求解结构型凸优化问题的一种算法改进  

An Improved Algorithm for Solving Structural Convex Optimization Problems

作　　者：张宇婷 李锋[1] 

机构地区：[1]云南师范大学,云南 昆明

出　　处：《应用数学进展》2021年第12期4352-4364,共13页Advances in Applied Mathematics

摘　　要：本文考虑一个求解结构型凸优化问题的新算法——带随机步长的并行分裂ADMM下降算法,它是基于并行分裂ADMM下降算法和改善步长的收缩算法产生的。新算法改变了步长和更新公式,同时利用服从独立同分布的随机数来扩张步长,提高了ADMM下降算法中因固定步长因子带来的收敛较慢的结果。然后在适当的条件下,证明了该方法是依概率收敛的。最后,通过对来自金融和统计中的一类问题进行的一系列数值实验,验证了新算法的高效性。This paper considers a new algorithm for solving structural convex optimization problems: parallel split ADMM descent algorithm with random step size, which is based on parallel split ADMM descent algorithm and shrinkage algorithm with improved step size. The new algorithm changes the step size and updates formula, and expands the step size by using random numbers subject to independent and identically distributed, which improves the slow convergence result caused by fixed step size factor in ADMM descent algorithm. Then, under appropriate conditions, it is proved that the method converges according to probability. Finally, a series of numerical experiments on a class of problems from finance and statistics verify the efficiency of the new algorithm.

关 键 词：并行分裂算法 ADMM算法 随机分布 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]