作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2011, Vol. 37 ›› Issue (19): 153-156. doi: 10.3969/j.issn.1000-3428.2011.19.050

• 人工智能及识别技术 • 上一篇    下一篇

基于CSMDEM算法的GMM学习方法

贾可新,何子述   

  1. (电子科技大学电子工程学院,成都 611731)
  • 收稿日期:2011-03-22 出版日期:2011-10-05 发布日期:2011-10-05
  • 作者简介:贾可新(1982-),男,博士研究生,主研方向:通信信号分选,模式识别;何子述,教授、博士生导师

GMM Learning Method Based on CSMDEM Algorithm

JIA Ke-xin, HE Zi-shu   

  1. (School of Electronic Engineering, University of Electronic Science and Technology, Chengdu 611731, China)
  • Received:2011-03-22 Online:2011-10-05 Published:2011-10-05

摘要: 基于Mahalanobis距离的EM(MDEM)算法存在过分裂问题。为此,提出一种竞争结束MDEM(CSMDEM)算法。该算法将最小描述长度准则作为竞争结束条件嵌入到MDEM算法中,能够在估计混合模型参数的同时选择模型阶数。实验结果表明,该算法具有较低的平均EM迭代次数,能够较好地拟合高斯混合模型。当其被应用到跳频网台分选时,能够以较高的正确率分选跳频信号。

关键词: 高斯混合模型, Mahalanobis距离, EM算法, 最小描述长度准则

Abstract: To solve the over-splitting problem suffered in Mahalanobis distance based EM(MDEM) algorithm, a Competitive Stop MDEM (CSMDEM) algorithm is proposed. By regarding Minimum Description Length(MDL) criteria as a competitive stop condition and embedding it into MDEM algorithm, the CSMDEM algorithm can select model order while estimating the parameters of GMM. Experimental results show that the proposed CSEM algorithm has an increased capability to fit GMM while maintaining a low average number of EM iterations. By applying it to signal sorting, the proposed EM algorithm can sort FH signals with high correctness.

Key words: Gaussian Mixture Model(GMM), Mahalanobis distance, Expectation Maximization(EM) algorithm, Minimum Description Length (MDL) criteria

中图分类号: