Project Grant 2232907

This three-year, $500,000 Project Grant from the National Science Foundation's Computer and Information Science and Engineering program will fund research at the University of California, Davis to develop trustworthy machine learning systems through adversarial robust reinforcement learning. Specifically, the award supports investigating potential vulnerabilities in reinforcement learning models and algorithms, developing robust RL approaches that mitigate impacts from adversarial attacks, and analyzing the costs and benefits of different mitigation strategies. The research aims to address open questions around whether decisions from RL agents can be trusted, whether adversaries can mislead RL agents, and how to design robust RL algorithms. While prior work has examined adversarial attacks on supervised learning, understanding of RL model vulnerabilities and corresponding defenses is less developed due to differences from supervised learning, such as long-term rewards, lack of future data access, and unknown dynamics. Outcomes from this research should advance the security and reliability of RL applications in safety-critical domains.