基于多相关系数分组HMM2的学习算法  

The learning algorithm of HMM2 based on grouping multiple observations by multiple correlation coefficient

作　　者：杜世平[1] 

机构地区：[1]四川农业大学生命科学与理学院,四川雅安625014

出　　处：《西北大学学报（自然科学版）》2007年第2期183-186,共4页Journal of Northwest University（Natural Science Edition）

基　　金：国家自然科学基金资助项目(30300219)

摘　　要：目的为了得到一种基于多相关系数分组二阶隐马尔可夫模型(second-or-der HMM:HMM2)的学习算法。方法最大似然准则,Lagrange乘子法。结果给出了在观测噪声和马尔可夫链不相互独立条件下二阶隐马尔可夫模型(second-or-der HMM:HMM2)的结构,获得了在多观测序列不相互独立的情况下HMM2的Baum-Welech学习算法。结论为得到充足数据,以对所有参数可靠估计,必须使用多观测序列。所获算法避免了直接计算条件概率的困难,考虑了训练序列间的相关性,故使计算过程更为便捷,在观测序列分组均匀相关情况下非常有用。Aim To obtain a training algorithm of second-order HMM （HMM2） which is based on grouping multiple correlation coefficient. Methods The maximum likelihood criterion and Lag-range multiplier. Results It proposes the structure of second-order HMM （HMM2） on condition that observation noise is not independent of the Mark- ov chain, and obtain the Baum-Welch algorithm of the model on condition that multiple observations is not independent. Conclusion It generally requires multiple observations with aim to obtain a large number of data to train the model. The new algorithm avoids computing the conditional probabilities directly and considers the correlativity between successive observation vector, it is very useful for training HMM when the group of multiple observations are uniformly dependent.

关 键 词：二阶隐马尔可夫模型(second-or-der HMM:HMM2) 多观测序列 多相关系数 Baum-Welech算法 

分 类 号：O212.7[理学—概率论与数理统计]