Abstract

Hierarchical decision-making often ignores context and sequential interactions. This talk presents two complementary frameworks: Contextual Stochastic Bilevel Optimization (CSBO) and Contextual Bilevel Reinforcement Learning (CB-RL). CSBO solves stochastic bilevel problems where the lower-level solution depends on both side information and upper-level variables—with applications in meta-learning and robust optimization—and we provide an efficient gradient method with optimal complexity. CB-RL extends this to sequential Stackelberg games, where a leader configures the environment while followers solve contextual MDPs; our hypergradient-based algorithm learns from observed trajectories without needing access to followers’ internal training. Together, these works establish a unified theoretical and algorithmic foundation for contextual bilevel decision-making, bridging static optimization and reinforcement learning under uncertainty.

Time

2026-08-03  10:30 - 11:30

Speaker

胡逸凡是罗格斯大学统计系助理教授。他的研究兴趣集中于不确定性下的决策，涉及随机优化、强化学习、统计学和因果推理的交叉领域,旨在构建新模型并开发易于实施且有理论保证的方法。他曾担任ICML和ICLR的Area Chair，其论文发表于Operations Research,SIAM Journal on OptimizationNeurlPS,ICML等期刊与会议。胡逸凡曾在洛桑联邦理工学院和苏黎世联邦理工学院跟随Daniel Kuhn和Andreas Krause教授从事博士后研究，在此之前他在伊利诺伊大学厄巴纳-香槟分校获得运筹学博士学位，师从陈新和何袅教授。

Room

Room 104