机器学习与数据科学博士生系列论坛（第五十五期）—— A Distributional Perspective on Reinforcement Learning
报告人：Yang Peng (PKU)
地点：腾讯会议 723 1564 5542
Distributional reinforcement learning (DRL) is a new mathematical framework for sequential decision making, which has achieved success in various fields such as risk management and computational neuroscience. Going beyond the expectation of cumulative return G=\sum_t \gamma^t R_t considered in classical RL, DRL focuses on the probability distribution of G. This allows us to define agents' behaviors that depend on the full distributions of G. For example, agents may avoid states that carry a high probability of failure or penalize decisions that have high variance. However, distirbution is an infinite-dimensional object, which poses new challenges for algorithm design and theoretical analysis. In this seminar, we follow the recent book, "Distributional reinforcement learning", to give a concise overview of the basic concepts, theoretical framework, and practical algorithms.