Universities of Leeds, Sheffield and York

http://eprints.whiterose.ac.uk/





Institute of Transport Studies
University of Leeds


This is an ITS Working Paper produced and published by the University of
Leeds. ITS Working Papers are intended to provide information and encourage
discussion on a topic in advance of formal publication. They represent only the
views of the authors, and do not necessarily reflect the views or approval of the
sponsors.


White Rose Repository URL for this paper:
http://eprints.whiterose.ac.uk/2415/



Published paper
Willumsen, L.G. (1978) Estimation of an O-D Matrix from Traffic Counts A
Review. Institute of Transport Studies, University of Leeds, Working Paper 99




White Rose Consortium ePrints Repository
eprints@whiterose.ac.uk

Working Paper 99
ESTIMATION OF AN 0- D MATRIX
FROM TRAFFIC COUNTS - A REVIEW
L G Willurnsen
August 1978
ITS Working Papers are intended to provide infirmution and encourage discussion
on a topic in advance of formal publication. They represent only the views of the
authors, and do not necessarily reflect the views or approval of the sponsors. This
review is part of a research project on the estimation of trip matrices from t mf i c
counts financed by the Science Research Council.
ABSTRACT
WILLU:.EEN, L.G. (1978) Estimation of an 0-D matrix from traffic counts
- a review. Leeds: University of Leeds, Inst. Transp. Stud., WP 99.
Following a review of conventional methods,for estimating a
trip matrix,the possibility of using cheaper methods based on traffic
counts is discussed. Three broad groups of models for such estimation
have been identified. The first one assumes that trips follow a gravity
type pattern and the problem is reduced to calibrating the parameters
of such a model from the observed counts. Depending on the assumed
form of the gravity model this technique leads to linear or non-linear
regression solutions.
The second group of models attempts to estimate the 0-D matrix
through a network equilibrium approach based on Wardrop's first principle.
The third group follows an entropy maximising approach in which the most
likely trip matrix compatible with the observed flows is sought.
Finally, the advantages of each approach are discussed and the
most promising technique and areas of application are identified.
CONTENTS
Abstract
1. The problem of estimating a trip matrix
Objectives
Conventional methods
Basic concepts and definitions
Assignment
Reliability of traffic counts
Estimability of a trip matrix from counts
Criteria for selecting a method of estimation
Overview of traffic count based methods
2. Linear models
2.1 Introduction
2.2 Low's model
2.2.1 Description
2.2.2 Scope and applications
2.2.3 Further developments
2.3 Overgaard's model
2.3.1 Description
2.3.2 Applications
2.4 Lamarref s model
2.4.1 Description
2.4.2 Results
2.5 .Commonwealth Bureau of Roads model
2.5.1 Description
2.5.2 Application
2.6 Danish Road Directorate approach
2.6.1 Description
2.6.2 The iterative process
2.6.3 Applications
3. Non-linear regression approach
3.1 Introduction
3.2 Hogberg's model
3.3 Robillard's model
3.4 Will's model
3.4.1 Description
3.4.2 Applications
4. Network equilibrium approach
4.1 Introduction
4.2 Nguyen's method, case I
4.3 Nguyen's method, case I1
5. Entropy maximising approach
5.1 Background
5.2 Van Zuylen's model
5.2.1 Description
5.2.2 Inconsistent constraints
5.3
An alternative entropy maximising model
5.3.1 Description
5.3.2 Extensions and limitations
6. Discussion
6.1 Introduction
6.2 Gravity type approaches
6.3 Equilibrium assignment approach
6.4 Entropy maximising approach
6.5 Extensions to include congestion
7. References
8. Appendix 1. Notation
9. Appendix 2. A least square estimation of the trip
factor for the DRD model
9.1 Method
9.2
Application to the DRD model
ESTIMATION OF AN 0-D MATRI X
FROM TRAFFIC COUNTS - A REVI EW
1. THE PROBLEM OF ESTIMATING A TRIP. MATRIX
1.1 Objectives
This review i s part of a research project financed by t he Science
Research Council on t he estimation of origin - destination t r i p matrices
from t r af f i c counts.
For a large number of applications t he existence of a rel i abl e and
rel at i vel y inexpensive method for estimating 0-D matrices would be very
valuable. For example:
- modelling transport demand i n towns
- modelling assessment and design of t r af f i c management schemes
i n r ur al and urban areas
- cross-checking other methods
- updating old 0-D matrices
- modelling r ur al or i nt er urban transport demand i n sparsely
populated areas.
- modelling transport demand where other data i s non existent,
unreliable or out of date as perhaps in developing countries.
A t t hi s stage we are almost exclusively concerned with journeys by car,
as are a l l t he methods reviewed here. The plan of t hi s report i s as follows:
The remainderof'%ction . one bri efl y reviews current methodologies
for t he estimation of 0-D matrices. This is followed by a discussion of some
basic concepts l i ke t r i p generation, distribution and assignment models. The
Section is closed with a comment on t he practical and t heoret i cal problems
of estimating a t r i p matrix from counts.
Section two reviews t he first group of techniques, namely l i near
models. Section t hree discusses an extension of these i nt o non-linear models.
Section four desribes methods based on t he assumption of equilibrium
assignment and Section five discusses entropy maximising models.
Finally Section s i x compares t he theoretical and pract i cal implications
of these four groups of models-and makes some recommendations for pract i cal
applications. Appendix I summarises the notation used i n t hi s paper.
1. _
1.2 Conventional methods
There ar e a number of methods for estimating an or i gi n - dest i nat i on
t r i p matrix. A l l but one of them imply sampling and even i f a l l t r i p s on a
day ar e recorded t he problem of s t a bi l i t y of t he 0-D matrix over time s t i l l
remains. "Rea1"tri.p matrices then ar e never avai l abl e; we only have more
or l es s r el i abl e methods f or estimating them.
-
Some of t he techniques avai l abl e a t present ar e:
( a )
Road si de interviews of dri vers, usually a t t he exi t of a
cordoned area. Sampling fract i ons of 1 i n 5 may be considered
t ypi cal but t hey cert ai nl y depend on t r a f f i c l evel s and manpower
avai l abl e. They ar e usually asynchronic, t ha t i s, t hey ar e
carri ed out on di fferent days but a t si mi l ar times. They tend
t o be expensive i n manpower, delays t o vehi cl es and processing.
(b) Home interview. Although usually not considered r el i abl e enough
f or a det ai l ed design of , f or example, a t r a f f i c management
scheme, they may be t he r es ul t of a l a r g w study and be avai l abl e
f or cross-checking. The method i s expensive on manpower and
time consuming.
( c ) Flagging methods. It i s possible t o i dent i f y vehi cl es a t cordon
and i nt er nal points by means of r egi st r at i on numbers, st i cker s
or by asking dri vers ent eri ng at a given poi nt t o switch t he i r
l i ght s on. This l a s t method has only been t r i e d f or very small
areas (roundabouts), see f or example BebEe (1959). The number
pl at e method i s of wider usage and al so requi res a good deal
of manpower and processing ef f or t . Usually dri vers do not
incur ext ra delays.
( d) Aeri al photography. To a l arge extent t h i s method i s experimental,
it requi res a good deal of processing t o i dent i f y cars and t rack
them i n a computer readable form. It is only successful f or
small areas but improvements i n t he automatic i dent i f i cat i on of
vehicles ma y increase t he i r appl i cabi l i t y. It i s t he only
method which does not necessarily require sampling.
( e ) Car following. Suggested by Wright (1977) and t r i e d i n Westminster,
t h i s method has been claimed t o be cheaper than ot her methods.
Its appl i cat i on i s probably r es t r i ct ed t o t he cent r al areas of
l ar ge c i t i e s .
I n cont rast , we ar e i nt erest ed i n t h i s review i n methods f or
t he estimation of 0-D matrices based on a descri pt i on of t he network
and t r a f f i c counts. I n addition some of t he methods requi re some widely
avai l abl e zonal dat a, l i ke population and employment. Traffi c counts are
r el at i vel y simple t o obtain and furt her progress i n automatic recording
and cl as s i f i cat i on i s envisaged. We suspect t hat not enough use i s
made of t h i s information.
1.3 Basic concepts and defi ni t i ons
Most of t he vocabulary i n t h i s review i s borrowed from conventional
t ransport demand models.
I n a conventional t ransport modelling exercise t he study area i s
usually divided i nt o zones which ar e considered t he generators and
at t r act or s of t r i ps . The modelling process usually proceeds i n a sequence
of four submodels:
- Trip generation which uses zonal data t o model t he number
of t r i p s being generated from and at t r act ed t o each zone.
- Tri p distrLbution which synthesises .the origin-destination matrix,
t hat i s t he number of t r i p s from each zone t o each zone. It i s
common pr act i ce t o use here a gravity type model. This, i n
general, assumes t ha t t he number of t r i p s between i and j ar e
proportional t o t he generation and at t r act i on power of each zone
and proportional t o some decreasing f'unction of t he cost ( i n
time, di st ance or monetary terms) of t r avel l i ng between them.
This decreasing function i s usually cal l ed t he deterrence
function of t he model.
- Modal s p l i t i n which t he choice of mode made by each t r avel l er
i s simulated. We s ha l l not be concerned with t h i s submodel
i n t h i s review, as we r e s t r i c t our at t ent i on t o car t r a f f i c .
- Assignment, i n which t he rout e followed by each t r i p is
modelled. The output f r omt hi s stage includes l i nk flows and
a revised measure of t he cost s of t r avel l i ng between each pai r
of zones.
As this assignment stage is very relevant to our problem we will discuss
it in some detail.
1.4 Assignment
A road network is usually described in terms of a collection of
points, called mdes and a collection of ordered pair of mdes called
'!real links" or simply "links".. A link is usually deemed to represent
a real section of a road, for example from one important intersection to
another. Most of the delays in urban areas occur at intersections. Thus,
in some cases it is worthwhile to define a network in such a way that
certain links represent particular turning movements at an intersection.
Figure 1.1 represents a way in which an intersection could be represented
by links rather than modes.
Figure 1.1 "NODE" AND "LINK" REPRESENTATION OF AN INTERSECTION
intersection represented intersection represented
by one node. by 6 two way links and 4
nodes.
Of course with a "link" representation the number of nodes and links
is greatly increased. On the other hand, it is customary to count traffic
precisely at intersections recording each turning movement independently,
thus this representation is of interest.
Origin and Destination zones may be considered as a sub-net of the
nodes, generally referred to as "centroids". The arc linking a "centroid"
with the "real link" network is usually called a "centroid connector".
Each link has a travel cost associated with it (time, distance, money
or generalised cost) which, due to congestion, depends on the total flow
on that link expressed as
Ca = ca(Va) (1.1)
C = C ( ~ a )
where a i s a l i nk
Ca is a l i nk cost
Va
i s t he t r a f f i c volume on l i nk a
Ibr a good discussion of t he problem of t r a f f i c assignment t he
reader i s referred t o Van Vl i et (1976). Here we s hal l only describe
br i ef l y t hree al t er nat i ve approaches.
( a ) All-or-nothing assignment vi a t he minimum cost route. If
we assume t hat all drivers t r avel l i ng between two zones choose
t he same rout e it seems only l ogi cal t o assume t hey t r avel
vi a t he (apparently) cheapest route ( a sequence of l i nks ) .
I n t he absence of congestion or i f t her e ar e no r e a l
al t er nat i ve routes it i s possible t o determine t he s et of
minimum cost routes before making any assumption about t he
number of t r i p s using them.
( b) Proportional assignment. If we accept t ha t not al l dri vers
,I
perceive" a rout e cost t o be t he same we must allow i n t he
assignment process f or t hese differences. I n t h i s case t he
proportion of dri vers choosing each of a reduced s e t of near
optimum routes w i l l be determined by assumed dri vers and
route charact eri st i cs. These methods current l y do not consider
congestion effect s so t hese proportions ar e t r eat ed as
independent of l i nk flow l evel s.
Following Robillard (1975) it i s possible t o s t r i c t l y define
proportional assignment as a technique i n which:
( i )
The total. assigned flow on a l i nk equals t he sum of
t he assigned flows obtained when t he method i s applied
t o each 0-D pai r separately.
( i i )
If a l l t he ent r i es of t he 0-D matrix ar e mul t i pl i ed
by a constant fact or A al l t he assigned flows on each
l i nk w i l l increase by t hat fact or.
Clearly t hese conditions ar e not r e a l i s t i c when congestion
plays an important r ol e i n assignment. On t he ot her hand,
t he advantage ofJhese assumptions i s t ha t it i s possible t o
aetermine t he proportion of t he t r i ps between i and j using
a l i nk before any guess i s made about t he 0-D matrix. These
proportions depend only on t he network and t he parameters of
t he assignment algorithm but not on flow l evel s a t each l i nk.
( c ) Equilibrium assignment. Wherever congestion ef f ect s ar e thought
t o be more important t han t he differences i n perceived t r avel
cost s a di fferent type of assignment is considered r eal i s t i c.
This type t r i e s t o sat i sf y Wardrop's f i r s t pri nci pl e which
can be st at ed as:
Traffi c on a network di st r i but es i t s e l f i n such a way
t ha t t he t r avel costs on a l l routes used from any or i gi n t o
any dest i nat i on are equal while a l l unused rout es have equal
or great er cost s.
I n ot her words an assignment i s sai d t o have reached equilibrium
when no dri ver can reduce hi s t r avel cost by switching t o another
rout e.
Methods based on both heuri st i c and mathematical programming techniques
ar e usea t o obtain t hi s type of assignment. I n t h i s case t he proportion of
t r i p s hetween an 0-D pai r using each rout e depends on t he act ual flows on
each l i nk.
1. 5 Rel i abi l i t y of t r a f f i c counts.
There ar e a number of problems associated with t r a f f i c counts and we
s ha l l review some of them.
( a) Independence. Not a l l t r a f f i c counts contain t he same amount
of 'information". For example i n Figure 1.2 t r a f f i c l i nk 3-4
i s made up of t he sum of t r a f f i c on l i nks 1-3 and 2-3.
Counting t r a f f i c on l i nk 3-4 i s then redundant and only two
counts t here can be sai d t o be independent.
Figure 1.2 DEPENDENT COUNTS
I n mathematical terms t h i s can be expressed by wri t i ng f or each l i nk i n
t he network:
Dependence i n t he flows occurs where it i s possible t o express one of t hese
equations as a l i near combination ofsome of t he other equations.
This i s l i kel y t o happen i n si t uat i ons such as t ha t depicted i n
Figure 1.2 where a flow cont i nui t y equation appl i es
The existence of a source or sink of flow (cent roi d connector) a t
3 w i l l immediately remove t h i s dependence.
(b) Inconsistency. Counting er r or s and asynchronous counting
are l i kel y t o l ead t o inconsistencies i n t he flows. I n other
words, t he expected cont i nui t y rel at i onshi ps w i l l not be met.
As we w i l l see some models can readi l y cope with t h i s problem
while ot hers requi re t hese inconsistencies t o be removed
beforehand.
( c ) Counting periods. Depending on t he counting method it is
possi bl e t o obtain t r a f f i c flows for periods as short as 15
minutes or as long as 24 hours. An 0-D matrix derived on t h i s
basi s w i l l only be relevant i f t he t r i p time within t he study
area i s si gni f i cant l y short er than t he counting period. For
i nt e r urban st udi es it w i l l usually be possi bl e t o estimate
t he average Annual Daily Traffi c ( MT ) and use it i n t he
models.
(d) Symmetry. Flows are in general asymmetric, that is
V.. # Vji. In most inter urban applications, though,
13
it is possible to assume symmetry in the counts and in the
0-D matrix thus reducing the complexity of the models.
1.6. Estimability of a trip matrix from counts.
a
Let Tij be the number of trips from zone i to zone j, and pi j be
the proportion of trips from i to j travelling on a link. For each link
we can now write the equation
a
where p is the proportion of trips from i to j using link a.
i j
a
In a study area with N centoids (zones) and assumingE.. to be
2 I J
known we would need N independent equations of the type (1.3) to uniquely
define the whole set of L-~ij-i If we assume that Tij = 0 for i = j we
would only need N(N-1) equations. However, in most cases the number of
independent link counts will be less than this figure. The most important
implication of this is that traffic counts will normally be insufficient
to uniquely determine an 0-D matrix L-~ij-7. Thus, in order to obtain an
estimation of the 0-D matrix some assumption will have to be made about
trip making behaviour.
As we will see when reviewing the proposed models, a frequent
approach is to assume some sort of gravity type behaviour in trip making.
In this way the actual number of unknowns is reduced to the number of
parameters for calibration in the assumed model.
If the number of unknowns is smaller than the number of independent
equations (link counts) errors have to be assumed in the model or the
observations in order to solve the system. For example
where gais an error term with assumed mean 0. The solution is then found
by-minimising the supposed errors, usually by means of multiple (linear
-
or otherwise) regression anal ysi s.
Most computer implementations include software t o solve t hi s type
of problem.
1. 7
Cr i t er i a for sel ect i ng a method of estimation.
For short term appl i cat i ons, (when t he assumption of a fi xed t r i p
matrix is r e a l i s t i c ) , t he main cr i t er i on i s probably accuracy.
I n order t o t e s t t he accuracy of a model t he i deal case would be t o
compare t he estimated 0-D matrix with t he "real 0-D matrix" which produced
t he counts used by t he method. However, we did not fi nd any such t e s t
i n t he l i t er at ur e, no doubt because a "real" 0-D matrix i s di f f i c ul t t o
come by. Instead, a more common i ndi cat or of t he accuracy of a model,
i s t o compare observed and predicted l i nk flows vi a, f or example
,&:
RMS er r or = /- 1 (OBSERVED - PREDICTED) /
/ n-1
(Root mean square er r or )
/
/
- -
/
expressed ei t her i n absolute value or as a percentage of t he average
observed value. This measure could al so be used t o t e s t t he accuracy
of each c e l l i n t he 0-D matrix and t hi s t e s t would be much more si gni fi cant .
I n addition, we would l i ke t o know how acceptable i n each case ar e
t he assumptions i mpl i ci t i n each method. For example, i s it possible t o
expect a gravi t y type o f t r a v e l behaviour? Is it sensi bl e t o disregard
congestion ef f ect s and assume rout e choice t o be independent from flow
l evel s?
Fi nal l y, i f we ar e i nt erest ed i n medium t o long term implications
we would l i ke t o know i f t he i mpl i ci t t r avel demand model i s suf f i ci ent l y
pol i cy sensi t i ve f or our purposes and i f it i s possible t o have a r el i abl e
forecast fi gure f or t he planning parameters i n t he model (Population,
employment, et c . ) .
We s ha l l ret urn t o t hese considerations i n s ect i ons i x.
1. 8 Overview of t r a f f i c count based methods.
We have only found a r es t r i ct ed number of publ i cat i ons i n t he f i el d.
We suspect, though, t hat other researchers and pr act i t i oner s have developed
and used si mi l ar methods without reporting them i n t echni cal journals.
We have grouped t he proposed methods i nt o t hr ee broad groups:
The f i r s t assumes t hat a gravity model i s capable of explaining
most of t he t r avel behaviour i n t he study area. The underlying gravity
model can be presented with several degrees of sophi st i cat i on and t h i s
has l e d us t o consider two sub-groups:
-.Gravity models leading t o l i near equations on t he l i nks
- Gravity models leading t o non-linear equationson t he l i nks.
Most of t he proposed methods and g t h e appl i cat i ons with r eal
dat a f a l l i nt o t hi s- br oad first category.
The second group of models attempts t o estimate an O-D matrix through
a network equilibrium approach based on Wardrop's first pr i nci pl e. These
ar e described i n Section 4.
The t hi r d group of models considers methods f or estimating t he l i kel y
t r i p matrix compatible with t he obser vedt r af f i c counts. These methods
could be r ef er r ed t o as entropy maximising or information minimising approaches
and ar e described i n Section 5.
2. LINEAR MODELS
2.1 Introduction
The i dea of using a simplified model f or t r avel demand i s not new.
Probably t he first such model cal i br at ed from l i nk flows wa s proposed by
Low (1972). The basi c i dea i s t o use a simple l i near model of t he gravity
t ype l i k e
where Pi, Ej = population, employment i n zones i and j.
d. . =
distance or t r avel time between i and j.
1 J
a = parameter f or cal i br at i on.
In this model a plays a combined role, as trip generation-attraction
parameter. Then, in order to calibrate a the modelled trips must be
translated into link flows to compare them with observed counts. This
implies an assumption about the type of trip assignment considered to be
realistic. The simplest assumption is to accept all-or-nothing assignment,
that is
(
0
if trips from i to j do not travel on link a.
a (
' j ( 1. if they do.
Then
a
Va = 1 T. .p.
- 2 a
1~ lj = ~C.(P.E.~.. )pij
i j i.2 1 J 1J
in which all parameters except a are known. It is then necessary to determine
a in such a way that the difference (or the sum of the squares of the
differences) between modelled and observed flows is minimisea. This can be
achieved here using standara linear regression techniques.
This chapter reviews different linear models as suggested by Low,
Overgaard, Lamarre and Jensen et al. They differ mainly with respect to:
- The independent variables
- The deterence function
- The type of assignment.
This is the family of models with the greatest number of applications,
mostly in rural or inter-urban areas.
2.2 D.E.Low1s Model
2.2.1 Description. The main objective of Low's approach is to "effectively
combine into one process what is usually handled in a series of three
or four sub models, each with its own set of errors." (LOW 1972)
Attempts to develop a model whose direct output is traffic volumes
and where modelling errors,which appear in the same terms, can be
easily described and assessed.
Low sees his model as applicable, in principle, to a large range of
study areas but especially to small ones (around 50,000 inhabitants).
-.
1.8 Overview of t r a f f i c count based methods.
We have only found a r es t r i ct ed number of publications i n t he f i el d.
We suspect, though, t hat other researchers and pr act i t i oner s have developed
and used si mi l ar methods without reporting them i n t echni cal journals.
We have grouped t he proposed methods i nt o t hr ee broad groups:
The first assumes t ha t a gravi t y model i s capable of explaining
most of t he t r avel behaviour i n t he study area. The underlying gravity
model can be presented with several degrees of sophi st i cat i on and t hi s
has l ed us t o consider two sub-groups:
-.Gravity models leading t o l i near equations on t he l i nks
- Q.avity models leading t o non-linear equationson t he l i nks.
Most of t he proposed methods and s t h e applications with real
dat a f al l i nt o this' broad first category.
The second group of models attempts t o estimate an 0-D matrix through
a network equilibrium approach based on Wardrop's f i r s t pri nci pl e. These
ar e described i n Section 4.
The t hi r d group of models considers methods f or estimating t he l i kel y
t r i p matrix compatible with t he observed t r a f f i c counts. These methods
could be referred t o as entropy maximising or information minimising approaches
and ar e described i n Section 5.
2. LINEAR MODELS
2.1 Introduction
The i dea of using a si mpl i fi ed model f or t r avel demand i s not new.
Probably t he f i r s t such model cal i brat ed from l i nk flows was proposed by
Low (1972). The basic i dea i s t o use a simple l i near model of t he gravity
type l i ke
where Pi, Ej = population, employment i n zones i and j.
-.
d- .
=
distance or t r avel time between i and j.
13
a = parameter f or cal i brat i on.
H i s approach consi st s of t hr ee st eps:
( a ) Assignment of ext ernal t r i ps . Current ext ernal t r i p s as
obtained from ext ernal cordon roadside interviews are
assigned t o t he exi st i ng network t o produce estimates of
"current ext ernal volumes" throughout t he network.
( b) Estimation of "current i nt er nal volumes". The so cal l ed
"current i nt er nal volumes" ar e then obtained from t he act ual
counted flows l e s s t he assigned "current ext ernal volumes".
These "current i nt er nel volumes" ar e t he r es ul t of t r i p s
11
wholly explainable i n terms of area char act er i st i cs. "
( c ) I nt er nal volume forecast i ng model.
(i)
I nt er zonal t r i p opportunity matrices of t he form
OPj ar e developed. For t he journey t o work Population
and Employment ar e used.
( i i ) Each element of t he t r i p opportunity matrix i s then
-m
mul t i pl i ed by a f r i ct i on fact or o f t h e type Ci j . I n
general time i s suggested as measure of cost . The product
111
0. D. C. . i s cal l ed then an "i nt er zonal t r i p pr obabi l i t y
1 J 1J
f act or " and t he full s et of f act or s i s t he t r i p
pr obabi l i t y matrix. Several t r i p pr obabi l i t y f act or s,
say f or di fferent journey purposes and person t ypes, may
be so developed.
-m
x . = 0 1,. c. .
1 J 1J
Where X = t r i p probabi l i t y f act or between i and j f or
n i j
person t yqe or journey purpose n.
( i i i )
Tri p pr obabi l i t y matrices ar e then assigned separately
t o t he current network ( i n pr act i ce an all-or-nothing
assignment) j ust as if t hey were t r i ps .
For a l i nk a
(1 i f t he l east cost route from i t o j
where p = ( passes through l i nk a
(0 otherwise.
( i v) . Multiple regression techniques are used t o develop
equations of tlie following form
where $ i s t he internal t r af f i c volume on l i nk a and bk
are t he constants t o be obtained.
(T)
Low suggests t hat separate equations can be developed
for different types of roads or areas or t hat additional
parameters defining t he characteristics of t he l i nk may be
included.
2.2.2 Scope and applications. It is suggested t hat t he model could be
used for forecasting purposes as well as replicating t he present
pattern of t r i ps. To t hi s end t he value of t he socio economic
parameters should be obtained for t he design year as well as t he
al t ernat i ve networks t hat would be tested.
Low al so suggests t hat more sophisticated versionsof hi s model
can be developed, especially i n rel at i on t o t he way i n which t he
t r i p probability matrix i s obtained.
No provision i s made for public transport t r i ps or for modal s pl i t .
The model was used during t he Monongalia County Transportation Study
i n West Virginia i n 1970-71. Conventional techniques were used i n
par al l el t o t hi s model.
The basic information used was
-
Base year population and employment by zone (52 zones)
-
Base year road t r af f i c assignment network
- Base year t r af f i c counts
- Base year external-trips t abl e
Finally, tests were made with multipath assignment techniques
to see if it was possible to improve the fit of the model. The
model used was the standard UMTA multipath assignment program
UROAD. Only marginal improvements were obtained with this
approach.
A conventional model was not applied to the area so it was not
possible to compare their accuracies. Nevertheless, the level
of accuracy of the IVF model in reproducing base year link
volumes was "certainly within the limits of conventional models."
This is to some extent surprising as only population and zone
to zone travel times are included in the model. The reported
ccrrtelation coefficient was R = 0.87 and the RMS values were
209 in absolute terms and 53% in percentage terms.
In order to evaluate the forecasting ability of the model
empirically the model was applied to "forecast" 1960 traffic
volumes using the 1960 Fond du Lac network and census data.
The results of this exercise showed that overall, the "absolute
error of the 1960 estimate of assigned voXumes was less than
that of the 1970 estimates."
The researchers, though favourably impressed by the accuracy
of the IXF mode1,were critical of its theoretical basis.
Smith and McFarlane pointed out three basic mis-specification
errors in the model.
(a)
Changes in the trip making propensity of the study area
population over time cannot be considered in the model.
For instance, the impact of car ownership is not reflected
in the model.
(b) The unconstrained nature of the model. If population is
doubled traffic flows will be multiplied by a factor of
four instead of just doubling as would be expected. This
criticism applies mainly to the model used in the
forecasting mode. This objection is easily overcome by
the inclusion of a "normalising factor" equal to
1 0 in equation (2.3).
j
(c)
The different measures of the trip probability factors
are likely to be co-linear in which case only one
probability factor should be included in the equation.
The researchers considered several possible improvements to the
model, some of which are described later in this working paper
as they appear in the literature. They finally concluded that
the inclusion of these improvements would probably made the
IVF model far too similar to conventional techniques and its
original attractiveness would be lost.
2.3 Overgaard's Model
2.3.1.Description Overgaard's,Lowls and the Danish Road Directorate methods
were discussed in the OECD Report on possibilities for the simplification
of urban traffic models OECD (1974).
Overgaard's model, though probably developed independently, can be
considered as an improved version of Low's approach. As it includes
a prow measure for trip making propensity it goes some way to answer
some of the criticisms of Smith and McFarlane. This model has been
applied to Silkeborg, a Danish town of 44,000 inhabitants.
External trips are obtained in the same way as in Low's method. The
main change is regarding trip generation:
Car ownership levels were known for Silkeborg so that trip generation
equations were stated in terms of trips per car per day (It has to be
remembered these methods are only concerned with car movements).
Furthermore, as a proxy for socio economic level, the type of dwelling
("single family house" or "apartment") was also introduced in the
trip generation term.
The general trip generation (and attraction) term was then expressed
in the form
where 0. = generation/attraction force in zone i
1
E. Pi = employment and population in zone i
1
-.
6i. = percentage of population in "one family houses"
This t r i p generation model wa s cal i brat ed independently a t two
di fferent zones.
The other main difference with Low's approach i s t ha t Overgaard
uses two di f f er ent deterrence functions f or motorised t r i p s of
l e s s and more than 90 seconds of duration.
2.3.2 Applications. When t h i s model was applied i n Silkeborg it was found
t ha t
bl = 1.75 - /-trips per workplace-7
b2
= 0.7 - /-trips per inhabitant-7
b3 = 0.008 - /-trips per inhabitants i n "one family houses" - 7
gave t he best r esul t s.
Also, an exponent of -1.8 f or t r avel time wa s cal i br at ed f or t r i p s
with t > 90 secs. The 'hccuracy" of t he model i n terms of t he
comparison of observed and predicted l i nk flows was measured by an
RMS (percentage) error. This was found t o be around 20% and bet t er
f or higher flows.
2.4 Lamarre's Model.
2. 4. 1. ~es cr i ~t i on. I n general l i near models are r es t r i ct ed i n terms of t he
number of parameters f or cal i brat i on. For example, t he cal i brat i on
of t he t r i p generation - at t r act i on and di st ance exponent parameters
i s made i n st eps. A sui t abl e, and sounder, al t er nat i ve method
would be t o use a non-linear regression technique as we s hal l see
i n Section 3.
Lamarre (1977) developed and t es t ed a simple l i near model incoroporating
cal i br at abl e parameters affect i ng socio-economic and t r a ve l time
i ndi cat ors. Lamarre developed two al t er nat i ve formulations. The
f i r s t one is an "0-D based" formulation as follows
Ti j
= bo + bl(Oi + Oj) + b2 Cij
(2.8)
and
where O i , Oj ar e indicators of t he generation and at t r act i on
"power" of each zone and ci j i s t he t r avel time between zones.
His second model i s a "link based" formulation where
It can be seen t ha t t he main difference between t he two formulations
l i e s i n t he assumption of a constant term bo i n ei t her each 0-D
pai r or each l i nk. Lamarre developed t h i s model f or t he anal ysi s
of i nt e r urban t r i ps . He used data from c i t i e s at t he i nt er i or of
Quebec and t he relevant network description and t r a f f i c counts of
t hat area. Lamarre apparently performed most cal cul at i ons without
using an. exi st i ng computer programme ( f or example hi s short est
route i s determined by hand and t he number of t r i p s between cert ai n
0-D pai r s ar e more or l es s ar bi t r ar el y assigned 50% t o each of
two pat hs) .
2.4.2.Results. Lamarre t est ed several socio-economic i ndi cat ors as
independent vari abl es including population, employment, t o t a l income,
car ownership, French speaking population, et c. He al so t es t ed hi s
model f or summer and winter t r i p making. H i s main findings can be
summarized as follows
( a )
The "l i nk based" formulation always gave a bet t er f i t t han
t he "0-D based" one.
(b) The t r avel time el as t i ci t i es b2 were found t o range between
-2.0 f or t he "0-D based" and -0.4 f or t he "l i nk based"
formulations.
( c ) The e l a s t i c i t i e s associated with population and car ownership
were not sat i sfact ory (wrong si gn) and t h i s made Lamarre
doubt t he soundness of t h i s l i near model. H i s best coeffi ci ent
2
of determination was R = 0.72.
He acknowledges t hat t he r es ul t s , on t he same network, of a non-linear
model by W i l l s (see section 3. 4) were superior t o hi s two l i near models.
2.5
Commonwealth Bureau of Roads model
2.5.1 Description.
This approach was developed f or t he Commonwealth Bureau
of Roads (cBR) i n Aust ral i a by Symons, Wilson and Paterson (1976).
The objective of t he exercise was "to devise a rigorous methodology
f or disaggregating t he Australian National System of Urban Centres
i nt o di s t i nct zones of economic act i vi t y; then, t o use, t hi s zonationto
construct a model of t r a f f i c generation for t he National Highway
System". This approach i s only applicable t o i nt er - ci t y t r avel .
The main innovations of t h i s approach resi des a t t he t r i p generation
stage. A rel at i onshi p i s assumed between t r a f f i c generation and t he
urban hierarchy which i s determined by t he provision of servi ces t o
l es s er urban centres. A hierarchy of urban cent res i s produced using
Central Place Theory considerations and it i s assumed t hat t r i p s w i l l
be generated from a low order centre t o a higher order cent re only.
The t r i p generation r at es per head of population ar e expected t o vary
from l evel t o l evel . Each of t he Australian urban cent res was then
assigned t o one of seven broad ranks using population and employment
type s t a t i s t i c s . As explanatory variables both population and a
recreat i onal at t ract i veness index were used. The model could then be
expressed as a group of sub models.
Tri p Purpose Functional Form (Fr)
Seeking l evel 7 services
Seeking l evel 6 services
Seeking l evel 5 servi ces
Seeking l evel 4 services
Recreation
Int er-capi t al
Where Pi = Population urban centre i.
Ri = Recreational index urban centre i.
-k
C. . = Deterence function.
1 J
-k
'ij
P. Cij
-k
1
-k
Pi Cij
P. C. .
-k
1 1.l
P. R. C. .
-k
1 J 1 J
P. P. C
-k
1 J i j
F = Functional form f or journey purpose r.
r
-. ..
The general model is then
a
and Va = 1 T. .P..
ij IJ 1J
that is
-
a
b FP..
' a - 2 1 r r l j
~j r
where equations (2.12) can be used to determine the calibration
parameters b for each journey purpose. Of course an implicit
r
assumption in this model is that route choice does not depend on
journey purpose and that proportional assignment is sufficiently
realistic.
In order to simplify calculations only trips between State capitals
and trips between a centre and the nodes in its market area were
considered. It was felt that this set of trips would capture the
bulk of inter urban traffic.
2.5.2 Application. In the Australian case assignment was performed manually
and multiple linear regression techniques were used to calibrate the
model. In addition different exponents k for the deterrence function
for each journey purpose were tested and those giving best fit were
adopted. They found that the smallest exponent was that of trips to
level 5 and suggested that perhaps the most essential and irreplaceable
goods and services were supplied by these centres. The researchers
also found a high exponent for inter-capital trips. They explained
this by the impact of competition from other modes (air and rail services).
Symons et a1 concluded that the model was to be of great interest to
State Highway Authorities for it indicates how much information may be
extracted from census information and road counts. They were satisfied
that the results were encouraging and statistically sound.
They
envisaged no conceptual difficulties in transforming the model into a
"robust predictive tool". The researchers mentioned a number of desirable
improvements in that direction mainly in terms of automating processes
within the model.
2.6 Danish Road Directorate approach.
2.6.1.Description. This i s probably t he method with t he great est number
of reported appl i cat i ons. A t l eas t two ful l scal e applications i n
r ur al areas, ens en' LNielsen (1973) .md Holm e t a1 (1976), have
been reported i n English. The method has been improved si nce i t s
f i r s t conception but only t he version described by Holm e t a1 (1976)
w i l l be discussed here.
I n general t he method i s again seen as a way of di r ect l y cal i brat i ng
a trip-generation-distribution-assignment model from t r a f f i c counts.
Probably t he main point of i nt er es t i s t he use of an i t er at i ve
assignment technique i nst ead of t he simpler all-or-nothing vi a
minimum cost path approach of previous researchers.
I n general "external t r i ps " ar e not obtained independently but as
a r es ul t of t he use of t he model. I n ef f ect , t he "external zones"
ar e coded i n t he same way as the' finterndfones and t eat ed i n t he
same way.
The basic t r i p generation-distribution model is:
where b i s t he t r i p generation fact or.
The problem i s then t o determine a value f or b consi st ent with t he
i t er at i ve assignment and which minimises t he di fference between
observed and modelled l i nk flows.
An i t er at i ve assignment has been chosen, hopefully converging
a
towards a Wardrop's equilibrium, and consequently t he p. . values
1.J
change at each i t er at i on. The Danish Road Directorate chose
Smock's (1968) i t er at i ve algorithm f or assignment and developed
a method for cal cul at i ng b as part of t he i t er at i ve process.
2.6.2 The i t er at i ve process. The "cost-flow" rel at i onshi p f or a l i nk
a i s assumed t o be
Then a l l or par t of t he t r a f f i c between each pai r of zones i s
assigned for routes i n such a way t ha t , f or each l i nk, t he flow
a t i t er at i on n i s
1
where V a = t r a f f i c volume resul t i ng from all-or-nothing
assignments t o t he l i nk af t er t he nth i t er at i on.
We can now creat e an auxi l i ary vari abl e X (andogous t o
a
Low's Xa) such t hat
. .-m, a
Xa = 10iO.. C ~ J p. .
i j J
1J
a
where pij = proportion of t r a f f i c T. . which uses l i nk a.
13
A t t he same time
1 1
van = bxan = b(1- --) xan-' + - bXa
n (2.16)
by making
A t t hi s s t a ge , it would be possible t o cal cul at e, af t er each i t er at i on,
t he value of b using l i near regression, and we could accept t hat
convergence has been achieved when t he estimations of b do not
di f f er si gni fi cant l y i n two consecutive i t er at i ons .
The Danish Road Directorate however preferred a maximum likelihood
..
method. It i s assumed t ha t t he observed flows Va ar e mutually
independent, normally di st ri but ed variables with mean van, t hat i s
A
Va = van + &a (2..18)
wher e. gai s a normally di st ri but ed vari abl e
2 -
wi t h mean 0 and variance [(Vanfo -/ t
t$= o means t ha t t he standard deviation i s independent
from t he mean.
f.1 means t ha t mean and variance ar e i n proportion.
y = 2 means t ha t average flow and standard deviation ar e
proportional.
Using t he maximum l i kel i hood method it i s possi bl e t o f i nd
t he best estimator f or b as
and al so
The deri vat i on of t hese rel at i onshi ps i s shown i n Appendix 11.
The algorithm developed by t he DRD can be sumnarised as:
Step 1 = Calculate f r ee flow t r avel times.
Step 2
=
Determine minimum cost routes f or a l l 0-D pai rs.
Step 3 = Assign following Smock's algorithm, t he X! .' =
13
-m
0.. 0 . c fact ors t o t he network.
1 J i j
Step 4 = Estimate b* with t he maximum l i kel i hood formulae.
Step 5
=
Calculate t r a f f i c volumes on al l links as bnxa.
Step 6 = Calculate new t r avel times.
Step 7
=
I f b* bas not converged proceed t o s t ep 2,
otherwise stop.
This algorithm is t es t ed wi t h di fferent values of t h e exponent
and m i n a manner si mi l ar t o Smith and McFarlane (1978).
The model i s intended mainly as a replacement f or more
sophi st i cat ed and expensive forecasting models.
2.6.3 Applications. The model has been applied, i n s l i ght l y di fferent
versions, i n t he Aarhus and i n t he South Zealand r ur a l areas i n
Denmark. I n t h i s last case t he dat a used consisted of a road
network, with seven types of roads (and speed-flow rel at i onshi ps) ,
pari sh populations and about 40 counting points. The network
included 73 zones and 334 links. It was found that f = 1 and
m = 3.25 gave the best fit.
The researchers found that the model tended to over estimate
large flows and under estimate small volumes. This is certainly
not very desirable from a planning view point. They thought the
source of this error was the fact that intrazonal trips were not
considered by the model.
The researchers estimated the percentage error in flows on the
links -gas around 17% and that this did not deteriorate much when
the model was calibrated with only half the counts.
The researchers also prepared confidence intervals for the curves.
They found that the half-width of the 95% confidence interval
was, for example, 5200 pcu/day at 10.000 pcu/day and about 3000
pcu/day at 5.000 pcu/day levels. They concluded that this accuracy
was no worse than that obtained through the use of more expensive
traditional modes.
In a personal communication one of the researchers involved
informed the author that the Danish Road Directorate was now
working on a non-linear regression approach, similar to those
described in the next section.
3. NON-LINEAR REGRESSION APPROACH
3.1 Introduction.
Several techniques for solving non-linear regression problems
are available and implemented in most computers. A non linear regression
problem normally takes longer to solve than a linear one with the same
number of variables. However, the non linear regression approach, allows
the use of more realistic and generally accepted implicit gravity models
in the estimation of Origin-Destination matrices.
We w i l l discuss here t hree such methods. The f i r s t one by Hogberg
(1976) assumes t hat an all-or-nothing minimum cost assignment i s r eal i s t i c
enough but incorporates a very general deterrence function and up t o
t hree "journey purposes" i n t he gravity type di st r i but i on model. The
second approach by Robillard (1975), uses a generalized gravi t y model
and accepts proportional assignment as r eal i s t i c. The t hi r d model has
been developed by W i l l s (1977) and al so includes a generalized gravi t y
model but of a di fferent kind.
It i s possible t o extend a l l l i near and non-linear models (Sections
2 and 3 i n t h i s work) t o t he proportional assignment case but even then
t hey f a i l t o consider congestion as an important f act or i n rout e choice.
This seems t o i ndi cat e t hat t hese methods could be more appropriate t o
r ur al or uncongested urban areas.
3.2 Hogberg's Model
Hogberg assumes t hat t he j oi nt generation-distribution model i s
of t he form
1 1 1 1 2 2 2 2 2
T i j
=
b 0. A. D. f(c. . .)+ b 0. A. D. f(c; . )+ b O i Ai D. f ( c . .)
1 1 1 J 1 J 2 1 1 J 1J 3 J 13
(3.1)
where
2
oil and 0. ar e di fferent types of t r i p generation parameters
1
2
D . ~ and D. ar e di fferent types of t r i p at t r act i ons parameters
J J
-
1 1
A. and A: ar e balancing fact ors of t he form A.' =/Ti. f(e:. . j "etc.
1
3. -- J J U-1
.b4 2
and f(c:. .) = c- .
exp (b5 (loge(cij) ))
1 J 1J
a
Using t he i ndi cat or vari abl e Pij *we can s t a t e t he flow a t
each l i nk as
-. .
a
* I n t he or i gi nal work Pij could only take values 0 or 1. Here,
it can t ake any value p with 0351.
For a subset of l i nks we can then compare t he observed with t he
modelled flows and minimise t he square of t he difference.
A
a
(va - l. pij Ti j )
2
Hogberg carri ed out a desk study of hi s method on a synt het i c network
with 16 modes and 22 two way l i nks using population and employment as
Ois and D. s . He assumed t he t hr ee elements i n t he di st r i but i on model
J
stood f or home-work, home-home and work-work t r i ps . Hogberg t hen used
an algorithm f or non-linear regression suggested by Marquant
(1963)
and from a synthetic O-D matrix he sampled flows from hal f of t he l i nks.
After introducing an a r t i f i c i a l error component he used t h i s sample t o
obtain a minimum squared difference between observed and modelled
t r a f f i c . He then compared t he flows predicted f or t he r e s t of t he 22
l i nks and found t he model "acceptable".
Hogberg (1975) has carri ed out some t heor et i cal work on t he
"contribution" of each l i nk t o t he accuracy of t he model. He found
t hat aFter t he inclusions of a few l i nks i n t he ~~~~~~~~~~~~~there was
l i t t l e gain i n t he accuracy of t he model. In a personal communication
t o t h i s author Mr . Hogberg has confirmed t hat pr act i cal applications
ar e bei rgcarri ed out i n Sweden with t hi s model but t hey have not yet
been reported i n English.
3.3 Robi l l ardl s Model
Robillard (1975) assumes a generalized gravi t y model leading t o
where Ea i s an error term and
3 , s . ar e t he parameters t o be estimated by t he non-linear
J
regression approach.
f ( C. . ) i s a known cost f'unction.
1 J
It is noted that the estimation of Ri and S. need not be unique; what
J
matters is that their product is. It is always possible to solve this
problem so that the Ri. and S' are positive.
j
Robillard indicates that the method suggested by Lawton and
Sylvester (1971) can be used to solve this model in order to minimise
the difference between observed and synthesized flows. Of course the
number of independent links with traffic counts available should be
greater than the number of the variables R. and S;. In terms of
1 J
Hogberg's previous example it means that at least 33 of the 44 one way
links should be sampled. The problem of minimising
can be solved by replacing the S's by their linear regression solution
in terms of the R's and then solve by the R's only. Of course other
algorithms are also possible but it is claimed that this approach should
reduce computer time. It is interesting to note this method does not
require any information regarding population or other trip generating
parameter. These are all embodied in the R's and S's. The model is,
in effect, only a generalised gravity model. One can compare the
standard doubly constrained gravity model.
T O.A.D.B. f(C.. )
i j = 1 1 ~ ~ IJ
with
Tij
= R.S. f(c.. )
1 J 1J
to find . the equivalences
-
R. = O.A.
1 1 1
-
S. = D.B.
J J J
It must be mentioned that Robillard's method is applicable only
where p. .a is known, that is when there are no congestion effects. In
1J
other words the applications of this model are restricted to cases in
which a proportional assigfient technique is considered realistic.
3.4 W i l l s t Model.
3.4.1. Description An al t ernat i ve non-linear model f or t he estimation
of an 0-D matrix i n i nt er urban cases has been proposed by W i l l s (1977).
H i s s t ar t i ng point i s a generalised gravi t y model of t he form.
where
k t h
A. = The k vari abl e corresponding t o or i gi n i,
1
f or example population a t i.
and b ---b parameters f or cal i brat i on from t r a f f i c counts.
0 k
As usual t hese t r i ps ar e t ransl at ed i nt o modelled l i nk flows as
The parameters b ---\ ar e cal i brat ed so as t o minimise
0
There ar e several algorithms which could be used t o minimise
( 3. 9) . Wills i n f act used a quasi-Newton optimizing method due t o
Fletcher and ot hers.
The main innovation i n W i l l ' s approach i s t he wide range of vari abl es
t hat can be introduced i n t h i s model. He t r i ed, f or example, t hree
groups of vari abl es for an i nt er urban system:
- Socio economic vari abl es l i ke population and employment
- Accessi bi l i t y vari abl es; he used a cent r al i t y index assuming
t hat towns i n t he periphery wi l l have a di fferent propensity
t o generate t r i ps than those near t he "centre".
- Separation, time or distance, variables.
Provision can al so be made f or a non-symmetric version of t he
model. This t akes t he form
replacing equation (3. 6).
3.4.2 Applications Two applications of t hi s model ar e known t o t hi s
author. The f i r s t one by W i l l s t o a road network i n Br i t i s h Columbia
including 37 centroids and only 3 parameters. As t r a f f i c counts f or
several years were avai l abl e it was possible t o compare t he performance
of t he model i n t he forecasting mode.
The second appl i cat i on has been carri ed out by Lamarre (1977)
using W i l l s ' computer programs. This exercise was aimed a t comparing
t he performance of W i l l s ' non-linear model with Lamarne's l i near one.
As reported i n 2.4 Wills' model gave a bet t er f i t .
In both cases t he comparisons were made agai nst observed flows
and not against "real " 0-D matrices. I n both cases t he predi ct i ons
of t he model were found "satisfactory". I n t he Br i t i s h Columbia case
2
an R of around 0.82 was reported. Lamarne report s an R~ of 0.88 for
2
W i l l s ' model against an R of 0.72 for t he l i near approach.
4. NETWORK EQUILIBRIUM APPROACH
4. 1 Introduction
With t he s ol e exception of t he Danish Road Di rect orat e method
al l ot her approaches described so far ar e r es t r i ct ed t o t he proportional
a
assignment case, t ha t i s where pij is independent from
This assumption may be good f or a l arge number of appl i cat i ons,
especi al l y where t he number of al t er nat i ve routes between two points
i s small and where congestion does not gky an important r ol e i n rout e
choice. Most i nt e r urban t ransport networks probably satisfy t hese
conditions.
Sang ~ g y y e i (1977) has extended some of Robi l l ard' s ideas f or t he
equilibrium assignment case which i s more rel evant t o urban areas.
Nguyensuggests two methods. The f i r s t one i s f or t he case i n which
consi st ent t r a f f i c counts ar e avai l abl e f or a l l l i nks. The second
method requi res only cost s f or a l l 0-D pai r s t o be known and does
not expl i ci t l y ask f or t r a f f i c counts a t a l l .
4.2 Nguyen's method, case I.
According t o Nme n (1977) t he f i r s t method can be described as
follows :
. . . .
Let Vd observed flows l i nk a and c . . t he observed cost s of
- IJ
t r avel l i ng between i and j by any used rout e. For each 0-D pai r
A A
only one c i j w i l l exi st since V i s assumed t o be i n Wardrop's
a , .
equilibrium. The cost-flow rel at i onshi p i n any l i nk is Ca(va).
,. A
I f Ti j i s t he t r i p matrix corresponding t o t he V .then t he equilibrium
' i ~
s t a t e i s expressed by
-
Hence t he necessary conditions f or a t r i p matrix LrTij/ t o be i dent i cal
.. _
t o _-Tij-/ ar e
and
..
b) C. . = C. . f or al l 0-D pai rs
1J =J
(4. 3)
where C. . i s t he cost on a l l used routes between i and j when L-T-. 7
1 J 1 J-
is assigned onto t he network according t o Wardrop's f i r s t pri nci pl e.
Nguyen goes on then t o prove t hat t he Ti j matrix s a t i s e i n g (4.2)
and (4.3) can be obtained by solving P1 where
(P1) min F ( v ~ ) = kl Va~a( x) dx o (4.4)
subject t o
( T i - p i = o for dl i j
I
i tpi j>o f or al l p, i j
c. (
I ( T: , . >o for al l i j
I 1 J
and
Where t - ;
=
flow on route p connecting i with j.
Pl J
Using Kuhn and Tucker conditions Nguyen (1977) est abl i shes t he
equivalence between PI and equations (4. 2) and ( 4 . 3 ) . Nguyea then
describes an algorithm t o solve P1.
The problem P1 without t he constraint C2 has t he same st ruct ure
as a network equilibrium problem. Accordingly, t he following algorithm
i s suggested by Nguyen.
Step 1 - Select an i n i t i a l feasi bl e - / - ~ i j - 7,
. .
f or example T i j = K/rjcij where
A *
K = 1 ~ a ( ~ a ) V a . Determine an i n i t i a l flow pat t er n Va using Ti j
a
Step 2 - Determine t he short est route between each 0-D pai r and l e t
C . . be t he t r avel cost on t hi s route.
1J
Step 3 - Find t he 0-D pai r r s for which
Crs/ C = min
rs i , j Ci j / C. . 1:
and load onto t he short est route from r t o s.
i j
"
Let - /-va - 7 be t he resul t i ng t r a f f i c pat t ern.
Step 4 - Test convergence.
t he sol ut i on i s E optimal and stop.
Step 5 - Find an optimum combination between previous flow pat t ern and t he
new one. Determine A, minimising
F(Va + X (;a-~a))
subject t o o<A<1.
Step 6 - Revise t he t r i p matrix and flows as follows
-. .
T. . = T . . -A T. . for a l l i j # r s
IJ 1 J 1J
Return t o st ep 2.
4. 3 Npyen' s : method, Case 11.
The Case I method i s only sui t abl e where it i s possi bl e t o obtain
t r a f f i c counts f or a l l l i nks. This would r e s t r i c t i t s appl i cat i on
t o small networks. I n addition, t he algorithm i s l i kel y t o be slow
f or a l ar ge number of 0-D pai rs. Nguyen then proposed a second method
which reduces t he dat a requirements.
He s t a r t s by proving t ha t t he solution t o t he following problem
P2 al s o i s a sol ut i on t o t he i n i t i a l model.
subject t o
t . . = 0 f or a l l i j
'ij - 6 p l j
(
(
Tp i j > 0 f or al l p, i , j (4.16)
"1
T i j > 0 f or a l l i j
(
..
It can be noted t hat here only t he observed cost s Ci j ar e required
and not t he t r a f f i c counts.
It should al so be noted t hat P2 has t he same form as a formulation
of a t r a f f i c assignment problem with el ast i c demand. By anal om Nguyen
considers t hat t he increase of t he demand function f or an 0-D pai r i j
..
i s constant and equal t o Ci j . He then suggests t ha t an algorithm as t he
one s t at ed i n Ngugen (1976) could be used t o solve t he problem.
We found only one t e s t of both methods on a synt het i c network
with 4 centroids and 18 one way l i nks as i n Figure 4.1. This is a
system with 4 T. t o estimate from 18 equations. The er r or s obtained
l j
were of about 16% f or t he T. . cel l s , and about 3% on l i nk flows. Both
1 J
methods gave an answer i n tliis case with approximately t he same
precision.
0 Node (7 Centroid
Figure 4.1 : NGUYENS TEST PROBLEM NETWORK
It i s i nt er est i ng t o note t hat Nguyeu has only i dent i f i ed t he
necessary conditions t o solve t he problem and not t he suffi ci ent ones.
That i s , t he solutions found with hi s method meet t he equilibrium
conditions but they w i l l not be normally unique sol ut i ons. This issue
w i l l be furt her discussed l a t e r .
5. ENTROPY MAXIMISING APPROAGH
5. 1 Background
The concept of entropy originated i n physics. I n a closed physical
system i t s elements tend t o an arrangement which can be organised i n as
many ways as possible compatible with t he system const rai nt s (energy,
mass, e t c ) . This arrangement i s t he most l i kel y one and al s o t ha t of
t he great est "disorder".
The i dea of entropy i s al so linked t o t hat of "information".
It can be seen i nt ui t i vel y a t l eas t , t hat a s t a t e of maximum disorder,/
i s al so one containing a minimum of information. The pot ent i al information
content of a message grows as t he sequence of symbols depart s from a
purely random (high di sorder) sequence.
The use of t he concept of entropy (or information) has found several
appl i cat i ons i n t ransport , urban and regional systems, see f or example
Wilson (1970). I t s best known application i s i n t he deri vat i on of a
f ul l y constrained gravity model.
The i dea behind t he use of an entropy maximising (or information
minimising) approach t o our problem can be described as follows:
We would l i ke t o fi nd t he most l i kel y ori gi n - dest i nat i on matrix
compatible with t he available s e t of l i nk counts. I n ot her words,
we would l i k e t o "exploit" a l l t he information contained i n t he
observed l i nk flows t o determine t he most l i kel y 0-D matrix compatible
with them.
Wilson and MacGill (1977) have observed t hat t he entropy minimising
method i s par t i cul ar l y sui t ed M.this type of system (systems with l arge
numbers of components with apparent disorganized complexity).
-. ,
It is possible to argue that when one is assuming an underlying
transport demand model (i.e. gravity type) and seeks to calibrate it
from traffic counts one is probably not fully using the information
content of the flows. This is to some extent confirmed by a theoretical
study by Rogberg (1975).
Re studied the contribution of each extra
link count on the accuracy of his non-linear model (for his 16 nodes
44 one way links network). He observed that after including the 4
most important one way links "the gain in precision is very small".
Two very similar entropy maximising models will be discussed
here. The first one put forward by van Zuylen (1977) uses a particulsr
measure of information by Brillouin. Van Zuylen also extended his
solution to the inconsistent flows case.
The second one by the author
of this paper follows a more traditional approach but reaches a very
similar formal solution. This one is also extended for the incomplete
counts case.
5.2 Van Zu.ylents model
5.2.1 Description. Van Zuylen (1977) approaches the problem from
the information minimising view point. He is trying to develop
an algorithm to obtain the trip matrix without losing more
information than necessary.
Among the several information measures discussed by Walsh and
Weber (1977) he chooses Brillouin's definition. This assumes
that an "a priori" solution t
is known. Only a brief summary
ij
of van Zuylen's analysis will be given here.
A more complete
description appears in van Zuylen (1977).
n
If we have a set of observed flows Va the modified Brillouin's
measure of information for each counted . link is given by:
using Sterling's approximation
loge X! - " X log X - X + log=
and neglecting t he constant term
a
I a = 1 TI . p. .
loge Tij
~j 1J l j
Va t i j
f or a l l t he l i nks i n t he uetwo~-k
This has t o be maximised . subject t o t he l i nk counts const rai nt s
-
a
Ga - E j pij T. .
1 J
The formal sol ut i on can be obtained by maximising t he Lagrangean.
where Aa ar e Lagrange mul t i pl i ers corresponding t o t he flow constraints.
According t o van Zuylen di f f er ent i at i ng L t he sol ut i on i s founa
t o be
T i j = tij .
A - ( l + ~ a )
where Xa = Vae
a
and t he Xa can be calculated from t he const rai nt s (5.5)
Formally equations (5.7) and (5.5) ar e a case of t he multiproportional
problems as discussed by Murchland (1977). He has shown t ha t t hi s
model converges t o a unique solution and has developed a simple
algorithm f or it.
5.2.2 Inconsistent constraints. In the inconsistent constraints case
the flow equations can be rearranged as
where &a is an error term for each link. Van ~ G e n then suggests
that the algorithm should now be modified so that the Xa factors
for equation (5.7) must be such that minimize the sum of squares
This is a problem of more difficult solution.
In a recent paper Van Zuylen (1978) suggests that a better approach
would be to remove inconsistencies in the counts first and then
to solve the problem using Murchlandls algorithm. He used a
maximum likelihood method as adapted by Hamerslag and Huisman
(1978) to remove inconsistencies in the observed flows in a
small example. Van Zuylen found then that the quality of the
solution depended a great deal on how good the "a priori"
estimation of the trip matrix L-tij] was. Re concluded his method
would be ideal to update an old Origin-Destination matrix.
5.3
An alternative entropy maximising model.
5.3.1 Description. The author has developed a model similar to that of
Van Zuylen, but which does not require an "a priori" estimation
of the Origin-Destination matrix. As this derivation has not
been published it will be described in some detail here.
According to the usual entrypy assumption, the most probably
arrangement of trips l-~ij-i , is formed by the configuration
with the greatest number of states associated with it. The
number of ways in which individual trips can be assigned to
a particular trip matrix is
w( Ti ] =
-
l l . T. . !
i~ 1~
The most l i kel y t r i p matrix w i l l be t hat maximising
w(Ti j ). Since w(Tij) i s i n variant with a monotonicriLly
increasing transformation, t he problem can be described as
max l ogw( Ti j ) = l og (Jj Tij) ! - , jlog Ti j !
and using St i r l i ng' s approximation
ma*: log W(T ... ) = l og (1 T:. .).. - 1
(T:. . l og T. . . - T.. .)
1J
i j 1~ i , j J 1 J 1~
Since log (1 T . .) ! i s a constant dependent only on t he t o t a l
l j
1J
number of t r i ps , it need not be considered
i n t he maximisation problem.
Then t he problem t o be solved i s t o maximise
l o g e W(T ... ) = - l o g Tf - T...)
1 J IJ e i j
1, J
13
subject t o t he l i nk flow const rai nt s
f or al l counted l i nks a.
This i s easi l y obtained by means of t he Lagrangean
Taking t he deri vat i ves with respect t o T i j we have
6L
-
a
6 Ti j
= - l og T i j - ha p. =
~j
by making e -A& = xa
t he formal model i s found t o be
9.
Ti j = r xa&j
a (5.19)
where again t he Xa ar e calculated f r omt he r es t r i ct i ons
It can be seen t h i s i s van Zuylen's model if we make a l l hi s
I,
a pr i or i " estimations L- t i j - 7 equal t o 1.
5.3.2 Extensions and l i mi t at i ons. It has not been yet possi bl e t o fi nd
an adequate i nt erpret at i on t o t he vari abl es Xa.
They ar e certaimly
associated with t he contribution t he flow on each counted l i nk
makes t o explain each par t i cul ar 0-D pai r .
The conventional gravity model can be derived by maximising
(5.14) subject t o
(5.21) and (5.22) being t he ori gi n and dest i nat i on const rai nt s
and (5.23) t he t o t a l cost const rai nt . I f i n our model we count
t r a f f i c on al l l i nks, including centroid connectors, we ar e
forci ng our sol ut i on t o meet a t l eas t t he const rai nt s (5.21)
and (5.22) as well. In a cordoned urban area t h i s would imply
counting t r a f f i c a t car parks and other major sources and
dest i nat i ons of t r i ps .
If, i n addition, one assumes t hat a l l t ransport cost s ar e
only t r i p end and l i nk cost s (this excludes t hi ngs l i ke modal
penal t i es or parking cost s which depend on t he l engt h of t he
st ay f or example) it can be sai d t hat t he cost const rai nt (5.23)
i s al so met by t he model. But it must be remembered t hat t here
ar e only 2n + 1 const rai nt s for t he gravi t y model, when n i s
t he number of zones. O u r model has as many const rai nt s as
independent counts, each one increasing t he information content
of t he sol ut i on.
When only a par t i al s et of t r a f f i c counts is avai l abl e t he
model i s l i kel y t o underestimate t he t o t a l number of t r i ps .
I f one has an independent estimate of t h i s number t he sol ut i on
can be improved. I n ef f ect , it i s possible t o include a new
const rai nt t o t he model imposing
where T i s t he t o t a l number of t r i ps .
Maximising (5.l)i) subject t o (5.15) and (5.24) l eads t o t he
sol ut i on
where B i s t he Lagrange mul t i pl i er f or const rai nt (5.24).
This may be i nt erpret ed as t he inclusion of an ext r a (not i onal )
l i nk which al l t r i ps use and whose flow equals t he t o t a l number
of t r i ps .
A t t he time of writing t he author i s experimenting with t h i s model
on a synt het i c network on which encouraging r es ul t s have been
obtained. These wi$l.be reported l a t e r .
a
Fi nal l y, it i s worth noting t hat both models assume p. . t o
1 J
be constant thus r es t r i ct i ng t he i r applications t o t he
proportional assignment case.
6. DISCUSSION
6.1 Introduction
The main advantages and disadvantages of each group of models w i l l
be discussed here. We w i l l consider applications i n t hr ee types of
areas
- I nt er urban or r ual areas
- Free standing urban areas
- Sub or inner areas i nsi de an urban area.
We s hal l discuss each group of models i n term but we must admit,
from t he out set , t hat l i t t l e can be sai d about t he r el at i ve accuracy of
t he proposed methods of estimating a t r i p matrix. I n ef f ect , very few
t e s t s have been made with r e a l data and i n t hese cases all comparisons
have been made against observed l i nk flows and not against a "real /
observed" 0-D matrix.
Nevertheless, something can be sai d about t he most promising
models and areas of application from t heor et i cal considerations. I n
par t i cul ar we s ha l l discuss each model i n terms of
- t he conditions under which t he assumptions ar e l i kel y t o be
acceptable
- t he amount of ext ra dat a ( i n addition t o t r a f f i c counts) required
- i nt er nal consistency
- t he i r eventual use f or forecasting purposes and not j ust short
term estimation of a t r i p matrix.
- t he i r f l exi bi l i t y f or incorporating information already
avai l abl e about t r i p making behaviour.
6.2 .Gravity type approaches
This group includes t he models reviewed i n chapters 2 and 3.
It must be sai d t hat most planners ar e familiar with t he idea t hat
t r avel demand may follow a "gravity type" of behaviour. This assumption
is more l i kel y t o be val i d i n l ar ge areas, where t r i p cost or t r i p
length i s an important fact or i n t he formation of a t r avel demand
pat t ern.
Of t hi s group, very general approaches l i k e t he ones suggested
by Hogberg and W i l l s , ar e i n general more t heor et i cal l y sound. On
t he other hand, t hey tend t o requi re more dat a and great er ef f or t i n
cal i brat i on. I f we assume here t hat mispecification er r or s disqualify
t he use of a model f or medium term planning we w i l l have t o accept
t hese ext ra cost s as a reasonable pri ce t o pay i n t h i s case.
The ext ra dat a used i n t he model w i l l depend on t he type of
data readi l y avai l abl e or obtainable with reasonable ef f or t . Robillard' s
model, f or example, passes most speci fi cat i on t e s t s , requi res no
ext ra dat a, but implies a very l ar ge cal i brat i on effort (The minimum
number of parameters t o determine by minimum squares techniques
i s twice t he number of zones plus one).
It i s not by chance t hat most applications of t h i s approach have
been i n i nt e r urban areas. The method seems i deal l y sui t abl e f or
problems where each town represents a si ngl e zone.
Applications t o f r ee standing towns w i l l usually imply a more
det ai l ed zoning system and an independent roadside interview t o
determine "external t r i ps". Moreover, congestion effect s ar e l i kel y t o
be more important i n t h i s case and an all-or-nothing assignment
w i l l not suffi ce.
We do not t hi nk t hi s approach t o be appropriate t o inner c i t y
areas. There seems t o be two reasons f or t hi s . One is t he short
t r i p lengths involved and t he ot her t he f act t ha t probably most of
a t r i p w i l l t ake place outside t he study area. Hence, a gravi t y type
of t r i p making behaviour i s not l i kel y t o be a sensi bl e assumption.
Fi nal l y, t h i s type of model can cope perfect l y well with
inconsistencies i n t he t r a f f i c counts.
6 . 3 Equilibrium assignment approach
We have already mentioned t hat we t hi nk both Nguyen's methods
w i l l determine a t r i p matrix but , with t he exception of pecul i ar
circumstances, t he sol ut i on. offered w i l l not be unique. It w i l l
actually depend on the initial, feasible, solution assumed in each case.
Nevertheless, Nguyen has rightly pointed out the importance of
introducing an allowancd for congestion effects in assignment. We
will discuss in section 6.5 some of the ways in which this might be
achieved.
6.4 Entropy maximising approach
This approach appears as the one of greater generality so far.
In principle it could be applied to the three areas of interest
mentioned before. We have already pointed out, though, that the model
appears sensitive to the total number of trips, a variable usually
not available in most cases. Failing this,counts on links to major trip
productions or attractions are required and these are more likely to
be available in urban areas (car parks, roads representing only links
with major residential areas, etc).
Furthermore, this approach, being independent of planning
parameters like population and car ownership, does not seem suitable
for medium to long term planning applications.
Computationally spe&ing the algorithm for solving this type
of problem appears faster and more convenient than the corresponding
methods for solving non linear regression models. Nevertheless,
inconsistencies in the traffic counts must be removed previous to an
application of the algorithm.
6.5 Extensions to include congestion
We have seen that both gravity type and entropy maximising
models can be solved for the proportional assignment case. There are
qases in which congestion effects are unimportant compared with
differences in driver's perceived route costs.
The Danish Road Directorate has developed their gravity model
so that it can introduce some congestion effects via Smock's iterative
approach. It is possible to extend this type of approach to all the
other methods as well (with the obvious exception of Nguyen's models).
The following generalised algorithm might be used
Step 1.
Assume traffic volumes in the network to be zero
and calculate travel costs for all 0-D pairs.
Step 2. Find the set of minimum cost routes im?licitly
a -
producing a new set of p. .a, L- $. . /. Set n = n+l.
1J 1J =
-
..,, a -
Step 3. Combine this L pij / with the previous one
-
a - n-1 -
L Pij - /
in the following way
Step 4. Estimate an 0-D matrix under proportional assignment
assumptions (using L-p. .a 7) and any selected approach.
13 -
II
Step 5. Assign L-T~ j-T" using /-p. . / and find new route costs.
- 1J -
Step 6. Test for convergence using a suitable convergence
criterion. For example test if 1-c.. f" is sufficiently
-n-1
1 J-
similar to /-c.. / . If algorith has "converged" stop,
- 1J-
otherwise proceed to step 2.
The main problem with this approach is that there is no guarantee
of achieving convergence. Other approaches are possible but their
analysis escape the purpose of this report.
There is also an extra cost associated with any method incorporating
congestion effects. All of them require good flow-cost relationships
for the links. These are not easy to obtain and may involve considerable
ad hoc data collection. There are two other alternative sources:
- The standard DTp speed flow curves which can only cater for
very coarse "types" of links but may provide a simple and
inexpensive allowance for congestion effects. They are
probably quite good for inter urban areas.
- For urban areas most delays occur at intersections and
hence any model giving a good estimate of delays under
different conditions there could be adapted. Typical
examples would be models like TRANSYT or S A W . Bolland
et a1 (1977).
7. References
1. BEBEE, M.A. (1959) Novel t r a f f i c survey method ut i l i s e s
vehi cl e l i ght s . Bulletin of t he Highway Research Board 224.
-
2. BOLLAND, J., HALL, M. , VAN VLIET, D. , WILLUMSEN, L., (1977)
A model for t he simulation of t r a f f i c management schemes
PTRC Summer Annual Meeting, Warwick.
3.
IWERSLAG, R. , and HUISMAN, K .C . , ( 1978) Bir,aire Cal i brat i e ,
Verkeers Kunde 166-168.
4. HOGBERG, P. (1975) Planning of t r af f i c counts. University of
Goteborg, Department of St at i s t i cs Research Report
1975:3
5:
HOGBERG, P. (1976) Estimation of parameters i n models f or
t r a f f i c prediction: a non-linear regression approach.
Transportation Research Vol.10, pp. 263-265.
6. HOLM, J., JENSEN, T:, NIELSEN, S.,CHRISTENSEN, A., JOHNSEN, B.,
RONBY, G (1976) Calibrating t r af f i c models on t r a f f i c census
r es ul t s only. Traffi c Engineering and Control, Vo1.17,
Apri l 1976.
7.
JENSEN , T. , NIELSEN, S. (1973) Calibrating gravi t y models and
estimating t he i r parameters using t r a f f i c volume counts.
5th Conference of Universities Transport Study &oup. London.
8, LAMARRE, L. (1977) Une methode l i neai r e simple d' estimation de
l a matrice des origines e t des dest i nat i ons a pa r t i r de
comptages sur l e s l i ens d'un reseau: une applicat20n
au reseau r out i er du Quebec. Centre de Recherche Sur Les
Transport, Universite de Montreal, Publication No. 76.
g . LOW, D. (1972) A new approach t o t ransport at i on systems modelling.
Traffi c Quart erl y, Vol. 26, 391-404.
10. WJRCELAND, T .P. (1977) The multiproportional problem. London:
University College, London, Transport Studies Group,
Notes JDW63 (unpublished).
11. NGUYEN, S. (1976) Procedures for equilibrium traffic assignment
with elastic demand. Centre de Recherche sur les Transports,
Universite de Montreal, Publication No. 39.
12. NGUYEN, S. (1977) Estimating an 0-D matrix from network data:
a network equilibrium approach. Centre de Recherche sur
les Transports, Universite de Montreal, Publication No. 87.
13. OECD Road Research Group (1974) Urban Traffic Models: Possibilities
for Simplification Paris; OECD.
14. ROBILLARD, P. (1975) Estimating an 0-D matrix from observed link
volumes. Transportation Research 9, 123-128.
15.
SMITH, R., McFARLANE, W. (1978) Examination of a simplified travel
demand model Transportation Engineering Journal of ASCE 104, (TE1) .
16. SMOCK, R. (1962) An iterative assignment approach to capacity
restraint on arterial networks. Highway Research Board Bulletin
NO. 347, 60-66.
17.
SYMONS, J., WILSON, R. and PATERSON, J. (1976) A model of inter-city
motor travel estimated by link volumes. Australian Road Research
Board Proceedings, Vo1.8 Session 20.
18. VAN VLIET, D (1976) Road assignment; parts I, 11 and 111
Transportation Research 10, pp. 137-143.
19. VAN Z m E N , H. J. (1977) Some remarks on the information minimising
method Yerkeersa.kademie, Tilburg, Mimeographed notes, Oct.1973.
20. VAN ZUYLEN, H.J. (1978) A method to estimate a trip matrix from
traffic volume counts PTRC Summer Annual Meeting, University
of Warwick, July 1978. London: PTRC Education & Research
Series Ltd. (unpublished)
21. WALSH, J., WEBBER, M. (1977) Information theory: some concepts
and measures. Environment and Planning, A.9, 395-417.
22. WILLS, M. (1977) Bon linear estimation of origin-destination flows
and gravity model parameters from traffic counts on links.
Department of Geography, University of British Columbia
mimeographed notes.
23. WILSON, A.G. (1970) Entropy in Urban and Regional Modelling
London : Pion.
24. WILSON, A.G., MacGILL, S.M. (1977) A systems analytical framework
for comprehensive urban and regional model building. Leeds:
University of Leeds, School of Geography Working Paper No. 209.
25. WRIGHT, C. OIIROM, H. (1976) The Westminster route choice survey:
a new technique for traffic studies. Traffic Engineering and
Control 17. (819).
Notation.
We have t r i e d t o keep notation uniform throughout t h i s paper.
Nevertheless, some i nevi t abl e but minor departures from t h i s norm w i l l
be apparent t o t he reader.
I n general
Ti j
t r i p s between i and j , where i and j ar e origin-destination
poi nt s, t hat i s zone centroids.*
1-~i j] = The whole t r i p matrix
-
a a l i nk
Va t o t a l flow on l i nk - a
Ca(Va) Cost flow rel at i onshi p on l i nk a
'
Oi
A t r i p end charact eri st i c f or t one i, i n par t i cul ar
P. ,E Part i cul ar population and employment a t i and j.
1 j
'i j
=
cost of t r avel l i ng between i and j
r
i j
=
a rout e or path from i t o j, defined by
( 0 i f l i nk a i s not on r..
ha I-. . ( '-J
1j ( 1 i f l i nk a & on r t hat i s rij uses l i nk a.
i j '
tri j
= number of t r i p s from i t o j on path rij ; i n general T. . = Zt
.
IJ r tij
a
= number of t r i p s from i t o j on path rij mi ng
tri j
l i nk a
a
Pi j
= ztr&ij ITij = proportion of t r i p s between i and j
using l i nk a
i n general 0 9 ." < 1 and t he extreme values are
i j -
taken when t her e i s only one path between i and j .
*
We do not di f f er ent i at e between person and car t r i p s i n t he understanding
t hat a mul t i pl i cat i on by an index corresponding t o an average car occupancy
r a t e is t he only operatism required t o convert one i nt o t he other.
,.
A as in ( ~ a ) is used to represent observed data (flows)
A* as in (rijn) is used to represent an optimum figure
(like optimum path between i and j)
9. APPENDIX 2.
A l e a s t square estimation of t he t r i p f act or f or t he DRD model
9.1 Method
The method i s closely linked t o t he normal di st r i but i on. I n f act ,
f or t he normal di st r i but i on t he maximum l i kel i hood sol ut i on i s found
when t he variance i s minimised.
I f we make E = X - mean it i s possible t o represent t he normal
di st r i but i on by
2 2
1
-E / 2t
f ( ~ ) = -
6&
and dP = f ( ~ ) d ~
For n observations we get t he corresponding j oi nt
pr obabi l i t y density
2
if we make pla = p2< = --- = pn dn2 = 4'
p. represents a scal ar defining t he r el at i onshi p between t he
1
ant i ci pat ed ( a pri ory) variances
then
n
-
taking l og
e
2
L = -n l og 4 - n l og .J27;- - +
e
-. . n i n c
2 2
Where L i s t he l i kel i hood function. The maximum l i kel i hood i s obtained
2
f or 1' p = minimum.
where p i s normally cal l ed "weight".
9.2 Application t o t he DFXJ model
The model estimates t he flow on a l i nk as
A
a = b Xa + sa where
Ill a
Xa = J o . O . . C . . p - wi t hm known
1 J 1 J 1 J
-
i j
ard t he er r or term .sa i s normally di st ri but ed with mean zero and variance
For example y = 1 means t hat mean and variance ar e i n proportion and
y = 2 means t ha t t he standard deviations of t he er r or i s
proportional t o t r a f f i c flow.
y w i l l be chosen independently of t he method.
I n t he absence of bet t er information an i n i t i a l guess can be made
regarding y and t hat improved by t r i a l and er r or i n order t o obt ai n t he
best f i t .
In our case we may define
.I t
and p a - - - &a with a s i f ( ~ , ( b ~ x ~ - ~ ) 2,
Xa
Now, our best unbiased l i near estimate would be t h a t which minimises
cpE2
In our case
I
a-b = 'a
2
we compute (yi-b) pi = C p . ~ .. 2
1
i
1 1
Making t he f i r s t der i vat i ve equal zero.
but our weights were pi- -6
sr
(b%"-A) 4
1
which does coincide with Dm's solution.