Reproduce “Toward Evaluating Robustness of Reinforcement Learning with Adversarial Policy” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiang Zheng, Xingjun Ma, Shengjie Wang, Xinyu Wang, Chao Shen, Cong Wang
Researchers at City University of Hong Kong and collaborators developed an Intrinsically Motivated Adversarial Policy (IMAP) framework for efficiently evaluating the robustness of black-box Reinforcement Learning agents. This approach, which incorporates novel intrinsic regularizers and an adaptive bias-reduction mechanism, demonstrated superior performance in attacking victim policies and evading state-of-the-art defense mechanisms across a variety of single- and multi-agent tasks.
Key insights
- Principled adversarial intrinsic regularizers significantly improve exploration efficiency in black-box AP learning, enabling the discovery of subtle vulnerabilities in victim policies.
- An adaptive bias-reduction mechanism effectively manages the trade-off between intrinsic exploration and the primary adversarial objective, ensuring convergence to optimal adversarial policies.
- The most effective type of intrinsic regularizer for uncovering vulnerabilities can be task-dependent, highlighting the need for diverse attack strategies for comprehensive robustness assessment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.