Mixed Hierarchical Oracle and Multi-Agent Benchmark in Two-player Zero-sum Games

Hongsong Tang, Bo Chen, Yingzhuo Liu, Yaodong Yang, Junge Zhang, Liuyu Xiang, Jianchun Xu, Zhaofeng He

Keywords: reinforcement learning, game theory, deep learning

Abstract: Self-play methods have achieved remarkable success in two-player zero-sum games, attaining superhuman performance in many complex game domains. Parallelizing learners is a feasible approach to handling large-scale games. However, parallelizing learners often leads to suboptimal exploitation of computational resources, resulting in inefficiencies. In this study, we introduce the Mixed Hierarchical Oracle (MHO), designed to enhance computational efficiency and performance in large-scale two-player zero-sum games. MHO enables the parallelization of reinforcement learning tasks through a hierarchical pipeline that balances exploration and exploitation across oracle levels. It also avoids cold-start issues by using a "model soup" initialization strategy. Additionally, we present MiniStar, an open-source environment focused on small-scale combat scenarios, developed to facilitate research in self-play algorithms. Through extensive experiments on matrix games and the MiniStar environment, we demonstrate that MHO outperforms existing methods in terms of computational efficiency and performance.

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: 7527