统计学习方法第十章——隐马尔可夫模型

10.2 概率计算算法

10.2.1 直接计算法

$$P(I\mid \lambda )=P\left(i_1,i_2,\dots ,i_T\mid \lambda \right)=P\left(i_T\mid i_1,i_2,\dots ,i_{T-1},\lambda \right)P\left(i_1,i_2,\dots ,i_{T-1}\mid \lambda \right)$$

根据齐次一阶马尔可夫假设： $P\left(i_T\mid i_1,i_2,\dots ,i_{T-1},\lambda \right)=P\left(i_T\mid i_{T-1},\lambda \right)=a_{i_{T-1},i_T}$所以
$$\begin{array}{rl}P(I|\lambda )& =a_{i_{T-1},i_T}P(i_1,i_2,\cdots ,i_{T-1}|\lambda )\\ & =a_{i_{T-1},i_T}{a}_{i_{T-2},i_{T-1}}P(i_1,i_2,\cdots ,i_{T-2}|\lambda )\\ & ={\pi }_{i_1}\prod _{t=2}^T{a}_{i_{t-1},i_t}\end{array}$$
又因为：

$P(O\mid I,\lambda )=P\left(o_1,o_2,\dots ,o_T\mid i_1,i_2,\dots ,i_T,\lambda \right)=P\left(o_T\mid o_1,o_2,\dots o_{T-1},i_1,i_2,\dots ,i_T,\lambda \right)P\left(o_1,o_2,\dots o_{T-1}\mid i_1,i_2,\dots i_T,\lambda \right)$

根据观察独立性假设
$$\begin{array}{rl}P(O|I,\lambda )& =P(o_T|i_T)P(o_1,o_2,\cdots ,o_{T-1}|i_1,i_2,\cdots ,i_T,\lambda )\\ & =b_{i_1}(o_1)P(o_1,o_2,\cdots ,o_{T-1}|i_1,i_2,\cdots ,i_T,\lambda )\\ & =b_{i_1}(o_1)b_{i_2}(o_2)P(o_1,o_2,\cdots ,o_{T-2}|i_1,i_2,\cdots ,i_T,\lambda )\\ & =\prod _{t=1}^T{b}_{i_t}(o_t)\end{array}$$
所以$O和I$同时出现的联合概率为：
$$\begin{array}{rl}P(O,I\mid \lambda )& =P(O\mid I,\lambda )P(I\mid \lambda )\\ & ={\pi }_{i_1}{b}_{i_1}\left(o_1\right)a_{i_1{i}_2}{b}_{i_2}\left(o_2\right)\cdots a_{i_{T-1}{i}_T}{b}_{i_T}\left(o_T\right)\end{array}$$
然后, 对所有可能的状态序列 $I$ 求和, 得到观测序列 $O$ 的概率 $P(O\mid \lambda )$, 即
$$\begin{array}{rl}P(O\mid \lambda )& =\sum _{I}P(O\mid I,\lambda )P(I\mid \lambda )\\ & =\sum _{i_1,i_2,\cdots ,i_T}{\pi }_{i_1}{b}_{i_1}\left(o_1\right)a_{i_1{i}_2}{b}_{i_2}\left(o_2\right)\cdots a_{i_{T-1}{i}_T}{b}_{i_T}\left(o_T\right)\end{array}$$

10.2.2 前向算法

首先，根据贝叶斯公式又有：
$$\begin{array}{rl}& P(A,B,C)=P(A)P(B|A)P(C|A,B)\\ & P(A,B,C|\lambda )=P(A|\lambda )P(B|A,\lambda )P(C|A,B,\lambda )\\ & P(A|\lambda )=\sum _{B}P(A,B|\lambda )\end{array}$$

$$\begin{array}{rl}{\alpha }_{t+1}(i)& =P(o_1,\cdots ,o_{t+1},i_{t+1}=q_i|\lambda )\\ & =\sum _{j=1}^{N}P\left(o_1,\dots ,o_{t+1},i_t=q_j,i_{t+1}=q_i\mid \lambda \right)\\ & =\sum _{j=1}^{N}P\left(o_1,\dots ,o_t,i_t=q_j\mid \lambda \right)P\left(i_{t+1}=q_i\mid o_1,\dots ,o_t,i_t=q_j,\lambda \right)P\left(o_{t+1}\mid o_1,\dots ,o_t,i_t=q_j,i_{t+1}=q_i,\lambda \right)\\ & =\sum ^{N}P\left(o_1,\dots ,o_t,i_t=q_j\mid \lambda \right)P\left(i_{t+1}=q_i\mid i_t=q_j,\lambda \right)P\left(o_{t+1}\mid i_{t+1}=q_i,\lambda \right)\\ & =\sum _{j=1}^N{\alpha }_t(j)a_{ji}{b}_i(o_{t+1}),i=1,2,\cdots ,N\end{array}$$

10.2.3 后向算法

$$\begin{array}{rl}{\beta }_t(i)& =P(o_{t+1},o_{t+2},\cdots ,o_T|i_t=q_i,\lambda )\\ & =\sum _{j=1}^{N}P(o_{t+1},o_{t+2},\cdots ,o_T,i_{t+1}=q_j|i_t=q_i,\lambda )\\ & =\sum _{j=1}^{N}P(i_{t+1}=q_j|i_t=q_i,\lambda )P(o_{t+2},\cdots ,o_T|i_{t+1}=q_j,i_t=q_i,\lambda )P(o_{t+1}|o_{t+2},\cdots ,o_T,i_{t+1}=q_j,i_t=q_i,\lambda )\\ & =\sum _{j=1}^{N}P(i_{t+1}=q_j|i_t=q_i,\lambda )P(o_{t+2},\cdots ,o_T|i_{t+1}=q_j,\lambda )P(o_{t+1}|i_{t+1}=q_j,\lambda )\\ & =\sum _{j=1}^N{a}_{ij}{\beta }_{t+1}(j)b_j(o_{t+1})\\ & =\sum _{j=1}^N{a}_{ij}{b}_j(o_{t+1}){\beta }_{t+1}(j)\end{array}$$

合并公式:
$$\begin{array}{rl}P(O\mid \lambda )& =\sum _{i=1}^N\sum _{j=1}^N{\alpha }_t(i)a_{ij}{b}_j\left(o_{t+1}\right){\beta }_{t+1}(j)\\ & =\sum _{i=1}^N{\alpha }_t(i)\sum _{j=1}^N{a}_{ij}{b}_j\left(o_{t+1}\right){\beta }_{t+1}(j)\\ & =\sum _{i=1}^N{\alpha }_t(i){\beta }_t(i)\end{array}$$
又因为
$$\begin{array}{rl}{\alpha }_t(i){\beta }_t(i)& =P\left(o_1,o_2,\dots ,o_t,i_t=q_i\mid \lambda \right)P\left(o_{t+1},\dots ,o_T\mid i_t=q_i,\lambda \right)\\ & =P\left(o_1,o_2,\dots ,o_t,i_t=q_i\mid \lambda \right)P\left(o_{t+1},\dots ,o_T\mid o_1,o_2,\dots ,o_t,i_t=q_i,\lambda \right)\\ & =P\left(o_1,o_2,\dots ,o_T,i_t=q_i\mid \lambda \right)\\ & =P\left(O,i_t=q_i\mid \lambda \right)\end{array}$$
所以
$$\sum _{i=1}^N{\alpha }_t(i){\beta }_t(i)=\sum _{i=1}^{N}P(O,i_t=q_i|\lambda )=P(O|\lambda )$$

10.3 学习算法

按照Q函数的定义：
$$\begin{array}{rl}Q(\lambda ,\overline{\lambda })& =E_I[logP(O,I|\lambda )|O,\overline{\lambda }]\\ & =\sum _{I}P(I|O,\overline{\lambda })logP(O,I|\lambda )\\ & =\sum _I\frac{P(O,I|\overline{\lambda })}{P(O|\overline{\lambda })}logP(O,I|\lambda )\end{array}$$
略去对$\lambda$而言的常数因子$\frac{1}{P(O|\overline{\lambda })}$，于是得到式子(10.33)
$$Q(\lambda ,\overline{\lambda })=\sum _{I}P(O,I|\overline{\lambda })logP(O,I|\lambda )$$
式子(10.35)求偏导得结果是：
$$\frac{P(O,i_1=i|\overline{\lambda })}{{\pi }_i}+\gamma =0$$
然后两边同时乘以${\pi }_i$得到书上的结果：
$$P(O,i_1=i|\overline{\lambda })+\gamma {\pi }_i=0$$
式子(10.37)按照上面的方法计算一遍：

注意到 $a_{ij}$ 满足约束条件 ${\sum }_{j=1}^N{a}_{ij}=1$, 利用拉格朗日乘子法, 写出拉格朗日函数:
$$\sum _{i=1}^N\sum _{j=1}^N\sum _{t=1}^{T-1}log{a}_{ij}P(O,i_t=i,i_{t+1}=j|\overline{\lambda })+\beta (\sum _{j=1}^N{a}_{ij}-1)$$
对其求偏导数并令结果为0得
$$\begin{gathered} \sum _{t=1}^{T-1}\frac{1}{a_{ij}}P(O,i_t=i,i_{t+1}=j|\overline{\lambda })+\beta =0 \\ \sum _{t=1}^{T-1}P(O,i_t=i,i_{t+1}=j|\overline{\lambda })+\beta a_{ij}=0 \\ {a}_{ij}=-\frac{1}{\beta }\sum _{t=1}^{T-1}P(O,i_t=i,i_{t+1}=j|\overline{\lambda }) \\ \sum _{j=1}^N{a}_{ij}=\sum _{j=1}^N-\frac{1}{\beta }\sum _{t=1}^{T-1}P(O,i_t=i,i_{t+1}=j|\overline{\lambda }) \\ \beta =-\sum _{t=1}^{T-1}P(O,i_t=i|\overline{\lambda }) \end{gathered}$$
然将得到的$\beta$代入最开始那个式子就得到(10.37)
$$a_{ij}=\frac{{\sum }_{t=1}^{T-1}P\left(O,i_t=i,i_{t+1}=j\mid \bar{\lambda }\right)}{{\sum }_{t=1}^{T-1}P\left(O,i_t=i\mid \bar{\lambda }\right)}$$
同样对式子(10.38)进行一样的操作，同样用拉格朗日乘子法, 约束条件是 ${\sum }_{k=1}^M{b}_j(k)=1$ 。注意, 只有在 $o_t=v_k$ 时 $b_j\left(o_t\right)$ 对 $b_j(k)$ 的偏导数才不为 0 , 以 $I\left(o_t=v_k\right)$ 表示。求得拉格朗日函数为：
$$\sum _{j=1}^N\sum _{t=1}^{T}log{b}_j(o_t)P(O,i_t=j|\overline{\lambda })+\eta (\sum _{k=1}^M-1)=0$$
对其求偏导数并令结果为0
$$\begin{gathered} \sum _{t=1}^T\frac{P(O,i_t=j|\overline{\lambda })I(o_t=v_k)}{b_j(o_k)}+\eta =0 \\ {b}_j(k)=-\frac{1}{\eta }\sum _{t=1}^{T}P(O,i_t=j|\overline{\lambda })I(o_t=v_k) \\ \eta =-\sum _{t=1}^T\sum _{k=1}^{M}P(O,i_t=j|\overline{\lambda })I(o_t=v_k) \\ \eta =-\sum _{t=1}^{T}P(O,i_t=j|\overline{\lambda }) \end{gathered}$$
将$\eta$代回上式得式子(10.38)
$$b_j(k)=\frac{{\sum }_{t=1}^{T}P\left(O,i_t=j\mid \bar{\lambda }\right)I\left(o_t=v_k\right)}{{\sum }_{t=1}^{T}P\left(O,i_t=j\mid \bar{\lambda }\right)}$$
下面对(10.39)~(10.41)进行推导
$$\begin{array}{rl}{a}_{ij}& =\frac{{\sum }_{t=1}^{T-1}P\left(O,i_t=i,i_{t+1}=j\mid \bar{\lambda }\right)}{{\sum }_{t=1}^{T-1}P\left(O,i_t=i\mid \bar{\lambda }\right)}\\ & =\frac{{\sum }_{t=1}^{T-1}P(O,i_t=i,i_{t+1}=j|\overline{\lambda })\frac{1}{P(O|\overline{\lambda })}}{{\sum }_{t=1}^{T-1}P(O,i_t=i|\overline{\lambda })\frac{1}{P(O|\overline{\lambda })}}\\ & =\frac{{\sum }_{t=1}^{T-1}{\zeta }_t(i,j)}{{\sum }_{t=1}^{T-1}{\gamma }_t(i)}\end{array}$$

$$\begin{array}{rl}{b}_j(k)& =\frac{{\sum }_{t=1}^{T}P(O,i_t=j|\overline{\lambda })I(o_t=v_k)}{{\sum }_{t=1}^{T}P(O,i_t=j|\overline{\lambda })}\\ & =\frac{{\sum }_{t=1,o_t=v_k}^{T}P(O,i_t=j|\overline{\lambda })\frac{1}{P(O|\overline{\lambda })}}{{\sum }_{t=1}^{T}P(O,i_t=j|\overline{\lambda })\frac{1}{P(O|\overline{\lambda })}}\\ & =\frac{{\sum }_{t=1,o_t=v_k}^T{\gamma }_t(j)}{{\sum }_{t=1}^T{\gamma }_t(j)}\end{array}$$

$${\pi }_i=\frac{P(O,i_1=i|\overline{\lambda })}{P(O|\overline{\lambda })}={\gamma }_1(i)$$