Second Workshop on Applications of Discrete Choice Models

August 31 - September 1, 2006

Room CM 113 Ecole Polytechnique Fédérale de Lausanne

organized by

Michel Bierlaire

After the success of the 2005 workshop, the 2006 workshop will be organized in the same spirit: the exhange of ideas around discrete choice models, with the objective to trigger new collaborations, or strenghten existing ones. At the end of the workshop, a list of potential collaborations will be identified, with specific objectives.


14:00 - 14:10 : Michel Bierlaire Welcome
14:10 - 14:50 : Stefan Mabit Representation of taste heterogeneity in willingness-to-pay indicators
14:50 - 15:30 : Carlo Prato Evaluation of choice set generation methods for route choice modelling 
15:30 - 16:10 : Matteo Sorci Discrete Choice Models for Static Facial Expression Recognition
16:10 - 16:40 : Break
16:40 - 17:20 : Mogens Fosgerau A practical test for the choice of mixing distribution in discrete choice models
17:20 - 18:00 : Enide Bogers Modeling day-to-day learning in route choice
18:00 - 18:40 : Nadine Schuessler Capturing similarities in route, mode and destination choice problems
19:30 : Dinner
09:00 - 9:40 : Michel Bierlaire A Latent Route Choice Model in Switzerland
9:40 - 10:20 : John Polak Towards a unification of expected utility theory and random utility theory
10:20 - 10:50 : Coffree break
10:50 - 11:30 : Otto Nielsen Multi-modal route choice models - utility functions and choice set generation
11:30 - 12:10 : Amaya Vega Residential location and travel behaviour in the Greater Dublin Area
12:10 - 14:00 : Lunch
14:00 : Workshop meeting
09:00 : Meeting in Lausanne
09:30 : Meeting in Orbe
10:00 - 12:00 : Hiking
12:00 : Genuine swiss fondue in a chalet up in the mountain


There is no registration fee. Everyone interested is invited to attend. Presentations are upon invitation only.

All participants, including speakers, must register by sending an Email to Michel Bierlaire



Because of the Triathlon World Championship, you may have difficulties to find available rooms in Lausanne. Look below for other possibilities.

EPFL-rates apply to the following hotels:

Swiss Youth Hotels
Bois-de-Vaux 36, 1007 Lausanne
tél: +41 21 626.02.22
fax: +41 21 626.02.26
adresse email:
single: 78.00

Hôtel Elite
Avenue Ste-Luce 1
1003 Lausanne
tél: +41 21 320 23 61
fax: + 41 21 320 39 63
single: 117.00; double: 174.00

Hôtel Alagare
Minotel Suisse
Rue du Simplon 14
1006 Lausanne
tél: +41 21 617 92 52
fax: +41 21 617 92 55
single 105.00; double: 150.00

Hôtel Alpha-Palmiers
Fassbind Hotels
Rue du Petit.Chêne 34
1003 Lausanne
tél: +41 21 555 59 99
fax: +41 21 555 59 98
single: 158.00; double: 220.00

EPFL area

Hotel Pré-Fleuri***, Rue du Centre 1, 1025 St-Sulpice.
Tél. 021  691 20 21
Fax 021  691 20 20
Price for a single room  around CHF 150.- 

Motel des Pierrettes**, St-Sulpice, 10 minutes walk to EPFL, Route cantonale 19, 1025 St-Sulpice
It has no web-site but you can call at +41 21 691 25 25. It has no restaurant. 
Price for a single room, around CHF 110.- (special price for EPFL hosts)

Hostellerie du Débarcadère, Chemin du Crêt 7, 1025 St-Sulpice, 
It belongs to "Relais& Châteaux" and its web-site is:
Price for a single room around CHF 170.- (special price for EPFL hosts)

If you come by car:
Novotel Lausanne Bussigny, 35, Route de Condémines, 1030 Bussigny
(15 minutes by car, no bus possibilities) 
Price for a single room, around CHF 200.- (special price for EPFL hosts)


The easiest way to get to EPFL is to take the train from Geneva Airport to Renens. In Renens, take the light-rail (called TSOL) towards Lausanne. There is a stop at EPFL. The travel time is about 1 hour.

Check the Swiss Federal Railways website.

To navigate within EPFL, use The room where the workshop will take place is CM 113.

UpList of presentations

A Latent Route Choice Model in Switzerland
M. Bierlaire, EPFL, Switzerland (with E. Frejinger and J. Stojanovic)

Modeling travelers' route choice behavior is important in many contexts (e.g. intelligent transport systems and transportation planning) but presents several difficulties, like the large size of the choice set, and the structural correlation among alternatives. Various authors, including Cascetta et al. (1996), Ben-Akiva and Bierlaire (1999), Vovsha and Bekhor (1998) and Frejinger and Bierlaire (2006) have proposed several solutions to handle these difficulties.

In this paper we present a route choice modeling approach with latent chosen routes which can be combined with the models proposed in the above mentioned literature. A latent chosen route is such that an exact description is not available. Instead, travelers describe their choice in terms of a sequence of locations and cities that they have traversed, without the need to relate the actual network used by the analyst. On the one hand, this technique improves the quality of the responses provided by the interviewees. Indeed, providing an accurate description of chosen routes is difficult and subject to structural errors. On the other hand, it complicates the role of the analyst.

In a study of long-distance route choice behavior in Switzerland, revealed preferences data has been collected through telephone interviews. The respondents were asked to describe their last long distance (over 20 km) car trip with the name of the origin and destination cities, as well as intermediate cities or locations on the chosen route. Hence, the observed choices correspond to sequences of zones in the route network, and the exact chosen routes are unknown, that is the choice is \emph{latent}. A direct application of existing route choice models is therefore not possible.

The problem of analyzing choices that are not directly observed has been of interest in other studies. Ben-Akiva and Lerman (1985) discuss the problem in the context of shopping destination choice. Ben-Akiva et al. (1984) estimate choice models for label paths but where the choice of physical path is observed. Toledo et al. (2003) analyze lane changing behavior where the choice of target lane is unknown and only the lane changing action is observed.

The observations at hand can be viewed as an aggregate of alternatives (a set of paths) from the route choice model. There are two ways of modeling this situation. First, the route choice model could be replaced by a model capturing the choice of ``aggregates''. Such a model would however be of limited use in practice. Second, the likelihood of an (aggregate) observation can be computed with an underlying route choice model using a detailed network description and actual paths. We adopt the latter.

In this context, not only several routes can correspond to the same observation, but the exact origin-destination pair is not necessarily known. We therefore consider several possible origin-destination pairs and their associated set of routes, generated by a choice set generation algorithm. We derive from this list the likelihood of each observation, in order to perform the maximum likelihood estimation of the route choice model.

We present estimation results of different route choice models based on 1200 observations from the survey conducted in Switzerland. In this application, the paths are defined by sequences of postal codes. The lengths of the sequences are between 3 and 5 zones, including the origin and the destination.

Modeling day-to-day learning in route choice
E. Bogers, TU Delft, The Netherlands (with M. Bierlaire and S. Hoogendoorn)

When travelers perform the same trip a number of times, they can learn about available routes from their experiences. In psychological learning theory two types of learning are found that we think play a role in day-to-day route choice: implicit (reinforcement) and explicit (belief based) learning. Memory decay plays a major role in this. Although a lot of progress has been made in modeling learning in route choice, a model that captures both learning types and for which the parameters are empirically underpinned was not found. In this paper such a model is developed and a large data set from experimental research is used to validate it and estimate its parameters. The developed model uses a Markov formulation for the day-to-day updating of the belief a person has about the travel time (the perceived travel time) on a route. Reinforcement was modeled by including the latest ten route choices in the model. The results showed that the most recent experience makes up 20% of the perceived travel time. This means that formulations that use either the mathematical mean of all past experienced travel times or only the last experienced travel times are not accurate. Furthermore, the reinforcement part of the model can make up a significant part of the utility of a route and should therefore be a standard component in route choice models. In sum, the results indicate the validity of the theoretical and mathematical model.

A practical test for the choice of mixing distribution in discrete choice models
M. Fosgerau, Danish Transport Research Institute, Denmark (with M. Bierlaire)

The choice of a specific distribution for random parameters of discrete choice models is a critical issue in transportation analysis. Indeed, various pieces of research have demonstrated that an inappropriate choice of the distribution may lead to serious biases in model forecast and in the estimated means of random parameters. In this paper, we propose a practical test, based on seminonparametric techniques. The test is analyzed both on synthetic and real data, and is shown to be simple and powerful.

Representation of taste heterogeneity in willingness-to-pay indicators
S. Mabit, DTU, Denmark

Existing treatments of taste heterogeneity in willingness-to-pay (WTP) indicators such as the valuation of travel time savings (VTTS) fall into three main categories, a) discrete segmentations, b) continuous interactions and c) random variations.

All of the above treatments can be specified directly in WTP space as opposed to preference space. As such, the WTP measures are estimated directly from the models, avoiding a calculation on the basis of independently estimated coefficients. However, while potentially avoiding significant amounts of bias, these approaches are still only used very sparingly.

The case-studies in this paper make use of various real-world datasets, as well as a set of custom-generated synthetic datasets. While the objectives differ slightly depending on the nature of the data, the various case-studies all present a comparison of the results in terms of variations in the WTP indicators depending on the treatment of taste heterogeneity, the base specification of the utility function and the decision to work in WTP or preference space. Preliminary results not only support the theoretical claims of the paper in showing significant differences in WTP indicators across modelling approaches, but importantly also indicate significant differences in the conclusions across the various datasets and scenarios used.

Multi-modal route choice models - utility functions and choice set generation
O. A. Nielsen, DTU; Denmark

At the first workshop on applications on discrete choice models, it was agreed upon to continue with joint research projects within the field of multimodal route choice models (Sacha Hoogendoorn-Lanser and Otto Nielsen) and to discuss choice set generation algorithms (also DTU and Sacha Hoogendoorn-Lanser). The lecture will present the preliminary findings of these two activities which are to some extent interrelated; 1) the type of choice set generation methods put on restrictions on the types of choice set models to be used - both on correlation structure and the use of non-linear utility functions; and 2) the size of the choice problem and network put on restrictions on feasible choice set generation methods. The discussions on choice set generation are largely based on initial work on this between Piet Bovy, TUDelft and Otto Anker Nielsen.

The public transport choice context in both the Netherlands and in Denmark involves choice situations between public transport feeder modes (buses) versus walk and bicycling. There seams to be a highly non-linearity in the utility function for walk/bi-cycle, which can be dealt with by moth choice models and choice set generation methods - although this non-linearity is usually ignored in most applied models.

In the estimation context linear models typically compensate for the lack of proper treatment of non-linearity by use of alternative specific constants (in discrete choice models) and/or a transfer penalty for transferring between bus and train (in large network based models). The transfer penalty in the choice between bus and bi-cycle can be a proxy for an alternative specific constant - i.e. the two cannot be differentiated from each other in the estimation context. Furthermore, the transfer penalty can be biased from the fact, that this may indeed is also being non-linear it selves with the number of transfers, and that the time of transferring (waiting time). Accordingly, a number of highly correlated coefficients are estimated.

Practitioners often interpret these coefficients too narrowly behaviourally, and therefore deciding modifications of these which are not justified from the data/estimation (e.g. the overlooking of transfer penalty in the OTM model for Copenhagen).

The presentation shows, that the formulation of the utility function with respect to these non-linearity and transfer penalties indeed have a very high impact on route choice probabilities in multi-model networks. While this can be dealt with in large-scale network-based models quite easily with respect to walk/bicycling, it can only be solved efficiently for medium-sized problems with respect to the other variables, while it is challenge to solve for large networks as the Copenhagen network. In this respect, there is a clear difference between transport systems of a feeder-mode hub-structure (back-bone network), opposite a network with many alternative main routes.

Towards a unification of expected utility theory and random utility theory
J. Polak, Imperial College, London, UK.

Recent work in the field of travel demand modelling has sought to address decision making under risk and uncertainty by bringing together elements of Expected Utility Theory (EUT) and Random Utility Theory (RUT). The aims of this paper are first to identify and discuss the key theoretical issues associated with the merging of EUT and RUT in a transport context and second, to explore the number of empirical issues associated with the specification of attitudes to risk. In particular, using data from a large SP exercise in which respondents were faced with a series of choices between alternative unreliable train services, we investigate the performance of alternative specification of attitudes to risk.

Evaluation of choice set generation methods for route choice modelling
C. Prato, Technion, Israel (with S. Bekhor)

The representation of individual route choice behavior should answer a simple question: which route does the traveler choose to move from the origin to the destination of the trip? The ideal route choice model is simple and elegant mathematically, easy and fast to estimate, correct and precise about prediction of actual choices from travelers moving on real size networks. Naturally, this model accounts for similarities among alternative routes… wait a second: which alternative routes? How a researcher is supposed to generate alternative routes? Do the nature and the number of the routes have any effect on the model estimates? And which model specification is more sensitive to the choice set composition? In order to answer these questions, several path generation methods are applied and compared to actual choices of individuals driving habitually in an urban environment, and different route choice models are estimated for different choice set compositions. Results suggest guidelines in the generation of alternative routes, important not only from the modeling perspective, but also from the traffic assignment applicative perspective.

Capturing similarities in route, mode and destination choice problems
N. Schuessler, ETHZ, Switzerland

Overcoming the "Independence of irrelevant alternatives" property of the basic Multinomial Logit model is still a prevalent research issue. Some approaches introduce a suitable measure of similarity in the systematic part of the utility function, others respecify the variance-covariance matrix. But none of them is completely satisfactory for transportation choice problems, which are characterised by large sets of alternatives with diverse variables. Models that open the variance-covariance structure are flexible and able to describe even complex correlation structures but they require a lot of effort in terms of specification and computation and are hardly suitable for large choice sets. Therefore an approach of introducing a similarity factor in the systematic part of the utility function will be followed in the work presented here. Whereas most similarity factor models developed so far have been designed to solve route choice problems, in this work destination will be scrutinised in detail. Destination choice alternatives are characterised by a variety of correlations between different attributes including immanent interdependencies with route and mode choice characteristics. Therefore a similarity coefficient that accounts for all these effects will be formulated. It will be applied in a combined route, mode and destination choice model. The final aim is to generalise the approach for a broader range of choice problems. In the workshop a detailed problem definition and initial ideas on how to solve this problem will be presented.

Discrete Choice Models for Static Facial Expression Recognition
M. Sorci, EPFL, Switzerland (with G. Antonini, J-Ph Thiran and M. Bierlaire)

In this paper we propose the use of Discrete Choice Analysis (DCA) for static facial expression classification. Facial expressions are described with expression descriptive units (EDU), consisting in a set of high level features derived from an active appearance model (AAM). The discrete choice model (DCM) is built considering the 6 universal facial expressions plus the neutral one as the set of the available alternatives. Each alternative is described by an utility function, defined as the sum of a linear combination of EDUs and a random term capturing the uncertainty. The utilities provide a measure of likelihood for a combinations of EDUs to represent a certain facial expression. They represent a natural way for the modeler to formalize her prior knowledge on the process. The model parameters are learned through maximum likelihood estimation and classification is performed assigning each test sample to the alternative showing the maximum utility. We compare the performance of the DCM classifier against Linear Discriminant Analysis (LDA), Generalized Discriminant Analysis (GDA), Relevant Component Analysis (RCA) and Support Vector Machine (SVM). Quantitative preliminary results are reported, showing good and encouraging performance of the DCM approach both in terms of recognition rate and discriminatory power

Residential location and travel behaviour in the Greater Dublin Area
A. Vega, University College Dublin, Ireland.

The extraordinary growth in population and employment that the Greater Dublin Area has experienced during the last decade had significant repercussions for travel behaviour and residential location patterns. The spatial distribution of employment in the region has moved away from its traditional location in the city centre, contributing to the consolidation of a polycentric urban structure. We present an analysis of the simultaneous choice of residential location and mode of travel to work in Dublin region where a number of closed-form discrete choice model structures are examined and estimation results are presented. Geographical Information System (GIS) visualisations and geostatistical analyses are used to generate the model choice set based on the definition of spatially aggregated alternatives that accommodate the new existing polycentric urban structure. This research aims to offer a more accurate picture of the relationship between residential location and travel to work patterns in the Greater Dublin Area and to contribute to the development of innovative solutions for the implementation of effective policies of transportation and land use planning.

UpList of participants

  1. Gianluca Antonini, EPFL, Lausanne, Switzerland
  2. Michel Bierlaire, EPFL, Lausanne, Switzerland
  3. James Birdsall, EPFL, Lausanne, Switzerland
  4. Enide Bogers, TU Delft, The Netherlands
  5. Mogens Fosgerau, Danish Transport Research Institute, Copenhagen, Denmark
  6. Stefan Mabit, DTU, Lyngby, Denmark
  7. Otto A. Nielsen, DTU, Lyngby, Denmark
  8. John Polak, Imperial College London, UK
  9. Carlo Prato, Technion, Israel
  10. Matteo Salani, EPFL, Lausanne, Switzerland
  11. Nadine Schuessler, ETHZ, Zürich, Switzerland
  12. Matteo Sorci, EPFL, Lausanne, Switzerland
  13. Jelena Stojanovic, EPFL, Lausanne, Switzerland
  14. Jean-Philippe Thiran, EPFL, Lausanne, Switzerland
  15. Amalia Vega, University College Dublin, Ireland