报告人
He Chen
Chinese University of Hong Kong
时间
2025年11月25日 星期二
下午 14:00-15:00
地点
602会议室
Abstract
Solving Stackelberg games (SG) to global optimality is generally intractable. A common surrogate is to compute a stationary strategy for the leader. Existing methods, however, only provide weak notions of stationarity when the follower’s optimal response is not unique. This talk will show that strong (Clarke) stationarity is still computable even when the follower admits multiple optimal solutions. The key ingredient is a new structural property, called set smoothness, which captures the variational dependence of the follower’s solution set on the leader’s strategies. Building on this notion, we establish the first non-asymptotic complexity guarantee for computing Clarke stationary strategies in general SG. Beyond this specific application, the set smoothness property emerges as a structural concept of broader interest. It offers a new lens for analyzing hierarchical, multi-agent, and parametric optimization problems that arise in economics and modern machine learning.
Biography
He Chen is currently a final-year PhD candidate at Chinese University of Hong Kong (CUHK), advised by Professor Anthony Man-Cho So. His research focuses on first-order optimization, bilevel optimization, and their applications in machine learning and economic computation. He is particularly interested in developing new analytical tools to resolve challenging open problems in these areas. His work has been published in top-tier conferences including NeurIPS, WINE, and ICML, and is currently under revision at the leading economics journal Games and Economic Behavior.
