基于GEP和Baum-Welch算法训练HMM模型的研究  

HMM training method based on GEP and Baum-Welch algorithms

作　　者：张增银[1] 元昌安[1] 胡建军[2] 蔡宏果[1] 王文栋[1] 杨立志[1] 

机构地区：[1]广西师范学院计算机与信息工程学院,广西南宁530001 [2]华南理工大学计算机科学与工程学院,广东广州510006

出　　处：《计算机工程与设计》2010年第9期2027-2029,2069,共4页Computer Engineering and Design

基　　金：国家自然科学基金项目(60763012);广西新世纪十百千人才工程专项基金项目(2006220);广西高等学校优秀人才计划基金项目(RC2007022);广西研究生教育创新计划基金项目(2009106030774M03)

摘　　要：传统的向前-向后算法或Baum-Welch算法训练HMM的转移概率aij和发射概率ai(ot),使观察序列的O概率恰好达到最大值往往很难,虽然在理论上训练HMM的这两个网络结构是可能的,但仅能保证局部的最大值,而基于全局搜索的基因表达式编程(GEP)的一个主要的特点就是可以高效快速的发现全局最优解。把GEP引入到HMM的训练中去,提出一种改进的训练方法GBHA。实验结果表明,该算法比传统算法的系统效率更高、更稳定。According to the transfer probability aij and transmit probability ai （ot） of HMM by traditional forward-backwards algorithm or Baum-Welch algorithms , which is possible to training this two network structure, but is extremely difficult to estimate parameter a and b by which make sure the observational sequence O is the max. Because this method can only lead to local optimization. Since gene expression programming （GEP） based on globel search is introduced for converging to global optimization, it is applied to HMM training, a modified training method GBHA is proposed. Experimental results show that the training method proposed is more effective than tranditioal methods, the experiments also show that GBHA is not sensitive to initial value of the model, thereby the stability of the algorithm is enhanced.

关 键 词：隐马尔科夫模型 基因表达式编程 遗传算法 Baum-Welch算法 参数最优化 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]