The workshop is an informal meeting for the exchange of ideas
around discrete choice models, with the objective to trigger new
collaborations, or strengthen existing ones, and to expose PhD
students to the international community. The participation to the
workshop is by invitation only.
Registration fee: 300 CHF.
The registration fee includes: dinner on Thursday, lunch on Friday, lunch on Saturday and coffee breaks.
Integrating Attention and Response Time Data into Cognitive Psychology Models to Understand Discrete Choices
Traditional choice data captures outcomes, not processes. Response times and attention patterns, increasingly available at scale through webcam-based eye-tracking, reveal how decisions are made, not just what was chosen. Standard utility models struggle to absorb these signals or explain phenomena like the decoy effect and intra-individual heterogeneity. This talk makes the case for sequential sampling models as a more natural framework: one where attention and response time emerge directly from the decision process, improving both the interpretability of estimates and the efficiency of inference
Prateek Bansal is a Presidential Young (Assistant) Professor at the National University of Singapore (NUS). Before joining NUS in 2022, he was a Leverhulme Trust Early Career Fellow at Imperial College London and did a Ph.D. from Cornell, an MSc from UT Austin, and a BTech from IIT Delhi. Prateek leads the Behavioural Computational Science Lab at NUS and the Adaptive Mobility module at Future Cities Laboratory Global. His research group is interested in creating new methods to address challenging questions related to mobility behavior and the adoption of emerging technologies at an individual level and on an urban scale. His research has led to over 80 journal articles. Apart from the top Transportation journals, he regularly publishes in interdisciplinary journals like Nature Communications and Statistics and Computing. He serves as the Associate Editor of Transportation Research Part A: Policy and Practice and Transportation Research Part B: Methodological. He is a member of the TRB's standing committees on Travel Data & Methods (AED17) and an elected board member of the International Association of Travel Behavior Research (IATBR).
Venue
The workshop will take place at Ecole Polytechnique Fédérale de Lausanne.
Room:
GC B1 10
A Simple Numerical Method for Rational Inattention with Posterior-Separable Costs
Posterior-separable cost formulations of rational inattention provide a flexible framework for modeling costly information acquisition. While these models have appealing theoretical properties, numerical methods for computing exact solutions remain less developed. We use convex duality to reformulate the problem as an unconstrained convex optimization problem. This allows the solution to be computed by standard numerical methods, without directly optimizing over information structures. We show how to recover the primal solution and illustrate the algorithm in an investor application with common and idiosyncratic uncertainty.
Schmid
Basil
Swiss Federal Office for Spatial Development (ARE)
The Swiss value of travel time (VTT) is estimated from a series of national stated preference (SP) surveys conducted since 2010. While pooling these surveys increases sample size and enables more robust model estimation, substantial differences in estimated VTT have been observed across survey waves. This research investigates whether these differences reflect genuine changes in preferences or methodological artifacts. Particular attention is given to changes in the calculation of travel costs, which directly affect the estimated time-money trade-offs. Building on reference-dependent choice theory, a novel utility specification is proposed in which respondents evaluate travel costs relative to the average cost level presented in the SP experiment. This formulation separates behavioral parameter estimation from the subsequent application of VTT to representative travel demand data. Using Swiss national SP route choice data from 2010 to 2025, the proposed model substantially improves comparability and model fit for car route choice and reduces intertemporal inconsistencies in estimated VTT. The results provide strong evidence for reference-dependent valuation of car travel costs, whereas little support is found for public transport. The proposed framework offers a promising approach for improving the behavioral consistency and transferability of VTT estimates across surveys and over time.
Ortelli
Nicola
Transports publics genevois (TPG)
The transfer penalty is (mostly) not about transfers: mode-specific discomfort in intermodal urban trips
Customer crowdsourcing has emerged as a promising fulfillment approach that leverages in-store customers to perform online order picking activities during their regular shopping trips. While existing studies primarily focus on operational efficiency, successful implementation also depends on customers’ willingness to participate.
This research investigates customer crowdsourcing in omnichannel grocery retail from both operational and behavioral perspectives. First, we develop a real-time crowd-picking framework that combines order decomposition and learning-based task assignment to manage fulfillment operations dynamically. Using real-world data, we show that customer-based picking can substantially improve fulfillment efficiency and reduce operational costs.
Second, we extend this perspective by examining customer acceptance of crowd-picking tasks through stated preference experiments and discrete choice models. We identify how task characteristics, incentive mechanisms, and customer heterogeneity influence participation decisions.
Together, these studies provide a broader understanding of customer crowdsourcing by addressing both the operational feasibility and behavioral acceptance of crowd-picking systems and offer insights for designing scalable customer-based fulfillment models.
Prateek Bansal
Prateek
National University of Singapore
Integrating Attention and Response Time Data into Cognitive Psychology Models to Understand Discrete Choices
Traditional choice data captures outcomes, not processes. Response times and attention patterns, increasingly available at scale through webcam-based eye-tracking, reveal how decisions are made, not just what was chosen. Standard utility models struggle to absorb these signals or explain phenomena like the decoy effect and intra-individual heterogeneity. This talk makes the case for sequential sampling models as a more natural framework: one where attention and response time emerge directly from the decision process, improving both the interpretability of estimates and the efficiency of inference
Picard
Nathalie
BETA, Unistra
Migration as an adaptation to climate change
Global warming induces heat waves, floods and other natural disasters, which makes it more and more difficult to leave in some places, especially urban, and especially in flats (by contrast to houses), and especially for older people.
Rather than relying on international migration to "solve the global warming problem" in the very long run (Cruz & Rossi-Hansberg, RES, 2024), thanks to improved productivity and local amenities in Northern countries such as Canada), we study within-country migrations, focusing on heterogeneity in households’ reactions to climate change.
We find that households significantly react to climate events (heatwaves, floods and other natural disasters) in their (re-)location decisions. Moreover, the magnitude of the reactions vary significantly over the life cycle, and across tenure status, dwelling types and urban/rural environment.
This opens avenues for research on the redistributive and behavioural effects policies related to adaptation to climate change.
El Yaakoubi
Youssef
BETA, Unistra
Negotiation of joint
discrete decisions in dual-earner couples
We develop a collective discrete-choice framework for
joint decisions in dual-earner households. Observed
outcomes are Pareto efficient and represented by a
weighted aggregation of individual utilities, where
the Pareto weight captures intra-household bargaining
power. The key identification challenge is that this
weight and individual utility scales are not separately
identified without additional structure. We show that
combining egotistic preferences, the private-good
structure of individual-specific attributes, and a
money-metric normalization delivers point
identification of the baseline level of bargaining
power, rather than only its variations across
households as in existing collective models. The
framework requires no observations on prices, income,
or budget shares, which opens it to a wide range of
empirical settings beyond traditional consumption
data. We embed the collective structure in a mixed
logit with shared error components, estimated by
Bayesian methods, and illustrate the framework on joint
commuting and vehicle ownership choices of dual-earner
couples from the Paris region.
DJAFON
Kokouvi Joseph
University of Strasbourg
Whose Marriage Is It Anyway? Intergenerational Preference Gaps, Mutual Misunderstanding, and Norm Misperception in Senegalese Marriage Markets
When parents hold authority over their daughters’ marriages but preferences diverge, the resulting match fails to maximise the welfare of the woman whose life is at stake. We study this conflict using the largest paired Discrete Choice Experiment on marriage preferences conducted to date. We interview 978 parent–daughter pairs in four Senegalese regions. Each pair independently evaluates 12 hypothetical marriage scenarios that vary on eight attributes (education, employment,husband income, family wealth, polygamy, and endogamy). Combining a stacked conditional logit, a bilateral projection module, and a norm-misperception analysis, we find that preference weights are broadly similar across generations. Both strongly value female education and husband income. Parents are, however, significantly more accepting of polygamous marriage as a second wife. Mutual understanding is imperfect. Parents correctly predict their daughter’s gender-attitude responses only 57% of the time, despite a self-projection rate of 87%, consistent with egocentric imperfect empathy. Within-household divergence in perceived community norms independently predicts DCE disagreement, showing that pluralistic ignorance compounds preference conflict.
Gomez
Luis
BETA, Unistra
You cannot drive my car without me: Car ownership and mode choice within couples
This paper analyzes the interactions between car ownership and commuting mode choices within couples. We develop a hierarchical discrete choice model based on a non-cooperative Stackelberg framework, where household decisions are modeled as a sequence of interconnected choices. At the first level, the probability that each partner assumes the role of leader or follower within the household is estimated. At the second level, vehicle acquisition decisions are modeled sequentially, with the first vehicle being acquired by the leader and the second by the follower. At the final level, partners jointly choose their modes of transport for commuting to work through a strategic interaction process consistent with a non-cooperative Stackelberg framework. The model is estimated using 2021 population census data from the Île-de-France region provided by the INSEE.
de Palma
andré
CYU CERGY PARIS UNIVERSITE
TBA
TBA
Seger
Marcel
University of Oxford
Commuter preferences for smart and bidirectional EV workplace charging: Evidence from a stated choice experiment in the UK and Germany
With the rise of ride-hailing platforms, rich transaction-level data and high-frequency passenger decisions provide a new opportunity to study the value of time. This paper asks why some passengers revise their choices after requesting a ride, for example by cancelling the request or paying to receive faster service. We argue that these decisions can be understood as responses to passengers’ evolving beliefs about how long they will have to wait before being assigned a driver. We first develop a model of ride assignment that maps how long the passenger has waited for a ride to how much longer he expects to wait. We then embed these expectations in a dynamic discrete choice model in which passengers choose whether to continue waiting, upgrade, or cancel. Using detailed data from the Iranian ride-hailing platform Tapsi, we estimate the ride-assignment component and outline how the resulting expected waiting times can be embedded in a dynamic discrete choice framework to recover passengers’ value of time.
Bierlaire
Michel
EPFL
Preto
Anne-Valérie
EPFL
Heterogeneity in Mobility: Latent Classes from Long-Term Tracking Data
Most methods for generating synthetic populations rely on cross-sectional snapshots or pseudo-panels, which do not track individuals consistently over time. This paper proposes a general framework for constructing synthetic populations whose panel structure is specified by design. Individuals are represented through life-based trajectories defined independently of calendar time, and a deterministic mapping recovers their state at any time t, allowing the reconstruction of population distributions at arbitrary points in time. The framework is model-agnostic and enforces internal consistency through structural constraints embedded in the life representation.
We further introduce a Bayesian updating mechanism that incorporates information from observed cross-sectional datasets. When data are available, the synthetic population is sampled from the posterior distribution, combining prior knowledge with the evidence contained in the observations. This allows cross-sectional information, such as census data, to inform the generation of coherent longitudinal populations.
Wang
Siyu
EPFL
A Transformer-Based Activity Schedule Generation Framework
The generative modelling of activities requires joint representation of multiple dimensions, including activity type, frequency, start time, and sequence, to capture individuals’ daily activity schedules. Transformers based on self-attention mechanisms have shown strong
capabilities in modeling complex nonlinear dependencies in sequential data, making activity sequence generation a natural application scenario and new insights into travel demand modelling. This study incorporates the OASIS (Optimisation-based Activity Scheduling
with Integrated Simultaneous choice dimensions) framework as a complementary section. We first evaluate the predictive performance of the Transformer model for key scheduling dimensions, including activity type and start time-of-day, and compare the modelling result with OASIS. The sample data from the 2021 Swiss Mobility and Transport Microcensus (MTMC) of non-commuters is utilized. Simulation results are assessed at the group-level aggregated distribution, examining the ability to reproduce observed activity-travel patterns. This framework provides another perspective for data-driven activity schedule generation and insight into the mechanisms underlying activity-travel behavior.
Paschalidis
Evangelos
EPFL
Henriquez-Jara
Bastian
Universidad de Chile
Identification of Q-Learning models under incomplete data and myopic specifications
The implementation of reinforcement learning (RL) models requires very specific datasets containing the entire choice history of participants, in addition to the outcomes/rewards obtained after each choice. However, typical datasets collected outside controlled laboratory conditions do not satisfy these require- ments. Usually, either the choice history is incomplete, or there is a partial or total lack of information about the obtained outcomes. In this paper, we explore how a simple RL algorithm, multiattribute Q-learning, can be identified under incomplete data conditions, accounting for panel structure. Then, we explore model simplifications that allow us to retrieve consistent parameters or parameter ratios. Ongoing results show that: 1) myopic specifications (i.e., ignoring the dynamic Markov structure of Q-learning) result in inconsistent parameter estimates and parameter ratios; and 2) spurious state dependence (Heckman, 1981) is found with MNL models that account for inertia when the data-generating process is Q-learning.
Kouwenhoven
Marco
Significance
How effective were the electric car purchase subsidies in the Netherlands?
This study employs discrete choice analysis within a comprehensive modeling framework to assess how effective purchase subsidies have been in driving electric vehicle (EV) adoption in the Netherlands. Central to the methodology is SPARK – a new national car fleet model – which integrates multiple discrete choice models (estimated from full-population vehicle transaction data and EV-focused stated preference surveys) to simulate consumer vehicle choice under varying conditions. The model also incorporates Bass diffusion theory to represent network effects and distinguishes between private and corporate car markets, reflecting their different drivers and sensitivities. To validate SPARK’s behavioral realism, we compare its implied price elasticities with international findings and conduct an independent longitudinal vehicle-type choice analysis using a multinomial logit model on historical data to gauge past policy impacts. By synthesizing these perspectives, the analysis disentangles key drivers of the EV share surge from 5% of new cars in 2018 to 34% in 2024 – including falling EV prices, improved range, and government incentives (purchase subsidies and tax breaks). The results suggest that while incentives did contribute, their role in the EV transition has been more modest than some policymakers expected.
Bhaskar
Ashish
Queensland University of Technology
Real time simulation of large urban networks: State-of- the-practice and research needs
Real-time simulation is becoming a key enabler of intelligent traffic management, allowing agencies to move beyond monitoring towards prediction, optimisation, and proactive network operations. Significant advances in sensing technologies, connected vehicle data, cloud computing, and artificial intelligence have made city-scale digital twins increasingly feasible. However, deploying and maintaining real-time simulation for large urban networks remains a complex scientific and engineering challenge.
In this presentation we discuss the current state of practice in real time simulation of large urban networks, with a case study on the development of Aimsun Live model for the city of Logan, Queensland, Australia. The framework integrates multiple components including, heterogeneous data sources, automated data processing, and dynamic demand estimation.
The presentation discusses key research challenges that continue to limit widespread deployment, including scalable calibration, uncertainty quantification, data fusion, computational efficiency, behavioral adaptation, and the integration of AI with simulation for network-wide prediction and control.
The presentation aims to stimulate discussion on future research directions in advancing real-time digital twins for intelligent urban mobility.
Borboën
Nicolas
EPFL
Test
Test
The workshop is organized by the Transport and
Mobility Laboratory, EPFL.
