Lecture 4

Trip Distribution Modeling

Purpose:
Develop a procedure that synthesizes the trip linkages
between traffic zones. Or a distribution model tries to
estimate the number of trips in each of the matrix cells
on the basis of any information available.
Content:

 Definition and Notation

 Growth Factor Model
 Gravity Models
 Entropy-maximizing Approach
---Z
 Definition and Notation

Inter-zonal trips and Intra-zonal trips.

Inter-zonal: tij , Tij , i  j

Intra-zonal: t jj , T jj , j  1,2,..., J

Base year trip matrix and horizon year trip matrix

Base-year: tij
Horizon year: Tij
 Dis-aggregation of O-D Matrices

 The matrices can be further disaggregated by person type (n)

(e.g., income level, age level) and/or by mode (k):

Tijkn : trips from i to j by mode k and person type n;

Oikn : total number of trips originating at zone i by mode k

and person type n, and so on.

 Summation over superscripts:

Tijn   Tijkn T   Tij t   tij

k ij ij
 Dis-aggregation of O-D Matrices

 The proportion of trips using a particular mode and the cost of traveling
between two points:

pijk : the proportion of trips from i to j by mode k;

cijk : the cost (time & money) of traveling between i and j by mode k.

 Trip production and attraction conditions:

Tj
ij  Oi T
i
ij  Dj

 Doubly constrained model: with information about both constraints

 Singly constrained model: with information only about one of these

constraints.
 Generalized Cost of Travel
One possible representation of generalized cost of travel for
mode k is (omit superscript “k” for simplicity):

Cij  a1tijv  a2tijw  a3tijt  a4tij  a5 Fij  a6  j  

where
t ijv : in-vehicle travel time between i and j;

t ijw : walking time to and from stops (stations);

tijt : waiting time at stops (stations);

t ij : interchange time, if any;
Fij : fare charged to travel between i and j;
 Generalized Cost of Travel

Cij  a1tijv  a2tijw  a3tijt  a4tij  a5 Fij  a6 j  

j: terminal (typically parking) cost associated with the
journey from i to j;
: modal penalty, a parameter representing all other
attributes not included in the generalized measure so
far, e.g. safety, comfort and convenience;
a1 ,..., a6 : weights attached to each element of cost; they have
dimensions appropriate for conversion of all attributes
to common units, e.g. money or time.
 Generalized Cost of Travel

Cij  a1tijv  a2 tijw  a3tijt  a4 tij  a5 Fij  a6  j  

 If the generalized cost is measured in money units then a1

is sometimes interpreted as the value of time (or more
precisely the value of in-vehicle time) as its units are
money/time.

 As generalized costs may be measured in money or time

units it is relatively easy to convert one into the other.
 Growth-Factor Methods

Uniform growth factor

Tij  tij for each pair i and j

T
Of course   , i.e. the rate of growth (ratio of
t
expansion) over previous total number of trips.
 Singly Constrained Growth-Factor Methods

Suppose information is available on the expected growth

in trips originating in (or attracted to) each zone.

Tij  i tij for origin-specific factors

Tij   j tij for destination-specific factors

Uniform Growth Factor: An Example:
1

2
 Example 1. The singly constrained
growth-factor methods

Table 3. Origin-constrained growth trip table

Using the singly constrained growth-factor methods, the
problem can be solved immediately by multiplying each row by
the ratio of target Oi over the base year total   , thus giving
j

the results in the following Table 4

Table 4. Expanded origin-constrained growth trip table

Total attraction
 Doubly Constrained Growth Factors

 When information is available on the future number of

trips originating and terminating in each zone, different
growth rates for trips in and out of each zone can be
obtained and consequently having two sets of growth
factors for each zone, say, i and  j .

 The application of an ‘average’ growth factor, say

Fij  0.5  i   j 
For a particular cell,
it can be used as an
initial approximation
is only a poor compromise, as none of these two targets or
trip-end constraints will be satisfied.
 Iterative balancing method

Introducing ‘balancing factors’ Ai and B j as follows:

Tij  tij i  j Ai B j

or incorporating the growth rates into new variables ai and b j

Tij  tij ai b j

with ai  i Ai and b j   j B j .

The factors ai and b j or Ai and B j must be calculated so that

the trip end constraints are satisfied.
 Bi-proportional Algorithm

1. set all b j  1.0 and solve for ai ; in this context,

‘solve for ai ’ means ‘find the correction factors ai
that satisfy the trip generation constraints’;

2. with the latest ai , solve for b j , e.g. satisfy the trip

attraction constraints;

3. keeping the b j ’s fixed, solve for ai and repeat Step

2 and 3 until the changes are sufficiently small.
  Oi
 ai 

Horizon year trip
 Tij  Oi  tij aibj  Oi tijbj
end constraints
 j  j 
 j
 or   
Tij  Dj tij aibj  Dj  Dj
i i bj  t a
 i ij i
Step 0: b j  1.0 ; k=1
(1)

k  Oi
Step 1: ai  ;
 tij b j
k 
j
k: number of iteration
 k  1 Dj
Step 2: bj 
t ai
ij
k 
i

Step 3: If a convergence criterion is met, then Stop

Otherwise, let k : k  1 and go to Step 1
t a b
j
ij i j  Oi ; t a b
i
ij i j  D j ; i  1, 2,..., I ; j  1, 2,..., J

Note that ai and b j are not unique

O   D
i j
I  J variables i j

I  J  1 independent equations
I: number of origin zones;
J: Number of destination zones

But ai b j is unique, so Tij is unique

 Bi-proportional Algorithm

 The bi-proportional method tries to produce the

minimum corrections to the base-year matrix tij 
necessary to satisfy the future year trip-end
constraints.

 Condition for convergence:

 t   t
i
i
j
ij
j
j
i
ij T
Example 2. Doubly constrained growth factors
The following table represents a doubly constrained growth
factor problem:
Table 5. Doubly constrained matrix expansion problem
 The solution to this problem, after three iterations on rows and
columns (three sets of corrections for all rows and three for all
columns), can be shown to be:
Table 6. Solution to the doubly constrained matrix expansion problem

Note this estimated matrix is within 1% of meeting the target trip ends,
more than enough accuracy for this problem.
 Advantages and Limitations of
Growth-Factor Methods

Advantages:
 simple to understand and make direct use of observed trip
matrix and forecasts of trip-end growth.

Limitations:
 Any error in the base-year may well be amplified by the
application of successive correction factors.
 The methods do not take into account changes in transport
costs due to improvements (or new congestion) in the
network.
 Additional Limitation

When there is change in land-use pattern, for

example, when there is a new trip generation or
attraction zone, the growth factor model is unable
to provide a result for the new zone
 The Gravity Distribution Model

Model structure:

Magnitude of trip-interchanges between two zones i-j is

directly proportional to the number of trips produced in zone i,
trips attracted to zone j, and inversely proportional to some
function of spatial separation between the zones.
Pi Aj
Tij 
dij 
Pi : Trips produced at i,
Aj : Trips attracted to j.
dij  : Separation function
 General Forms of Gravity Model

Tij  Oi D j f  cij 

f  cij  is a generalized function of the travel costs with one or more

parameters for calibration. Popular versions for this function:

f (cij )  exp  cij  exponential function

f (cij )  cij  n power function

f (cij )  cijn exp  cij  combined function

Figure 1. Different deterrence functions
Figure 2. Typical trip length distribution in urban areas
Gravity models can be:

 Singly constrained

 Doubly constrained

 Unconstrained.
 Singly and Doubly Constrained Models

 To ensure the trip-end constraints are met, two sets of

balancing factors Ai and B j are introduced to replace the
single proportionality factor , yielding:

Tij  AO
i i B j D j f ( cij )

 Singly constrained version, either origin or destination

constrained, can be produced by making one set of
balancing factors Ai or B j equal to one.

Bj=1.0 for all j if origin only constrained!

 Singly and Doubly Constrained Models (cont’d)

 For an origin-constrained model, B j =1.0 for all j, and

1 i i B j D j f  cij 
Tij  AO
Ai 
 D j f cij 
j

For a doubly constrained model the values of the balancing factors are:

1 1
Ai  , Bj  WHY?
 B j D j f cij 
j
i i f  cij 
 AO
i

 The balancing factors are interdependent. An iterative process is applied:

Given set of values for the deterrence function f  cij  , start with B j =1.0 for all
j, solve for Ai and then use these values to re-estimate the B j ’s; repeat until
convergence is achieved.
 Entropy-Maximizing Approach

 Micro states of a system: identifying each individual

traveler, its origin, destination, mode and so on.
Tijk ,m : person m
 Meso-state (more aggregate): a meso state may just specify
the number of trips between each origin and each destination
(O-D distribution).
Tij

 Macro state (even higher level of aggregation): the total

number of trips on particular links, or the total trips
generated and attracted to each zone.
Oi , D j
 Entropy-Maximizing Approach

 For many practical purposes, it may be sufficient to

work on the basis of meso-state specification (e.g.
O-D distribution matrix.
Tij
 Estimates about the future are usually restricted to
macro-state descriptions because it is easier to make
observations at this higher macro level of
aggregation.
Oi , D j
 Basic Consideration of Entropy-Maximizing
Approach

 Entropy method is to accept that all micro states consistent

with our information about macro states are equally likely to
occur.
 A good way of enforcing consistency with our knowledge
about macro states is to express our information as equality
constraints in the system.
 As we are interested in the meso-state descriptions of the
system (O-D trip matrix), we would like to identify those
meso states which are most likely, given our constraints
about macro states.
 TOA  TOB  4 O  4 is given 

Assuming all micro states The most probably meso state

are equally likely, with a (generated with the greatest
probability 1/16 number of ways) or the most
(Total: 16 micro-states) probably O-D distribution
 Mathematical Expressions of Entropy-
Maximizing Approach

 The number of micro states W Tij  associated with the

meso state Tij is given by:

 
W Tij  
T!
 T   Tij 
 ijTij !  ij 

 As it is assumed that all micro states are equally likely,

the most probably meso state would be the one that can
be generated in a greater number of ways. Therefore,
what is needed is a technique to identify the values of
Tij  (O-D distribution), which maximize W Tij .
Given T   Tij , consider OD pairs ij , ij, ij,.... .
ij

The ways of allocating Tij , Tij , Tij  ,... are given by

W Tij   CTTij  CTij ij  CTij ij

T T T T Ti j 
 


T!

 T  Tij !

 T  Tij  Tij !
 
T  Tij !Tij ! T  Tij  Tij !Tij ! T  Tij  Tij  Tij !Tij !
T! T!
 
Tij ! Tij ! Tij !   ijTij !
 Mathematical Expressions of Entropy-
Maximizing Approach

 Because W and ln(W) have the same maximum. Thus

 T! 
ln W  ln W {Tij }  ln 
  ijTij ! 
 ln T !   ln T !
ij
  ij

With stirling’s approximation for ln X !  X ln X  X , to make it

easier to optimize:

ln W  ln T !  T ij
ij ln Tij  Tij 

 Omitting the constant term log T !, the rest of the equation is often
referred to as the Entropy function:

ln W *   T
ij
ij ln Tij  Tij 
Mathematical Expressions of Entropy-
Maximizing Approach

The most likely matrix Tij  can thus be obtained:

maximize ln W *  
Tij 
T
ij
ij ln Tij  Tij 

subject to
T j
ij  Oi (if Oi is known)

and/or
T i
ij  D j (if D j is known)
 Entropy-Maximizing Approach
 When additional information in the form of prior or old values for the
meso states, for example an outdated trip matrix tij , the problem
may be recast with this information and the revised objective function
becomes:

  Tij  
ln W    Tij ln    Tij  tij 
**
 tij 
ij    

 If Tij  tij then ln W **  0 .

 It can be further shown that

1 Tij  tij 
2

 ln W  **

2 ij tij

Therefore,  lnW ** is a good measure of the difference between Tij 

and tij .
  Tij  x
Consider Tij ln    Tij  tij and let f  x   x ln    x  tij
t  t 
 ij   ij 
When Tij  tij for all  i , j  , then, from Taylor’s expansion formula
1
f  x   f  x0   f   x0  x  x0   f   x0  x  x0 
2

2!
 tij 
Let x0  tij ; x  Tij , we have f  x0   f  tij   tij ln    tij  tij  0
t 
 ij 
x  tij 
f   x0   f   tij   ln    x    1  ln    1  1  0
tij 1
t  x t t 
 
ij ij  ij 

f   x0   f   tij     
tij 1 1 1
x tij x tij

1  Tij  tij   Tij  tij 

2 2

Therefore, f  Tij  
1
or  ln W  
**

2 tij 2 ij tij
(like Weighted Least Squares Method)
Approximation of the Entropy-Maximizing Model with prior
(or outdated) O-D matrix distribution information

1  Tij  tij 
2

min 
Tij  2 ij tij
subject to
T
j
ij  Oi (if Oi is known)

and/or
T
i
ij  D j (if D j is known)