时间:2019年03月20(星期三)11:00-12:00
地点:学院南路校区,图配楼506
报告题目: ParallelizableAlgorithms for Optimization Problems with Orthogonality Constraints
报告人:刘歆,中国科学院数学与系统科学研究院副研究员,博士生导师,国家优秀青年科学基金获得者。
报告摘要:To constructa parallel approach for solving optimization problems with orthogonalityconstraints is usually regarded as an extremely difficult mission, due to thelow scalability of the orthogonalization procedure. However, such demand isparticularly huge in some application domains such as material computation. Inthis talk, we propose two infeasible algorithms, based on augmentedLagrangian penalty function, for solving optimization problems withorthogonality constraints. Different with the classic augmented Lagrangianmethod, our algorithms update both the prime variables and the dual variablesby new strategies. The orthogonalization procedure is only invoked once as thelast step of the above mentioned two algorithms. Consequently, the main partsof these two algorithms can be parallelized naturally. We establish globalsubsequence convergence results for our proposed algorithms. Worst-casecomplexity and local convergence rate are also studied under some mildassumptions. Numerical experiments, including tests under parallel environment,illustrate that our new algorithms attain good performances and a highscalability in solving discretized Kohn-Sham total energy minimizationproblems.
报告人简介:刘歆2004年本科毕业于北京大学数学科学学院;2009年于中国科学院研究生院获得博士学位,导师是袁亚湘院士;毕业后留所工作至今。期间分别在德国ZIB研究所、美国RICE大学、美国纽约大学Courant研究所进行过长期访问。他的主要研究方向包括:正交约束矩阵优化问题,线性与非线性特征值问题,及其在电子结构计算中的应用;非线性最小二乘的算法与理论,分布式优化算法设计,及其在机器学习中的应用。刘歆在2016年8月获得国家自然科学基金委优秀青年科学基金;2016年10月获得中国运筹学会青年科技奖;2017年2月入选中国科学院北京分院“启明星”优秀人才计划。于2015年7月起担任《MathematicalProgramming Computation》编委;2016年10月起担任中国运筹学会理事;2017年7月起担任《计算数学》编委;2018年5月起担任中国科学院青年创新促进会数理分会副会长; 2018年6月起担任《物理学报》特约栏目编辑。
本次活动受威尼斯wns8855662019专题学术讲座项目资助。