题     目：Inexact Sequential Quadratic Optimization with Penalty Parameter Updates within the OP Solver

报 告 人：王浩

单     位：上海科技大学

时     间：2022年11月18日15:30

地     点：腾讯 632-440-885

摘 要：This talk focuses on the design of sequential quadraticoptimization (commonly known as SOP) methods for solving large-scalenonlinear optimization problems. The most computationally demandingaspect of such an approach is the computation of the search directionduring eachiteration, for whichwe consider the use of matrix-freemethods. In particular we develop a method that requires an inexact solveof a single OP subproblem to establish the convergence of the overall SOpmethod. It is known that SOP methods can be plagued by poor behavior ofthe global convergence mechanism. To confront this issue, we propose theuse of an exact penalty functionwith a dynamic penalty parameterupdating strategy to be employed within the subproblem solver in such away that the resulting search direction predicts progress toward bothfeasibility and optimality. We present our parameter updating strategy andprove that, under reasonable assumptions, the strategy does not modify thepenalty parameter unnecessarily. We close the paper with a discussion ofthe results of numerical experiments that illustrate the benefits of ourproposed techniques.

报告人简介: 王浩博士，于2015年5月在美国Lehigh University工业工程系获得博士学位，导师为Frank E.Curtis，并于2010年和2007年在北京航空航天大学数学与应用数学系分别获得理学硕士和学士学位主要成果在优化顶级期刊SIAM Journalon Optimization上发表。王浩博士于2016年3月以助理教授加入上海科技大学信息与技术学院。当前研究领域主要为惩罚算法、非精确算法、正则化问题等。