Offline-to-Online Reinforcement Learning with Prioritized Experience Selection
Keywords: Reinforcement Learning; Offline-to-Online Reinforcement Learning; Prioritized Experience Selection
TL;DR: This paper proposes a simple yet effective fine-tuning method named Prioritized Experience Selection (PES) for offline-to-online RL.
Abstract: Offline-to-online reinforcement learning (O2O RL) offers a promising paradigm that first pre-trains an offline policy and fine-tunes it with further online interactions. Nevertheless, the distribution shift between the offline and online phase often hinders the fine-tuning performance, sometimes even incurring performance collapse. Existing methods mitigate this by enhancing training robustness with Q-ensemble, training a density ratio estimator to balance offline and online data, etc. But they often rely on components like ensemble and have higher training costs. In this paper, we address this issue by establishing a concrete performance bound for the optimal policies between two consecutive online steps. Motivated by the theoretical insight, we propose a simple yet effective fine-tuning method, \textbf{P}rioritized \textbf{E}xperience \textbf{S}election (PES). During the online stage, PES maintains a dynamically updated priority queue containing a portion of high-return trajectories, and only selects online samples that are close to the samples in the queue for fine-tuning. In this way, the distribution shift issue can be mitigated and the fine-tuning performance can be boosted. PES is computationally efficient and compatible with numerous approaches. Experimental results on a variety of D4RL datasets show that PES can benefit different offline and O2O RL algorithms and enhance Q-value estimate. Our code is available and will be open-source.
Supplementary Material: zip
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/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 1634
Loading