Delphos: A reinforcement learning framework for assisting discrete choice model specification

01 ABSTRACT

The paper introduces Delphos, a framework that treats discrete choice model specification as sequential decision-making and uses a Deep Q-Network to learn modeling actions. The agent can add alternative-specific constants, include generic and alternative-specific taste parameters, apply non-linear transformations, and add covariate interactions; each candidate model is estimated and evaluated by a modeler-defined reward function. The authors evaluate Delphos on simulated data and two empirical datasets, reporting that it learns effective strategies while exploring a small fraction of the feasible space and generates behaviorally plausible candidates; these results come from their specific experimental settings and represent the authors' demonstration of the framework's potential.

02 KEY FINDINGS

1. Model specification is cast as sequential decision-making, with a Deep Q-Network learning sequences of modeling actions.
2. The reward function is modeler-defined and can reflect both statistical fit and behavioral expectations.
3. Evaluation on simulated data and two empirical datasets demonstrates exploration efficiency and plausible candidate models.