Go to ICLR 2025 Conference homepageMinimax Based Fast-training Defense against Adversarial Policy in Two-player Competitive Games
Keywords: Deep reinforcement learning, adversarial policy, adversarial defense, Nash equilibrium
TL;DR: This work presents a novel robust policy training method based on the Minimax Theorem, which significantly reduces computational overhead and outperforms existing methods in both efficiency and robustness in two-player game environments.
Abstract: Adversarial policies have been shown to exploit vulnerabilities in agents during two-player competitive games, significantly undermining their performance. While existing approaches model the challenge of training robust policies in such environments as the search for Nash equilibrium points in the policy space, this often leads to substantial computational overhead. In this work, we propose MM-FATROL, a novel robust policy training method grounded in the Minimax Theorem, which significantly reduces computational overhead by efficiently identifying promising policy updates. We provide a formal analysis of the speedup achieved by our method. Extensive experiments demonstrate that MM-FATROL not only enhances efficiency but also surpasses the state-of-the-art method in terms of generalization and robustness. Additionally, we discuss the limitations of our approach and the challenges that remain in developing robust policies for more complex game environments.
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.
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: 13310
Loading