电子学报 ›› 2022, Vol. 50 ›› Issue (11): 2765-2772.DOI: 10.12263/DZXB.20210870

• 学术论文 • 上一篇    下一篇

基于遗传非参数MDL-BW方法的HMM结构优化

徐佳伟1,2, 罗倩1,2   

  1. 1.北京信息科技大学信息与通信工程学院, 北京 100101
    2.北京信息科技大学光电测试技术及仪器教育部重点实验室, 北京 100101
  • 收稿日期:2021-07-09 修回日期:2021-12-02 出版日期:2022-11-25
    • 作者简介:
    • 徐佳伟 男,1996年12月出生于湖南省株洲市.现为北京信息科技大学硕士研究生.主要研究方向为机器学习.E‑mail: wangyiemail163@163.com
      罗 倩 女,1965年12月出生于山西省太原市.现为北京信息科技大学副教授.主要研究方向为信号与信息处理,大数据处理.E‑mail: luoqian@bistu.edu.cn
    • 基金资助:
    • 企业委托基金 (9151524108)

HMM Structure Optimization Based on Genetic Nonparametric MDL-BW Method

XU Jia-wei1,2, LUO Qian1,2   

  1. 1.College of Information and Communication Engineering,Beijing Information Science & Technology University,Beijing 100101,China
    2.Key Laboratory of the Ministry of Education for Optoelectronic Measurement Technology and Instrument,Beijing Information Science & Technology University,Beijing 100101,China
  • Received:2021-07-09 Revised:2021-12-02 Online:2022-11-25 Published:2022-11-19

摘要:

隐马尔科夫模型(Hidden Markov Model,HMM)广泛用于语音信号等时序信号的建模.HMM的结构优化包括模型参数个数优化和参数值的优化.针对传统的用于训练HMM的鲍姆-韦尔奇(Baum Welch,BW)算法在寻求最优解时容易陷入局部极值以及无法优化HMM参数个数的问题,本文提出了遗传非参数MDL-BW方法.该方法通过结合遗传(Genetic Algorithm,GA)算法随机搜索的特点和自适应思想来扩大HMM参数值解的搜索空间,结合非参数思想帮助自动寻求HMM的合适参数个数,同时以最小描述长度MDL(Minimum Description Length,MDL)作为模型优化准则来寻求HMM在全局上的最优结构.仿真数据、语音数据以及人体动作数据的仿真结果表明遗传非参数MDL-BW方法相较BW方法等同类方法在HMM结构的寻求上具有更好的效果.

关键词: 随机搜索, MDL准则, 非参数, 结构优化, 全局优化, 隐马尔科夫模型, BW方法

Abstract:

Hidden Markov model(HMM) is widely used for modeling time series signals such as speech signals. The structural optimization of HMM includes optimization of the number of model parameters and parameter values. Aiming at the problem that the traditional Baum Welch(BW) method used to train HMM is easy to fall into local maxima and the number of parameters cannot be optimized when seeking the optimal solution, genetic nonparametric MDL-BW method was proposed. This method expanded the search space of parameter values of HMM by combining the characteristics of stochastic search of genetic algorithm(GA) with adaptive ideas, and combined nonparametric ideas to help automatically find the appropriate number of HMM parameters, and used minimum description length(MDL) as optimization criterion to find the global optimal structure of HMM. Based on simulation data, speech data and human action data, the results show that the genetic nonparametric MDL-BW method has a better performance in searching for the structure of the HMM comparing with the BW method and other similar methods.

Key words: stochastic search, minimum description length(MDL) criterion, nonparametric, structural optimization, global optimization, hidden Markov model(HMM), Baum Welch(BW) method

中图分类号: