Optimal Sample Complexity of Single-Time-Scale Actor-Critic, and Non-rectangular Robust RL скачать в хорошем качестве

Optimal Sample Complexity of Single-Time-Scale Actor-Critic, and Non-rectangular Robust RL

Optimal Sample Complexity of Single-Time-Scale Actor-Critic, and Non-rectangular Robust Reinforcement Learning, by Navdeep Kumar Date : 09 Feb 2026 Abstract: In this talk, I will discuss our work on single-timescale actor–critic methods, improving the sample complexity for finding an $\varepsilon$-close global optimal policy from $O(\varepsilon^{-4})$ to the optimal $O(\varepsilon^{-2})$ using STOchastic Recursive Momentum (STORM). In the second part, I will present our work on robust reinforcement learning, with applications in high-stakes domains such as robotics and healthcare. We introduce a novel value-regularizer and, for the first time, develop an efficient algorithm for a class of non-rectangular uncertainty sets, which are NP-hard in general. Link to the talk: https://www.csa.iisc.ac.in/theorysemi...