必赢3003am
通知与公告

【上海财经大学】Efficient first order methods for convex and nonconvex functional constrained optimization

2022年10月31日 08:47  点击:[]

报告题目Efficient first order methods for convex and nonconvex functional constrained optimization

报告人: 邓琪 副教授(上海财经大学)

报告时间: 2022111日上午10:00-11:00

报告地点: 腾讯会议(线上)

会议ID297-863-741     会议密码:202210

报告校内联系人:肖现涛   联系电话:84708351-8307


报告摘要: Functional constrained optimization is becoming more and more important in machine learning and operations research. Such problems have potential applications in risk-averse machine learning, semi-supervised learning and robust optimization among others. In this talk, we will present some new first order algorithms for solving convex or nonconvex functional constrained problems. First, we present a primal-dual method (ConEx) which utilizes linear approximations of the constraint functions to dene the extrapolation (or acceleration) step. We show that this method is a unied algorithm that achieves some of the best-known rate of convergence for solving different functional constrained convex composite problems. We will discuss how to extend our algorithm to the nonconvex setting in a proximal point framework. Second, we will discuss how to further improve the optimization performance by exploiting the problem structure. In particular, we develop new adaptive and accelerated primal-dual methods which obtain even better convergence rates on problems with strongly convex constraint functions. We illustrate the advantage of these new algorithms with an application in sparsity-inducing optimization.


报告人简介: 邓琪,现为上海财经大学信息管理与工程学院常任副教授,博士生导师。本科毕业于上海交通大学计算机系,博士毕业于美国佛罗里达大学计算机信息科学与工程系,主要研究兴趣为连续优化算法及其在机器学习和中的应用。其研究成果发表在数学优化优化、运营管理和机器学习等领域的期刊和会议上,如Mathematical Programming, Production and Operations Management, NeurIPS 等。



上一条:【北京大学】万物皆数学 下一条:稀疏恢复及其应用研讨会

关闭