Learning & Decisions under Heterogeneity
Transfer RL, latent structure discovery, bandits, and dynamic pricing
Modern decision systems rarely operate in a single homogeneous environment: firms learn across markets, customer segments, and platforms; AI agents face related but non-identical tasks; and large populations contain hidden behavioral types. This stream studies two complementary regimes — transfer, when subgroups are observable and we borrow strength with finite-sample guarantees, and latent-structure discovery, when subgroups must be identified jointly with the decision problem. The methodological core is low-rank and mixture models, bias-variance analysis of when borrowing helps, and finite-sample risk and regret guarantees, spanning reinforcement learning, bandits, dynamic pricing, and assortment optimization.
Representative work: Data-Driven Knowledge Transfer in Batch Q-Learning (JASA, 2026), RL in Latent Heterogeneous Environments (JASA, 2024), Transfer Q-Learning (EJS, 2025; ICML, 2025), Transfer Faster, Price Smarter (NeurIPS 2025 Spotlight). See publications for the full list.