[2Q1-OS-27a] 強化学習の新展開

[2Q1-OS-27a-04] (OS invited talk) On Various Problem Settings of Sequential Decision Making and How Prior Knowledge of Problems Affects Theoretical Guarantees

〇Tadashi Kozuno¹, Toshinori Kitamura², Yuki Ichihara^3,4, Makoto Hagiwara⁵ (1. OMRON SINIC X Corporation, 2. University of Tokyo, 3. Nara Institute of Science and Technology, 4. Advanced Telecommunications Research Institute International, 5. pluszero)

Keywords:Sequential Decision Making

Recently, various problem settings of sequential decision making are proposed, and various types of performance guarantees are given. Examples are non-stationary MDPs and constrained MDPs. In this paper, we summarize and report recent development in performance guarantees for those settings, and we explain what kind of prior knowledge is useful for a better performance guarantee. Then, finally, we discuss open problems and interesting future study directions in the theory of sequential decision making.