Efficient Sequential Policy Optimization via Off-Policy Correction in Multi-Agent Reinforcement Learning

Pei Xu, Junge Zhang, Kaiqi Huang

Keywords: trust region policy optimization, multi-agent learning

Abstract:

Although trust region policy optimization methods have achieved a lot of success in cooperative multi-agent tasks, most of them face a non-stationarity problem during the learning process. Recently, sequential trust region methods that update policies agent-by-agent have shed light on alleviating the non-stationarity problem. However, these methods are still less sample-efficient when compared to their counterparts (i.e., PPO) in a single-agent setting. To narrow this efficiency gap, we propose the Off-Policyness-aware Sequential Policy Optimization (OPSPO) method, which explicitly manages the off-policyness that arises from the sequential policy update process among multiple agents. We prove that our OPSPO has the tightness of the monotonic improvement bound compared with other trust region multi-agent learning methods. Finally, we demonstrate that our OPSPO consistently outperforms strong baselines under challenging multi-agent benchmarks, including StarCraftII micromanagement tasks, Multi-agent MuJoCo, and Google Research Football full game scenarios.

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: 8734