基于高斯混合模型的EM学习算法  

A Study of EM Learning Algorithm Based on Gaussian Mixture Model

作　　者：王源[1,2] 陈亚军[3] 

机构地区：[1]西华师范大学计算机学院微机应用研究所 [2]淮南师范学院信息技术系,安徽淮南232001 [3]西华师范大学物理与电子信息学院

出　　处：《山西师范大学学报（自然科学版）》2005年第1期46-49,共4页Journal of Shanxi Normal University(Natural Science Edition)

基　　金：四川省教育厅重点项目基金资助(2004A102)

摘　　要：本文研究了一类基于无监督聚类学习的算法———EM算法的算法实现.EM算法通常用于存在隐含变量时的聚类学习,由于引入了隐含变量,导致算法难以保证收敛和达到极优值.本文通过将该算法应用于高斯混合模型的学习,引入重叠度分析的方法改进EM算法的约束条件,从而能够确保EM算法的正确学习.In this paper,We conducts a theoretical analysis into the method of Machine leaning with EM algorithm which is an unsupervised-clusting one.The EM algorithm used to estimate some clusting-learning parameters including hidden variables, Which lead to difficulties of converging correctly and obtaining to the local maximum points . We use EM algorithm to learn some parameters of Gaussian Mixture Model and demonstrate that the analysis of the mixture density′s overlap measure can enforce the restrict conditions of EM algorithm, as a result ,this analysis can assure the efficiency of learning.

关 键 词：学习算法 EM算法 高斯混合模型 无监督聚类 算法实现 变量 约束条件 收敛 重叠度 正确 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]