PhD Dissertation Proposal by GE Zichang | Harmony without Uniformity: Recovering Individual Decision Models from Heterogeneous Behavioral Data

ABOUT THE TALK

This dissertation studies how to recover behaviorally faithful individual decision models from heterogeneous behavioral data. An individual may be an artificial agent governed by a policy or a human participant producing a sequence of choices, and the relevant evidence may consist of complete state--action trajectories or trial-by-trial human decisions. Observations from a single target are often limited, while indiscriminate pooling across a heterogeneous population can obscure target-specific behavior or transfer incompatible structure. The central challenge is therefore to use population data to improve individual model recovery without averaging away the behavior that defines the target. Variational Trajectory Encoding (VTE) addresses the first stage of this problem by learning unsupervised representations of complete state--action trajectories without reward, goal, or policy-identity labels. These representations preserve meaningful variation among trajectory-generating policies and enable a shared policy to recover distinct policy behaviors. Population-to-Individual Cognitive Synthesis (PICS) addresses the second stage by first discovering reusable program structures from population behavior and then specializing them into executable models of individual human participants. Across ten decision-making datasets, PICS achieves the best overall average held-out performance and the highest average participant-win count among the evaluated methods.

The proposed research addresses the third stage by examining how reusable behavioral structure can be transferred across populations and how sparse target observations can be supplemented with behaviorally compatible evidence. Together, the completed and proposed studies show population data can support accurate and data-efficient individual model recovery without suppressing target-specific behavior. The dissertation ultimately pursues harmony without uniformity: learning common behavioral structure from population data while preserving the distinctions required to model each individual decision-making source.

ABOUT THE SPEAKER

GE Zichang is a Ph.D. candidate in Computer Science at the School of Computing and Information Systems, Singapore Management University, under the supervision of Professor Pradeep VARAKANTHAM. His research focuses on decision making, particularly representation learning and cognitive program synthesis.