1

2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46

SIMULTANEOUS TRAVEL MODEL ESTIMATION

FROM SURVEY DATA AND TRAFFIC COUNTS

Vincent L. Bernardin, Jr., Ph.D.

Senior Consultant, RSG
2709 Washington Ave, Ste. 9
Evansville, IN 47714
Vince.Bernardin@RSGinc.com
Ph: 812-200-2351
Steven Trevino
Analyst, RSG
2709 Washington Ave, Ste. 9
Evansville, IN 47714
Steven.Trevino@RSGinc.com
Ph: 812-200-2351
Greg Slater
Senior Planner, Michiana Area Council of Governments
227 West Jefferson, Blvd., Room 1120
South Bend, IN 46601
gslater@macog.com
Ph: 574-674-8894
John Gliebe, Ph.D.
Senior Consultant, RSG
2200 Wilson Blvd., Suite 205
Arlington, VA 22201
John.Gliebe@RSGinc.com
Ph: 888-774-5986

Submitted: August 1, 2014

Word Count: 6,142 words + 2 tables/figures @ 250 words = 6,642 equivalent words

Bernardin, Trevino, Slater, Gliebe

47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65

ABSTRACT
This paper presents the successful application of a new method to improve travel demand
forecasting models by taking advantage of cheap and readily available traffic count data and
using them together with household travel survey data to inform the models parameter
estimates. Although traffic counts are frequently used in an ad hoc manner in the validation of
travel model components, this paper presents a more rigorous, structured and statistically
efficient method for allowing the information contained in traffic counts to influence the
selection of model parameters simultaneously with household survey data. This formal process
allows traffic counts to inform indirectly but importantly related parameters such as destination
choice utility functions through formal statistical inference where human inference would be
difficult if not impossible due to the complexity of the system and manual random trial and error
would be time and cost prohibitive. The approach used a genetic algorithm metaheuristic to
implement composite log-likelihood and pseudo-composite log-likelihood maximization for the
development of a new travel model for the South Bend, Indiana, urban area for the Michiana
Area Council of Governments. The process made use of and began from a set of parameters
transferred from another region and resulted in new parameters which result in significantly
better consistency of the model with both local survey data and counts. While computationally
intense, this paper demonstrates the feasibility and promise of this exciting new approach, at
least for mid-sized urban areas.

Bernardin, Trevino, Slater, Gliebe

66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111

INTRODUCTION
This paper reports a the successful application of a novel approach to the classic challenge of
calibrating travel demand forecasting model parameters against both household travel survey
data and traffic counts. Traditionally, various components of the demand model are statistically
estimated sequentially from the survey data using a statistical software package; then the
estimated model is implemented in application and the parameters are adjusted in an ad hoc,
judgment based process to improve the fit of the model against the count data.
In the case reported here a genetic algorithm meta-heuristic which calls the model
application code was used to simultaneously statistically estimate nearly all the model
parameters from both the survey data and traffic counts using a composite pseudo-likelihood
maximization framework. This framework harnesses the information available on network
conditions to help understand underlying travel patterns and behavior. The method employed
here extends the typical way of estimating demand model parameters from maximizing the
likelihood of observing limited household survey data by maximizing this likelihood while
simultaneously minimizing the squared error between observed network conditions and those
predicted by the travel model.
The traditional approach makes very poor use of traffic count data and is overly reliant on
household survey data despite the abundance of traffic count data and issues with household
survey data including both possible bias due to significant non-response rates, well-documented
under-reporting of trips or stops and frequently severely limited sample sizes, particularly for
spatial inferences. Although numerous methods have been and are commonly employed to
attempt to address issues with household surveys including targeted oversampling, GPS subsamples for developing under-reporting correction factors, etc., the potential to supplement
survey data with traffic count data has not generally been previously considered in any serious,
structured or rigorous way.
Although in practice, it is quite common for modelers to use judgments to adjust surveybased model components in light of traffic count data on the network (e.g., increasing surveyestimated trip rates in trip generation in response to under-loading on the network versus traffic
counts), these adjustment processes are largely ad hoc, based on very limited exploration of the
models parameter space due to the time required for manual model runs, and generally limited
to fairly simple inferences which humans can easily grasp like the example of global underloading and trip-rate under-reporting.
In contrast, the rigorous statistical method developed here allows for traffic counts to be
used in a systematic way to influence demand model parameters. This permits indirect
inferences difficult for a human analyst to make, such as traffic count data informing trip length
distributions or other aspects of origin-destination flow patterns represented in destination choice
model utilities, etc.
The process made use of and began from a set of parameters transferred from another
region and resulted in new parameters which result in substantially better consistency of the
model with both the local survey data and counts. In this case a large count database (1,536
counts) supplemented a limited household survey sample (681 households). However, it is easy
to see how this same technique may be extensible to a provide a statistically rigorous method for
transferring models to regions with no household survey data, using traffic count data to adjust
demand model parameters to reflect local conditions.
The method employed here may also save some amount of analyst time in model
calibration making manual model runs, allowing more time to be spent on sensitivity testing and

Bernardin, Trevino, Slater, Gliebe

112
113
114
115
116
117
118
119
120
121
122
123
124

dynamic validation of model response properties. That said, this method is not a replacement for
manual calibration efforts which remain necessary for addressing issues with centroid connector
placement, errors in the representation of the roadway network, debugging of model scripts, etc.
This paper presents the data, methodology and basic results and insights from the
application of this method to the estimation of demand model parameters for a new hybrid tourbased model (1) for the Michiana Area Council of Governments (MACOG). It should be noted
that the MACOG case is in many ways ideal for this approach with a large, detailed and high
quality traffic count dataset and limited number of traffic analysis zones (576). While the
method worked well in this case, given its computational intensity even for a problem of this
size, it may not feasible for larger, more complex models without substantial further
improvements to the computational efficiency of the method. Even so, this initial application
opens the door to new, improved methods for incorporating traffic count data in travel demand
model development.

125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156

LITERATURE REVIEW
The estimation of model parameters from survey and count data overlaps and is similar to the
estimation of origin-destination trip matrices based on traffic count data, which has received far
more attention. However, the problem explored here of model parameter estimation is much
broader (matrix estimation from counts alone can be viewed as a special case) and generally a
different problem with its own distinct mathematical characteristics. The problem of travel
demand model parameter estimation from counts (PEC), or more generally, estimating
parameters from traffic counts together with household survey data or other information, has
received limited direct attention in the literature but is related to a significant portion of the
literature on transportation modeling in general.
Only fifteen journal articles have been published on the PEC problem. Other recorded
work on the problem includes two papers presented at conferences and two reports. Seven
distinct groups of papers can be identified. The first is comprised of the work in the 1970s
following Low (2) using least squares estimation techniques with count data exclusively. The
second group includes the work by Jensen, Nielsen and collaborators in the 1970s using
maximum likelihood estimation from count data exclusively. The third group includes the work
of Willumsen and his collaborators in the decade from 1981 to 1991. Fourth is the work of Fisk
and Boyce in the 1980s on the calibration of a combined distribution-assignment model using
traffic count data exclusively and the subsequent extension by Lam and Huang (3). The fifth
group is the work by Cascetta and Russo from the mid-1980s to 2011. Sixth was a lone paper
by Ottomanelli and last and most recently a dissertation by Shrewsbury.
The possibility of estimating the parameters of a travel demand model was first explored
by Low (2). His paper emphasized the cost of extensive surveys and proposed instead using
traffic counts to estimate the parameters of a three step model (regression based trip generation,
trip distribution using a gravity model and an unspecified proportional network assignment) by
means of multiple regression. The emphasis of the paper is on the application of the technique
to the Morgantown, West Virginia, area, where extensive surveys were also conducted and used
in parallel to estimate the model parameters. The fit of the two models (survey based and count
based) to the observed flows was comparable.
Hogberg (4) published a very brief paper which developed a somewhat more formal
theoretical framework for the least squares (regression) estimation proposed by Low and
demonstrated its application on a test network. Symons et al. (5) also reported on the application

Bernardin, Trevino, Slater, Gliebe

157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202

of a similar technique to Lows for an intercity model in Australia. Smith and McFarlane (6)
pointed out the absence of balancing factors in Lows gravity model (it was not constrained to
reproduce the number of trips from each zone estimated in trip generation). They also applied
their constrained version of Lows model in another small American urban area and obtained
results similar to Lows.
Contemporaneously with the development of least-squares estimation, maximum
likelihood techniques were also explored, beginning with an unpublished paper by Jensen and
Nielsen in 1973 which applied the technique to the estimation of a gravity model parameter.
Jensen and Nielsen later collaborated with Holm et al. (7). Together they applied the technique
to estimate a rural regional model for South Zealand, Denmark, and extended the technique to
incorporate the estimation of a trip generation parameter (with a feedback loop of travel times
from a capacity restrained assignment to trip generation and distribution).
Willumsen (8) synthesized the various work on the problem up to that point. He
presented a more developed formulation of the three-step models being assumed and used it to
point out weaknesses in earlier work, such as the lack of balancing factors in Lows gravity
model and the failure to account/control for either long distance through trips or short distance
intra-zonal trips in Holms model. Willumsen also reviewed work on the problem of origindestination matrix estimation from traffic counts and clarified the dual problems of independence
and inconsistency in traffic counts for either matrix or parameter estimation.
Willumsen published again with Tamin (9). Their paper presented three estimation
techniques: non-linear least squares error, weighted non-linear least squares error (in effect
similar to least absolute error), and maximum likelihood. They applied each of the three
techniques to estimate the parameters of several trip distribution models (gravity, intervening
opportunity and a combined gravity opportunity model) for a small English town. The results for
the various estimation techniques (and distribution models) were compared with parameters
estimated from an extensive survey and revealed that non-linear least squares and maximum
likelihood performed comparably and produced good results robustly over the various models,
but the weighted non-linear least squares error performed notably worse.
Ortuzar and Willumsen (10) also published together a broad paper which included some
work on updating model parameters using count data as well as explorations of updating model
parameters using stated preference data and the calibration of disaggregately estimated choice
models to aggregate data. They considered a simultaneous (multinomial) destination-mode
choice model in conjunction with a proportional assignment model. They consider the problem
of updating the choice model parameters with both traffic count and observed mode share data.
They consider both maximum likelihood and generalized least squares approaches and identify
two advantages of the latter: it requires no distributional assumptions on the data and it allows
the possibility of explicitly discounting for differing accuracy of data items prior to estimation.
Fisk and Boyce (11) introduced a different framework for parameter estimation from
count data based on the combined distribution-assignment models of Erlander et al. (12) without
reference to the work of Low, Jensen and Nielsen or Willumsen, presenting the technique as a
method for estimating origin-destination matrices as much as model parameters, but noting the
advantages of estimating a model which can be used for forecasting. Their model represented an
user equilibrium between destination and route choices in marked contrast to both previous and
subsequent work on the problem. Their work also differs significantly in that their approach
does not fall into any of the categories of least squares, maximum likelihood or Bayesian
estimation. However, as the emphasis is on the matrix resulting from the model rather than the

Bernardin, Trevino, Slater, Gliebe

203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248

model and its parameter, the parameter estimation itself is not precisely specified, but rather two
potential general heuristics are presented. In combining distribution and assignment in an
equilibrium framework, the model is superior to any of the other models presented for areas
where congestion is significant. However, the model neglected mode choice and accommodated
only one trip purpose or class of traveler and, like Jensen and Nielsens models, suffered from
the inability to account for intrazonal or external travel observed in counts.
Fisk (13) went on to prove that given the observation of true user equilibrium in the
counts their method of calibrating a combined distribution-assignment model to estimate an
origin-destination matrix was equivalent to two families of models for matrix estimation from
counts, those following Nguyen (14) and those following Van Zuylen and Willumsen (15). Lam
and Huang (1992b) also applied Fisk and Boyces technique to a multi-class combined model
with a gravity parameter for each class and added a random search heuristic in a brute
computational force attempt to improve the parameter estimates.
Cascetta (16) began his work on the problem at the University of Montreal. Cascetta and
Russo collaborated on a conference paper on the subject and eventually published (17). They
considered parameter estimation from count data for a sequential four step model (without
feedback) comprised of regression models of trip generation, a gravity trip distribution model,
multinomial logit mode choice model and probit-based stochastic user equilibrium (SUE)
assignment model. They examined both the direct (simultaneous estimation using survey data
and counts) and indirect (updating / conditioning assumed parameters using counts) estimation of
the parameters of the trip generation, distribution and mode choice models. They examined
possible maximum likelihood, generalized least squares, & Bayesian approaches and applied the
generalized least squares approach to both a test network and to a small Italian city, with
promising results. They asserted the convexity of their objective function based on numerical
experiments and used a solution approach based on first derivatives which are a composite of
analytic and numerical derivatives. Russo later collaborated with Vitetta (18) to extend the least
squares framework to include parameters in the assignment link cost function.
Ottomanelli (19) conducted some numerical experiments using a least squares framework
examining assumptions about the accuracy of both the traffic counts and prior household survey
based parameter estimates.
More recently, Shrewsbury (20) used the generalized linear model variant of least squares
to calibrate gravity models and demonstrated that counts along four screenlines had the same
significance as a thousand household survey.
Nearly all of the papers emphasized the tremendous potential cost savings of using traffic
count data as compared to survey data; many pointed out the particular appropriateness for
smaller areas or developing countries where budgets are not adequate for surveys. Others
suggested the usefulness of the technique for more frequent updates of model parameters or in
transferring parameters from one region to another. Six of the papers (Low, Holm et al., Symons
et al., Smith and McFarlane, Tamin and Willumsen, Cascetta and Russo) included an application
of the technique to real data, and all reported very good results. However, nearly half of the
papers were published before 1980, and most were limited by significantly less computational
resources than those available today. The models employed were therefore somewhat limited in
complexity. Only gravity (and in one case intervening opportunity) models have been studied
and more general logit destination choice models have not been explored. Only the two most
recent papers include any mode choice model and neither model it as both a distinct choice from
destination and conditional on destination choice (a nested destination-mode choice).

Bernardin, Trevino, Slater, Gliebe

249
250
251
252
253
254
255
256
257

Deterministic user equilibrium was only implemented as the network assignment model by Fisk
and Boyce. Hence, mode choice has never been incorporated in the equilibrium with destination
and route choices. Ignoring the non-cooperative equilibrium and/or the multiple information
sources, the problem has always been simplified as a single level optimization problem. Most
papers consider parameter estimation exclusively from traffic count information, and when
information from survey data has been included together with counts, it has always been as an
initial parameter estimate to be updated with the count information. Simultaneous estimation of
parameters using both survey and count data has never been considered and certainly not for
complex models incorporating user equilibrium, mode choice, etc.

258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274

DATA
The Michiana Area Council of Governments (MACOG) began collecting traffic counts in 1976,
and over the years have continued to evolve and improve the program. Today, MACOG collects
counts at over 4,000 stations and traffic monitoring data is a key source of management
information for the regions infrastructure. Information from the collected data is used for
transportation planning, engineering, development of crash rate statistics, estimating mobile
source emissions, and calibrating travel demand models to forecast future traffic. MACOGs
traffic counting staff consists of four, 2-person teams working Monday through Thursday, from
approximately 8:00 AM to 1:00 PM. Clusters of sites are assigned weekly to most efficiently
utilize resources while avoiding the impacts of construction on the flow of traffic. Validation
checks verify the data is within acceptable thresholds based on the circumstances of the location
and a post-processor exports a record of each vehicle counted to a web-accessible database.
Utilizing post-processors, MACOG has been able to automate the capture of a robust set of data
from the traffic counting equipment. These datasets include the classification of vehicles,
estimation of vehicular travel speeds, as well as the volume of traffic passing by the location by
the time of day. All of this in less time and with more accuracy than manually preparing and
recording 24-hour traffic volumes one site at a time.

275
276

FIGURE 1 MACOG Traffic Count Locations in St. Joseph and Elkhart Counties

Bernardin, Trevino, Slater, Gliebe

277
278
279
280
281
282
283
284

The two-county model area included 1,536 count stations. For model estimation, the
counts were processed to produce counts by direction, three vehicle classes (automobiles, single
unit and multi-unit trucks) and three daily time periods (AM peak, PM peak and off peak). The
result were 27,648 observed volumes for comparison with the models network loadings.
MACOG also conducted a small household travel survey in the fall of 2013 with a
sample of 518 households. The results of this survey were combined with Indianas add-on
sample from the National Household Travel Survey, yielding a total sample of 681 households,
1,486 individuals and 5,858 trips.

285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321

METHODOLOGY
MACOG adopted the hybrid trip-based, tour-based framework developed by Bernardin (1) and
first applied to Knoxville, Tennessee, and subsequently, Evansville, Indiana. The model
architecture includes population synthesis, vehicle ownership choice, tour and stop generation,
tour mode choice, destination (stop location and stop sequence) choice models, trip mode choice,
and departure time choice together with truck and external demand models and user equilibrium
assignment. Parameters were estimated simultaneously from MACOGs traffic count and
household survey data for all model components except population synthesis and the truck and
external demand models. The same genetic algorithm was used to estimate the truck model
parameters from MACOGs counts alone. Not considering the truck model, the joint count and
survey dataset was used to estimate 213 parameters.
The parameter estimation problem can generally be formulated as a bilevel program, with
the upper level consisting of likelihood maximization, the minimization of squared errors or a
combination thereof. The lower level is the combined travel forecasting model. This bilevel
program formulation can be interpreted as a Stackelberg (21) leader-follower game between two
groups of decision-makers. This particular problem also involves a simultaneous noncooperative game among the followers. The analogy can be made to a gambling game in which
a crooked referee bets on a particular outcome of a sports game which he will referee and
manipulates the rules (point awards for various goals) of the game to maximize the payoff of his
bet.
The leader, or referee, in this case, is the transportation modeler whose objective is to
make his simulated population of travelers (the Stackelberg followers or athletes) behave as
similarly as possible to his observations of real travelers on the roadway network and from travel
surveys. Each simulated travelers objective is to maximize her own joint utility of her
destination, mode and route choices in the competition among travelers. The modeler decides
how significant each variable is to the simulated travelers utilities and the simulated travelers
react by choosing their destination, mode and route.
The problem was solved in this case using an application-based genetic algorithm. The
code to implement the estimation routine is entirely native in the models application software,
developed in its scripting language, and calls the actual model application code in the estimation
process. In addition to being able to accommodate and incorporate attraction constraints, this
approach also has the practical advantage of greatly reducing the opportunity for inconsistencies
between the estimation and estimation data sets and the final resulting forecasting tool.
The genetic algorithm begins by generating an initial population or set of solutions, each
solution being a complete set of parameters. The algorithm then uses the application code to
apply each candidate solution. Another module evaluates the fitness of each solution, calculating
its pseudo log likelihood by comparing the predicted probabilities from the model application to

Bernardin, Trevino, Slater, Gliebe

322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365

the observed choices from MACOGs survey data and the models network volumes with
observed traffic counts. Once the fitness of all the candidate solutions in the population has been
evaluated, the least fit solutions die or are removed from the population, the best solution is
cloned to start the next generations population and the remaining members of the new
population are created by either mating or randomly recombining two solutions from the parent
population (whose probability of reproducing is a function of their fitness) or by mutating a
single solution from the parent population (again, whose probability of being selected for
mutation is a function of its fitness). The process iterates until no further progress can be made
and a maximum likelihood solution is produced.
This genetic programming approach has significant advantages and disadvantages
compared to traditional analytic gradient-based methods. The key disadvantage is its
computational intensity. Even the incomplete run accomplished in this initial application
required nearly 16 days of CPU time. This is clearly a significant impediment to its widespread
use. However, the version of the algorithm used in this study was not fully multi-threaded, and it
is hoped that better parallel processing may be able to reduce the run time by as much as an order
of magnitude.
Despite the runtime, the genetic programming approach has a number of very attractive
advantages over traditional estimation methods. The method is robust to multiple optima which
are possible for constrained models. It obviates the need for sampling of alternatives common in
destination choice model estimation, simplifying the estimation and improving the statistical
efficiency of the estimator, making better use of limited survey data. It allows for and in fact
benefits from the use of inequality constraints on parameters which can be value in practice to,
for instance, enforce that a parameter has the logical sign. It allows for otherwise non-linear-inparameters utility specifications, such as the estimation of embedded distance decay parameters
in accessibility terms important for controlling for spatial autocorrelation, trip-chaining effects,
etc. (22) Also, as previously noted, it greatly reduces the opportunity for inconsistencies
between estimation and ultimate application of the model for forecasting.
The unique challenge of this problem is the specification the fitness function for the
whole travel demand model compared to both the survey and count data. Initially, a true
composite log-likelihood approach was attempted. However, ultimately, due to the difficulty in
finding the proper distributional assumption for network loading errors, a pragmatic shift was
made to a pseudo-composite log-likelihood combining the log-likelihood of the demand models
against the survey data with the inverse of the squared error between the network assignment and
the traffic counts.
In both cases, the overall fitness function or grand likelihood in the original specification
was comprised of three terms.

Here LLGen denotes the log-likelihood of tour and stop generation versus the combined survey
data; and LLDist denotes the log-likelihood of the distribution of trips between locations and
across modes and times of day. FitNet was initially the log-likelihood of assigned volumes
versus counts, but was ultimately a negative factor of the squared error of assigned volumes
versus counts, scaled approximately to the original log-likelihood of assigned volumes versus
counts.

Bernardin, Trevino, Slater, Gliebe

366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390

10

One of the key challenges in constructing this fitness function is the handling of the units
of observation which are not consistent between the two data sources. In this case, the original
weights on the trips observed from the survey were factored to convert observed trips into an
equivalent number of likely traffic count observations. This was done by skimming the path
along the network between observed trips, counting the number of links along the path and
multiplying the number of links in the path by the fraction of links in the network with traffic
counts. This factor was then multiplied by the sample weight from the survey to produce new
estimation weights for trips. Non-motorized and transit trips were weighted using the auto paths
over the network consistent with and as though they were auto trips. Intrazonal trips were
treated as though they consisted of a single network link, i.e., used the fraction of links in the
network with traffic counts.
Other schemes for handling the units of observation are clearly possible. Considerable
thought was given to using distance-based as opposed to link-based weighting and further work
might consider the comparison of various approaches. In this preliminary case, however, it was
decided that the segmentation of the network into links was significant, with short links in
downtown areas and long links on freeways and in rural areas, such that the likelihood of a trip
being observed in a traffic count was more likely a function of the number of links in the path
than the length of the path with the thought that a length-based approach might over-weight rural
counts and under-weight urban ones. The actual number of count stations along the path could
also be used, but this was avoided as it was thought to be more an artifact of the path finding
algorithm and parameters than the number of links or length.
With this definition of the estimation weights for observed trips, Wobs, the likelihood
function for generation was specified assuming Poisson distributed errors, owing largely to their
simplicity:

391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407

Here k denotes the observed (unweighted) number of tours or stops, and denotes the predicted
tour or stop rate (per household) from the generation model. The relative size of LLGen and LLDist
suggested that assumption of Poisson distributed errors may be incorrect, and that the variance is
likely greater. Future efforts might consider the use of negative binomial errors or other more
flexible distributions. However, for this study, LLGen was simply scaled to account for this
misspecification.
The distributional assumption for the distribution of trips between zones and across
modes and times of day is implicit in the specification of these demand models. In the case of
MACOGs hybrid tour-based model, these component demand models are logit discrete choice
models. However, owing to the inclusions of attraction constraints on work and school location
choice models as well as the inclusion of accessibility variables in the destination choice utilities,
these models are not strictly generalized extreme value (GEV), although the implied distribution
may well be a near variant thereof.
The log-likelihood for the multi-dimensional distribution can be calculated simply using
the probability distribution implied by the trip tables produced by the model:

Bernardin, Trevino, Slater, Gliebe

11

408
409
410
411
412
413
414
415
416
417
418
419
420
421

Here, again Wobs is the estimation weight described above, and Pmod is the probability of a trip
between two zones by a particular mode during a particular time of day. This probability can be
obtained by simply normalizing the models trip tables by their grand sum. In the case of
MACOGs hybrid model, tour mode choice occurs prior to stop location choice and destinations
and time of day are not modeled for non-automobile modes. For these modes, the model simply
produces a vector of trips by residence zone and so vector probabilities are calculated. Given the
models three modes, three times of day and three tour types, there are would be 27 matrices, but
since non-auto modes are only partially modeled, there are only 12 matrices and 3 vectors. Since
there are 576 zones in the MACOG model, this amounts to 2,987,136 discrete probabilities.
Originally, the network fitness was to be modeled as the log-likelihood of network
loadings versus the traffic counts through the assumption of an error distribution, similar to nonlinear regression methods.

422
423
424
425

Here fDist would be the error distribution. Initially, the errors were assumed to be normally
distributed.
(

426
427
428
429
430
431

Where x denotes the modeled traffic volume, denotes the observed volume from the count and
is the assumed standard deviation. However, assuming a normal distribution made the
probability for low volume counts go to zero and dominate the log likelihood. Therefore, the
distributional assumption was shifted to assume log-normally distributed errors.
(

432
433
434
435
436
437
438
439
440
441

Here is the log-normal scale parameter which was approximated by assuming that the mean
approaches the median. This significantly improved over the normal distribution. However,
problems still remained. Tracking both LLNet and the squared error, it became evident that
although lower squared error always corresponded to a higher LLNet, a higher LLNet did not
always correspond to a lower squared error, and these cases proved problematic. Given the
constraints of the project schedule, finding the appropriate distributional assumption for the
network errors was left for future research and pseudo-composite likelihood function was created
by replacing the LLNet term with the squared error of between the observed and modeled
volumes, scaled to the order of magnitude suggested by the log-normal distribution.

Bernardin, Trevino, Slater, Gliebe

12

442

(
443
444
445
446
447
448
449
450
451
452
453
454

Although the squared error was ultimately used in the fitness function for estimation, the true
composite log-likelihood (assuming log-normal network errors) was still calculated and tracked.
RESULTS
The genetic algorithm was run for over 1,500 iterations. It was stopped not because it had
clearly converged upon a final and globally optimal solution but based on the requirements of the
project schedule. On the one hand, it seemed clear that algorithm was making and could have
made further progress resulting in a better fit model. On the other hand, its progress had clearly
slowed compared to earlier in the run, indicating diminishing returns. This initial application is
therefore only a preliminary indication of what this approach could accomplish given more
computing time, improvements to the computational efficiency of the algorithm or both.

455
456

FIGURE 2 Pseudo-Log-likelihoods from Genetic Algorithm

457
458
459

After 1,500 iterations, the pseudo-composite log-likelihood had improved by 5.5%. The
actual estimate of the log-likelihood assuming Poisson generation errors and log-normal network
errors improved 1.4%. LLGen only marginally improved by 0.2%. LLDist improved a modest

Bernardin, Trevino, Slater, Gliebe

13

460
461
462
463
464

2.6%. LLNet improved only 1.7%, but the %RMSE had improved by 8.1% relative to its starting
value (from over 34% to almost 31%). While these improvements are perhaps modest, they are
not insignificant either, and represent a resulting model that conforms better to both the available
survey and count data rather than compromising the original fit of the model against the survey
to better fit the counts as often happens in ad hoc calibration.

465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491

CONCLUSIONS
This research makes a number of important contributions that are both theoretical and practical.
On the theoretical side, the main contribution of this research is the incorporation of information
from traffic counts into the estimation of travel demand models considerably more complex than
have been previous considered in the literature and the first application of a genetic programming
metaheuristic to this problem.
On the practical side, the approach presented here allows a more efficient estimate of the
model parameters by incorporating additional data. Given the limited data with which travel
models are often currently estimated in practice, this by itself is not insignificant. Moreover, the
traffic count data that it incorporates are in most cases already available, and in any event, much
cheaper to collect than household survey observations. Hence, one of the great advantages of
this approach is that it promises to improve travel models without (or only marginally) increasing
the cost of data collection to support them.
Further, traffic count data could be used through this technique not only in initial model
estimation, but to update or transfer models. Annual traffic count programs could, for instance,
be used to check and update the travel model each year, or household survey data could be
borrowed from one community and local count data used to adjust and fit it for a new (similar)
areas model. The technique offers a formal and automated approach to estimating model
parameters like special trip generation factors, survey under-reporting correction factors, intrazonal trip shares and other parameters that currently must be developed through labor intensive
trial and error.
Although improvements in computational efficiency are necessary to make extensive
application of this technique feasible, this initial application promises a real and significant
improvement in travel modeling. The approach developed here is not only more rigorous, and
statistically efficient. With economies of scale it will also likely become cheaper to apply. The
result is to open the way forward for travel modeling that is better grounded in theory and data, is
less costly and more accurate.

492
493
494
495
496
497
498
499
500
501
502
503

REFERENCES
1. Bernardin, V. L., and M. Conger. From Academia to Application. Transportation Research Record:
Journal of the Transportation Research Board. No 2176, 2010, pp. 50-58.
2. Low, Dana E. New Approach to Transportation Systems Modelling. Traffic Quarterly, Vol. 26, No. 3
1972.
3. Lam, William HK, and Hai-Jun Huang. Calibration of the combined trip distribution and assignment
model for multiple user classes. Transportation Research Part B: Methodological, Vol. 26, No. 4,
1992, pp. 289-305.
4. Hgberg, P. Estimation of parameters in models for traffic prediction: a non-linear regression
approach. Transportation Research, Vol. 10, No. 4, 1976, pp. 263-265.
5. Symons, J., R. Wilson, and J. Patterson. A Model of Inter-city Motor Travel Estimated by Link
Volumes. Australian Road Research Board (ARRB) Procedings. Vol. 8, No. 20, 1976, pp. 53-65.

Bernardin, Trevino, Slater, Gliebe

504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541

14

6. Smith, Robert L., and William J. McFarlane. Examination of Simplified Travel Demand Model.
Transportation Engineering Journal, Vol. 104, No. 1, 1978, pp. 31-41.
7. Holm, J., T. Jensen, S. K. Nielsen, A. Christensen, B. Johnsen, and G. Ronby. Calibrating traffic
models on traffic census results only. Traffic Engineering & Control, Vol. 17, No. 4, 1976.
8. Willumsen, L. G. Simplified transport models based on traffic counts. Transportation, Vol. 10, No. 3,
1981, pp. 257-278.
9. Tamin, O. Z., and L. G. Willumsen. Transport demand model estimation from traffic counts.
Transportation, Vol. 16, No. 1, 1989, pp. 3-26.
10. Ortzar, Juan De Dios, and Luis G. Willumsen. Flexible long range planning using low cost
information. Transportation, Vol. 18, No. 2, 1991, pp. 151-173.
11. Fisk, Caroline S., and David E. Boyce. A note on trip matrix estimation from link traffic count data.
Transportation Research Part B: Methodological, Vol. 17, No. 3, 1983, pp. 245-250.
12. Erlander, Sven, Sang Nguyen, and Neil Frederick Stewart. On the calibration of the combined
distribution-assignment model. Transportation Research Part B: Methodological. Vol. 13, No. 3,
1979, pp. 259-267.
13. Fisk, C. S. On combining maximum entropy trip matrix estimation with user optimal assignment.
Transportation Research Part B: Methodological, Vol. 22, No. 1, 1988pp. 69-73.
14. Nguyen, S. Estimating and OD Matrix from Network Data: a Network Equilibrium Approach.
Montral: Universit de Montral, Centre de recherche sur les transports, 1977.
15. Van Zuylen, Henk J., and Luis G. Willumsen. The most likely trip matrix estimated from traffic
counts. Transportation Research Part B: Methodological, Vol. 14, No. 3, 1980, pp. 281-293.
16. Cascetta, Ennio. Class of Travel Demand Estimators using Traffic Flows." CRT Publication No. 375.
Universit de Montreal, Montreal, Canada (1986).
17. Cascetta, Ennio, and Francesco Russo. Calibrating aggregate travel demand models with traffic
counts: Estimators and statistical performance. Transportation, Vol. 24, No. 3, 1997, pp. 271-293.

18. Russo, Francesco, and Antonino Vitetta. Reverse assignment: calibrating link cost functions
and updating demand from traffic counts and time measurements. Inverse Problems in
Science and Engineering, Vol. 19, No. 7, 2011, pp. 921-950.
19. Ottomanelli, Michele. Effects of Data Accuracy in Aggregate Travel Demand Models Calibration
with Traffic Counts. In Mathematical Methods on Optimization in Transportation Systems, pp. 225247. Springer US, 2001.
20. Shrewsbury, John Stephen. Calibration of trip distribution by generalised linear models. 2012.
21. von Stackelberg, H. The Theory of the Market Economy. Oxford University Press, New York, New
York, 1952.
22. Bernardin, Vincent L., Frank Koppelman, and David Boyce. Enhanced destination choice models
incorporating agglomeration related to trip chaining while controlling for spatial competition.
Transportation Research Record: Journal of the Transportation Research Board, No. 2132, 2009, pp.
143-151.