<ul><li>Online safe reinforcement learning is crucial in dynamic environments like autonomous driving, robotics, and cybersecurity.</li><li>Existing methods for constrained Markov decision processes struggle in adversarial settings with unknown, time-varying constraints.</li><li>The Optimistic Mirror Descent Primal-Dual (OMDPD) algorithm is introduced to handle online CMDPs with anytime adversarial constraints.</li><li>OMDPD achieves optimal regret O(sqrt(K)) and strong constraint violation O(sqrt(K)) without requiring a strictly known safe policy, providing practical guarantees for safe decision-making in adversarial environments.</li></ul>

An Optimistic Algorithm for online CMDPS with Anytime Adversarial Constraints

Discover more