分布式在线共轭对偶梯度算法  

Distributed online conjugate dual gradient algorithm

作　　者：吕净阁 LV Jingge(School of Mathematics and Big Data,Anhui University of Science and Technology,Huainan Anhui 232000,China)

机构地区：[1]安徽理工大学数学与大数据学院,安徽淮南232000

出　　处：《阜阳师范学院学报（自然科学版）》2018年第4期63-68,共6页Journal of Fuyang Normal University(Natural Science)

基　　金：安徽省高校学科(专业)拔尖人才学术资助重点项目(gxbjZD2016049)资助

摘　　要：针对如何对动态环境下的流数据进行实时处理的问题,研究了一种基于权重平衡和共轭对偶梯度算法(CDG)的分布式在线学习优化算法—分布式在线共轭对偶梯度算法(DO-CDG)。首先,针对分布式在线优化问题,在CDG算法的基础上建立了数学模型并设计了DO-CDG算法,并进行求解;其次,给出算法的Regret界用于表征在线算法的优劣性,证明了当本地损失函数是强凸函数时,DO-CDG算法的收敛性以及本地估计的Regret界关于时间的次线性;最后,经过数据仿真实验,证明了算法的收敛性。Aiming at the problem of real- time processing of flowing data in dynamic environment, a distributed online learning optimization algorithm—distributed online conjugate dual gradient algorithm （DO-CDG） was studied based on weighted balance and conjugate dual gradient algorithm （CDG）. Firstly, for the problem of distributed online optimization, a mathematical model was established based on the CDG algorithm and an DO-CDG algorithm is designed to solve it; Secondly, a Regret bound of distributed online optimization algorithm was given to evaluate the performance of online algorithm and proved the convergence of the DO-CDG algorithm for the cases that the local loss functions are strongly convex, and the Regret bound is upper bounded sublinearly in the time ; Finally, simulation experiments proved the convergence of the DO-CDG algorithm.

关 键 词：在线学习 Regret界 分布式优化 权重平衡 

分 类 号：TP306.1[自动化与计算机技术—计算机系统结构]