Keywords: Offline Reinforcement Learning, Stationary Distribution Correction Estimation
TL;DR: We propose a DICE algorithm with weighted f-divergence regularization for offline RL, which enables state-action dependent regularization.
Abstract: One of the major challenges of offline reinforcement learning (RL) is dealing with distribution shifts that stem from the mismatch between the trained policy and the data collection policy. Prior offline RL algorithms have addressed this issue by regularizing the policy optimization with $f$-divergence between the state-action visitation distributions of the data collection policy and the optimized policy. While such regularization provides a theoretical lower bound on performance and has had some practical success, it is not affected by the optimality of state-actions and can be overly pessimistic, especially when valuable state-actions are rare in the dataset. To mitigate the problem, we introduce and analyze a weighted $f$-divergence regularized RL framework that can less regularize valuable but rare state-actions as long as sampling error allows. This leads to an offline RL algorithm with iterative stationary distribution correction estimation while jointly re-adjusting the regularization for each state-action. We show that the presented algorithm with weighted $f$-divergence performs competitively with the state-of-the-art methods.
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.
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: Yes
Please Choose The Closest Area That Your Submission Falls Into: Reinforcement Learning (eg, decision and control, planning, hierarchical RL, robotics)