报告题目：A Gradient-Based Algorithm for Constrained Bilevel Optimization

报 告 人：张进 教授 （南方科技大学）

报告时间：2025年5月23日（星期五）16:30-17:30

报告地点：数学科学学院111A      

校内联系人： 刘永朝 教授   联系方式：84708351-8619

报告摘要：This talk presents new approaches and single-loop, Hessian-free gradient-based algorithms for solving a class of constrained bilevel optimization (BLO) problems, where the lower-level problem involves constraints that couple both upper- and lower-level variables. Such problems have recently attracted considerable interest in machine learning due to their wide applicability. However, the nonsmoothness introduced by the lower-level coupling constraints complicates the design of efficient gradient-based methods. To address this challenge, we introduce a smooth reformulation of the constrained lower-level problem based on a doubly regularized gap function. This reformulation transforms the original BLO problem into an equivalent single-level optimization problem with smooth constraints. Building on this reformulation, we develop a single-loop Hessian-free gradient-based algorithm for constrained BLO problems.  Numerical experiments demonstrate the efficiency of the proposed algorithm.

报告人简介：张进，南方科技大学数学系/深圳国家应用数学中心教授，从事最优化理论和应用研究，代表性成果发表在 Math Program、 SIAM J Optim、 Math Oper Res、 SIAM J Numer Anal、 Informs. J. Comput、 J Mach Learn Res、 IEEE Trans Pattern Anal Mach Intell，以及 ICML、 NeurIPS、 ICLR 等最优化、机器学习期刊与会议上。