【统计学习方法】第10章 隐马尔可夫模型

隐马尔可夫模型（hidden Markov model，HMM）是可用于标注问题的统计学习模型，描述由隐藏的马尔可夫链随机生成观测序列的过程，属于生成模型。

1、隐马尔可夫模型的基本概念

隐马尔可夫模型的定义

隐马尔可夫模型是关于时序的概率模型，描述由一个隐藏的马尔可夫链随机生成不可观测的状态随机序列，再由各个状态生成一个观测而产生观测随机序列的过程.隐藏的马尔可夫链随机生成的状态的序列，称为状态序列（state sequence）；每个状态生成一个观测，而由此产生的观测的随机序列，称为观测序列（observation sequence）。序列的每一个位置又可以看作是一个时刻。

状态集合 Q = { q 1 , q 2 , … , q N } ∣ Q ∣ = N \begin{aligned} & Q=\left\{q_{1},q_{2},\ldots ,q_{N}\right\} \quad \left| Q\right| =N \end{aligned} ​Q={q1​,q2​,…,qN​}∣Q∣=N​

观测集合 V = { v 1 , v 2 , … , v M } ∣ V ∣ = M \begin{aligned} & V=\left\{v_{1},v_{2},\ldots ,v_{M}\right\} \quad \left| V\right| =M \end{aligned} ​V={v1​,v2​,…,vM​}∣V∣=M​

状态序列 I = { i 1 , i 2 , … , i t , … , i T } i t ∈ Q ( t = 1 , 2 , … , T ) \begin{aligned} & I=\left\{i_{1},i_{2},\ldots ,i_{t},\ldots,i_{T}\right\} \quad i_{t}\in Q \quad \left(t=1,2,\ldots,T \right)\end{aligned} ​I={i1​,i2​,…,it​,…,iT​}it​∈Q(t=1,2,…,T)​

观测序列 O = { o 1 , o 2 , … , o t , … , o T } o t ∈ V ( t = 1 , 2 , … , T ) \begin{aligned} & O=\left\{o_{1},o_{2},\ldots ,o_{t},\ldots,o_{T}\right\} \quad o_{t}\in V \quad \left(t=1,2,\ldots,T \right)\end{aligned} ​O={o1​,o2​,…,ot​,…,oT​}ot​∈V(t=1,2,…,T)​

状态转移矩阵$$\begin{array}{rl}& A={\left[a_{ij}\right]}_{N\times N}\end{array}$$

在$t$时刻处于状态$q_i$的条件下，在$t+1$时刻转移到状态$q_j$的概率$$\begin{array}{rl}& a_{ij}=P\left(i_{t+1}=q_j|i_t=q_i\right)\left(i=1,2,\dots ,N\right)\left(j=1,2,\dots ,M\right)\end{array}$$

观测概率矩阵$$\begin{array}{rl}& B={\left[b_j\left(k\right)\right]}_{N\times M}\end{array}$$

在$t$时刻处于状态$q_i$的条件下，生成观测$v_k$的概率$$\begin{array}{rl}& b_j\left(k\right)=P\left(o_t=v_k|i_t=q_j\right)\left(k=1,2,\dots ,M\right)\left(j=1,2,\dots ,N\right)\end{array}$$

初始概率向量$$\begin{array}{rl}& \pi =\left({\pi }_i\right)\end{array}$$

在时刻$t=1$处于状态$q_i$的概率$$\begin{array}{rl}& {\pi }_i=P\left(i_1=q_i\right)\left(i=1,2,\dots ,N\right)\end{array}$$

隐马尔科夫模型$$\begin{array}{rl}& \lambda =\left(A,B.\pi \right)\end{array}$$

隐马尔科夫模型基本假设：

1. 齐次马尔科夫性假设：在任意时刻$t$的状态只依赖于时刻$t-1$的状态。$$\begin{array}{rl}& P\left(i_t|i_{t-1},o_{t-1},\dots ,i_1,o_1\right)=P\left(i_t|i_{t-1}\right)\left(t=1,2,\dots ,T\right)\end{array}$$
2. 观测独立性假设：任意时刻$t$的观测只依赖于时刻$t$的状态。$$\begin{array}{rl}& P\left(o_t|i_T,o_T,i_{T-1},o_{T-1},\dots ,i_{t+1},o_{t+1},i_t,i_{t-1},o_{t-1},\dots ,i_1,o_1\right)=P\left(o_t|i_t\right)\left(t=1,2,\dots ,T\right)\end{array}$$

观测序列的生成过程

观测序列生成算法:

• 输入：隐马尔科夫模型$\lambda =\left(A,B.\pi \right)$,观测序列长度$T$;
• 输出：观测序列 O = { o 1 , o 2 , … , o t , … , o T } O=\left\{o_{1},o_{2},\ldots ,o_{t},\ldots,o_{T}\right\} O={o1​,o2​,…,ot​,…,oT​}；

1. 由初始概率向量$\pi$产生状态$i_1$；
2. $t=1$；
3. 由状态$i_t$的观测概率分布$b_j\left(k\right)$生成$o_t$；
4. 由状态$i_t$的状态转移概率分布$a_{i_t{i}_{t+1}}$生成状态$i_{t+1}\left(i_{t+1}=1,2,\dots ,N\right)$；
5. $t=t+1$；如果KaTeX parse error: Expected 'EOF', got '&' at position 2: t&̲lt;T，转至3.；否则，结束。

隐马尔可夫模型的3个基本问题

隐马尔科夫模型的3个基本问题：

1. 概率计算：已知$\lambda =\left(A,B.\pi \right)$和 O = { o 1 , o 2 , … , o t , … , o T } O=\left\{o_{1},o_{2},\ldots ,o_{t},\ldots,o_{T}\right\} O={o1​,o2​,…,ot​,…,oT​}，计算$P\left(O|\lambda \right)$
2. 学习：已知 O = { o 1 , o 2 , … , o t , … , o T } O=\left\{o_{1},o_{2},\ldots ,o_{t},\ldots,o_{T}\right\} O={o1​,o2​,…,ot​,…,oT​}，计算 ${\lambda }^{\ast }=\arg \max P\left(O|\lambda \right)$
3. 预测（编码）：已知$\lambda =\left(A,B.\pi \right)$和 O = { o 1 , o 2 , … , o t , … , o T } O=\left\{o_{1},o_{2},\ldots ,o_{t},\ldots,o_{T}\right\} O={o1​,o2​,…,ot​,…,oT​}，计算 $I^{\ast }=\arg \max P\left(I|O\lambda \right)$
   前向概率$$\begin{array}{rl}& {\alpha }_t\left(i\right)=P\left(o_1,o_2,\dots ,o_t,i_t=q_i|\lambda \right)\end{array}$$
   给定模型$\lambda$，时刻$t$部分观测序列为$o_1,o_2,\dots ,o_t$且状态为$q_i$的概率。

2、概率计算算法

前向算法

前向概率递推计算$$\begin{array}{rl}& {\alpha }_t\left(i\right)=P\left(o_1,o_2,\dots ,o_t,i_t=q_i|\lambda \right)＝P\left(i_t=q_i,o_1^t\right)\\ & =\sum _{j=1}^{N}P\left(i_{t-1}=q_j,i_t=q_i,o_1^{t-1},o_t\right)\\ & =\sum _{j=1}^{N}P\left(i_t=q_i,o_t|i_{t-1}=q_j,o_1^{t-1}\right)\cdot P\left(i_{t-1}=q_j,o_1^{t-1}\right)\\ & =\sum _{j=1}^{N}P\left(i_t=q_i,o_t|i_{t-1}=q_j\right)\cdot {\alpha }_{t-1}\left(j\right)\\ & =\sum _{j=1}^{N}P\left(o_t|i_t=q_i,i_{t-1}=q_j\right)\cdot P\left(i_t=q_i|i_{t-1}=q_j\right)\cdot {\alpha }_{t-1}\left(j\right)\\ & =\sum _{j=1}^N{b}_i\left(o_t\right)\cdot a_{ji}\cdot {\alpha }_{t-1}\left(j\right)\end{array}$$

概率计算$$\begin{array}{rl}& P\left(O|\lambda \right)=P\left(o_1^T|\lambda \right)\\ & =\sum _{i=1}^{N}P\left(o_1^T,i_T=q_i\right)\\ & =\sum _{i=1}^N{\alpha }_T\left(i\right)\end{array}$$

观测序列概率计算的前向算法：

• 输入：隐马尔科夫模型$\lambda$,观测序列$O$;
• 输出：观测序列概率$P\left(O|\lambda \right)$；

1. 初值$$\begin{array}{rl}& {\alpha }_1\left(i\right)={\pi }_i{b}_i\left(o_1\right)\left(t=1,2,\dots ,N\right)\end{array}$$
2. 递推 对$t=1,2,\dots ,T-1$$$\begin{array}{rl}& {\alpha }_{t+1}\left(i\right)=\sum _{j=1}^N{b}_i\left(o_{t+1}\right)\cdot a_{ji}\cdot {\alpha }_t\left(j\right)\left(t=1,2,\dots ,N\right)\end{array}$$
3. 终止$$\begin{array}{rl}& P\left(O|\lambda \right)=\sum _{j=1}^N{\alpha }_T\left(i\right)\end{array}$$

后向算法

后向概率$$\begin{array}{rl}& {\beta }_t\left(i\right)=P\left(o_{t+1},o_{t+2},\dots ,o_T|i_t=q_i\lambda \right)\end{array}$$
给定模型$\lambda$，时刻$t$状态为$q_i$的条件下，从时刻$t+1$到时刻$T$的部分观测序列为$o_{t+1},o_{t+2},\dots ,o_T$的概率。

后向概率递推计算$$\begin{array}{rl}& {\beta }_t\left(i\right)=P\left(o_{t+1},o_{t+2},\dots ,o_T|i_t=q_i,\lambda \right)＝P\left(o_{t+1}^T|i_t=q_i\right)\\ & =\frac{P\left(o_{t+1}^T,i_t=q_i\right)}{P\left(i_t=q_i\right)}\\ & =\frac{{\sum }_{j=1}^{N}P\left(o_{t+1}^T,i_t=q_i,i_{t+1}=q_j\right)}{P\left(i_t=q_i\right)}\\ & =\sum _{j=1}^N\frac{P\left(o_{t+1}^T|i_t=q_i,i_{t+1}=q_j\right)\cdot P\left(i_t=q_i,i_{t+1}=q_j\right)}{P\left(i_t=q_i\right)}\\ & =\sum _{j=1}^{N}P\left(o_{t+1}^T|i_{t+1}=q_j\right)\cdot \frac{P\left(i_{t+1}=q_j|i_t=q_i\right)\cdot P\left(i_t=q_i\right)}{P\left(i_t=q_i\right)}\\ & =\sum _{j=1}^{N}P\left(o_{t+2}^N,o_{t+1}|i_{t+1}=q_j\right)\cdot a_{ij}\\ & =\sum _{j=1}^{N}P\left(o_{t+2}^T|i_{t+1}=q_j\right)\cdot P\left(o_{t+1}|i_{t+1}=q_j\right)\cdot a_{ij}\\ & =\sum _{j=1}^N{\beta }_{t+1}\left(j\right)\cdot b_j\left(o_{t+1}\right)\cdot a_{ij}\end{array}$$

概率计算$$\begin{array}{rl}& P\left(O|\lambda \right)=P\left(o_1^T|\lambda \right)\\ & =\sum _{i=1}^{N}P\left(o_1^T,i_1=q_i\right)\\ & =\sum _{i=1}^{N}P\left(i_1=q_i\right)\cdot P\left(o_1|i_1=q_i\right)\cdot P\left(o_2^T|i_1=q_i\right)\\ & =\sum _{i=1}^N{\pi }_i{b}_i\left(o_1\right){\beta }_1\left(i\right)\end{array}$$

观测序列概率计算的后向算法：

• 输入：隐马尔科夫模型$\lambda$,观测序列$O$;
• 输出：观测序列概率$P\left(O|\lambda \right)$；

1. 初值$$\begin{array}{rl}& {\beta }_T\left(i\right)=1\left(t=1,2,\dots ,N\right)\end{array}$$
2. 递推 对$t=T-1,T-2,\dots ,1$$$\begin{array}{rl}& {\beta }_t\left(i\right)=\sum _{j=1}^N{\beta }_{t+1}\left(j\right)\cdot b_j\left(o_{t+1}\right)\cdot a_{ij}\left(t=1,2,\dots ,N\right)\end{array}$$
3. 终止$$\begin{array}{rl}& P\left(O|\lambda \right)=\sum _{j=1}^N{\pi }_i{b}_i\left(o_1\right){\beta }_1\left(i\right)\end{array}$$

$P\left(O|\lambda \right)$的前向概率、后向概率的表示
$$\begin{array}{rl}& P\left(O|\lambda \right)＝P\left(o_1^T\right)\\ & ＝\sum _{i=1}^N\sum _{j=1}^{N}P\left(o_1^t,o_{t+1}^T,i_t=q_i,i_{t+1}=q_j\right)\\ & ＝\sum _{i=1}^N\sum _{j=1}^{N}P\left(o_1^t,i_t=q_i,i_{t+1}=q_j\right)P\left(o_{t+1}^T|i_{t+1}=q_j\right)\\ & =\sum _{i=1}^N\sum _{j=1}^{N}P\left(o_1^t,i_t=q_i\right)P\left(i_{t+1}=q_j|i_t=q_i\right)P\left(o_{t+1}^T|i_{t+1}=q_j\right)\\ & =\sum _{i=1}^N\sum _{j=1}^{N}P\left(o_1^t,i_t=q_i\right)P\left(i_{t+1}=q_j|i_t=q_i\right)P\left(o_{t+1}|i_{t+1}=q_j\right)P\left(o_{t+2}^T|i_{t+1}=q_j\right)\\ & =\sum _{i=1}^N\sum _{j=1}^N{\alpha }_t\left(i\right)a_{ij}{b}_j\left(o_{t+1}\right){\beta }_{t+1}\left(j\right)t=1,2,\cdots ,T-1\end{array}$$

一些概率与期望值的计算

给定模型$\lambda$和观测$O$，在时刻$t$处于状态$q_i$的概率$$\begin{array}{rl}& \\ & {\gamma }_t\left(i\right)=P\left(i_t=q_i|O,\lambda \right)\\ & =\frac{P\left(i_t=q_i,O|\lambda \right)}{P\left(O|\lambda \right)}\\ & =\frac{P\left(i_t=q_i,O|\lambda \right)}{{\sum }_{j=1}^N\left(i_t=q_i,O|\lambda \right)}\\ & =\frac{P\left(o_1^t,i_t=q_i\right)P\left(o_{t+1}^T|i_t=q_i\right)}{{\sum }_{j=1}^{N}P\left(o_1^t,i_t=q_i\right)P\left(o_{t+1}^T|i_t=q_i\right)}\\ & =\frac{{\alpha }_t\left(i\right){\beta }_t\left(i\right)}{{\sum }_{j=1}^N{\alpha }_t\left(i\right){\beta }_t\left(i\right)}\end{array}$$

给定模型$\lambda$和观测$O$，在时刻$t$处于状态$q_i$且在时刻$t+1$处于状态$q_j$的概率
$$\begin{array}{rl}& \\ & {\xi }_t\left(i,j\right)=P\left(i_t=q_i,i_{t+1}=q_j|O,\lambda \right)\\ & =\frac{P\left(i_t=q_i,i_{t+1}=q_j,O|\lambda \right)}{P\left(O|\lambda \right)}\\ & =\frac{P\left(i_t=q_i,i_{t+1}=q_j,O|\lambda \right)}{{\sum }_{i=1}^N{\sum }_{j=1}^{N}P\left(i_t=q_i,i_{t+1}=q_j,O|\lambda \right)}\\ & =\frac{{\alpha }_t\left(i\right)a_{ij}{b}_j\left(o_{t+1}\right){\beta }_{t+1}\left(j\right)}{{\sum }_{i=1}^N{\sum }_{j=1}^N{\alpha }_t\left(i\right)a_{ij}{b}_j\left(o_{t+1}\right){\beta }_{t+1}\left(j\right)}\end{array}$$

在观测$O$下状态$i$出现的期望$$\begin{array}{rl}& \sum _{t=1}^T{\gamma }_t\left(i\right)=\sum _{t=1}^{T}P\left(i_t=q_i|O,\lambda \right)\end{array}$$

在观测$O$下由状态$i$转移的期望$$\begin{array}{rl}& \sum _{t=1}^{T－1}{\gamma }_t\left(i\right)=\sum _{t=1}^{T－1}P\left(i_t=q_i|O,\lambda \right)\end{array}$$

在观测$O$下由状态$i$转移到状态$j$的期望$$\begin{array}{rl}& \sum _{t=1}^{T－1}{\xi }_t\left(i,j\right)=\sum _{t=1}^{T－1}P\left(i_t=q_i,i_{t+1}=q_j|O,\lambda \right)\end{array}$$

3、学习算法

Baum-Welch 算法

将观测序列作为观测数据$O$,将状态序列作为隐数据$I$，则应马尔科夫模型是含有隐变量的概率模型
$$\begin{array}{rl}& P\left(O|\lambda \right)=\sum _{I}P\left(O|I,\lambda \right)P\left(I|\lambda \right)\end{array}$$

完全数据$$\begin{array}{rl}& \left(O,I\right)=\left(o_1,o_2,\cdots ,o_T,i_1,i_2,\cdots ,o_T\right)\end{array}$$

完全数据的对数似然函数$$\begin{array}{rl}& \log P\left(O,I|\lambda \right)\end{array}$$

$Q\left(\lambda ,\overline{\lambda }\right)$函数
$$\begin{array}{rl}& \\ & Q\left(\lambda ,\overline{\lambda }\right)=E_I\left[\log P\left(O,I|\lambda \right)|O,\overline{\lambda }\right]\\ & =\sum _I\log P\left(O,I|\lambda \right)P\left(I|O,\overline{\lambda }\right)\\ & =\sum _I\log \frac{P\left(O,I|\lambda \right)P\left(O,I|\overline{\lambda }\right)}{P\left(O|\overline{\lambda }\right)}\end{array}$$
其中，$\overline{\lambda }$是隐马尔科夫模型参数的当前估计值，$\lambda$是隐马尔科夫模型参数。

由于对最大化$Q\left(\lambda ,\overline{\lambda }\right)$函数，$P\left(O|\overline{\lambda }\right)$为常数因子，
以及$$\begin{array}{rl}& P\left(O,I|\lambda \right)={\pi }_{i_1}{b}_{i_1}\left(o_1\right)a_{i_1{i}_2}{b}_{i_2}\left(o_2\right)\cdots a_{i_{T-1}{i}_T}{b}_T\left(o_T\right)\end{array}$$
所以求$Q\left(\lambda ,\overline{\lambda }\right)$函数对$\lambda$的最大$$\begin{array}{rl}& \lambda =\arg \max Q\left(\lambda ,\overline{\lambda }\right)\iff \arg \max \sum _I\log P\left(O,I|\lambda \right)P\left(O,I|\overline{\lambda }\right)\\ & =\sum _I\log {\pi }_{i_1}P\left(O,I|\overline{\lambda }\right)+\sum _I\left(\sum _{t=1}^{T-1}\log a_{i_t{i}_{t+1}}\right)P\left(O,I|\overline{\lambda }\right)+\sum _I\left(\sum _{t=1}^T\log b_{i_t}\left(o_t\right)\right)P\left(O,I|\overline{\lambda }\right)\end{array}$$

对三项分别进行极大化：

$$\begin{array}{rl}& \max \sum _I\log {\pi }_{i_1}P\left(O,I|\overline{\lambda }\right)=\sum _{i=1}^N\log {\pi }_{i_1}P\left(O,i_1=i|\overline{\lambda }\right)\\ & s.t.\sum _{i=1}^N{\pi }_i=1\end{array}$$

1. 构造拉格朗日函数，对其求偏导，令结果为0$$\begin{array}{rl}& \frac{\partial }{\partial {\pi }_i}\left[\sum _{i=1}^N\log {\pi }_{i_1}P\left(O,i_1=i|\overline{\lambda }\right)+\gamma \left(\sum _{i=1}^N{\pi }_i-1\right)\right]=0\end{array}$$
   得$$\begin{array}{rl}& P\left(O,i_1=i|\overline{\lambda }\right)+\gamma {\pi }_i=0\\ & \sum _{i=1}^N\left[P\left(O,i_1=i|\overline{\lambda }\right)+\gamma {\pi }_i\right]=0\\ & \sum _{i=1}^{N}P\left(O,i_1=i|\overline{\lambda }\right)+\gamma \sum _{i=1}^N{\pi }_i=0\\ & P\left(O|\overline{\lambda }\right)+\gamma =0\\ & \gamma =-P\left(O|\overline{\lambda }\right)\end{array}$$
   代入$P\left(O,i_1=i|\overline{\lambda }\right)+\gamma {\pi }_i=0$，得$$\begin{array}{rl}& {\pi }_i=\frac{P\left(O,i_1=i|\overline{\lambda }\right)}{P\left(O|\overline{\lambda }\right)}\\ & ={\gamma }_1\left(i\right)\end{array}$$
   $$\begin{array}{rl}& \\ & \max \sum _I\left(\sum _{t=1}^{T-1}\log a_{i_t{i}_{t+1}}\right)P\left(O,I|\overline{\lambda }\right)=\sum _{i=1}^N\sum _{j=1}^N\sum _{t=1}^{T-1}\log a_{ij}P\left(O,i_t=i,i_{t+1}=j|\overline{\lambda }\right)\\ & s.t.\sum _{j=1}^N{a}_{ij}=1\end{array}$$
2. 得$$\begin{array}{rl}& \\ & a_{ij}=\frac{{\sum }_{t=1}^{T-1}P\left(O,i_t=i,i_{t+1}=j|\overline{\lambda }\right)}{{\sum }_{t=1}^{T-1}P\left(O,i_t=i|\overline{\lambda }\right)}\\ & =\frac{{\sum }_{t=1}^{T-1}{\xi }_t\left(i,j\right)}{{\sum }_{t=1}^{T-1}{\gamma }_t\left(i\right)}\end{array}$$
   $$\begin{array}{rl}& \\ & \max \sum _I\left(\sum _{t=1}^N\log b_{i_t}\left(o_t\right)\right)P\left(O,I|\overline{\lambda }\right)=\sum _{j=1}^N\sum _{t=1}^T\log b_j\left(o_t\right)P\left(O,i_t=j|\overline{\lambda }\right)\\ & s.t.\sum _{k=1}^M{b}_j\left(k\right)=1\end{array}$$
3. 得$$\begin{array}{rl}& \\ & b_j\left(k\right)=\frac{{\sum }_{t=1}^{T}P\left(O,i_t=j|\overline{\lambda }\right)I\left(o_t=v_k\right)}{{\sum }_{t=1}^{T}P\left(O,i_t=j|\overline{\lambda }\right)}\\ & =\frac{{\sum }_{t=1,o_t=v_k}^T{\gamma }_t\left(j\right)}{{\sum }_{t=1}^T{\gamma }_t\left(j\right)}\end{array}$$

Baum-Welch算法：

• 输入：观测数据$O=\left(o_1,o_2,\cdots ,o_T\right)$
• 输出：隐马尔科夫模型参数

1. 初始化
   对$n=0$，选取$a_{ij}^{\left(0\right)},b_j{\left(k\right)}^{\left(0\right)},{\pi }_i^{\left(0\right)}$，得到模型${\lambda }^{\left(0\right)}=\left(a_{ij}^{\left(0\right)},b_j{\left(k\right)}^{\left(0\right)},{\pi }_i^{\left(0\right)}\right)$
2. 递推
   对$n=1,2,\cdots ,$
   $$\begin{array}{rl}& \\ & a_{ij}^{\left(n+1\right)}=\frac{{\sum }_{t=1}^{T-1}{\xi }_t\left(i,j\right)}{{\sum }_{t=1}^{T-1}{\gamma }_t\left(i\right)}\\ & b_j{\left(k\right)}^{\left(n+1\right)}=\frac{{\sum }_{t=1,o_t=v_k}^T{\gamma }_t\left(j\right)}{{\sum }_{t=1}^T{\gamma }_t\left(j\right)}\\ & {\pi }_i^{\left(n+1\right)}=\frac{P\left(O,i_1=i|\overline{\lambda }\right)}{P\left(O|\overline{\lambda }\right)}\end{array}$$
   其中，右端各值按观测数据$O=\left(o_1,o_2,\cdots ,o_T\right)$和模型${\lambda }^{\left(n\right)}=\left(A^{\left(n\right)},B^{\left(n\right)},{\pi }^{\left(n\right)}\right)$计算。
3. 终止
   得到模型${\lambda }^{\left(n＋1\right)}=\left(A^{\left(n+1\right)},B^{\left(n+1\right)},{\pi }^{\left(n+1\right)}\right)$
   在时刻$t$状态为$i$的所有单个路径$\left(i_1,i_2,\cdots ,i_t\right)$中概率最大值
   $$\begin{array}{rl}& \\ & {\delta }_t\left(i\right)=\underset{i_1,i_2,\cdots ,i_{t-1}}{\max }P\left(i_t=i,i_{t-1},\cdots ,i_1,o_t,\cdots ,o_1|\lambda \right)i=1,2,\cdots ,N\end{array}$$

得递推公式$$\begin{array}{rl}& \\ & {\delta }_{t+1}\left(i\right)=\underset{i_1,i_2,\cdots ,i_t}{\max }P\left(i_{t+1}=i,i_t,\cdots ,i_1,o_{t+1},\cdots ,o_1|\lambda \right)\\ & =\underset{1\le j\le N}{\max }\left[\underset{i_1,i_2,\cdots ,i_{t-1}}{\max }P\left(i_{t+1}=i,i_t=j,i_{t-1},\cdots ,i_1,o_{t+1},o_t,\cdots ,o_1|\lambda \right)\right]\\ & =\underset{1\le j\le N}{\max }\left[\underset{i_1,i_2,\cdots ,i_{t-1}}{\max }P\left(i_{t+1}=i,i_t=j,i_{t-1},\cdots ,i_1,o_t,o_{t-1},\cdots ,o_1|\lambda \right)P\left(o_{t+1}|i_{t+1}=i,\lambda \right)\right]\\ & =\underset{1\le j\le N}{\max }\left[\underset{i_1,i_2,\cdots ,i_{t-1}}{\max }P\left(i_t=j,i_{t-1},\cdots ,i_1,o_t,o_{t-1},\cdots ,o_1|\lambda \right)P\left(i_{t+1}=i|i_t=j,\lambda \right)P\left(o_{t+1}|i_{t+1}=i,\lambda \right)\right]\\ & =\underset{1\le j\le N}{\max }\left[{\delta }_t\left(j\right)a_{ji}\right]b_i\left(o_{t+1}\right)i=1,2,\cdots ,N\end{array}$$

在时刻$t$状态为$i$的所有单个路径$\left(i_1,i_2,\cdots ,i_t\right)$中概率最大值的路径的第$t-1$个结点$$\begin{array}{rl}& \\ & {\psi }_t\left(i\right)=\arg \underset{1\le j\le N}{\max }\left[{\delta }_{t-1}\left(j\right)a_{ji}\right]i=1,2,\cdots ,N\end{array}$$

4、预测算法

维特比算法

维特比算法：

• 输入：模型$\lambda =\left(A,B,\pi \right)$和观测数据$O=\left(o_1,o_2,\cdots ,o_T\right)$
• 输出：最优路径$I^{\ast }=\left(i_1^{\ast },i_2^{\ast },\cdots ,i_T^{\ast }\right)$

1. 初始化
   $$\begin{array}{rl}& \\ & {\delta }_1\left(i\right)={\pi }_i{b}_i\left(o_1\right)i=1,2,\cdots ,N\\ & {\psi }_1\left(i\right)=0\end{array}$$
2. 递推
   对$t=2,3,\cdots ,T$$$\begin{array}{rl}& \\ & {\delta }_t\left(i\right)=\underset{1\le j\le N}{\max }\left[{\delta }_{t-1}\left(j\right)a_{ji}\right]b_i\left(o_t\right)i=1,2,\cdots ,N\\ & {\psi }_t\left(i\right)=\arg \underset{1\le j\le N}{\max }\left[{\delta }_{t-1}\left(j\right)a_{ji}\right]i=1,2,\cdots ,N\end{array}$$
3. 终止$$\begin{array}{rl}& \\ & P^{\ast }=\underset{1\le j\le N}{\max }{\delta }_T\left(i\right)\\ & i_T^{\ast }=\arg \underset{1\le j\le N}{\max }\left[{\delta }_T\left(i\right)\right]\end{array}$$
4. 最优路径回溯
   对$t=T-1,T-2,\cdots ,1$$$\begin{array}{rl}& \\ & i_t^{\ast }={\psi }_{t+1}\left(i_{t+1}^{\ast }\right)\end{array}$$
   求得最优路径$I^{\ast }=\left(i_1^{\ast },i_2^{\ast },\cdots ,i_T^{\ast }\right)$

5、概要总结

1．隐马尔可夫模型是关于时序的概率模型，描述由一个隐藏的马尔可夫链随机生成不可观测的状态的序列，再由各个状态随机生成一个观测而产生观测的序列的过程。

隐马尔可夫模型由初始状态概率向$\pi$、状态转移概率矩阵$A$和观测概率矩阵$B$决定。因此，隐马尔可夫模型可以写成$\lambda =(A,B,\pi )$。

隐马尔可夫模型是一个生成模型，表示状态序列和观测序列的联合分布，但是状态序列是隐藏的，不可观测的。

隐马尔可夫模型可以用于标注，这时状态对应着标记。标注问题是给定观测序列预测其对应的标记序列。

2．概率计算问题。给定模型$\lambda =(A,B,\pi )$和观测序列$O＝(o_1，o_2,\dots ,o_T)$，计算在模型$\lambda$下观测序列$O$出现的概率$P(O|\lambda )$。前向-后向算法是通过递推地计算前向-后向概率可以高效地进行隐马尔可夫模型的概率计算。

3．学习问题。已知观测序列$O＝(o_1，o_2,\dots ,o_T)$，估计模型$\lambda =(A,B,\pi )$参数，使得在该模型下观测序列概率$P(O|\lambda )$最大。即用极大似然估计的方法估计参数。Baum-Welch算法，也就是EM算法可以高效地对隐马尔可夫模型进行训练。它是一种非监督学习算法。

4．预测问题。已知模型$\lambda =(A,B,\pi )$和观测序列$O＝(o_1，o_2,\dots ,o_T)$，求对给定观测序列条件概率$P(I|O)$最大的状态序列$I＝(i_1，i_2,\dots ,i_T)$。维特比算法应用动态规划高效地求解最优路径，即概率最大的状态序列。