Improving choice model specification using reinforcement learning
Authors:
Gabriel Nova,
Sander van Cranenburgh,
Stephane Hess
Abstract:
Discrete choice modelling is a theory-driven modelling framework for understanding and forecasting choice behaviour. To obtain behavioural insights, modellers test several competing model specifications in their attempts to discover the 'true' data generation process. This trial-and-error process requires expertise, is time-consuming, and relies on subjective theoretical assumptions. Although meta…
▽ More
Discrete choice modelling is a theory-driven modelling framework for understanding and forecasting choice behaviour. To obtain behavioural insights, modellers test several competing model specifications in their attempts to discover the 'true' data generation process. This trial-and-error process requires expertise, is time-consuming, and relies on subjective theoretical assumptions. Although metaheuristics have been proposed to assist choice modellers, they treat model specification as a classic optimisation problem, relying on static strategies, applying predefined rules, and neglecting outcomes from previous estimated models. As a result, current metaheuristics struggle to prioritise promising search regions, adapt exploration dynamically, and transfer knowledge to other modelling tasks. To address these limitations, we introduce a deep reinforcement learning-based framework where an 'agent' specifies models by estimating them and receiving rewards based on goodness-of-fit and parsimony. Results demonstrate the agent dynamically adapts its strategies to identify promising specifications across data generation processes, showing robustness and potential transferability, without prior domain knowledge.
△ Less
Submitted 6 June, 2025;
originally announced June 2025.
Understanding the decision-making process of choice modellers
Authors:
Gabriel Nova,
Sander van Cranenburgh,
Stephane Hess
Abstract:
Discrete Choice Modelling serves as a robust framework for modelling human choice behaviour across various disciplines. Building a choice model is a semi structured research process that involves a combination of a priori assumptions, behavioural theories, and statistical methods. This complex set of decisions, coupled with diverse workflows, can lead to substantial variability in model outcomes.…
▽ More
Discrete Choice Modelling serves as a robust framework for modelling human choice behaviour across various disciplines. Building a choice model is a semi structured research process that involves a combination of a priori assumptions, behavioural theories, and statistical methods. This complex set of decisions, coupled with diverse workflows, can lead to substantial variability in model outcomes. To better understand these dynamics, we developed the Serious Choice Modelling Game, which simulates the real world modelling process and tracks modellers' decisions in real time using a stated preference dataset. Participants were asked to develop choice models to estimate Willingness to Pay values to inform policymakers about strategies for reducing noise pollution. The game recorded actions across multiple phases, including descriptive analysis, model specification, and outcome interpretation, allowing us to analyse both individual decisions and differences in modelling approaches. While our findings reveal a strong preference for using data visualisation tools in descriptive analysis, it also identifies gaps in missing values handling before model specification. We also found significant variation in the modelling approach, even when modellers were working with the same choice dataset. Despite the availability of more complex models, simpler models such as Multinomial Logit were often preferred, suggesting that modellers tend to avoid complexity when time and resources are limited. Participants who engaged in more comprehensive data exploration and iterative model comparison tended to achieve better model fit and parsimony, which demonstrate that the methodological choices made throughout the workflow have significant implications, particularly when modelling outcomes are used for policy formulation.
△ Less
Submitted 6 June, 2025; v1 submitted 3 November, 2024;
originally announced November 2024.
