Title: Automated discovery of interpretable cognitive models

Abstract:
Much research in human and animal decision making uses hand-designed reinforcement learning models to capture trial-by-trial choice behavior and associated neural signals during learning. Despite a long literature and many refinements, these theories represent a relatively narrow class of models restricted by core assumptions that are not well justified and remain controversial. Recent work using data-driven methods show that additional structure remains to be captured in the data, but offer little insight or interpretability about it. I present two projects that adopt techniques from modern AI to discover new models automatically. In particular, we scale up datasets and computation, and leverage more flexible model classes to discover more accurate theories across several human and animal datasets. The resulting models fit comparably well as those from black-box methods, but use novel approaches --- first, hybrid neuro-symbolic networks and second, LLM-driven program synthesis --- to preserve interpretability, by explicitly manipulating the tradeoff between interpretability and fidelity. The discovered theories offer a new perspective on learning in classic laboratory tasks and on the promises and limitations of AI-assisted scientific discovery.

The Daw Lab at Princeton University studies how people and animals learn from trial and error (and from rewards and punishments) to make decisions, combining computational, neural, and behavioral perspectives. They focus on understanding how subjects cope with computationally demanding decision situations, notably choice under uncertainty and in tasks (such as mazes or chess) requiring many decisions to be made sequentially. In engineering, these are the key problems motivating reinforcement learning and bayesian decision theory. They are particularly interested in using these computational frameworks as a basis for analyzing and understanding biological decision making.