学术报告
当前位置: 首页 >> 合作交流 >> 学术交流 >> 学术报告 >> 正文
学术报告——Andreas Themelis教授(日本九州大学)


主讲人:Andreas Themelis  教授

邀请人:王宪福 教授

地点:汇贤楼江南娱乐体育官网入口326会议室

主办单位:江南娱乐体育官网入口


专家简介

日本九州大学教授,意大利IMT LUCCA大学计算机与决策系统科学专业和比利时KU LEUVEN 大学的电气工程专业双博士学位。SIAM J. Optim., Math. Program. IEEE Trans. Automat. Control优化领域的权威期刊上发表论文12篇。


时间:2023813日   9:00

腾讯会议号:445 811 226

报告名称:Globalized Newton-type algorithms for nonsmooth nonconvex structured optimization

报告摘要

"Splitting" algorithms, such as proximal gradient and ADMM, have proven advantageous for tackling complex problems by breaking them into manageable subtasks. However, due to their first-order nature, they often encounter challenges with ill conditioning, leading to frustratingly slow convergence rates that hinder their practicality. Several attempts have been proposed to address this issue, but most are either problem specific, offer only local convergence guarantees, or compromise the algorithms' simplicity.This talk offers a tutorial on the employment of "proximal envelopes" to provide a solution that does not suffer from any of the aforementioned limitations. The tutorial focuses on the proximal gradient algorithm, but the methodology easily extends to any other splitting method that possesses an "envelope" function, such as the ADMM and the Douglas-Rachford splitting.


时间:2023815   9:00

腾讯会议号:173 542 225

报告名称:Adaptive proximal algorithms for convex optimization under local Lipschitz continuity of the gradient

报告摘要

Gradient-based proximal algorithms have traditionally been bound to global Lipschitz differentiability requirements. Attempts to widen their applicability or reduce conservatism typically involve wasteful trial-and-error backtracking routines. Extending recent advancements in the smooth setting, we show how for convex problems it is possible to avoid backtrackings altogether and retrieve stepsizes adaptively without function evaluations. We demonstrate this with an adaptive primal-dual three-term splitting method that includes proximal gradient as special case. Finally, these findings are extended to the even wider class of (simple) bilevel programs.