Adaptive Offline Data Replay in Offline-to-Online Reinforcement Learning
Keywords: offline-to-online reinforcement learning, offline data replay
TL;DR: We propose a method to adaptively leverage offline data in the online training of offline-to-online reinforcement learning.
Abstract: Offline-to-online reinforcement learning allows agents to benefit from both sample efficiency and performance by integrating offline and online learning stages. One central challenge in this setting is how to effectively combine the rich experiences collected offline with real-time online explorations. Previous works commonly adopted predetermined mixing ratios for offline data replay as a primary approach to harness offline data in online training. However, determining the best mixing ratio that suits a specific environment and offline dataset often demands empirical adjustments that are context-specific. To address this, we propose a new approach for offline data replay that dynamically adjusts the mixing ratio based on the exploration reward of the agent in the online learning phase. Specifically, our method employs a bandit model to explore and exploit various mixing ratios, subsequently establishing a dynamic adjustment pattern for these ratios to enhance offline data utilization. Empirical results demonstrate that our approach outperforms conventional offline data replay methods, consistently proving its effectiveness across various environments and datasets without the need for targeted, context-specific adjustments.
Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3208
Loading