DIFFUSION POLICIES FOR OFFLINE REINFORCEMENT LEARNING

The Hong Kong University of Science and Technology (Guangzhou)

数据科学与分析学域

PhD Thesis Proposal Examination

By Mr. Linjiajie FANG

摘要

Reinforcement learning (RL) has made significant progress in decision-making problems, but most advancements have been tested in games or simulations that require millions of interactions. This limits the practical application of RL methods in real-world scenarios. To address this, offline RL aims to learn effective policies from previously collected data without needing online interactions. However, relying solely on offline data poses challenges. Prior data, such as human demonstrations, is often suboptimal and covers only a small part of the state-action space. Learning policies that surpass the behavior policy requires evaluating the value function of actions not present in the dataset. These out-of-distribution actions can increase bootstrapping error in value function estimation,leading to overestimated action values and poor performance. This proposal provides a brief overview of recent advancements in offline RL and discusses potential solutions through improved policy representations. Specifically, we explore the use of powerful diffusion models to enhance the expressiveness of policies and solve the challenge of offline RL.

TPE Committee

Chair of Committee: Prof. Xiaowen CHU

Prime Supervisor: Prof. Dong XIA

Co-Supervisor: Prof. Wenjia WANG

Examiner: Prof. Zixin ZHONG