modified network. This equivalence will enable us to
employ the experience accumulated with the assignment model in order to study the
traffic behavior predicted by the integrated transport model as well as to compute
the traffic flow equilibria in a transportation system.
 108

2. Preliminaries

A network G is a set of nodes {x} and a set of links {(x,y)} connecting

pairs of nodes. Some of the nodes are origins others are destinations. By a path

(xn- l'x)
n , where
xn are distinct nodes, Xl is an origin and xn is a destina-

tion. We let 9' denote the set of all paths of G , 9 x the set of all paths

originating at the origin X !2 the set of all paths terminating at the des-
y
tination y and !Jf the set of all paths connecting the origin-destination pair
w
w = (x,y) Traffic flow produced at the origin nodes and traveling along the

paths of the network terminates at the destination nodes thus generating a flow

pattern jV= {F} where F denotes the traffic fldw on path p. The total
p p
number of trips generated in the origin node x (trip productions) will be de-

noted by o The total nu~her of trips terminating at the destination node y

x
(trip attractions) will be denoted by D Finally, the travel demand associated
Y
with the origin-destination pair w will be denoted by d
w
It is clear that d 0 must satisfy the following flow conserva-
and D
w x
d = I y F (1)
tion equations: w
PEg£' p
W

0
x L F
pE9 p
(2)
x
L Fp D
y (3) =
PE!2y
If T denotes the total number of trips produced at all origin nodes (and

equal to the total number of trips terminating at all destination nodes) we must
also have T = L 0X = L Dy (4)
orig. dest.
We assume that each user traveling on a path p incurs a travel cost c
p
which depends on the flow patterns, i.e.,

cp = c
p
(§)

The above assumption (congestion effect) is common in the assignment model and

we refer the reader e.g. to [8J and [2J for a detailed discussion on reasonable

forms and properties of the functions c (jV).

p
 109

3. The Integrated Transport Problem

In the current transport model, at the generation stage one estimates the

trip productions and/or attractions; in the distribution stage, using as inputs

the trip productions and/or attractions, one predicts the origin-distination

travel demands and, finally, in the assignment stage, one determines how the

origin-destination travel demands will be distributed among the available paths.

Here we develop an integrated transport model based on a general behavioral

assumption and capable of combining the functions of generation. distribution and

assignment.

The following situations, each referring to different circumstances. are con-

sidered:

Situation 1: Users of the network have predetermined origins (e.g. places of resi-

dence) and destinations (e.g. places of work) and they are free to choose their

travel paths.

Situation 2: Users of the network have predetermined origins and they are free to

choose their destinations as well as their travel paths.

Situation 3: Users of the network have predetermined destinations and they are

free to choose their origins as well as their travel paths.

Situation 4: Users of the network are free to choose their origins, their destina-

tions as well as their travel paths.

In terms of the terminology introduced in Section 2, the above four situations

lead to the following transport problems:

Problem 1: The travel demand d is given for each origin-destination pair w •

w
Determine the flow pattern i7

Problem 2: The total number 0 of trips produced in each origin node x is

x
given. Determine the origin-destination travel demands d and the flow pattern Yo
w
Problem 3: The total number D of trips attracted in each destination node y
Y
is given. Determine the origin-destination travel demands dw and the flow pat-

tern Y.

Problem 4: The total number T of trips produced in all origin nodes of the net-

work (and equal to the total number of trips attracted in all destination nodes)
 110

is given. Determine the trip productions Ox ' the trip attractions D • the
y
origin-destination travel demands d and the flow pattern Y.
w
Thus Problem 1 is the classical assignment problem, Problem 2 is a combination

of production-constrained distribution and assignment, Problem 3 is a combination

of attraction-constrained distribution and assignment and Problem 4 is a comb ina-

tion of gereration, distribution and assignment.

Another situation of interest is the cowbination of the attraction-production-

constrained distribution and assignment where the trip productions °x and the
trip attractions D (satisfying eqs. (4)) are given and one has to determine the
Y
origin-destination travel demands d and the flow pattern Y. However, this
w
problem cannot be treated with the methods of this article (see the footnote of

pg. 7) and it will be studied elsewhere.

 111

4. The Behavioral Assumption

The behavioral assumption we will adopt here has been motivated by a similar

assumption generally accepted in traffic assignment models. It states that each

user of the network chooses (subject to constraints) his origin, destination and

path so as to minimize his "travel cost" given that all other users have made their

choices.

Of course, in the above statement, "travel cost" should be interpreted in a

very liberal fashion. In reality, additional factors such as "attractiveness" of

the origins and the destinations have to be taken into account but these can be

incorporated into the model as "travel costs" by a straightforward modification of

the network, e.g. by adding artificial links with travel cost representing attrac-

tiveness. With the statement of the behavioral assumption the description of the

integrated model is complete.

The choice mechanism described above leads to a flow pattern which, in addition

to the appropriate flow conservation equations, must satisfy the "equilibrium prop-

erty" that no user has any incentive to change his origin, destination or path given

that all other users adhere to their present choices. A flow pattern with the above

property will be called user-optimized.

(~.)
If the travel cost functions c (jV) are reasonable •. , a user-optimized flow
p
pattern satisfies the following equilibrium conditions:
Problem 1. A flow pattern jV, satisfying the flow conservation equations (1),

is user-optimized if and only if for each origin-destination pair w there is an

ordering, PI' P2' ••• , Ps' Ps+l' ••• ,Pm of Yew such that

U,)
Roug~ly speaking, one requires that if a user traveling on path p changes his
path wh:le ~ll,others adhere to their present choices, ~ decreases. Under this
a~s~mpt~on ~t.~s easy to establish, as in [1], the validi~y of the equilibrium con-
d~t~ons. It ~s a remarkable feature of Problems 1 2 3 and 4 that if
d c h. t 1 ' , , no user can
e re~se ~s rave cost by a unilateral decisiqn, then he cannot decrease his cost
even ~: he makes his decision ~n cooperation with a "small" group of users. It is
for th~s reason that the behav~ora1 assumption is "stable" a1thou h ' t ' b d
purel 'd' 'd 1" , g ~ ~s ase on
y :n ~v~ ua ~~t~c cons~d:ratio~s. ,On the contrary, in the case of the co~hined
product~on-attract~on-constra~ned d~str~bution and assignment the b h . 1
tion pro d h . ab ' e anora assump-
. pose e:e ~s unst le since it is clear that groups of users can consider-
ably ~mprove the~r travel costs through mutual cooperation.
 112

= c
Ps
(g) = ).
w ~ c
Ps+l
(9) :;; ... ~ c (9)
Pm

F
Pr
> 0 , r = 1, ... , s (I)

F
Pr =0 , r = s+l, ... , m

Problem 2. A flow pattern g, satisfying the flow conservation equations (2), is

user-optimized if and only if for each origin node x there is an ordering,

PI ' P2 ' Ps ' Ps +l ' ••• , Pm of ~x such that (I) holds.

Problem 3. A flow pattern g, satisfying the flow conservation equations (3), is
user-optimized if and only if for each destination node y there is an ordering,

PI ' P2 ' ••• , Ps ' Ps +l ' ••• ,Pm of fly such that (I) holds.
Problem 4. A flow pattern g , satisfying the flow conservation equations (4), is

user-optimized if and only if there is an ordering, PI ' P2 ' ... , Ps ' Ps +l '
,Pm of ~ such that (I) holds.
Thus equilibrium conditions(1) and the appropriate flow conservation equations

(I), (2), (3) or (4) determine completely user-optimized flow patterns for Problems

1,2,30r4.
 113

5. Reduction of the Integrated Model to Assignment

Here we show that the integrated transport models 2, 3 and 4, discussed in

the previous section, can be reduced to assignment problems.

We begin with transport Problem 2. We modify the original network G by

adding an artificial node ~ and joining every destination node y of G with

~ by an artificial link (y,~) . The resulting network with the same origin

nodes {x}, as the original network G, the same productions Ox' but a single
destination node ~ will be denoted by G' • Every path of G' will be of the

form p' = (p,(y,~» where p is a path of G and (y,~) is one of the artifi-

cial links joining destination node y of G with ~. It is clear now that

each flow pattern _f7 for G gi ves rise to a flow pattern Y' for G' with

r'p' =F
P
Moreover, we assume that the travel cost over the artificial links is

zero so that cp ' C9') = cp (ff). In this setting we observe that the flow pattern

3'satisfies equilibrium conditions (I) for Problem 2 for G if and only if the

flow pattern 5(' satisfies conditions (I) for Problem 1 for G' Thus, the prob-

lem of determining a user-optimized flow pattern for the transport Problem 2 over

the network G has been reduced to the problem of determining a user-optimized

flow pattern for an assignment problem over the network G' •

Similarly, transport Problems 3 and 4 over a network G can be reduced to

assignment problems over a network G' constructed by analogous procedures. Thus,

for transport Problem 3, G' will be constructed from G by adding an artificial

origin node X , which will be the only origin node of G' and then joining every

origin node x of G with X by an artificial link <X,x) with zero travel cost.

In transport Problem 4, G' is constructed as follows: We add an artificial ori-

gin node x, which will be the only origin node of G' , and we join every origin

node x of G with X by an artificial link (x, X) with zero travel cost. We

also add an artificial destination node ~,which will be the only destination

node of G' , and we join every destination node y of G with ~ by an arti-

ficial link (Y,o/) with zero cost.
 114

6. Conclusions

The equivalence established between various integrated transport models and

the assignment model enables us to employ the experience accumulated on the theory

of traffic assignment in order to establish for transport Problems 2, 3 and 4,

provided that travel cost functions are reasonable, existence, uniqueness and

stability of user-optimized flow patterns as well as efficient methods for com-

puting these patterns.

Various extensions of the integrated transport model are possible. For example,

one may consider elastic demands where the trip productions Ox or the trip attrac-

tions D are not fixed but are decreasing functions of travel cost. In this case
y
the user-optimized flow pattern can be computed using the algorithm developed in

[4] for the traffic assignment problem with elastic demands.

Another possible extension is when the paths of the network are shared by

users of several classes each of which has an individual cost function and at the

same time contributes to its own and other classes' cost function in an individual

way. In this case the equilibrium flow patterns can be studied using the methods

presented in [3] for the multi class-user traffic assignment model.

 115

7. Appendix

As an illustration of the proposed method we consider Problem 2 for the

simple network consisting of a single origin x and three destinations Yl , Y2 ,

Y3 shown in Fig. 1.

Fig. 1

We assume that the users travel cost on link i, i = 1, ••• ,6 of the netowrk is

of the form

where

The total travel demand produced at the origin x is Ox = 1000.

vie now add the imaginary destinaticn 1/1 and we join y l' Y2' Y3 with 1/1 by the

links (Yl' 1/1), (Y 2 , 1/1), (Y3' 1/1) (sketched by dotted lines in Fig. 1) with zero

travel cost. We have thus reduced the above Problem P2 for the original network

to an assignment problem for a network with a single origin-destination pair

(x, 1/1). with associated travel demand Ox.

Nodes x and 1/1 are connected by six paths which will again be numbered

by 1, ••• ,6 where path i is the (uniquely determined) path containing link i.

We also observe that the flow and user's travel cost on path i coincide with

those on a link i so that we will use f i , ci to denote the flow and the user's

cost both on link i and its corresponding path.

 116

For a network of the present type the assignment problem can be solved

in closed form (without applying any interative scheme) by the method developed

in ([1], 2.2 p. 106). We review briefly the procedure. In general, one has to

renumber the paths so that the corresponding sequence of h's becomes non-

decreasing. For our example the sequence is already nondecreasing and no re-

numbering is necessary. Next we compute a decreasing sequence of numbers >-1'

k h.
1.
0
x + E gi
i=l
\ k
1
i=l
I ~

and we determine the index s for which

Then, the solution of the assignment problem is

k=l, .•• ,s

o k :: s + 1

The determination of s is given in Table 1.

Table 1. Determination of Critical s

k k h.
1
k I
i=l gi
I
k=l
1.

gi \ hk hk +l

1 0.25 95 4300 380 400

2 1.25 495 1196 400 410

3 1. 75 700 971 410 400

4 2.75 1130 725 430 440

5 3.00 1240 747 440 800

 117

Hence s =5 and
747-380 747-400
fl = 4 91, f2
1
347

747-410 747-430
f3 = 2 = 168, f4 = 1
317

747-440
fS 4 = 77, f6 0

The resulting origin-destination travel demands are

438, 168
 118

8. References

[lJ Dafermos, S. C. and F. T. Sparrow. The Traffic Assignment Problem for a

General Network. Journal of the National Bureau of Standards, 37B, 91-118
(1969).

[2J Dafermos, S. C. An Extended Traffic Assignment Model with Applications to

Two-Way Traffic. Transportation Science, ~, 366-389 (1971).

[3J Dafermos, S. C. The Traffic Assignment Problem for Multiclass-User Trans-

portation Networks. Transportation Science. ~, 73-87 (1972).

[4J Florian M. and S. Nguyen. A New Method for Computing Network Equilibrium
with Elastic Demands. Publication No. 140, Department d'Informatique.
Universite de Montreal. 1973.

[5J Leventhal. T •• Nemhauser G. and Trotter. L•• Jr. A Column Generation Algorithm
for Optimal Traffic Assignment. Transportation Science. Z, 168-176 (1973).

[6J Manheim, M. L. Practical Implications of Some Fundamental Properties of

Travel Demand Models. M.I.T. Research Report No. R72-23. Department of Civil
Engineering. M.I.T •• Cambridge. Massachusetts.

[7J Nguyen. S. An Algorithm for the Traffic Assignment Problem. Transportation

Science. ~. 203-216 (1974).

[8J Potts. R. B. and R. M. Oliver. Flows in Transportation Networks. Academic

Press. New York, 1972.

[9J Quandt, R. E. and W. J. Baumal. The Abstract Mode Model: Theory and Measure-
ment. Northeast Corridor Transportation Project. Technical paper No.4,
June, 1966.
 EQUILIJRIUM VERSUS OPTIMUM
IN PUBLIC TRANSPORTATION SYSTH1S

Martin J. Beckmann
Brown University
Providence, Rhode Island, USA
and
Technische Universitat Munchen
Munich, West Germany

1. It is tempting to believe that public transportation systems can be completely

planned. Ready techniques such as Nonlinear Programming could be applied to predict
and control movements of people between given origins and destinations in a pres-
cribed manner. But a closer view of affairs shows that even in the most perfectly
planned system the public retains certain freedoms, and that it may choose to ignore
the intentions of planners and play games of its own. Transportation analysis cannot
afford to loose sight of "how people use transportation". [1]

Thus there is a need for behavioral analysis of user decisions in public trans-
portation systems as well as in automobile traffic. This is the motivation for the
following attempt to develop an equilibrium model of transport user decisions in
public modes. It should be emphasized that this is a first cut which leaves out many
interesting aspects of this difficult problem area. However, we believe that some
important problems are captured in the following simple model.

2. We consider a transportation system in which carriers (pods, buses, trains,

planes, etc.) travel at a frequency of Vik departures per hour on links (i,j) of a
network whose nodes are denoted by single letters i,j,k and whose directed links are
denoted by the ordered pairs of vertices connected. In the following we ignore all
problems of "making connections". It is assumed either that the frequency of depart-
ures is large enough to make the connection delays negligible or that scheduling is
(near) perfect. The problem of connection delay is certainly an important one in prac-
tical situations but we have chosen not to pursue it in this paper.
 120

In a first version of the problem the demand x ik for

trips between origins i and destinations k is given and
fixed. We consider the problem of allocating carriers to
routes in such a way that total transportation cost of
carriers and people is minimized. Because of capacity limita-
tion on links or through intersections the flow of carriers
through certain links or nodes may be restricted. Moreover,
there may be a bound on the total available number of
carriers.

The allocation problem to be formulated is a modification

of the classical problem of allocating aircraft to routes
studied by Dantzig [2]. The capacity d of the carrier is
assumed to be the same on all routes. This restriction is
easily relaxed, however. The basic decision variables are

v .. flow of carriers in links (i,j) per unit

1J
of time,
x ik demand for trips from i to k per unit
of time,
travel by passengers with destination k on
link (i,j) per unit of time.

The flow equations for the systems are as follows.

For passengers:

arrivals and trip generations departures

(1) I x J1,
· · k + x ·1k = Ij x ·1J,
· k i cJ k
j
(As a matter of convention we interpret all summations
as extending over the existing links in the network.)

For carriers: arrivals = departures

(2) Ij v ..
1J
Ij v ..
J1
all i.
 121

The required carrier flows to support passenger travel are

(3 ) d·v iJ· > I

k
x .. k·
~J,

Let t .. be the travel time on link (i,j). The departures

~J
v .. per hour require a stock of I v .. t.. carrier hours per
~J ij ~J ~J

hour, and this must not exceed the available stock s.

(4 ) I t .. v .. < s.
~J ~J =
ij

Let the capacity of link (i, j ) be c .. and of node i be c ~..

~J

The carrier flow is then further restricted by

(5) v .. < c .. and

~J ~J

(6) Ij v .. + v .. < c i .
~J J~

The objective function is composed of the money cost r .. of

~J

carrier transportation and of the time cost t .. of

~J

travellers. Time is converted into money by means of a

coefficient ~ denoting the money value of time. It is
assumed that the money cost of carrier transportation is
passed on to travellers so that transactions need not be
considered explicitly.

(7 ) Min I r .. v .. + a
~J ~J
I t ..
~J xij,k
ij ij,k

subject to (1 ) - (6 ) and v .. > 0, xij,k > 0.

~J
 122

This is a linear program whose solution represents the

ideal plan. The general question to be examined now is whether
individual travellers will be motivated to behave according to
this plan.

An answer may be found by considering the efficiency

conditions of this linear program. To begin with, let us ignore
the network capacity constraints and the constraints on the
stock of carriers and consider simply the best use of whatever
carriers are needed in an uncapacited network. The Lagrangian
function of this simple problem is

(8) L l r. . v.. - a ij,k

ij 1J 1J
l t.. x. . k + l A. k
1J 1J, ik 1
[lj x. . k - x ik
1J,

- lj x .. kl + L Y1' l
J1, i j
(v .. -v .. )
1J J1
+lij S .. [d·v. '-l x .. kl
1J 1J k 1J,

where Ai' Yi and Sij are

Lagrangian multipliers or dual variables or efficiency prices.

The efficiency conditions are now

(9 ) v ..
1J {;} o <=> - r .. +y.-y.+S .. d
1J 1 J 1J {:} 0

(10) x .. k
1J, {:} o <=> - a t .. + Aik - Ajk - Sij
1J {:} O.

Suppose now that there is a positive volume of travel from

every point i to every destination k. Then in condi-
tion (10) variables xij,k must be positive for every i and
k and some j.
 123

The inequality may, therefore, be written

for some j all i,k.

This may be rewritten as

(12) Aik Min {a t .. + B.. + Ajk } .

j 1J 1J

Observe that Akk = 0 since no constraint (1) occurs for

i = k. Now Aik may be interpreted as the minimum trip cost
from i to k, namely the combined money and time cost along
the shortest (most economical) route, if we interpret B.. to
1J
be the fare charged to travellers on link ij.

To verify this interpretation of B·1J. consider now

condition (9). Summing the right-hand expressions over any
closed path we have

- v .. + dB .. + y. or
1J 1J 1

since the y terms vanish

and

= if the closed path is an optimal one on

with v .. > 0 throughout. Condition (13) states that with

1J
fare B..
1J
collected from travellers on all segments where
the carrier is full, the money costs of every round trip is
just recovered while on no other closed paths will fare
 124

revenues ever exceed costs. A further efficiency condition

(14) Sij
{:}
0 <=> L xij,k
k {:} dv ..
~J

states that positive fares may be charged only on links where

carriers are full.
The analysis of the efficiency conditions shows that
travellers can indeed be induced to conform to the system
optimizing plan provided that fares are set at the proper levels
determined by equations (9), (13) and (14).

Let now the capacity constraints (4), (5) and (6) be re-
introduced. Their effect is to add efficiency prices a,~ ..
~J

and ~i respectively for the use of carriers and for the passage
through links and intersections, to the money cost r .. of a
~J
link trip. Thus r .. is replaced by
~J

(15) r .. -+ r .. + at .. +~ .. +~; + ~ .•
~J ~J ~J ~J ~ ]

These efficiency prices vanish whenever the corresponding

constraint is not binding.

3. So far, it was assumed that carriers move at a given speed.

Suppose, however, that with increasing flow, travel times also
increase

(16) t;J' = f .. (v .. )
~ ~J ~J

where f!.
~J
> o.
 125

The optimum problem becomes a nonlinear program.

Max Iij r iJ· v .. - a

~J
Iij fiJ·(viJ·)xiJ·,k

subject to (1) - (6) and v .. > 0, x .. k > O.

~J ~J,-

The efficiency conditions (9), (10) are then modified

o
(9 I ) v ..
~J
{;}
+ s ..~J
g

The behavior equations for travellers are unchanged.

Travellers continue to follow shortest paths in terms of travel
time and money cost. But the routing of carriers is affected
by the condition that a charge of

kI x ~J,
.. k' equal to the marginal cost of

the delay to travellers, be added to the money cost of

transportation. Since carriers must be compensated for this
extra cost by charges S.. sufficient to recover all costs on
~J

each circular trip, the traveller bears this cost eventually.

But the system of charges is designed to induce optimum link
frequencies v .. for carriers.
~J
 126

4. As a next step consider given demand functions for trips.

Let

be the demand function for trips. As part of the objective

function we now consider consumers' surplus defined as

where

8ik {qik) is the inverse function to (17). We obtain a convex

program [3]

(18) Max r .. v ..
~J ~J
L
ij,k
f .. (v .. )x .. k
~J ~J ~J,

subject to (l), (2), (3) and x'~ k => 0, x ~J

.. , k ~ 0, v;J'
...
~ O.

The efficiency conditions for v ij and xij,k (g') and {lO'l

are unchanged. With respect to the new decision variable x ik
we have

(19) x ik {:} 0 <=> O.

This means that demand, when positive, is based on the trip

cost Aik by shortest (cheapest) routes.

5. As a final step reconsider the problem when travellers'

preferences are described by utility functions. Let utility
depend on consumption of goods other than travel, on leisure
 127

and on trips to various destinations. Denote net income after

fares by y and leisure before travel by z. The utility
function for a representative traveller at location i is of
the form

u(Yi' z.1 - kI t.1, k X·1 k ' .•• X'1 k ••• )

where Yi' zi are income and leisure of residents at location i·,

t. k denotes the travel times between origin and destination.
1,
For simplici ty assume again that travel times t. . on links are
1J
fixed. The objective (or social welfare) function W to be
maximized consists of the sum of individual utilities (measured
in money equivalents) net of cost carriers

(21) W Ii u. - L roo v ...

1 ij 1J 1J

In order to convert trip frequences x ik into flow variables,

it is now necessary to attach a subscript for both origin and
destination to the passengers traffic on every link [4].
Let links be denoted by g,h and travellers flows by
Xi,gh,k·

The utility function assumes the form

u (y i' z. -
1
I t
gh,k gh
x. h k'
1,g ,
• •• X ik •.• )

and the welfare function is (21) as before.

The flows of passengers and carriers are related as follows
g = i
(1' ) I Xi,gh,k - Xi,hg,k
{:ik
h
g f. i
 128

(2 ') t\' Vgh - v hg 0

(3') l..\' x hk < d v gh ·

°
1,
k 1,g,
°

The Kuhn-Tucker conditions for a maximum with respect to

and are, respectively

(9") v
gh
{:}
o <=> - Ygh - Yg + Yh + dS gh
{:} 0

(10" )
Xi,gh,k {:}
o <=> - u il tgh + A - 'ih,k - 6gh {:} 0

(18') x ik
{:}
0 <=> u ik - AUk
tJ 0

where uio is the marginal utility of time.

These correspond to the previous conditions (9), (10) and (18).

Equation (10") may be rewritten

(12') Aigk = Min

h
[u
11
o t h + B h + A1 hk ]·
g g
o

This shows that travellers seek shortest paths for the distance
function

on links gh.

It is tempting to interpret Sgh again as money fare to

be charged uniformly to all travellers on links gh. But
 129

actually Sgh must be considered the marginal utility of such a

money fare.

(say) •

Unless the marginal utilities of money (or general consumption)

are the same for all travellers--the assumption that underlies
the use of consumers' surplus as a welfare criterion--the fares
p.~g h must be made to depend on the type of traveller, his
origin i and his income level. Such discriminatory charges
are, of course, highly impractical. But without these,
travellers do not have the proper incentives to choose levels
and routes--and times--of travel that are consistent with a
system's optimum.

The underlying reason is, of course, economic inequality

and the differences in the relative valuation of time and money
implied. But since transportation is a poor instrument for
achieving economic equality, we shall ignore inequality as an
economic problem in this context. In the following policy
discussion we assume, in effect, that marginal utilities of
money are equal among transport users. In other words we
shall return to consumers surplus as the appropriate welfare
criterion.

6. The policy conclusions are now straightforward. In some

ways the implementation of an optimum allocation of travellers
to routes is simpler for public transportation systems than in
the automobile case since the problem becomes part of the
bigger question of setting user charges. Tolls are not
collected additionally but are incorporated in fares.
 130

These fares should be based on the marginal cost principle.

Thus, when there is excess carrier capacity on a link, users
should not be charged at all for travel on that link.

However, charges must be sufficient so that all carrier

costs are recovered on completed round trips. This implies in
particular that on at least one link of a carrier's round trip,
traveller demand must equal carrier capacity. This is achieved
by selecting the proper frequency of departures v ..•
1J
Guidance of travellers over the optimal paths (optimal
for the system) requires that fares should be based on links
traversed and thus be made to depend on the route chosen.
Whether this is practical depends on the institutional arrange-
ments. Thus it is quite practical for long distance bus,
railroad and plane travel. For subway systems the present
methods of zonal or unit fares is not conducive to economic
choices of routes by travellers, since routes which
maximize individual utility may cause crowding to others and
thus generate negative externalities. This passing of some
of the cost to others does not achieve a system's optimum.
The problem of inefficiency in the present use of congested
subway systems both with respect to the timing and routing of
passengers is certainly a real one. While the economists'
preferred solution through differential charges appears clear
and straightforward enough, in practice it may have to be
replaced by other methods which are second best. This might
include prescription of routes to certain destination and
prohibitions of certain types of trips at certain times. But
as this analysis shows there will always be incentives for
travellers to deviate from prescribed routes and times and
beat the planners at their game.
 131

REFERENCES

[1] Charles Townes, Public Policy Committee General Motors as

quoted in Fortune, January, 1972, page 176.
[2] G.B. Dantzig, Linear Programming and Extensions, Princeton
University Press, Princeton, New Jersey, 1963, 568-91.
[3] M. Beckmann, C.B. McGuire and C.B. Winsten, Studies in the
Economics of Transportation, New Haven, Yale University
Press, 1956,

[4] M.J. Beckmann, and T. Golob, "Traveler Decisions and Traffic

Flows: A Behavioral Theory of Network Equilibrium,"
6th International Symposium on Transportation and Traffic
Theory, Sydney, Australia, 1974.
 STATISTICAL EQUILIBRIUM

Arne Jensen
The Technical University of Denmark
Lyngby, Denmark

The most powerful method in the development of traffic theories was

formulated and used by the famous A.K. Erlang who died in 1929. He
was the initiator of the theory of telephone traffic at the beginning
of this century in his using the working hypothesis statistical equili-
brium, first published in 1917. On the assumption that equilibrium
exists in tele-traffic during the busy period of the day, Erlang was
able to develop in a very precise and accurate way the total probabi-
lity distribution for the states of the traffic process under consi-
deration, and by doing this also the forecasts of the relevant service
criteria for the system for a period of time, e.g., busy signals and
waiting-time, etc.

With this methodology it was possible to respect the technical func-

tioning of the telephone exchanges down to the smallest detail, a
working method which - due to its fundamental and deep understanding
of the process in question - time has shown was a very effective tool.
An elucidation of A.K. Erlang's methods on statistical equilibrium
through the theory of stochastic processes is given by me in "The
Life and Works of A.K. Erlang" (1948).
 133

It is my impression that he in his work got inspiration from

the theories of the Brownian motions, put forward during the
last century.

It is remarkable how well telephone traffic fits into this

simple description by Markov processes. The validity of the
processes has been confirmed through extensive measurements
in Sweden, in the United States, and also through the very
detailled measurements I together with the assistance Villy
B~k Iversen of my Institute have carried out - in Denmark during
the later years. New electronic equipment has made it possible to
reveal every detail in the process. The simple fundamental
exponential law for time distances between change of states
could pass even tests developed in the theory of extreme val-
ues. Even observed showers of calls during a period of 2 weeks
was in good agreement with the assumption of telephone traf-
fic being a Poisson process. (The main report, in Danish,
"Holb~kmalingerne", I!1S0R, 1972, and special topics at the

6th International Teletraffic Congress, Munich, 1970, at the

7th International Teletraffic Congress, Stockholm, 1973 and
by Villy B~k Iversen in ERICSSON TECHNICS, Number 1, 1973).

In 1943 the late Conny Palm in his dissertation "Intensitats-

schwankungen im Fernsprechverkehr" expanded this methodology
of statistical equilibrium, which presupposed equilibrium at
every point of time, to "processes where equilibrium was only
supposed to exist at discrete points of time". He expanded
the process consisting of equilibrium intervals in the Erlang
work to a process with equilibrium points.

Erlang required for a traffic process to enter into statisti-

cal equilibrium, without saying so, that it in the infinity
point reaches a distribution independent of the initial and
tha.t the mean integral of the distribution, over time, from
the initial time to infinity was equal to this distribution.
Later on, in the 1930'ies, A. Kolmogoroff and A. Khintchine
 134

defined more precisely the ergodic process.

In many traffic situations Conny Palm's expansion of the

notion equilibrium has been very useful by opening up for the
possibility to discuss the overall picture of the process as
a series of equilibrium points connected by "bridges". Now-a-
day we talk about regeneration or recurrence points, such as
put forward, among others, by Willy Feller who worked togeth-
er with Conny Palm in the 1930'ies in Stockholm. In my work
in the 1950'ies ("INGENI~REN, Nos. 17 and 19, 1957, "Trafik-
teoriens hj~lpemidler ved p1anl~gning og drift af vejnet", p.
490),"bridgebuilding-technique" was formulated within the theory
of road traffic. I used the case waiting-time at intersections
illustrating my point by applying this technique on vehicles
waiting on a byroad to cross a major road. Let us divide the
traffic stream in blocks and anti-blocks in the following way:

By the term anti-block we understand an interval of time be-

tween 2 vehicles following each other which may be used for
crossing the main road, while a block is a section with vehi-
cles, possibly only one vehicle, which renders the crossing
of the main road impossible for the traffic on the byroad by
its critical distance L. The number of anti-blocks per time
unit is calculated by the exponential distribution as Ne- NL
when we assume that the line of traffic on the main road which
we are regarding is Poisson dist~ibuted. The law of distribu-
tion for the length of anti-blocks is the exponential distri-
bution.

The number of blocks is of course equivalent to the number of

anti-blocks, and since the blocks, in the case of the driver
on the byroad, have the effect of a quantity of traffic
1 - e -NL , the average duration,!),
1 of a block is determined
by this quantity of traffic divided by the number of blocks
per time unit, so that we get:
 135

1 1 - e -NL
o=
Ne -NL

Some investigations of American material seem to show that it

is defensible to approximate the law of distribution for the
length of the blocks by an exponential distribution with the
correct mean figure. In the following we shall assume this
approximation for the distribution of the length of the blocks.

To determine the delay caused to the traffic on the byroad,

we must realize that the probability of e.g. more than two
vehicles waiting to cross the main road depends on time. It
is declining during the anti-block situation, and rising dur-
ing the block situation, since, in the former case, there may
be both passage across the main road and addition to the num-
ber of vehicles, while, in the latter case, there can only be
the question of addition to the number of vehicles increasing
the queue waiting to cross the mainroad. On the basis of the
assumption that the length of the blocks and the length of the
anti-blocks are independent of one another, we may use Erlang's
working hypothesis, stastical equilibrium, to determine the
law of distribution for the number of waiting vehicles at the
time when a block situation has just ceased (po(s»; likewise
to determine the law of distribution for the number of waiting
vehicles on the byroad at the time when the block has just
commenced, (p (s-~».

As illustrated in Fig. 1, attached, the law of distribution

for the number of waiting vehicles at the end of a block situa-
tion is determined by assuming that the law of distribution for
the number of waiting vehicles at the commencement of the block
interval is known, and then calculate the distribution we are
looking for by the exploitation of our knowledge that the in-
crease of the length of the queue derives from a Poisson dis-
tributed addition to the traffic, and that the law of distribu-
tion Pb(t) for the duration of the length of the block is known.
 136

It may be expressed in the following relationship:

p(s) =

For Pb(t)=e -bt bdt we get

It is somewhat more difficult to determine the relationship

between po(s) and Pl(s) by counting from Po to Pl merging the
two quantities into one, while the traffic pattern may be ex-
posed to both additions to the traffic on the byroad and de-
partures from the traffic on the byroad leaving the queue to
cross the main road. By applying the aids which the author
has explained in his work in 1954, it is nevertheless rela-
tively easy when regarded from the angle of calculation tech-
nique; I will revert to this matter later on in this paper.

The procedure here given rests upon' the hypothesis of points

of equilibrium and statistical equilibrium, and it employs
relatively simple bridge building executed between the points
that mark the beginning and the end of the anti-block and the
end of the following block. Thus we arrive at the following
results:

The probability D that there will be a delay in crossing is as

follows:

D = l_e- NL +!!
II

in which 1/11 is the average time which a vehicle on the byroad

wishes to occupy the main road for the purpose of crossing,
before another vehicle on the byroad will be able to take ad-
vantage of a possibly free main road.
 137

The average length of queue will be

s = ~(l-Drl
1I

Since the volume of the traffic is n, the mean waiting time

at the main road becomes

1
m = -(l-Dr1
1I

It is to be noted that the results arrived at in the two

border line cases, in which either the main road traffic is
0, or the byroad traffic is 0, lead to results that are in

agreement with the results from a Poisson distributed traf-

. fic. If we examine the formula evolved by Raff, that which
has been in use hitherto in practice

2,Sn 2NL
D = -3600
l_~e__ ~~~ e-____
_ 2,Sn -NL
l-e 3600 (l-e )

we find that while in relation to the byroad traffic equal

to 0 it leads to the classical results, it does not lead to
the well known results from the queuing theory in relation to
the main road traffic when the latter is equal to o.

In Fig. 2, attached, we see a comparison between the results

arrived at here and those. of Raff's on the assumption that
the crossing of the main road requires an average of 1,8 sec-
onds, and with a safe distance L of 4,6 seconds. Since the
individual driver has his safe distance, it will be most
correct just as in the case of the pedestrians, to assume a
law of distribution for the safe distance L. Observations in
American material tend to show that Erlang's distribution for
f=4 may be used. Through this we get the following results for
the delay D:
 138

while the expression for the waiting time remains unchanged.

Uncontrolled Crossroad; a few Horking Hypotheses.

One must bear in mind that on a main road the traffic is not
always moving so freely that it may be regarded as a Poisson
distributed traffic, but one must then pay a little attention
to how these results would look in a traffic which is to some
extent equalized by a narrow road grid.

It is to be noted that the time intervals between vehicles

following one another are not exponentially distributed in
an over-crowded road grid. The law distribution for the dura-
tion of an anti-block is in that case normally exponentially
distributed, as the safe distance in general will be greater
than the sphere affected by delays.

Thus, it has been ascertained that the bridge building de-

scribed here may be generally applied with respect to block
and anti-block, taking its departure in Conny Palm's notion
equilibrium points: It is a very central device when it comes
to the characteristics of the traffic stream.

The above,by making use of the matrix notation,may (in a more

"compressed" form,e.g.,Arne Jensen, "A Distribution Hodel,
1954) - be generally expressed in the following way:

In the anti-block period a "birth and death" Markov process

on the byroad with the intensity matrix AOl : {aij}. In the
block on the byroad with the intensity matrix Alo : {b ij }.

The distribution for the anti-block period t being Pa(t)dt :

e-NLNdt.
 139

The distribution for the block period t being Pb(t)dt which

. 1 .
approx~mat y ~s e
-bt b d t.

aij = ai,i+l = n
ai,i-l = \1
ai,s = 0 s+i-l, i, i+1

r.
a·~,~. = -n-\1

b ij = ~,~+
=n
a. 1 = 0 s+ i, i+1
~,s

a ~,~
.• = -n

The traffic intensity on the main road being N, on the byroad

n.

1
The mean period for anti-block mol = N

1 1_e- NL
The mean period for block m10 =b =
Ne- NL

Let the equilibrium distribution for the number of waiting

cars s on the byroad the beginning of the anti-bJock interval
be notated by the vector Po = {po(s)}.

Let the equilibrium distribution for the number s of waiting

cars at the byroad at the beginning of the block interval be
the vector P1 = {P1(s)}.

We get the following equation systems for the determination

Po and P1 (ref., e.g., Arne Jensen, 1954).
 140

or

Po . ..Ioe A01
t -NT
e Ndt = PI

.. Alot -bt
PI Ioe e bdt = P0

that is

or

The distribution P t at a given lag of time after the begin-

0,
ning of the anti-block will be

and PI , t in the block period

A pioneering work on road traffic from this year, 1974

"The multi-lane traffic flow process. An evaluation of
queueing and lane-changing patterns, based on a Markov model"
by Jens R¢rbech has after detailled measurements at three cuts
on higway traffic by electronic equipment shown that it will
also be' possible, as far as road traffic is concerned, to
apply Markov processes as an adequate tool for the description
of vehicles changing from lane to lane. This opens up for
 141

introduction of knowledge from the behavioral sciences with

respect to the driver's behaviour direct into the matrix in-
tensities for lanechanging from one state to another. The above
theory could be expanded to cover this point of view.

Conditions for a process to enter into statistical equilibrium.

In my work "A Distribution Model" I tried to throw light on the
conditions for a process to enter into statistical equilibrium
and reached the results shown below, using quasi-processes as
point of departure. In the quasi-process the probability mass
is less than or equal to one. By expanding the process to cov-
er the process in question, you are able to find the conditions
for the probability mass = 1 for the wanted process.

In order for the process to enter into statistical equilibrium,

it is a necessary and sufficient condition that there exists
a solution Vo = {V o,s }

Vo,s >os = 0,1, ••• ,

IV <co
so,s

and sI(-as,s )Vo,s <co

A = {a •. }
l.J

The equilibrium distribution {pes)} will then be determined

as
v
pes) = ~ s = 0,1, •••
IV
so,s
 142

These may be summarized in the following terms: If you are

able to find the equilibrium distribution and the mean number
of changes from state to state in a given time interval is
limited, the working hypotheses statistical equilibrium is
valid. That means, that you will have problems when the pro-
cess becomes exaggerated during rapid changes. As a consequenee,
it might disappear from the space in question in final time.
The result is that the probability- mass might diminish to
measure less than one, creating risks of different sorts, such
as Pascal's case of ruin.

For this reason you also find contributions, which putlight on

equilibrium, among the works emanating from other schools
dealing with theory of risks, and especially from the famous
Cramer Stockholm school. Let us mention, besides W. Feller,
O. Lundberg, N. Arley, and V. Borchsenius.In the work by R.
von Mises in 1931, you will find instructive contributions
trying to classify processes. From each original description
of states and their connection by jumps he was able to find
out if the final contribution depends on, or does not depend
on, the initial state of the process. Also for the development
of traffic equilibrium, it is important to look at the famous
work by Volterra about the two competitive species in the
Adriatic Ocean, which work at a later date was expanded by,
e.g., R.A. Fisher, to a stochastic, but incomplete picture of
this fight - incomplete, because it overlooked the risk of
extinction of the species during the fight.

When I mention this, it is for the reason that the abstraction

behind the formulation of the mathematical models seems at a
too early stage to use the assumption equilibrium without being
sure that no part of the probability space disappears in ruin
cases. You might still have a solution - and a correct condi-
tional solution - for the rest of the space, but if you are
faced with ruin cases, it is very important for the application
to be aware of it in due time - ecology and world models not
to be forgotten.
 143

A large portion of the literature has been concentrated on

that part of the processes which we call Markov processes, and
the like, and a large portion of reality, fortunately, may be
covered by this.

As I mentioned earlier, Conny Palm expanded this group of

processes by notion equilibrium points. If you want even more
general results you have to look for others, such as F.Pollaczek
and A. Khintschine, from the old school, using the theory of
function as primary tool. But I must admit that many of these
works are not easily accessible and are not particularly meant
for the illustration of equilibrium notion.
 144

BIBLIOGRAPHY.

ARLEY, N. & BORCHSENIUS, V". (1945): On the Theory of Infinite

Systems of Differential Equations and their Application to
the Theory of Stochastic Processes and the Perturbation
Theory of Quantum Mechanics. Acta Mathematica, 76.
BROCKI1EYER, E., HALSTR0M, H.L. & JENSEN, Arne (1948): The Life
and Works of A.K.Erlang. Transactions of the Danish Academy
of Technical Sciences, No.2, Copenhagen.
FELLER, W. (1949): Fluctuation Theory of Recurrent Events. Trans.
Amer. Math. Soc. 67, p. 98.
IMSOR, KTAS and LME (1972): "Holbeek-malingerne". Indsamling og
statistisk analyse af data fra en transitcentral".
IVERSEN, Villy Beek (1973): Analyses of Real Teletra=fic Processes
Based on Computerized Measurem~nts. ERICSSON Technics.
JENSEN, Arne (1957): "Trafikteoriens hjeelpemidler ved planleegning
og drift af vejnet". INGENI0REN Nos. 17 and 19, 1957.
("Traffic Theory as an Aid in the Planning and Operation of
the Road Grid" - offprint INGENI0REN - Int'l Edition No.2,
vol. 1).
JENSEN, Arne (1954): "A Distribution Model" - dissertation.
KOLMOGOROFF, A. (1933): Grundbegriffe der Wahrscheinlichkeits-
rechnung. Ergebnisse der Mathematik und ihrer Grenzgebiete.
LUNDBERG, O. (1940): On Random Processes and their Application
to Sickness and Accident Statistics. Uppsala. Dissertation.
MISES, R. VON (1931): Wahrscheinlichkeitsrechnung und ihre An-
wendung in der Statistik und teoretischen Physik. 1. Band.
PALM, Conny (1943): "Intensitatsschwankungen im Fernsprech-
verkehr". ERICSSON Technics, 44. Dissertation.
R0RBECH, Jens (1974): "The multi-lane Traffic Flow Process. An
Evaluation of Queueing and Lane-Changing Patterns, based on a
Markov Model". Dissertation.
 145

Figure 1

Bridge building

" --
~y:: :'>i

Block Anti-block

o Equilibrium point

Fig. 1. Bridge building is undertaken between points of

equilibrium by exploitation of knowledge of the law of
distribution for the number ~ of vehicles arriving.
 146

Figure 2

" (per hour)

·•·•,
"DO .,,
,
, ,,
,,
,, 0) RAFF:
.,,.
\

..,
.,
1) A.J.:
aoo \
\

...
\

. A.J.:
\

fOOD
2)
~
II.\ ..\
~ .. L = li,6 sec.
\

\\ \ . 1
,.,..... = 1,8 sec.
\

'. .. 0, = O,7J' p

, .
,,..,,.
"~,

"" '''-- ...

o N per "hour)

Fig. 2. The probability that a driver in the byroad traffic

will be delayed, Db' depends on the volume N of the main road
traffic, the volume n of the byroad traffic, the safe distance
L, and the time required to cross the main road lip. In the
illustration, Raff's descriptive mathematical formula is com-
pared with the theory submitted in this article on the one
hand with a fixed safe distance A.J. 1, on the other hand with
varying safe distance A.J. 2.
 TRANSPORTATION MODEL WITH ELASTIC CONSTRAINTS

R. Hamerslag
Dwars, Heederik en Verhey B.V.
ABSTRACT

The study describes static and dynamic versions of a transportation model with
elastic constraints.

The model was developed on the basis of a consumer behaviour theory and on the
assumption that there is an equilibrium between supply and demand of employment and
between demand and employment of working population in every traffic zone. In order
to provide the elastic constraints, extra equations were added to those used in tra-
ditional distribution models. One of the characteristics of the model is that it
generalized the transportation models with none, one, or two constraints.

A number of calculations which were carried out to investigate the sensitivity

of the model showed that traffic and population distribution for employment is depen-
dent on the assumptions used for the constraints in the static version of the model.
In the dynamic model the traffic flows are iess sensitive to assumptions on elasticity.

The deterrence function is determined by mode. The value of the function depends
on trip time and trip costs, the effort in the sensitivity, analysis applied to the
dynamic model the travel effort values range from high to low. The traffic flow and
the geographical distribution of population and employment proved to be dependent to
a great extent on the value of the travel effort.

The growth of population and employment development was rather realistic. In

spite of the relatively simple assumptions which has been used, it proved possible
to describe the growth of towns, and of commuter saldi, and the extent of suburban
growth and in the exploding towns.

The geographical location in which these developments occur is strongly influen-

ced by the infrastructure.
 A UNIFIED APPROACH TO EQUILIBRIUM METHODS
FOR TRAFFIC ASSIGNMENT

Sang Nguyen
Universite de Montreal
Montreal, Quebec, Canada

ABSTRACT

In this paper we address the methodology of the equilibrium

traffic assignment on congested networks. The paper presents a
unified approach to a class of iterative solution procedures
within a general mathematical programming framework. A prelimi-
nary comparative analysis of three basic algorithms for the equi-
librium traffic assignment is reported.
 149

INTRODUCTION

In the context of transportation planning, traffic assign-

ment is the process of distributing given interzonal trip demands to

alternate routes between these zones. This paper is concerned with

the equilibri~ model of the traffic assignment problem. In this

model, it is assumed that the travel time, or the generalized travel

cost, along each link of the network is a known monotone increasing

function of the link flow. It is further assumed, that, there is

only one class of travelers, and each individual travels along his

minimum cost route. A traffic pattern, which satisfies these assump-

tions, corresponds to an equilibrium state, in which, the travel costs

on all the utilized routes between any given origin-to-destination (OD)

pair are equal, and less than those on the nonutilized routes (Wardrop's

first principle [28]). In addition, should the trip demand be elastic,

and given by a monotone decreasing function of the associated inter-

zonal travel cost, one may obtain the classic economic supply-demand

equilibrium, where the interzonal travel cost implied by the demand

function is equal to the actual travel cost of the utilized routes.

It has been shown, that the problem of computing equilibrium

flows may be formulated as a convex cost minimization problem on a network.

The latter formulation has attracted in recent years the attention of many

researchers, [ I, 2, 3, 5, 12, 13, 20, 21, 30 ] among others.

 150

Solution procedures for the traffic assignment problem may

be divided roughly into two classes, the capacity restraint methods

[17,18,19, 22, 24, 27], and the equilibrium methods [2,5,11, 15,

20, 30, 35J. The basic idea on all these methods is to obtain an equili-

brium traffic pattern, from an initial trial solution, by iteratively

adjusting the travel costs and the traffic flows. However, the capa-

city restraint methods are generally based on intuitive arguments,

while the equilibrium methods are based on rigorous mathematical argu-

ments which ensure the convergence of the calculations to the solution

of the convex cost minimization problem. For a detailed review of the

existing methods, the reader is referred to a recent survey by Ruiter [34].

Although the advantages of equilibrium methods over the capacity

restraint methods have been generally recognized, there does not exist

any operational equilibrium method at the present stage. There is a need

for studying the efficiency of the most promising methods, and for expe-

rimenting these methods with real-life traffic assignment problems. This

paper presents a unified approach to a class of iterative equilibrium

methods based on the general approach of feasible-direction methods in

nonlinear programming. This approach not only sheds some new insight

into the main existing equilibrium methods, by bringing out formally their

common structure, but also allows to construct new efficient algorithms

both in the fixed and elastic demand cases. A brief comparative

study that examines computer coding considerations and presents computa-

tional tests complete the report.

 151

NOTATION AND PROBLEM FORMULATIONS

Consider a transportation network consisting of n nodes and

m oriented links (arcs). A node represents either a zone or an intersection

of roads. A two-way traffic road link is represented by two oriented

links of opposite directions. For convenience. it is assumed that nodes

are numbered from 1 to n and links from 1 to m the first

1.2 •...• n o nodes are the origins. A path (route) connecting an origin-

destination pair ij is defined as a sequence of links. all oriented

in the same direction. and free of loops. Let denote the inter-

zonal trip demand between origin i and destination j • h~ the flow

on the path ~ • and the flow on the link a coming from origin s .

The total flow on link a is therefore defined as

n
0
(1) f
a L fS
a
s=l

or as

(2) f
a L L 0a~ h~
ij ~£Qij

where denotes the set of paths connecting the OD pair ij , and

0a~ equals 1 if link a belongs to path ~,and 0 otherwise. On

the network the relationships between the different flow variables are

expressed by the flow conservation equations:

(3) (all ij)

 152

(4) o,
or

-gsi if i is a destination node,

(5) L fS
a L fS
a
a£Wi a£V i
0 otherwise

(i ~ s i 1,2, ... ,n s = 1,2, ... ,n )

o

(6) (a 1,2, ... ,m s

(7) o (all ij ) ,

where Wi and Vi are respectively the sets of arcs beginning at i,

and ending at i.

Let ca denote the link travel cost on a link a, and v ij the inter-

zonal travel cost for OD pair ij , then the relationships between the

travel costs and the flows are given as follows:

(8) c
a

where the function Ca is assumed monotone increasing,

(9)

where the demand function is assumed monotone decreasing, and

is the inverse function of

 153

The convex cost minimization problem, which can be viewed as a

minimum convex cost mu1ticommodity flow problem, may be defined

using, either the arc-chain formulation, or the equivalent node-

arc formulation, as in the following:

minimize

(10) Z Z (f)
a a L
a ij

subject to:

a) arc-chain constraints (3)-(4);

or

b) node-arc constraints (5)-(6)-(7)

where f is defined in (1), or in (2), and

a

f
(11) Z (f)
a a r
0
C
a
(x) dx ,

gij
(12) Zij (gij) f Gij (x) dx •
0

In the fixed-demand case the objective function (10) reduces to:

(13) Z Z (f) .
a a
a

Note that the convexity of (10) is implied by Ca monotone increasing,

and Gij monotone decreasing. It should be noted, that, in the case of

strict convexity, the uniqueness of the solution to problem (10) is

only with respect to the total link flows f and to the demands
a
while the path flows hR,' and the pattial link flows fS , are not
a
necessarily unique.
 A GENERAL SOLUTION FRAMEWORK

We proceed now to develop a general solution procedure to

the convex cost multicommodity flow problem. For convenience, we

rewrite problem (10) in the standard programming form:

(14) min F (x) subject to AX - b =0 , x ~ 0 .

where F (x) is convex. Specially relevant to the linearly constrained

convex programming problem (14) are primal methods in nonlinear programming.

In particular, methods that cast into the general framework of the well

known feasible-direction method originally proposed by Zoutendijk [33].

The method consists of generating, from a given starting feasible solu-

tion, a sequence of feasible solutions, each of these improves the objective

function, and is obtained from the previous one by determining, a direction

that leads to an improved solution, and the optimal step length to be

taken in that direction. A precise statement of a feasible-direction

method, as applied to the linearly constrained convex program (14),

is as follows:

A feasible-direction method
Step 1. Given a feasible xl to (14). Set q = 1.
Step 2. Determine a direction dq ~ 0 which satisfies

(15) o

(16) d~
~
~ o if o
 155

and

(17) < o

where V denotes gradient. If no such rlirection exists, terminate.

Step 3. Determine the step length tq that solves:

min F (x q + t d q )
tEL

where L { t I x q + t dq ~ o}

q+1
Step 4. Compute x set q q +1
and return to Step 2.

Note that (15)-(17) imply that there exists a scalar e > 0 , such that

for any 0 < t S e

b o
and

This general scheme becomes a distinct algorithm whenever a precise direction

generation procedure is chosen for Step 2. To guarantee the convergence

of the method, the direction generation procedure that may be implemented

in Step 2 must possess the property that equations (15)-(17) hold in the

limit, q + 00. Under this assumption, the method, either terminates at

an optimal solution, or generates an infinite sequence {xq }, such

that the limit of any convergen~ subsequence of { x q } is an optimal

solution.
 156

In the following, we shall describe several appropriate

direction generation procedures that satisfy the convergence require-

ment.

A direction obtained by a linear approximation to the objecti~ function.

This direction is suggested in the Frank-Wolfe (FW) [9]

algorithm, originally proposed for a quadratic programming problem.

Given a nonoptimal x q , let yq be a feasible solution that minimizes

the linear approximation to F at xq

yq - x q , it is clear that dq satisfies (15)-(17). Hence,

the direction generation procedure consists essentially in solving the

following linear program:

(18) minimize V F (x q ) , y

subject to

Ay b o
y ~ 0

Should xq be an optimal solution to (14), one obtains V F (x q ) , (yq-xq) o.

Directions obtained by a simplex-like approach.

Given a feasible x q , let Bq be a basis of the constraint

matrix and the basic and nonbasic components of

Taking into consideration this partition of variables, we may rewrite

(15)-(17), after several simple manipulations, as:

 157

q
(19) dq
i -L
je:J
a ij dq
j
i e: I

(20) dq
i
~ 0 if x q
i
o, i e: I

(21) yq
j
dq
j
s o, j e: J

(22) dq
j
~ o, if xq
j
o, j e: J

where are elements of the simplex tableau Aq

and the relative cost factors y~ are given by

J

(23) L j e: J.
i

To generate a direction here, it suffices to choose the nonbasic components

dj , such that conditions (21)-(22) are satisfied. The basic components

dq are then determined by equations (19), whereas conditions (20) are

i
satisfied, either by invoking the nondegeneracy assumption, or by imposing

a partition of variables that ensures the nonnegativity of all

The choice of the nonbasic components is clearly dictated, in this

approach, by the values of the relative cost factors y~ Basically,

J
whenever the cost factor y~ is negative, the nonbasic component
J
is given a positive value, and vice-versa, in order to obtain a negative

product. When y~ is positive, the value given to dq must be propor-

J j
tional to the corresponding nonbasic variable in order to satisfy

(22). One also has the choice, in constructing the direction d q , to

modify, several nonbasic variables, or only one, at a single iteration.

 158

We shall now describe two closely related directions, that

are based upon the above scheme. The first is a modified version of

the reduced-gradient (MRG) direction [31], and the second is taken

from the convex-simplex (CS) method [32]. These directions are defined

by the following choices:

MRG direction -yq if y~ < 0

j J
dq
J
-y~ x jq otherwise
J

CS direction

Let yq min { y~ } and yq xq max y~ x~ } ,

R, J k k J J

then

a) if yq yq x q 0 dq 0
R, k k

b) if yq <: yq x q d R,q 1 d~ 0 for all j ;t R,

R, k k J

c) otherwise, dq -1 dq 0 , for all j ;t k .

k j

In the following we shall describe three basic algorithms for ca1cu1a-

tion of the equilibrium traffic flow, that result from the application

of the above developed feasible directions, to solve the mu1ticommodity

flow problem (10).

Most of the times, results will be stated informally, without

the supporting justifications. For a rigorous and detailed development

of the algorithms, the reader is referred to [36].

 159

ALGORITHMS EMPLOYING CALCULATION OF SHORTEST ROUTES BETWEEN ZONES.

In this section, we describe the application of the

feasible-direction method, using successively the FW direction

and MRG direction, to the arc-chain formulation (lO-a). At first

view, it seems that the a priori specification of all the routes

between each 00 pair, required by the arc-chain formulation, makes

any solution procedure based on this formulation untractable. How-

ever, we will show that the a priori generation of all the routes may

be replaced advantageously by a repetitive generation of shortest

routes between 00 pairs.

Algorithm Al an application of the FW direction.

Consider the arc-chain formulation (lO-a), in the fixed-

demand case. To determine the FW direction d q , at iteration q ,

one has to solve a linear program, subject to the same constraints

than (lO-a), and where the cost coefficients of the objective function

are the partial derivatives of (13) with respect to the path flows

evaluated at the current flow values. By noting that the partial

derivatives of (13) with respect to the path flows are merely the travel

costs of the paths, this linear program may be interpreted as a mathe-

matical formulation of the all or nothing traffic assignment. There-

fore, it may be solved efficiently with any existing shortest route

algorithm. The determination of the step length t q , which follows,

 160

does not present any particular problem. It can be done with any

one-dimensional search procedure. The main appeal in this approach

arises from the fact that, though, the arc-chain formulation is used,

the path flows are not involved in the different computational steps,

and the explicit knowledge of the paths is not necessary. In other

words, only the link flows f have to be kept along the solving
a
process, therefore, reducing considerably the memory requirement in

the implementation of this type of algorithm.

We give, in the following, the main lines of the resulting

algorithm.

Algorithm Al

Step 1. Determine initial link flows f }.

a

Step 2. Determine an all or nothing traffic pattern

{W }, using the link costs c

a a

Step 3. If ~ £ (£ 2: 0 predeter-

mined tolerance), terminate.

Step 4. Compute the step length t* that minimizes

I Za (fa + t (Wa '- fa» ; 0 ~ t ~ 1

a

(Za (.) is defined in (11».

Step 5. Update f and

a

c C (f) Return to Step 2.

a a a
 161

By using the convexity of the objective function (13), it is easily

proven that (L
Z (f) - E) is a lower bound on the value of the
a a
a
objective function upon completion of algorithm Al •

An algorithm similar to Al has been first proposed by Bruynooghe,

Gibert and Sakarovitch [2]. However, the development of their algo-

rithm is based on an intuitive approach, rather than on a matpematical

programming framework. Recently, Leblanc [37] has rediscovered algo-

rithm Al by applying the Frank-Wolfe algorithm to a node-arc

formulation of the equilibrium model.

In the elastic demand case, the development of an elastic

demand algorithm,within the feasible-direction framework is very

similar to the foregoing discussions. The resulting algorithm differs

from algorithm Al in the following way: for a given OD pair, if the

interzonal travel cost corresponding to the current level of demand is

less than the current minimum cost route, the application of the FW

direction produces a decrease in the demand, and the route flows; in

the opposite case, it produces an increase in the demand, and the new

demand is added to the current shortest route. It should be noted that

the adjustment of the demands and the adjustment of the link flows are

done simultaneously. This is a new algorithm, though, slightly similar

to an elastic demand algorithm proposed by Gibert [11]. For a precise

statement of the algorithm see [36].

 162

Algorithm A2 a sequential approach - an application of the MRG

direction.

Since the interaction between the traffic associated with

different OD pairs, is only explicitly expressed in the objective

function of problem (10), it may be advantageous to decompose the

original problem into a sequence of smaller subproblems, consisting

each one of a single aD pair. Basically, the sequential approach

consists in determining separately in turn an equilibrium traffic

pattern for each OD pair, until an equilibrium traffic pattern for

the entire network is obtained. This approach has been suggested

by Dafermos and Sparrow [ 5] •

We shall now discuss the application of the feasible-direction

framework, using the MRG direction, to solve a typical subproblem asso-

ciated with a given aD pair. We shall first consider the fixed-demand

case. Consider the subproblem associated with the aD pair ij The

objective function is given by (13), and the constraint set of this problem

consists of one flow conservation equation (3), and the nonnegativity cons-

traints (4). The flows associated with all other OD pairs are kept invari-

able. Assume that at the beginning of a typical iteration all the route

and the route costs uR, I Ca ,have been updated.

a£ path R,

Let b specify a minimum cost route. Assume now that ~ has been chosen

to be the basic variable.

 163

In this case, the MRG direction is given by:

(24) for all t ~ b

(25)

Note here that db ~ 0 , since d t SO, for all t ~ b. Therefore

the application of this direction produces a shift of traffic from

all the utilized routes, onto the minimum cost route. The amount of

traffic to be moved is determined by the step length t*. The calcu-

lation of t* can be done, as before, with a one-dimensional search

procedure.

The arc-chain formulation requires in principle a complete

enumeration of all possible routes between the OD pair. Since the

number of routes increases exponentially with the number of nodes in

the network, the storage of these routes becomes usually prohibitive

for real size problems. However, as noted previously,

it is possible to circumvent this difficulty by applying a column

generation technique. Basically, this strategy enables utilized routes

to be generated progressively during the solution process. This strategy

has been proposed recently by Leventhal et al [15].

We shall now state the computational steps related to the

determination of an equilibrium traffic for one OD pair. The complete

algorithm consists essentially in cyclic iterative applications of those

steps, until equilibrium is reached for all OD pairs.

 164

Algorithm A2

(Computational steps for the subproblem associated with

OD pair ij)

Step 1. Given the current set of generated routes

between OD pair ij (at the beginning Qij may contain

only one route), and the route flows hi' i £ Qij .

Step 2. Compute the route costs ui ' i £ Qij , and

determine the minimum cost route b.

Step 3. Determine the MRG direction according to (24)-(25).

If db S £ (predetermined tolerance), go to Step 6, otherwise

continue to Step 4.

Step 4. Determine the step length t*, that minimizes

L Z
a
(f
a
+ t (wa fa» ; o S t & min {-ht / d t I d t < O}
a

where W
a L Qat dt (Qat is defined in (2».
i

Step 5. Update the route flows h t = h t + t* d t ' t E Qij ,

the link flows fa ' and the link costs c C (f)

a a a
Return to Step 2.

Step 6. Generate a minimum cost route, say route k I f the

route cost is less than the cost of the current utilized

routes t £ Qij , then add route k to the set and

return to Step 2; otherwise, terminate.

 165

With this algorit~,the user has a more refined control on the

accuracy of the equilibrium traffic pattern, via equation (24),

than with algorithm AI. Indeed, in AI' the accuracy of the final

solution is given by a lower bound on the value of the objective

function, while in A2 ' the accuracy is directly given in terms

of the differences of route costs.

Should the CS direction be used here instead of the MRG direction,

one may obtain a similar algorithm to the one proposed by Dafermos

and Sparrow [5] for linear travel costs. In this case, the adjust-

ment of flow at each iteration is a transfer of traffic from the

highest cost utilized route onto the least cost route.

Extension to the elastic demand case is straightforward,

and does not present any particular problem. The adjustment of the

demand is done in the same manner as in the previous elastic demand

algorithm (using the Frank-Wolfe direction). However, the fraction

of the demand to be increased, or to be decreased, is different.

A precise statement of this new elastic demand algorithm is given

in [36].

A SEQUENTIAL APPROACH USING THE NODE-ARC FORMULATION

Until now, we have introduced algorithms in which the calcu-

lation of the shortest routes plays a predominant part. Although, shortest

path algorithms have been made very efficient computationally, the calcula-
 100

tion of shortest routes for large networks, a large number of times,

may become time consuming. However, the shortest route calculation

is not inherent to the equilibrium traffic assignment model, as it

is generally conceived, and the following solution approach will

clearly demonstrate this assertion.

Consider the node-arc formulation (lO-b) of the min-convex-

cost multicommodity flow problem. As noted earlier, the absence of

explicit capacity constraints, allows a decomposition of the original

problem into a sequence of smaller subproblems; each one of these

consists of one origin and all destinations. We may then single out

a subproblem associated with a given origin, and proceed to develop

a solution procedure for this problem using the convex-simplex

direction.

We examine first the fixed-demand case. Consider a sub-

problem associated with a particular origin, say origin s This

subproblem is obtained from (lO-b) by fixing all variables, but the fS , at

a
their current values. It is well known that a basis of the node-arc

incidence matrix corresponds on the network to a spanning tree. Let

Y = B-lA denote a simplex tableau, where B is a basis of the node-

arc incidence matrix A. It is also well known, that a column Yt '

where is a nonbasic variable, corresponds on the network to a

unique cycle formed by the out-of-tree arc t and the in-tree arcs

defined by the nonzero elements of Yt More specifically, the one-

 167

to-one correspondence is as follows:

+ 1 then the in-tree arc i is oriented in the

same direction as R. on the cycle

YiR.
- 1 then the in-tree arc i is oriented in the

opposite direction to R, on the cycle.

Let ER. denote the cycle associated with the out-of-tree arc R..

We denote by the set of arcs of the cycle which are oriented in

the same direction as R" and by E~ the remaining arcs of the cycle.

In this case, the relative cost factor YR, (expression ~3» may be

interpreted as the cost of the cycle ER,' and may be computed as

follows:

c c
a a

where c is the cost associated with link a (8). Therefore, the

a
calculation of the convex-simplex direction can be achieved directly

on the network by exploiting the spanning tree structure of the basis.

The efficiency of this approach is greatly enhanced by the fact that,

in the application of the feasible-direction method, it is possible to

restrict oneself to those "directed" spanning trees rooted at origin s

where all in-tree arcs are oriented along the paths connecting s to

all other nodes (arborences with root s). This key result has been

proven in [3~. With this property, the cost of a cycle may be inter-

preted as the difference of the cost of two alternate paths connecting

the origin s to a certain given node. Let Pi be the path connecting

 168

the origin s to node i consisting only of in-tree arcs. For a

given "directed" spanning tree rooted at s , Pi is uniquely defined.

Let u(i) denote the cost of path Pi ' u(i) = L c

a (in the programming
a£Pi

terminology, the u(i) are the dual variables).

Therefore, the cost Yt of the cycle Et is alternatively computed

as follows:

+ u(i) u(j)

/
where t = (i,j) . We are now able to compute all the cycle costs Yt ,

without having to trace back any cycle, by using the n-l (n=number of

nodes) path costs u(i) , i ~ s , i = 1,2, •.. ,n. Only one cycle is

traced back to provide the links on which the traffic flow must be

adjusted. It should also be noted that, with a "directed" spanning tree,

each cycle is made up by two alternate paths connecting a given pair of

nodes (not necessarily an OD pair). Therefore, the transfer of flow

is from the high cost path to the low cost path.

We shall summarize the foregoing discussions in the following:

Algorithm A3

(Computational steps for the subproblem associated with

origin s)

Step 1. Given the current traffic flows fS a = 1,2, ... ,m,

a
and the current set of in-tree links T(s) •
 169

Step 2. For each node i ~ s , compute the cost u(i)

of the path Pi ~ T(s) , that connects the origin s

to node i , u(i) =~ c
a
• Set u(s) =0 •
aEPi

Step 3. For each out-of-tree link a, compute the cycle

cost Ya = c a + u(i) - u(j) ,where i and j , are the

initial and terminal nodes of link a.

Step 4. Determine

Yt = min { Ya } and Yk f~ = max { Ya f: } •

If Iytl, Yk f~ S E (predeterminedto1erance), terminate.

If set j = t, otherwise set j =k •

Step 5. Trace back the cycle Ej Let L1 and L2 be the

two paths that make up the cycle Ej Assume that the transfer

of flow is from L1 to L2 • Determine

If e > 0 go to Step 6. Revise the set of in-tree links

T(s) by interchanging the in-tree link r with an out-of-

tree link, chosen among the set of links incident at the end

node of link r. Return to Step 2.

Step 6. Determine the step length t* that minimizes

Z (f + t)
a a
oS t s 9 .
 170

Step 7. Update the flows and the costs on the cycle Ej

fS fS - t* a f.:
a a Ll

fS fS + t* a £ L2
a a

f
a
fS +
a i;tsL fi
a
a £ Ej

c C (f ) a £ Ej
a a a

Return to Step 2.

To solve the original multicommodity flow problem (la-b), it suffices

to solve each subproblem in turn with the above computational procedure,

until a global equilibrium traffic pattern is reached. Note that the

accuracy of the final equilibrium traffic is also given in terms of

route costs.

In the elastic demand case, the subproblem associated with

origin s may be considered as a minimal cost circulation problem by

considering that the variable demands gsj are the flows on the

fictitious links (j,s) with negative costs c js

Therefore, a generalized version of algorithm A3 can be used to

solve the elastic demand traffic assignment.

 171

CODING CONSIDERATIONS AND COMPUTATIONAL RESULTS

The algorithms AI' A2 , and A3 were coded in Fortran and

tested using the FTN compiler on a CDC CYBER 74 with a maximum memory

of 130,0008 words of 60 bits.

In this section, we examine the memory requirements of the

codes, and analyse some of the test results obtained.

The memory required by the codes may be divided roughly into

storage zones associated with, the network geometry, the assignment data,

the cost coefficients, and finally the working block. The first three

divisions are similar for all the codes, while the working memory space

is specific to each code.

The network is represented in memory by two vectors of lengths

m (number of links), where the ith elements contain the origin node

and the end node of link i Links are presorted in increasing order

of origin node, and the address of the first link leading out of node

i is stored in the ith element of a third vector of length n (number

of nodes). The assignment data includes the origin nodes, the destination

nodes, and the OD demands. However, in AZ and A3 ' as the path flows,

and the partial link flows related to each origin, must be stored in

memory, the OD demands are only used for the generation of the initial

traffic pattern, and may be discarded thereafter. The size of the memory

zone allocated to the storage of the cost coefficients depends, of course,

 172

on the cost functions used. Usually travel time functions require at

least two parameters: the free flow travel time and the link capacity.

Concerning the size of the working memory zone, the algorithms

may be ranked in the order Al , A3 , and A2 • Indeed, Al operates with

the total link flows, therefore requires a relatively small number of

memory words. The number of memory words required is approximately

equal to 3(n + m). In A3 ' one has to store, for each origin, the

related link flows, and the in-tree links. The number of memory words

required here is approximately equal to (n + 3) (n + m) , where n

o 0

is the number of origins. However, as one origin is processed at a

time, the use of auxiliary memory can be easily implemented. In the

code corresponding to A2 ' the main part of the working memory zone

is allocated to the storage of the generated paths. a simple list

structure has been used to cope with the unequal number of links on

each path, and the unequal number of paths related to each OD pair.

The number of memory words required is approximately equal to 3(n + m)

+ 2p + 3q + tq , where p is the number of OD pairs, q is the maxi-

mum number of paths allowable for all OD pairs, and t is the expected

number of links on each path. Even for moderate size networks, the size

of this block is quite large, and is responsible for the limited capacity

of the in-cove code. However, as A2 is based on a sequential approach,

it is again possible to use an auxiliary memory without paying

a serious penalty in the execution time.

 173

Two real life traffic assignment cases have been used to test the

different algorithms.

The first case is a daily-average traffic assignment over the mo-

delling network of the city of Hull, Quebec, Canada. The network has

376 links, 155 nodes and 27 centro!ds. The trip matrix has been obtained

with a gravity model distribution program. The cost function considered

for this case is the travel time function suggested in the Bureau of Pu-

blic Roads assignment manuel [271 with the following functional form:

f 4
t
a
t a [1.0 + 0.15 (-1!.) 1
k
a

where t is the free flow travel time, and k the practical capacity
a a
of link a By varying the number of OD pairs of the original problem,

we produced a set of test problems of increasing complexity.

The second case is a peak-hour traffic assignment over the model-

ling network of the city of Winnipeg, Manitoba, Canada. This network

has 2788 links, 1035 nodes and 146 centro!ds. The work-trip matrix has

been obtained from a survey done in 1971, and the total trip matrix is

obtained from the latter by applying appropriate conversion factors. The

cost functions considered are again the travel time functions. These

functions have been developed by the Traffic Research Corporation for

the city of Winnipeg. There are 19 different curves, all sharing the

same functional form:

t =
a
d
a
[0
f
+ a(-1!.
n
- y) 1 +
\Va.'2 f
(-1!. - y)
n
2
+S
a a
 1M

where d is the length in miles, and n is the number of lanes of

a a
link a 0, a, Sand yare parameters specific to each of the 19 cur-

ves.

The shortest path algorithm implemented in the codes is Dijkstra's

algorithm [ 81, and the determination of the step length t* is achieved

with the Golden Section method [141. For algorithm A2 and A3 , we

have multiplied the MRG direction, and the CS direction by appropriate

factors in order to obtain the same interval for the step length t,

o~ t ~ 1 , in all the codes. The tolerances in the stopping criteria

are suitably chosen in each code, so that one obtains for the same test

problem, approximately a same final value of the objective function with

different codes.

To emphasize the relative efficiency of the algorithms, we plot

in figure 1, the execution times of the algorithms versus the number of

OD pairs in the test problems constructed from the traffic assignment

data of the city of Hull. The graphs in figure 1 clearly indicate that

for this type of problems, algorithm A3 is the most efficient, often

by a significant margin in computer times. The computer execution time

ratio between A3 , A2 and Al is roughly: 1:2:4. In A2 ' the deter-

mination of the step length t* comsumes slightly more time than the

calculation of shortest paths, while in Al the major portion of the

execution time is spent in the calculation of shortest paths. The re-

lative efficiency of A3 may be explained by the fact that adjustments

of flow are generally restricted to portions of paths, thereby reducing

considerably the number of links involved in each flow adjustment. This

 175

is clearly demonstrated by a relatively small amount of time consumed

in the determination of the step length t*. A debated discussion of

these test results may be found in reference [~61. An interesting detail

to be noted from figure 1 is the fact that the execution time of algo-

rithm Al , which is consistently higher then those of A2 and A3 ' does

not, however, increase linearly with the increasing number of OD pairs.

In fact, the curve corresponding to Al seems to bend slighly at the top;

this could possibly indicate the existence of an assymtote with a small po-

sitive slope. If this is true, then based on the slope of the curve

corresponding to A3 , we may infer that for large size networks the run-

ning times of Al and A3 would not be as different as indicated by these

test problems.

In figure 2, we plot the objective function values at different

stages versus the cumulative execution times of algorithms Al and A3

on obtaining the peak-hour equilibrium traffic pattern over the network

of the city of Winnipeg. These graphs give a good representation of the

empirical convergence of the algorithms with respect to real life traffic

assignments over large size networks. The empirical convergence of both

algorithms Al and A3 is quite reasonable. However, algorithm Al still

exhibits a tailing off phenomenon, in the vicinity of the equilibrium re-

gion, which is a well know property of the Frank-Wolfe method. Even

though A3 has a computional advantage in terms of running time, the

actual cost to the user is nearly the same (to obtain an objective value

of .8640 x 10 6 , the cost of running Al is 314.89 dollars and that of

A3 is 291.38 dollars), since A3 requires more memory than Al

 176

This preliminary comparative study of the basic equilibrium traffic

assignment methods strongly indicates that these methods can solve at

reasonable cost real life traffic assignments. For small and moderate

size networks, algorithms A3 and A2 are preferable to Al For large

size networks, both algorithms Al and A3 may be used without any sub-

stantial difference in actual user cost.

 177

80

,...
(/)

~
()
70
Ai

/A,
~
'J

~
E-t
60

a
~ 50
~
40

30

20

10

100 200 300 400

NUMBER OF OD PAIRS

FIGURE 1

Relative efficiency of the algorithms (network of HULL)

 178

a0

;]
a

OJ I
~ A1.

a
0
A3
a
(\J

....
~
en

0
.--<

-.
X- C>
0
0
W
:::l m
.-J
IT
>
-,.
"'-- D
0 0
0
I- a
L) en
-.,.
:::>
LL
0
W 0
> 0
>-i
en
I- ID
L)
W
J
co
.
0 0
0
0
II)
ID

0
0
~
r---
ID

a0
a
ID
0:>
1.000 3.000 5.000 7.000 9.000 11.000
EXECUTION TIME(XIO 2)

FIG. 2 EXPERIMENTRL CONVERGENCE

 179

REFERENCES t; BIBLIOGRAPHY

[1] Beckmarm, M.J., McGuire, C.B., Winsten, C.B., STUDIES IN THE

ECONOMICS OF TRANSPORTATION, Yale University Press, New Haven,
Corm., 1956.

[ 2] Bruynooghe, M., Gibert, A., Sakarovitch, M., "Une methode d' affec-
tation du trafic", in FOURIH SYMPOSruM ON THE THEORY OF TFAFTIC
FlOW, Karlsruhe, 1968.

[ 3] O1arnes, A., and Cooper, W. W., ''Multicopy Traffic Network Models",

PROCEEDINGS OF THE SYMPOSruM ON TRAmC FlOW (R. He:nnan, ed.),
1961.

[4] Dafermos, S.C., ''Network Models for Traffic Assignment", Cornell

Uni versi ty, Iepartrnent of Operations Research, Technical Report
No. 90, Ithaca, N.Y., 1969.

[5] Dafermos, S.C., and Sparrow, F.T., "The Traffic Assignment Problem
for a General Network", JOURNAL OF RESEARCH, NATIONAL BUREAU OF
STANDARDS, Vol. 37 B, No.2, 1969.

[ 6] Dafermos, S. C., "Toll Patterns for Multiclass-User Transportation

Networks", Cornell University, Iepartment of Operations Research,
Technical Report No. 143, Ithaca, N.Y., 1972.

[ 7] Dial, R.B., "A Probabilistic Multipath Traffic Assignment Model

which Obviates Path Enumeration", TRANSPORTATION RESEARCH, Vol. 5,
pp. 83-111.

[8] Dijkstra, E.W., "A Note on Two Problems in Cormexion with Graphs",
NUMERISCHE MA'llID1ATIK, Vol. 1, 1959, pp. 269-271.

[ 9] Frank, M., and Wolfe, P., "An Algori thIn of Quadratic Prograrruning",
NAVAL RESEARCH LOGISTICS QUARTERLY, Vol. 3, 1956, pp. 95-110.

[ 10] Geoffrion, A.Mo" PERSPECTIVES ON OPTIMIZATION (A.M. Geoffrion, ed.),

Addison Wesley, 1972.
 180

[ 11] Gibert, A., "A Method for the Traffic Assigrunent when Demand is
Elastic", LBS-TNT- 85, Transport Network Theory Unit, lDndon Busi-
ness School, lDndon, 1968.

[12] Jewell, W.S., ''Models for Traffic Assigrunent", TRANSPORTATION RE-

SEARca, Vol. 1, 1967, pp. 31-46.

[ 13] Jorgensen, N.O., "SCJJe Aspects of the Urban Traffic Assignment

Problem", Graduate Feport, Institute of Transportation and Traffic
Engineering, lliiversi ty of California, Berkeley, 1963.

[14] Ka.vallk, J., and Osborne, M.R., ME'IHOrn FOR UNCONSTRAINED OPTIMI-
ZATION PROBLEMS, Elsevier, New York, 1968.

[15] Leventhal, T.L., Nemhauser, G.L., and Th:>t'ter, L.E., "A Column
Generation Algorithm for Optimal Traffic Assigrunent", TRANSPORTA-
TION SCIENCE, Vol. 7, No.2, 1973.

[16] lllenberger, G.G., INTRODUCITON TO LINEAR AND NONLlNEAR PROGRAMMING,

Addison Wesley, 1973.

[17] Martin, B. V., "A Canputer Program for Traffic Assigrurent Fesearch",
M. 1. T. Civil Engineering Systems Laboratory, Fesearch Feport RS4-41,
1964.

[18] Martin, B.V., and Manheim, M.L., "A Fesearch Program for C01!parisrn.
of Traffic Assignment Techniques", HIGHWAY RESEARCH RECORD 88, 1965.

[l9] Mosher, W.W., "A Capacity Festraint Algoritlun for Assigning flow
to a Transport Network", HIGHWAY RESEARCH RECORD 88, 1963.

[20] Murchland, J.D., "Road Network Traffic Distributioo in FquilibrilDll",

paper presented at the conference ''Mathematical Methods in the Eco-
nomic Sciences", Mathematisches Forschungsinsti tut, Oberwolfach,
1969.

[ 21] Netter, M., "Affectation de trafic et tarification au cont marginal

social: Critique de quelques idees admises", TRANSPORTATION RE-
SEARCH, Vol. 6, No.4, 1972.
 181

[ 22] Overgaard, K. R., "Testing a Traffic Assignment AlgorithIn", VEHICU-

lAR TRAFFIC SCIENCE, Elsevier, New York, 1967.

[23] Robillani, P., "Probabilistic Multipath Assignment Models", Univer-

site de Montreal, r.epartement d' infornatique, Publication No. 112,
1973.

[ 24] Schneider, M., "A Direct Approach to Traffic Assignment", HIGHWAY

RESEARCH RECORD 6, 1963.

[25] Schneider, M., ''Probability Maximization in Networks", paper pre-

sented at the International Conference on Transportation Research,
Bruges, 1973.

[ 26] TanJin, J.A., "Minimum Cost Multicamnodity Network TIoos", OPERA-

TIONS RESEARrn, Vol. 14, 1966, pp. 45-51.

[27] U.S. D=parbnent of Ccmnerce, Bureau of Public Roads, Traffic Assign-

ment Manual, Washington, D.C., 1964.

[28] War'Clrop, J.G., "Some Theoretical Aspects of Road Traffic Research",

PROC. INST. CIVIL ENGINEERS, PART II, Vol. 1, 1952, pp. 325-378.

[ 29] Wigan, W.R., "Benefit Assessment for Network Traffic Models and
Application to Road Pricing", ROAD RESEARrn IABORATORY REPORT LR
417, Croothome, Berkshire, 1971.

[ 3D] Wilkie, D. F., and Stefanek:, R.G., "Precise Determination of Equili-

brium in Travel Forecasting P:roblems Using NUJrerica1 Optimization
Techniques", HIGI-MAY RESEARCE RECORD 369, 1971.

[ 311 Wolfe, P., "Methods of Nonlinear P:rogramming", NONLINEAR PROGRAMMING

(J. Abadie, ed.), John Wiley, New York, 1967, pp. 97-131.

[ 32] Zangwill, W. I., NONLINEAR PROGRAMMING: A UNIFIED APPROArn, Prentice

Hall, Englewood Cliffs, N. J., 1969.

[ 33] Zoutendijk, G., METHODS OF FEASIBLE DIRECTIONS, Elsevier, 1960.

 182

[34] Ruiter, E.R., "The Prediction of Network EquilibrilDll -

The State of the Art", Transportation System Division of
Civil Engineering, M. I. T ., paper presented at the Inter-
national Conference on Transportation Research, Transporta-
tion Research Forum, Bruges, 1973.

[35] Nguyen, S., "An Algorithm for the Traffic Assignment Problem",
publication #123, nepartement d'infomatique, Universite de
Montreal, 1973.

[36] Nguyen, S., ''Une approche unifiee des methodes d' equilibre
pour l'affectation du trafic", publication #171, neparte-
ment d'infomatique, Universite de Montreal, 1974.

[37] Leblanc, L.J., "Mathematical Prograrrnning Algorithms for Large

Scale Network EquilibrilDll and Network Design Problems", Ph.D.
dissertation, Depar1Jnent of Industrial Engineering and Manage-
ment Sciences, Northwestern University, 1973.

[38] Peterson, E.R., Fullerton, H. V. and Cloutier, J .E., "Railcar

Network Model Feasibility Report", School of Business, Queen' 6
University, Kingston, Ontario, March 1971.

[39] Cloutier, J .E., "Traffic Assigrunent Algorithm ODFLOW", School

of Business Working Paper Series No. 71-1, Queen's University,
July 1971.
 NETWORK EQUILIBRIUM CAPABILITIES
FOR THE UMTA TRANSPORTATION
PLANNING SYSTEM

Earl R. Ruiter
Cambridge Systematics, Inc.
Cambridge, Massachusetts, USA

ABSTRACT

This paper describes the network equilibrium capabilities which have been
specified for incorporation into a revision of the UMTA Transportation Planning
System (UT~S) highway assignment program, UROAD. These capabilities will provide a
number of options for the calculation of network equilibrium when travel demands are
fixed. The options are all based on the data structure limitation that only total
link flows may be saved between iterations; link flows are not identified by their
origin and/or destination.

Within the general framework of an equilibrium algorithm based on the feasible

direction strategy, equilibrium options are provided in three areas. The options in
each of these areas are as follows:

1. Initial network flows and costs

a. Use of the results of a previous assignment

b. Use of zero flows and the corresponding travel costs
c. For existing facilities, use of observed travel costs, and zero flows
d. Use of previously predicted travel costs, and zero flows
e. Use of travel costs corresponding to previously predicted flows, and
zero flows.
 184

Whenever zero flows are specified above, the initial solution can be completed by
using the loading method chosen in 2, below; combined, if desired, by a CATS-type
loading with link updating periodically during the loading process.

2. Alternative trial solution generation methods, following the computation of

minimal cost paths:

a. All-or-nothing loading
b. Loading using Dial's probabilistic method
c. Level-up loading -- a new approach developed specifically for the UTPS
program and described in the paper.

3. Alternative methods for the combination of trial solution and previous

solution:

a. Objective function minimization using a search procedure.

b. Approximate objective function minimization using direct evaluation.

The new equilibrium version of UROAD will retain both the production orienta-
tion and the flexibility of the present program. It will include a set of default
options and the associated parameters for the typical transportation planner. In
addition, it will allow the sophisticated user to add additional options such as
those described above and including alternate link supply functions.
 185

INTRODUCTION

Specifications for a set of modifications to the UMTA Transportation

Planning System (UTPS) highway assignment program, UROAD, have been developed.

The purpose of these modifications is to expand UROAD's existing network

equilibrium capabilities. The modified program incorporates a number of

algorithms for computing network equilibrium. The program provides the

transportation analyst with various ways to trade off assignment accuracy with

assignment cost, and at the same time assures that additional computations

are used most efficiently to improve previously-obtained assignment results.

In addition, all of the flexibility of previous UROAD versions is retained,

providing the sophisticated user with an opportunity to create additional

network equilibrium options. Also, the program allows the user to choose

from standard link volume-time (supply) functions of the linear segment type,

or allows him to provide his own functional forms.

As it has been developed to date, UROAD is a very flexible program for

analyzing highway networks by computing minimum paths, by using either multi-

path or all-or-nothing trip loading procedures, and by applying a number of

classes of capacity restraint methods, including one approach based on network

equilibrium theory. Much of this flexibility is achieved by providing a num-

ber of user entry points, at which specialized processing can be performed.

The present version of UROAD, as well as the modifications specified here,

are being developed as part of the Software Systems Development Program of

 186

the Urban Mass Transportation Administration [8]*. Functional and detailed

specifications for the UROAD modifications have been prepared by Cambridge

Systematics, Incorporated under subcontract to Peat, Marwick, Mitchell & Co.

In addition to the author, the project staff includes John D. Murchland,

whose considerable contributions, both in the form of earlier development of

the concepts of network equilibrium and of the specifics of the modifications

described here, have been crucial to the success of this effort. These con-

tributions are gratefully acknowledged.

The modified UROAD program is seen as a further extension of the analy-

sis capabilities of UTPS, improving the system's ability to accurately model

flows in congested networks. As part of UTPS, the program is integrated

with the data structures of the other programs in the system, as well as

with the highway network data structures of the Federal Highway Administra-

tion's (FHWA) Urban Transportation Planning System. The program operates

in the computer hardware environment which is standard for UTPS: an IBM 360

or 370 computer using the full Operating System/360 and having both tape and

disk input/output devices. Core requirements vary from one UTPS program to

another and, within programs, by the problem size.

This paper describes the equilibrium-oriented modifications of UROAD

and relates them to the general framework of an equilibrium algorithm based

on the feasible direction strategy. Because the paper is being presented at

a symposium on traffic equilibrium methods, it will not dwell on the mathe-

matical formulation of the network equilibrium problem, or on a mathematical

justification for the general algorithm. These subjects are adequately

covered in the existing literature [2, 5] and will undoubtedly be restated

* References are listed at the end of this paper.

 187

in a number of papers presented at the symposium.

This paper will briefly place the present effort within its historical

perspective and then state the general framework within which the modifications

have been developed. The major portion of the paper will descripe the options

to be provided in each of the following areas:

• Initial network solutions

• Trial solution generation methods

• Solution combination methods

• Link supply functional forms

The paper concludes with descriptions of the opportunities provided in

the program for the insertion of additional options developed by the sophis-

ticated user, as well as the default algorithm provided for the user who makes

no specific choices of the available options.

THE HISTORICAL PERSPECTIVE*

The assignment of travellers to specific facilities, or links, of a trans-

portation system is a critical step in the total transportation planning process.

Past recognition of this criticality has led to the early development and con-

tinued improvement of traffic assignment computer programs. These programs

quickly progressed from all-or-nothing assignments based on fixed link times or

distances to diversion and capacity-restrained assignments designed to split

the flows among parallel routes. Later, assignment procedures were developed

which adjusted both link times and interzonal demands as the network was loaded.

All of these refinements of all-or-nothing assignments were very ad hoc

by nature. Until they were tried, no one knew how closely they would converge

to a satisfactory solution. Indeed, the nature of a satisfactory solution itself

*This section is a general summary of Reference [7].
 188

was very poorly understood by the transportation analysts.

While the transportation analysts were developing a whole range of ad hoc

assignment procedures, a number of applied mathematicians and economists were

carefully defining the traffic assignment problem in mathematical terms, and

proposing methods for its accurate solution. Unfortunately, interaction

between these two groups has only recently become sufficient to result in an

assignment method which is both mathematically correct, in that its convergence

properties are known, and properly efficient for the processing of large networks.

Now, however, the bridge between accuracy and efficiency can be built. Net-

work equilibrium algorithms have been developed which are accurate, and at the

same time, have the efficiency necessary for application to large networks. The

purpose of the modified UROAD program is to provide a production-oriented high-

way network equilibrium capability as part of the larger UTPS system.

The highway network equilibrium problem has been chosen to be addressed

initially because congestion is much more critical in the highway case than it

is in the transit case. The problem addressed has been further limited to fixed-

demand traffic assignment rather than variable-demand assignment because the

fixed case is by far the most cornmon problem now addressed using production-

oriented systems. Extensions of equilibrium methods to the multi-modal, variable-

demand cases have therefore been left as possible future tasks.

Because of the limitation to fixed-demand highway networks, the modified

UROAD program can be thought of as providing a capacity restraint procedure as

part of UTPS which will include a number of alternative approaches, all of which

will have the following advantages over traditional capacity restraint methods:

• The procedure insures that additional assignment iterations will

result in increased accuracy in representing equilibrium flows
in the transportation network.

• The procedure includes measures of convergence which can be used

by the analyst to determine how accurate a given solution is and
how much improvement has been gained over the previous solution.
 189

A NETWORK EQUILIBRIUM SOLUTION STRATEGY

A general strategy for the solution of programming problems, such as the

network equilibrium problem, which can be stated as the minimization of a con-

vex objective function with linear constraints is the feasible direction strategy.

As described by Florian and Nguyen [2], this strategy starts with an initial

feasible solution - one which satisfies the constraints of the problem. From

this starting point, a sequence of feasible solutions is generated, each of

which improves the objective function and is obtained from the previous solution

in the following process:

1. By finding a direction that leads to an improved solution.

2. By finding an optimal step length to be taken in this new

direction.

The network equilibrium problem lends itself to the feasible direction strategy

because: a) improved solution directions can be found efficiently using minimum

path and other similarly efficient procedures, and b) optimal step lengths can

be found by minimizing a single-variable function of link flows and costs. The

entire process is very similar to existing iterative capacity restraint proce-

dures, which can be thought of as replacing the determination of an optimal step

length (in 2 above) with the use of a step length pre-specified by the analyst.

A general algorithm for the application of the feasible direction strategy

to the network equilibrium problem can be stated. This algorithm is the one

developed by Murchland [5], as it can be specialized to the fixed-demand problem.

The four steps of the algorithm are the following:

Step 1: Initialization: Develop an initial feasible network solution, S.

Step 2: Develop a trial solution, St' which proceeds in the best feasible

direction.
 190

Step 3: Use an optimization procedure to obtain the best next solution,

as a combination of S and St. Symbolically, where C is some

combination: S = C(S,St).
Step 4: Determine whether S is a satisfactory final solution. If it is

not, return to Step 2.

ALGORITHM OPTIONS IN UROAD

Within each of the steps of this general algorithm, there are a number

of alternative approaches available. Those provided in UROAD are described in

this section. Each of these options represents a minimum departure from exist-

ing traffic assignment procedures, allowing the maximum use of the existing

production-oriented procedures already available in UROAD. This restriction to

minimum departures from existing procedures eliminates the inclusion of ap-

proaches which involve significantly more flow variables than those traditional-

ly used in traffic assignments.

Step 1. Initialization: Initialization involves all of the operations nec-

essary to provide a basis for the iterative sequence of Steps 2 through 4 which

follow. Conceptually, the major task is to obtain an initial solution involving

placement of each O-D (origin-to-destination flow) on one or more paths connecting

the O-D zone pair.

Three methods of doing this are provided:

• Use a flow pattern generated in a previous assignment to the same network,

performed either by an equilibrium procedure or by another assignment
technique.

• Assume a set of initial link travel costs and generate an initial solution
in the same way additional trial solutions will be generated in Step 2.

• Assume a set of initial link travel costs and generate an initial solution
using a traditional assignment technique.
 191

If the first alternative is chosen, then this step simply involves input

of the previous flow pattern. If the second or third alternative is chosen,

than a further decision must be made concerning the choice of a set of initial

travel costs. The user may choose from the following alternative sources of

initial travel costs:

a. For networks representing existing facilities, observed travel

costs.

b. For networks on which previous assignments have been performed,

previously predicted travel costs.

c. For networks on which previous assignments have been performed,

travel costs corresponding to previously predicted link volumes.
(This alternative will -differ from £ only if previously predicted
link costs and volumes were not in agreement with the link supply
functions.)

d. Travel costs corresponding to link volumes equal to some fraction

of link capacity.

e. Travel costs corresponding to zero link volumes.

If the third alternative is chosen, then a decision must be made on the

assignment technique. UROAD provides for CATS-type loading (updating of link

volumes periodically during assignment - after the processing of each N zones,

where N can be specified by the user.)

If either the second or third alternative initial solution generation method

is used, then this step concludes with an updating of all link and O-D cost varia-

bles. The updated O-D cost variables are set equal to the cost on the new minimum

path for the O-D pair.

Step 2. Develop a Trial Solution: The greatest local improvement in the net-

work solution is provided when the direction (the new paths for each O-D pair) for

the trial solution is based on the minimum cost paths. However, it is also known

that equilibrium solutions generally involve the use of a number of paths for each

O-D pair. Therefore, multipath trial solutions have the potential advantage of
 192

more closely matching final solutions, leading to more rapid convergence in spite

of the fact that "steepest descent" trial solutions are foregone. UROAD allows

the choice of trial solutions based on loading flows, either on the new minimum

paths only (the traditional all-or-nothing procedure) or on multiple paths cho-

sen based on the minimum paths. Two multipath procedures are available in the

modified UROAD. The first is Dial's probabilistic method, termed STOCH, which

is presently available [1]. The second multipath method has been developed speci-

fically with equilibrium procedures in mind. Termed level-up loading, it is des-

cribed in detail in Reference [3].

Level-up loading is based on the observation that after two or three itera-

tions of the equilibrium procedure, the solution being developed is much closer

to equilibrium than any all-or-nothing, single path trial solution can be. There-

fore, it is reasonable to combine the guidelines provided by minimum cost path

information with information on the present state of the solution to divert some

travel from minimum paths to almost-minimum paths in a way which will match the

available solution as closely as possible. Because the minimum paths change from

one iteration to another, it is never possible to match the available solution

exactly, which is to be avoided because it prevents improvement of the available

solution.

Level-up loading is similar to STOCH loading in that it processes each net-

work node just once for each origin zone in order of decreasing travel cost.

Paths via each link into the node are formed by assuming that they proceed from

the origin zone to the link's origin node via a minimum path, and then via the

link to the node being considered. Each of these paths which has a cost less than

some multiple (typically slightly larger than one) of the minimum cost to the
 193

node being considered is a potential path for the volume passing from the origin

zone through or to the node being considered. Volumes are then split among the

links being considered in decreasing proportion to the costs of the potential

paths and in increasing proportion to the difference between the initial link's

previous volume and the current assigned volume on the path

When either multipath load method is used, a number of parameters are available

to provide for control of the degree of "spreading" allowed for in minimum paths.

Step 3. Combine Sand St to Obtain a New Solution:This is the critical step,

because it is here where traditional capacity restraint methods fall short of being

network equilibrium procedures with proven convergence properties. To guarantee

convergence, it is necessary that the proportions of old and trial solutions be

determined by the procedure itself rather than by the analyst. The combination

method can be described as having two components; the characteristics of the com-

bination function itself, and the procedure for choosing the parameter of this

function:

The Combination F.unction -- When the number of link flow variables is limited

to one per link~hen the only feasible combination function is one which increases

the volumes in the trial solution by a fraction and at the same time proportional-

ly decreases the volumes on all previous paths.

The Combination Function Parameter -- The fraction of the new flow to be add-

ed to the remaining portion of the old flow can be obtained by using an iterative

search procedure to maximize the solution improvement. Alternatively, efficiency

may be gained by using a direct solution method on an approximate form of the pro-

blem.
Both of these combination methods - combination by objective function mini-

mization using a search procedure and combination by approximate objective func-

tion evaluation using a direct analytic procedure - are available as options in

UROAD. When the search procedure method is used, an attempt is made to reduce the
 194

number of iterations required by using the results of the approximation method

as an initial estimate of the combination fraction, for input to the search

procedure.

The existing user subroutines available in UROAD can be used to perform

alternative combination methods devised by the program's users. These methods

mayor may not provide solutions which always converge to equilibrium. For

example, analyst-specified fractions for each iteration can be used for combin-

ations, or for incremental-type assignments.

Step 4. Apply a StOpping Rule: UROAD provides three limits on the num-

ber of assignment iterations to be performed. These limits may be used either

singly or together (in 'which case if any of the specified conditions is met,

the process is stopped). The three limits are:

a. The procedure is stopped when the number of iterations equals

a specified maximum.

b. The procedure is stopped when an error measure, e l , based on

the fractional change in the solution objective function be-
tween iterations, is less than a specified value.

c. The procedure is stopped when an error measure, e 2 , based on

the degree to which flows occur on the current paths,is less
than a specified value.

The error measures e l and e 2 require further explanation. The value of e l is

obtained as follows:
~
Z

where:

Z is the solution objective function which is defined as follows:

f
Z = E
a
!
0
a s
a
(x)dx

and:

f is the flow on link a

a
 195

sa{x) is the supply function: the cost of using link a as a function

of its flow x.

~z is the change in Z from one iteration to the next.

The value of e 2 is obtained as follows:

Lc f
a a

L c f
a a
a
where:

a refers to a link

k refers to a O-D pair

c is a cost or time

f is a flow.

The numerator of this error measure is Murchland's 8, which represents the sum

of the solution objective function and its dual [51. By dividing by total system

cost, a normalized form of this measure is obtained.

Link Volume-Time Functions: A final equilibrium algorithm-related feature

of the modified UROAD program is the form of the link supply, or volume-time func-

tions provided. UROAD provides a single volume-cost relationship which includes

a general volume-time relationship. The structure of the volume-cost relation-

ship is:

COST =[TIMEF (link parameters) x CT/60 + CD]X DIST

where:

TIMEF is a volume-time function. It provides time in minutes per mile

as a function of link and link type parameters.

CT is the cost of time in dollars/hour.

CD is the cost of distance in dollars/mile.

 196

DIST is the distance of the link, in hundredths of miles.

COST is the cost of travelling over the link, in cents.

Flexibility is provided in UROAD by using a very general linear-segment

form for the TIMEF volume-time functions. In addition, UROAD's user-coded sub-

routines can be used to add additional functions which need not be piecewise

linear. There may be up to 20 of these volume-time functions, each defined using

a maximum of five points. If linear segment volume-time functions are not satis-

factory in a particular application, provision is made for the user to supply his

own functions in a user-coded subroutine.

DEVELOPING ADDITIONAL OPTIONS

In order to accomodate a wide range of additional equilibrium options, and

to give the user the opportunity to intervene at various points in the operation

of the UROAD program, a number of entry points are provided at which user-supplied

subroutines may be invoked. These entry points can be used to provide the follow-

ing types of alternatives to the equilibrium- options discussed above:

1. Initial solutions based on special operations applied to volumes and

times provided in the input historical record.

2. Combination methods based on pre-specified combination fractions

which can represent either an iterative solution method (old and

new solution fractions always sum to 1) or an incremental method

(old fraction always 1, new fraction less than 1, sum of new frac-

tions for all iterations sum to 1). These alternatives, combined

with specialized link time and cost updating procedures, can dupli-

cate a number of traditional capacity restraint methods.

3. Link volume-time functions which are alternatives to the linear

segment form provided. One alternative is the FHWA form.

 197

A DEFAULT ALGORITHM

The default values of each of the options and parameters described in the

previous section and their effects on the algorithm provided by UROAD are des-

cribed in this section. The default options have been chosen because they have

proven to be most efficient and realistic in the network testing done to date.

However, since no testing effort can encompass all eventualities, the non-default

options have been provided because of their potential usefulness for some net-

work equilibrium problems. The conditions under which these options may be pre-

ferred are also indicated.

Initial Solution: The default alternative is the use of travel times which

correspond to zero travel volumes. However, if a previous assignment has been

made of the same trips to the network (i.e., if a previous assignment is being

improved by performing more iterations), then the previously predicted flow pat-

tern should be used. If a new assignment is to be performed after minor network

or trip table modifications, then either the previously predicted travel times or

travel times and costs corresponding to previously predicted link volumes should

be used. Finally, if link capacities have been carefully specified then travel

times and costs corresponding to a fraction of capacity can be used.

Trial Solution Generation Method: The default alternative is level-up load-

ing which has proven to be very efficient in approaching a good estimate of the

optimal objective function with three to five iterations. Allor nothing loading

may be desirable when only two or three iterations are to be performed on net-

works so large that running time is an overriding factor. STOCH loading with

all diversion parameter values equal to 0.1 can be recommended if a final flow

pattern is desired which approaches agreement both with the STOCH flow splitting

rule, and with the link volume-cost functions.

 198

Combination Method: The default alternative is objective function minimi-

zation using a search procedure. This alternative foregoes some of the poten-

tial efficiencies of the approximation method, but is safe for even the most

extreme volume-cost functions. If network congestion is not expected to be high-

ly critical, or if very large networks are being analyzed, then the approximate

objective function minimization using direct evaluation alternative should be

considered.

Link Volume-Time Functions: The default alternative is the use of stan-

dard linear segment functions. This option has been chosen for implementation

both due to flexibility of the functional form and because this form has been

used successfully in previous methods 14,6]. By providing for the definition

of up to twenty linear segment functions, flexibility is greatly multiplied.

Provision has been made for user-supplied volume-time functions in order to

address unique situations beyond the range of the linear segment functions. These

situations might include volume-time functions which depend on the volumes for

more than one link, unique functional forms not well approximated by linear seg-

ments, or replication of the functions used in other assignment packages, such as

the FHWA battery.

 199

REFERENCES

1. Dial, R.B., "A Probabilistic Multipath Traffic Assignment Model which

Obviates Path Enumeration", Transportation Research, Volume 5, 1971.

2. Florian, M., and S. Nguyen, "A New Look at Some Old Problems in
Transportation Planning", prepared for presentation at the PTRC Summer
Award Meeting, 1974, University of Montreal, Montreal Canada.

3. "Functional and Detailed Specifications for a Highway Network Equilibrium

Program", prepared for UMTA Planning Methodology and Technical Support
Division by Cambridge Systematics, Inc., Cambridge, Massachusetts (forth-
coming) •

4. Irwin, N.A., N. Dodd, and H.G. von Cube, "Capacity Restraint in Assignment
Programs", Bulletin 297, Highway Research Board, Washington, D.C. 1961.

5. Murchland, J.D., "Road Network Traffic Distribution in Equilibrium", paper

for the Tagung ueber "Mathematische Methoden in den Wirtschaftswissen-
schaften", Mathematisches Forschungsisstitut, Oberwolfach, October 1969.

6. Ruiter, Earl R., "ICES TRANSET I: Engineering Users' Manual", Research

Report R68-10, M.I.T. Department of Civil Engineering, Cambridge, Mass.,
1968.

7. Ruiter, Earl R., "The Prediction of Network Equilibrium: The State of

the Art", in Proceedings of the International Conference on Transportation
Research, 1973, held in Bruges, Belgium; Transportation Research Forum,
June, 1973.

8. "UMTA Transportation Planning System (UTPS) Reference Manual", UMTA

Planning Methodology and Technical Support Division, Washington, D.C.,
August, 1974.

9. Wardrop, J .G., "Some Theoretical Aspects of Road Traffic Research",

Proc. Institute Civil Engineers, Part II, Volume I, 1952, pp. 325-378.
 APPLICATION OF ERROR MEASURES
TO 'TRAFFIC ASSIGNMENT RESULTS

Richard E. Nestle
Cambridge Systematics, Inc.
Cambridge, Massachusetts, USA

ABSTRACT

Several measures of accuracy have been applied to the results

of traffic assignments as part of research into cost vs. accuracy
tradeoffs involved in incremental traffic assignment methods. The
theoretical and practical considerations of formulating useful
accuracy measures lead to a priori conclusions about the behavior
of the measures being considered. These conclusions are generally
borne out by the results obtained when they are applied to assign-
ments on a fairly large (800 links) highway network. The behavior
of the measures leads to conclusions about the sensitivity of the
assignment procedure itself. The results also demonstrate the
limitations of certain accuracy measures and suggestions are made
as to alternative formulations which should provide more useful
measures, both for the incremental assignment methods considered
and also for iterative, full equilibrium procedures.
 SOME t40DELS FOR COMBINING THE TRIP DISTRIBUTION AND TRAFFIC
ASSIGNt1ENT STAGES IN THE TRANSPORT PLANNING PROCESS.

Suzanne P. Evans
University of Bristol
Bristol, England

1. INTRODUCTION

The transport planning process as it is usually carried out consists of a

number of stages. This paper considers commonly used models for two of these stages,
trip distribution and traffic assignment, and derives models combining them into a
single stage.

The trip distribution stage of the transport planning process is concerned with
the estimation of the number of trips per unit time which will be made under certain
circumstances between each pair of zones in an area to which the process is being
applied. In most models the estimated pattern of trips depends on the costs of
travel between the various pairs of zones and these costs are usually calculated
beforehand from fixed costs associated with the links of the transport network.
It is known, however, that the cost of travelling along a link increases with the
amount of traffic using the link, and this is usually taken into account when the
trip demands obtained from a distribution model are allocated to routes through the
network according to a traffic assignment model. The link costs which correspond to
the final estimated traffic flows obtained from the traffic assignment model are,
however, not in general the same as those assumed at the trip distribution stage.
In this paper this problem is overcome by combining trip distribution and traffic
assignment into one stage and describing them by one model.

We shall begin by discussing briefly the nature of the separate models for trip
distribution and traffic assignment which will later be combined to form single models
for the combined distribution and assignment stage.
 202

Several different kinds of model for trip distribution have been studied but

in this paper we restrict ourselves to the widely used class of models called

gravity models. A gravity model for trip distribution describes the number of

trips made per unit time between two zones as a product of three factors; one

is associated with the zone in which a trip begins, one with the zone in which

it ends and the third with the separation between the zones. The separation or

deterrence factor is usually a decreasing function of the cost of travelling between

the zones, where by 'cost' is meant some combination of the travel time, the

distance travelled and the direct monetary costs. Different kinds of gravity

model are appropriate to different circumstances and the most important

distinctions relate to the availability of estimates for the number of trips which

begin and end in each zone, and to the form of the cost or deterrence function.

In this paper two different gravity models are combined with a particular

assignm~nt model.

The purpose of traffic assignN'ent is to provide an estimate of the amount of

traffic which will use each part of a transport network under certain conditions.

In general this is done by taking the estimated trip demands obtained by solving

a trip distribut LO n model and allocating them to routes through the network

according to some hypothesis about the way in which such a pattern of trip demands

will be distributed between the various possible routes through the network. If

travellers are free to choose their own routes through the network then it is

natural to assume that each traveller, in choosing his route, attempts to minimize

his own journey cost taking account of the traffic conditions resulting from

other travellers' decisions. If we assume that the resulting decisions produce

an equilibrium flow pattern such that no traveller can decrease ms journey cost

by changing his route then each traveller has indeed minimized his journey cost

i.n the light of other travellers' decisions. Such a flow pattern is sometimes

called a selfish equilibrium or is said to satisfy the selfish equilibrium

condition. In this paper we consider an assignment model in which the traffic

flow pattern is required to satisfy the selfish equilibrium condition. The

model will be specified completely when a mathematical description of the trans-

 203

port network has been given and various assumptions about the characteristics of
the individual transport links have been made.

It has been known for some time that the usual models for trip distribution

and traffic aSSignment can be equivalently formulateJas optimization problems

subject to certain constraints and mathematical programming techniques have

frequently been used to establish the properties of these models and to find

solutions to them. In the author's Ph.D. thesis, Evans (1973), this common

mathematical programming background is used to show how the separate models may

be combined into a single model in a consistent way, and once the combined

models are formulated the usual mathematical problems of establishing existence

and uniqueness of solutions, devising algorithms for solving the models and proving

that these algorithms converge to unique solutions, are all solved by the

application of suitable mathematical programming techniques. The purpose of this

paper is to present the results contained in Evans (1973) without going too

deeply into the mathematics involved in obtaining them. For detailed proofs

the reader is referred to Evans (1973) or to Evans (1974).

A number of other attemps have been made to combine trip distribution and

traffic aSSignment. The most common resort to repeated use of distribution and

assign~~t computer programs or to slightly more complicated devices to provide

feedback between congested aSSignment and trip distribution within a suite of

transport planning programs. However this is usually very expensive to do and

the process does not necessarily converge, in any sense, to an oVArall solution.

A few other authors have proposed single models for the combined distribution

and assignment problem, notably Bruynooghe (1969), Tomlin (1971), and, most

recently, Florian et al (1974). A brief discussion of these models is given

in the final section.

The st~cture of the paper is as follows. Most of the paper is taken up

with the formulation and examination of just one model for combining trip

distribution and traffic aSSignment. This is called the first combined model

and is defined by the following three sets of assumptions. First, the demand

for trips is described by a gravity model in which the numbers of trips per

unit time which begin and end in each zone are known, and the cost function either
 204

takes on the value zero or has the form exp(-c(.<:') for each origin-destination pair.
This trip distribution model is essentially that model known as the dou;Jly

constrained gravity model with exponential cost function. It is formulated

mathematically in Section 2 and some of its properties are discussed. Secondly,

it is assumed that the cost of travelling along each link of the transport network

is a known, strictly increasing function of the traffic flow on that link such that

the cost increases indefinitely as the capacity of the link is approached.

Thirdly it is assumed that trips are assigned to the network in such a way that

the resulting traffic flow pattern satisfies the selfish equilibrium condition.

The mathematical model for assigning a fixed pattern of trips on to a transport

network under these last two assumptions is formulated in Section 3 and its

properties are discussed. In Section 4 the combined model is formulated mathemat-

ically as an existence problem and two minimization problems are presented which,

under certain circumstances, are equivalent to this existence problem. By

appealing to the theory of Rockafellar (1967), which is concerned with just such

equivalent existence and minimization problems, sufficient conditions are given

for the existence of a solution to the first combined distribution and assignment

problem. This solution is unique in the sense that it defines a unique value for

the number of trips per unit time between each origin and destination and a unique

value for the total traffic flow on each link of the network. However, the number

of travellers using each particular route is not in general unique.

Instead of attempting to solve the first combined distribution and assignment

problem directly, it is convenient to consider one of the equivalent minimization

problems, and in Section 5 an algorithm is presented for solving the chosen

problem. It has been shown (Evans, 1973) that, in the limit, this algorithm

converges to the required solution and in Section 6 several different criteria

for terminating the calculations are suggested.

Section 7 outlines several other combined models which can be formulated

and solved in an exactly analogous way, and the final Section 8 discusses briefly

the work of Bruynooghe (1969), Tomlin (1971) and Florian et al (1974).

 205

2. THE TRIP DISTRIBUTION MODEL

In this section we formulate mathematically the doubly constrained gravity

model with exponential cost function, which was introduced in Section 1.

Consider an area in which there is only one mode of transport and suppose it

has been subdivided into smaller areas or zones. Suppose the number of zones

in which trips begin is I and the number of zones in which trips end is J.

If the number of trips per unit time beginning in origin zone i is 0.." (~'" I t ~ I ••. , 1.)
and the number of trips per unit time ending in destination zone j

then
.r. T
L Q.~ ::. L.. b·
d
::. IN
~ ~ / i='
where w is the total number of trips being made per unit time. We shall denote

by t~i the estimated number of trips per unit time which begin in zone i and

end in zone j. The trip distribution process is concerned with obtaining a

suitable matrix (~b) of such estimates, which is called the trip matrix.

Clearly any possible trip matrix (t LJ J must satisfy certain conditions. The

row sums must agree with the ai' the column sums must agree with the b j , and

each element t~ must be non-negative. In addition, where there are particular

origin and destination pairs (i,j), between which, for some reason, trips are

impossible, the corresponding t'i must be zero. For those tti that are not

constrained to be zero we assume that

where r i is a factor associated with the Origin~, Si is a factor associated

wi th the destlnation j, c.i.J is the cost of travelling from zone i to zone j and.

Oi.. is a parameter. Since the exponential terms exp(-f>(.,.jl)t~ a i and the b j are

all strictly positive, so too are all the t.~ that are not constrained to be

zero.

Thus if we define the set'tl. by setting

t(:. 'd ): trips between the origin i and the destination j

are impossible.}

then the trip distribution model that we are considering is defined by the

following constraints.
 206

;rt'i :.
r~ S. el<p \-01.. C.~i) for all (i,j) ~ 1L
U)
~)
1: t .. = a..~ for all i
4.' r t.d (3)
L.t .. :: b'
~
for all j
.". L4
t- ..
~~
.... 0 for all (i,j) (; u.. ~)

and t 'A.. > 0 for all (i,j)

+ tl If the
~)

where~ is a parameter which is assumed known. G\.' b· and c. ..

"; 'J
are given as usual, then from Bacharach (1970) we have the following results.

1. If there exists any matrix (t4) satisfying conditions

(2) - (5) then there exists a unique matrix (tti> satisfying conditions

(1) - (5).

2. If there exists a solution to conditions (1) - (5) then it can be found by

the following iterative procedure called the biproportiona1 process. The

I
procedure starts with any matrix (tt~) of the form

(( ,i )
\

t~i 0 For", l\ ~ 11
t ~d..
t
and -::.. r.' S ~ e)( p ( - 01.. c·· ) for all (i,j) ~ 1..l
l d ~
At each stage 2n the row sums of the current matrix are made to agree with the

a i and at each stage 2n + 1 the column sums of the current matrix are made to

agree with the b j i the process continuing until convergence is achieved.

(Note that this procedure is identical to the Furness iterative procedure for

solving doubly constrained gravity models, which is described in Martin et al

(1961»

It has been shown by a number of different authors, notably Murchiand (1966),

Wilson (1967) and Beckmann (1971), that the problem of finding a matrix (tii )

satisfying a gravity model for trip distribution is equivalent to a constrained

optimization problem. We shall give here a minimization problem which is

equivalent to the problem of solving the particular gravity model defined by

conditions (1) ~). In this context it will be convenient to regard the IxJ

matrix (t~a) as an IJ-dimensiona1 vector denoted by ! and henceforth both

notations will be used interchangeably. We define first the objective function

F by letting
 207

The expression ti.j log ti.J is defined for all t~d ') 0 and

since
\ i iY\ t·· loa t·· = o
t.~· ~ 0 to I.J oJ \d
we can extend its domain of definition to include the origin by assigning it

the value zero at that point.

Thus the function F is defined for all ~having

o
The problem we shall consider is that of minimizing ~(S) subject to the

constraints
.-
Lt··
~ ~ L .~
= a·
• for all i (6)

t:.
~sl
t.·~ :::. b·<I for all j (7)

t··~~ = 0 for all (i,j) ~ 'U. (8)

and t li >,.- 0 for all (i,j) {- 1L (9)

Conditions (6) - (9) define a region in IJ-dimensional euclidean space

which will be denoted by A. Thus it is required to minimize the objective

function F over the set A. Note that, of the conditions

(6) - (9) defining~, conditions (6) - (8) are identical to conditions

(2) - (4), but condition (9) (tl~ ~ 0 for all (i,j) ~ tl) is not

the same as condition (5) (t':J) 0 for all (i,j) ~ 1L).

Now, by using the method of Lagrange multipliers and some standard results

from analysis, the following theorem is easily proved.

Theorem 1. If it is possible to find a matrix (ti.j) satisfying conditions

(2) - (5) then the problem of minimizing the function F over the set A is

equivalent to the problem of solving the gravity model defined by conditions

(1) - (5), and both have the unique solution (ti.d)'

.
3. THE TRAFFIC ASSIGNMENT PROBLEM

In this section we describe and then formulate mathematically the traffic

assignment model which is to be combined with the trip distribution model of

Section 2.

3.1, Assumptions and discussion

In this section we shall describe in some detail the various assumptions

 208

that define the traffic assignment model with which we are concerned. The

model is restricted to one means of transport and is described in terms of

roads. It is assumed that estimates of the trip demands have been obtained

previously by solving a trip distribution model. It remains, therefore, to

discuss the representation of the network and the assignment criterion.

The road network is represented mathematically by a directed graph consisting

of a collection of points called vertices and a collection of ordered pairs of

vertices called directed links. We shall denote vertices by the single letters

i,j,k,l, and links by ordered pairs of the form (k,l). In general the

vertices represent major intersections and interchanges, or zone centroids, which

are the points at which journeys may begin and end. A directed link (k,l)

indicates a stretch of road'without junctions which connects vertices k and 1

and on which traffic may flow in the direction k to 1. A two way road is

represented by two oppositely directed links. We assume that there is at most one

link joining any pair of vertices in each direction. We also assume that there are

no links of the form (k,k), an assumption which will be discussed later in this

section.

All origins and destinations must be at vertices which represent zone

centroids and therefore no traffic will enter or leave the network except at

such a vertex. In general the letter i is used to indicate a vertex considered

as an origin and the letter j to indicate~vertex considered as a destination.

We denote the set of all vertices by V', the subset of vertices representing

zone centroids by x., and the set of all links by i... Let V, Z and L be

respectively the number of elements in 'V)~ and L. The zone centroids are

identified by numbers running from 1 to Z and the other vertices by numbers

running from Z + 1 to V.

Associated with each link (k,l) is a function which expresses the cost of

travelling on link (k,l) as a function of the amount of traffic per unit time

using (k,l). Let X~i be the number of vehicles per unit time travelling on

(k,l) and denote the cost or congestion function on (k,l) by C~tl:x.\q(.)' Each

c~~ is assumed to be a continuous, strictly increasing function on some half-

open interval (0) X kt) such that C~t.(x) increases without limit as x tends
 209

to )(.~l from below. The number Xkt will be known as the capacity of the link

(k,l) and it is assumed that ~~l can only take on values which are strictly less

than this capacity.

In this paper it is assumed that each vertex in 7v is both an origJll and a

destination and that trips can be made between any two distinct vertices in~.

For simplicity we assume that there are no trips which begin and end at the same

vertex and this assumption is discussed later in the section. For each vertex

i in X it is convenient to define the set t.~ = t ~: ~ ~ X, I J' :f. ~ J

Then, following the notation of Section 2, for each origin-destination pair

(i,j) ~ C!: X) i Eo 'X.;) t~J is the number of trips per unit time which begin at

i and end at j.

We come now to the question of the assignment criterion. As mentioned in

Section 1 we assume that trips are assigned to the network in such a way that

the resulting traffic flow pattern satisfies the selfish equilibrium condition.

If traffic flows are regarded as continuous variables and if the cost or congestion

functions c~~ are as we have assumed them to be, then the selfish equilibrium

condition is satisfied if and only if traffic is aSSigned in such a way that the

journey costs on any used routes between an origin and a destination are equal

and are not greater than the journey cost on any unused route. This is Wardrop's

first assignment principle (Wardrop, 1952) the principle of equal journey costs.

In the preceding description of the assumptions that define the model two

important restrictions were imposed. Both links and trips that begin and end

at the same vertex were excluded. This means that int~~o~l trips are excluded

from the analysis and this simplifies the ~~Uematical formulation of the problem

considerably. However the analysis can be extended to include such trips and

links and this is done in the author's Ph.D. thesis (Evans,1973).

3.2. Mathematical formulation of the problem

In this section we formulate mathematically the assignment model defined by

the assumptions of Section 3.1.

Cons ider the set of known travel demands, t i.~. The problem is to assign

these demands on to routes through the network in such a way that Wardrop's
 210

principle of equal journey costs is satisfied.

Although transport planners may only be interested in the total am"'lllt of

traffic which will use each link, that is in the valueSof the flows "Itl I to

formulate the problem mathematically it is necessary to distinguish traffic on

each link by the vertex from which it came or by the vertex to which it is going.

We therefore let ~~,~ be the number of vehicles per unit time which use the link

(k,l) and which originated at vertex i, and for convenience we define XR~,L to be

zero for all k,l such that there is no link (k,l).

Suppose now that we have found a set {.:x:b.0 (~,~),,!} of link flows which
corresponds to an assignment of traffic such that the demand for trips is satisfied

exactly. Clearly there must exist a set of XkQ.,~ such that

For each ~ E; 'l.. X~e.,.: ~ 0 for all (k,l) E t (10)

For each \. Ci: 'A. XkQ..,c

z
= 0 for all (k,l) ~s.. (11)

.:x:R~ = lL:.
.. , X:kt II.. for all (k,l) ~ 1.. (12)

X.kl < X'H for all (k,l) ~ :. (13)

L. X:k' .
~E:'\I' .,1. .L. x'elk,...,
~E1J'
': t .~.. for 0. U ~ t k, ~'~ 'l.. ~ lilt)

L:..
k4iV
X M ,~· L.. X tR '
R~'If ,I.
:. 0 for o..lI ~t't, t ( 'X. liS-)

A set {x.\l.Q.. : (~,~) (; s. } can be regarded as a vector ~ w~os,

components are the JCIU arranged in a suitable order. Any vector;{: for which

there exist :CR~C satisfying (10)-(15) will be called feasible or will be

said to represent a feasible assignment of the trip demandst~ on to the net-

work. Suppose now that we have a set of ~~l which represents a feasible aSSignment

and suppose that this assignment satisfies Wardrop's first principle, the

principle of equal journey costs. For each i." 1.. and ~ ~ 'If (~'" t) we let <$~e.

be the cost of travelling from i to 1 on a minimum cost route. These ~~t must

satisfy certain conditions which will now be derived.

Consider a link (k,l). By the definition of the ~L~

'(t ~ '(~ +' c.Rt. bC.R,l) for all I. E; ~ where

t *' ~ ~.,cl. ~ ~ k

If traffic from i ( ~ Eo 'l.) uses link (k, 1), that is if ::c.Rt,~ >0 , then, since

the principle of equal journey costs is satisfied

+
 211

Moreover, on setting 6•• equal to zero for all i in~, these two sets of conditions

remain true for i =k and i = 1. The ~~~ must therefore satisfy

)( LI..
'.
= 0 ~r Cl.ll ~ E r... (16)

'6~t ~
'6~k + C.\':Illxlq,t. ) ~or a.\l ~t 'k) t~,e.) (i (17)

and ~~'L = 'IS' ~p' +- CRllx.Rt.) if :X:~-lIL. > 0 (18)

To summarize, necessary conditions for the existence of an assignment of the

fixed demands t~~ on to the network according to the prinCiple of equal journey

costs, are that there should exist a set of x:.~l ) (k,t)" t, for which there exist

XRQ.,~ (~E:tl R,hlf) and numbers '6Lt (l.~'X.ltE:V') satisfying conditions (10)-

Q8). Conversely it can be shown that these conditions are also sufficient. The

problem of finding a set of 'X.\\l for which there exist Xlll,' and l(.:t satisfying

conditions (10)-(18) is an existence problem. We give now an equivalent

minimization problem which has been derived in various ways by many authors and

has been used to prove the existence and uniqueness of a solution to the assign-

ment model and to devise algorithms for finding that solution.

We begin by defining a new set of functions For each link (k,l) we

set
:x.I!.t.
S o C RIL <.x:.) d x.
The integral is certainly defined for all .x.llt in the half-open interval (0) X ~Il )
and it is also defined at the point X Rl ~ X~ if the limit of C\telx) as X " - X~;
exists and is finite. Throughout this paper we assume that for each link (k,l)

C~e.(x.) tends to infinity as x tends to XP.t from below, but essentially the

same results can be proved for the case where ~Rl~e.) is defined and finite at
~i~ )(kl' although the appropriate formulation of the equivalent minimization

problem is slightly different. For each (k,l) the function C;~~ is continuous

on (o,X\lt). Il is o..\So Sh-ic.H'j c.on"e,X S\n~e its de.l"ivo..Hve" CRQ.

is Q. ih"ic.tl';l inc........ o.sin~ fu..I'1c.\-ion on Co) X"1<t41).

Define now the function G by setting G <.~) :::. l~ f. C.u (.:oc"Re.)

and consider the prOblem of minimizing G{~~ by adloice of:!So for which there

exist 'X\I."~ such that conditions (10)-(15) are satisfied.

 212

If we define the set S by

S = [x: -J
~he ...<. CI'I'st "XI..~.
1<,,) ....
~,,(.h ~hoJ co""d.i~iof\S (lo)-tlS-) ore S4Iisf,ed.]

then the problem becomes:

Minimize the function G(~ subject to ::s ~ 63.

(Note that in this case all the constraints, without exception, that define

the setG over which G must be minimized, also appear as constraints in the

existence problem. This is in contrast to the situation in Section 2 where

the constraints (9), which appear in the minimization problem, replace the

corresponding constraints (5) in the trip distribution model. It is also in

contrast to the situation which will arise in Section 4 in the case of the

combined model).

The following theorems establish the equivalence of the two. ,problems and

the existence of a unique solution to both of them.

Theorem 2. The existence problem defined by conditions (10)-(18), and the

problem of minimizing the function G over the set~ are equivalent.

~. This theorem can be proved by using the Kuhn Tucker theorem as given

in Vajda (1961), taking account of the fact that eRe Cx.) _ ~ CA.$ X. ~ XR~
(see, for example, Jorgensen(1963».

Theorem 3. If the set ~ is non-empty then there exists a unique ~* in e such

that
G( ~*)

Proof. This theorem is an easy consequence of the strict convexity of G and the

fact that for each (k,l) Eo 1..) C~t (x) -. 0() Q.S X ~ XIeQ. - •

The following corollary is a direct consequence of Theorems 2 and 3.

Corollary 1. If the set ~ is non-empty then there exists a unique flow vector ~*
.. vir
for which there exist Xkl,~ and Ott satisfying conditions (10)-(18).

It should be noted that although the solution ~* is unique, and so therefore

are the total flows Xkt on the links, the corresponding values of ",t X.~QI~ are not

in general unique.
 213

4. THE COMBINED DISTRIBUTION AND

ASSIGNMENT PROBLEM.

In Section 3 we saw that if a given set of fixed trip demands t~~ L: t X, i E; 1.. )
can be assigned to the network in some feasible way, so that all the link flows

are less than the link capacities, then they can be so assigned in such a way as

to satisfy the principle of equal journey costs. This assignment is unique with

respect to total flows on links and hence the associated travel costs ~~ are

also unique. In the situation where the travel demands ti.J (i. (; t,J (; Xi.) are

not fixed but are required to satisfy a trip distribution model involving the

costs ~~A ( ~ "" X, d' (; 1..: ) the question arises whether i t is possible to find

travel demands t~ which, when assigned according to the principle of equal

journey costs,give rise to costs ~~i which would have generated exactly those

demands t~4 .
From this discussion it is easy to see how necessary and sufficient

conditions may be derived for a set of link flows x RQ. (k,l)(;.t) and a set of

demands t\.~ (~e ~'ieA.~) to solve, in a consistent way, both a doubly constrained

model with exponential cost function of the type described in Section 2, and the

assignment model derived in Section 3. To begin with the :x.. RQ must solve the

assignment model of Section 3 with the t"L~ regarded as variables instead of

given constants. These tt~ will then be required to satisfy
t,·
L4 0: \".L S·4 ~)(.p ..
(-0<.. ~ Lei ) for all .: t A:, d' to h'L (19)

L. 1: I.d.. =- a.'• for all ~ <:: 'l.. (20)

JG- X ..
L. t Ld.' ::. b'~ for all
cI
E: 't.. (21)
i.fi:~i
and t ~d" > 0 for all .: f; ~Ji t ~~ (22)

where the '(, . are the associated interzonal minimum trip costs from the
~A

assignment model. Clearly the conditions (19)-(22) define a trip distribution

model of the type defined by conditions (1)-(5) in Section 2, in which

I = J Z, the set 'U. is defined by

1.l = \. (i,i);
so that no intrazonal trips are permitted, the costs c.'j have been replaced \oj

the ,,~~ and the zero elements t" have been omitted.
 214

To summarize, therefore, a set of link flows x lrtt «kf) Eo i..) and a

set of trip demands t(.~ «( Eo X) j €; Z~) solve this combined distribution

and assignment problem i f and only i f there exist x.~Q.,L. (Lf; XI ~,~E-1f» ¥i.t (~t~, i£ln
f" l0 €-~) and 5~ (J E; 'X, ) such that conditions (10) - (22) are satisfied.

The conditions (10)-(22) will henceforward be known as conditions I and the

model which they define will be known as the first combined distribution and

assignment model. It is worth noting that, even if transport planners are

only interested in predicting the total traffic flow on each link in the network

and the total number of trips per unit time between each possible origin and

destination, the first combined distribution and assignment model cannot be

formulated in terms of the variables ~RQ and t~~' only; it is necessary to

introduce the further variables XRQ)': ) ~':t I r'~ ""c\ s.~'

In considering the problem of finding sets of .x.ko. and l'4' which solve

the first combined distribution and assignment problem, there are three

important questions to be answered. These are:

1. Under what circumstances, if any, do such x.1t~ and 1:0:4 exist?

2. If there exist such 'X.\lQ and l'J are they uniquely defined by

Conditions I?

3. If there exist unique values of X'I\Q. and t:.:~ such that the first

combined distribution and assignment problem is solved, how can these

values be found?

Answers to these questions are obtained in Evans (1973), using the

equivalence which exists between the existence problem defined by Conditions I

and a number of constrained minimization problems. By appealing to the theory

of Rockafellar(1967), which is concerned with separable convex programming

problems with linear constraints, it can be shown that, under certain conditions,

the existence problem is equivalent to:

(i) a primal minimization problem formulated in terms of link flows and

trips per unit time between all origin-destination pair's, and

(ii) a dual minimization problem for"'" "I ed in terms of link costs, trip costs

and two other sets of variables which are related to the r, and s •
~ j
 215

The answer to Question 1 is contained in the exact statement of

Rockafellar's equivalence result which is as follows:

If conditions (10)-(15) and (20)-(22) can be satisfied at all then the

existence problem, the primal problem, and the dual problem are all

equivalent and all have solutions.

That is to say, a sufficient condition for the existence of a solution

to the first combined distribution and assignment problem is that there

should exist a set of t~ satisfying conditions (20)-(22), and a set of

~k~ representing a feasible assignment of these tlj on to the network.

The conditions (10)-(15) and (20)-(22) will be denoted by conditions I(a)

and henceforward we shall assume that they can be satisfied.

We give now the equivalent primal problem. It is required to minimize

the primal objective function ~(~)~) -=!.R~l.CR~(:(t?Q.l-;- ~ ~t.~}:':j(I03~Lj -I}

by a choice of x.~~ and I:.:~ such that there exist XRJ,i. satisfying conditions

I(b) where these conditions I(b) are obtained from the conditions I(a) by

replacing condition (22) by the new condition

o ~or 0.. II Q.3)

(fhis is analogous to the situation in Section 2, where the constraints of

the minimization problem are not all identical to corresponding constraints in

the existence problem). If we define the set ID, by

SJ\ :: i (~1 £.) : there exist :x:~~, ~ such that conditions I(b)

are satisfied}

then the problem becomes:

Minimize the function PI l~J t ) subject to (~) t) f ~I

We have already seen that the Cl:!e are strictly convex functions of the

'X.l:l.t and it is not difficult to show that the functions t L~ (log t'i -1) are

also strictly convex with respect to the tti'

The primal objective function PI l~ ) !::. ) is therefore strictly convex and

the solution to the primal problem must be unique. It follows that the solution

to the first combined distribution and aSSignment problem must be unique with
 216

respect to the link flows X~.t. and the demands t:,~. and this unique solution

will be denoted by
It is the primal problem that we shall eventually choose to solve instead of the

equivalent existence problem, but before turning to the question of how to find

the unique solution (~. J ~.) we shall give the third equivalent problem, the

dual problem.

We need first of all to introduce the vectors 1) /! of dual

variables, and a new set of functions C~e'

*"
The vectors of dual variables are defined as follows:

(i) The vector ."I.., is a I" L row vector of link costs "lqt I U~, t)E t.
(11) The vector (5
,.., is a I 1< LV row vector of trip costs ~'t ~E: '/.. J {E 1/,
~ I

(iii) The vector~ is a I~ ~ row vector with components o(~ I ~ ~ ~ I which

are related to the r i , and

(i v) The vector j: is a I)' Z. row vector wi th components ~ J ,i E X. I which are

related to the SJ'

The new functions Ctl are closely related to tbe C IU. Essentially each
~
C lI.i may be regarded as the area, not under the curve cRt. but to the left of

it, More formally. for each (k, l) Ii i.. 1 the funct ion C ~ ~ is defined by

setting
-= C I'u lo)

> C.RA. <..0 I

The dual problem can now be given as follows:-

Minimize the dual objective function

11'
])I PI'" I ~.J (3) : L
(R.,VE-J.
C.. , ( '7 iu. ) -t- ..LL:. L e.)(.p(-o<..(o(;t-f3i +'(~I'))
_)';) I \~ 1< ... 0<. 'l''};..dt:~ 1<1 ~

+ .r:...
," 'X.
Q.~ oI.~ ... ~~
j
bi ~i
by a choice of ") '" ('1 k~ I, ~::. ( '(~t ) . ~ ~ (ct.: ), ri. '" (~d') $"bJt.~d ~o
o tor oJI ~ ~ %
and for a.ll ~~ h-, LR,e.)e-! l2S)

Conditions (24)-(25) restrict the ~~t so that each 't,( is less than or

equal to the minimum cost of travelling from i to I when the link costs are
 217

given by the'1kt.
The relationships between the solutions to Rockafellar's primal, dual and

existence problems are essentially as follows:-

If (~, t) solves the primal problem and ('1) 'f., ~ I r-) solves the

dual problem then:

1. For each (k,l) E J.. '1 ~e =

2. The ~~t are the minimum journey costs from i to 1 when link costs are given

by the "'11le. (= Clu (xite. ) ) •

3. For all " ~ k, i E: A~
t ~d = e )( p ( - 0( (.,(.: 1- ~.i +- '6 ~i ))
Y'i S ~ e. j( P ( - 0(. ~ ':J )
where r i exp (- ..c.co<..) for all ~ ~ 't..
and Sj = exp(-«. ~i ) for all j E; ~
and

4. There exist:x.lI.t,~ (~Ie- t., R,t €- 11') such that the x'Rt (01.,.t)E /...),

:LR1,~ (lE~) ~,~ ~ 1/), t~j (4." x.,~'(;~)' ri (~E; X) and s~ ( 4' '" 'l- ) satisfy

conditions I I thus implying that (;G) t ) solves the first combined distribution

and assignment problem.

Finally we give a general result from Rockafellar (1967) concerning the

relationship between the possible values which can be taken on by the primal

objective function and the possible values which can be taken on by the dual objective

function.

5. If (~, t ) satisfies the constraints (conditions I (b») of the primal problem and

( 1 , 1l) ~, ~ ) satisfies the constraints (conditions (24)-(25» of the dual

problem then

(26)

and equality holds in (26) if and only i f (~, t ) solves the primal problem and

( 1. ,1, ~) p. . ) solves the dual problem.

 218

5. AN ALGORITHM FOR THE FIRST COMBINED

DISTRIBUTION AND ASSIGNMENT PROBLEM

To solve the first combined distribution and assignment problem we consider

instead the equivalent primal problem. The following is the algorithm proposed in

Evans (1973) for solving the primal problem.

The algorithm starts from any (Z, I \;;.., ) in SJ, and generates a sequence

in ~I in the following way. Suppose that the sequence

has been constructed as far as l~""t,,)·

Then

1. For each (k, 1) calculate the costs C. ~A. <. ':)C'" ~ ~) which correspond to

the flows :x:: V\ IH •

2. For each origin vertex find the minimum cos t ~0 ~ i of

travelling to each destination i in Z.: and choose a minimum cost route fron i to j.

*'
3. (See below for footnote.)

Find a new set of demands a. " Q' (; J( l ) by solving the

v.'" 4
doubly constrained gravity model defined by

'L-n'4 :: r
""
S ' e.lC:p (-0<. 'tn':d )
"'4
for all ~E:: X, d' E 7...; (27)
L '(", = el'~ for all L EO 7... (28)
'.~ 'h,"'J
.L
,Eo Xj
'(,h;~
.:. b'4 for all C 1... (29)

and
~n~j > 0 for all l- E. h, d' (; 'A.,' (30)

4. tI!>sign the new demands ~n (.~ to the minimum cost routes chosen in

Step 2 to obtain a new flow vector

5. Find the linear combination (\-)q ( ~"') ~,,) + >-. l~,",) 't . . ),
o ..::. ~ ~ l) of (~n) £.,,) and ( ~'" '<t.... ) that minimizes the

objective function PI and denote this combination by ( ::;f t'\ ~ I

6. Return to Step 1 replacing n by n + 1.

* To reduce the time taken at each iteration in solving a doubly constrained

gravity model, the solution at each iteration (n-l) can be used in iteration n

to provide improlfed sYo.,.tin<j \lo.lves \(""-'1' Sl"-llj el<.f' (-0(. ))',,~) in the Fu'-'H~~

iterative procedure. (See the second result of Bacharach (1970) given in

Section 2).
 219

Note that since we have already assumed that conditions lea) can be satisfied,
it must always be possible to satisfy conditions (28)-(30), It follows tnen, from

the discussion of Section 2, that there must always exist a unique set of demands

q,,., cd ( C " A. I J' ~ A.. ~ ) solving the doubly constrained gravity model of

Step 3,

It is proved in Evans (1973) that the sequence £(. ::: . . ,/;.,,)} in ~') generated
by the above iterative scheme, converges to the unique solution, (;S~ ~) ) to the

primal minimization problem,

6, MONITORING THE CONVERGENCE

Al though it has been shown that the sequence £C:f", t ..J} converges to

(~ ... ) £. *") as n tends to infinity, in practice the algorithm has to be stopped

after a finite number of iterations, and therefore it is important to have a

cri terion for deciding when the current (:1-

_n ) t
.....," ) is sufficiently close to the

solution ( ~ If) t'" ) for the algorithm to be stopped, The decreasing value of

the objective function PI provides one means by which the progress towards the

solution may be judged, and we shall show that a lower bound to PI can be found

at each iteration. Consider the point (:f ... It,., ) in:i)" Let <;." be the 1)( L
row vector with components e nh e. = '-I< Il. l X V'\ \<.(. ).

Let ,{" be the I,. Z V row vector with components

where each 't r'\~ t (l t ~) -t Eo 1f} l +- t) is the minimum cost of

travelling from i to 1 when the flows on links are given by the components of

~~ and each '( V\'~ ~ E: A- ) is defined to be zero,

Finally let ~ nand be the II' Z. row vectors with components ci. n • ::: -.!- I0 5 r-,,~

±
01..

(~Ii:~) and ~,,~ =- 103 S'"'~ (i ~ :z.) respectively, where the r ni and s nJ. are

found in the course of Step 3 of the algorithm of Section 5, Now (Sf'):f" )~n I ~"')
certainly satisfies the constraints of the dual problem given in Section 4 and

it follows from Result 5 at the end of Section 4 that

P, ( ~ ,t ) ~ - 3>, (£", 'f . . '~n I (i,,) tor- Q.\l (z It) In $, (31)

That is -D l ( <;"'''1 ~ .... , ~I"\ I ~,,) is a lower bound to PI'

Now it is easily shown that

 220

where ::. , and it follows, therefore, that the

lower bound can be calculated from the quantities CnkQ.. 1""". and S .... i· which are

found in the course of the n'th iteration of the algorithm. (It should be noted

that in an exactly analogous way -PI ( ~ ('\, t" ) is a lower bound to Dl ).

Equality holds in (31) if and only if (~,!) is the unique solution (~*,!;) to

the primal problem and ( C

""",t"'\. )
¥
",..,.," 1
J..
_1"\' (l,
l-:;. .....
) solves the corresponding dual

problem.

Using these facts we define two different quantities either of which can

be used to test for convergence.

Test quarr\'lI',! 1. We begin by defining the quanti ties c1"n ( i\ ,,~ I) •

Let

j), (£.,) ~, ~, , ft, )

and '= 'thin { e-n-, ]), lSI'I' ~t"\, ~n, (L,)}
If we now define.

E. ..... + for

then from (31) we have the following results.

(i) E.. n :? 0 for all n,

(ii) fEn} is a decreasing sequence and converges to zero as n ~ 00.

(iii) (", = () if and only if (?£ j\ , !n) is identically ( =!$~) f" ) .

and (iv) "PI t~n' t,,, ) PI t-£ \OJ t. . . ) .c:. c::.. n
(It can also be verified that

~~nn D, (~"") 'ir"'\, 06 .... ~"") I - 1), ('1"', '{ ... ~ If) @..II- )
- "n ]), ('2 *', t *', ~ It- / a"') ~ en
wh ere ( ':'."")
~ ~.
~~ :.
~•*) (3_
I~ * ) is a solution to the dual problem).

The quantity e~ can be used to test for convergence in various different

ways. For example (::'n,tJ may be accepted as a good enough approximation

to (3t*, 1.*) if

(32)

or if (33)
 221

where £ is some small quantity chosen according to the level of

accuracy required. The choice of E- is discussed later in this section

In order to use £.1'\ in this way, the quantities 0-.., and PI ~ .... I ~" ) have

to be calculated at the end of each iteration. Thus both the primal and dual

objective functions have to be evaluated at each iteration. Now

[") (:x:: t ):. L

r, .-'" -" (,~,Q),i
r
0
xnkQ.
qa tx)
J
<l.X +-
-.L
<I..
r... Lt., t \oa t ~.d -
~E;'A. ;fA..
•
la J
I
)

:n:(;
1 ~n)
~
-'"
d..
- " ' \-"
B ):. L (nk~ ~~ l'1 )d.1 + ~
C
l~,Q.)~J.. }en!!. (0)
-r f--x. Q.i. ~n~ +- {i--r-- bi r"i
and the integrals may require more time to calculate than is desil"'a.ble.

t quantity, analogous to one proposed by

We therefore describe another tes

Murchland (1969) for the case of th e assignment problem with elastic demand,

which requires less calculation.

Test quantity 2. Let

=-
It is shown in Evans (1973) that this quantity simplifies to

Now

( i) o for all r\.

(ii) o if and only if (~n ,,t .... ) is just (~* ,,t,*) .

and (iii) Pl(~n ,t .... ) Pl(~*'!:) -e:. bn for all n >.... \ .

(Similarly Dl ( £")~"l~"'rt-) Dl('l"") ~ ... } ~ .... ) (l.*') ~ b...,.)

Unfortunately [bn 1 is not necessarily a decreasing sequence and since it

is clear that S", ~ £", for all n ~ I I ~n is a less accurate estimate

than [. n of the amount by which the current value of PI exceeds the

minimum possible value and may therefore be considered less desirable as a test

quantity. However, it is easier to calculate than En Like c: 1"\ the

quantity ~ 1"\ can also be used to test for convergence in various different ways.

For example (.?En ,t",> may be accepted as a good enough approximation to

- -
(x*,t*) if
COn
or i f max t'D n/ \ 1', l~ n , t .. )I
 222

In using either £0 or ~ (l in the convergence criterion the choice of a

suitable ~ such as appears in expressions (32) - (35), is obviously crucial.

The quanti ties [." and b '" are both upper bounds to the amount by which the

current value of PI' P j (~'~n)' exceeds the minimum possible value PI (~*,t*)·

It is sometimes possible to find, in terms of [", and b "'- I an upper bound to

the amount by which the :x: '" 11. t and the t" "i may be different from the optimal

values X. ~t and t:j , and such information can be used to set a value for £ .

If we let
(x
""""n)
t _r'\

then by expanding PI (~"'!n) about (~*,~*), and using the conditions I that

define the first combined distribution and assignment problem, we obtain

P,l:f",t",) - P,l?f:~£") .>;, L D ck.~(:x:.tt j- e~:Crd~~)(6.xnRQ)2.­
(R,~)t'"

+ _I L. ~ flo I ) (l\tni.~·)2. l2b)

0<. .to'};. dO:., (hJ +&~l:"c~
where o e and prime denotes differentiation. We shall

suppose that we have found a quantity S such that

(37)

Clearly ei ther E '" or ~'" can replace in (37). For all ~ C 7:., J- E. 'f..~'

t."(~ t- e D. tn.:~ I
>,.,.. ""i ..... t 't"",;('\A..
_I
a.-... ) (\"
4(: i h-
'"
~J = KI > 0

and since each term on the right hand side of (36) is non-negative it is

certainly true that

d. S (38)
K,
(Note that KI is independent of the current (~" ,,!.,».
If, in addition, the cost functions c kl are such that
()'\ in I
C~t LX) ;> 0 for all (k,l) E: -J... (39)
:x. Ii. [0, Xh)

then for each (k,l) " i...

c~~ (X kt + e 'b::n~Q.) 4- mil") -
1"1"" V\
IN.,O) d, :><. f [0, >< It~ I
I
Co 1< LX)
Q.
K..:t > 0
and it follows that
L:. (6. x n \Qt):l. _I S (40)
llt,t)c: l. K:2,
(Note that K2 is independent of the current (~I"I ,!,,»
 223

Relations (38) and (40) set upper bounds to the amounts by which IIle

t n1'j and x may be different from the t * and xk*l'

nkl ij

How likely it is that condition (39) will be satisfied1 From the

assumptions that have been made about the c~~ it is easily seen that (39) is

satisfied if and only if

> o (41)

Many of the commonly used cost functions do satisfy (41), and therefore (39),

but in practice ~~t(O) may well be zero. Some suggestions for overcoming this

difficulty are given in Evans (1973). This completes our discussion of possible

convergence tests. It also completes our analysis of the first combined

distribution and assignm"'nt problem. In the next Section 7 we shall discuss

briefly a number of other combined models, in particular a second combined

distribution and assignment model in which the assignment model of Section 3 is

combined with a singly constrained gravity model.

7. FURTHER COMBINED MODELS

7.1. The second combined distribution and assignment problem

We consider a producti·on-constrained gravity model with exponential cost

function. The model is defined by the conditions.

t"l4 =- r· B~ e}(p(-"'Cl~)
~
for all
"
t: ~)
~
~ A..l (42)

j G: A.,
~ t·LJ, = a. . L for all E: % (43)

and t l<l.. > 0 for all l C f.., J't: 1..:. (44)

where for each i €: t) B·4 is a known parameter which is supposed to represent

some index of the 'attractiveness' of zone j.

Following exactly the procedure of Section 4 we can formulate a model which

combines this distribution model with the assignment model of Section 3. This

second combined model can be analysed and solved in exactly the same way as the

first combined model and in the following discussion only the main differences

will be pointed out.

 224

The conditions II which define the second combined distribution and

assignment problem are obtained from the conditions 1«10)- (22» defining the

first combined model by omitting the attraction constraint (21) and replacing

condition (19) by condition (42). Similarly, by simply omitting the attraction

constraint in each case, the primal constraints)conditions II(b» that define the

set ~ll. are obtained from the corresponding conditions I (b) that define ro" and

the conditions II(a), which must be satisfied if solutions to the three equivalent

problems are to exist, are obtained from the corresponding conditions I(a). The

new primal objective function P2 is given by

R.- (~) ~_)

-- -
L
lkMf.l.
C. III
b:Ln)
.....
+ I)
and the new dual objective function D2 is given by
'( J) --
D!I. (j_I ~)~ "\
L--
lR.,tJE: i.
c· (1'1
~I!. 'RQ.
) + _I -r
tt4.
",...
L
L-L"L'Y j""'L'
h..
B·4 e.)(.of l-oI. (0<.."t-~."
J
»)
"t-
The constraints of the dual problem remain the same. The algorithm for solving

the equivalent primal problem, that is the problem of minimizing the function

P2(~'1) subject to (~'l)~ S)~fs obtained from the corresponding algorithm of

Section 5 by requiring a singly constrained gravity model to be solved in

Step 3 instead of a doubly constrained gravity model. The proof of convergence is

contained in Evans (1973) and convergence criteria can be devised as in

Section 6.

7.2, The assignment problem with elastic demand

The assignment problem with elastic demand combines the assignment model

of Section 3 with a trip demand model consisting of a set of separate demand

curves, one for each possible origin - destination pair {i,j), such that the

number of journeys made per unit time between each given origin and destination

is a decreasing function of the cost of travelling between that origin and

destination. Thus the resulting model combines in a particular form the three

usually separate stages of the transport planning process: trip generation,

trip distribution and traffic assignment. This model was first propoaed by

Beckmann (1956) and has since been studied by a number of authors including

Gibert (1968) and Murchland (1969).

 225

The algorithm proposed by Murchland for solving the problem is analogoll9 to

the algorithm given in Section 5 for solving the first combined distribution

and assignment problem, and the same methods can be used to show that his

algorithm converges to a unique solution. For the details of the proof see

Evans(1973).

7.3. Combining modal split, trip distribution and traffic assignment

It should be noted that the methods of this paper can very easily be

extended to develop and solve a model combining the distribution and modal

split model proposed by Wilson (1967) with the assignment model of Section 3,

provided that the following assumptions are made.

1. A separate transport network carries the trips made by each separate

mode of transport.

and 2. The cost of travelling on any link of any of the transport networks

depends only on the amount of traffic using that link.

These restrictions would be satisfied, for example, if the only

methods of transport available were public transport by rail, and the

private car.

8. SOME OTHER MODELS COMBINING DISTRIBUTION AND ASSIGNMENT

In this section we discuss very briefly the work of Bruynooghe (1969),

Tomlin (1971) and Florian et al (1974). For a more detailed discussion of

the work of the first two of these authors see Evans (1973).

The combined model derived by Bruynooghe (1969) is similar to the first

combined distribution and assignment model derived in Section 4 of this paper,

but rather more general. However the proof given by Bruynooghe that solving

his model is equivalent to solving a constrained minimization problem does

not extend to the special case of the combined model incorporating a doubly

constrained gravity' model with exponential cost function, that is to the

model of Section 4. It is the use of Rockafellar's theory that makes the

results of this paper potentially more general than those derived by Bruynooghe.
 226

Two algorithms are proposed by Bruynooghe for solving the general form

of his equivalent minimization problem. Each is an adaptation of an algorithm

presented in Bruynooghe et al (1968) and under certain circumstances each is

shown to converge to a unique solution. The first algorithm involves the

finding of a set of minimum cost routes and the solving of a transportation

problem (a standard problem in linear programming) at each iteration. The

second involves exactly twice as much work per iteration since at each iteration

both the minimum cost routes and the maximum cost used routes are found and

two transportation problems are solved.

Tomlin (1971) considers the problem of combining a doubly constrained

gravity model with exponential cost function, with a traffic assignment model

which seeks a traffic flow pattern minimizing the total cost of all journeys

when link travel costs are constant and each link has an associated capacity.

This combined model is formulated directly as a nonlinear minimization problem

which is solved by using the Dantzig-Wolfe decomposition principle. By

applying this principle, the problem is reformulated as a linear programme with

the columns to be generated by two types of subproblem, a trip distribution

subproblem and a cheapest route subproblem. The calculations required are

therefore similar to those required in the algorithm of Section 5.

Finally Florian et al (1974) have suggested another algorithm, similar

to that proposed by Bruynooghe (1969), for solving what is essentially the

primal minimization problem given in Section 4. Their method is an

adaptation of the Frank and Wolfe algorithm (see Frank and Wolfe, 1956) for

quadratic programming and falls within the class of feasible direction

algorithms of nonlinear programming.

Acknowledge.n&~U - This work was carried out in the Research Group in

Traffic Studies at University College London while the author held a

Hackett Studentship awarded by The University of Western Australia. The author

is very grateful to Dr. R.E. Allsop for his most valuable advice and comments

given at every sta~e of the work. She is also grateful to Dr. J.D. Murchland

for a number of useful and stimulating discussions.

 227

REFERENCES.
Bacharach M. (1970). Biproportiona1 Matrices and Input-Output Change.

Cambridge University Press.

Beckmann M., McGuire C.B. and Winsten C.B. (1956).

Studies in the Economics of Transportation.

Cowles Commission Monograph. Yale University Press, New Haven,

Beckmann M. (1971). On the Metaphysical foundations of traffic theory;

entropy revisited.

Traffic Flow and Transportation. Proceedings of The

Fifth International Symposium on the Theory of Traffic Flow and

Transportation.

Elsevier, New York.

Bruynooghe M., Gibert A. and Sakarovitch M. (1968).

Une Methode d'Affectation du Trafic. Report from Service d'Et~d~S

Techniques des Routes et Autoroutes SETRA, Paris.

Bruynooghe M. (1969). Une Mod~le Integr6 de Distribution et d'Affectation

de Trafic sur un R~seau. Report from Institut Recherche Transport

I.R.T., Department Recherche Operationnelle et Informatique, Paris

Evans Suz~nne P. (1973). Some Applications of Mathematical Optimisation

Theory in Transport Planning. Ph.D.Thesis, University of London.

Evans Suzanne P. (1974). Some Models for Combining the Trip Distribution

and Traffic Assignment stages in the Transport Planning Process.

Unabridged and unpublished version of the paper presented at the

Symposium on Traffic Equilibrium Methods, Universite de Montreal,

November 1974.

Florian M., Nguyen S. and Ferland J. (1974). On the Combined Distribution-

Assignment of -Traffic. Publication ~8 - Centre de recherche sur

les transports, Universit~ de Montr~al.

Frank M. and Wolfe P. (1956). An algorithm for quadratic programming.

Naval Research Logistics Quarter1~ ~, 95 - 110.

 228

Gibert A. (1968). A Method for the Traff~c Assignment Problem when Demand

is Elastic.

LBS - TNT - 85. Transport Network Theory Unit, London Graduate

School of Business Studies.

Jorgensen N.O. (1963). Some Aspects of the Urban Traffic Assignment Problem.

Graduate Report, Institute of Transportation and Traffic Engineering,

University of California, Berkeley.

Martin B.V., Memmott F.W. and BONE A.J. (1961).

Principles and Techniques of Predicting Future Demand for Urban Area

Transportation.

M.I.T. Report No.3. Massachusetts Institute of Technology,

Cambridge, Mass.

Murchland J.D. (1966). Some remarks on the Gravity Model of Traffic

Distribution and an Equivalent Maximization Formulation.

LSE - TNT - 38. Transport Network Theory Unit, London Graduate

School of Business Studies.

Murchland J.D. (1969). Road network traffic distribution in equilibrium.

Paper presented at the conference Mathematical Methods in the Economic

Sciences.

Mathematisches Forschungsinstitut, Oberwolfach.

Rockafellar R.T. (1967). Convex programming and systems of elementary

monotonic relations.

J. Math.Anal.Appl. ~, 543 - 564.

Tomlin J.A. (1971). A mathematical programming model for the combined

distribution - assignment of traffic. Transpn. Sci. ~, 122 - 140,

Vajda A. (1961). Mathematical Programming.

Addison - Wesley, Reading, Mass.

Wardrop J.G. (1952). Some theoretical aspects of road traffic research.

Proc. Inst. of Civil Eng. Part II, 325 - 378.

Wilson A.G. (1967). A statistical theory of spatial distribution models.

Transpn Res. J, 253 - 270.

 MULTIPLE ROUTE ASSIGNMENT: A COMPARISON OF TWO METHODS

J.E. BURRELL

Burrell &Associates
London, England

INTRODUCTION

In the real world, the processes of trip generation, distribution, modal split
and assignment, interact with each other in complex ways. There is a growing reali-
sation, among those who desigh mathematical models of traffic, that for these models
to be fully effective, they must take account of these interactions. To do this
requires procedures which include iteration and which must therefore be made to con-
verge.

For the models to converge, not only must the stages of generation, distribu-
tion, modal split and assignment be fitted neatly together, but also each of them
must be free of anomalies and instability. In the past it has often been found that
the process of assignment has been a major source of difficulty.

The purpose of this paper is to compare two methods of multiple route assignment,
one previously described by the present author (reference 1) and one described by
Robert Dial (reference 2). The results given by the two methods in various types of
situation are described and compared.

THE DIAL METHOD

Dial sets out five requirements for his method:

i) The method should give all reasonable paths between a given origin and
destination a non-zero probability of use, while all unreasonable paths
should be given a probability of zero.
ii) All reasonable paths of equal length should have an equal probability
of use.
 230

iii) When there are two or mor~ reasonable paths of unequa' length, the
shorter shoul d have the h"1 gher probabil ;+;y of Lise.
iv) The model's user should have ~ome control over the path diversion
orobabil iti es.
v) The assignment a'goritrm ~hould not explicit'y enumerate paths.

A path is defined as Reasonable if every 1ink on it has ·ts ini:o:al node nearer
to the origin than its final node and the final node closer :0 the destination than
the initial node. (An alternative method, us·ng less computer time is offered, in
which the second restriction is removed.

The ratio of the amount of traffic on a given path to that on the shortest path
between the origin and destinatior:, is determ~ned by the difference in length between
the given path and the shortest ~atr.

The above characteristics of the D';al method provide the essential specification
of what it does.

THE LTNK TIME PERTURBATION METHOD

When findi'lg ordinary s"ngo:e routes thro'Jgr a 'letwor-<, H is supposed that each
1ink has associated witr it a :rip time, and that each tr~p uses the route that mini-
m; ses its tota 1 t';me. (Tre term~ t~melnd 1ength al"e used interchangeably throughout
thi s paper. \

It ~s assumed in the pel"turbation multip:e route model that the user does not
know ':he actual 1ink times, 'Jut that he ass0c' ate5 with every 1ink in the system a
supoosed link t'me, which is drawn at randrnn from a distribut~on of times, having as
its mea'1 :he ac:ual link time, and a mean deviatiop, which might by say 20% of the
actua' 1;nk time. It i~ assumed that the user then finds and uses the route which
m·n"mises the s~m of ~is supposed link times.

T+ is 'lot ~oss;ble ir practice, for reasons of computation, to allow each user

te· have a different set of supposed link times, and thus is necessary to have a group
 231

of users all having the same set of supposed link times. It is convenient to take
as a group the trips generated in a part'cular zone. This enables routes to be found
ir the conventiona1 way by bu~lding trees,

When the trees are plotted and examined, it is found that t~ey g~ve ~outes which
in practice might wel, be used. Although in some cases these are quite d'fferent
from those given by ordinary single route ~rocedures, in other cases, where alterna-
tive routes may not be available, exactly the same route is obtained as is given by
single routing.

It is possible to use the multiple route procedure in a more advanced form, in

which the trips from each zone are divided ~nto several equal parts, each of which
has its own tree built from a different set of supposed lin~ times. ~owever, this
requires additional computer time to build the addit~onal trees and assign traffic
to them.

It must be stressed that, in the form in which there is only one :ree +or each
zone of generat~on, the computer time required is very little differen~ from that
required for ordinary single routing. The only extra ~ime, that required for calcu-
lating the supposed link times from actual link times, is quite ~mall.

Ordinary single routing can of course be considered as a specia~ case of this

type of multiple routing, given by taking distributions of lir.k times WhiCh have zero
deviation from the actual link times, thus causing each supposed link time to exactly
equal the actual time for the link.

In order to apply the method it is necessary to be able to choose at random .~

supposed link time from a distribution which has as its mean the actua' link time.
For reasons of computation, it is found convenient to allow this d~stribution to have
eight possible values. In principle, one might wish to use a Normal D~stribution of
link speeds, but with only eight values it is not possible to represent the tails of
a normal distribution. It has been decided in practice to use a simple rectangular
distribution, with each of the eight possible values having an equal change of being
chosen.
 232

In order that the spread of supposed link times on routes of a given length
shall be independent of the number of nodes on the route, it is desirable that the
ratio of the width of the distribution of supposed link times to the actual link
time shall be proportional to the square root of the actual link time.

PERTURBATION METHOD: ROUTES BETWEEN NODE PAIRS

It will be appreciated that, in practice, the traffic passing between a pair of

nodes will include trips generated from many different zones. Thus the very simple
and economical assumption that there is only one set of supposed link times, and
therefore only one tree, for each zone of generation, in fact allows many different
routes between pairs of nodes. It is this that is the fundamental feature of the
link time perturbation method of multiple routing.

In a large network, the maximum proportion of trips on a link generated from

anyone zone is not often more than 10%, and is usually much less than this on
important roads carrying long-distance traffic.

Thus the effect of the random numbers used in obtaining the supposed link times
for any particular zone of generation has only a small effect on the traffic assigned
to any particular road. It was found, in one case studied, that the average change
in link loadings caused by changing the random factors for all zones of generation
was of the order of 2%.

The main object of an assignment procedure is usually to find the traffic load-
ings on each link. Although in the method here described, the route taken by a par-
ticular trip may be arbitrary, in the sense that it is determined by random numbers,
nevertheless, the total traffic loadings on each road link are insensitive to these
random number. This is a necessary feature of the validity of the method.

In all the examples discussed below, illustrated by Figures 1 to 5, it is

assumed that the traffic passing from A to B is generated from many different zones.
 233

The purpose of the next several sections is to illustrate the results given by
the two methods in various simple network situations, and in particular, to show that
the axioms stated by Dial can have unexpected consequences.

ROUTES OF VARIOUS LENGTHS

It seems to be inherent in the Dial method that the split of traffic between two
routes depends only on the difference between their lengths. Thus for example, if
between a pair of points there are two routes of length 3 and 6, and between a second
pair of points there are two routes of length 30 and 33, then the split of traffic
will be the same in the two cases.

Considering now the perturbation method, it is unlikely that the random numbers
used in obtaining each supposed link length along a route will cause each of them to
take an extreme value from its distribution of possible values. Thus for long routes,
the total of supposed link lengths will , on average, differ from the total of actual
link lengths by a smaller percentage than for shorter routes.

It is thought that this corresponds well to what happens in practice. For

example, on a journey for which the best possible time is five minutes, a person
might choose a route taking seven minutes, which is 40% greater than the best time.
On the other hand, it seems much less likely that on a journey for which the best
time is 100 minutes, a person would choose a route taking 140 minutes. The perturba-
tion method adopted reflects what seems to be the likely actual situation, namely
that on shorter routes people will accept an excess which is greater in percentage
terms, but smaller in absolute terms, than the excess they will accept on longer
routes.

REASONABLE ROUTES

Consider the network configuration shown in Figure 1. Suppose that ED equals EC

and that each of them is slightly less than CD. Then, in reality, it cannot be ex-
pected that any significant amount of the traffic between A and B will use the route
ACEDB.
 234

However, in the Dial method, ACEOB may well be a reasonable route, and may carry
a substantial amount of traffic round the detour CEO.

In the perturbation method, with the sorts of spread of supposed link times that
are normally used, there will be no set of supposed link times which will make the
supposed time for route CEO the shortest route between C and D.

For example, suppose the actual link times are CE = 9, ED = 9, CD = 10. Suppose
also that the spread of link times gives the supposed link times for each of CE and
ED in the range 7 to 11, and the supposed link times for CD in the range 7.9 to l2.l.
Then there is no set of supposed link times giving a total of less than 14 for the
route CEO or a time of more than 12.1 for the direct route CD.

Thus in the perturbation method, the route CEO will never give the lowest sup-
posed time between C and 0 and it will never be used.

UNREASONABLE ROUTES

Consider the network shown in Figure 2. This type of configuration commonly

arises in practice when the route AEOB is a by-pass or ring route. Suppose that
AED is slightly longer than ACB. In reality, if DB is fairly small compared with
AED, it must be expected that some of the traffic moving from A to B will use the
route AEDB.

However, the Dial method will give no traffic on route AEDB, for the link DB
has its final node nearer to the origin than its initial node and thus the route is
classed as unreasonable.

Using the perturbation method, then provided that the spread of supposed link
times on AC and AED is sufficiently high relative to the length of DB, then some of
the traffic will be assigned to the route AEDB.
 235

ROUTES OF EQUAL LENGTHS

Consider the network shown in Figure 3. Suppose that each of the three routes
from A to B is of equal length, and that each of the two routes between A and C is
of negligible length compared with CB. Then it must be expected that, in reality,
about half the traffic between A and B would use the route ADB, and that each of the
two routes via C would carry about a quarter of the total.

However, in the Dial method, if both paths through C are classed as reasonable,
then each of the three routes between A and B will carry one third of the total
traffic, because of the rule that routes of equal length carry equal flows.

Using the perturbation method, then as the actual lengths of the two links AC
are negligible compared with the actual length of CB, the supposed length of the two
links AC will be compared the supposed length of CB. Thus about half the traffic
will find the sum of its supposed link times is minimised by the route through D and
will thus use that route. The other half of the traffic will divide equally between
the two routes through C.

ROUTES OF SHORTER LENGTH

Consider now that in Figure 3 the two routes from A to B via C are equal, and
that each is slightly shorter than ADB. Suppose, as before, that each of the two
routes AC is of negligible length. Then it must be expected that, in reality, each
of the two routes via C will take just over a quarter of the traffic.

However, using the Dial method, if all three routes are classed as reasonable,
then each of the routes via C will take just over a third of the traffic, because of
the rule that a shorter route takes more traffic than a longer route.

The perturbation method gives slightly over a quarter of the traffic on each
of the two routes via C,
 236

PARTIALLY DISTINCT ROUTES

Consider Figure 4, in which each of the three routes from A to B is of equal

length, with CB being of negligible length. Then it must be expected that, in
reality, each of the three routes will carry one third of the traffic.

Consider now Figure 5, in which each of the three routes from A to B are again
equal, but with each of the two routes AC now equal to CB. This Figure can be consi-
deredas intermediate between Figures 3 and 4, and thus it must be expected that, in
reality, the proportion of traffic between A and B that uses the route ADB will be
about midway between a third and a half.

In Figure 5, the Dial method will put one third of the traffic on rou~e ADB,
because of the rule that traffic divides equally among routes of equal length.

It can be shown - either theoretically or by using the ordinary computer pro-

grams - that the perturbation method will put onto route ADB an amount of traffic
which is about midway between a third and a half of the total traffic between A and B

DEFECTS IN DIAL METHOD

The criticis~s of the Dial method expressed in this paper are of a fundamental
sort, in that they require the rejection of his first three axioms, namely those on
1) reasonable and unreasonable routes, 2) routes of equal length, and 3) routes of
shorter length.

Multiple routes are a more subtle and paradoxical matter than may appear at
first sight. Perhaps the Dial axioms would be acceptable in cases where routes are
wholly distinct from each other, but in practice this is rarely true in road networks
For a multiple route method to be realistic, it is essential that it is able to deal
with cases of partially distinct routes, such as those shown in Figur~ 5.

Looking at the Dial method in a broad perspective, perhaps its greatest failing
is that it lacks robustness. In other words, small changes in input data can cause
large changes in output data. Thus in the section above which discusses routes of
 237

equal length, illustrated by Figure 3, tbe flow on ADB will be changed from one
third to one half of the total if a tiny change is made in the network coding by
deleting one of the two routes between A and C.

Perhaps an even more serious example of the lack of robustness of the Dial
method arises from the definition of Reasonableness. When a multiple route method
is used as part of a capacity restraint procedure, there are changes in link times
from one stage of the iteration to the next. Thus for example, in Figures 1 and 2,
a route which is reasonable at one stage may be unreasonable at another stage. There
are cases in which this causes severe instability.

CONCLUSIONS

Although the examples used in this paper to illustrate the two methods are of an
extremely simple sort, it must be emphasised that situations of these types occur
frequently in real networks. Only if a method deals accurately with simple networks
can it be trusted to deal adequately with those that are complex.

In the view of the present author, the Dial method has only one advantage over
the perturbation method, namely that it is fully determinate. However, although the
results produced by the perturbation method are affected by the random numbers which
happen to be used, the effect of these numbers on the link loadings is found to be
slight.

The perturbation method has been used in a great many studies in various coun-
tries, and has been found to be simple to operate, and economical in computer time.
It works well, not only in unrestrained applications, but also as part of capacity
restraint procedures, and of other complex iterative models.
 238

REFERENCES

1} Burrell, J.E., Multiple Route Assignment and its Application to Capacity

Restraint. Fourth International Symposium on the Theory of Traffic Flow,
held at Karlsruhe, June 1968. (Proceedings published as Heft 86 of
Strassenbau und Strassenverkehrstechnik.)

2} Dial, Robert B., A Probabilistic Mu1tipath Traffic Assignment Model which

Obviates Path Enumeration. Transportation Research, Vol. 5, No.2, June
1971.
 239

Figure 1. .
A ..._ _ _ _ _ ~
~----~--
B

C 0

Figure 2.

A B
Figure 3.

Figure 4. A B

A,
Figure 5.
 EMULATION OF DYNAMIC EQUILIBRIUM IN TRAFFIC NETWORKS

Sam Yagar
University of Waterloo
Waterloo, Ontario, Canada

ABSTRACT

Some notions of equilibria, both static and dynamic, are discussed relative to
traffic assignment. Existing types of assignment techniques are examined. Their
relative merits and deficiencies in predicting the self-assignment of drivers to a
road network according to the principle where each individual minimizes his own
travel time are discussed. A technique is proposed, which combines the incremental,
iterative and queueing procedures employed in previous techniques. For each time
period with its assumed time-stationary demands, a sequence of iterations is per-
formed, each consisting of a complete incremental assignment. In each iteration the
final link costs from the previous iteration are given a relative weight W, with the
costs normally used in the incremental technique having a weight of l-W. This trades
off some of the multipath capability of incremental techniques in order to avoid
irreversible assignments to ultimately irrational paths in the early increments.
Respective weights for l~W of :~~ were found to work sufficiently well for a real
network in Ottawa, and the technique formed a basis for an evaluative model which was
applied to that network for testing traffic control schemes.
 241

Introduction

Traffic assignment techniques attempt to predict the flows and queues

in a traffic network. They generally assume Wardrop's First Principle, that

each traveller chooses his path in such a way as to minimize his own travel

cost. Travel time is commonly used to represent travel cost, as discussed

in the Appendix. The principle of individual path minimization is conceptually

simple and its rationality generally appeals to those modelling the network

loading phenomena. The optimal paths of individuals are often mutually

dependent. However, drivers faced with the same situation on a repetitive

basis will generally converge to an equilibrium solution, i.e. the actual

flows and queues in the network tend toward equilibrium as a result of the

drivers competing to optimize their own paths. Analysts do not seem to be as

successful in obtaining this equilibrium as are the drivers themselves. This

is partly because we have tended to model these dynamic processes as if they

were static. Also the search for a practical technique for emulating

Wardrop's First Principle even for static processes has taken various directions.

In order to emulate equilibrium we must first understand it.

Notions of Equilibrium

A theoretical equilibrium assignment can be defined as a saddle point

type of solution where no user can better himself by deviating from the

assignment unless others also deviate. This solution has the characteristic

that each path which is used between any given pair of points has a cost

which is no greater than any other path between those points. An equilibrium

solution need not be unique. Further to this, it may be dynamic, with flows

and queues varying with time due to varying demands and/or link characteristics.
 242

For the time-stationary case, equilibr!um can be formulated mathematically

as that network loading process which minimizes the cumulative individual

marginal costs encountered. Although very difficult to reduce, this

formulation indicates that a user can not decrease his cost by deviating

from that solution, but may incur an additional cost. Iterative and

incremental algorithms give solutions, but have generally failed to produce

equilibrium assignments. Some reasons are discussed later.

In a theoretical equilibrium situation each driver can be treated as the

marginal user in the network. Each driver should be rational in his choice

of path. A system is generally stable against the actions of relatively few

irrational drivers. The rational drivers can be thought of as the marginally

assigned demands, exploiting any potential inequity in travel costs caused by

the irrational drivers. This tends to bring the link costs into equilibrium.

Background on Assignment Techniques

Table I illustrates a hierarchy of existing types of assignment

techniques. No attempt is made herein to compile a bibliography of assignment

techniques or give an exhaustive review of the many techniques which have

been proposed. Rather, the background portion of the paper is restricted to

the information which is needed as a build-up to the work described in this

paper. Only the above general types of assignment techniques are discussed,

with example methods referred to when it is felt appropriate.

 243

Table 1

Hierarchy
of Techniques Characteristics

1. Desirabilty Demands, capacities and link costs are

all constant.

2. Constrained: Demands and capacities are constant.

(i) Incremental Link costs are flow-dependent.
(ii) Iterative
(iii) Multipath

3. Dynamic Link costs pre flow- and queue-dependent.

Demands and capacities are time-
dependent.

The simplest form of assignment, by desirability only, is useful for

networks whose travel times are independent of assigned flow. With this

method the assignments of demands between each origin-destination pair are

made on the basis of the tree of minimum paths from each origin to all

destinations. The minimum path trees can be obtained quite easily from the

early tree algorithm of Dijkstra (1959). Although travel times are rarely

independent of assigned flows, an assignment by desirability indicates an

ideal network which would give each user a choice of paths that is not

constrained by the flows of other users. It can, therefore, be used as a

planning tool for indicating optimal choices of location and amount for

capacity additions to a network.

Constrained assignment methods have been proposed to model existing

traffic flows in real networks where travel times are flow-dependent and

flows are themselves restricted by capacity. These can be sub-classified

into incremental, iterative and multipath techniques.

 244

The basic philosophy of the incremental methods is to assign

fractions of the total demand matrix, increasing the costs on some of

the links after each incremental assignment, so that the costs reflect

the cumulative assignment of flows to each link. Typical variations

of this general method are those of Steel (1965) and Homburger (1969).

Their philosophy is the same, that of increasing the travel time on a

link as the flow increases, but they differ in their procedures. Homburger

increases the travel time on a link by a preset amount when its assigned

flow reaches a given value, and eliminates the link when its assignment

reaches capacity. Steel assigns a set percentage of the demand at each

step and then updates the travel times. Although their methods vary only

in computational procedure they will generally yield different results.

Another variation of the general class of "constrained" assignment

techniques is that of Yagar (1971), which restricts flows at capacity but

allows further assignment into queues at estimated costs which include time

in queue. It is a dynamic method which can treat time-varying demands

having short-term peaks through its capability for storing any excess demands

in queues until they can be served. This technique will be pursued further

when the proposed method is synthesized.

Iterative techniques such as the Wayne Method of Smock (1963) and

Smulick (1965) assume a set of costs for the links and assign all demands

by desirability. On the basis of the assumed link costs and the costs

corresponding to the desirability assignment, a new set of costs is

calculated and the assignment repeated in another iteration. The iterations

continue until convergence is obtained. Convergence may be defined as a

 245

state where all of the flows obtained do not differ significantly from

those of the previous iteration, or as a state where the flows obtained

on the links correspond to the costs assumed for those links. The latter

state will be referred to as a state of equilibrium. The former is easier

to obtain, as will be illustrated in a simple example, but is not necessarily

meaningful. Although Overgaard (1966) found the iterative method to converge

to an equilibrium state in a single test case, there is no theoretical proof

that it will indeed do so. In fact, example 2 which will follow is a

simple illustration that convergence to an equilibrium state is not

achievable in some cases.

The common multipath method allows the user a choice of greater

than one path in any iteration. It does not increase the cost on a link

incrementally during an iteration. By assigning all of the volume to a given

link at the same speed, it avoids the major disadvantage with incremental

assignments, that of permanent assignment to unrealistic paths. The Toronto

Method of Irwin, Dodd and von Cube (1961) allows up to nine alternate routes

for each O-D pair, each of which is generated in an iteration whose travel

costs are based on the assignment of the previous iteration. This method

has combined multipath-and equilibrium-type capabilities. The method of

Dalton and Harmelink (1972) introduces a normally distributed random component

into the link costs to overcome the problem of generating trivial alternate

paths which are minor deviations from one another. The random component also
 246

tends to decrease the drastic all-or-nothing type of effect caused if

piecewise-constant approximations to the unit cost functions are used.

Their method uses an iterative procedure for improving estimates of speed.

It has provision for counting only the final iteration, or for weighing the

assignments of the various iterations. The authors refer to the latter

option as an incremental procedure. This option has the common disadvantage

typical of incremental loading procedures, that any incorrectly loaded

demands are permanent. This is discussed later.

Discussion of Incremental-Type Methods

These methods provide a multipath capability consistent with the

first principle of Wardrop in that the basic shortest-tree algorithm can

alter its choice of best path between a given pair of points as the link

travel times change. However, a change is made only when the flow reaches

a level which is inconsistent with the cost used.

•
It is never made just to

spread the load around~ The assignments may flip back and forth between a

pair of paths as their travel times take turns increasing. The net effect

is that the travel times tend to equalize on those parallel paths between a

pair of pOints which are used for trips between that pair of points.

The incremental methods have the following desired effects:

(i) They find all optimal paths between each pair of points.

(ii) All O-D pairs can be considered simultaneously in the

assignment process, so that there is not a bias in favour

of certain nodes.

(iii) The incremental loading and flow-updating procedures employed

by this general type of method make it conducive to a straight-

forward simulation of the queueing which takes place after

 247

capacity flow has been reached on a link, such as employed by

Yagar (1971). This is important in operations characterized

by short-term excesses of demand over capacity, as occur in

urban traffic networks during peak periods. The modelling

of path choice and point-to-point travel times is seriously

hampered if there is no consideration for the peaking

characteristics of demand and their resultant effects of

queueing and increased delays.

Incremental methods have the major disadvantage of permanently

assigning" flows on the basis of costs which reflect the initially low flows.

These increasing flow-dependent costs are used for the purpose of loading

the network in such a manner that the lowest cost path between each pair of

points is chosen in the light of the demands already assigned to the various

links. The disadvantage is that the basic incremental technique does not

have provision to consider equilibrium costs in the path finding procedure.

That is, during the network loading process the model considers users to be

making their path choices on the basis of the artificially low flow and cost

levels which are used in the network loading process rather than on the more

realistic equilibrium values that each user would really encounter. The

real user knows how the loaded network behaves, but this information is not

available to the simulation process during the artificial network loading

process. The process could use some form of information feedback (really a

feedforward) mechanism. For in the real equilibrium situation all of the

drivers are effectively simultaneously marginal users. This leads to a

very complicated game-type process which is very difficult to model

accurately.
 248

As noted above an incremental assignment technique can assign users

at the beginning to paths that they would not take in the equilibrium

situation, i.e., paths that will ultimately be other than their minimum cost

paths. This gives a type of multipath effect, one of using greater than one

path between two points, even though the costs of the paths between the two

points differ. It therefore has the effect of using illogical non-optical

paths. This is illustrated by example 1 using the method of Hamburger which

is described by Yagar (1971). In Fig. 1 links (1,2) and (2,4) represent

freeway links while (3,2) is an on-ramp and (3,4) is a surface street route.

The dashed link (2,3) does not exist for this example. The demand has 2000

units wishing to go from 1 to 4 and 1000 from 3 to 4. Without loss of

generality, links (1,2) and (3,2) can be given zero cost and infinite

capacity. Link (2,4) has a low cost for low flows and high cost for high

flows, the latter reflecting queueing. Link (3,4) is not as sensitive to

flow in the range of flows that it might encounter. For simplicity, the

cost vs flow curves are each approximated by o~ly two piecewise-constant

components as shown in Figs. 2(a), (b), respectively. All of the demand

from node 1 must use links (1,2) and (2,4) while that from node 3 can use

either the pair of links (3,2) and (2,4) or just link (3,4). Initially it

prefers (3,2) and (2,4) with a cost of 1 as opposed to the alternative cost

of 2. Since all of the 3000 units want link (2,4),which can accept only
2000
2000 at the initial cost of 1, a fraction equal to 3000 of each of the

attempted assignments is made; 1333 from node 1, and 667 from node 3. Now

the cost on link (2,4) increases to 5 and the remaining demand from node 1

must still use it, while the remainder from node 3 takes link (3,4) which

still has a cost of 2.

 249

When all of the demands have been assigned link (2,4) has a flow of

2667 and cost of 5, while link (3,4) has a flow of 333 and cost of 2. The

cost of the path from node 1 to node 4 is 5 units, while the two paths which

are used from node 3 to node 4 have respective costs of 5 and 2. The 667

units assigned from 3 to 4 by way of link (2,4) would really rather use

link (3,4) at the cost of 2. In fact all of them could use (3,4) with the

cost remaining at 2. In this case, link (2,4) would only have the 2000

units from 1 to 4 with a unit cost of 1, and this would be an equilibrium

solution as no user would be able to better himself by changing paths. In

fact, for any demand from 1 to 4 of at least 2000 units a simple equilibrium

solution can be obtained by inspection, as no user travelling from 3 to 4

will pay 5 units to use link (2,4) according to the principle of individual

path minimization.

It is seen that if those travelling from 3 to 4 are provided with some

foresight into the equilibrium travel costs, they will not pay complete

attention to the temporary unit costs used in the network loading process.

An iterative method for providing this insight to the model using the costs

at the end of each previous iteration is proposed later.

Discussion of Iterative Methods

The major problem encountered with iterative techniques is that even

if such a technique converges it will not necessarily converge to flows which

predict the actudi flows. The reason for this is that the method assigns all

of its demand between any pair of points to a single path. Therefore it can

turn out that the resultant travel times on two paths between a pair of

points are different, with the utilized path having the higher cost. This
 250

is in contradiction to Wardrop's principle that each user chooses his

minimum cost path. Some iterative methods are therefore only convergence

techniques and not necessarily equilibrium techniques. Since the sinr,Ie-path

convergence method works with only one cost value for each path in any

given iteration it has the consolation that all of the trips between any

pair of points will have the same, though not necessarily realistic, cost.

In real networks the equilibrium situation ,where no person can improve his

trip time by changing paths,commonly has more than one path used between

some pairs of points,with the parallel paths having approximately equal

travel times.

It is now convenient to illustrate the problem of convergence to an

irrational set of flows with the iterative assignment technique. The

example network described by Figs. 1, 2 is used again for example 2, but

this time the dashed link (2,3) in Fig. 1 is included and has zero cost.

Suppose that the initial iteration uses respective costs of 5 and 3 for

links (2,4) and (3,4) respectively. That iteration will assign the 3000

units of demand to link (3,4) and none to link (2,4). Depending upon the

logic used to estimate link costs for successive iterations, the total

demand will continue to oscillate between links (2,4) and (3,4) or will

ultimately converge to link (3,4). Suppose that for the second, third, etc.,

iterations, the cost for each link be taken as the average,over all of the

previous iterations,of the final costs for those iterations. The final

cost for each iteration is that cost value from Fig. 2 which corresponds

to the flow assigned to the link(s) in that iteration. When the flows for

two successive iterations correspond, convergence is said to have occurred.

In iteration 1 the flows on links (2,4) and (3,4) are 0 and 3000 respec-

tively. The result is that the total flow of 3000 units is assigned to
 251

link (2,4) in iteration 2. This oscillating effect continues until the

assigned flows of iterations 5 and 6 are found to be equal, at which point

the process is defined to have converged to flows of 0 on link (2,4) and

3000 on link (3,4). The results of the iterations are illustrated in

Table 1.

TABLE 1

Iteration
1 2 3 4 5 6

Assumed Cost 5 0/1 5/2 5/3 10/4 10/5

Link (2,4) Assigned Flow 0 3000 0 3000 0 0
Resultant Cost 0 5 0 5 0 0

Assumed Cost 3 3/1 3/1 6/3 6/4 9/5

Link (3,4) Assigned Flow 3000 0 3000 0 3000 3000
Resul tan t Cos t 3 0 3 0 3 3

As in example 1 which was used in discussing incremental methods, a

true equilibrium solution, characterized by no driver being able to better

himself by deviating from the assignment, would have had 2000 units on

link (2,4) and 1000 on link (3,4). The proposed method which will be out-

lined later in the paper accomplishes this aim for both examples. It is

interesting that the incremental method would have given an equilibrium

solution to example 2, while the above iterative method would have done the

same for example 1, on which tHe incremental method was employed.

 252

Dynamic Equilibrium

Equilibrium in traffic assignments has generally been treated as

being static. This seems to have some intuitive appeal, as the stability

of an equilibrium state is easily confused with time-stationarity. The

actual stability of an equilibrium process stems from the fact that no

individual is able to benefit by personally deviating from the equilibrium

pattern. However, an equilibrium process need only remain in any singLe

state for as long as conditions remain the same. Traffic demands are

generally ti~e-dependent, with the equilibrium in a constant state of

transition. The flows and queues in a network may remain in equilibrium

while their values change over an extended period of time due to time-

varying demands and/or network characteristics. A prime example of this is

road traffic, where the demands are generally time-dependent, so that any

equilibrium ~akes en a dynamic form.

A dynamic form of equilibrium can be defined in terms of a saddle

point solution with time-dependent demands. The added complexity introduced

by time variation makes it more difficult to attain an equilibrium solution.

However, dynamic equilibrium is readily described in terms of a sequence of

short-term static equilibrium states, each corresponding to a portion of

the transient time period. This technique was employed by Yagar (1971) to

simplify the treatment of time-varying demands, and contributes to a

proposed technique which is described below.

Proposed Method - Combining Incremental and Iterative Proceduces

As discussed previously, the incremental assignment procedure may

assign users to paths which turn out not to be their shortest paths at the
 253

end of a complete iteration of incremental assignments. Therefore it is

difficult to achieve an equilibrium assignment by applying the simple

incremental technique. It is proposed here that an equilibrium state may

be approached in some cases by iterating on successive incremental assign-

ments. In each iteration an estimate of the equilibrium travel cost is

combined with the (increasing) costs of the incremental assignment. The

travel costs at the end of the previous iteration are used as estimates of

the equilibrium costs for the present iteration. The incremental technique

is chosen as the network loading device within each iteration because of its

simple provision for queueing. The iterative procedure is added to minimize

the premature loading to irrational paths. To allow for the assignment of

time-varying demands the queueing procedu.re of Yagar (1971) is used. The

following total method results:

(1) The simulated time period is broken up into time slices (of

the order of 15 minutes) which are short enough that the demand

can be considered to be stationary with respect to time. The

remaining steps (2) to (4) apply to each time slice.

(2) The demand which was queued in the previous time slice is

added to the exogenous demand of the present time slice to

obtain the total demand wishing to be served in the present

time slice.

(3) The total demand is assigned to the network by an incremental

technique which restricts flows to capacity levels but allows

the amount of demand assigned to any path to exceed capacity,

with the cost used for the path reflecting the cost of

queueing. Excess demands are queued at the first point of

congestion in their path.

 (4) When the first iteration of total incremental assignment

has been completed the final link costs Xi corresponding

to assigned flows and queues are noted and the assignment

is erased. A second iteration of incremental loading is

performed, followed by further iterations as deemed

necessary. For the second and succeeding iterations, the

costs Xi at the end of the immediately previous iteration

of incremental assignment are weighted in with the costs

corresponding to partially assigned flows during the

present iteration. Denoting the former by Xi and the

latter by Yi , with the weight to the former equal to W,

the unit cost for link i in the succeeding iteration is

W,Xi + (I-W).Yi ·

It is noted that Yi may increase during the iteration. Also 0 ~ W ~ 1.

W = 0 gives a straight incremental assignment, while W = I gives a single-

path iterative assignment. The single-path iterative technique is commonly

known as an all-or-nothing assignment method. The disadvantages of each

have been discussed and were illustrated in examples land 2 respectively.

The level of compromise between the two extremes is governed by the choice

of W. Iteration 2 in example 1 yields the equilibrium solution for W > .25,

the incremental solution for W < .25, and something in between for W = .25.

Iteration 2 in example 2 gives the equilibrium solution for W < .75, and

an all-or-nothing solution for W> .75.

Although the proposed method was observed to yield multipath equili-

brium solutions in the above simple examples there is no guarantee that this

will be accowplished in any given application. However, it can be seen

 255

that there is a trade-off between the levels of achievement in terms of

(i) finding all of the competing paths, and (ii) assigning flows to their

ultimate shortest paths. The balance in this trade-off is governed by the

choice of the weight W given to the estimated equilibrium costs.

In an application to the Ottawa Queensway corridor, Yagar (1973), W = .20

and W = .40 were found to give reasonably good results, with the latter perhaps

better. This might be used as a starting rule-of-thumb, although the optimal

value of W will depend upon the relative cost vs flow relationships of

the links in one's own network. Although an analysis of the effects of

varying W for the above corridor would have been interesting, it was not

feasible to do so at the time.

It is also possible to weigh the travel costs of the previous time

slice into the first iteration of the present time slice if desired. This

may have merit for some networks and not for others. It tpnds to inhibit

stability of results when the demands vary greatly from one time slice to the

next. A test was performed in which, for each time slice, costs from the

previous slice were weighed in but only one iteration was used. An unstable

solution was obtained which was not considered worthy of illustration.

Testing and Application of the Method

The method was tested on the Ottawa Queensway Corridor and found to

predict actual flows reasonably well. The number of iterations used in

each time slice was constrained between 4 and 6 inclusive. The model

chose to use 4 iterations for each time slice. In spite of the fact that

many of the network users were sensitive to small variations in estimated

trip times the results are encouraging. For purposes of comparison the flows
 256

on a small but critical part of the corridor - the freeway and its ramps

in the most congested section - are illustrated in Figs. 3-5 for the five

peak l5-minute time slices (slices 2-6) from 7:15 a.m. to 8:30 a.m. Figure

3 shows the observed demands at the freeway links. These demands are

reflected in flows and queues (queues in brackets), their relative values

having been altered to preserve a constant capacity - e.g., 950 veh/15 mins.

for link (43,44). The queue has been allowed to take up the actual capacity

variation (e.g., if a flow of 968 and a queue of 10 were observed, this was

adjusted to a flow of 950 and a queue of 28, as some fixed capacity had to

be defined for link (43,44».

A set of O-D demand matrices, one for each time slice, was obtained

from the counts and an O-D postcard survey. These demands were assigned to

the network with a straight incremental technique which corresponds to W = 0,

and the results are shc~~ in Fig. 4. The proposed technique was then applied

to the same demands and network but with W = 0.40. A total of four itera-

tions of the incremental technique were' performed in the latter case. The

results of the assignment are shown in Fig. 5. The model's ability to

predict was deemed sufficient for it to be used in subsequent testing of

network improvement and control schemes.

Discussion of Test Results

Although both methods seemed to predict actual flows and queues

reasonably well, the proposed iterative-incremental method was found to be

superior to the simple incremental technique. The former gave an average

deviation from observed values for flows entering and leaving the freeway

of 16.5 vehi15 mins., while the latter gave an average deviation of 21.5.

The former also predicted queues much better. It is interesting that the
 257

former tended to shy away from the on-ramp at node 40 between 7:15 a.m.

and 8:00 a.m., using a parallel path instead, while the latter tended to

over-assign to that ramp during this period. The travel times on the

freeway and main alternate path are nearly identical in this period. The

distances are also about equal. All of the deviation from observed counts

at that ramp could therefore be due to the aggregation at a single origin

point of all users originating near node 40. The best path for some of

them is via the freeway, while for others it is via the alternate path.

The shortest path from that concentration point to downstream destinations

was quite sensitive to the incremental technique's variation in the relative

trip costs used for freeway and alternative paths.

Discussion of Other Possibilities for

Multipath Equilibrium Assignment

The multipath methods of Irwin et al (1961) and Dalton and Harmelink

(1972) do not presently have the capability to deal with queueing situations,

but this type of extension could perhaps be added. First the capacity

constraints would be eliminated for the assignment portion of the procedure.

Costs would be assigned to each link in each iteration on the basis of the

link assignments at the end of the previous iteration. These costs could

include queueing if any link assignments exceeded capacity. However, when

flows in excess of capacity are assigned to a sequence of links on a common

path, only the link(s) with the actual queueing should reflect this cost,

(i.e., if the vehicles are queued on the first link, the flow on the

remaining links would be governed by the output of that first link). This

problem of representing queue costs is compounded by the fact that vehicles

on different paths may share a link. If this problem and others that may
 258

crop up can be solved, the multipath method might be made superior to

the one described herein.

Conclusions

The described assignment technique combining the incremental loading

and iterative procedures seems to have the ability to predict flows

sufficiently well for practical purposes. It appears to be an improvement

on the straight incremental technique, especially on critical links where

there is queueing. Links with queueing are quite volatile in terms of

travel time and are very susceptible to early acceptance of trips for

which they would not be able to compete in the equilibrium situation. This

major deficiency of the incremental techniques has been countered by th~

proposed method. Although it has been found to give satisfa~tory results

there is no theoretical proof that it will converge to equilibrium solutions.

It is felt that further research into dynamic assignment methods would be

worthwhile. The time slice and queueing approach is promising. Although

the secondary effects such as queue spillbacks are logically modelled by

techniques based on incremental loading, other techniques merit some

attention, especially for non-roadway assignments. It is conceivable

that the multipath equilibrium techniques could rival and even supersede

the proposed method if they could be given the provision to handle queueing

so that they could accept time-varying demands.

 259

REFERENCES

B"nesh, A.H. (1972), "Traffic Assignment Witll0ut Bias", HRR 392, pp. 36-46.

D,d.ton, P.M., Harmelink, M.D. (1972), ''Development and Testing of a Mu1ti-

path-Assignment Technique", HRR 392, Abridgment, pp. 136-138.

Dij ks tra, E. W. (1959), "A Note on Two Problems in Connexion with Graphs",
Numerische Math 1, pp. 269-271.

Homburger, W.S. (1969), "Traffic Estimation-Computer Programs for

Educational Purposes", 2nd edition, ITTE, Course Notes, University
of California, Berkeley.

Irwin, N.A., Dodd, N., von Cube, H.G. (1961), "Capacity Restraint in
Assignment Programs", HRB Bull. 297, pp. 109-127.

Moskowitz, Karl (1956), "California Method of Assigning Diverted Traffic

to Proposed Freeways", HRB Bull. 130, pp. 1-26.

Smock, R. (1963), '~ Comparative Description of a Capacity Restrained

Traffic Assighment", HRR 6.

Smulick, J. (1965), '~Mathematical Model for Traffic Assignment",

M.A. Thesis, Wayne State University.

Steel, M.A. (1965), "Capacity Restraint - A New Technique", Traffic Eng.

Control 7, pp. 381-384.

Wardrop, J. G. (1952), "Some Theoretica 1 Aspec ts of Road Traffic Research",

lnst. Civil Engrs. Road Paper 36.

Yagar, S. (1971), ''Dynamic Traffic Assignment by Individual Path

Minimization and Queueing", Transp. Res., pp. 171-196.

Yagar, S. (1973), "An Analysis of the Morning Peak Traffic Problems in

the Eastbound Corridor Serving Ottawa", report submitted to the
Ministry of Transportation and Communications of Ontario,
Downsview, Ontario.
 260

Fi~. 1 A Fvoothetical SiJ11u1e Net\10rk -- Dashed Link (2,3) is Used

for Example 2 But Not for Example 1.

51--------------- ,..-----

I
~
2000 f

(0) FREEWAY LINK (2,4)

3 t------~----

2t-----~

I
I
I
I
1000 f
(b) SURFACE STREET LINK (3,4)

Fi~' s 2 r.ost-flOH Pe1ationshios for the FVDothetical Net',1Ork

of rig. 1.
 261

37 31
TIME \.
SLICE :58Q.h.~~~
2 -~L8A 3~~~
104 39

(140)
3 41~
187 \45 /225(2)

4
24

89 78
14

185
( 102)
'12 725.

35
r 950~

225 (12)

22 14

37 26

Fig, 3 De..-uands Counted on the nttmva ()ueensway between 7: 15 am and

8.30 am -~ queues (in Brackets) Have Been Adjusted to Give
a Downstream Capacity of 950 Veh/15 Mins.
 262

62 48
(119)

950~

82 27 33 152 (15)

36 30

[1.L
(180)
3 754
40

169 \94 182 (20)

106

4
49

81
.~
68 33

55
29

177 90
(160 )

42 727. r 950.

223 (30)

42 25

114 56 130 71 240(27)

59 50

115 45 132 117 238

Fig. 4 Flows and Queues (in Brackets) Predicted by the Increwental

Apsignment Technique Extended to Provide for (\ueuing -- for
Comparison with Observed Values of Fig. 3.
 263

TlME*
SLICE :597
2 37
36

633
~ 38~--+{
o

82

(125)

4 <12 72 0. ~ 95"-

19 220 (25)

Fig. 5 Flowsand Oueues (in Brackets) Predicted by the Pronosed

Technique -- for Comparison with Observed Values of Fig. 3.
 264

APPENDIX

Rationalization for Choosing Paths Solely on the Basis of Travel Time

Each user chooses his own preferred path. His choice may be based

on time, distance, comfort and any other factors. The total cost function

will vary with the individual, and would likely defy quantification even if

a typical individual could be defined. It is generally narrowed down to no

more than the two factors, time and distance. For the simple choice

between a freeway and arterial road linking a single pair of points, a

diversion curve such as the California Method of Moskowitz (1956) which

considers both distance and time may be useful. Benesh (1972) has given

an example where traffic assignment procedures using the shortest time

path tend to over-assign to high speed facilities. He suggested a pair

of formulae, based on both time and distance, which for his particular case

reproduced the counted volumes with reasonable accuracy. This is one of

many possible ways of weighing in distance and other factors with average

or "non-incident" travel time. The decision of some drivers not to use

the "faster" freeway may be linked more to the fact that its travel time is

affected greatly by incidents such as accidents and they must have a reliable

trip time.

In any case, in dealing with more complicated networks having many

origin-destination pairs, where a given O-D pair may be satisfied by a

number of paths, it is preferable to reduce the travel cost to a simple

function and if possible a single variable which is easy to measure. Since

detailed assignment is required mainly for situations where capacity is

critical, with queueing and delays common, trip time tends to be the dominant

component in trip cost and is therefore used herein to represent cost.

 A MODEL AND AN ALGORITH~l FOR THE
DYNAMIC TRAFFIC ASSIGNMENT PROBLEM

Deepak K. Merchant
University of Rochester
Rochester, New York, USA

George L. Nemhauser l
Cornell University
Ithaca, New York, USA

SUMMARy 2

ABSTRACT

A discrete time model is presented for dynamic traffic assignment. Congestion

is treated explicitly in the flow equations. The model is a nonlinear and non-convex
mathematical programming problem. A piecewise linear version of the model, with addi-
tional assumptions on the objective function, can be solved for a global optimum using
a one-pass simplex algorithm--branch-and-bound is not required. The piecewise linear
program has a staircase structure and can be solved by decomposition techniques or
compactification methods for sparse matrices. The nonlinear model, with the same
additional assumptions on the objective function, has the property that all local
optima are global.

Supported by National Science Foundation Grant GK-32282X to Cornell University.

2 Full paper was issued as Technical Report No. 247, Department of Operations
Research, Cornell University, Ithaca, NY 14853 and as Working Paper No. 7504,
Graduate School of Management, University of Rochester, Rochester, NY 14627.
 266

Summary

In the traffic assignment problem we are given a road network and a set

of ordered pairs of points on the network that are called origin-destination

(OlD) pairs. For each OlD pair (i,j), there is a given function Rij(t),

o~ t ~ T, where R.• (t) is the rate at which vehicles leave i at time t

~J

to go to j, and T is the planning horizon. The assignment problem is to

determine the traffic pattern or flo~s on the links of the network satisfying

specified optimality or equilibrium conditions.

When, for all OlD pairs, R•• (t) is a constant function over a long
~J

horizon, the assignment problem is called static. There has been an abundant

literature of models and methodology for the static problem. Unfortunately,

the static assumption that the external .inputs do not vary with time may not

be applicable in many realistic traffic situations.

We develop a discrete-time macro-mode~which is a nonlinear, nonconvex

mathematical programming problem. The model treats congestion explicitly

in the flow equations. It has the capability of handling multiple origins

and general topology, but is limited to one destination. In a macro-sense,

the one destination situation may be reasonable for the morning home-to-work
rush period.

We represent the traffic network by a directed graph G = (N,A), where

N is the set of nodes and A is the set of directed arcs (multiple arcs

joining a pair of nodes are permitted). The planning horizon is divided into
equal time intervals of suitably small length and we denote the individual

periods by {iii = O,I, ••• ,N}. The external inputs are approximated by a

constant for each time period and Fi(q) denotes the external input at node q

for the ith time period. Each arc j is assigned a cost function hij and an

exit function g .• When x is the amount of traffic on arc j at the beginning

J
 267

of time period i, we assume a cost h .. (x) is incurred and during period i

1.)
an amount of traffic g.(x), o g.(x) _< x, exits from arc j. (Since
) -< )
represents a physical phenomenon, it is assumed not to depend on time, though

such a generalization poses no difficulty.) To represent the physical process

of congestion~ {gj} are nondecreasing, continuous, concave functions, and

to represent the disutility of congestion, {h.} are nonnegative, nondecreasing,

)

continuous, convex functions.

The fundamental transformation equations (also known as state or coupling

equations) are

x1..+1,). = x1.)
.. - g.(x .. ) + d .. ,
) 1.) 1.)
i = 0, ... ,N-l, (1)

where the state variable x .. is the total flow on arc j at the beginning
1.J
of the ith period, and the decision variable d .. is the amount of flow that
1.J
is admitted onto the arc j during the ith time period.

Let A(q) = {j £ Alj points out of node q} and B(q) = {j £ Alj points

into node q}. Then for each node q (q # sink), at each time period, we

have the conservation equation

L d ••
1.J
F.(q) + L g.(x .. ). (2)
j£A(q) 1. j£B(q) J 1.J

Let a = IAI denote the cardinality of A. Assuming that there is a

single destination or sink (node n), which corresponds in the terminology of

network flows to a one-commodity problem, the N-period, single-sink problem,

denoted by P, is
 268

minimize
N
Lr
a

i=l j=l
h .. {x .. )
~J ~J
(P.O)

subject to

X~+l,J' = x .. - g.(x .. ) + d .. , i = 0, ... ,N-l, V j e: A (P.l)

- ~J J ~J ~J

L d .. = F.(q) + L
gJ'(x iJ·), i O, ..• ,N-l, Vqe:N\{n} (P.2)
je:A(q) ~J
~ je:B(q)

xOj = R. given, V j e: A (P.3)

J

d .. 0, i = O, ... ,N-l, V j e: A (P.4)

~J ->

x •• > 0,
- i = 1, ..• ,N, V j e: A. (P.S)
~J

In problem P, the equ.ations (P.l) correspond to (1), the equations (P.2)

correspond to (2), (P.3) are the given initial conditions, and (P.4) and (P.S)

are the natural nonnegativity constraints. We have not explicitly included

upper bounds on the flows. We prefer instead to handle capacities implicitly

by having dg.(x)/dx =0 for large x.
J

Cost Function Assumption (CFA): We say that {hij(x)} satisfies CFA when

a. hij(x) is nonnegative, nondecreasing and convex for all i and j (as

above), and,

b. the right-hand derivative ah:.(x)/ax exists for all i,j and x, and
~J

for each simple path containing arcs jl and j2 such that j2 is the arc

closer to the sink, sup[ah:. (x)/ax] < inf[ah:. (x)/ax] for all i.
x ~J2 - x ~Jl
Two cost functions of practical value that satisfy CFA are

(i) hij(X) = x for all i and j.

i =N
(ii) h .. (x) -- {XO"
~J
otherwise.
 269

With the objective function specified by (i), the total cost is the number

of vehicles summed over all time periods and with (H) the total cost is the

number of vehicles that have not reached the destination by the end of the

planning horizon. crA imposes larger costs on arcs further from the sink. The

effect of this cost structure is to produce decisions that tend to move the flow

towards the sink as quickly as possible, which is a quite natural goal for the

system.

With computational efficiency as the main goal, we consider a piecewise

linearized version of P. Correspondingly, for each j g A, we partition

k k
the nonnegative segment of the real line into K(j) subintervals [R.j,u j ],

where
I
R.. = 0, u.KC")
J = 0> and R.~ = u.k-l for k = 2, ••. ,K(j). In the
J J J J
k k k
interval [R.. ,u.], we approximate g. by an affine function with slope gj'
J ~ J
k k+l k
where 0 < g. < I and gj > g., and h .. by an affine function with
J - J ~J
k k
slope h~ .. The variable x~. is identified with the interval [R.. ,u.]. Thus
~J ~J J J
o < x~. < u~ - R.~ =.c~ for k l, ... ,k(j).
~J - J J J
With these approximations, P becomes the problem LP given by

N a K(j)
minimize L L L h~.l. (LP.O)
i=O j=l k=l ~J ~J

subject to

K(j) k K(j)
k k
L xi+l,j L (l-g.)x .. + d .. ,
J ~J ~J
i O, ... ,N-I, vj e: A (LP.l)
k=l k=l

d ..
~J
= F. (q) + L i = 0, .•. ,N-I, (LP.2)
~ jgB(q)

K(~) k
~ x Oj Rj given Vj g A (LP.3)
k=l

d .. > 0, i = 0, ... ,N-I, V j g A (LP.4 )

~J

° _ x~.
< ~J:5.. c kj , i=O, ... ,N, VjgA, k=l, ... ,K(j), (LP.5)
 270

and

(asp) l:-l>
~J
0 => l. ~J
k = 1, ••• ,K(j)-l, i = O, ••• ,N, vj E: A.

Except for the last relationship, which we call the ordered solution

property asp, LP is a linear programming problelu. It is necessary to satisfy

asp, otherwise the piecewise linear approximations are not faithful repre-

sentations of the exit functions. When no optimal solution to the linear

programming problem given by (LP.0-LP.5) satisfies asp, the only feasible

alternatives to attain asp are:

(a) branch-and-bound techniques.

(b) ordered (or restricted) basis entry procedures in the

simplex method.

Nei~her of these alternatives is very attractive since (a) is quite expensive

computationally and (b) does not yield a global solution.

(LP.0-LP.5) does not satisfy well-known conditions that guarantee asp.

However, the main result of this paper is

Theorem l~ If {h •. (x)} are piecewise linear and satisfy the cost function
~J

assumption CFA, then there exists a basic (extreme point) optimal solution to

LP.O-LP.5 that satisfies asp.

We also establish the closely related result

Theorem 2: If {hij(X)} satisfies the cost function assumption CFA, then

every local minimum solution to problem P is a global minimum.

Finally we show that the constraint matrix specified by (LP.I-LP.5) has a

staircase structure so that decomposition techniques or compactification methods fl

treating sparse structured matrices can be applied in solving LP. Thus

 271

high-speed memory requirements will be substantially smaller and computation

time possible faster than would result from the use of standard linear
programming codes.
 272

BIBLIOGRAPHY

1. S. C. DAFERMOS and F. T. SPARROW, "The Traffic Assignment Problem for a

General Network," J. of Res. of Nat. Bur. Stds. 73B, 91-118 (1969).

2. R. S. GARFINKEL and G. L. NEMHAUSER, Integer Programming, Wiley, New York,

1972.

3. C. R. GLASSEY, "Nested Decomposition and Multi-Stage Linear Programs,"

Man. Sci. 20, 282-292 (1973).

4. G. HADLEY, Nonlinear and Dynamic Programming, Addison-Wesley, Reading, Mass.,

1964.

5. A. R. HEERSTEMAN and J. SANDEE, "Special Simplex Algorithm for Linked Problems,"

Man. Sci. 11, 420-428 (1965).

6. J. K. HO and A. S. MANNE, "Nested Decomposition for Dynamic Models," Math.

Prog. 6, 121-140 (1974).

7. L. J. LEBLANC and E. K. MORLOK, "An Analysis and Comparison of Behavioral

Assumptions in Traffic Assignment," Presented at the International Symp.
on Traffic Equilibrium Methods, Universite de Montreal, Nov., 1974.

8. E, LOUTE, "A General Inverse Compactification Method for Solving Large

Scale Linear Programs," in preparation.

9. D. K. MERCHANT, "A Study of Dynamic Traffic Assignment and Control,"

Ph.D. Dissertation, Cornell University, August, 1974.

10. J. L. MIDLER, "A Stochastic Multiperiod Multimode Transportation Model,"

Transp. Sci. 3, 1-7 (1969).

11. S. NGUYEN, "A Unified Approach to Equilibrium Methods for Traffic Assignment,"
Presented at the International Symp. on Traffic Equilibrium Methods.
Universite de Montreal, Nov., 1974.
12. R. B. POTTS and R. M. OLIVER, Flows in Transportation Networks, Academic
Press, New York, 1972.
13. S. ~TIDHAM ~R., "Or~ere~ Basis Entry in Linear Programs with Upper-bounded
VarJ.ables WJ.th Appl:catJ.ons to Separable Non-linear Programs," Tech. Rep.
No. Ill, Cornell UnJ.versity, May, 1970.
14. C. S. TAPIERO and M. A. SOLIMAN, "Multi-Commodities Transportation Schedules
Over Time," Networks 2, 311-328 (1972).
15. S. YAGER, "Dynamic Traffic Assignment by Individual Path Minimization and
Queuing," Transp. Res. 5, 179-196 (197.1).
16. S. YAGER, "Emulation of Dynamic Equilibrium in Traffic Networks," Presented
at th; International Symp. on Traffic Equilibrium Methods, Universite de
Montreal, Nov., 1974.
 273

17. W. I. ZANGWILL, Nonlinear Programming: A Unified Approach, Prentice~all,

Englewood Cliffs, New Jersey, 1969.

18. R. A. ZVIAGUINA, "Multilevel Decomposition in Linear Programming," (in

Russian), Mathematische Operationsforschung und Statistiek 6, 427-443 (1973).
AREA TRAFFIC CONTROL AND NETWORK EQUILIBRIUr1

Nathan Gartner
Massachusetts Institute of Technology
Cambridge, Massachusetts, USA

ABSTRACT

Area traffic control systems play an important

role in determining the equilibrium between demand
and supply in an urban highway network. The paper
describes some of the methods that have been developed
for the operation of these systems and derives the
level-of-service that would result from a given set
of flows. Currently used techniques attempt to opti-
mize network performano6 assuming a fixed pattern of
demands. It is shown, via an example, that control
measures can be used to affect the demand pattern in
such a way that total network performance is improved.
A model for achieving this objective is discussed.
 275

1. INTRODUCTION

According to a recent survey(l) more than 100 urban areas in the United
States and Canada are in the process of planning, designing and implementing
area-wide computerized traffic control systems. It is anticipated that with-
in the next few years every city of more than 100,000 people will have such
systems in operation for controlling the traffic flow through its inter-
sections and freeway corridors. (2) These systems are bound to have a pro-
found impact on the performance of the traffic network in those areas in
terms of 1evel-of-service and capacity. It is the purpose of this paper to
describe some of the methods that have been developed for the operation of
these systems and to discuss possible further developments.
The methods that are used for setting signals to control the flow of
traffic through the intersections of a highway network are, perhaps, the most
practical application of the concepts of traffic equilibrium in a transporta-
tion network. The systems in which they are implemented have the additional
capability of monitoring the traffic flow, keeping track of its time-varying
dynamics in great detail via vehicle detectors, and responding quickly to
changes in demand via the communication system linking detectors, signals,
and computer altogether. Unfortunately, the electronic hardware seems to be
far more advanced than the engineering knowledge is capable of making use of
it.
The basic paradigm of equilibrium in a transportation network is illus-
trated in Fig. lAo (3) Let us define:
L = level-of-service (such as trip time) on a particular facility
V = volume of flows on this facility
T = specification of the transportation system (including its control
measures)
 276

A = specification of the activity system

Then the supply function
L = 5(T, V) (1.1 )
shows an increase in the 1eve1-of-service as volume increases, and the de-
mand function
V = D(A, L) (1.2)
a decrease in volume as the 1eve1-of-service increases (in the negative sense).
The resulting equilibrium point E(L o , V0 ) occurs at the intersection of the
two curves.
In the context of computing the traffic equilibrium in a signal-control-
led highway network the simplifying assumption is made that demand is an in-
elastic function, fixed at a flow pattern F. This assumption is a good
o
approximation when estimating short-run equilibrium. The equilibrium values
in this case, E(L o , F0 ), represent the 1eve1-of-service at which the given de-
mand is serviced (Fig. 1B). It is also useful in calculating long-run equi-
1ibrium, such as in the traffic assignment phase of the transportation plan-
ning process, and has been used widely in previous studies. (4,5) Models for
determining this equilibrium are described in the following sections.

2. THE CONTROLLED INTERSECTION

The most important element determining the level-of-service of traffic

in an urban area is the at-grade signal-controlled intersection. The effect
of traffic flow on travel time between intersections is usually minor com-
pared to its effect on the delay time incurred at the intersection itself. (6)
Therefore, the primary determinant of the level-of-service variable L becomes
 277

the delay time. This section describes traffic performance at a single iso-
lated signal-controlled intersection.
Let us consider fir.st one approach to a signalled intersection. Assuming
that arriving traffic is not modulated by any nearby controlling devices, the
average delay per vehicle on the approach, d, can be regarded as the sum of
two components

(2.1)

where dd is the delay that would result if the flow were uniform and ds is
the additional delay caused by the stochastic nature of traffic flow. Let
us define,
C = the signal cycle time (sec)
G = effective green time for the approach (sec)
g = G/C, proportion of cycle which is effectively green
q = arrival flow on approach (veh/sec)
s = saturation flow at the signal stop line (veh/sec)
x = q/gs, degree of saturation; ratio of flow to maximum possible flow
under given setting.
According to Webster, (7) the average delay per vehicle on an- approach can be
approximated by
2 2
d = K[C(l - g) + x ] (2.2)
2(1 - q/s) 2q(1 - x}

where K is a constant (usually 0.9). The predictions made by this formula

were found to correlate very well with field data. Similar expressions were
also derived by Miller(S) and Newell.(9)
The way the delay varies with traffic flow (or, with the degree of satura-
tion) is illustrated, for a typical case, in Fig. 2. It is seen that at high
degrees of saturation the delay rises steeply. Theoretically, the delay in-
 278

creases to infinity as the flow approaches capacity (x + 1.0). But in prac-

tice the flow does not sustain a high value for a long period; it falls off
at the end of the peak period, and the queue does not reach a length required
to cause excessive long delays. Furthermore, it has been observed that drivers,
upon seeing a long queue-up away from the intersection, will turn off and seek
alternative routes.
To derive the level-of-service (i.e., the delay) at which traffic through
the intersection will be served, both the green times G and the cycle time
C have to be determined and all flows must be considered. Let us examine,
for example, the signalled intersection analysed in Fig. 3. The signal has two
phases corresponding to the two possible directions of movement, N - Sand
E - W. The sum of the effective green times for the phases is

(2.3)

where L, in this case, is the total 'lost time' for the intersection. Using
equ. (2.2) to calculate the average delay per vehicle on each approach, and
noting the symmetry in the flows, we obtain for the ra~e of delay, D, for each
phase,

(2.4)

The rate of total delay, D, considering all flows through the intersection
is

D = DEW + DNS (2.5)

The three functions in equ. (2.5) are illustrated in Fig. 3 for various com-
binations of effective green times satisfying equ. (2.3) and a fixed cycle
time C. (G)min is the minimum effective green time that still can accommodate
 279

the demand on the approach, though at a very high rate of delay, and is given
by

(G)min = qC/s i.e., x = 1.0 (2.6)

The apportioning of green time among the conflicting streams at the inter-
section can be formulated as the following optimization program:

Given qi' si for all approaches i, and C

Find G. for all phases j to
J -
Min D = ~Dj
J
Subject to (2.7)

EG. =C- L
j J

Gj ~ (Gj)min
The optimal solution is obtained as an equilibrium point where the marginal
rate of delay for the conflicting phases is equalized. In the simple two-phase
example this would be

(2.8)

Webster gives an approximate rule for determining the optimal splits of green
time,

(2.9)
where Yj is the maximum ratio of flow to saturation flow for the different
approaches having simultaneous right-of-way during phase j, and

v= EY. (2.10)
j J

To determine the optimum cycle time C* for the intersection, capacity

considerations play an important role. For each approach i we must have
 280

q.C
1 -
< S.G.
1 1
(2.11 )

When the approaches are assigned to phases j we obtain,

y. < G./C . (2.12)

J - J

Summation over all phases j at the intersection yields

r,y. < r,gJ.

J -
(2.13)

Using equs. (2.3) and -(2.10), we obtain

Y < 1 - L/C, or, C ~ L/(l - y) (2.14)

The minimum cycle time, Cmin , i.e., the cycle time that will serve all phases
at a degree of saturation of unity, is

C.
mln = L/ (1 - Y) (2.15)

Obviously, because of the randomness in arrivals, such a cycle will cause an

intolerable amount of delay. When plotting the total rate of delay at an in-
tersection, D, as a function of cycle time (splits being optimized indepen-
dently for each cycle), one obtains a family of curves as shown in Fig. 4.
It is seen that the rate of delay is asymptotic to the minimum cycle ordinate.
It decreases toward a minimum at higher cycle times as the increased capacity
causes a reduction in the stochastic delay component. At still higher cycle
times the rate of delay increases again. At this stage the deterministic
delay component becomes dominant because of the larger red times on each ap-
proach. Webster shows that the optimum cycle time is roughly twice the mini-
mum cycle time, i.e.,

C* ~ 2C.
mln = 2L/(1 - Y) (2.16)

Mi11er(10) derives an expression for the optimal cycle time that gives simi-
lar results. On the other hand, A11so~(11) develops an iterative procedure
 281

to determine delay-minimizing settings to all traffic passing through the in-

tersection.

3. AREA TRAFFIC CONTROL

When two or more intersections are in close proximity, some fornl of 1ink-
ing is necessary to reduce delays to traffic and prevent frequent stopping.
A signal-controlled intersection has a platooning effect on the traffic leaving
it, and it is advantageous to have the signals synchronized, i.e., operating
with a common cycle time. It also becomes necessary to coordinate the signals,
i.e., to establish an offset between the signals, so that loss to traffic is
minimized. A typical platoon profile that was measured at the downstream end
of a signalized traffic link is shown in Fig. 5A. Fig. 5B shows the associated
link performance function, i.e., the delay incurred by the platoon when pass-
ing through the downstream signal as a function of the offset between the up-
stream and downstream signals. The maximum of the function corresponds rough-
1y to the head of the platoon arriving at start of downstream red and the mini-
mum corresponds roughly to the tail of the platoon arriving at the end of down-
stream green. (12)
The usual procedure for setting signals on arterials and in networ~s in-
volves three steps. (13) First, a common cycle time is determined according to
the requirements of the most heavily loaded intersection. Then splits of green
time are apportioned at each intersection according to the interacting f10w/
capacity ratios. Lastly, a computer optimization procedure is used to deter-
mine a set of offsets throughout the network. Several computer programs have
been developed for determining offsets in a network [e.g., references 14-21],
 282

the main differences among them being in the way the traffic flow process
is modeled and the optimization technique employed. Practically all these
methods have shown an improvement in network performance, when compared
with the earlier manual methods that were used by traffic engineers. (22,23)
However, the three-step sequential decision process used for setting
the signals is not entirely satisfactory. Similar to the single inter-
section case where all the interacting flows have to be considered when
determining optimal cycle time and splits, the network case too requires
that all demands be considered simultaneously when determining the con-
trol variables. A model for such a program is presented below.
The signal-controlled traffic network consists of a set of links
(i, j) connecting the adjacent signals Si and Sj. Let,

Gij(R ij ) effective green (red) time at Sj facing link (i, j)

1.. lost time at signal phase serving link (i, j)
lJ
~ij = offset time between Si and Sj along (i, j)
qij(sij) = average flow (saturation flow) on link (i, j)

The link performance function is composed of a deterministic delay

component and a stochastic delay component. The deterministic component,
z .. (~ .. , R.. , C), is given by the average delay incurred per vehicle in
lJ lJ lJ
a periodic flow through Sj. The stochastic component, which arises from
variations in driving speeds, marginal friction, and turns, is expressed
by the occurence of an overflow queue Qij(R ij , C) at the stop line of Sj.
An estimation of the expected overflow queue was given by wormleighton,(24)
who considered traffic behavior along the link as a non-homogeneous Poisson
process with a periodic intensity function represented by the flow pattern
on the link. Therefore, we can consider the total delay in the network,
 283

0, to be composed of two components

(3.1 )

where,

Dd = 1: q .. z .. (~ .. , RiJ·, C) (3.2)
i,j lJ lJ lJ
°s =i ,j Q.. (R .. , C}
1:
1J 1J
(3.3)

A number of constraint equations involving the decision,variables are

necessary to model the network. First, the algebraic sum of offsets around
any loop of the network must equal an integral multiple of the cycle time, (25)
i. e. ,

1:~ .. =nC (3.4)

(i ,j)d lJ R.

where nR. is an integer number associated with loop R.. Effective green and
effective red are related by,

G.. +R .. =C (3.5)
lJ lJ
In order for the network to be able to handle the given flow we must have
for each link the capacity constraint

(3.6)

Practical considerations, including pedestrian crossing times and driver be-

havior prescribe

(3.7)

C.
mln -< C -< Cmax (3.8)

Assuming, for simplicity, two-phase intersections we also have

(3.9)
 284

where (i, j) and (k, j) are assigned conflicting phases at S ..

J
Thus, the network signal setting problem can be stated in a general
form as the following nonlinear optimization program:
Find $ .. , R.. , C to:
-- lJ lJ -

Subject to:
E $ .. = n C
(i ,j)e:R. lJ R.

Gij + Rij = C

Rij - lij = Gkj + \j

(3.10)

This program is nonlinear in the objective function and in the loop con-
straint equations, and has integer decision variables. By a suitable representa-
tion of the objective function, the program can be solved by mixed-integer pro-
gramming. (26) In principle, the program can also be solved by dynamic pro-
gramming, though, for computational reasons the splits have to be determined
independently rather than simultaneously with the other decision variables. (27)
It is important to study the sensitivity of the network objective func-
tion with respect to the network cycle time. A typical relationship is shown
in Fig. 6. Each point in the graph represents network performance optimized
with respect to splits and offsets at a fixed cycle time. It is evident that
 285

the optimal cycle time for the network constitutes a least-cost equilibrium
point between delays occuring in a deterministic situation and delays con-
tributed by stochastic factors. While the first component of delay usually
increases with cycle length (though at a decreasing rate), the latter com-
ponent decreases with it because of the increase in capacity at higher cycle
times. The stochastic component, and the total delay, is asymptotic from
above to the minimal cycle time for the network, which is the theoretical
minimum for the most heavily loaded intersection if all flows were uniform
periodic. These characteristics are analogous to those observed at a single
intersection, however, the implications regarding signal settings in a net-
work are different and a single intersection analysis would virtually never
give the optimum settings for the network.

4. TRAFFIC CONTROL AND ROUTE CHOICE

The foregoing discussion has considered methods for determining traffic

control settings that minimize total cost, given a fixed pattern of traffic
flows. It is assumed that all route choices are fixed, resulting in constant
flows on each link regardless of the controls imposed on that link and, hence,
regardless of the level-of-service that is offered by the link. This assump-
tion would be correct only in the event that the level-of-service on the con-
trolled links is insensitive to the control settings, which is, of course,
incorrect. It seems, therefore, to be of fundamental importance to have a
model which incorporates both traffic controls and route choice and provides
a tool for establishing a system-optimized traffic flow pattern. A practi-
cal example illustrates the argument.
 286

Fig. 7A shows the intersection of Commonwealth Avenue/Boston University

Bridge in Boston with an advanced green phase allowing for west-to-north
left-turns. The average delay for the left-turners is 63.0 seconds and the
rate of total delay for all traffic passing through the intersection is 41.6
veh x sec/sec. Turning to Fig. 7B it is seen that west-to-north movements
can also be accomplished by traveling an extra 75 seconds through route
Al - E - F - B2· The signal can now be operated on a shorter cycle time by dis-
allowing the left-turn movement. The total expected travel time on the ex-
tended route is 22.0 + 75.0 + 16.7 = 103.7 seconds, i.e., an increase of 64%
in travel time. However, the total rate of delay for all users of the inter-
section is now only 32.3 veh x sec/sec, i.e., a reduction of 22.3% with res-
pect to the previous figure. It should be noted that despite the fact that
no left-turn arrow exists anymore, it is physically possible to make the
turn, through the opposing traffic, and occasionally some vehicles are making
it. But taking into account that there is a good chance of getting a costly
ticket, which can be regarded to be equivalent to a very high delay, most of
the drivers will choose the longer route, thus achieving a system-optimiza-
tion at this location.
The question still remains, how to incorporate route choice as part of
the traffic control optimizing program (for a general discussion on this sub-
ject, see a recent paper by Allsop(28)). One possible approach might be to
use the general formulation given by equs. (3.10), and let the qij'S be de-
cision variables rather than input parameters. The objective function would
have to be modified to include also the cost of traveling on the links,
E q .. l(t.., where t .. is the travel cost on link (i, j). And to the con-
;,j lJ lJ lJ
straints, a set of equations would have to be added, respresenting the 0 - D
demands and the nodal continuity of flow equations. This set will be of the form,
 287

(4.1)

where,
A = node-link incidence matrix of the network
g = link flow vector
~ = nodal trip productions (attractions) vector
A solution to this nonlinear optimization program would represent a system-
optimized flow pattern and control program. Recent results on traffic assign-
ment indicate that such a pattern may not be significantly different from a
user-optimized pattern. (29) Alternatively, the settings obtained with this
program could be used in conjunction with a suitable simulation program to
arrive at a user-optimized equilibrium flow pattern that would also constitute
a minimization of community costs.
Route choice and route control are also connected with the time-vary-
ing characteristics of traffic flow, and the possibilities of exerting dyna-
mic, traffic-responsive, control. In theory, the potential savings could be
substantial. However, no successful schemes have been reported to-date,(30)
and the area is open for further research.
 288

5. REFERENCES

1. J.F. Schlaefli, "Street Traffic Control by Computer--a state-of-the-art

discussion," Workshop on Computer Traffic Control, University of Minne-
sota, March, 1974.

2. N.R. Kleinfield, "Computerized Traffic Control," The Wall Street Journal,

August 13, 1974
3. M.L. Manheim, "Search and Choice in Transport Systems Analysis," High-
way Research Record No. 293: Transportation Systems Planning, 1969.
4. N.A. Irwin and H.G. Von Cube, "Capacity Restraint in Multi-Travel Mode
Assignment Programs," Highway Research Board Bulletin 347: Trip Charac-
teristics and Traffic Assignment, 1962.
5. Comsis Corp. ,"Traffic Assignment--methods, applications, products,"
U.S. Department of Transportation, Federal Highway Administration, 1973.
6. "Hi ghway Capacity ~lanua 1," Hi ghway Research Board, Speci a 1 Report 87,
1965.
7. F.V. Webster, "Traffic Signal Settings," Road Research Technical Paper
No. 39, H.M. Stationery Office, London, 1961.
8. A.J. Miller, "Settings for Fixed-Cycle Traffic Signals," Operational
Research Quarterly 11, 1963.
9. G.F. Newell, "Approximation Methods for Queues with Application to the
Fixed-Cycle Traffic Light," SIAM Review I, 1965.
10. A.J. Miller, "The Capacity of Signalized Intersections in Australia,"
Australian Road Research Board Bulletin No.3, 1968.
11. R.E. Allsop, "Delay-Minimizing Settings for Fixed-Time Traffic Signals
at a Single Road Junction," J. Instr. Maths. Applics. ~, 1971.
12. N. Gartner, "Microscopic Analysis of Traffic Flow Patterns for ~lini­
mi zi ng Delay on Si gna 1 Controlled Li nks," Hi ghway Research Record No.
445: Traffic Signals, 1973.
13. J. Holroyd, "The Practical Implementation of Combination Method and
TRANSYT Programs," Transport and Road Research Lab Report LR 518,
Crowthorne, 1972.
14. J.T. Morgan and J.D.C. Little, "Synchronizing Traffic Si9nals for Maxi-
mal Bandwidth," Operations Research 11, 1964.
15. J.D.C. Little, "The Synchronization of Traffic Signals by Mixed-Integer
Linear Programming," Operations Research 11, 1966.

16. Traffic Research Corp., "SIGOP: Traffic Signal Optimization Program,"

PB 173 738, 1966.
 289

17. 0.1. Robertson, "TRANSYT: A Traffic Network Study Tool," Road Research
laboratory Report lR 253, Crowthorne, 1969.
18. J.A. Hillier, and R.S. lott, "A Method of Linking Traffic Signals to
Minimize Delay," 8th International Study Week in Traffic Engineering,
Theme V: Area Control of Traffic, Barcelona, 1966.
19. K.W. Huddart and E.D. Turner, "Traffic Signal Progressions--GLC Combina-
tion Method," Traffic Engineering and Control, 1969.
20. R.E. Allsop, "Choice of Offsets in Linking Traffic Signals," Traffic
Engineering and Control, 1968.
21. N. Gartner, "Optimal Synchronization of Traffic Signal Networks by Dyna-
mic Programming," Traffic Flow and Transportation (G.F. Newell, Ed.),
American Elsevier, New York, 1972.
22. J. Holroyd and J.A. Hillier, "Area Traffic Control in Glasgow: a sum-
mary of resu lts from four control schemes," Traffic Engi neeri ng and Con-
trol, 1969.
23. F.A. Wagner, F.C. Barnes, and D.L. Gerlough, "Improved Criteria for
Traffic Signal Systems in Urban Networks," NCHRP Report 124, Highway
Research Board, 1971.
24. R. Wormleighton, "Queues at a Fixed Time Traffic Signal with Periodic
Random Input," Canadian Operations Research Society Journal 1., 1965.
25. N. Gartner, "Constraining Relations Among Offsets in Synchronized Signal
Networks," Transportation Sci ence .§.' 1972.
26. N. Gartner, J.D.C. Little, and H. Gabbay, "Optimization of Traffic Sig-
na 1 Sett i ngs in Networks by Mi xed- Integer Linea r Prograrrrni ng ," Techni-
cal Report No. 91, Operations Research Center, M.I.T., Cambridge, 1974.
27. N. Gartner and J.D.C. Little, "The Generalized Combination Method for
Area Traffic Control," Transportation Research Record (to appear).
28. R.E. Allsop, "Some Possibil ities for Using Traffic Control to Influence
Trip Distribution and Route Choice," Proceedings 6th Intern. Symp. on
Transportation and Traffic Theory, Sydney, 1974.
29. l.J. Leblanc and E.K. Morlok, "An Analysis and Comparison of Behavioral
Assumptions in Traffic Assignment," Intern. Symp. on Traffic Equilibrium
Methods, Montreal, 1974.
30. J. Holroyd and D.1. Robertson, "Strategies for Area Traffic Control Sys-
tems: Present and Future," Transport and Road Research Lab. Report
LR 569, Crowthorne, 1973.
SERVICE SERVICE
L L

o S o S

La La
~

VOLurlE L-_ _ _ _ _ _--'-_ _ _ _..VOLurlE

I .. V
Va Fa • V

Fioure lA Fi(lUre lB

Figure 1: Basic paradigms of equilibrium in a transportation network.

 TRAFFIC FLOW (veh/hr)
0 300 600 900
I II
60

.s::
cv 50
>
....... c = 60 sec
(.)
cv G = 30 sec
1/1
40 ~ s = 1800 veh/hr
~
...J 30
w
0 I\.)

L ~
w
t!) 20
«
0::
W
~ 10

o 0.2 0.4 0.6 0.8 1.0

DEGREE OF SATURATION, X

Figure 2: Average delay to traffic on a signal-controlled intersection approach.

 RATE OF DELAY (veh x sec/sec)
N
0 (J1 o (J1 o
a o b b b
0 -.J
o
fT1 en fT1
0 0
.,.,
(;) fT1
fT1 fTl
(") (")
""
-; -
:e - - - - - - - - - - - - - - - - - - -;
N (J1
3 <
< 0 :::J 0 rr1
fT1

G) G)
:::0 I\,)
:::0 VJ ~ CD
rr1 I\,)
fT1 0 0
fT1 rr1
Z
o Z
G)
G) ~ VJ Z
fTl 0 0 CJ)
:e G)
-z
CJ)
(J)
(J) (1)
(1) (J1 N n
n 3
- 0
0 5'

en o
0
Figure 3: Apportioning of green time at a four-leg, two-phase signalled intersection.
(Data: qE=qW= 400 veh/hr; qN=qS= 600 veh/hr; s = 1800 veh/hr for all legs;
C = 80 sec; L = 10 sec)
 75

U
<l>
~ 60 I
u
<l>
IJ') I
X I
.r;
~ 45 I Eqi = 3000 vph
I
o I
I Cmin C*
~ 30
--1 I I\,)
W Eqi 2000 vph ~
o I
LL
I\
o 15 I
w Cmin
~ C..
0:::
o
o 20 40 60 80 100 120 140 160
CYCLE TIME (SECONDS)

Figure 4: Sensitivity of the rate of delay with respect to cycle length. (Two-phase. four-leg
intersection: equal flows on all legs; equal green times; si= 1800 veh/hr; L = 10 sec)
 294

2.0
(Bloor Street, Toronto, 1972)
Fig. 5A
-
f- -'
-~
I-
z
1.5 - r- -~
;-
;-
-
-
i-
..... f-
r l-
t- t-
-
:J: t- t-
W
- -
-> 1.0
I-
;-
~
-
== ;-
-
9I.L 0.5 - t- -
t-

[[h[
~

o ~Lb-r
I 5 10 15 20 25 30 35
ASPECT CYCLE IN 2 SECON D INTERVALS

40~----~----~------~-----'------'------'------'

Fig. 5B
:J:
W
> CYCLE LENGTH 70 SECS
..... 30 EFFECTIVE GREEN 32 SEeS
(/)
o RELEASE RATE 1. 76 VEH/INT

~
frl
~
20
>-
<r
...J
ow
o
W
~ 10
0:
W
>
<r

0~--~7-----~----~~--~~--~~----~--__~
o 30 40 50 70
OFF S ET (SECONDS)

Figure 5: A. Platoon profile; B. Link performance function.

 140

120

u
---
Q)

.........
'"
u
Q) 100--
II>

X
.r:
Q)
>

>-
80
co::
-'
w
o
W-
o J\.)
60 CD
V1 tn
w
I-
co::
c:r:

~
o
40
3:
I-
w
Z

Os
20 I
Cmin = 45 sec Copt = 75sec
I I

o 40 50 60 7080 90 I 00 II 0 120
C -NETWORK CYCLE TIME (SECONDS)

Figure 6: Rate of total delay vs. cycle time in a synchronized network.

 296

QJ
Jf

t
Cl'
"0
.~

!...
OJ
81
::>

1
OJ

Commonwealth Avenue

J
2

.
A' A B

. I j 1I
~
Phase Sequence:
•

C = 100 sec; L = 9 sec

Approach q s 9 d 0
(vph) (vph) (cycles) x (sec) (veh x sec/sec
A1 240 1500 0.18 0.88 63.0 4.2

A1 1500 3000 0.54 0.925 30.7 12.8

A2 1200 3000 0.36 0.75 28.8 9.6

Bl 1050 3000 0.38 0.92 42.0 12.2

82 500 3000 0.38 0.44 20.4 2.8
TOTAL: 41.6

Figure 7A: Commonwealth / Boston University (3 phases).

 297

OJ .If

t
0-
....
"0
~
co Bl

.
:::>

co
1

Al .. -_----_
~E

1 I
Commonwealth Avenue

I
I
l
Phase Sequence:
I
A B I
}
-- --""/F

c = SO sec; L = 9 sec

Approach q s 9 d D
(vph) (vph) (cycles) x (sec) (veh x sec/sec

Al 1740 4500 0.46 0.S5 22.0 10.6

A2 1200 3000 0.46 0.59 16.0 6.S

Bl 1050 3000 0.43 0.815 22.6 6.5

B2 740 3000 0.43 0.58 16.7 2.d.

(sub-total) 27.3

(travel time on extra route) 5.0

TOTAL: 32.3

Figure 7B: Commonwealth / Boston University (2 phases).

 EVALUATION OF FREEWAY CORRIDOR ASSIGNMENT EQUILIBRIA*

Maxence P. Orthlieb
and
Adolf D. May

ITTE, University of California

Berkeley, California, USA

ABSTRACT

In recent years, ramp controls and advanced freeway designs to improve freeway
traffic operations have received considerable attention. Concurrently, work has
been undertaken to develop coordinated design and/or control systems to improve
street operations. However, unless an integrated approach to design and/or control
of both freeway and the adjacent street network (Urban Freeway Corridor) is developed,
an effective balancing of traffic against available capacity cannot be attained.

The main thrust of the present work is twofold. It determines both control
improvement strategies (signal setting and ramp metering plans) and resulting assign-
ment equilibria. The control strategies are selected so that they attempt to reduce
peak-hour congestion on the freeway without creating significant delays to motorists
on the surface streets. Since alternate routes are available, an original traffic
assignment procedure is employed to predict how corridor drivers will react to the
above control strategies. This procedure is an extention of the user-optimized traffi
assignment algorithm. The equilibrium between individual travel time and assignment
is considered.

* The contents of this paper reflect the views of the authors who are responsible
for the facts and the accuracy of the data presented herein. The contents do
not necessarily reflect the official views or policies of the State of California
or the Federal Highway Administration. This paper does not constitute a standard,
specification, or regulation.
 299

The present methodology combines two existing simulation models with a decision
model. The simulation models are deterministic and macroscopic and they predict,
respectively, freeway and street traffic performances as functions of: (1) freeway
and street designs, and (2) allowable ramp inflows and street link flows. The
decision model selects the corridor control settings and predicts the outcoming
traffic assignment.

It is expected that the results of this research will provide traffic engineers
and highway officials with a beneficial tool to be used for evaluating corridor per-
formances and assignment equilibria under different control alternatives. It will
also encourage improved usage of existing facilities.
 300

1. INTRODUCTION

The proposed paper is related to the Freeway Operations Study (FOS) project carried
on by the Institute of Transportation and Traffic Engineering (ITTE), at the University
of California, Berkeley. The project is sponsored by the California Department of
Transportation, Division of Highways, in cooperation with the U. S. Department of
Transportation. The main goal of this project is to investigate various operational as-
pects of urban freeways. Within the six-year research work, the following major
stages have been accomplished:

(1) Development, validation and refinement of a freeway model (FREQ3,

formerly called FREEQ) to evaluate freeway design and control improve-
ments (l to g,..!&. to 19, ~. *

(2) Use of this model to evaluate cost-effectiveness of freeway design alterna-

tives QQ.,W and to analyze freeway on-ramp control strategies Qi,,ll).

(3) Collection and analysis of traffic data from the Eastshore Freeway Corri-
dor @i,g§).

The present stage involves the development and the application of analytical techniques
for evaluating design and/or control improvement plans on a freeway corridor basis
~ 26 to ~. This stage presents an integrated approach to design and/or control of
both freeway and the adjacent street network which results in an effective balancing
of traffic against available capacity. Within this integrated approach, this paper focuses
on the evaluation of corridor assignment equilibria. ** Such an evaluation provides a
better understanding of the interaction between corridor control and corridor assign-
ment.

This paper is divided into three main sections. The first section gives the basic defi-
nitions and assumptions relative to an urban freeway corridor, interprets how it per-
forms, and introduces the proposed methodology. The second section describes the
idea of drivers' behavior, the settings of the control elements, and the interaction of
corridor control with corridor assignment. Finally, the last section presents a for-
mulation for the control-assignment interaction.

* Underlined numbers at the end of a sentence or following a name are reference

numbers.
**A corridor assignment equilibrium is defined, in this paper, as a set of flows, on
subsections and links, which satisfies all the users.
 301

2. THE URBAN FREEWAY CORRIDOR

2.1 CORRIDOR SYSTEM DEFINITIONS

The urban freeway corridor is defined in terms of function, configuration, elements,

and boundaries.

The main function of the corridor is to serve users safely, efficiently, and convenient-
ly.

The configuration of the corridor is shown as a strip of land including a portion of

directional freeway, with associated ramps, along the major axis, a major parallel
arterial street, and a set of connecting cross-streets (Figure 1).

The main elements of the corridor are grouped into three categories:

(1) The design elements which determine the capacities of the various
parts of the corridor.

(2) The traffic elements which determine the demands per time-slice* in
the various parts of the corridor.

(3) The control elements which are the corridor control variables to be
optimized.

These categories are detailed in Figures 2, 3, and 4.

The boundaries of the corridor are so chosen that:

(1) Traffic conditions outside the corridor will not be significantly affected
by the control inside the corridor.

(2) The input demand at the corridor boundaries can be considered inde-
pendent of the control inside the corridor.

In other words, excluding changes in demand, whatever happens inside the corridor has
little effect on what happens outside the corridor and vice versa. Some effort might be
made to check this assumption which attempts to balance between simplicity and realism.

2.2 CORRIDOR MODEL ASSUMPTIONS

In building a model of the corridor system, assumptions must be introduced. The

most important ones are presented here:

*Time-slice: significant fraction of an hour. Demand and all traffic characteristics

are assumed to remain constant over a time-slice.
 302

Fig. 1 - Corridor Configuration and Boundaries

 303

DESIGN

FREEWAY STREET

1. Subsection 1. Link (or approach)

- length - length

- number of lanes - number of lanes

- capacity - capacity

- design speed - average speed

2. On-ramp 2. Intersection

- location - design

- capacity - capacity

- queue storage - queue storage

3. Off-ramp

- location

- capacity

Fig. 2 - Corridor Design Elements

 304

TRAFFIC

FREEWAY (FRWY) STREET (STRT)

1. O-D Table 1. Flow on Link

- Type I, II, III and IV-a users - Type IV-b users

(see table below) (see table below)

2. Proportion of Right Turns

2. Vehicle Occupancy
3. Proportion of Left Turns
3. Mainline Input Fluctuation
4. Trucks and Through-
buses

5. Number of Local Buses

6. Pedestrian Activity

LEAVES
Type of Users
FRWY STRT

FRWY I II
ENTERS
STRT III IV-a IV-b

IV-a: driver may use freeway

IV-b: driver does not use freeway
- - - - - - - - - - -------- ------ ----------

Fig. 3 - Corridor Traffic Elements

 305

CONTROL

FREEWAY STREET

Ramp Control Intersection Control

- ramp metering rate - common cycle length

- green time (or split)

- offset

Fig. 4 - Corridor Control Elements

 306

(1) The corridor system will always be able to handle the total corridor
traffic during the period of study.

(2) During some time-slice, the freeway subsystem might not be able to
handle the total freeway traffic.

(3) For a given time-slice, the arterial street subsystem will always be
able to handle the total street traffic.

(4) Any demand not served during a given time-slice is added to the demand
of the next time-slice.

(5) Freeway trips having their origin and/or their destination outside the
corridor boundaries (Types I, n, and III users in Fig. 3) will not
change their pattern in space, time, or mode.

(6) Freeway trips having their origin and their destination within the cor-
ridor boundaries (Type IV-a users) may only change their spatial
pattern by selecting the street route from their origin to their destina-
tion.

(7) Street traffic (Type IV-b users) will not change its pattern in space,
time, or mode.

2.3 CORRIDOR SYSTEM DESCRIPTION

In describing the corridor system, particular attention must be given to the Type IV-a
users. Indeed, they represent the only portion of the corridor traffic on which control
might affect assignment. It is assumed that these drivers have the following charac-
teristics:

(1) They select their route (via freeway or surface street) solely on the
basis of minimizing their individual trip travel time.

(2) They Imow the actual trip travel time via freeway and surface street
routes.

Therefore, Type IV-a drivers' behavior will be to select the alternate route which
decreases their trip travel time, based on their past experience of the corridor sys-
tem. Such an attitude is reflected in Wardrop's first principle which states: "The
journey times on all routes actually used are equal, and less than those which would
be experienced by a single vehicle on any unused route" @Q). Now, given a drivers'
behavior model using the above characteristics, assignment equilibria can be deter-
 307

mined and the interaction between corridor control and corridor assignment can be
studied. In brief, this interaction can be presented as follows:

Assume there is a corridor assignment equilibrium with its corre-

sponding corridor control settings. On the top of this equilibrium,
add a number of new corridor users. Whatever their assignment is,
the original equilibrium is likely to be broken. To get back to an
equilibrium, a change in corridor control settings must be made.
Such a change is going to affect the trip travel times of all types of
drivers. However, only Type IV-a users have the potentiality to
react to this change. They react according to the characteristics
stated earlier. Doing so, they modify the actual assignment which,
in turn, would lead to another change in corridor control.

This control-assignment interaction is shown on Figure 5. To simulate this interaction

and evaluate the assignment equilibria, three major needs might be foreseen:

(1) A freeway model to predict trip travel times through the freeway
subsystem.

(2) A surface street model to predict trip travel times through the
street subsystem.

(3) A decision model to picture the drivers' behavior, select the corridor
control settings and predict the most probable corridor assignment.

These needs are discussed in more detail in the following paragraph.

2.4 CORRIDOR MODEL METHODOLOGY

The methodology used in developing the corridor model combines two existing simula-
tion models with a decision model, as shown on Figure 6. The freeway and surface
street simulation models are similar in nature. The models are deterministic and
macroscopic, and predict traffic performances as functions of design, demand and
control. The decision model simultaneously selects the corridor settings and pre-
dicts the resulting assignment taking into account drivers' behavior. The two simula-
tion models are presented below, and the decision model is described in the following
section of this paper.

2.4.1 THE FREEWAY MODEL

The freeway model, called FREQ3, was developed at ITTE as the result of the second
 308

~ MORE USERS

determines Jr----------,I__" _affects

, -, ASSIGNMENT .

l DRIVER'S BEHAVIOR J l SEARCH FOR

CONTROL IMPROVEMENT

,-------,.
---"'in"'=fl=-u-en-c-e-s--I: TRIP TRAVEL TIME [1'4"-c:-::h":a:"ng:":e":s--OW---

)
NEW CONTROL ~

Fig. 5 - Control-Assignment Interaction

 309

ST RT MODEL
(TRANSYT) t - - - - - - - - + i S T R T TRIP TRAVEL TIME

SEARCH FOR
ASSIGNMENT
CONTROL IMPROVEMENTS

F RWY MODEL
I-------~ F R W Y TRIP TRAVEL TIME
(FREQ3)

STRT ' SURFACE STREET FRWY, FREEWAY

Fig. 6 - Corridor Model Methodology

 310

phase of the Freeway Operations Study. It simulates the freeway traffic performance
under given design and demand characteristics. Briefly, the model takes the design
and the demand parameters, and computes the traffic performance and the actual
capacity. The design parameters include subsection length, "straight-pipe" capacity,
ramp location, lane configuration, and design speed. The demand parameters are in
the form of origin-destination tables for every time-slice. The traffic performance
is expressed in terms of speed, vic ratio, density, travel time, vehicle-hour expended,
and vehicle-miles expended. The actual capacity is determined as the "straight-pipe"
capacity minus the weaving effect.

This model has been validated, and latest details are described in report No. 73-1 of
the Freeway Operations Study QID. Extensions of this model, including cost-effective-
ness evaluation of freeway design alternatives and analysis of freeway on-ramp strate-
gies, are discussed in reports No. 73-2 and 73-3 @,m.

2.4.2 THE SURFACE STREET MODEL

The surface street model, called TRANSYT, was developed by Dennis Robertson, and
it was field tested in the United Kingdom. The first detailed description of this model
was given in a TRRL (Transport and Road Research Laboratory) report dated 1969 ill).
TRANSYT is a method of modelling traffic behavior in a network of urban roads and
optimizing the fixed-time settings of the traffic signals. It has two main elements -
the simulation model and the hill-climbing optimization process. The simulation model
is used to calculate the performance index (defined as any chosen balance between total
delay and number of stops) of the network for a given set of signal timings. It allows
for the flow interaction between successive road sections, for the dispersion of platoon
and for flow control by signals or another traffic stream which has the right of way.
The optimization process leads toward optimal signal phasing and offsets which mini-
mize the performance index.

3. THE DECISION MODEL

As mentioned above, both freeway and surface street models have been tested and are
operational. The decision model is still in its development stage. The methodology
followed in developing this decision model is given below. It includes three major parts.
First, the drivers' behavior is expressed in terms of constraints; second, the deter-
mination of the corridor control settings is presented; finally, the control-assignment
interaction is described.
 311

3.1 THE DRIVERS' BEHAVIOR MODEL

The drivers' behavior model determines, for some given system objective, corridor
control settings, and initial assignment, the most probable Type IV-a users' route
selection which results in satisfactory traffic conditions for all corridor drivers. This
model is expressed as a set of constraints which insures the Type IV-a users' assign-
ment to be such that:

(1) No more than the Type IV-a users' demand is assigned.

(2) No freeway subsection is congested.

(3) No street link is congested.

(4) No total on-ramp delay is larger than the total aggregated saving in
trip travel-time for all the users of the corresponding on-ramp.

The first constraint prevents the model from assigning more than the Type IV-a users'
demand. The second and third constraints protect the freeway and street subsystems
from being congested. The last constraint distributes, at every on-ramp, the Type
IV-a users so that they will queue up only if they can save time even including the ramp
delay.

3.2 CORRIDOR CONTROL SETTINGS DETERMINATION

The corridor control settings are grouped into four classes: (1) common cycle length,
(2) green time, (3) offset, and (4) ramp metering rate. It is completely infeasible to
determine all the elements of these four classes at once. Therefore, they are resolved
by class.

The common cycle length is either fixed by the model user or determined as a function
of the street design and the pedestrian activity. Although the corridor model does not
include a cycle length optimization, it can be used to derive some optimum traffic
performance of a corridor system for a number of different cycle lengths and hence
the most suitable cycle to use.

For a given peak period analysis, it is assumed that the green times and the offsets
are fixed over the entire period. To determine the green times and the offsets, the
following steps are considered. First, a prevailing hour must be selected. It is sug-
gested to choose the hour during which the sum of all the link flows in the corridor sys-
tem is the highest. Then, for the existing link flows corresponding to this prevailing
hour, the values of green times and offsets are obtained by using the TRANSYT opti-
mization routine. These values minimize the performance index of the street subsys-
tem for the prevailing hour.
 312

The ramp metering rates are updated at the beginning of every time-slice. For a
given time-slice, these rates are set to optimize one of the measures of effectiveness
listed here: (1) total saving in trip travel time for Type IV-a users, (2) total system
travel time, (3) total number of vehicles diverted to the surface street, (4) total
vehicle-miles diverted to the surface street, (5) total ramp delay, (6) ramp queue
length, (7) freeway input rate, or (8) freeway throughput. Such optimization is subject
to the constraints mentioned earlier, and to the following:

(1) The ramp metering rate at each on-ramp must lie between preselected
minimum and maximum values.

(2) No ramp metering rate exceeds the corresponding on-ramp demand.

3.3 THE CONTROL-ASSIGNMENT INTERACTION

The control-assignment interaction determines the ramp metering rates and the re-
sulting assignment which optimize one of the proposed measures of effectiveness,
subject to the previously mentioned constraints. In formulating the objective functions of
most of the eight measures of effectiveness, one encounters non-linear relationships
due to the flow-travel time dependence. A short-cut usually described in the trans-
portation planning literature proposes travel times independent of flows. It is assumed
here that travel time is a step function of flow for every subsection and every link in
the corridor system.

The Type IV-a users' demand is broken down into incremental loads. It is the incre-
mental loading of the Type IV-a users' demand onto the corridor system which defines
the step function. The first step of the function is determined by loading Type I, II,
III and IV-b users. The next step is deduced by assigning an incremental Type IV-a
users' load on the top of the previous loading. This load is made of incremental flows
which are distributed between freeway and street subsystems. These incremental
flows and the resulting ramp metering rates are determined in order to optimize the
selected measure of effectiveness subject to the constraints discussed earlier. The
process keeps iterating until all the Type IV-a users' demand has been assigned. The
use of this loading procedure allows the formulation of the objectives as linear func-
tions of flows and/or ramp metering rates within every incremental load. It is then
possible to apply the well-known Linear Programming technique to determine the
optimum set of flows and ramp metering rates @' Refinements of this process
might include a loading and unloading procedure as described by Steel @.:D.
 313

To summarize this control-assignment interaction, the following algorithm is pro-

posed, for a given time-slice:

Step 0 Load Types I, II, III and IV-b users' demand.

Step 1 Calculate travel times along the freeway and the surface street
routes for the proposed load.

Step 2 Determine the incremental portion of the Type IV-a users' demand
to be loaded at the end of the present iteration.

Step 3 Solve the linear program with the selected objective function and
the appropriate constraints.

Step 4 If the refinements described in reference 33 are used, modify the

flows and the rates accordingly; otherwise, go to Step 5.

~ Update the various flow and rate variables, and adjust the con-
straints.

Step 6 If the Type IV-a users' demand is not completely loaded,

go to Step 1; otherwise, stop.

4. CONTROL-ASSIGNMENT FORMULATION

First, the definition of the variables used is given; then, the constraints and the objec-
tive functions are expressed.

4.1 DEFINITION OF VARIABLES

The following variables are used in the formulation of the constraints and objective
functions.

index referring to the origin (decision point) of a trip; i varies

from 1 to N; i = 1 is the freeway mainline origin.

index referring to the destination (decision point) of a trip; i

varies from 1 to M; j = M is the freeway mainline destination.

k index referring to a subsection of the freeway.

q index referring to a link of the street.

J3 kij (Yqij) variable taking the value 1 if subsection l.s. (or link g) is on the
freeway (or street) path from origin ito destination i. It takes
the value 0 otherwise.
 314

j=l to M; Types I and II users' demand in vph.

i=2 to N: Types III users' demand in vph.

i=2 to N; j=l to M-l: Types IV-a users' demand in vph.

FLOWq Type IV-b users' demand in vph.

LOAD .. incremental load of the Type IV-a users' demand to be assigned

IJ
at the end of the iteration

cumulative portion of the Type IV -a users'demand loaded at the

end of the iteration.

Type IV-a users' demand, in vph, already assigned on the

freeway route between I and j, at the beginning of the iteration.

X.. incremental flow, in vph, of the Type IV-a users' demand to be

IJ
assigned on the freeway route between I and j, at the end of the
iteration.

Ri ramp metering rate, in vph, at on-ramp I (corresponding to

decision point j).

FR.. trip travel time, in min., from origin I to destination j through

IJ
the freeway route, including on-ramp delay.

ST.. trip travel time, in min., from origin I to destination 1 through

IJ
the surface street route.

6 .. saving time, in min., by using the freeway route, fromi tOi.

IJ
6 .. = ST .. - FR..
IJ IJ IJ

TRAV travel time, in min., through link 9..

q
TS length of a time-slice, in min.

available freeway capacity, in vph, on subsection Is. at the

beginning of the iteration.

SCAP a vailable street capacity, in vph, on link 9. at the beginning

q
LFRij length, in miles, of a trip between origin I and destination i
via the freeway.

LFF .. length, in miles, of the freeway portion of a trip between

IJ
origin 1. and destination 1. via the freeway.
 315

LST .. length, in miles, of a trip between originl and destina-

1J
tion j via the street.

The fo11owing variables must be initialized as shown below:

CLDij set to O.

ASGNij set to O.

FCAPk set to the "straight-pipe" capacity of subsection k minus

N
i':/kiM DiM •

SCAPq set to capacity of the corresponding approach at the down-

stream intersection minus FLOWq'

4.2 FORMULATION OF THE CONSTRAINTS

The constraints mentioned in the control-assignment interaction are expressed as fo11ows,

for every incremental load (or iteration):

(1) No more than the incremental load of the Type IV-a users' demand is
assigned.

0::: Xij ::: LOADij for a11 type IV-a trips.

(2) No freeway subsection is congested.

N-1 M-1
~ ~ {3k.. • X.. ::: FCAPk for a11 subsections.
i=2 j",=i 1J 1J

(3) No street link is congested.

N-1 M-1
~ ~ 'Y.. • (LOAD .. - X .. ) ::: SCAP for all links.
i=2 j",=i q1J 1J 1J q

(4) No total on-ramp delay is larger than the total aggregate saving
in trip travel time for all users of the corresponding on-ramp.

M-l
~ (ASGN. j + X .. ) . 0 .. + D. M . VAL
j"'=i 1 1J 1J 1

TS2 . (~-1 (ASGN ..

j "'=1 1J
+ X .. ) + D. M -
1J 1
R.) for all on-ramps
1

where VAL max

j "'=i
 316

(5) No ramp metering rate can be outside of a preselected range of

values.

VAL :5 Ri :5 max Rate i for all on-ramps

where VAL = max (min Rate i , DiM)

(6) No ramp metering rate exceeds the corresponding on-ramp demand.

M-1
~ (ASGN .. + X .. ) + D' M ?: R. for all on-ramps
1J 1J 1 1
j ?:i

The following variables must be updated at the end (or at the beginning in the case
of CLD .. ) of each iteration as shown below:
IJ
N-1 M-1
FCAPk = FCAPk - ~ ~
13 k ij X1J..
i=2 j ?:i

N-1 M-1
SCAPq SCAP ~ ~ (LOAD ij - Xij )
q
i=2 j ?:i

ASGN .. ASGN .. + X ..
IJ IJ 1J

CLD .. CLD .. + LOAD ..

IJ IJ IJ

4.3 FORMULATION OF THE OBJECTIVE FUNCTIONS

For each of the eight measures of effectiveness mentioned earlier, the corresponding
objective function may be expressed as follows:

(1) Total saving in trip travel time for Type IV-a users, in min. per hour:
N M-1
maximize ~ ~ (ASGN ij + X ij )· 0ij
i=2j?:1

(2) Total system travel time, in min. per hour:

N M-1
minimize CST - ~ ~ (ASGN .. + X .. ) • 0 ..
i=2 j?:i IJ IJ IJ

M-1 N
where CST L D.. ' FR.. + l' ~ 1 D1' M . FR1' M
j=1 IJ IJ

N M-1
+ L FLOW TRAV + ~ L CLD ..• ST ..
q q q i =2 j?: i IJ IJ
 317

(3) Total number of vehicles diverted to the surface street, in vph:

N M-l
minimize ~ ~ (CLD"" - ASGN""
1) 1)
i=2 j~i

(4) Total vehicle-miles diverted to the surface street, in vehicle-miles

per hour:
N M-l
minimize ~ ~ (CLD i )" - ASGN i )" - Xi)") • LST i )"
i=2 j ~i

(5) Total ramp delay, in vehicle-min. per hour:

N 1 2 (M-l )
minimize "~ (120)· TS· "~" (ASGN ij + Xij) + DiM - Ri
1=2 J~I

(6) Ramp queue length, in vehicles:

N (M-l
minimize ~
i =2
(TS).
60
~
j ~i
(ASGN + Xl")") + DI"M - RI")
. ij

(7) Freeway input rate, in vph:

N M
maximize ~ R. + ~ D1)"
1
i=2 j=l

(8) Freeway throughput, in vehicle-miles per hour:

N M-l
maximize ~ ~
i=2 j ~i

M
where CST ~ D 1)"
j=l

4.4 EXAMPLE

In summary, the complete formulation of the interaction might be written as follows,

using in this example, freeway input rate as the measure of effectiveness:

Find, at the end of each iteration, X""' s and R"' s which

1) 1

N M
maximize ~ R. + ~ D 1)"
i=2 1 j=l
 318

subject to:

o :s X .. :S LOAD .. for i=2 to N; j=1 to M-1

1) 1)

for i=2 to N

where VAL = max (min Rate i , DiM)

N M-1
~ ~ i3 ki).· Xi)· :S FCAPk for all .!i
i=2 j 2':i

N M-1
~ ~ r qi ).· (LOAD i ). - Xi).):S SCAPq for all .9.
i=2 j 2':i

M-1 1 2
~ (ASGN .. + X .. )· 6 .. + DiM· VAL 2': (120) . TS
j2':i 1) 1) 1)

• (~-1
j 2':i
(ASGN .. + X.:) + DiM - Ri)
1) 1)
for i=2 to N

where VAL = max (6 .. )

j 2':i 1)

M-1
~ (ASGN ij + Xij ) + DiM 2': Ri for i=2 to N
j 2':i

5. CONCLUSION

This paper proposes a methodology for evaluating freeway corridor assignment equil-
ibria and their effects on corridor control. The freeway corridor is presented in
terms of definitions, assumptions, and description. Then, the corridor model is
introduced as a combination of a freeway model (FREQ3), a street model (TRANSYT)
and a decision model. Each of these models is discussed and particular attention is
given to the decision model which pictures the control-assignment interaction. Final-
ly, the formulation of the control-assignment interaction is presented.

This corridor model is in its development stage. Future work includes the program-
ming of the decision model and the integration of freeway, street, and decision models.
 319

It is expected that such work will provide traffic engineers and highway officials with
a beneficial tool to be used for predicting the performance of alternative control
policies and corridor configuration prior to field implementation. It will also encour-
age improved usage of existing facilities.

The authors wish to acknowledge the assistance of Wayne Kittelson in making numerous
comments about the ideas expressed here.

REFERENCES

1. May, A. D., and W. E. Gillfillan, Bay Area Freeway Operations Study. First
Interim Report. Berkeley: University of California, Institute of Transportation
and Traffic Engineering, January 1968.

2. Ybarra, W. A., and A. D. May, Bay Area Freeway Operations Study. Second
Interim Report. Berkeley: University of California, Institute of Transportation
and Traffic Engineering, July 1968.

3. Salam, Y. M., W. E. Gillfillan, and A. D. May, Bay Area Freeway Operations

Study, Third Interim Report. Berkeley: University of California, Institute of
Transportation and Traffic Engineering, December 1968.

4. Jacobs, R. R., and W. E. Gillfillan, Bay Area Freeway Operations Study. Fourth
Interim Report. Berkeley: University of California, Institute of Traffic and Trans-
portation Engineering, December 1968.

5. Jacobs, R. R., Bay Area Freeway Operations Study. Fifth Interim Report.
Berkeley: University of California, Institute of Transportation and Traffic Engi-
neering, February 1969.

6. Hom, R. E., and W. E. Gillfillan, Bay Area Freeway Operations Study. Sixth In-
terim Report. Berkeley: University of California, Institute of Transportation and
Traffic Engineering, March 1969.

7. Jacobs, R. R., Bay Area Freeway Operations Study. Seventh Interim Report.
Berkeley: University of California, Institute of Transportation and Traffic Engi-
neering' March 1969.

8. Gillfillan, W. E., Bay Area Freeway Operations Study. Eighth Interim Report.
Berkeley: University of California, Institute of Transportation and Traffic Engi-
neering, March 1969.

9. Bay Area Freeway Operations Study, Summary of the Comprehensive Phase.

Berkeley: University of California, Institute of Transportation and Traffic Engineer-
ing, March 1969.

10. Makigami, Yasuji, Lee Woodie, and Adolf D. May, Bay Area Freeway Operations
Study - Final Report. Part I of III: The Freeway Model. Berkeley: University of
California, Institute of Transportation and Traffic Engineering, August 1970.
 320

11. Allen, B. L., and A. D. May, Bay Area Freeway Operations Study - Final
Report. Part II of III: On the San Francisco-Oakland Bay Bridge. Berkeley:
University of California, Institute of Transportation and Traffic Engineering,
August 1970.

12. Aidoo, J. F., R. W. Goedhart, and A. D. May, Bay Area Freeway Operations
Study - Final Report. Part III of III: On the Eastshore Freeway (1-80) North-
bound. Berkeley: University of California, Institute of Transportation and Traffic
Engineering, August 1970.

13. Faravardeh, Mohsen, Freeway Operations Study - Phase III. Interim Report
No.1: A Suggested Freeway Corridor Model. Berkeley: University of California,
Institute of Transportation and Traffic Engineering, July 1972 (unpublished).

14. Wang, J. J., and A. D. May, Freeway Operations study - Phase III. Interim
Report No.2: Analytical Techniques for Ramp Control. Berkeley: University of
California, Institute of Transportation and Traffic Engineering, June 1972.

15. May, A. D., and J. H. James, Freeway Operations Study - Phase III. Interim
RruJort No.3-I. Cost-Effectiveness Evaluation of Freeway Design Alternatives.
Berkeley: University of California, Institute of Transportation and Traffic Engi-
neering, July 1972.

16. Torpp, O. C., and A. D. May, Freeway Operations Study - Phase III. Interim
Report No. 3-2: Extended Analytical Techniques for Evaluating Freeway Opera-
tions Improvement. Berkeley: University of California, Institute of Transporta-
tion and Traffic Engineering, June 1972.

17. Blankenhorn, R. C., and A. D. May. Freeway_9Jlerntions Study - Phase III.

In.terim Report No.4-I: FREQ2 - A Revision of the "FREEQ" Freeway Model.
Berkeley: University of California, Institute of Transportation and Traffic Engi-
neering, June 1972.

18. Yeh, San Fu, and A. D. May, Freewl!Y Operations Study - Phase III. Interim
Report No. 4-2: Investigation of Capacity and Flow-Speed Relationship of Free-
way Model. Berkeley: University of California, Institute of Transportation and
Traffic Engineering, June 1972.

19. Stock, W. A., R. C. Blankenhorn, A. D. May, Freeway OperfJ.lions Study -

Phase III. Report 73-1: The FREQ3 Freeway Model. Berkeley: University of
California, Institute of Transportation and Traffic Engineering, June 1973.

20. Eldor, M., and A. D. May, Freewav Operations Study - Phase III. Report 73-2:
Cost-Effectiveness Evaluation of Freeway Design Alternatives. Berkeley: Univer-
sity of California, Institute of Transportation and Traffic Engineering, September
1973.

21. Wang, J. J., and A. D. May, FreewayQperations Study - Phase III. Report 73-3:
Analysis of Freeway On-Ramp Control Strategies. Berkeley: UniverSity of Cali-
fornia, Institute of Transportation and Traffic Engineering, June 1973.

22. Orthlieb, M., S. Roedder, and A. D. May, Freeway Operations Stygy - Phase III.
Report 73-4: Progress Toward a Freeway Corridor Model Berkeley: University
of California, Institute of Transportation and Traffic Engineering, September 1973.
 321

23. May, A. D., Freeway Operations Study - Phase III. Report 73-5: Optimization
Techniques Applied to Improving Freeway Operations. Berkeley: University of
California, Institute of Transportation and Traffic Engineering, June 1973.

24. Dowling, R. G., and A. D. May, Freeway Operations Study - Phase IV. Report
74-1: Data Collection and Analysis for the Bastshore Freeway (1-80) Corridor.
Berkeley: University of California, Institute of Transportation and Traffic Engi-
neering' June 1974.

25. Dupuy, A., and A. D. May, Freeway Operations Study - Phase IV. Report 74-2:
Data Processing Computer Programs. Herl{eley: University of California, Insti-
tute of Transportation and Traffic Engineering, June 1974

26. Dupuy, A., Kittelson, W. K., and A. D. May, Freeway Operations Study -
Phase IV. Report 74-3: Evaluation of Selected Surface Street Network Models.
Berkeley: University of California, Institute of Transportation and Traffic Engi-
neering, June 1974.

27. Kittelson, W. K., and A. D. May, Freeway Operations Study - Phase IV. Re-
port 74-4: Development of a Freeway Diversion Algorithm. Berkeley: Univer-
sity of California, Institute of Transportation and Traffic Engineering, August
1974.

28. Orthlieb, M., and A. D. May, Freeway Operations Study - Phase IV. Report
74-5: Freeway Corridor Control Strategies. Berkeley: UniVersity of California,
Institute of Transportation and Traffic Engineering, December 1974.

29. May, A. D., Freeway Operations Study - Phase IV. Report 74-6: Summary of
Freeway Corridor Research. Berkeley: University of California, Institute of Trans-
portation and Traffic Engineering, December 1974.

30. Wardrop, J. G., "Some Theoretical Aspects of Road Traffic Research," Proc.
Inst. Civil Engineers, Road Paper No. 36, London, 1952.

31. Robertson, D. I., "TRANSYT: A Traffic Network Study Tool," Transport and
Road ReRearch Laboratory Report. LR 253, Crowthorne, 1969.

32. Lasdon, L. S., Optimization Theory for Large Systems. MacMillan Co., 1970.

33. Steel, M. A., "Capacity Restraint - A New Technique," Traffic Engineering and
Control, October 1965, pp. 381-384.
 SUPPLY FUNCTIONS FOR PUBLIC TRANSPORT:
INITIAL CONCEPTS AND MODELS*

Edward K. Morlok
University of Pennsylvania
Philadelphia, Pennsylvania, USA

ABSTRACT

In this paper, supply functions are developed for describing urban public
transit services. These functions relate the level of service as it would be per-
ceived by a traveler to the flow on a route, in a manner conceptually similar to
that of the automobile - road supply curves which are widely used in network models.
Models are derived for the supply characteristics of conventional scheduled, fixed
route, transit services. The models appear very plausible and seem to explain
generally observed differences among various transit systems. Data on a few routes
in Chicago and Philadelphia permitted limited testing of the models, and the models
performed very well.

* The research underlying this paper was partially supported by the U.S. Depart-
ment of Transportation, through Urban Mass Transportation Administration grant
no. URT-20 (70), to Northwestern University, and Program in University Research
contract no. DOT-OS-40092 to the University of Pennsylvania. This support in
no way implies endorsement of the findings or conclusions by the sponsors.
 323

Purpose

In this paper supply functions are developed for describing urban

public transit services. These functions relate the level of service as
it would be perceived by a traveller to the quantity of flow on the transit
route. These functions were developed with a view toward their inclusion
in the network equilibrium models commonly used in the urban transportation
planning process. Such supply functions for public transit are necessary
in order that both major urban travel modes are treated using the same
paradigm: that of supply and demand interacting to yield an equilibrium
flow pattern.
After an initial discussion of the need for supply models for public
transit, the various forms which these might take are identified and dis-
cussed. Then attention is focused upon building a theory of supply of
transit services on a given route and with a given transit technology.
Theoretical functions are developed, primarily with the bus mode in mind.
These are then tested using data from the Chicago area on bus services and
also with data for Philadelphia area railroad commuter services. Finally,
the use of these models in urban transportation planning is discussed and
aspects requiring further research are identified.

Background
Need
Almost all of the transportation system models used in urban trans-
portation planning as well as in many other contexts are based on the conceptual
 · 324

framework of network equilibrium in order to predict the traffic flow

pattern and the level of service associated with each alternative future
transportation system. The underlying concept of supply and demand inter-
acting to yield equilibrium flows has evolved in a period in which the primary
focus of transportation planning was on road investment planning. In this con-
text demand and supply functions have been extensively developed. Since
these planning efforts usually are focused upon alternative fixed plants,
such models usually treat the road fixed plant as determined exogenously.
For any such specified alternative, the network equilibrium model is used
to estimate the resulting flow on each element of the system and the quality
of service associated with it. Since the volume of traffic on each individ-
ual facility will determine the level of service provided by that facility,
given its design characteristics, it is necessary to represent the level of
service as a function of volume for each link in the system. This is done
by using what is commonly referred to as a supply function, a function re-
lating level of service to volume. The underlining theory for such relation-
ships has been well developed by traffic engineers and flow theorists and
such relationships seem to be quite reliable for purposes of estimating
quality of service.
In the case of public transit, it would be expected that some aspects
of the quality of transit service would be determined, at least in part, by
the volume of traffic on the route or system. Thus, in order to treat
public transit within the same conceptual framework as the road system, it
is necessary to be able to specify supply functions for public transit
services. However, no such functions exist. Rather, public transit is
 325

usually modelled in such a manner that all quality of service and capacity
characteristics are determined exogenously to the network equilibrium model.
Then traffic is assigned to the network and an equilibrium flow determined,
assuming that these characteristics are fixed. The previously assumed transit
service and capacity then must be checked for consistency with the resulting
traffic flow, and corrections made through iteration. Such iterations usually
are performed only rarely and consist primarily of adjusting transit capacity
to demand.
There are two major problems with this approach to the modelling of
transit. The first is that often it is impractical to iterate to the extent
required to achieve compatibility or consistency between the assumed level
of service and capacity and the actual flow. The second is that there is no
basis for assessing the compatibility between the transit service provided
and the volume of traffic with the exception of comparing capacity with
traffic. Thus it is not possible to adjust transit service to the traffic
on other than an ad hoc basis.

Supply Function Concepts

The supply of transportation services in network equilibrium models
depends upon the values chosen for two different types of variables, exog-
enous variables and endogenous variables. In modelling the road network,
the exogenous variables are those variables which the planner assumes to
have directly under his control, the construction of new links and improve-
ment of existing links. These exogenous variables can then be varied so
that alternative improvement schemes can be evaluated. The endogenous
variables describe the level of service on each facility as it responds
 326

to the volume placed upon it. In the highway case, this response is described
by the supply function which relates the level of service to the volume of
traffic on each link or other element of the network. The design is specified
by the exogenous variables. Underlying such supply relationships usually are
assumptions regarding the operational aspects of road system management, such
as the timing of traffic signals to optimally acommodate the traffic which
flows through an intersection. In principle, various operational choices such
as the timing of signals could be treated as exogenously determined variables,
in order to evaluate different timing strategies, but this is not normally
done because planners are not interested in questions of such a detailed nature.
Presumably any intersection supply function is based upon expected adjustments
of signal timing to traffic load, thereby in effect assuming certain behavior
on the part of road traffic engineers.
In the case of public transit there are even more options with

respect to the range of supply variables and hence options as to whether

they are treated as exogenous or endogenous. The traditional focus of urban
transportation planning for public transit has been upon aggregate route
structure options, with plans specifying only the most general characteristics
such as major route location, primary station stop locations, overall travel
time, and technology to be employed. In such a situation supply functions
would be needed to specify the quality of service and the capacity which actual-
ly would be provided by vehicle movements on each route. Presumably, these
functions would represent the behavior of transit management in selecting among
various available route operation options. In the extreme situation where the
only concern of the transportation study is to choose among alternative transit
 327

fixed plants such as rail and busway routes, then presumably even the route
structure for the portion of the bus system which uses the road network in
mixed traffic would be determined by the transit management and hence would
logically be treated as part of or endogenous to the supply functions. How-
ever, more recently urban transportation planners have been much more con-
cerned with developing plans which are immediately implementable and which
would yield immediate benefits. Thus they are increasingly concerned with
the details of transit services, including the route structure and certainly
to some extent the pricing and service provided on each route. In this con-
text, fewer of the service variables would be modelled as endogenous to the
supply function, with more of these variables being exogenously determined
so that alternative choices could be evaluated.
The various possible forms for transit supply functions,with particular
emphasis on the division between endogenous and exogenous variables, are shown
in Figure 1. Among these possible forms, the most immediately useful are probably
those which describe the supply of transit service on a single route, given the
location and the technology. With such a supply function it would be possible
for those using the equilibrium model to explore explicitly alternative route
structures and the technologies or modes to be used on those routes. But it
would relieve them of the requirement of specifying detailed transit schedules
and capacity levels. Planners could then focus upon questions of routes and
technologies, questions which seem most appropriately addressed in a planning
effort, since planning studies are usually the source of new investment pro-
grams and also usually must approve any specific investment. However, the
supply functions built into the network model would attempt to describe the
actual service to be provided by the transit management on the given route
 328

structure and with the given technologies.

Thus a model which describes the supply of transit services on a
particular route,with the routing and the transit technology to be used
specified, seems to be the most appropriate form for initial development
of public transport supply functions. This type of transit supply model
will be the focus of this paper.

Transit Management's View

Most transit services in North America are provided by public
authorities and the cost of these services usually exceeds by a consider-
able amount the revenues derived from transit services. Strong pressures
exist to minimize the deficit, which must be met primarily out of general
tax revenues. This induces a very strong cost minimization orientation
on the part of transit operators. The cost is largely determined by the
weekday peak period operation, because that determines both the number of
vehicles and the number of drivers or other operating personnel required.
Thus much of management's attention will be directed toward the peak period.
Regulatory control of the services provided by transit authorities is minimal,
although sometimes a minimum service frequency or a requirement to operate
a particular unremunerative route is imposed. While management usually has
a great deal of discretion over the service provided, typically, it has
little control over prices, these being largely politically determined
along with the amounts of subsidies.
Transit management usually has relatively little information on the
demand curve which it faces for its services. Rather, it has information
on the actual flow on its various routes, in the form of counts of
 329

passengers past the peak load pOints on each route during the peak period
and sometimes during other periods. Usually no information is available on
the actual origin and destination pattern of traffic or on other demand
characteristics such as the trip length distribution. Also, there
is very little information on the sensitivity of traffic on a route to
service or price changes. This is partly due to the few changes which are
usually made in transit services over a long period of time, and also part-
ly to an apparent distrust or lack of knowledge of demand models and partic-
ularly of the mode choice models used in transportation planning.
Discussions with those responsible for the scheduling of transit routes
in a few large cities - Chicago, New York and Philadelphia - revealed they at-
tempt to schedule bus routes in the following manner. During the week-day peak
period, an attempt is made to operate a number of buses just sufficient to ac-
commodate the traffic. This implies a standard for the capacity of a bus, in
most cases a nominal capacity of 75 passengers for the most common 53 seat bus
was expressed. However, it was stated that a smaller value for capacity is
used in cases where the typical trip is quite long, and where there seems to be
a large number of passengers boarding and alighting (this requiring relatively
free movement in the aisles). However, such adjustments are made on a rather
ad hoc basis. Also influencing the scheduling is a requirement to operate ve-
hicles at headways no greater than a specified maximum. In Chicago this maximum
headway was expressed as ten minutes during the peak periods. In the other
cities the maximum headway seems to depend upon the type of route. Outside of
the peak periods the attempt to fully utilize the capacity of each vehicle was
far less strong, the minimal frequency requirement usually resulting in route
capacity greater than actual traffic.
 330

No particularly generalizable expressions of policy regarding the

introduction of any type of skip-stop or express service were made. Most
comments in this regard were that such services have existed for such a long
period of time that the question of introducing new ones or eliminating old
ones rarely arises. Moreover, the service innovations along these lines seem
to be treated as special cases with the planning of them requiring special
studies and often the collaboration of the regional transportation planning
agencies.
While these statements of general guidelines in the scheduling of
buses provide some information and insight into the scheduling process, and
hence into the supply of public transit services, they are certainly far from
complete in dealing with the various options available to transit management.
Thus, it seems appropriate to turn to the development of plausible rules of
managerial behavior in supplying transit service and then attempt to test
these against data on services actually provided.
 331

Local Service Only Bus Routes

Definition
In developing quantitive models and hypotheses about the supply
of transit service it is useful to begin with the simplest possible
case, that of a single bus route in which all of the buses operate in
local service from one end of the route to the other, using public
streets in mixed traffic. Such a route is shown in Figure 2a (2b
will be referred to in a later section). In this case it will be as-
sumed that transit management has no control over the running time of
buses from one end of the route to the other, this time being deter-
mined by the transit vehicle's acceleration and speed capabilities
and the prevailing speed of traffic on the roads comprising the route.

Management Behavior
In this context, the only decision variable open to management is
that of how frequently to operate buses. The general discussion of
managerial behavior above suggest that the management would follow the
following two rules: First, buses would be operated with a frequency
at least equal to the minimum frequency considered acceptable for the
transit service.
f~ F =~ (1)

where: f = frequency of bus departures in one direction, buses per hour

F = minimum acceptable frequency, buses per hour
H = maximum acceptable headway, hours per bus
 332

Since the volume of passengers could exceed that which can be ac-
commodated in buses operated at the minimum frequency, the frequency
also has to be greater than or equal to that which is required to ac-
commodate the passenger flow during any particular period.

f~ ~ (2)

where:
p = passenger flow past peak load point on route, pass. per hour
Q = capacity of bus, passengers per bus

Assuming that management operates the minimum number of buses in

order to meet these two service-related criteria, the rule which manage-
ment should follow in determining the ,frequency of busses would be the
following:

f = max (F, ~) (3)

Following such a rule would probably minimize the cost of

operating the service, since the number of buses, operators, and ve-
hicle-miles would be minimized for the given route conditions.

Although the rule stated in equation (3) above is expressed in

the form applicable to the weekday peak period, the same
rule could be applied to any other period in which service is operated.
However, it is likely that the acceptable vehicle capacity, Q, would
be smaller in the weekday base period and on weekends and holidays due
to a desire to have fewer or perhaps no passengers standing. Of course,
as a practical matter, transit routes are rarely so heavily used in such
non-peak periods that vehicles are rarely operated at capacity. Also,
the minimum frequency, F, may vary by periods of the day.
 333

A further complication is that the rate of passenger flow may vary with-
in any period. This would lead to a non-uniform headway if more than the mini-
mum frequency service is required, i.e., passenger flow is greater than FQ, and
if management policy permitted non-uniform headways. If a constant headway
were required, then headway would be adjusted for the peak passenger flow with-
in each schedule period. This would lead to an average load less than vehicle
capacity, and could be incorporated in the model by appropriate selection of
the value of Q. It also should be noted that the number of bus trips made
over the entire day must be an integer, thereby possibly requiring a slight ad-
justment in the frequency of operation in each period, but the adjustment is
small in comparison to typical total frequency values and therefore is ignored
here. Similarly, variations in the passenger flow during anyone period will be
ignored here, in effect presuming that periods are selected to correspond to
roughly constant flow rates.
Supply of Services
Given the managerial behavior described in equation (3) above, is is then
possible to estimate the supply of transit service on a route as it would be
viewed by a traveller. One important aspect of the supply of service would be
the travel time from origin to destination, for any particular traveler, in-
cluding the waiting time as well as on-vehicle time. Assuming uniform or random
passenger arrivals at the origin stop and constant headways, the average waiting
time would be one half the headway.
1 60
w = 2" h = 2f (4)
where:
h = bus headway, minutes per bus
As can be seen from this expression, as the passenger traffic in-
creases above that amount required to fill the minimum frequency of buses,
travelers' waiting time would decrease. Assuming that the volume of traffic
 334

is sufficient to fill all the vehicles at the minimum frequency, the

average number of passengers per vehicle past the peak load point will
be independent of the traffic volume. Assuming the origin-destination
pattern of traffic does not vary with volume, the same number of passen-
gers will board each vehicle regardless of volume. Therefore it may be
assumed that the number of stops and the dwell time is independent of
volume and hence the travel time will be independent of volume. In this
case travel time between any points and j equals
1 (5)
Vij + 2 h
I
tij
where:
V~j vehicle running time between stops i and j, plus one-half
the dwell time required for alighting at j, minutes
tij = total travel time between stops i and j, minutes

The Vij term must include one-half the time required for the
alighting of passengers at j, since this is part of their time on the
system. However, a comparable inclusion of boarding time
at i is not required,since this is already accounted for in the waiting
time.
Two points regarding the Vij term should be made. First, in the case
where the passenger traffic is less than sufficient to fill the minimum number
of vehicle trips operated (the minimum frequencYh then the travel time
presumably will be slightly less, reflecting a diminished value of Vj, due
to fewer stops and shorter dwell time for loading and unloading. Of course,
this assumes a timetable adjustment by the management which may not be made.
Such an adjustment will not be considered further here. The portion of Vij
due to the time required for unloading at stop j is also likely to be very
 335

small, particularly since most urban buses have two or more doors, both
of which are usually used for unloading whenever the number of alighting
passengers is large. The maximum time which would be included would then
be the time required to unload 75 passengers through two doors, about 40
seconds*. Since most values would be less than this, in further discussion
this unloading time will be ignored, and Vij replaced with Vij , the running
time of the bus from departure at i to arrival at j.
Virtually, all studies of modal choice have revealed that passengers
value waiting time considerably more than on-board vehicle time. Thus the
perceived average travel time for any traveler between stops i and j will
be the following:
_ W60 _ + 30W (6)
t ij - Vij + ~ - Vij max (F, p/Q)
where:
t ij = total perceived travel time between i and j, equivalent minutes
W relative weight of waiting to on-board vehicle time
VlJ
.. = vehicle running time from ito j, minutes
Most studies indicate that the value of Wis 2.5 to 3.0 (e.g., Pratt, 1970).

Example
An understanding or appreciation of how total travel time as perceived
by a traveler would vary with volume on an all-stop local service bus route
can be obtained from an examination of relationships for a particular hy-
pothetical route. Assuming a route which is 11 miles long, for 10 miles of
which the bus can operate at a speed of 12 miles per hour (including stops) and
one mile of which is in the central business district with an average speed
*This is based on the assumption of 38 passengers alighting via each
door, using the recently estimated regression formula for the A.M. peak
dwell time (eqtn. A3) presented in Kraft and Bergen (1974).
 336

of 6 miles per hour, and a minimum service frequency of 6 buses per hour
(equivalent to a maximum headway of 10 minutes), the relationship between
travel time and volume is as shown in Figure 3. Passengers are assumed
to originate with a uniform distribution in the 10 mile portion and be
destined with a uniform distribution along the one mile portion within
the central business district. Buses are assumed to have maximum capacity
of 75 passengers.
Figure 3 shows the relationship between average total travel time as
measured in equivalent minutes and the volume of passenger traffic on the
route. The average of all passengers follows the line for the travel time
between points Band D. As can be seen, the general shape of the travel
time vs. volume relationship is one in which the travel time remains con-
stant until the capacity provided by the minimum frequency of vehicle trips
is fully utilized, after which the increasing frequency required will result
in a reduction in waiting time. This reduction can be fairly substantial.
In this example, for volumes up to 450 passengers per hour the average total
equivalent travel time is 45 minutes; a traffic increase to 1,500
passengers per hour reduces the time to 34.5 equivalent minutes, a 23.3%
reduction. Of course, because the reduction in travel time with increasing
volume is the same regardless of the length of the trip, the percentage de-
crease is much more significant for shorter trips. This example serves to
illustrate the substantial influence that increasing frequency will have on
reducing waiting time and hence equivalent passenger travel time.

Cost to Operator
Many models of bus cost have revealed that there are basically three
determinants of the cost of operating any given bus route. These are: the
number of buses, the number of drivers or driver hours, and the bus
 337

vehicle-miles of operation. Continuing to focus only on week-day peak

periods, the number of buses required to operate the peak period service
is given by the following relationship:

B
f<Rf) , R~ P
l<Pf) , R~ P
(7)

where:
B = number of buses required
R = round trip time of buses on the route, hours
P = duration of each peak period, hours
() denotes the smallest integer containing the number inside the brackets

Assuming we are only interested in the cost associated with or the marginal
cost of peak period operation, the relevent number of drivers required is
equal to the number of buses.*
The number of vehicle miles operated for the service durin!] the t\10 peak
periods is given by the following formula, again assuming all trips are from
one end of the line to the other and return:

M = 2 P f (2L) (8)

where:
M= vehicle-miles operated
L = one-way route length, miles
For application of equations (7) and (8), fP should be an integer, presumably
the smallest integer containing the required number of trips.
The cost for each of these items for transit service in Chicago has been deter.-

*It is interesting to note at this point that reductions in the number of ve-
hicles and drivers required for peak period operation will result from speed
increases (reductions in round trip times) only if R ~ P. However, if R > P
additional vehicles will be required to operate the base period service, the
number being «(R-P) fQ>J fo being the base.frequency. Thus reducing R can reduce
total service vehicle requirements, but if R > P the reduction in fleet per
unit reduction in the round trip time will generally be less than if R , P.
 338

mined (Mor1ok et.a1, 1971). Since the new purchase price of a modern 53 seat
bus is $ 40,000, and this vehicle typically has a life of 12 years and
there is an 8% maintenance reserve, the cost per available vehicle day
is approximately $ 20.31. The cost of hiring a driver is $ 41.13 per
day, this being a minimum which must be paid to the driver regardless of
the number of hours worked. If he works more than eight hours within a
period of 10 1/2, then overtime must be paid. The average cost of oper-
ating newer vehicles, per vehicle mile, is $ 0.423. Knowing these values
it is then possible to estimate the total cost to the owner or operator
of the transit system for service on this hypothetical route, using the
formula
C = o(.B + (3 B + ¥ M (9)

where:
0(. = cost per ava i1 ab 1e bus day
~ = cost per driver day
~ = cost per vehicle-mile
The results for the hypothetical route are shown in Figure 4. As
can be seen, the cost per passenger mile begins at low volumes at a
fairly high value of 20 cents but this drops as traffic increases, up
to 450 passengers per hour, reflecting the fact that more passengers are
using the system (passenger miles generated) while costs are remaining
constant (because the number of buses operated remains constant). Beyond
450 pass. per hour the cost remains approximately 10 cents per passenger
mile, the minor variations reflecting the requirement of an integer number
of vehicles and vehicle trips.
 339

Total Transportation Costs

A very commonly employed concept in transportation analysis is that
of total transportation cost, which includes the actual monetary cost of
providing the service and an evaluation of the costs borne by the user or
traveler. The latter costs consist primarily of travel time. Assuming
that passengers value time at a rate of $ 3 per hour, and continuing with
the assumption that waiting time is valued at three times the on-vehicle
time, the total cost· of providing transport on this bus route is also
shown in Figure 4. The total transportation cost drops steadily with in-
creasing volume, although declining at a decreasing rate with increasing
volume. This pattern is primarily due to the effect of increasing fre-
quency on reducing the waiting time of passengers, although the reduced
operator cost per passenger mile in the lower volume range also exerts an
influence. The reduction in cost is significant, total transportation costs
per passenger mile at 2,000 passengers per hour being approximately one-
third less than that at 225 passengers per hour.
 340

Zone Express and Local Bus Service

Definition
In operating a bus service along a particular route, transit management
can choose not only the frequency of service to be provided with a given
stopping policy, but also select the particular pattern of stops which buses
will make. Many bus routes in North America are divided into two or more
sections during heavy volume periods. Buses may be operated between the
central business district and the closer of the two sections in local service,
rraking all stops ,as shown in Figure 2 b. In order to provide service to the
outer section, some buses may operate from the CBD, passing through the inner
zone without stopping (operating as an express), then making all stops in
the outer zone. This wpe of service is called zone service. There may be
more than two zones.
In this manner, the running time of buses to the outer section is sub-
stantially reduced. In general, this benefits the travelers between the
outer section and the central business district. Also,by increasing the
average speed of buses, more revenue trips can be obtained from each vehicle
in any particular time period and vehicle and driver costs are reduced. How-
ever, there are di sbenefits to the zone servi ce, since travel between the
two zones requires a vehicle transfer and since the frequency of service
operated in anyone zone usually will be less than that which would prevail
with an exclusively all-stop route.
Management Behavior
If we assume that the route is divided at a point such that one-half
of the traffic to the central business district originates in one zone and
one half in the other, then the previously applied rules regarding the
frequency with which vehicle trips will be operated become the following:
 341

(10)

where:
e and t refer to the zone express and zone local services, respectively.
This formula is analogous to equation (3), and assumes that the same service
frequency and bus capacity standards will be applied.
Supply of Services
The total travel time of travelers using the zone services will be the
following from either zone

z Vz + 30W z 30W
ti j ij fz Vij + max (F,p/2Q) (11 )

where:

t~j total perceived travel time between and j on appropriate zone

service, min.

Vfj running time of vehicles between and j on appropriate zone

service, min.

fZ frequency of service in each of the zones (here all assumed equall

fZ = fe = f~ , buses per hour.

Provided the traffic is greater than that required to just fill the minimum
frequency service, all travelers will face a reduced frequency of service in
comparison to that which would occur with the all-local service. This results
in an increased waiting time, but the longer distance travelers will ex-
perience a reduced on-board vehicle running time. In general, the average
speed which vehicles can maintain in the inner-zone local service will be
somewhat less than which they could maintain in the non-zone local service,
because more passengers would be boarding in a shorter distance than in the
non-zone case and hence the number of stops and dwell times would be longer.
 342

However, the gains for the longer distance traveler can be very substantial.
Example
The same numerical example as used earlier can be used to illustrate
the effects of the introduction of zone services. These computations are
performed for the same numerical assumptions as in the non-zone all-stop
service case, with the exception that it is assumed that the bus operating
in the purely express mode could average 45 miles per hour, reflecting
freeway type operation. Figure 5 shows the average travel time for persons
traveling between each of the two zones and the central business district.
These times are also shown for the all-stop local service and also for the
zone express or the zone local service. The average time for travelers be-
tween the outer zone (which extends between points A and B and the central
business district) is equal to the time between points F and D. As can be
seen from the figure, the reduction in average travel time for these due to
the zone service is very substantial, reductions of one-third being common.
The increases in time for the travelers in the near zone(which extends be-
tween points B and C) are significant, although the increase is the least
at large volumes. The average travel time between the inner zone and the
central business district is identical to that between pants G and D. At
1,000 passengers per hour, the average times on non-zone and zone services
are 24 and 31 minutes, respectively, while at 2,000 passengers per hour, the
values are 21 and 24 minutes , respectively.
Operator Costs
The costm the owner or operator in providing the service following the
zone pattern can be easily calculated using the same basic formulae used for
the non-zone all-stop service. The formulae for the number of buses (and
the number of drivers) are as follows:
 343

BR, 1(R, f ') , RR, ~ P

t
(12)
(Pf z >, R~P

Be = e r'). Re~ P ( 13)

{Pfz >, R~P

BZ BR, + Be (14 )

The number of vehicle- miles operated is

MZ =2 P fZ (2 Le + 2 LR, ). (15)

The same integer considerations as in the non-zone case apply here.

Using the same unit costs for buses, drivers, and vehicle miles, the
cost incurred by the operator in providing this peak period service has been
calculated. Figure 6 contains the results of these computations for this
particular example. As would be expected,in the lower volume range the all-
stop non-zone service is less expensive, reflecting the greater penalty of
the minimum frequency requirements on the zone-type service. However, once
the volume exceeds 700 passengers per hour, the zone-type service becomes
substantially less costly than the all-stop service, the difference or
saving ncreasing with volume. Thus, it appears that a cost minimizing
transit management would attempt to introduce some type of zone local and
express service on high-volume routes. It is important to nota that the
traffic need not reach that point required to completely fill buses operated
at the minimum frequency on the two zone service for that zone service to
be cheaper than the all-stop service. This is because at some lower volume
the increase in vehicle utilization made possible by the higher average speed
of the zone express service will more than outweigh the increased number of
trips required for that service. Thus the zone service may be less expensive
than all-stop service even if the former requires the operation of partly
empty vehicles.
 344

The management of a transit agency may behave in such a manner that it

attempts to minimize its own cost. If this is the case, it is easy to ascer-
tain that volume of traffic at which the cost of zone and non-zone services are
exactly equal. At greater volumes, 1he zone service is preferred, and at
lower vol umes the non-zone servioe is ..L\ssumi ng p > FO and R~ P, the costs of
the all-stop non-zone and zone services are, respectively

C = (oe. +(3) R ~ + ~ ( 4 PL ~ ) (16)

C z = (oc.+(3){R;, F + Re F) + ~(2PF)(Ll + Le) ( 17)

The critical value of p is then

p = (<x'+I3) (R~l + Re F) +(\'(2 PF)(Lg + Le) (18 )

c (0£.+ ) RIO + 'g (4 PLIO)

As would be expected, as the minimum frequency increases, or the vehicle

capacity increases, Pc increases. As the speed of the express service in-
creases, as might result from priority bus treatment, Pc decreases.
Total Transportation Cost
The total transportation cost for the zone express service can be
compared with that of the all-stop service, using the unit costs presented
earlier in the discussion of the all-stop service. This total transportation
cost per passenger-mile is shown in Figure 7. It is interesting to note
that, for this particular example, the zone service is less costly than the
all-stop service at almost all volumes. This is undoubtedly due to the
high average speed we have assumed for buses in express operation. The
intersection of the total cost curves would surely move closer to the volume
at which the owner-operator cost curves cross (at 725 passengers per hour in
Figure 7), as the speed difference between local and express operation dimin-
ishes. Nevertheless, this illustrates potential benefits through the use of
some type of zone express service instead of a purely all-stop service.
 345

Testsof Bus ServiceSupply Models

There appear to be two primary aspects of the theoretical models of the

supply of bus transit service which are directly testable. The first is that
the frequency of transit service on any given route is related to the passen-
ger traffic volume in the manner described by equation (3). The second is
that zone or express and local service would be introduced at high
volumes, due to the benefit to the service operator and owner
in the form of reduced costs. This latter rule must of course be tempered
by the fact that the express service makes sense only if a very large frac-
tion of the travelers (ideally all of them) are traveling to or from one
end of the 1ine.
Empirical Frequency Relationships
Data were obtained from the Chicago Transit Authority (CTA) on the volume
of peak period traffic and the frequency of service operated on all of their
bus transit routes. These data were collected in May 1972. The data on
flows were manual counts of passengers on board vehicles past the peak load
pant on each route, these peak load points having been determined by earlier
studies. These data were for all 133 distinct bus routes operated by the
CTA at that time. Some of these routes are in effect related, in the sense
that one represents a zone express and the other a zone local, sharing the same
basic route. For purposes of this frequency analysis, any such routes were
considered to be two distinct routes.
These data are plotted in Figure 8. The data for the two-hour morning
peak are represented by dots, for the two-hour evening peak by X's. This
figure reveals a rather consistent linear relationship between the bus trips
operated in each two hour peak period and the number of passengers past
the peak load point in each such period. Also,d~spite the statement of
CTA officials that they attempt to provide a minimum of 12 bus trips in
 346

each two hour period on each route, 27 of the total of 266 observations were
of a frequency of fewer than 12 trips. The goal of filling each bus to capacity
also was not being achieved. As measured over the entire two-hour peak
period, the average ifor all route~ number of passengers per bus past the
peak load point of each route was 43.5 in the morning and 41.0 in the
evening. Most buses seat 45 to 53 persons, and have standing room for 20
to 25 additional people. Perhaps this is due to peaking of passengers within
the two hour period, anda desire to provide an essentially uniform headway ser-
vice such that more than 75 persons are not on anyone bus trip. \~ere this the
case, the average load per bus would be less than the nominal capacity.
Because of the appearance of a linear relationship between bus trips
and passengers in Figure 8, a linear regression was fitted to these data.
In order to ascertan whether there was any difference in management behavior
between the two periods, regressions were separately performed for the
data for each period.
The resulting equation for the morning peak period was

f = 4.24 + 0.0140 p. (19 )

f = 10.92 buses per hour p = 473 passengers per hour
R2 = 82.2%
SE = standard error = 2.27 buses per hour
The high value of R2 and the low standard error (which is 20.8% of 1l1e mean frequency),
indicate that this line represents a reasonably good fit to the data. Also,
the distribution of residuals from the regression line was essentially normal.
The PM peak period regression exhibited very similar characteri~ics, the
estimating relationship being
f = 4.59 + 0.0134 P (20)
f = 10.14 buses per hour p = 415 passengers per hour
R2 = 79.7%
 347

SE = 1.98 buses per hour

The high R2 value,79.7%, and the standard error value, 19.5% of the
mean frequency, indicate that this relationship is also a good fit.
Again the residuals followed the normal distribution.
It is interesting to note that, in both the morning and evening peak
periods, the eTA seems. to be operating a minimum of approximately nine
vehicle trips per two hour period. Also, it seems as though the number
of trips is increased above this value as traffic increases at a rate
corresponding to approximately 70 to 75 passengers per vehicle. However,
the average load per bus is always less than 75 persons,because of the
initial threshold vehicle trips operated.
These results seem to confirm that there is a strong relationship
between the number of buses operated and the volume of traffic on the route.
These data would seemm indicate that the number of buses operated is
more nearly represented as a linear function of volume rather than the
piecewise linear function hypothesized earlier. However, other factors
which are not readily taken into account probably are involved. For example,
there is a definite tendency toward higher bus loadings as the volume of
traffic increases. Secondly, there are numerous routes on which the average
bus loadings are considerably below the 75 passengers per vehicle stated
capacity and even less than the 53 passengers per bus seated capacity which
is found on most eTA buses. The operation of many more vehicle trips than
would appear to be necessary to accommodate the peak period traffic seems
to warrant further inquiry. A possible reason is that the passenger traffic
data are subject to considerable error, thus implying a linear relationship
which really does not exist. Another is that the 27 (of 266) routes and
periods, inWhich the frequency is below the stated policy minimum of six
per hour represent unusual situations, such as a route serving a particular

factory where following the stated policy is not warranted due to some anomaly
in passenger demand.
 348

In any event, the general form of the transit supply function is con-
firmed. It is basically a function in which the total equivalent travel
time decreas~with increasing volume. Table 1 presents the average travel
time for the all-stop route used earlier, assuming that the buses are
operated with the frequency estimated by the eTA peak period functions. Also
presented are the times based on equation (3), letting Q equal 45 persons per
bus, close to the eTA average. As can be seen, the difference between the
eTA data and those of the theoretically-based frequency function is rather
minimal.
Zone Express Services
In order to ascertain at what volume it is advantageous to the operator
to introduce zone service in preference to all-stop local service, it is
necessary to know the travel times of buses with both operating plans. Since
these running times for express and local operation could not be obtained
for most eTA routes (this requiring information on travel time by possible
express routes which generally are not operate~, it is not possible to
evaluate this service criterion precisely. However, it will always be
advantageous for the operator to introduce express service before that volume
of traffic is reached which will just fill the minimum frequency service on
the two zone ser-vices. This can be used an an upper approximation to the critical
value of volume between all-stop local and zone services. This critical
value is 900 passengers in two hours assuming a minimum frequency of 12 buses
per two hours and a 75 passenger capacity for buses. However, the eTA seems
to operate buses consistant1y with fewer than 75 passengers in the peak
period, average loads being in the 40 to 45 passenger range. Therefore,
a critical value of 650 passengers per hour was arbitrarily selected.
First, those routes which the eTA currently operates as zone services were
examined to determine if the flows did in. fact exceed this critical value.
 349

There were 18 routes which when combined appropriately yielded seven zone
express and local service routes. Then the total volume which would have
occurred on a single all-stop local route serving the same points was calcu-
lated, this being the sum of the traffic on the component routes.
This volume was compared to the 1,300 passengers per two hours which would
be required in order to warrant the introduction of zonal express and local
service. In all seven cases the total volume of traffic occurring on the
combined route exceeded this figure, indicating that express service was
provided only on routes where the volume exceeded this critical value.
Then all the remaining data (266 rout~minus the 18above) were examined to
identif9 routes on which the volume during the morning or evening peak
period exceeded 1,300 passengers per two hours. Twelve such routes were
identified. On all of these, only purely local service is operated. Seven
of these twelve routes were clearly not central business district oriented.
On these it is likely that no express and local zone service is provided due
to the fact that traffic is not oriented to a single traffic generating area.
Of the five other routes, one was a very short route feeding a rail rapid
transit station, and the other four paralleled rail rapid transit routes.
It might reasonably be presumed that the routes paralleling rail rapid
transit lines would be used primarily by short distance travelers, while the
longer distance travelers would use the rapid transit line. Hence, the
service benefits to be achieved from express operations would not be expected
to accrue on these four routes.
Thus, the eTA services seem to be consistent wtth the theory presented
earlier regarding those conditions under which express service ought to be
introduced. Although the data are inadequate for a thorough examination of
this question, the results certainly do not contradict the theory.
 350

Suburban Railroad Service

While the theory developed above is primarily directed toward the
provision of bus services, it was felt likely that conventional fixed
route, scheduled rail passenger service would exhibit somewhat similar
supply characteristics. The primarly differences between rail and bus
services for purposes of supply analysis would seem to be (l)the option
of providing express service is much more limited in the case of rail
systems due to limited multiple trackage, (2) the longer typical trip
lengths of rail might make zone services very attractive, and (3) the
capacity of a vehicle unit is not fixed in the case of rail because any
number of cars (up to a technological maximum) may be coupled together
to form a traln.
Data on total daily passenger traffic in both directions on the commuter
railroad lines in the Philadelphia metropolitan area were obtained from the
Southeastern Pennsylvania Transportation Authority (1972). This agency has
responsibility for providing local rail passenger service in this region,
contracting with two area railroads for the services provided. For all
routes data on the number of trains operated per day on each route was
also obtained (Penn Central, 1972, and Reading Lines, 1972).
These data are plotted in Figure 9. A distinction is made in this
figure between the services provided on the Reading Company and the ser-
vices provided on the Penn Central Railroad. As can be seen, from this
figure, there seems to be a consistent relationship within each company's
services between the volume of traffic and the daily train frequency.
The Reading Company services seem to be operated generally with higher
frequencies at any given traffic level than Penn Central services. There
 351

also are some rather substantial deviations from any approximating linear
functions. A possible reason for such deviations may be the different
character of the services provided. Some of the lines are very short, the
shortest being 8.0 miles in length, while others are much longer, up to
58.1 miles. It is reasonable to assume that higher frequencies of service
would be provided on short routes, because the waiting time would be such
a large portion of travel time on these and hence a relatively more im-
portant determinant of the attractiveness of the service.
In order to test the hypothesis that frequency is related to passenger
flow and line length, a multi-linear regression was performed between daily
train frequency as the dependent variable and passenger traffic and line
length as the independent variables. This was done separately for the
services provided on each company. The resulting relationship for the
Reading Company lines was
r = 45.57 + 0.00340 u - 0.647 L (21 )
where:
r = two-way train frequency, trains per day
u = total passengers on the route, passengers per day
L = route length, miles
r = 41.8 trains per day u = 4,452 pass. per day L = 29.2 miles
R2 = 86.16%
SE = 9.825 trains per day
The resulting relationship for the Penn Central routes was
r = 38.24 + 0.00349u - 1.03 L (22)
r = 52.8 trains per day u 8,186 pass. per day [= 21.2 miles
R2 = 99.38%
SE = 3.24 trains per day
 352

These relationships seem to estimate the frequency of trains very well

on the basis of passenger traffic and line length. The coefficients are of
the expected sign. Also, it can be seen that the Reading Company seems
to provide a more frequent service, ceteris parabus, than the Penn Central
R.R. Although the two services are now operated for the same public author-
ity, it is likely that the two railroads had different policies regarding
service frequency when the lines were privately operated and these service
patterns have been continued under the public authority.
On only one of these routes is there any significant amount of express
train service provided. This is on the highest volume rail line, carrying
about 27,392 passengers per weekday. This line has four tracks, thus al-
lowing considerable express service. Also, the trains operating over the
longer distances on the Reading Company routes do not stop at many stations
close to Philadelphia, where other routes also share the same tracks and
serve those omitted stations.
Thus, it appears as though the same rather simple explanatory relation-
ships can be used for rail service. In order to explain other variations in
rail service, such as the stations at which trains stop, much greater variance
in traffic volume, in particular higher volumes with more examples of express
service, would be necessary than these data provide. However, they do defi-
nitely suggest that the same type of underlying rationale can be used for
explaining the frequency of rail service and bus service.
 353

Implications for Network Equilibrium Models

The preceeding analysis has revealed that the supply of public transit
service on fixed routes is such that the user's travel time is generally a
decreasing function of passenger volume. Theory suggests that waiting time
will decrease once the minimum frequency of vehicle trips are filled, while
the empirical evidence suggests waiting time is a decreasing function of
volume over the entire range. The only exception is due to the introduction
of zone express and local service, wherein longer distance travelers will
experience further reductions in travel time due to higher vehicle speeds
while shorter distance travelers may experience some increase in travel time
due to reduced frequency. Nevertheless the overall trend is definitely to-
ward reduced travel time with increasing volume. This would be further re-
inforced if fixed plant changes were considered, such as special lanes for
busses, priority signals for busses, busways, and conversion to rail rapid
transit - all changes which are more likely as traffic increases.
Virtually all existing network equilibrium models are solved using
algorithms or heuristic rules which are derived using the assumption of in-
creasing user costs as a function of volume. This assumption is of course
quite correct in the case of automobiles or roads, but is inappropriate in
the case of public transit. For the most part, transit time will decrease
with volume, given that management has sufficient time to adjust service.
Since most carriers revise schedules at least twice a year, the time lag is
quite short in comparison to the duration of most planning horizons.
This is not to imply that transit time can not be increasing with volume.
Obviously in the very short run - a period in which management can not respond
to traffic changes - travel time will increase if the traffic increase is
 354

sufficient to lengthen dwell times for loading and unloading and perhaps
affect running speeds. In a system with little slack in the schedules at
turn around points, such lengthening of the running time of one trip will
affect later trips, resulting in longer headways, more passengers boarding
at each stop, resulting in further delays, etc. Of course, management will
attempt to minimize such deviation of service from that planned.
Of perhaps more significance for bus services is the congestion on the
streets used. The running times of transit vehicles will be affected by the
prevailing traffic flow conditions, and hence these must be considered in
estimating the running times. From the standpoint of transit supply, other
traffic can be considered an independent input. Transit vehicle flow is usu-
ally so small compared to total vehicle flow that the volume increases pro-
bably will not significantly affect overall flow and hence bus speeds. How-
ever, if the bus volume becomes very large, running time will be affected.
It may increase", as would perhaps be expected on intuitive grounds. On the
other hand, the heavy volume of buses may in effect prempt a lane, and perhaps
move more rapidly than if the bus volume were less. The question of bus runn-
ing time and the factors which influence it requires much research.
Although consideration of th€se factors will further complicate supply
functions in public transit, it is clear that service will for the most part
improve with increasing volumes. This is provided that appropriate schedule
adjustments can be made, as seems reasonable in the time frames for most
transportation planning efforts. In this context, the current procedures and
underlying theory for finding network equilibrium will have to be examined for
applicability to modes in which service can improve with volume. Although it
is impossible to prejudge which methods will be applicable to public transit
 355

and multi-modal networks,it is possible to conjecture some of the problems

that may be encountered. One is the likelihood of multiple equilibria,
which raises interesting mathematical as well as social policy questions
about the selection of a "best" equilibrium and how to achieve it. It is
likely that realistic modelling of the demand for transport will require
consideration of a vector of level of service measures, including time,
price, comfort and perhaps others. This will require the seeking of an
equilibrium in many dimensions, a problem undoubtedly much more complex
than that for a single dimension, such as the travel time in most models
developed for the road context. Thus there is a wealth of interesting
and challenging problems posed for network equilibrium modelling by the
treatment of public transportation.
 356

Conclusions

The supply of public transportation or mass transit service in an

urban area seems to be amenable to quantitative treatment in a manner
broadly compatible with current urban transportation network models.
However, the supply functions are quite different from those used for
automobile transport. Transit management determines the supply of
transit service, subject to factors beyond its control such as road
traffic conditions and regul atory constraints.
This research has just touched the surface of potential research
into this aspect of urban transportation modelling. 0 Much more needs
to be done in the testing and refinement of models for bus services in
which the route structure is given, the major focus of the work reported.
Data for other cities and operating agencies need to be explored, as do
questions of estimating bus journey times in mixed traffic. Also, ques-
tions of route structure and possibly choice of technology should be
addressed.

Acknowledgements

The author wishes to acknowledge the assistance of the following

persons who provided data and assisted in other ways: Joseph Adams of
the Chicago Area Transportation Study, Thomas Luglio of the Southeastern
Pennsylvania Transportation Authority, and Oliver Sch~eftan of the Uni-
versity of Pennsylvania.
 357

References
Chicago Transit Authority (1972), Bus System Operating Facts, R.P.-y72144,
4-19-72.

W. H. Kraft and T. F. Bergen (1974), Evaluation of Passenger Service Times

for Street Transit Systems, Transportation Research Record No. 505, pp.13-20.

E.K. Morlok et.al. (1971), Transit Cost Models, Appendix C in The Effect
of Reduced Fares for the Elderly on Transit System Routes, Northwestern
University Transportation Center Research Report.

Penn Central Transportation Co. (1972), various timetables of Philadelphia

area suburban services, April 29, 1972.

R. H. Pratt (1970), A Utilitarian Theory of Travel Mode Choice, Highway

Research Record No. 322, pp. 40-53.

Reading Lines (1972), various timetables of Philadelphia area suburban

services, July 29, 1972.

Southeastern Pennsylvania Transportation Authority (1972) Philadelphia

and Suburban Railroad Passenger System: May 1972 Railroad Statistics,
mimeographed.
 358

Table 1

Comparison of Average Perceived Travel Times on Example

All-Stop Route* Using CTA and Theoretical Frequency Functions

Average Travel Time, Minutes

Passengers CTA Functions Theoretical

~er Two Hours A.M. P.M. Function**
250 44.9 44.4 45.0
500 41.6 41.3 45.0
750 39.4 39.4 40.8
1000 38.0 38.1 38.1
2000 34.9 35.0 34.1
3000 33.6 33.6 32.7

*See Figure 3.
**Using equation (6) and Q = 45 passengers per bus.
 359

Variables Primary Factors Influencing

Exogenous Endogenous Endogenous Variable Choices
Route Structure Schedule Regulatory Constraints
Technology Capacity Operating Subsidies
Stop Location

Route Structure Technology All above, plus

Stop Location Regional Plans
Schedule Capital Grants
Capacity

Route Structure All above

Technology
Stop Location
Schedule
Capacity

Figure 1. Possible Forms of Public Transportation Supply Functions.

 360

• •,'---...---""',•
CBD

( < :>
Bus Operations
- All-Stops
- No Cut-back Trips

2a. All-Stop Local Service

• • • ,•
CBD
Bus Operations

( -...
Inner Zone Local
>
-All-Stops

< :;;.
\ "\ ", I

All Stops No Stops A11 Stops

Outer Zone Express

2b. Zone Type Service

Figure 2. Types of Bus Service Operation.

 361

70
Time Between A and D
VI
QJ 60
+'
::s
c:
:f
+' 50
c:
,....
QJ

rtl Time Between Band D

.....>
::s
40
.
0-
W

QJ
.....E
I- 30
Qj
>
rtl
s..
I-
,.... 20 Time Between C and D
rtl
+'
0
I-
QJ 10
Ol
rtl
s.. Waiting Time
QJ
>
c:x:
. 0
1000 2000
........
+'
0 500 1500

P, Flow, Pass./Hr.

A B C D E
• • • • •
Points Along Route CBD

Figure 3. Travel Time on All-Stop Bus Route,

 362

.70

.60 Total Cost, Owner-Operator

Plus Pass. Time
.50

.40

.30

.20
...,
en
~
\ Owner-Operator Cost
o
U .10

o
o 500 1000 1500 2000
P, Flow, Pass. Per Hour

Figure 4. Cost On All-Stop Bus Route For Peak Period Operation.

 363

70

til
....,C1J 60 Time Between F and 0
'"c:
.~

All-Stop
:E:
....,
c:
~
C1J 50
ttl
>
'"0-
w

C1J
E
.~
~
40
F Zone Express
I-
~
30
C1J
> Zone Local
ttl
~
I-
~ 20
....,
ttl
Time Between G and 0
0 All-Stop
I-
C1J
C>
ttl
~
10
C1J
>
c:(

...., ~

0
500 1000 1500 2000 2500
0

P, Corridor Flow, Pass./Hr.

G c o E
.
A F B
'"_ _ _-----1

CSD

Zone } { Express Servi ce: Stops A-B, Non-Stop B-C, Stops C-E
Service: Local Service: Stops B-E
All-Stop Service: Stops A-E

Figure 5. Comparison of Travel Times of All-Stop and Zone-Type Services.

 364

6000

5000

>,
c'"

-
s... All-Stop Service
OJ
4000 0-

.
.....Vl
0
U
~

'"
.....
0
I-
3000

Zone Local and

2000 Express Service

1000

o , .......
o 500 1000 1500 2000 2500

P, Peak Flow, Pass. Per Hour

Figure 6. Comparison of All-Stop and Zone Service Costs.

 365

.70

.60

Cost, Owner-Operator Plus Travel Time

<II
..- .50
i:I
All-Stop Service
III
III
<0
0.. .40
"-
<II
0..

..... 30
+'
III
Zone Service
o
U

~ .20
<0 ( Owner-Operator Cost
"-
<II
>
ex:
All-Stop Servi ce
.10
~~~-~-~--.~~,==~~.~~~===========
- - -I •
- Zone Servi ce
o
o 500 1000 1500 2000 2500

P, Peak Flow, Pass. Per Hour

Figure 7. Comparison of User and Operator Costs of All-Stop and Zone Services.
 366

o o o o o
¢ M N

Bus Trips Operated in Each Two-Hour Peak Period

Figure 8. Chicago Peak Period Bus Frequency and Traffic Data and Regression Lines.
 367

co
N

®
~

'">-
+'
C
OJ O'l
"""
N

U C
c -C
C
OJ
0-
'"
OJ
0:: >,
'"
0
>- 0
•
OJ
0 0-
N

'"'"'"
0-

0
0
C!
0
<0

• '"OJ>-
en
c
®
OJ
III

'"'" N

0 0-
>,

@ '"
Cl
~

'"
+'
0 co
~

•
•• •
•

0 • •
•• 0

0 0 0 0 0 0
LO M N
"""
Daily Train Frequency, Trains Per Day

Figure 9. Philadelphia Suburban Railroad Service Supply.

 A NOTE ON THE ECONOMIC INTERPRETATION OF
DELAY FUNCTIONS IN ASSIGNMENT PROBLEMS

~1arc Gaudry
Universite de Montreal
Montreal, Quebec, Canada

ABSTRACT

The relationship between the time delay curves of traffic assignment problems
and supply curves is examined in this paper. A claim is made that most delay curves
are simplified cost and/or production functions. The resulting distinction between
delay and supply curves is illustrated by differentiating between production and
market equilibria in networks, notably in cases in which the former equilibria are
observed while the latter are not.
 369

Introduction*

Time delay curves are often identified with supply curves

in assignment problems. Used in conjunction with demand curves,

they define problems which have the appearance of market problems

centered on the existence and stability of equilibrium market clearing

solutions. This note purports to explore the relationship between

delay and supply curves.

A first section derives delay curves from production and

cost functions for the usual travel output along a single street.

The second section describes a case in which delay curves might be

interpreted as supply curves and lists more realistic cases in which

delay and supply curves have 80 little in common that production

equilibria occur on delsy curves without supply curves being even

observable--due to the presence of stable market disequilibria. The

conclusion briefly extends these perspectives to the interpretation

of complex network equilibria.

The derivation of delay curves from production and coat functions

A. Travel production and cost functions

Let us assume that a single street goes from i to j and

that vehicle trips per hour between those points may be produced with

* Theauthor wishes to thank Michael Florian and Sang Nguyen of Universite

de Montreal and Harold F. Kuhn of Princeton University for helpful dis-
cussions and encouragement.
 370

v.hicl. trips
from itoj

Iv
I
I
II
I
IIE
l'.,,"'ci• .. t"
(a) '1tJ.

cost
6 total

I marginal

/
/
~/
// av.rag.
;~
,;

o v -"z
v.hicl. trips
(b) from i to j

figure 1
 371

inputs of vehicle-distance per hour and of vehicle-time per hour.

The resulting relationship between travel inputs and outputs is

represented in figure la where OV gives the value of output asso-

ciated with all values of inputs. OV is a curve and not a sur-

face because, at a given level of output, vehicle-miles cannot be

dissociated from vehicle-minutes (there is no factor substitution).

Up to the point v, the curve OV is a straight line and exhibits

constant returns to scale because vehicles are assumed to travel at

the legal speed if they can. Beyond the point v , the production

function exhibits rapidly decreasing returns to scale, and factor

proportions change.

If the prices of the vehicle-distance input (gas + main-

tenance + depreciation per mile + tolls) and of the vehicle-time

input (driver's wage rate + depreciation per time period) are known

constants, one may derive the total, average and marginal costs

associated with each level of outputl these curves are given in

figure lb. The total cost curve gives the minimum total cost for each

level of output at given input prices. The "firm's" expansion path

is of course simply the curve OE, the projection of OV on the XY

input place; it is independent from variations in relative input prices.

In a complex network, by contrast, the expansion path changes

with relative input prices because factor substitution is possible. If

there are four itineraries from i to j , as shown in figure 2a, input

 372

z y

vehicle trips
from itoj

v ec

.....
......... d/
.......... X
........'<*" ........
.............

y
(a)

a ec'

pz
/ (b)
ed'

figure 2
 373

requirements for v trips per time period vary with relative input

prices. As can be seen in figure 2b where the four points are pro-

jected on the XY plane, input combination a' minimizes total time

but is is the cheapest only if relative input prices are given by a

price line such as P2 rather than by a price line such as PI

(which embodies a relatively lower price of time than P2)' The

cost functions of this network would be step functions because higher

trip levels can be produced at equal or greater unit cost.

The computation of the cheapest input requirements for each

output level 1s what we may call the network e~uilibriumyroduction

problem. We will not discuss differences between system and user [see 1)

optimizing solutions or even ask whether these procedures mimic actual

behavior. We will be satisfied with the fact that assignment procedures

succeed in computing input requirements which respect the constraints

of the production technology and we shall limit our considerations to

the single street case for expository purposes.

B. Dela~ curves as s~ecial ~roduction/cost curves

There are two ways of obtaining delay functions, both of

which involve making the distance input irrelevant.

The first approach transforms the production function into

a one~rgument function by neglecting the distance input. This can

be effected by projecting the curve OV in figure la on the XZ plane

 374

and by adding the average and marginal curves, as is done in

figure 3. The delay curve is the average time curve.

Another approach maintains the production function but

derives cost functions by assuming that the distance input is

free and that the unit cost of time is one. The resulting time

cost curve is identical to that shown in figure 3 except that the

vertical axis is labelled "time cost".

Whether delay curves result from simplified production

functionsor from special cost curves, distance does not matter in

the construction of delay curves: it is either not an input, or

it is free.

Delay curves and supply curves

If we accept the exceedingly strong assumption that dis-

tance is irrelevant, the production equilibrium will certainly be on

the delay curve which associates to each output level its minimum

average time input requirement. But are these delay curves supply

curves?

We need information on market structure to answer that ques-

tion; let us then examine three classes of markets in which it might

be asked. If we need to use figures, we shall adopt the common usage

and label the vertical axis in time units; a parallel presentation

using the cost label would leave the argument unchanged.

 375

x
tim.

marginay
I
.#~
~r averag•
.,-"'--
--

o ,
v trips - Z

figure 3

_0
tim.

"
, Imarginal tim. cost
! jav.rag• t.cost

" ..,~/II
/ ,,0
t" ____________

--_. :-I . ~arginal tim. r.venu•

o Y:. v.
I ."
trips
...
- -Z

figure 4
 376

A. Taxi markets

If there are nothing but taxis serving the street in

question and if the taxi industry is perfectly competitive, the

delay curve is a supply curve: the situation is analogous to the

classical case in which each producer is a price taker who has no

effective influence on price so that the long-run supply curve for

the industry coincides with the average "cost" curve.

As soon as the firm is able to set its own "price", the

concept of the supply curve is not strictly applicable any more [2].

A profit maximizing taxi monopoly will choose a particular trip

volume such that its marginal (time) revenues and (time) costs are

equal. If there is congestion at that point, the chosen point, such

as point b in figure 4, is not even on the average delay curve.

The delay curve is not the supply curve but simply describes tech-

nically feasible choices and/or their costs.

B. Private car -trip markets

The private car trip market is characterized by the fact

that the suppliers are often consumers as well. It is a market in

which each driver is a "price" taker; but i t is also a market in

which consumers do not normally pay the producer. A fair representa-

tion of the process might be the following: assume that in each

household some members are drivers but all members demand car trips.

The suppliers' reaction would be such that they would drive consumers

(including themselves) to their destinations more often if the average

time required for the trip is small than if it is large: this means

that the supply curve -- like the demand curve -- is downward sloping.
 377

In that market, trips can occur if suppliers are willing

to drive and if there are enough consumers who want to travel in

the network. Trips will occur if that (time) price is sustainable

in the network. Figures Sa and 5b illustrate two likely equilibria.

The striped curves give possible outcomes and the delay curves give

technically feasible ones; points m and n cannot be maintained

because they would imply that drivers purposefully drive slowly in

empty streets. The sustainable solutions are those in which the

vehicles which do show up are driven as fast as possible. In Sa

the resulting values of v and t are on the demand curve; in 5b

they are on the supply curve. Both situations are production equilibria

but none are market equilibria. Markets would clear if by chance the

market equilibrium were identical with the production equilibrium.

This does not ease the task of estimating demand and supply

curves but it makes it clear that, even if distance is assumed not to

matter, delay curves have little to do with supply curves.

C. Public transit markets

In an urban context, the representative public transit author-

ity will face a trade-off between having buses travel as fast as cars

and having them stop for passengers. The chosen travel speed will

normally be smaller than the car speed. The optimal speed is chosen

among all the points situated above the delay function boundary, for

all of these are feasible and perhaps desirable points. A good demand

model, a formulation of the authority's objectives and constraints are

needed to find appropriate supply points in this considerable domain.

 378

x x
time

t* - - - - - - - - - - - - - -
s I
I
o , o - : I -,..z
v V V* trips
(a) (b)
figure 5

//1J
x x

.J :1'///j
//'t,!.// /~/ /
time

~ // / ~
//.t/~.# .0

0
v
-z 0 I
y
.... Z
trips trips
(a) (b)

figure 6
 379

In figure 6a the demand curve has not been drawn because it will

shift every time the distance between bus stops is changed in order

to increase travel speed. In figure 6b a demand curve has been

drawn on the assumption that the number of stops along the street

is fixed: in consequence the transit authority's area of choice

for a supply point has been reduced.

This difficult problem has received little attention in

the literature. One can however say that, in this transit context,

a time delay curve is nothing more than a boundary: as soon as speed

becomes an object of choice, the already tenuous link between delay

and supply curves vanishes altogether.

Conclusion

The above perspective distinguishes clearly between produc-

tion and market equilibria. The examples have been chosen from a single

street case for the sake of clarity. In more complicated problems the

relevant distinctions are not so easy to make.

If there exists a network between a single given origin-

destination pair, the analysis carries over: as in figure 2, one has

production processes rather than production functions. Processes

combine time and distance inputs (from different streets) in varying

proportions and the chosen process depends not only on relative input

prices but on control technology: our monopolistic taxi corporation

could not assign its cars in a system optimizing fashion without cen-
 380

tralized control of individual itineraries. Once the average cost

curve is found, the distinction between production and market

equilibrium easily reasserts itself.

But if one considers a city in which many origin-destination

pairs have access to a common network of streets, the situation is

complicated by the fact that there are as many distinct commodities

as there are distinct origin-destination pairs. One is in effect

dealing with a great number of interrelated production and market

equilibria without being able to associate a unique cost function to

each commodity. This indeterminacy arises because, in a congested

network, input requirements for the production of a given number of

trips between a particular pair of points depend on the output levels

already occurring among other pairs and influence input requirements

for those output levels. In consequence, unless the pairs are arbi-

trarily ordered, there is no unique way of imputing the technological

externalities -- and resulting increased production costs -- to specific

processes.

The difficulty of specifying interrelated production processes

should not deter us from concluding that delay curves are never in practice

good proxies for supply curves and that estimates of delay curves yield

a particular projection of the production function. Let us add that obser-

vations on which such estimates are based probably do not represent market

equilibria and that the task of finding the extent to which they do will

require the formulation and estimation of market supply curves -- a tho-

roughly underdeveloped area of transportation research.

 381

References

[1] Potts, Renfrey B. and Oliver, Robert M., Flpws in Transportation

Networks, Academic Press, New York, 1972, p. SO.

[2] Baumol, William J., Efonomic Theory and Operations Analysis,

Prentice-Hall, Second Edition, 1965, ch. 14.
 RECENT EXPERIENCE WITH EQUILIBRIUM METHODS
FOR THE STUDY OF A CONGESTED URBAN AREA

Michael Florian
and
Sang Nguyen

Universite de Montreal
Montreal, Quebec, Canada

INTRODUCTION

We report in this paper on the application of the algorithm of Nguyen [1],

that computes network equilibrium flows, to assign peak hour traffic on the road net-
work (1971) of the City of Winnipeg, Manitoba.

The Transportation Division of the City of vJinnipeg carried out an origin-

destination survey in 1971 and in addition, compiled a wealth of data on the use of
the road network. The road network of the city of 1035 nodes and 2788 arcs was coded;
for most of the links that were coded, observed volumes and observed travel times
were obtained; a family of volume delay curves for the links of the network, were
estimated by Traffic Research Corporation in a study carried out earlier, an origin-
destination matrix was estimated from the survey data for the A.M. peak.

Our task in this undertaking was to apply an "equilibrium type" assignment

technique and to provide the City of Winnipeg with the results of the output, if suc-
cessful. We were extremely fortunate to obtain invaluable help in establishing contact
with the City of Winnipeg from the staff at the Transportation Development Agency.

Thus, we can report on a large scale application of an equilibrium type

assignment algorithm that is guaranteed to produce flows that satisfy Wardrop's [2]
first principle. Also, due to the wealth of data available on the observed volumes,
we had the unique opportunity to attempt a validation of the behavioural hypothesis
imbedded in the method, and test, for instance, the differences between, an "all-or-
nothing" assignment and an equilibrium solution.
 P ra ctical (opacity Range
Speed Vehicles 'er HOur Per Lon.
Lim i t
0-249 1000

16~ r-.. 0-30

6 9
•
10 14 IS
14t
W // (3) '-" SO+ 16 17
--.J '.
2 12
0:::
w
Cl...
C/) 10
w
f-
:::::>
z
8
2
I
w
~ c..u
f- 6, oe
--.J
/ // / /// / /// / //' c..u
w
~ 4~
0::: I
f-
2
:....---::: ===
0 200 400 600 800 1000 1200 1400
VEHICLES PER HOUR PER LANE

RECOMMENDED TRAVEL TIME AND VOLUME RELATIONSHIPS

FOR THE WINNIPEG AREA TRANSPORTATION STUDY

FIGURE 1
 384

1. THE VOLUME-DELAY CURVES

The family of volume-delay curves estimated by Traffic Research

Corporation for the City of Winnipeg were of the functional form

v 2 va
S (v ) = IS + 0.(2 -y) +
a a n
a. (n - y) 2 + B
a a

where S (v ) is the travel time in minutes/mile, v is the total vo-

a a a
lume of the link, n is the number of lanes of the link and ex, B, y,
a
IS, are constants that vary for each of the curves. Figure 1 gives a gra-

phica1 representation of the curves.

The delays at stop signs and signalized intersections are not given

explicitly; rather, they are represented implicitly by the volume delay

curves. Thus, the time calculated from those curves is an indication on

the time taken on the average by a vehicle on the link. The time taken

to cross a link ta(v a ) is simply computed by the formula

where da is the length of the link in miles.

2. NETWORK CODING

The network was coded in a conventional manner, with the following

exception. Intersections for which the turning movements were detailed,

were coded as follows:

 385

FIGURE 2

All arcs that originate and terminate at a node in the set {I, 2,

3, 4} in an example are fictitious or dummy links. Thus, when computing

shortest routes, it is necessary to prevent a route that contains two

successive dummy links, such as (1,2) (2,3) (3,4), for instance~ Al-

though the algorithm may be applied with this modification we were not

certain whether some pathological cases may arise\that bring about in-

consistencies in the computation. Then we also converted the inter-

sections to the more usual representation:

* A similar argument must be applied when tracing a potentially

utilised route.
 386

Jt
-

FIGURE 3

However, the computational results that we report on are based only on

the representation of figure 2.

3. THE ORIGIN-DESTINATION MATRIX

The 1971 Auto Driver A.M. Peak origin-destination matrix included

also the students trips to University of Manitoba, University of Winnipeg

and the Red River Community College. As we shall point o~t later, there

are some special characteristics to these trips, as they contain a sig-

nificant amount of car-pooling. The O-D matrix excluded buses which will

be added on a link after the assignment. While this was the method chosen
 387

by the City of Winnipeg, it may be advantageous to add these trips to

the link volumes prior to the assignment. While we plan to do so, the

results on which we report here, do not include the buses. The O-D ma-

trix also contains a conversion factor that augments the flows by a ratio

of Total Vehicle Trips/Work Vehicle Trip obtained from a strictly selected

screen line.

4. THE COMPUTATIONAL RESULTS

We applied first the assignment algorithm to obtain the equilibrium

flows and plotted the computed volumes against the observed volumes for

all the links which has this information. Although the results were sa-

tisfactory, several unreasonable differences were observed.

We reasoned that some of the differences may result from differences

in network coding, the number of lanes, the volume delay curves or in

transcribing some of the ground counts. There is of course the possibi-

ltty that the behavioural hypothesis imbedded in the method gives erro-

neous results, however, we thought it best to check the data first.

We edited the data in various ways and returned to discuss these

with the transportation planners of the City of Winnipeg. A thorough

edit of the data was done and several modifications were made to the
specified number of lanes, the volume delay curves and several observed

volumes were found to contain transcription errors.

 388

Our confidence in the model increased appreciably, as the flows

obtained, were instrumental in refining the data used in the simulation.

The assignment algorithm was reapplied to recompute the equilibrium flows.

The results are given in figures 4a and 4b at 5% equilibrium. The "out-

lines" at the upper left hand side of the graph are still to be thoroughly

explained. They may still represent some differences in data coding and

network coding. For some of the other "high" points on the curve, the

transportation planners of the City of Winnipeg suggest the following

explanation: Many of the trips to the University are not taken along

the "current shortest" routes due to car-pooling. Thus, these trips were

intermediate destinations and the O-D matrix does not reflect these inter-

mediate destinations. Also an expressway that passes close to the Uni-

versity does not attract as many trips as the model predicts. This is

probably due to the fact that the volume delay curves only model the tra-

vel time and in certain extreme situation, travellers may also be sensitive

to the travel distance. Although we do not have conclusive evidence to

this effect, we plan to pursue this matter in our subsequent work.

For comparison we also plotted the "all-or-nothing" solution in

figures Sa and 5b and the 15% equilibrium solution in figures 6a and 6b

while there seems to be little difference between the 5% and the 15% so-

lution, the "all-or-nothing" solution is clearly unsatisfactory.

 389

+
+
+
+ +
+ +

+
o
o
+
o + + + + +
K3 .++ + + + +-tt-
+1-+
+ + +
+
N g
o + +
o...... 0

x
N
+

o
*
+
W
I-
~ 0 + +
a....
:r:
ou
0
0
0 + ++ +*
~
~
+I- +
+ +
o +
o
o
LJ)
.
+
o
o
o
o ~--~~~~------.-------.--------.--------,-------,

0 . 000 5.000 \0 . 000 \5.000 20 .000 25 .000 30 . 000

OBSERVED VOLUME(X10 2)

WINNIPEG 1971
FIGURE 4a The 5\ equilibrium solution
 390

0
0
0
0
(T) + + + ...
+ + + + ++ +
0
a + ++ + +
L? + +
+ +
II)
C\J + + +
+ + + +
+ + + + + :If. +
+ -t+
+ + +t + +++ + +
0
0 + + +
L? + +
a..--< a +
C\J + + + ++ + +
X + q-* + +
W + ++-t q- ++ +
L:
:::J 0
+ + -4-+
+
+ + +
+ +
0
-.oJ L? + + + + ++ ++ +
0
>
II)
*+
++ it-
++ + it-+ +
0 + ++ + + + + it- +
W
+ + +++
I- + + + ++
+ +
:::J
0.....
0
0
L? + +
+ +1;- + + +,.+ + ++
L:
a 0
+ + ++ + + + +
U ~++ + ++ + + ++ ....+ + +
t \t+ -++- +.j!j.. +
+
+ + + t+ +++t+ + + + +
.
0
+
t
0
a +++
(f) ++ + + +
+ ++ + + + +
/ + + ++
/ + ++=1= ++ + + +
++ + + T ++
+
0
0 -tt + + + +
0
0
+ +
0.000 6.000 10.000 16.000 20.000 26.000 30.000
OBSERVED VOLUME(X1O 1)

WINNIPEG 1971
FIGURE 4b The 5\ !9uilibrium solution
 391

+
+
++ ++ +
+ + +

++ + + +
0 + ++ +
""C? + + +
0
(T) + +
+ ++
++
-It- + +
0 +
""C? + + +++
Lf)
:+- ++ +
N
+ ++ + + -t+ +
+ + + ~-t
N 0
""
0 + +
0 0 +
--' N +
X
W
L 0
:::)
--.J ""
C?
a Lf)
+
>
0
W
I-
:::) 0
Il...
L
""
0
0 g
u

0
""C?
Lf)

oo
o
o -1'------"

0 .000 6.000 10 . 000 16.000 20 .000 26 .000 30 . 000

OBSERVED VOLUME(X 10 2)

WINNIPEG 1971
FIGURE Sa The "all-or-nothing" solution
 392

0
0
0
0 +
(T')
+
+ -t +
0
0
+ +
S' t +
lJ)
N + + ++
+ + .+
+ + t +
.....
~
0
0
++ +
S' + ++
+ #++
*+t +
0 0
....... N
+ ofT
X
+ ++ +
W +
1:: 0
:::)
---l
0
C?
+ -1-!- +
- +
0 II)
>
+ +
0
W ++
~
:::)
CL
0
0 +
1:: C?
0 0
-'
U

.
0
0 +
0
+
II)
+ ~+
+
-1+-
oo
o
o ~~Hfi--~-*~~~~~+H~~~~~~~~~~~~--h

0.000 5.000 10.000 15.000 20.000 25.000 30.000

OBSERVED VOLUME(X10 1)

WINNIPEG 1971
FIGURE Sb The "all-or-nothing" solution
 393

+
+
+
+
0 + +
0 +
C?
0
(T)

+
+ -t
0 + +++ + + +
0
0 + +
+
+ t +
If)
C\I

t+
++
+
"* +
+ + + +
+
~
0
C\I 0
C?
+ + + #" +
0
.- 0
C\I + +1/ +
X + .++
w
+
+
L
:::> 0
0
++
.-J
a
0 ~
+ ++
>
If)

+ t +
0
w -#- t* +
+
I-
:::> +
-t +
0
a.... 0
0
L
a 0 .....+"\,+ +
u '+- + +
it ++
++ ++
0
0 +*r-
0 + +
+ +..ft.
If)

+
0
0
0
0

0 .000 5 .000 10.000 15.000 :20 . 000 :25 .000 30 .000

OBSERVED VOLUME( XIO 2)

~JI NN I PEG 1971

FIGURE 6a The lS\ equilibrium solution
 394

o
o
~
o +
(J')
+ +
+ ++
+ + + +
o
o + + + + + + +
~ + + + +
II) + + +
C\J
+ + +
t +
+ + ++
oc + +
+ + ++ +
o
~
0 + +
....... C\J +
X
+ + +
W + +
~
::::>
0
c
+ -+ir + +1<+ t + +
---l
o>
0
II)
~
+ ++ ++ + ++
+
o + + +
w +
~
::::> 0
a.... <=>
~ 0 -tr + + +
o 0
+ + + +
u
+
1 + t + -if- + ++
+ + ++*~ ++ +
oo
t + ++ + + + +
o +++ 4+ +-t++
+ +++ ... + + +
l/)
+ + + + +
+ + + + ++
+ +
+++ + + + + + +
o +=!;t.. + + . + +
o
~ +:f +4 + +
o ~~~r-~~~~r#~~rr-r+p~--~+o~-ti~H++--+~~h

0.000 6.000 10.000 to.OOO 20.000 26.000 30.000

OBSERVED VOLUME(X10 1)

WINNIPEG 1971
FIGURE 6b The 15\ equilibrium solution
 395

REFERENCES

[l[ Nguyen, S., "An Algorithm for the Traffic Assignment Problem",

Transportation Science, Vol. 8, No.3, pp. 209-216, 1974.

[21 Wardrop, J. G. , "Some Theoretical Aspects of Road Traffic Research",

Proceedings Int. Civil Engineers, Part II, pp. 325-378,

1952.

[ 31 Von Cube, H.G. et aI, Winnipeg Area Transportation Study, Analysis

and Synthesis of Travel Patterns, 48 pp., 1966.

 396

THE CHOICE OF ASSIGrlMENT TECHNIQUES FOR LARGE NETWORKS

DIRCK VAN VLIET *

The Institute for Transport Studies
The University of Leeds
Leeds, England

As part of the Greater London Transportation Survey

of 1971/72 a road network of approximately 1,000 zones,
5,000 nodes and 15,000 one-way links has been built and
a study undertaken to develop assignment procedures suit-
able for a network of this size. This paper describes
the major conclusions of this study plus a summary of a
series of tests undertaken on a previous London network
of 1,800 nodes and 5,600 links. It was found that a
method incorporating four incremental loads plus one re-
assignment of selected trips previously assigned to par-
ticularly poor routes achieved a satisfactory degree of
convergence.

It also became increasingly apparent during the course

of this work that achieving a perfect or near perfect equili-
brium solution is not the sole criterion for a "good" assign-
ment method. A number of more practical considerations must
also be taken into account; e.g., the ease with which secon-
dary analyses may be undertaken and the size of the inherent
errors. There seems little point in developing sophisticated
assignment techniques which are orders of magnitu,::e more
accurate than the data on which they operate.

* Formerly with The Department of Planning and Transportation,

The Greater London Council, County Hall, London, England.
 397

INTRODUCTION

In canjunctian with the Greater Landan Transpartatian Survey (GLTS) carried aut
in 1971/72 by the Greater Landan Cauncil an imprO'ved transpartatian madel has been
develaped and is currently in the calibratian stage using the survey data. As part
af the develapment pragramme a study was undertaken af the methads far the assign-
ment af trips by raad. This paper discusses the criteria used to' chaase a methad,
the canclusians reached and the results af same tests an the methads selected uSing
previaus GLC netwarks.

Let us first say samething af the size and nature af the prablem. First, the
raad netwark being built far the GLTS cansists af appraximately 5,000 nades and
15,000 ane-way links; in additian there are 933 internal zanes and abaut 100 "exter-
nal zanes" at cardan paints abaut the baundary af the netwark. It is abviaus that a
netwark af this size presents a number af special prablems, in particular the campu-
tatianal requirements af time and space far even simple netwark manipulatians. Far
example, building a camplete set af minimum cast "trees" an the full netwark requires
af the arder af 30 minutes CPU time an an IBM 370/145. This has meant that a number
af assignment algarithms which have been praposed in the past have had to be exclu-
ded on practical grounds. Second, there are relatively very few high capacity, high
speed motorway links in the network and most links (during peak periods at least)
are aperating at or near capacity. The problem af incorporating capacity restraint
is therefore an important factor.

Sections II to V of the paper deal primarily with the criteria necessary to

evaluate different assignment pracedures. Thus Section II lists some af the objec-
tives of a road assignment, with particular reference to those of the GLTS; these
are seen to be somewhat wider than simply predicting link-by-link traffic flows.
Section III discusses the general principles of raad assignment, in particular
Wardrop's Principle and the conditions under which it is valid, while Section IV
describes the usual formulation of the assignment problem and Wardrop's Principle in
network terms. The potential errors in both Wardrop's Principle and the network re-
presentation are summarised in Section V; and their effect on the relative impor-
 398

tance to be attached to different objectives, analysed.

Our major conclusion from this discussion is that achieving a perfect or near
perfect equilibrium solution, while a very important objective in itself, is not the
sole criterion for a "good" assignment. A number of other practical considerations
should also be taken into account.

The final two sections, VI and VII, introduce the method recommended for the
GLTS - incremental loading with the option of re-assigning some selected trips - and
the reasons for the choice, and give the results of a series of tests carried out
using a "District Test" network (the fore-runner of the GLTS model).

It is anticipated that the :esults of some similar tests on various stochastic

methods of assignment which have also been developed for the GLTS will be published
later.

II ASSIGNMENT OBJECTIVES

The most obvious objective of any road assignment including the GLTS is to
correctly predict the volume of traffic on each network link (or, on a more macro-
*
scopic scale, across a corridor). If, however, this were all that we require it
would very often be easier (and very likely more accurate) to measure present-day
traffic flows directly and forecast changes by simple rate-of-growth assumptions.
The real advantage of a synthesised assignment is the ability to analyse the traffic
pattern that is built up. This leads one to identify a number of secondary, but
often extremely important, objectives; e.g.:
(1) Determining the route or routes used by each O-D pair as well as
the alternatives available.
(2) Analysing the composition of traffic on a given link or links,
e.g., in terms of its origin/destination, the split between
different trip purposes, etc.
* Another primary objective might be determining the minimum or average travel cost
between each origin-destination pair as required by the distribution stage.
 399

,
(3) Cordoning off external and through traffic for a sub-network to
be studied in greater detail.
(4) Forecasting and analysing turning movements at intersections.
(5) Visual presentation of data.

Another important property of an assignment method is the ability to predict

equally well not only the total pattern of traffic in a network but also marginal
changes due to mi nor a lterati ons on the network ("perturbati on methods").

In a slightly different sense it is also an "objective" to satisfy any physical

constraints, for example the amount of core storage and time available on the compu-
ter, the amount of manual work required, etc. For networks the size of the GLTS
constraints of this nature place severe limitations on the choice of an assignment
method.

III THE PRINCIPLES OF ROAD ASSIGNMENT

The basic principle of traffic assignment under congested conditions was ori-
ginally formulated by wardrop(l) to be:

Traffic on a network distributes itself in such a way that the

travel costs on all routes used from any origin to any destina-
tion are equal while all unused routes have equal or greater
travel costs.

This principle is founded on two basic premises: (1), that each driver chooses the
route which offers him the minimum perceived cost; and (2), that all drivers per-
ceive costs in a unique fashion. While it is difficult to fault the first premise
since it does not really define "perceived cost" the second is the weak link in the
chain. It is well known that not all drivers perceive costs in the same way and that
a wide variety of different routes are chosen for a wide variety of reasons. In fact
we will now argue that the degree of spread in perceived costs is crucial in deter-
mining whether or not Wardrop's Principle should be applied in a given situation.

Let us imagine that we have arrived at some definition of perceived cost (e.g.,
 400

by equating it to travel time) and have solved for the equilibrium flows which,
given all our assumptions and inputs, satisfy Wardrop's Principle. If we were now
to survey, say, all trips from a particular origin to a particular destination we
could work out the costs on our synthesised network for the routes reported. By
definition the costs measured along any route R, cR, must be greater than or equal
to our calculated minimum cost, cmin . Let us define the fractional excess cost along
R R min mi n . R
a route R as E = (c - c )/c ,and the welghted mean value of all E for all O-D
combinations as El . El is therefore a measure of the degree of confidence with which
*
we measure network costs; differences less than El tend to lose their significance.

A somewhat less strict formulation of Wardrop's Principle might therefore be:

All routes used are within El of the minimum cost while routes
with costs above El are unused **

We now require a comparable measure of the effect of congestion on the disper-

sion of routes. Rather arbitrarily let us define a parameter E2 for each O-D pair
as the fractional difference between the minimum cost as measured under free flow
conditions and the minimum cost measured at the equilibrium consistent with the total
load. More generally we can define a network-wide measure E2 by:
E = (total travel cost at equilibrium) - (total free flow cost) (1)
2 (total travel cost at equilibrium)

We now postulate as a general rule:

If E2 ~ El congestion is the major factor determining the diver-

gence in route choice and Wardrop's Principle is applicable;
alternatively if E2 < El differences in drivers' perceived costs
are the dominant factor and other principles must be enunciated.

* Note that E, is not an entirely intrinsic characteristic of the drivers since it

also depends on our definition of costs; the better our definition is the smaller
El should be.
** This is not to say that ~ routes less than El are used any more than the ori-
ginal formulation of Wardrop's Principle says that all minimum cost routes must
be used.
 401

Our reasoning is as follows: Suppose the free flow minimum cost were 20% less
than the final congested cost and that the spread in perceived cost were only 10%.
It seems reasonable to assume that a sizeable fraction of drivers would regard the
free-flow route as a viable choice disregarding congestion and that some of them
would divert onto it were it to be significantly underloaded. If, however, the
values above were reversed the free flow route would, in general, have only a margi-
nal advantage over the alternatives.

As a general rule of thumb we might expect £1 to be of the order of 10%. Since,

as any London driver will readily confirm, differences in travel time between free
flow and congested situations in London are considerably in excess of this, in deve-
loping assignment procedures for the GLTS we have assumed that Wardrop's Principle
is indeed valid. The majority of the work described here is therefore concerned with
congested assignment; Section VI summarises our conclusions and Section VII gives
the results of some tests on a congested assignment using incremental loading.

However, recognizing that the variation in cost perception is an important fac-

tor, even in congested situations, we have also developed stochastic assignment
methods similar to those of Dial(2) to deal with this problem; a survey of our tests
and conclusions will be given in a later paper.

IV ROAD ASSIGNMENT AS A NETWORK PROBLEM

The road network may be represented by a set of links (arcs) {a} corresponding
to the roads and a set of nodes (vertices) corresponding to the intersections.
Usually the origin and destination zones are considered as a sub-set of the nodes
referred to as centroids and their "centroid connectors", as a sub-set of the links.
For each link a there is a unique travel cost, ca ' generally assumed to be a function
only of the total volume of traffic on that link, Va * :

(2)

and the total travel cost along an O-D route is the sum of the individual link costs
which make up that route.

* But see point (2) in Section V.

 402

The total demand for (vehicle) trips between origin i and destination j is
represented by Tij . For the purposes of the GLTS and the present paper we regard
the {T ij } as fixed.

A feasible solution of link flows {Va} may be defined as one in which for each
O-D pair ij:

i) A number of routes through the network R~j) are defined,

ii) The total number of trips along each route, T~~) ,

lJ

designated; L T~~) = T1•J· ,

n lJ

and the total flow on each link is given by:

~ (n)
L l. (n) Tij (3)
i ,j n, a£R

A feasible solution satisfies Wardrop's Principle if, having calculated the set
of link costs {cal corresponding to the link volumes {Va} via (2), all routes R~j)
for which T~j) > 0 are minimum cost routes.

A large number of heuristic methods for generating successive feasible solutions

and combining them together so as to approach convergence have been proposed in the
past, incremental loading which we discuss below being one. These methods are very
often referred to as "existence methods".

An important advance in assignment theory has been to recognize that the assign-
ment problem can also be formulated as a multi-commodity flow minimization problem(3).
Finding an equilibrium solution satisfying Wardrop's Principle is identical to find-
ing a feasible solution {Va} which minimizes an objective function:

F = L fa(V a ) (4)
a
V
where fa(V a ) = foa ca(d) dv (4a)

given that ca(V a ) is a monotonic increasing function.

Formulating the problem in this way has many advantages; e.g., it allows more
 403

powerful mathematical methods to be brought into play. It also allows one to quan-
tify the degree of convergence; thus Murchland(4) has shown that an upper limit on
the difference between" F and its minimum value is set by a parameter <5 defined by:

<5 = L Va ca(V a ) - L Tij c~~n

lJ
(5)
a i ,j

where cmin
ij is the minimum cost from i to j given link costs {ca(V a )}
Since the first term in (5) is the total travel cost actually incurred on the net-
work and the second is the total travel cost which would result if all trips could
take the current mini"mum cost path (at fixed costs) <5 is therefore a measure of the
total excess travel cost.

In this paper I have adopted a slightly different definition of <5 to make it

less problem dependent by expressing it as a fraction of the total minimum travel
cost:

(6)
\ T c~~n
i ,j
l. ij lJ

Similarly a value of <5 may be defined for the nth set of routes:

cmin]
ij
(6a)
.L.
1,J
T~~) min
lJ c ij

where c~~)
lJ
is the equilibrium travel cost from i to j by route n.

It is easily seen that when Wardrop's Principle is satisfied <5 = 0 and other-
wise <5 > O.

V ERRORS IN THE ASSIGNMENT MODELS

While Wardrop's Principle and the equivalent minimization of the objective func-
tion F are extremely important conceptual steps in the congested assignment problem
it is important to remember that there are still a large number of inherent errors
present. Although these are well known it might be useful to list some of the most
 404

important:

(1) Not all drivers perceive costs in the same way.

(2) Even if they did we would still have trouble in defining the correct formulae;
e.g., even in the simplest form cost = a l x time + a2 x distance there is no
general agreement on the values to be used for a l and a2.

(3) Equally there are errors in forecasting the individual components of generali-
zed cost; e.g., link times and their relationship to link flows. In particu-
lar the normal assumption that link time is a function ~ of traffic on that
link (see equation (2)) may be a gross oversimplification, particularly in ur-
ban areas. If we accept that most delays occur at intersections there may very
often be a strong relationship between the delay and the volume of cross
traffic, e.g., at a roundabout or priority junction.

(4) Although we normally assume uniform rates of flow between O-D pairs in reality
there can be significant variations over even the smallest time period.

(5) It is doubtful that network flows ever really do reach equilibrium given day-
to-day variations, road works, etc.

(6) "End effect" errors in aggregating trip ends into zones.

(7) Generally not all links are included in networks, particularly large ones such
as the GLTS.

(8) Banned turns are very often not included.

(9) Large errors in the trip matrices being assigned.

(10) Intra-zonal trips are very often not assigned at all

Given such a large number of potential errors in our basic formulation of the
assignment problem it seems reasonable to assume that the resultant link flows will
not be particularly accurate either, independent of the method used to calculate them.
Without attempting to attach quantitative estimates to these errors several important
qualitative conclusions may be drawn:
 405

First, a small value of the convergence parameter 0 does not necessarily imply
accurate link flows nor, for sufficiently small values of 0, need there be any corre-
lation between the two. It follows that we should be more concerned with developing
methods that reduce 0 to an acceptable value commensurate with the level of errors
elsewhere rather than those that guarantee reducing 0 to O. Extremely sophisticated
assignment procedures may not necessarily be justified in terms of practical returns.

Second, in evaluating assignment techniques, greater emphasis should be placed

on the "other" objectives listed in Section II. For example, as we shall argue in
the next Section, we at the GLTS are prepared to accept an inherent lack of conver-
gence provided that it is counterbalanced by relatively simple procedure for second-
ary analysis.

Third, as a very general observation, more attention should probably be paid to

eliminating errors in the input data which specify the assignment problem than to
eliminating inconsistencies in the mathematical techniques.

VI THE CHOICE OF MODEL FOR THE GLTS

Having defined our objectives for the GLTS assignment model and established,
with reservations, a quantitative measure of the degree of convergence which would
prove acceptable, a number of possible methods (generally of the heuristic or exist-
ence school) were considered both from a theoretical point of view and by a number of
representative tests. Our conclusion was that,given the size of the GLTS network and
the expected shape of the cost/volume relationship (2), an incremental loading proce-
dure would be most suitable.

The basic procedure proposed was as follows:

1) Set the network link costs to their free flow values.
2) Calculate minimum cost routes for all O-D pairs using these link costs.
3) Load a pre-determined fraction of the total trip matrix, e.g., 25%, onto
these routes.
4) Add these link volumes onto those already accumulated.
 406

S) Calculate new link costs corresponding to the accumulated link volumes via
equation (2).

6) If the entire trip matrix has been loaded, terminate; otherwise return to
step 2).

(In general an additional tree build and load sequence is necessary in order to de-
termine the minimum cost matrix and to calculate the convergence parameter 0).

It is recognized that this procedure is unlikely to yield the correct equili-

brium solution (see, for example, Florian and Nguyen(S)) due to the fact that routes
selected on the early increments may prove to be congested and unreasonably "expen-
sive" in the light of the total travel pattern. However tests described in Section
VII indicate that, on the former GLC network at least, the effect was not particu-
larly significant. In addition, as we shall describe in detail in Section VII, many
of the worst routings may be eliminated by re-assigning some of the badly routed
trips as an extra stage in the procedure above. Our experience with incremental
loading has therefore lead us to believe that a reasonable, although not a perfect,
degree of convergence may be achieved in a relatively short time.

We have also found a number of other virtues in incremental loading, corres-

ponding more with the "other" objectives listed in Section II than to absolute con-
vergence; for example:

(1) Secondary analysis is relatively simple since the total load consists of
4 to 5 explicitly defined routes per O-D pair plus corresponding trip
matrices. Virtually any analysis required, e.g., extracting a cordon
matrix or analysing O-D pairs along a link, is feasible, although in most
cases it is necessary to repeat the basic operation for each set of routes.

(2) It is relatively economical in terms of the computer time and core space
required.

(3) It is relatively straight forward procedure which may be easily understood

by the non-specialist; e.g., the traffic engineer who must use the results
or the politician who must approve them.
 407

(4) The progressive build up of traffic across the network can be extremely
instructive in terms of network behaviour.

Although our present opinions may be altered by our practical experiences in cali-
brating the new GLTS network and assignment, in particular by comparing synthesised
loads with measured traffic counts, so far both our judgment and experience with test
situations have tended to confirm the advantages of incremental loading.

VII TESTS USING THE GLC DISTRICT TEST NETWORKS

As mentioned above the prime objection to incremental loading has always been
that routes selected on the early increments may prove to be unreasonable choices
once the final link costs are known. In order to test this presumption on a repre-
sentative large network a number of tests w~re carried out using a network and time
(=cost)/volume relationship developed as part of the former GLC transportation model,
the "District Test Model". The network selected consisted of 252 origins, 1786 real
nodes and 6,554 one-way links. The form of the relationship assumed to hold between
link time and volume is shown as curve 'A' in Figure 1. Note in particular that for
traffic volumes in excess of the link capacity C the link time is held constant at
twice the free flow time; this rather unusual feature was introduced as a surogate
for extra (unspecified) restraint policies which would limit all link traffic to the
capacity specified.

The trip matrix to be assigned represented the total 24-hour O-D movements prior
to restraint, about 12 million trips in all. Under free flow conditions the mean
trip time was 12.4 minutes and near equilibrium, about 18.15 minutes; hence £2 ~ 31%
and following the arguments presented in Section III we conclude that congestion
effects are probably dominant in dispersing trips across the network.

Our initial trial consisted of four equal increments, 25% each, following the
procedure described in Section VI. The final link times were calculated and used to
find (i), the minimum O-D times t~}n and (ii), the times t~j) as measured along the
four sets of routes on which the trips had been loaded. Table 1, Part I gives the
 408

distribution of the ratios t~~)/t~~n for each increment and summed over all incre-
lJ lJ
ments. For example, 11.5% of the routes chosen on the first increment are 5 to 10%
above the minimum, 7.0% of all routes are 5 to 10% too high, etc. Also listed are
the error parameters o(n) and 0 as defined by equations (6a) and (6).

These results show that the number of unreasonable routes is in fact relatively
small. 42% of all routes taken are actually minimum cost routes while only 2.9% are
more than 10% above their minimum, most of these occurring on the first increment.
Overall the error parameter 0 is only 1.4% which, given the inherent errors involved,
is probably not too significant.

However it should be pointed out that due to the shape of the time-volume curve
A trips using overloaded links (23% of all links have VIC> 1) are not as severely
penalized as they no doubt would be in practice. To overcome this objection we cons-
tructed a second, purely hypothetical time-volume curve, B in Figure 1. This curve
is more in line with current GLC thinking which assumes that in urban areas speeds
are relatively constant at low flows but tend to decrease rapidly as the capacity is
reached and congestion occurs.

Repeating the original procedure of 4 equal increments with curve 'B' gave con-
siderably worse results - Part II of Table 1. 27% of all routes are now more than
10% above the minimum (cf. 2.9% previously) and the error parameter 0 = 5.8%. By
increasing the size of the early increments * - 33%, 33%, 22% and 12% respectively -
we improved the convergence somewhat but only to the extent that 0 was reduced to
5.2% - Part III of Table 1.

However since most of the unsatisfactory routes arose on the first increment a
purely heuristic method for removing the worst of these from the network was devised.

* This was based on the observation that in general the early increments tended to
use the high speed, high capacity links which did not become quickly overloaded,
whereas the later increments used low-capacity secondary routes which did become
overloaded. If the fastest links had had a low capacity it might have been advis-
able to load smaller increments first. ---
 409

The extra steps were as follows:

7) Define a sub-set ~~}) of trips from the first increment depending on the
relative i-j costs (see below)

8) Remove these trips from the routes assigned then in the first increment
(by loading a negative matrix).

9) Re-calculate the link costs and build new minimum cost trees.

10) Load the matrix ~~}) onto these routes.

11) Re-calculate final link costs, a final cost matrix and the final conver-
gence parameters.

After considerable experimentation it was found that the best results were
achieved by defining the matrix of trips to be re-assigned as follows:

o t~P = t~~n
lJ lJ

t~~n < t~~) < (4/3) t~~n

lJ lJ lJ
t(~) ~ (4/3) t~~n
lJ lJ

This means that first increment trips using minimum cost routes were unaffected,
those on routes over 33% too long were entirely re-assigned with a linearly varying
fraction in between.

As a result of this procedure the error parameter 0 was reduced from 5.2% to the
more acceptable value of 3.4%. Perhaps surprisingly it also improved the goodness of
fit of routes in the other increments - Part IV of Table 1. For example 85.8% of
routes 2, 3 and 4 were within 10% of their minimum and 94.5% within 15%; the pre-
vious figures were 80.5% and 90.0%. Although it is difficult to give comparable
figures for the first and last set of routes based purely on ij costs since a large
number of ij trips are specifically excluded in each, by the very nature of the re-
assignment there should be relatively few unreasonable routes in either.
 410

While we have not tested other assignment algorithms in as great detail as

incremental loading, or indeed other methods of re-assigning trips, there did appear
to be advantages in re-assigning ~ those trips which had been assigned to bad
routes. In addition this method places greater emphasis on identifying "bad" routes
by their relative excess cost rather than the absolute excess cost. This is in
accord with our intuitive ideas that drivers will be more liable to change routes if
they can save 5 mi nutes in 10 than if they can save 5 mi nutes in 100.

REFERENCES

(1) Wardrop, J.G., Proc. Inst. Civ. Eng., Part 11,1, 325 (1952).

(2) Dial, R.B., Transportation Research i, 83 (1972).

(3) Beckmann, M., McGuire, C.B. and Winsten, C.B., "Studies in the Economics of
transportation", Yale University Press (1956).

(4) Murchland, J.D., "Road Network Traffic Distribution in Equilibrium", paper

presented at the conference "Mathematical Methods in the Economic Sciences",
Mathematisches Forschungs-Institut, Oberwolfach (1969).

(5) Florian, M. and Nguyen, S., "A New Look at Some Old Problems in Transportation
Planning", paper presented at the PTRC Summer Annual Meeting, Warwick (1974).
 411

Figure 1 Link Time-Volume Relationships A and B

I
/
/
/
/
B /
/
/
/
/
/ A
2 x tmin

/
A

I
/
I B
-'

QJ
E
.~

o
o 2

Link Volume / Link Capacity

 412

TABLE 1

Distribution of t~j) / t~}n from different incremental loads

(All figures are given as percents)

Load No.
&Size Identical
1.00
-
1.05
-
1.05
1.10
1.10
-
1.15
1.15
-
1.20
1.20
-
1.30
1.30
-
1.40
1.40
1.50
-
"7
1.50 0
(%)
PART I - TIME/VOLUME CURVE A, EQUAL INCREMENTS
1 (25%) 34.4 46.8 11.5 4.6 1.6 0.59 0.03 - - 1.85
2 (25%) 38.7 51.0 8.3 1.5 0.15 0.01 - - - 1.6
3 (25%) 44.6 50.1 4.0 0.2 0.01 - - - - 1.3
4 (25%) 50.7 44.2 4.1 0.5 0.005 0.001 - - - 1.2
TOTAL 42.1 48.0 7.0 1.7 0.5 0.16 0.007 - - 1.4

PART II - TIME/VOLUME CURVE B, EQUAL INCREMENTS

1 (25%) 17.5 25.9 22.4 13.5 8.5 8.7 2.7 0.7 0.16 6.4
2 (25%) 18.7 29.2 21.8 12.6 7.8 7.6 1.8 0.3 0.05 5.7
3 (25%) 23.6 30.8 20.5 12.2 7.0 4.9 0.9 0.15 0.00 5.8
4 (25%) 33.0 32.0 16.8 9.1 4.9 3.8 0.4 0.04 0.01 5.3
TOTAL 23.2 29.5 20.4 11.9 6.5 6.3 1.5 0.3 0.06 5.8
PART III - TIME/VOLUME CURVE B, VARIABLE INCREMENTS
1 (33%) 16.6 26.5 20.7 13.0 8.4 9.7 3.6 1.1 0.33 6.7
2 (33%) 19.1 29.8 22.2 12.4 7.3 7.3 1.7 0.2 0.02 5.6
3 (22%) 28.7 36.8 19.4 9.6 3.7 1.6 0.2 0.03 0.00 4.4 5
4 (12%) 51.4 40.9 6.1 1.3 0.2 0.03 0.00 - - 1.4
TOTAL 24.0 31. 5 19.3 10.7 6.1 6.0 1.8 0.4 0.11 5.2
PART IV - RE-ASSIGNED TRIPS FROM PART III
2 (33%) 23.7 39.2 21.3 9.6 4.1 1.9 0.15 0.01 -
3 (22%) 27.6 35.6 19.6 10.3 4.1 2.3 0.4 0.08 0.01
4 ( 12%) 41. 3 41. 9 12.5 3.3 0.8 0.2 0.00 - -
TOTAL 3.4
 AN ANALYSIS AND COMPARISON OF BEHAVIORAL
ASSUMPTIONS IN TRAFFIC ASSIGNMENT

Larry J. LeBlanc
Southern t1ethodist University
Dallas, Texas, USA

Edward K. Morlok
University of Pennsylvania
Philadelphia, Pennsylvania, USA

This paper compares modeling of traffic flow on

networks in such a manner that total user cost is mini-
mized with the modeling of flow so that an equilibrium
is achieved. A detailed comparison of the output from
these two models is presented for a 76 arc, 24 node
sketch planning network used to model the city of
Sioux Falls, South Dakota. The results indicate that
the difference between the two sets of flows is small
enough to justify the simplifying assumption that total
user cost is minimized. Such an assumption is often
made in network design or link addition models tp achieve
computational ease, particularly in the context of multiple
objectives.
 414

INTRODUCTION

In this paper we will compare the modeling of flow on networks in such a

manner that total cost is minimized with the modeling of flow so that an equili-
brium is achieved. An analysis and comparison of the output from these two models
applied to a 24 node, 76 arc network used to model the city of Sioux Falls, South
Dakota (described in the next section) is presented. This analysis supports the
use of a simplifying assumption in network design modeling regarding the assignment
of traffic, namely that total user cost is minimized. This assumption makes net-
work design modeling significantly more tractable from a computational point of
view.

One of the most common behavioral assumptions made in urban transportation

planning is that each vehicle on a highway network will take the path of least per-
ceived cost between his origin and destination. It is known that travel time is

a significant factor to the majority of travelers when they choose their routes
in the network, and most urban studies have utilized travel time as the basic
measure of travel cost. (Comsis, 1972)

The assumption that each driver takes the path of least perceived cost gives
rise to the concept of network equilibrium. A set of flows along the arcs of a
network is said to be at equilibrium (Wardrop, 1952) if the following two conditions
are satisfied for every origin-destination pair, r-s:

(il If two or more routes between node r and node s are actually travelled,
then the cost to each traveler between rand s must be the same for
each of these routes.
(ii) There does not exist an alternative unused route between nodes rand s
with less cost than that of the routes which are travelled.

The assumption is made that each user of the network seeks to minimize his
own travel cost, and that he experiments with different routes, eventually finding
the least cost one. Although this may not be completely true in reality, it is
 415

assumed that those drivers using more costly routes constitute a negligible
portion of the total. It is clear that if (i) or (ii) were not true, some dri-
vers would switch to the cheaper routes, congesting them, and causing a new flow
pattern to evolve. An equilibrium is the aggregate result of individual decisions;
at an equilibrium, no single driver can reduce his own cost by choosing an alter-
native route in the network.

The two equilibrium conditions above are equivalent to Wardrop's first prin-
ciple, the principle of equal travel times for all users. His second principle,
that of overall minimization, leads to a different assumption of driver behavior.
Wardrop's second principle states that flows are distributed over the arcs of the
network in such a manner that the sum of the travel times for all users is mini-
mized. Models based on the assumption of equilibrium are often referred to as
"user-optimal" models, while models based on Wardrop's second principle are re-
ferred to as "system optimal" models. Since the assumption of equilibrium seems
to represent reality more closely, the majority of traffic assignment models used
today are user optimal models.

The travel time functions used by the U.S. Federal Highway Administration are
shown below. The average travel time function, Aij(x ij ), gives the time (typi-
cally in minutes) required for each driver to travel the length of arc ij when
the flow rate along the arc is x ... Because of congestion, the function is taken
lJ
to be nonlinear. At low levels of flow, A.. (x .. ) is nearly constant. However,
lJ lJ
as the flow reaches some threshold level, congestion sets in, and each driver's
travel time begins to increase rapidly. The total travel time function, Tij(x ij ),
equals the sl.l11 of every driver's travel time on the arc; thus Ti(x;j) = XijAij(Xij).
 416

A•• (x .. )
lJ 1J

Travel Time Functions Used by the U.S.F.H.W.A.

Figure 1

A system-optimal model, which assumes that flows are distributed over the arcs
of the network so that total travel time is minimized is of the form

Min L T .. (x .. ) (1)
ijEA lJ lJ

p s
s.t. x·· = L x1' J' all ijEA (2)
lJ 5=1

D(j,s) + L x~. = L x~k all nodes j, all destinations s (3)

i lJ k J

X5ij ~ 0 all arcs ij, all destinations s, (4)

where A = set of arcs in the network

X~j = flow along arc ij with destination s

Xij = total flow along arc ij

D(j,s) = number of vehicles which must travel between origin j and

destination s.
 417

The objective function (1) is the sum over the arcs of the network of the total
travel time functions; thus we wish to minimize total congestion in the network.
Constraints (2) are definitional constraints, stating that the flow along each arc
ij equals the sum of the flo~/s along the arc traveling to the various destinations
s. Constraints (3) are conservation of flow equations for each node j in the net-
work, for each destination s in the network. Because we must distinguish between
flows traveling to the various destinations, a set of constraints (3) is required
for each destination s; see (LeBlanc, 1973) for a more detailed discussion. Cons-
traints (4) are the non-negativity requirements on the flows.

Beckmann (1956) has shown that the equilibrium flows (i.e., the flows satisfying
conditions (i) and (ii) above) are given by the solution to the problem

(4)

subject to constraints (2), (3) and (4). A detailed discussion and proof of this is
also given in (LeBlanc, 1973). Using the F.H.W.A. Functions, the objective function
(2) in the system optimal problem is:

Min L [a .. x .. + (b .. )(x .. )5]

ijEA lJ lJ lJ lJ

while the objective function in the equilibrium problem (4) is

L
Min ijEA

Thus the only difference between the equilibrium problem and the system optimal pro-
blem is in the coefficient of the second term for each arc in the objective function.
Therefore it is intuitive that the equilibrium flows may not be significantly differ-
ent from the system optimal flows; this is borne out in the comparison which follows.

In (LeBlanc, 1973), the Frank-Wolfe algorithm is shown to be a very efficient

technique for solving the equilibrium problem and the system optimal problem. This
technique was used to solve these two problems for the network shown in Figure 2,
 418

In Table 1, the parameters for the travel time functions of the 76 arcs and the num-
ber of trips per day between each pair of nodes are shown.

Although the assumption of equilibrium is more common in the literature, a

study of the Sioux Falls network in Figure 2 indicated that the difference between
the system optimal flows and the equilibrium flo~/s is not significant. In Tables 2
through 4, a detailed comparison of three sets of flows is given: the equilibrium
and system optimal flows given by the Frank-Wolfe algorithm, and the system optimal
flows obtained by linearizing the user cost functions as in Figure 3 and using the
simplex method to solve the resulting (approximate) linear programming problem. By
symmetry of the trip table and travel time function, the optimal flows along each
arc ij are equal to the flows along arc ji. Therefore these tables compare only the
flows on the first 38 arcs of the network, rather than all 76. For each of the
thirty-eight arcs in the network, the tables give the type (E-expressway, A-arterial,
C-CBD, F-fringe), flow, total vehicle hours and total vehicle miles for each set of
flows. The units are respectively thousands of vehicles/day, thousands of vehicle
hours/day, and thousands of vehicle miles/day. For example in Tables 2 and 3, arc
is an expressway on the fringe of the network. From the Frank-Wolfe algorithm
applied to the equilibrium problem, we see that arc 1 has a flow of 17,400 vehicles/
day at equilibrium. The system optimal flow level on arc 1 is 18,700, a 7% change
from the equilibrium flow value. Using the linear approximation to the cost functions
yields a flow of 17,300, a 7% change from flow value of 18,700.

Looking again at the table, we see that there are 400 vehicle hours on arc 1
at equilibrium. The vehicle hours at the system optimal flows on arc 1 are 400, an
8% difference from the equilibrium vehicle hours (the apparent discrepancy is due to
rounding in the printout). The vehicle hours on arc 1 using the linearized cost
functions are 600, a 36% change from the system optimal vehicle hours. In the "Totals"
row, we see that there are 49,200 vehicle hours in the network at equilibrium, 48,100
vehicle hours at system optimal flows, and 48,700 vehicle hours at the system optimal
flows using linearized travel time functions. Furthermore, there are 45,700 vehicle
hours at equilibrium on arterial arcs in the network and 15,900 vehicle hours in the
CBD at equilibrium.
 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24
1 C 1 1 5 2 3 5 8 5 13 5 2 5 3 5 5 4 1 3 3 1 4 3 1
2 1 0 1 2 1 4 2 4 2 6 2 1 3 1 1 4 2 0 1 1 0 1 0 0
3 1 1 0 2 1 3 1 2 1 3 3 2 1 1 1 2 1 0 0 0 0 1 1 0
4 5 2 2 0 5 4 4 7 7 12 14 6 6 5 5 8 5 1 2 3 2 4 5 2
5 2 1 1 5 0 2 2 5 8 10 5 2 2 1 2 5 2 0 1 1 1 2 1 0
6 3 4 3 4 2 0 4 8 4 8 4 2 2 1 2 9 5 1 2 3 1 2 1 1
7 5 2 1 4 2 4 0 10 6 19 5 7 4 2 5 14 10 2 4 5 2 5 2 1
8 8 4 2 7 5 8 10 0 8 16 8 6 6 4 6 22 14 3 7 9 4 5 3 2
9 5 2 1 7 8 4 6 8 0 28 14 6 6 6 9 14 9 2 4 6 3 7 5 2
10 13 6 3 12 10 8 19 16 28 0 40 20 19 21 40 44 39 7 18 25 12 26 18 8
11 5 2 3 15 5 4 5 8 14 39 0 14 10 16 14 14 10 1 4 6 4 11 13 6
12 2 1 2 6 2 2 7 6 6 20 14 0 13 7 7 7 6 2 3 4 3 7 7 5
13 5 3 1 6 2 2 4 6 6 19 10 13 0 6 7 6 5 1 3 6 6 13 8 8
14 3 1 1 5 1 1 2 4 6 21 16 7 6 0 13 7 7 1 3 5 4 12 11 4
15 5 1 1 5 2 2 5 6 10 40 14 7 7 13 0 12 15 2 8 11 8 26 10 4
16 5 4 2 8 5 9 14 22 14 44 14 7 6 7 12 0 28 5 13 16 6 12 5 3
17 4 2 1 5 2 5 10 14 9 39 10 6 5 7 15 28 0 6 17 17 6 17 6 3
18 1 0 0 1 0 1 2 3 2 7 2 2 1 1 2 5 6 0 3 4 1 3 1 0
19 3 1 0 2 1 2 4 7 4 18 4 3 3 3 8 13 17 3 0 12 4 12 3 1
20 3 1 0 3 1 3 5 9 6 25 6 5 6 5 11 16 17 4 12 0 12 24 7 4
21 1 0 0 2 1 1 2 4 3 12 4 3 6 4 8 6 6 1 4 12 0 18 7 5
22 4 1 1 4 2 2 5 5 7 26 11 7 13 12 26 12 17 3 12 24 18 0 21 11 -l
OJ
23 3 0 1 5 1 1 2 3 5 18 13 7 8 11 10 5 6 1 3 7 7 21 0 7 c:r
~
24 1 0 0 2 0 1 1 2 2 8 6 5 7 4 4 3 3 0 1 4 5 11 7 0 ro
:!:::
CD

ARC A B ARC A B ARC A B ARC A B

1 .02 .00000001 20 .04 .00000893 39 .02 .00000001 58 .04 .00000893
2 .06 .00000027 21 .03 .00000790 40 .06 .00000027 59 .03 .00000790
3 .04 .00000007 22 ,06 .00001373 41 ,04 .00000007 60 .06 .00001373
4 .04 .00000002 23 .05 .00001241 42 .04 .00000002 61 .05 .00001241
5 .06 .00000002 24 .02 .00000521 43 .06 .00000002 62 .02 .00000521
6 .04 .00000002 25 .03 .00000119 44 .04 .00000002 63 .03 .00000119
7 .03 .00000001 26 .04 .00001001 45 .03 .00000001 64 .04 .00001001
8 .04 .00000002 27 .02 .00000451 46 .04 .00000002 65 .02 .00000451
9 .05 .00000075 28 ,02 .00000401 47 .05 .00000075 66 .02 .00000401
10 .03 .00000003 29 ,04 .00001020 48 .03 .00000003 67 .04 .00001020
11 .03 .00000010 30 .04 .00000960 49 .03 .00000010 68 .04 .00000960
12 .08 .00001930 31 .05 .00001085 50 .08 .00001930 69 .05 .00001085
13 .10 .00002306 32 .02 .00000401 51 .10 .00002306 70 .02 .00000401
14 .06 .00001550 33 .02 .00000554 52 .06 .00001550 71 .02 .00000554
15 .05 .00000075 34 .04 .00000958 53 .05 .00000075 72 .04 .00000958
16 .03 .00000012 35 .04 .00001061 54 .03 .00000012 73 .04 .00001061
17 .03 .00000052 36 .05 .00001130 55 .03 .00000052 74 .05 .00001130
18 .06 .00001550 37 .05 .00001157 56 .06 .00001550 75 .05 .00001157
19 .02 .00000003 38 .04 .00001080 57 .02 .00000003 76 .04 .00001080
 420

\- -------;-:-----====
(i ._-_.. _..
'i'
r
--...!.----.-.----.---
',3

26

64

51
6
13

37
39

38

76

73 35
40

31

69

~5 7
67 29

68

30

21
I 20
J ~
___ --------~~--------~--------lJ~~-----------------'~
59
~B

Figure 2
 421

Piecewise Linearization of the Total Travel Time Function

Figure 3

Next, the table compares vehicle miles on each arc for each set of flows. On
arc 1, there are 20,900 vehicle miles/day at equilibrium, 22,400 vehicle miles at
system optimal flows, and 20,800 vehicle miles using linearized cost functions. In
the "Totals" row, there are 634,200 vechile miles/day in the network at equilibrium.
The number of vehicle miles on arterial streets at equilibrium is 442,300.

Finally, the number and types of congested arcs (more than 1,300 vehicle hours)
are shown. For example, there are 14 congested arterial arcs and five congested CBD
arcs in the network at equilibrium. In Table 4 the same comparisons described above
are presented for each congested arc in the network.

Observe that total vehicle hours and miles in the network ("Totals" row, Tables
3 and 4) using a system optimum model with linearized cost functions differ by only
1% from the equilibrium values. The difference in vehicle hours and miles in the CBD
for the two algorithms is 2 or 3% and the average difference in flows along the arcs
of the network from the three algorithms is insignificant. At the bottom of Table 2
and 3, we see that the number of congested arcs at equilibrium is approximately the
same as the number of congested arcs when we minimize the linear user costs. In con-
clusion, the assumption that flows will be distributed over the arcs of the network
such that the linearized user costs are minimized yields results very close to those
of the assumption of equilibrium,
 422

Table 2

A Comparison of Flows, Total Vehicle Hours and Total Vehicle Miles on the Arcs of
the Network (E = expressway, A = arterial; C = CBD, F = fringe).

Equilibrium Flows From System Optimal Flows From

Frank-Wolfe Algorithm Frank~·Wolfe Algorithm

Arc ~ Flow Time Miles Flow !.Q!ff. Time % Diff. Miles % Diff.
1 E, F 17.4 .4 20.9 18.7 7 .4 8 22.4 7
2 A, C 24.9 4.0 27.4 25.4 2 4.3 7 27.9 2
3 A, F 18.1 .9 32.6 19.5 7 1.0 13 35.1 7
4 E, F 10.5 .5 26.3 13 . 5 28 .6 29 33.8 28
5 E, F 6.3 .4 21.4 9.0 42 .5 43 30.6 42
6 E, F 14.5 .6 34.8 18.3 26 .8 29 43.9 26
7 E, F 15.9 .5 31.8 18.0 13 .6 14 36.0 13
8 E, F 22.7 1.1 56.7 23.4 3 1.2 4 58.5 3
9 A, F 19.1 2.9 15.3 18.8 1 2.7 5 15.0 1
10 A, F 18.1 .5 16.3 21.2 17 .7 26 19.1 17
11 A, C 19.6 1.0 21.6 19.4 1 1.0 2 21.3 1
12 A, C 8.7 1.7 13.9 9.0 3 1.9 12 14.4 3
13 A, C 8.3 1.7 15.8 8,5 2 1.8 7 16.1 2
14 A, F 7.2 .8 10.8 7.3 1 ,8 3 10.9 1
15 A, F 18.4 2.5 20.2 19.1 3 2.9 14 21.0 3
16 A, C 23.8 1.6 9.5 23.5 1 1.5 4 9.4 1
17 A, F 18.9 1.9 22.7 17 .9 5 1.6 17 21. 5 5
18 A, F 8.6 1.3 17.2 8.1 5 1.1 17 16.2 5
19 A, F 21.1 .6 21.1 21.0 0 .6 0 21.0 0
20 A, F 11.9 2.6 17.8 11.7 1 2.4 7 17 .5 1
21 A, F 10.5 1.4 10.5 10.6 0 1.4 3 10.6 0
22 A, F 7.5 .8 9.7 7.3 2 .7 6 9.5 2
23 A, F 6.8 .5 14.3 7.5 10 .7 28 15.7 10
24 A, F 13.7 2.8 12.3 13.6 0 2.7 3 12.2 0
25 A, F 14.1 1.0 15.5 14.8 4 1.2 19 16.3 4
26 A, F 8.8 .9 13.2 7.7 12 .6 33 11.5 12
27 A, F 8.4 .3 4.2 8.5 1 .4 3 4.3 1
28 A, F 9.1 .4 5.5 9.2 1 .4 3 5.5 1
29 A, F 8.5 .8 7.6 8.3 2 .8 7 7.5 2
30 A, F 10.3 1.5 10.3 10.1 1 1.4 7 10.1 1
31 A, F 10.1 1.6 11.1 9.8 2 1.4 10 10.8 2
32 A, C 11.7 1.1 4.7 11. 1 5 .9 19 4.4 5
33 A, C 10.0 .8 6.0 8.6 13 .5 41 5.2 13
34 A, F 9.4 1.1 10.3 9.0 4 .9 14 9.9 4
35 A, F 10.6 1.9 9.5 10.1 4 1.6 17 9.1 4
36 A, F 8.2 .8 11.5 7.9 3 .7 10 11.1 3
37 A, C 9.0 1.1 10.8 8.6 4 1.0 14 10.3 4
38 A, C 11.8 2.:.Q 13.0 11.4 3 2.6 13 12.5 3
Totals 49.2 634.2 48.1 668.4
Arterial 45.7 442.3 44.0 443.2
Expressway 3.4 191.9 4.1 225.2
CBD 15.9 122,6 15.4 121.7
Fringe 33.2 511.6 32.7 546.7
Number and Types of Congested Arcs
Arterial 14 14
Expressway 0 0
CBD 5 5
Fringe 9 9
 423

Table 3

A Comparison of Flows, Total Vehicle Hours and Total Vehicle Miles on the Arcs of
the Network (E;:; expressway, A;:: arterial; C = CBD, F = fringe),

System Optimal Flows From System Optimal Flows From

Frank-Wolfe Algorithm Linearized Cost Functions
Arc ~ Flow Time ~ Flow % Di ff. Ti.me % Diff. Miles % Diff.
1 E, F 18.7 .4 22.4 17.3 7 .6 36 20.8 7
2 A, C 25.4 4.3 27.9 26.5 4 5.1 17 29.2 4
3 A, F 19.5 1.0 35.1 19.0 2 1.2 20 34.3 2
4 E, F 13.5 .6 33.8 12.7 5 .9 47 31.9 5
5 E, F 9.0 .5 30.6 8.7 3 .8 49 29.6 3
6 E, F 18.3 .8 43.9 15.6 14 1.0 25 37.6 14
7 E, F 18.0 .6 36.0 16.5 8 .8 40 33.0 8
8 E, F 23.4 1.2 58.5 21.0 10 1.4 23 52.5 10
9 A, F 18.8 2.7 15.0 20.0 6 3.4 25 16.0 6
10 A, F 21.2 .7 19.1 15.9 24 .7 2 14.4 24
11 A, C 19.4 1.0 21.3 21.3 9 1.1 10 23.4 9
12 A, C 9.0 1.9 14.4 9.4 4 2.1 11 15.1 4
13 A, C 8.5 1.8 16.1 7.5 11 .9 50 14.2 11
14 A, F 7.3 .8 10.9 7.5 2 .6 25 11.3 2
15 A, F 19.1 2.9 21.0 20.0 4 3.4 18 22.0 4
16 A, C 23.5 1.5 9.4 26.0 10 2.2 45 10.4 10
17 A, F 17.9 1.6 21.5 18.1 1 1.8 11 21.7 1
18 A, F 8.1 1.1 16.2 7.5 7 .6 44 15.0 7
19 A, F 21.0 .6 21.0 20.3 3 .7 11 20.3 3
20 A, F 11.7 2.4 17.5 12.0 2 2.7 11 18.0 2
21 A, F 10.6 1.4 10.6 9.9 6 1.4 2 9.9 6
22 A, F 7.3 .7 9.5 7.5 2 .5 23 9.7 2
23 A, F 7.5 .7 15.7 7.5 0 .5 29 15.7 0
24 A, F 13.6 2.7 12.2 12.0 11 1.4 48 10.8 11
25 A, F 14.8 1.2 16.3 13.0 11 .8 31 14.4 11
26 A, F 7.7 .6 11.5 7.5 2 .4 36 11.3 2
27 A, F 8.5 .4 4.3 7.5 11 .2 34 3.8 11
28 A, F 9.2 .4 5.5 9.3 0 .4 3 5.6 0
29 A, F 8.3 .8 7.5 7.5 9 .4 43 6.7 9
30 A, F 10.1 1.4 10.1 10.0 0 1.2 13 10.0 0
31 A, F 9.8 1.4 10.8 10.0 2 1.4 3 11.0 2
32 A, C 11. 1 .9 4.4 12.0 8 1.2 39 4.8 8
33 A, C 8.6 .5 5.2 9.7 12 .6 40 5.8 12
34 A, F 9.0 .9 9.9 9.3 2 1.0 4 10.2 2
35 A, F 10.1 1.6 9.1 10.5 3 1.8 16 9.4 3
36 A, F 7.9 .7 11.1 7.5 5 .4 40 10.5 5
37 A, C 8.6 1.0 10.3 7.5 12 .5 49 9.0 12
38 A, C 11.4 2.6 12.5 11.5 0 2.7 3 12.6 0
Tota 1s 48.1 668.4 48.7 641.9
Arterial 44.0 443,2 43.2 436.6
Expressway 4.1 225.2 5.5 205.3
CBD 15,4 121.7 16.4 124.7
Fringe 32..7 546.7 32.3 1)17.2
Number and Types of Congested Arcs
Arterial 14 12
Expressway 0 1
CBD 5 4
Fringe 9 9
 424

Tqble 4,
A Comparison~of Congested Arcs (More than 1300 Vehicle Hours).
E = expressway, A = arterial; C = CBD, F = fringe.

Equilibrium Flows From System Optimal Flows From

Frank-Wolfe Algorithm Frank~Wolfe Algorithm

Arc ~ Flow Time Miles Flow %Diff, Time % Diff. Miles %Diff.
2 A, C 24.9 4.0 27.4 25.4 2 4.3 7 27.9 2
8 E, F 22.7 1.1 56.7 23.4 3 1.2 4 58.5 3
9 A, F 19.1 2.9 15.3 18.8 1 2.7 5 15.0 1
12 A, C 8.7 1.7 13.9 9.0 3 1.9 12 14.4 3
13 A, C 8.3 1.7 15.8 8.5 2 1.8 7 16.1 2
15 A, F 18.4 2.5 20.2 19.1 3 2.9 14 21.0 3
16 A, C 23.8 1.6 9.5 23.5 1 1.5 4 9.4 1
17 A, F 18.9 1.9 22.7 17.9 5 1.6 17 21.5 5
20 A, F 11 .9 2.6 17.8 11.7 1 2.4 7 17.5 1
21 A, F 10.5 1.4 10.5 10.6 0 1.4 3 10.6 0
24 A, F 13.7 2.8 12.3 13.6 0 2.7 3 12.2 0
30 A, F 10.3 1.5 10.3 10.1 1 1.4 7 10.1 1
31 A, F 10.1 1.6 11.1 9.8 2 1.4 10 10.8 2
35 A, F 10.6 1.9 9.5 10.1 4 1.6 17 9.1 4
38 A, C 11.8 3.0 13.0 11.4 3 2.6 13 12.5 3

System Optimal Flows From System Optimal Flows From

Frank-Wolfe Algorithm Linearized Cost Functions
Arc ~ Flow Time Mi 1es 1'1 ow % Di ff. Time % Diff. Miles of,
-~~.-
Diff.
2 A, C 25.4 4.3 27.9 26.5 4 5.1 17 29.2 4
8 E, F 23.4 1.2 58.5 21.0 10 1.4 23 52.5 10
9 A, F 18.8 2.7 15.0 20.0 6 3.4 25 16.0 6
12 A, C 9.0 1.9 14.4 9.4 4 2.1 11 15.1 4
13 A, C 8.5 1.8 16.1 7.5 11 .9 50 14.2 11
15 A, F 19.1 2.9 21.0 20.0 4 3.4 18 22.0 4
16 A, C 23.5 1.5 9.4 26.0 10 2.2 45 10.4 10
17 A, F 17.9 1.6 21. 5 18.1 1 1.8 11 21.7 1
20 A, F 11.7 2.4 17.5 12.0 2 2.7 11 18.0 2
21 A, F 10.6 1.4 10.6 9.9 6 1.4 2 9.9 6
24 A, F 13.6 2.7 12.2 12.0 11 1.4 48 10.8 11
30 A, F 10.1 1.4 10.1 10.0 0 1.2 13 10.0 0
31 A, F 9.8 1.4 10.8 10.0 2 1.4 3 11.0 2
35 A, F 10.1 1.6 9.1 10.5 3 1.8 16 9.4 3
38 A, C 11.4 2.6 12.5 11.5 0 2.7 3 12.6 0
 425

REFERENCES

1. Beckmann, M., McGuire, C,B, and Winsten, C., Studies in the Economics of
Transportation, New Haven, Connecticut, Yale University Press (1956).

2. Comsis Corporation, Traffic Assignment; Methods~Applications~Products.

Prepared for Urban Planning Division, Office of Planning, Federal Highway
Administration, Preliminary-- Subject to Review (1972).

2. Dreyfus, S.E., "An Appraisal of Some Shortest Path Algorithms", Operations

Research, 394-412 (1969).

4. Jackson, L.F. et al., "Equilibrium Models", Search and Choice in Transport

System Planning, Vol. IX, Massachusetts Institute of Technology (1968).

5. LeBlanc, L.J., "Mathematical Programming Algorithms for Large Scale Network

Equilibrium and Network Design Problems", Ph.D. Dissertation, Dept. of
Industrial Engineering and Management Sciences, Northwestern University (1973).

6. LeBlanc, L.J., Morlok, E.K, and Pierskalla, W.P., "An Accurate and Efficient
Approach to Equilibrium Traffic Assignment on Congested Networks", Transporta-
tion Research Record 491 - Interactive Graphics and Transportation Systems
Planning, 12-33, (1974).

7. Martin, B.V. and Manheim, M.L., "A Research Program for Comparison of Traffic
Assignment Techniques", Highway Research Record 88, 69-84 (1965).

8. Murchland, J.D., "Braess' Paradox of Traffic Flow", Transportation Research

391-394 (1970).
 AN INVESTIGATION OF SOt1E PROBLEMS IN THE
CALIBRATION OF TRAFFIC ASSIGNMENTS

E.J. Judge
Leeds University
Leeds, England

Introduction

The fitting or calibration of a simulated assignment of traffic on a network

model with the flows observed on a real network is a basic part of the valiaCLtion of
a transport planning model. This paper is concerned with examining some of the pro-
blems involved in this calibration process and with presenting the results of some
empirical work on inter-urbain traffic assignment which will illustrate the nature of
some of these problems. The discussion is prefaced by an outline of the traffic
assignment process and concludes with some ideas for further work in this area.
 THE USE OF TRAFFIC EQUILIBRIUM BASED PROBLEM-ORIENTED LANGUAGE
FOR A GOVERNMENTAL FEASIBILITY STUDY

Bernard-Andre Genest, David Burden,* Jacques Muguet &Esat Kardicali

Service des projets speciaux et du developpement
Ministere des Transports, Gouvernement du Quebec

* Societe de Mathematiques appliquees Inc.

THE STUDY

In 1973, the Quebec Department of Transport created a task force to conduct a

study on the feasibility of linking the new international airport at Mirabel to
Montreal with a high speed guided transport system that would also provide regional
and suburban service.

This ETCAR (for Itude de lransport en fommun ~eroportuaire et Regional) study

produced the following recommendations:

1. the use of an exclusive right-of-way based largely on existing rail lines,

for most of the network;

2. by way of.technology high speed (100 mph) electric trains using conventional
principles but new design features and construction method;

3. a 55 mile network (Fig. 1) comprising 1 airport station (under an air ter-

minal) plus 15 regional stations including 4 transfer stations with the
Montreal metro system of which two at city centre;

4. .a very high quality of service including high frequency, a very high degree
of complementarity with other transit services, ample parking at stations,
and privileged access lanes for kiss'n'ride travellers.

Despite the high investment required ($438 million, 1973 values), and since the
system, known as TRRAMM (Transport Rapide Regional Aeroportuaire Montreal-Mirabel)
or (Mirable-Montreal Airport and Regional Rapid Transit) is expected to attract
 429

112000 trips per day in 1981, financial and economic analyses showed that the propo-
sal had relatively high rates of internal return.

The Government of Quebec approved the recommendations in principle in the Spring

of 1974, released a public report, and, last September, named a coordinator for design
and implementation.

ANALYTIC REQUIREMENTS

The analysis process, datawise, was based on a series of 83 x 83 trip matrices

compiled from origin-destination surveys conducted in 1970, 1971, and, for air pas-
sengers, 1972. Due to data availability and institutional constraints (such as muni-
cipal boundaries), zone population varied widely (from 1070 to 97600 persons) and
zone area even more (from 0.35 to 137.4 square miles).

The computerized road network was comprised of 1400 links with a 3.1 connecti-
vity, and the transit network (all modes), of 1200 links, with a connectivity of 2.5.

The analysis process used in this study (Fig. 2) is an expansion of the trans-
portation problem-selving process formulated by Manheim in several of his publica-
tions.

The iterative nature of the process, the feedback relationship between prelimi-
nary results and subsequent analyses, and the need for extensive and numerous para-
metric analyses were important requirements for this project.

Essentially, to select among transportation options in order to determine the

key parameters of the transit system, it is necessary to predict direct impacts that
can and must be quantified: systems costs, changes in travel times, modal split,
attracted ridership, and system revenues and costs. This is the core of the itera-
tive process.

Furthermore, and this is perhaps more important in the decision process func-
tional impacts must be identified, appraised, and, to all possible extent, quantified.
 430

r SYSTEMES D' ACTIVITES .....--------OBJECTIFS :~----_.

J
DEVELOPP~ENT REGIONAL

Analyse de la
OBJECTIF: ~------I

r
cl ientele SATISFAIRE A LA DH1ANDE

CARACTERISTIQUES DES
Formulation desJ
objectifs-service
BESOINS DE DEPLACEMENTS

I )
CRITERrS DE
SERVICE ET ~E PERFORMANCE

Elaboration
Recherche
systematiqu.;,,:.,e.....l
AL TERNATIVES: ALTERN TIVES:
SCHEMAS D'AMENAGEMENT CONFIGURATIONS
PROGRA~1t1ES DE POLITIQUES D'OPERATION
DEVELOPPEI4ENT STRATE IES D' H1PLANTATION
Prediction
i nformati see

TEr·1PS ET REPARTITION DES DEPLACH1ENTS --------',~

UTILISATION COUTS ET REVENUS
..v
n

1.(,----------'mPACTS FONCTIONNELS----------~
-l,
Evaluation
Distribution
Ponderation
PREFERENCES------------~

Selection

CHOIX
t
Figure 2. Processus dynamigue d'analyse tenant compte de 1 'interaction,
des innovations et de la revision des objectifs
 431

To achieve these objectives requires a well-structured computerized set of pro-

cedures, that is easy to use and efficient. Within this group of procedures, traffic
assignment must be accurate, and, mainly, efficient because it will be done often.

At the time the study started, DODOTRANS appeared to be the problem-oriented

language for transportation analysis that had the required size capability and achie-
ved the largest number of analysis objectives with a reasonable degree of accuracy
and efficiency.

The remarks made on this model, it is believed, would apply to most transporta-
tion software packages based on traffic equilibrium methods when used for a govern-
mental project.

OBSERVATIONS ON THE USE OF DODOTRANS

Although theoretically satisfactory, the Baumol-Quandt demand model, the only

option available in DODOTRANS, presented some fundamental as well as very practical
difficulties.

In urban travel indeed, a good proportion of trip-making, such as for work and
studies (80% of all trips in Montreal), must be done irrespective of available trans-
port services. As a result, the introduction of a new mode, as far as these trips
are concerned, tends to affect modal split rather than increase travel, Of course,
discretionary trip-making may be increased.

Also, when a drastic improvement in transit is provided, such as the TRRAMM

system for the Montreal suburbs, there are doubts as to whether the hypothesis of
linearity still prevails. Namely, the elasticities to attributes * in the suburbs
in the current situation pertain to transit ridership more than 80% captive whereas
TRRAMM ridership would be 20% captive or less.

Finally, the estimation of coefficients for th.e Beaumol-Quandt model in this

case would have to have been based on close to 50,000 observations, many of which
were simply not available.
*
of transportation modes and services
 432

Hence DODOTRANS could not be used as an integrated model, but only to estimate
travel times by equilibrium assignment on the road and transit networks separately,
based on present and future modal trip matrices compiled exogeneously. This implied,
in effect, flat demand curves.

Despite the fact that, in theory, as well as due to the mechanics of the DOOO-
TRANS assignment algorithm, this should have resulted in prediction errors of link
volumes, it was possible to calibrate both networks satisfactorily compatible witb
the purpose of the study.

Significant technical problems were encountered at this step, however, mainly

related to input-output procedures and error-checking features. These are described
at some length in the full paper.

NOTE: The conclusions formulated in this report as well as the opinions expressed
herein are strictly those of the authors and do not represent an official
position of the Government of Quebec.
Vol. 59: J. A. Hanson, Growth in Open Economies. V, 128 pages.
1971.
Vol. 60: H. Hauptmann, Schatz- und Kontrolltheorie in stetigen
dynamischen Wirtschaftsmodellen. V, 104 Seiten. 1971.
Vol. 61 : K. H. F. Meyer, Wartesysteme mit variabler Bearbeitungs-
rate. VII, 314 Seiten. 1971.
Vol. 62: W. Krelle u. G. Gabisch unter Mitarbeit von J. Burger-
meister, Wachstumstheorie. VII, 223 Seiten. 1972.
Vol. 63: J. Kohlas, Monte Carlo Simulation im Operations Re-
search. VI, 162 Seiten. 1972.
Vol. 64: P. Gessner u. K. Spremann, Optimierung in Funktionen-
raumen. IV, 120 Seiten. 1972.
Vol. 65: W. Everling, Exercises in Computer Systems Analysis.
VIII, 184 pages. 1972.
Vol. 66: F. Bauer, P. Garabedian and D. Korn, Supercritical Wing
Sections. V, 211 pages. 1972.
Vol. 67: I. V. Girsanov, Lectures on Mathematical Theory of
Extremum Problems. V, 136 pages. 1972.
Vol. 68: J. Loeckx, Computability and Decidability. An Introduction
for Students of Computer Science. VI, 76 pages. 1972.
Vol. 69: S. Ashour, Sequencing Theory. V, 133 pages. 1972.
Vol. 70: J. P. Brown, The Economic Effects of Floods. Investiga-
tions of a Stochastic Model of Rational Investment. Behavior in the
Face of Floods. V, 87 pages. 1972.
Vol. 71 : R. Henn und O. Opitz, Konsum- und Produktionstheorie II.
V, 134 Seiten. 1972.
Vol. 72: T. P. Bagchi andJ. G. C. Templeton, Numerical Methods in
Markov Chains and Bulk Oueues. XI, 89 pages. 1972.
Vol. 73: H. Kiendl, Suboptimale Regier mit abschnittweise linearer
Struktur. VI, 146 Seiten. 1972.
Vol. 74: F. Pokropp, Aggregation von Produktionsfunktionen. VI,
107 Seiten. 1972.
Vol. 75: GI-Gesellschaft fUr Informatik e.V. Bericht Nr. 3. 1. Fach-
!agung Gber Programmiersprachen . MGnchen, 9.-11. Marz 1971.
Herausgegeben im Auftrag der Gesellschaft fUr Informatik von H.
Langmaack und M. Paul. VII, 280 Seiten. 1972.
Vol. 76: G. Fandel, Optimale Entscheidung bei mehrfacher Ziel-
setzung. II, 121 Seiten. 1972.
Vol. 77: A. Auslender, Problemes de Minimax via l'Analyse Con-
vexe et les Inegalites Variationelles: Theorie et Algorithmes. VII,
132 pages. 1972.
Vol. 78: GI-Gesellschaft fUr Informatik e.V. 2. Jahrestagung, Karls-
ruhe, 2.-4. Oktober 1972. Herausgegeben im Auftrag der Gesell-
schaft fUr Informatik von P. Deussen. XI, 576 Seiten. 1973.
Vol. 79: A. Berman, Cones, Matrices and Mathematicar Program-
ming. V, 96 pages. j 973.
Vol. 80: International Seminar on Trends in Mathematical Model-
ling, Venice, 13-18 December 1971. Edited by N. Hawkes. VI,
288 pages. 1973.
Vol. 81: Advanced Course on Software Engineering. Edited by
F. L. Bauer. XII, 545 pages. 1973.
Vol. 82: R. Saeks, Resolution Space, Operators and Systems. X,
267 pages. 1973.
Vol. 83: NTG/GI-Gesellschaft fUr Informatik, Nachrichtentech-
nische Gesellschaft. Fachtagung .Cognitive Verfahren und Sy-
sterne", Hamburg, 11.-13. April 1973. Herausgegeben im Auf trag
der NTG/GI von Th. Einsele, W. Giloi und H.-H. Nagel. VIII, 373
Seiten. 1973.
Vol. 84: A. V. Balakrishnan, Stochastic Differential Systems I.
Filtering and Control. A Function Space Approach. V, 252 pages.
1973.
Vol. 85: T. Page, Economics of Involuntary Transfers: A Unified
Approach to Pollution and Congestion Externalities. XI, 159 pages.
1973.
Vol. 86: Symposium on the Theory of Scheduling and its Applica-
tions. Edited by S. E. Elmaghraby. VIII, 437 pages. 1973.
Vol. 87: G. F. Newell, Approximate Stochastic Behavior of n-Server
Service Systems with Large n. VII, 11B pages. 1973.
Vol. BB: H. Steckhan, GGterstrome in Netzen. VII, 134 Seiten.
1973.
Vol. B9: J. P. Wallace and A. Sherret, Estimation of Product.
Attributes and Their Importances. V, 94 pages. 1973.
Vol. 90: J.-F. Richard, Posterior and Predictive Densities for
Simultaneous Equation Models. VI, 226 pages. 1973.
Vol. 91: Th. Marschak and R. Selten, General Equilibrium with
Price-Making Firms. XI, 246 pages. 1974.
Vol. 92: E. Dierker, Topological Methods in Walrasian Economics.
IV, 130 pages. 1974.
Vol. 93: 4th IFACIIFIP International Conference on Digital Com-
puter Applications to Process Control, Part I. Zurich/Switzerland,
March 19-22,1974. Edited by M. Mansour and W. Schaufelberger.
XVIII,544 pages. 1974.
Vol. 94: 4th IFACIIFIP International Conference on Digital Com-
puter Applications to Process Control, Part II. Zurich/Switzerland,
March 19-22, 1974. Edited by M. Mansour and W. Schaufelberger.
XVIII, 546 pages. 1974.
Vol. 95: M. Zeleny, Linear Multiobjective Programming. X, 220 pages.
1974.
Vol. 96: O. Moeschlin, Zur Theorie von Neumannscher Wachs-
tumsmodelle. XI, 115 Seiten. 1974.
Vol. 97: G. Schmidt, OberdieStabilitatdeseinfachen Bedienungs-
kanals. VII, 147 Seiten. 1974.
Vol. 98: Mathematical Methods in Oueueing Theory. Proceedings
1973. Edited by A. B. Clarke. VII, 374 pages. 1974.
Vol. 99: Production Theory. Edited by W. Eichhorn, R. Henn,
O. Opitz, and R. W. Shephard. VIII, 386 pages. 1974.
Vol. 100: B. S. Duran and P. L. O~ell, Cluster Analysis. A Survey.
VI, 137 pages. 1974.
Vol. 101: W. M. Wonham, Linear Multivariable Control. A Geo-
metric Approach. X, 344 pages. 1974.
Vol. 102: Analyse Convexe et Ses Applications. Comptes Rendus,
Janvier 1974. Edited by J.-P. Aubin. IV, 244 pages. 1974.
Vol. 103: D. E. Boyce, A. Farhi, R. Weischedel, Optimal Subset
Selection. Multiple Regression, Interdependence and Optimal
Network Algorithms. XIII, lB7 pages. 1974.
Vol. 104: S. Fujino, A Neo-Keynesian Theory of Inflation and
Economic Growth. V, 96 pages. 1974.
Vol. 105: Optimal Control Theory and its Appl;r.ations. Part I. Pro-
ceedings 1973. Edited by B. J. Kirby. VI, 425 pages. 1974.
Vol. 106: Optimal Control Theory and its Applications. Part II. Pro-
ceedings 1973. Edited by B. J. Kirby. VI, 403 pages. 1974.
Vol. 107: Control Theory, Numerical Methods and Computer
Systems Modeling. International Symposium, Rocquencourt, June
17-21,1974. Edited by A. Bensoussan and J. L. Lions. VIII, 757 pages.
1975.
Vol. 108: F. Bauer et aI., Supercritical Wing Sections II. A Hand-
book. V, 296 pages. 1975.
Vol. 109: R. von Randow, Introduction to the Theory of Matroids.
IX, 102 pages. 1975.
Vol. 110: C. Striebel, Optimal Control of Discrete Time Stochastic
Systems. III. 208 pages. 1975.
Vol. 111: Variable Structure Systems with Application to Economics
and Biology. Proceedings 1974. Edited by A. Ruberti and R. R. Mohler.
VI, 321 pages. 1975.
Vol. 112: J. Wilhlem, Objectives and Multi-Objective Decision Making
Under Uncertainty. IV, 111 pages. 1975.
Vol. 113: G. A. Aschinger, Stabilitatsaussagen uber Klassen von
Matrizen mit verschwindenden Zeilensummen. V, 102 Seiten. 1975.
Vol. 114: G. Uebe, Produktionstheorie. XVII, 301 Seiten. 1976.
Vol. 115: Anderson et aI., Foundations of System Theory: Finitary
and Infinitary Conditions. VII, 93 pages. 1976
Vol. 116: K. Miyazawa, Input-Output Analysis and the Structure of
Income Distribution. IX, 135 pages. 1976.
Vol. 117: Optimization and Operations Research. Proceedings 1975.
Edited by W. Oettli and K. Ritter. IV, 316 pages. 1976.
Vol. 118: Traffic Equilibrium Methods, Proceedings 1974. Edited by
M. A Florian. XXIII, 432 pages. 1976.
Okonometrie und Unternehmensforschung
Econometrics and Operations Research
Vol. I Nichtlineare Programmierung. Von H. P. Kunzi und w. Krelle unter
Mitwirkung von w. Oettli. Vergriffen
Vol. II Lineare Programmierung und Erweiterungen. Von G. B. Dantzig. Ins
Deutsche ubertragen urid bearbeitet von A. Jaeger. - Mit 103 Ab-
bildungen. XVI, 712 Seiten. 1966. Geb.
Vol. III Stochastic Processes. By M. Girault. - With 35 figures. XII, 126
pages. 1966. Cloth
Vol. IV Methoden der Unternehmensforschung im Versicherungswesen. Von
K. H. Wolff. - Mit 14 Diagrammen. VIII, 266 Seiten. 1966. Geb.
Vol. V The Theory of Max-Min and its Application to Weapons Allocation
Problems. By John M. Danskin. - With 6 figures. X, 126 pages. 1967.
Cloth
Vol. VI Entscheidungskriterien bei Risiko. Von H. Schneeweiss. - Mit 35
Abbildungen. XII, 215 Seiten. 1967. Geb.
Vol. VII Boolean Methods in Operations Research and Related Areas. By P.
L. Hammer (lvAnescu) and S. Rudeanu. With a preface by R. Bellman. -
With 25 figures. XVI, 329 pages. 1968. Cloth
Vol. VIII Strategy for R&D: Studies in the Microeconomics of Development.
By Th. Marschak. Th K. Glennan Jr., and R. Summers. - With 44 figures.
XIV, 330 pages. 1967. Cloth
Vol. IX Dynamic Programming of Economic Decisions. By M. J. Beckmann. -
With 9 figures XII, 143 pages. 1968. Cloth
Vol. X Input-Output-Analyse. Von J. Schumann. - Mit 12 Abbildungen. X,
311 Seiten. 1968. Geb.
Vol. XI Produktionstheorie. Von W. Wittmann. - Mit 54 Abbildungen. VIII,
177 Seiten. 1968. Geb.
Vol. XII Sensitivitatsanalysen und parametrische Programmierung. Von W. Din-
kelbach. - Mit 20 Abbildungen. XI, 190 Seiten. 1969. Geb.
Vol. XIII Graphentheoretische Methoden und ihre Anwendungen. Von W.
Knodel. - Mit 24 Abbildungen. VIII, 111 Seiten. 1969. Geb.
Vol. XIV Praktische Studien zur Unternehmensforschung. Von E. Nievergelt,
O. Muller, F. E. Schlaepfer und W. H. Landis. - Mit 82 Abbildungen.
XII, 240 Seiten. 1970. Geb.
Vol. XV Optimale Reihenfolgen. Von H. Muller-Merbach. - Mit45 Abbildungen.
IX, 225 Seiten. 1970. Geb.
Vol. XVI Preispolitik der Mehrproduktenunternehmung in der statischen Theo-
rie. Von R. Selten. - Mit 20 Abbildungen. VIII, 195 Seiten. 1970. Geb.
Vol. XVII Information Theory for Systems Engineers. Byl. P. Hyvarinen. - With
42 figures. VIII, 197 pages. 1970. Cloth
Vol. XVIII Unternehmensforschung im Bergbau. Von F. L. Wilke. - Mit 29 Ab-
bildungen. VIII, 150 Seiten. 1972. Geb.
Vol. XIX Anti-Aquilibrium. Von J. Kornai. - Mit 31 Abbildungen. XVI, 382 Seiten.
1975. Geb.
Vol. XX Mathematische Optimierung. Von E. Blum, W. Oettli. IX, 413 Seiten.
(davon 75 Seiten Bibliographie). 1975. Geb.
Vol. XXI Stochastic Linear Programming. By P. Kall. VI, 95 pages. 1976. Cloth.