一类凸优化的加速混合下降算法  

A ACCELERATED HYBRID DESCENT ALGORITHM FOR CONVEX MINIMIZATION

作　　者：徐海文[1] 孙黎明[2] 

机构地区：[1]中国民用航空飞行学院计算机学院,广汉618307 [2]南京审计大学理学院,南京211815

出　　处：《计算数学》2017年第2期200-212,共13页Mathematica Numerica Sinica

基　　金：国家自然科学基金(U1233105)资助项目

摘　　要：凸优化问题的混合下降算法利用近似条件的已知信息和随机数扩张预测校正步得到了一组下降方向.而前向加速收缩算法利用高斯赛德尔迭代算法的技术,结合邻近点算法和近似邻近点算法的思想,构造了富有扩张性的下降方向.本文借鉴混合下降算法和前向加速收缩算法的思想,利用已有近似规则信息改善了混合下降算法的下降方向,得到了一类凸优化问题的加速混合下降算法.随后利用Markov不等式、凸函数性质和投影的基本性质等,实现了算法的依概率收敛证明.一系列数值试验表明了加速混合下降算法的有效性和效率性.The set of descent directions of hybrid descent method（HD Method） for convex mini- mization is obtained by the known information of approximate conditions and random num- ber expansion prediction correction step. While the rich expansion descent directions of forward accelerated contraction method is constructed by the technology of Gauss-Seidel iterative algorithm and the thoughts of proximal point algorithm （PPA） and approximate proximal point algorithm. Inspired by the idea of the hybrid descent algorithm and the for- ward acceleration algorithm, the accelerated hybrid descent algorithm for convex minimiza- tion is obtained by using the history approximate rule information to improve the iterative step of hybrid descent algorithm, Subsequetly, the probability convergence of accelerated hybrid descent algorithm is introduced by the basic properties of Markov＇s inequality, the convex function and the projection. Some numerical experiments show the effectiveness and efficiency of the accelerated hybrid descent algorithm.

关 键 词：凸优化问题 混合下降算法 邻近点算法 前向加速收缩算法 依概率收敛 

分 类 号：O224[理学—运筹学与控制论]