ABOUT THE TALK

This proposal addresses the state of adversarial robustness in reinforcement learning (RL). The methods proposed thus far provide novel applications of regret to RL to improve fundamentally robust decision making under uncertain circumstances, particularly under adversarial attacks. The proposed body of work promotes risk assessment as an optimization criterion, and will further study and address the theoretical obstacles stemming from the application of game theoretic and information theoretic ideas to long-horizon decision making tasks.

ABOUT THE SPEAKER

Roman BELAIRE is a Ph.D. candidate at SCIS, under the supervision of Prof. Pradeep VARAKANTHAM. Roman's dissertation research focuses on new reinforcement learning methods for policies that are robust to adversarial perturbation, using concepts from game theory and information theory to motivate principled responses. General research interests include robustness and generalization in agent systems, information theory, and reinforcement learning.