基于进化计算的DHMM训练方法

New DHMM Training Method Based on Evolutionary Computation

隐马尔柯夫模型 (HMM)的传统训练方法—— Baum-Welch算法只能得到一个局部最优模型 ,从而影响最终的识别率。对于 CHMM,分段 K平均方法来取得一个初始值可以解决这一问题 ,但对 DHMM却改进不大。而基于全局搜索的进化计算的一个重要特点便是可以得到次优解乃至全局最优解。本文把进化计算引入到DHMM的训练中去 ,提出一种改进的进化训练方法 ,实验结果表明 ,这种训练方法初具了全局搜索和快速收敛的特点 ,得到的模型优于传统方法和直接用进化计算所得的模型 ,提高了系统的识别率。

Traditional training methods for hidden Markov models, such as Baum Welch method can only lead to local optima, which reduces the recognition rate. Segmental K means method might solve this problem for CHMMs and offer little help for DHMMs. Since evolutionary computation based on global search is introduced for converging to sub optima or global optima, it is applied to HMM training. A modified evolutionary training method is proposed. Experimental results show that the method balances well between global search and rapid convergence and the resulting models are superior to those obtained by traditional methods or simple evolutionary computation without modification, which contribute to the increase of recognition rate.