Distributionally Robust Reinforcement Learning for offline RL: Optimality and Scalability

Abstract

Offline reinforcement learning (RL) focuses on learning good decision-making strategies from pre-collected datasets, without further interaction with the environment. This is especially important in high-stakes areas like healthcare and transportation, where active exploration can be costly or risky. However, offline RL faces two major challenges: limited data coverage and potential mismatches between the environment where the data was collected and the one where the policy will be used. In this talk, I will introduce how distributionally robust RL offers a principled way to address both challenges. I will show that when the data accurately reflects the target environment, this approach can make highly efficient use of the data. Even when there is a mismatch, it can still provide strong performance guarantees. I will wrap up by sharing some ongoing work aimed at making these methods more scalable and practical for real-world applications.

Bio

Yue Wang is an Assistant Professor in the Department of Electrical and Computer Engineering at the University of Central Florida. His research focuses on robust and efficient reinforcement learning, with an emphasis on learning under uncertainty. His work spans offline RL, transfer learning, and safe decision-making, and has been published in top machine learning and AI venues. His research works are published in top machine learning conferences. He received the AAAI Distinguished Paper Award in 2023.

For more info, please follow this link.