Based On The Mixture Model Parameter Estimation To Improve The Em Algorithm And Clustering Analysis

EM algorithm (Expectation-Maximization) algorithm, as a breakthrough development of maximum likelihood estimation, was proposed in the article of the maximum likelihood from incomplete data via the EM algorithm by Dempster in 1977. It has been widely applied in data mining, image processing, bio-statistics and reliability analysis and the other fields, particularly in parameter estimation and cluster analysis under the model of mixed multi-branch. But, EM algorithm also has some deficiencies, such as iterative formula deriver complex, the convergence speed of the algorithm is very sensitive to iterative initial value and how to improve it to solve the problem of parameter estimation of mixture model when the true mixture number is unknown. The purpose of this paper is to solve the above question.First of all, in order to facilitate the readers understand of the next section, the first chapter generally introduced the background, the principle and the character of the EM algorithm. Secondly, the second chapter shows the superiority of principle simple, idea clear of the EM algorithm by setting mixed normal distribution model parameter estimation as an example and through specific algorithm implementation process.Meanwhile, it reflected the disadvantage of iterative formula derived complex. Based on the above, this paper gives a method that improved EM algorithm to simplify the process of iterative formula derived. Meanwhile, it proposed a method what is how to select iterative initial value. In chapter three, because of need data packet in the process of parameter estimation, but M-step involved the posterior probability calculate, it increases the computing complexity and affects the efficiency of the algorithm. This paper finishes alternatively parameter modification and data cluster at each iterative process by referring dynamic k-means cluster in order to reduce the computing complexity and convert complex decimal into 0-1 discrete data by introduce indicator function. How to determine the number of cluster is always an intractable problem. In this paper, we propose the rival penalization mechanism that enables the redundant densities in the mixture to be gradually faded Out during the learning. It can automatically select an appropriate number of densities mixture clustering. The experiments have the promising results.