HMM隐马尔可夫模型的数学推导（二）

前言

上一篇HMM高斯混合模型参数估计我们对learning问题进行了数学推导，但是由于本身隐马尔可夫模型的复杂性。即便算出了迭代式的公式，等号右边的概率该如何计算仍然式一个问题
数学基础：【概率论与数理统计知识复习-哔哩哔哩】

数学推导

Evaluation：

该问题就是要求出$P(X|\theta )$。

对于该问题，前面我们求Learning问题时，曾求出过
$$P(X|\theta )=\sum _Z\pi \prod _{i=1}^{T-1}{a}_{(i,i+1)}\prod _{j=1}^T{b}_{(j,j)}$$
这种解法的复杂度太高。因此引入复杂度相对较低的算法（前向算法和后向算法）。

前向算法(递推式算法)

$$\begin{array}{c}\begin{array}{rl}P(X|\theta )=& P(x_1,x_2,\cdots ,x_T|\theta )\\ =& \sum _{z_T}P(x_1,x_2,\cdots ,x_T,z_T|\theta )\\ =& \sum _{i=1}^{n}P(x_1,x_2,\cdots ,x_T,z_T=q_i|\theta )\end{array}\end{array}$$

令${\alpha }_{(z_t=q_i)}=P(x_1,x_2,\cdots ,x_t,z_t=q_i|\theta )$。注意该式子就是上面learning问题时所得。

找出其递推式，对于${\alpha }_{(z_{t+1}=q_i)}$
$$\begin{array}{c}\begin{array}{rl}{\alpha }_{(z_{t+1}=q_i)}=& P(x_1,x_2,\cdots ,x_t,x_{t+1},z_{t+1}=q_i|\theta )\\ =& P(x_{t+1}|x_1,\cdots x_t,z_{t+1}=q_i,\theta )P(x_1,\cdots x_t,z_{t+1}=q_i|\theta )\\ =& P(x_{t+1}|z_{t+1}=q_i)P(x_1,\cdots x_t,z_{t+1}|\theta )\\ =& b_{(z_{t+1}=q_i,x_{t+1})}\sum _{z_t}P(x_1,\cdots x_t,z_t,z_{t+1}=q_i|\theta )\\ =& b_{(z_{t+1}=q_i,x_{t+1})}\sum _{j=1}^{n}P(x_1,\cdots x_t,z_t=q_j,z_{t+1}=q_i|\theta )\\ =& b_{(z_{t+1}=q_i,x_{t+1})}\sum _{j=1}^{n}P(z_{t+1}=q_i|x_1,x_2,\cdots ,x_t,z_t=q_j,\theta )P(x_1,x_2,\cdots ,x_t,z_t=q_j|\theta )\\ =& b_{(z_{t+1}=q_i,x_{t+1})}\sum _{j=1}^{n}P(z_{t+1}=q_i|x_1,x_2,\cdots ,x_t,z_t=q_j,\theta )P(x_1,x_2,\cdots ,x_t,z_t=q_j|\theta )\\ =& \sum _{j=1}^n{b}_{(z_{t+1}=q_i,x_{t+1})}{a}_{(z_t=q_j,z_{t+1}=q_i)}{\alpha }_{(z_t=q_j)}\end{array}\end{array}$$
请注意区分$\alpha$和$a$

所以，我们找到了${\alpha }_{(z_t=q_i)}(x_t)$和${\alpha }_{(z_{t+1}=q_i)}(x_{t+1})$的关系式。然后就可以递归地完成实现上面的$P(X|\theta )$
$$P(X|\theta )=\sum _{i=1}^n{\alpha }_{(z_T=q_i)}$$

后向算法(递推式算法)

$$\begin{array}{c}\begin{array}{rl}P(X|\theta )=& P(x_1,x_2,\cdots ,x_T|\theta )\\ =& \sum _{z_1}P(x_1,x_2,\cdots ,x_T,z_1|\theta )\\ =& \sum _{z_1}P(x_1|x_2,\cdots ,x_T,z_1,\theta )P(x_2,\cdots ,x_T,z_1|\theta )\\ =& \sum _{z_1}{b}_{(z_1,x_1)}P(x_2,\cdots ,x_T,z_1|\theta )\\ =& \sum _{z_1}{b}_{(z_1,x_1)}P(x_2,\cdots ,x_T|z_1,\theta )P(z_1|\theta )\\ =& \sum _{i=1}^n{b}_{(z_1=q_i,x_1)}P(x_2,\cdots ,x_T|z_1=q_i,\theta ){\pi }_i\end{array}\end{array}$$

令${\beta }_{(z_t=q_i)}=P(x_{t+1},\cdots ,x_T|z_t=q_i,\theta )$。

找出其递推式，对于${\beta }_{(z_{t-1}=q_i)}$
$$\begin{array}{c}\begin{array}{rl}{\beta }_{(z_{t-1}=q_i)}=& P(x_t,\cdots ,x_T|z_{t-1}=q_i,\theta )\\ =& \sum _{z_t}P(x_t,\cdots ,x_T,z_t|z_{t-1}=q_i,\theta )\\ =& \sum _{z_t}P(x_t,\cdots ,x_T|z_t,z_{t-1}=q_i,\theta )P(z_t|z_{t-1}=q_i,\theta )\\ =& \sum _{z_t}P(x_t,\cdots ,x_T|z_t,z_{t-1}=q_i,\theta )a_{(z_{t-1}=q_i,z_t)}\\ =& \sum _{z_t}P(x_t,\cdots ,x_T|z_t,\theta )a_{(z_{t-1}=q_i,z_t)}\\ =& \sum _{z_t}P(x_t|x_{t+1},\cdots ,x_T,z_t,\theta )P(x_{t+1},\cdots ,x_T|z_t,\theta )a_{(z_{t-1}=q_i,z_t)}\\ =& \sum _{z_t}p(x_t|z_t)P(x_{t+1},\cdots ,x_T|z_t,\theta )a_{(z_{t-1}=q_i,z_t)}\\ =& \sum _{j=1}^n{b}_{(z_t=q_j,x_t)}{\beta }_{(z_t=q_j)}{a}_{(z_{t-1}=q_i,z_t=q_j)}\end{array}\end{array}$$
这仍然是一个递归式算法。和前向算法一个道理。所以
$$P(X|\theta )=\sum _{i=1}^n{b}_{(z_1=q_i,x_1)}{\beta }_{(z_1=q_i)}{\pi }_i$$

接下来就找出learning问题中如何对概率求解
$$\begin{array}{c}\begin{array}{rl}& P(z_t=q_i,X|{\theta }^t)\\ =& P(z_t=q_i,x_1,x_2,\cdots ,x_T|{\theta }^t)\\ =& P(x_{t+1},\cdots x_T|z_t=q_i,x_1,\cdots ,x_t,{\theta }^t)P(z_t=q_i,x_1,\cdots ,x_t|{\theta }^t)\\ =& P(x_{t+1},\cdots x_T|z_t=q_i,{\theta }^t)P(z_t=q_i,x_1,\cdots ,x_t|{\theta }^t)\\ =& {\beta }_{(z_t=q_i)}{\alpha }_{(z_t=q_i)}\end{array}\end{array}$$
从该问题可得到$P(z_1=q_i,X|{\theta }^t)$的求解。那么还剩下$\sum _t^{T-1}P(z_t=q_i,z_{t+1}=q_j,X|{\theta }^t)$
$$\begin{array}{c}\begin{array}{rl}& \sum _t^{T-1}P(z_t=q_i,z_{t+1}=q_j,X|{\theta }^t)\\ =& P(x_1,x_2,\cdots ,x_T,z_t=q_i,z_{t+1}=q_j|{\theta }^t)\\ =& P(z_{t+1}=q_j,x_{t+1},\cdots ,x_T|x_1,x_2,\cdots ,x_t,z_t=q_i,{\theta }^t)P(x_1,x_2,\cdots ,x_t,z_t=q_i|{\theta }^t)\\ =& P(z_{t+1}=q_j,x_{t+1},\cdots ,x_T|x_1,x_2,\cdots ,x_t,z_t=q_i,{\theta }^t){\alpha }_{(z_t=q_i)}\\ =& P(x_{t+2},\cdots ,x_T|z_{t+1}=q_j,x_1,x_2,\cdots ,x_{t+1},z_t=q_i,{\theta }^t)P(x_{t+1},z_{t+1}=q_j|x_1,x_2,\cdots ,x_t,z_t=q_i,{\theta }^t){\alpha }_{(z_t=q_i)}\\ =& P(x_{t+2},\cdots ,x_T|z_{t+1}=q_i,{\theta }^t)P(x_{t+1},z_{t+1}=q_j|x_1,x_2,\cdots ,x_t,z_t=q_i,{\theta }^t){\alpha }_{(z_t=q_i)}\\ =& {\beta }_{(z_{t+1}=q_j)}P(x_{t+1},z_{t+1}=q_j|x_1,x_2,\cdots ,x_t,z_t=q_i,{\theta }^t){\alpha }_{(z_t=q_i)}\\ =& {\beta }_{(z_{t+1}=q_j)}P(x_{t+1}|z_{t+1}=q_j,x_1,x_2,\cdots ,x_t,z_t=q_i,{\theta }^t)P(z_{t=1}=q_j|x_1,x_2,\cdots ,x_t,z_t=q_i,{\theta }^t){\alpha }_{(z_t=q_i)}\\ =& {\beta }_{(z_{t+1}=q_j)}{b}_{(z_{t+1}=q_j,x_{t+1})}{a}_{(z_t=q_i,z_{t+1}=q_j)}{\alpha }_{(z_t=q_i)}\end{array}\end{array}$$
所以，最终在Learning中所得结果应当是
$$\begin{gathered} {\pi }_i=\frac{{\beta }_{(z_1=q_i)}{\alpha }_{(z_1=q_i)}}{P(X|{\theta }^t)}; \\ {a}_{(z=q_i,z=q_j)}=\frac{\sum _{t=1}^{T-1}{\beta }_{(z_{t+1}=q_j)}{b}_{(z_{t+1}=q_j,x_{t+1})}{a}_{(z_t=q_i,z_{t+1}=q_j)}{\alpha }_{(z_t=q_i)}}{\sum _{t=1}^{T-1}{\beta }_{(z_t=q_i)}{\alpha }_{(z_t=q_i)}}; \\ {b}_{(z=q_i,x=v_j)}=\frac{\sum _{t=1}^{T-1}{\beta }_{(z_t=q_i)}{\alpha }_{(z_t=q_i)}I(x_t=v_j)}{\sum _{t=1}^T{\beta }_{(z_t=q_i)}{\alpha }_{(z_t=q_i)}} \end{gathered}$$

Decoding：

模型训练完成之后，我们求出$P(Z|X,\theta )$以实现预测

即
$$\hat{Z}=\underset{Z}{\max }P(Z|X,\theta )$$
传统方式方法的复杂度及其之高。故引入维特比算法

维特比算法(递推式算法)

所谓维特比算法。实际上就是采用了动态规划的思想。对于原式
$$\begin{array}{c}\begin{array}{rl}\hat{Z}=& \underset{Z}{\max }P(Z|X,\theta )\\ =& \underset{Z}{\max }\frac{P(Z,X|\theta )}{P(X|\theta )}\\ =& \underset{Z}{\max }P(Z,X|\theta )\end{array}\end{array}$$
设
$${\xi }_{z_t}(q_i)=\underset{z_1,z_2,\cdots ,z_{t-1}}{\max }P(z_1,z_2,\cdots ,z_t=q_i,x_1,x_2,\cdots ,x_t|\theta )$$
其中${\xi }_{z_t}(q_i)$表示处于第t个隐序列，状态为$q_i$的能够使得概率最大的概率。

找出其递推式，对于${\xi }_{z_{t+1}}(q_i)$
ξ z t + 1 ( q i ) = max ⁡ j ∈ { 1 , 2 , ⋯   , N } ξ z t ( q j ) a ( z t = q j , z t = 1 = q i ) b ( z t = 1 = q i , x t + 1 ) \begin{equation} \begin{aligned} \xi_{z_{t+1}}(q_i)=&\max_{j\in\{1,2,\cdots,N\}}\xi_{z_t}(q_j)a_{(z_t=q_j,z_{t=1}=q_i)}b_{(z_{t=1}=q_i,x_{t+1})} \end{aligned} \end{equation} ξzt+1​​(qi​)=​j∈{1,2,⋯,N}max​ξzt​​(qj​)a(zt​=qj​,zt=1​=qi​)​b(zt=1​=qi​,xt+1​)​​​​
可理解寻找前一个使后面能够最大的前一个值。

那么，依照这种迭代式算法，我们只需要记录每一个能够使得$t+1$时刻最大的$t$时刻所对应的$q_i$即可

结束

这就是HMM隐马尔可夫模型的数学原理推导了，推导并不严谨，如有问题，还望指出。另外，对于隐马尔可夫模型的应用隐马尔可夫模型中文分词，感兴趣的可以看看，阿里嘎多。