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

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.