More about HKUST
Sample Complexity in Reinforcement Learning
PhD Qualifying Examination Title: "Sample Complexity in Reinforcement Learning" by Mr. Zijun CHEN Abstract: Reinforcement Learning (RL) is a fundamental branch of machine learning in which an agent learns to make decisions by interacting with an environment to maximize its cumulative reward. In practice, collecting samples with interaction from real-world environments is often expensive and time-consuming. Therefore, understanding the sample complexity-the number of interactions required to learn a near-optimal policy or value function has become a central issue in both the theoretical and practical aspects of RL research. In this survey, we review the historical development and recent advances in the study of sample complexity in RL. We begin by examining the classical RL setting with discounted rewards, followed by a discussion of the average reward criterion. Next, we introduce recent progress in distributionally robust reinforcement learning (DR-RL). For each topic, we compare model-free and model-based methods, and analyze the underlying mathematical principles that govern the behavior of these algorithms. Finally, we highlight promising future research directions in this field. Date: Friday, 5 September 2025 Time: 3:00pm - 5:00pm Venue: Room 5501 Lifts 25/26 Committee Members: Prof. Ke Yi (Supervisor, Chairperson) Dr. Sunil Arya Dr. Nian Si (IEDA)