机器学习与数据科学博士生系列论坛（第一百零一期）—— Continuous-Time Reinforcement Learning with Risk-Sensitive Objectives
报告人：谢楚焓（国产自慰
）
时间：2026-04-30 16:00-17:00
地点：腾讯会议：928-6293-8217
摘要：
In this talk, we study and design RL algorithms in a setting with continuous-time dynamics and risk-sensitive objectives.
While the nonlinear nature of risk measures poses fundamental challenges, such as the lack of Markov optimality and direct construction of the Bellman equation, these difficulties can be overcome when the risk measure is one of the optimized certainty equivalents (OCE), via an augmentation approach based on a prior work. By introducing an appropriately constructed augmented environment, the original risk-sensitive problem can be reformulated as a risk-neutral one that closely resembles standard RL settings, and a meta-algorithm is proposed to link the solution in the augmented formulation with that of the original formulation.
We further adopt a martingale approach to solve the augmented problem. We demonstrate the effectiveness of the proposed approach through numerical experiments.
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。
