修正Gibbs采样的有限混合模型无监督学习算法  

Unsupervised Learning for Finite Mixture Models Via Modified Gibbs Sampling

作　　者：刘伟峰[1] 韩崇昭[1] 石勇[1] 

机构地区：[1]西安交通大学电子与信息工程学院,西安710049

出　　处：《西安交通大学学报》2009年第2期15-19,共5页Journal of Xi'an Jiaotong University

基　　金：国家自然科学基金资助项目(60574033;60602026);国家重点基础研究发展规划资助项目(2007CB311006)

摘　　要：针对传统有限混合模型无监督学习算法不能处理参数维数变化的问题,提出了一种基于修正Gibbs采样的无监督学习算法.该算法的关键是,在每一次完全采样之后引入分布元的合并和剔除技术,即将利用均值、协方差矩阵差值的2范数作为合并的判断准则,最小且小于阈值的分布元权重作为剔除规则.仿真实验表明,所提算法对于参数初值的选择是不敏感的,对于分布元个数的先验信息要求得更少,它不仅可以处理维数变化问题,而且不必计算跳变概率,同时能够很好地估计出分布元个数及其参数.Since the conventional algorithm can not deal with the variable parameter dimension in the unsupervised learning of finite mixture models （FMM）, an unsupervised learning algorithm based on the modified Gibbs sampling scheme is proposed. The key for the proposed algorithm is to adopt the component management techniques that include component combination and elimination after each complete iterative step. The 2-norm of the differences in the mean and covariance are used for the component combination rule, and the component elimination rule is that the component that has the least weight and is less than certain threshold will be discarded. Simulation results show that the proposed algorithm is robust for the parameter initialization and requires fewer prior information for the number of components. The proposed algorithm can deal with the variable dimension and avoid the calculation of the jump probability. Moreover, it can estimate the number of the components and parameters effectively.

关 键 词：无监督学习 有限混合模型 参数维数变化 跳变 分布元管理 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]