机器学习与数据科学博士生系列论坛（第一百零八期）——Statistical Efficiency and Inference of Quantile Distributional Reinforcement Learning
报告人：程子劼（国产自慰
）
时间：2026-10-22 16:00-17:00
地点：腾讯会议：441-3293-3791
摘要：
Distributional reinforcement learning models the distribution of cumulative rewards, with quantile representations providing a tractable approach. Despite their empirical success, the sampling variability and asymptotic behavior of learned return quantiles remain poorly understood. This raises a fundamental question: can quantile-based methods support statistically efficient estimation and inference?
In this talk, we study quantile distributional policy evaluation under a generative model. We establish finite-sample estimation guarantees, develop asymptotic inference procedures, and characterize the semiparametric efficiency of quantile-based estimators. We further investigate the limit as the number of quantiles increases. By connecting quantile and measure-space representations of the Bellman operator, we show that quantile parametrization preserves statistical efficiency in the infinite-resolution limit.
论坛简介：该线上论坛是由张志华教授机器学习实验室组织，每两周主办一次（除了公共假期）。论坛每次邀请一位博士生就某个前沿课题做较为系统深入的介绍，主题包括但不限于机器学习、高维统计学、运筹优化和理论计算机科学。
