基于经验分布函数的EM估计初值选取  

Selection of initial values for EM algorithm based on empirical distribution function

作　　者：代振 王平波[1] 卫红凯[1] DAI Zhen;WANG Ping-bo;WEI Hong-kai(College of Electronic Engineering, Naval Univ. of Engineering, Wuhan 430033, China)

机构地区：[1]海军工程大学电子工程学院,武汉430033

出　　处：《海军工程大学学报》2020年第5期16-20,共5页Journal of Naval University of Engineering

基　　金：国家自然科学基金资助项目(51109218)。

摘　　要：针对EM算法对初值较为敏感、容易收敛到局部极值点的问题,基于经验分布函数设计了两种新的初值方案,即标准差方案和点方案,仿真对比了新初值方案与传统随机单一初值方案、K初值方案的估计性能。结果表明:新的初值方案估计精度更高、估计速度更快,更适合工程应用。The EM(expectation maximization)algorithm is sensitive to initial values and easily converges to local extremism points.To solve this problem,two new initial value schemes were designed based on the empirical distribution function:standard deviation scheme and point scheme.The estimated performance of the new initial value scheme and traditional random single initial value scheme and K initial value scheme was compared through simulation.The results show that the new initial value scheme is more accurate and faster than the traditional scheme,and is more suitable for engineering practice.

关 键 词：混合高斯模型 EM算法 经验分布函数 初值选取 

分 类 号：TJ630[兵器科学与技术—武器系统与运用工程]