Seminar talk at Chicago Booth: "What Makes an RL Sample Transferable?"

On November 5, I gave a talk at the Econometrics and Statistics Workshop at the University of Chicago Booth School of Business:

What Makes an RL Sample Transferable? A Statistical Theory of Transfer Q-Learning

The talk asks when the standard transfer-learning recipe — pool source and target samples, then correct for a low-complexity difference — carries over from regression to reinforcement learning. In a Markov decision process, the regression response at each stage is a Bellman target that embeds the source’s future value under the source’s transition, so pooling introduces a bias with no supervised-learning counterpart. This leads to a precise notion of a transferable RL sample and a re-weighted targeting procedure that restores Bellman alignment, with convergence rates that show exactly when source data accelerate learning.

Many thanks to the workshop organizers for the invitation and to the audience for the engaging discussion!

Related papers:

• Deep Transfer Q-Learning for Offline Non-Stationary Reinforcement Learning. J. Chai, E. Chen, J. Fan. arXiv:2501.04870.
• Transfer Q-Learning for Finite-Horizon Markov Decision Processes. E. Chen, S. Li, M. I. Jordan. Electronic Journal of Statistics, 2025.
• Data-Driven Knowledge Transfer in Batch Q-Learning. E. Chen, X. Chen, W. Jing. Journal of the American Statistical Association, 2026.
• One-Step Bellman Alignment Enables Provably Efficient Transfer in Online RL. E. Chen, E. Zhang, J. Chai, Y. Yan. arXiv:2601.21924.
• Transition Transfer Q-Learning for Composite Markov Decision Processes. J. Chai, E. Chen, L. Yang. ICML, 2025.