正态逼近与基于覆盖宽度的EM估计

提出了非正态分布的有限混合正态分布的逼近思路.为了解决混合正态分布中正态分布个数的确定问题,针对基于极大似然估计的期望最大化(EM,Expectation Maximization)算法,提出了最大覆盖宽度的定阶原则.实证结果表明该方法的可行性.在阶数确定上,最大覆盖准则要优于赤池信息准则,而在宽度计算中,对于最大均值和最小均值的基于标准差的权重调整是必要的.

Approximation by normal distribution with covering width based EM estimation

Based on the comparison and analysis on existingnon-normal distributions, the approximation method by using mixture normal distributions was proposed. For overcoming the order determination difficultyin maximum likelihood estimation based expectation maximization (EM) algorithm, the maximalcovering width principle (MCWP) for order determining was designed. The experiment result supports the feasibility of the method. It shows that the fitting by MCWP has the advantage of the Akaike information criterion. In calculating the covering width, it is necessary to adjust the maximum mean and the minimum mean by the standard deviationbased weights.