Ecole Polytechnique Fédérale de Lausanne     
anglais seulement
    Place centrale > SB > IMA > ROSO > Workshop DCM

First Workshop on Applications of Discrete Choice Models

September 1-2, 2005
Room MA-12

Ecole Polytechnique Fédérale de Lausanne

organized by

Michel Bierlaire
Chaire de Recherche Opérationnelle ROSO
EPFL - Ecole Polytechnique Fédérale de Lausanne

The objective of the workshop is to identify joint research interests among research teams working on applications of discrete choice models. The 2005 workshop will be organized around three topics:

At the end of the workshop, a list of potential collaborations will be identified, with specific objectives (like writing a research paper or a research proposal)


14:00 - 14:10 : Michel Bierlaire
14:10 - 14:50 : Gianluca Antonini
14:50 - 15:30 : Sebastian Caussade
15:30 - 16:10 : Emma Frejinger
16:10 - 16:40 : Coffee break
16:40 - 17:20 : Andrea Papola
17:20 - 18:00 : Mogens Fosgerau
18:00 - 18:40 : Otto Nielsen
19:30 : Dinner
09:00 - 9:40 : Serge Hoogendoorn
9:40 - 10:20 : Stefan Mabit
10:20 - 10:50 : Coffree break
10:50 - 11:30 : Stephane Hess
11:30 - 12:10 : Sascha Hoogendoorn-Lanser
12:10 - 14:00 : Lunch
14:00 : Workshop meeting
09:00 : Meeting in Lausanne
09:30 : Meeting in Orbe
10:00 - 12:00 : Hiking (if weather allows)
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


EPFL-rates apply to the following hotels:

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: 147.20; double: 206.40


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.

The workshop will take place in room MA12 (see map).

Timetable. Check also the Swiss Federal Railways website.

UpList of presentations

A Discrete choice framework for acceleration and direction change behaviors in walking pedestrians
Gianluca Antonini and Michel Bierlaire, EPFL, Lausanne, Switzerland

In this work we propose a discrete choice framework for pedestrian walking behavior. We are interested in the short range behaviors of individuals, as reactions to their immediate environment. Walking behaviors are classified into direction change and acceleration patterns, both of them being constrained and unconstrained. An example of observed patterns for pedestrians fitting such a classification are the leader follower (constrained acceleration) and collision avoidance (constrained direction change) models. The spatial correlation structure in the choice set deriving from a simultaneous choice of speed regimes and radial directions is taken into account specifying a cross nested logit model (CNL). Quantitative results are presented, obtained by maximum likelihood estimation on a real data set with more than 10 thousands position observations, manually tracked from video sequences. Finally, a pedestrian simulator is presented for validation purposes.

Assessing the influence of design dimensions in stated choice experiments estimates and willingness-to-pay measures
Sebastian Caussade, Steer Davies Gleave Chile

Stated choice (SC) methods include a variety of ways to elicit consumer's preferences in addition to the possibility of estimating willingness-to-pay (WTP) for improving specific attributes. Even when choices provide useful information about consumer preferences, they contain what specialists have interpreted for years as noise or unexplained variance. Although modellers have developed methods to separate parameter estimates from noise, such efforts have tended to focus on the information supplied by the choices themselves ignoring the influence that the instrument design can have. This slant has arisen in spite of a growing body of research by both psychologists and specialists in decision making processes who have suggested that both task complexity and choice environment affect respondents facing complex situations. Given this evidence, the scope of this research was to investigate, reveal and assess the influence (on model estimates and on measures of WTP) of the complexity of a stated route choice experiment carried out in Santiago de Chile. The complexity of the experiment was encompassed by the number of available alternatives, number of attributes used to characterize the alternatives, number of choice situations presented to the respondent, number of attribute levels and variation range for those levels. The modelling proceeds by means of a Heteroskedastic Logit (HL) model (Swait and Adamowicz, 2001; DeShazo and Fermo, 2002), where the scale factor is allowed to be a function of the design dimensions.

Separating preferences and errors: a doubly seminonparametric model for binary panel data
Mogens Fosgerau, Danish Transport Research Institute

The mixed logit is a flexible model that can in principle approximate any RUM. Without justification, it is often interpreted such that the logistic errors are optimisation errors while the stochastic parameters are seen as preference variation. Furthermore the distribution of the stochastic parameters needs to be specified in advance with often drastic consequences for results and little possibility for testing the assumptions.

The paper investigates separate identification of the distribution of preferences and the distribution of errors in a simple setting using experimental binary stated choice panel data. The model investigated resembles a mixed logit model with a one-dimensional mixing distribution, except that here both the mixing distribution and the kernel are nonparametric, approximated by series expansions. The model is tested on both simulated data and a real dataset.

Capturing Overlap with Subpaths in Route Choice Models
Emma Frejinger and Michel Bierlaire, EPFL, Lausanne, Switzerland

There are two main difficulties in route choice modeling, namely generating choice sets and capturing the correlation among alternatives. In this presentation we focus on two ways of capturing correlation. First, we discuss deterministic corrections of the Multinomial Logit model; Path-Size Logit and C-Logit. We show that among the Path Size formulations, only the original one should be used, we also show that this Path Size formulation should be preferred to the different Commonality Factor formulations. Second, we introduce the novel concept of subpaths, designed to model the correlation among paths in a network. Considering subpaths instead of links in route choice modeling has two main advantages. First, it is behaviorally more realistic and second, it significantly reduces the complexity of the models. We present a factor analytic specification of the Logit Kernel model including subpath components. We estimate route choice models from Global Positioning System (GPS) data collected in the Swedish town of Borlänge.

Some issues in modelling the value of travel-time savings
Stephane Hess, Imperial College London & RAND Europe

The computation of the value of travel-time savings (VTTS), that is the willingness by respondents to pay for reductions in their journey-time, is one of the most important topics in travel-behaviour research in general, and discrete choice modelling by extension. However, while straightforward from a conceptual point of view, major issues with specification and interpretation arise. This presentation looks at two of these issues, namely the representation of random variations in the VTTS across respondents, and ways of accommodating the presence of individuals with a zero VTTS. The first part of the presentation shows how the use of alternatives to the commonly used Normal distribution can lead to important reductions in the risk of biased results that arise when relying on symmetrical distribution functions. The second part of the presentation highlights the important risk of biased results when not accounting for the potential presence of individuals with zero VTTS, and shows how discrete mixture models can help reduce this bias while also providing additional insights into the variations in tastes across respondents.

Home-end/activity-end approach for multimodal route choice modelling: A GNL-model distinguishing home-end and activity-end nests
Sascha Hoogendoorn-Lanser, Delft University of Technology, Faculty of Civil Engineering and Geosciences, Transport & Planning, The Netherlands (joint work with Rob van Nes)

Modelling multimodal travel behaviour is complex. Travellers make choices with respect to modes, boarding nodes, transfer nodes, and alighting nodes. A typical characteristic that should be considered is that there is a clear difference between the home-based part of the trip and the activity-based part. For the home-based part travellers have other modes available compared to the activity-end of the trip, especially with respect to private modes. Furthermore, choices made for the outbound trip might determine the return trip. Finally, the traveller's level of knowledge of the transport system (location of stops and timetables) and the road network (walking, cycling and car routes) will be higher in the neighbourhood of traveller's home compared to activity locations.

These are typical characteristics that might be dealt with in tour-based modelling approach. However, tour data is hardly ever available at the level of detail required to properly determine the impact of all kinds of relevant multimodal trip characteristics such as transport mode specific in-vehicle times, costs, transfer characteristics and alike. Therefore, single trip data is often used to model route choice behaviour. Traditionally, the direction of the trip, i.e. outbound or return trip, is explicitly accounted for by inclusion of so-called directional trip attributes, such as types and order of transfers, transfer waiting time and transfer-walking times, in the utility specification. However, such an approach cannot properly account for the distinction in vehicle availability and knowledge between the home-based part and the activity based part of the trip.

This paper proposes a so-called direction-free route choice model in which only variables that are independent of the direction of the trip are considered. Such a model allows to account for home-end and activity-end characteristics. For instance, separate variables are included for among other things, in-vehicle times and walking times, at home-ends and activity-ends, apart from the variables for the line-haul parts of the trip.

The differences between home-ends and activity-ends of trips are studied in detail using a specific type of Generalised Nested Logit (GNL) model (Wen & Koppelman (2001)). The flexibility of the GNL-model with respect to allocating alternatives and estimating logsum parameters is largely dependent on the choice problem. For multi-modal train trip making for instance, alternatives can naturally be grouped such that each alternative belongs to only one home-end and only one activity-end nest. Due to the fact that the number and characteristics of alternatives strongly differ among travellers, the extent to which an alternative is allocated to a home-end and an activity-end nest cannot be estimated, and should be the same for each alternative. Since allocation parameters are equal for all home-end nests and for all activity-end nests, different logsum parameters are estimated for each nest simultaneously with all attribute parameters in the utility function.

The GNL-model has been applied to multi-modal, inter-urban train trips resulting from a survey that has been conducted among train travellers in an urbanized corridor in The Netherlands (Hoogendoorn-Lanser (2005)). The survey data was extended with detailed data on all trip components, such as travel time and costs (on mode level), as well as with similar data for all other reasonable non-reported route alternatives for the same trips. The dataset contains 708 respondents who made a multimodal train trip and in total 23.494 multimodal route alternatives. Analysis results show that the GNL-model leads to a substantial improvement of the modelling performance compared to a very detailed MNL-model. Furthermore, there is indeed a large difference in valuation of home-ends and activity-ends. Home-ends appear to be approximately twice as important in the choice process than activity-ends. The analysis thus clearly shows that accounting for correlation between alternatives based on a nesting for home-end and activity-end trip parts leads to a better insight into multimodal travel behaviour.


Microscopic continuous time and space pedestrian simulation by optimal control
Serge Hoogendoorn, Delft University of Technology, Faculty of Civil Engineering and Geosciences, Transport & Planning, The Netherlands

This contribution deals with modeling pedestrian behavior in continuous time and space using optimal control theory and dynamic game theory. First, the main behavior assumptions underlying the model derivation will be discussed. Subsequently, we will show how optimal control theory can be applied to derive the mathematical model. Nota that in fact, optimal control theory can be perceived as some sort of 'continuous choice theory' where the number of control actions is in fact infinite. The relation with discrete choice theory will be discussed explicitly in the contribution.

The model is continuous in time and space and is described by a set of ordinary differential equations. These differential equations describe the dynamics of a pedestrian in relation to the other pedestrians, obstacles, preferred walking direction and speed, etc. We will discuss the properties of the model (both in static and dynamic sense, i.e. the resulting fundamental diagram, emerging spatio-temporal patterns, etc.).

In the final part of the contribution, we will discuss model calibration and validation using microscopic pedestrian data.

Correlation between different VTTS in public transport
Stefan Mabit, Centre for Traffic and Transport, DTU (Technical University of Denmark)

When computing the value-of-travel-time-savings(VTTS) for a journey using public transport there are different time aspects included in the journey. Therefore it is natural to seperate the journey time into the different aspects e.g walking to station, waiting at station and in-vehicle time. Using a mixed logit model approach would then give a VTTS distribution for each aspect.

Theoretically and intuitively these VTTS should be connected for the same person, because they both at a minimum include this persons value of time as a ressource. This ought to show up as correlation between the different VTTS distributions. So the project is concerned with capturing these correlations in a practical way.

Application of random coefficient models for route choice
Otto Anker Nielsen, Centre for Traffic and Transport, DTU (Technical University of Denmark)

The lecture will provide an overview of work at CTT concerning application of random coefficient models for route choice.

The first part will discuss special issues concerning route choice compared to other types of discrete choice models, and argue, why use of random coefficients make a huge difference on the chosen routes compared to models without. The difference between the estimation context - where the choice set and feasible number of alternatives are typical finite and discrete - and the application context is also discussed.

The second part will deal with solution algorithms in the application of route choice models. These can mainly be divided in two classes 1) pre-choice generation of a (sub)choice set of the possible choice and application of a discrete choice model on this, and 2) a simultaneous choice set and choice probability generation by simulation approaches. Pro- and cons of the two techniques are discussed. Also some special characteristics for applied models for dynamic assignment problems and freight are briefly discussed.

The third part will discuss estimation techniques for route choice models. Most models are estimated on discrete choice data, and examples are given for road transport and schedule-based transport. This results in some problems when transferring the utility function to the application context. Then some initial results on a project where models where estimated based on GPS-data are given, since this may be a promising future technique for estimation of route choice model.

Since the purpose of the seminar also is to discuss possible concrete project and co-operation, the lecture will mainly focus on providing an overview or application of discrete choice models on route choice modelling, rather than going into details on the specific model.

On the covariance structure of Cross-Nested Logit models
Andrea Papola, Dept. of Transportation Engineering "Luigi Tocchetti", University of Naples "Federico II"

CNL model represents so far one of the most flexible GEV models, allowing a very general structure for the underlying covariance matrix. In spite of the large number of applications available in literature, several theoretical and operative issues remain unsolved: this contribution aims to suggest some research developments with the relative preliminary results.

Firstly, the relationship between real covariance figures and model parameters cannot be expressed in closed form, and only an approximate formulation can be found in literature. Therefore, the degree of approximation of this conjecture is analyzed through numerical computation of actual covariances, and a specification procedure is suggested in order to reproduce, through a CNL model, some kinds of covariance matrices.

Secondly, some theoretical issues can be pointed out. From one hand, CNL capability to reproduce any homoskedastic covariance matrix should be investigated, i.e. whether the domain of reproducible covariances is strictly contained within the domain of all the feasible (positive definite) covariance matrices. Moreover, numerical results allow pointing out a non-injective correspondence between model parameters and covariance matrix, i.e. several vectors of model parameters might lead to the same covariance matrix. Actually, corresponding choice probabilities are in turn different among each other; therefore, there is not an injective correspondence between choice probabilities and covariance matrix, differently from Probit and Mixed Logit models. This issue leads also to operative problems: for instance, the range of variation of these probabilities should be investigated, as well as the effects on estimation procedures.

UpList of participants

  1. Emmanuel Abbé, MIT, Cambridge, Ma.
  2. Gianluca Antonini, EPFL, Lausanne, Switzerland
  3. Michel Bierlaire, EPFL, Lausanne, Switzerland
  4. Sebastian Caussade, Steer Davies Gleaves Chile
  5. Mogens Fosgerau, Danish Transport Research Institute, Copenhagen, Denmark
  6. Emma Frejinger, EPFL, Lausanne, Switzerland
  7. Stephane Hess, Imperial College, London, UK
  8. Sascha Hoogendoorn-Lanser, TU Delft, The Netherlands
  9. Serge Hoogendoorn, TU Delft, The Netherlands
  10. Stefan Mabit, TU of Denmark
  11. Vittorio Marzano, University degli studi di Napoli Federico II, Italy
  12. Otto Nielsen, TU of Denmark
  13. Andrea Papola, University degli studi di Napoli Federico II, Italy
  14. Michaël Thémans, EPFL, Lausanne, Switzerland
  15. Jean-Philippe Thiran, EPFL, Lausanne, Switzerland