HMM(隐马尔科夫模型的预测算法)

UTF8gbsn

The predication of HMM has two different methods. the approximation way
and the Viterbi algorithm. The predication problem of HMM is to find the
state sequence based on $\lambda =(A,B,\pi )$ and observation
$O=(o_1,o_2,\cdots ,o_T)$

approximation way

The probability of $i_t=q_i$ when we have
$\lambda =(A,B,\pi ),O=\left(o_1,o_2,\cdots ,o_T\right)$ is
${\gamma }_t=i$.
$${\gamma }_t(i)=\frac{{\alpha }_t(i){\beta }_t(i)}{P(O|\lambda )}=\frac{{\alpha }_t(i){\beta }_t(i)}{{\sum }_{j=1}^N{\alpha }_t(j){\beta }_t(j)}$$

Now, the most likely state $i_t^{\ast }$ of time t is :
$$i_t^{\ast }=\arg \underset{1⩽i⩽N}{\max }[{\gamma }_t(i)],t=1,2,\cdots ,T$$

Finaly, we can get a state sequence
$I^{\ast }=\left(i_1^{\ast },i_2^{\ast },\cdots ,i_T^{\ast }\right)$. One drawback
of this approximation algorithm is that the generated state sequence
might be impossible. For example, $a_{ij}=0$ for some $i,j$.

Viterbi

The Viterbi algorithm uses dynamic programming approach to solve the HMM
predication problem. Let’s first introduce two variables $\delta ,\Psi$.

$${\delta }_t(i)=\underset{i_1,i_2,\cdots ,i_{t-1}}{\max }P(i_t=i,i_{t-1},\cdots ,i_1,o_t,\cdots ,o_1|\lambda ),i=1,2,\cdots ,N$$

Now, the recursion of $\delta$ is:

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

The definition of $\Psi$ is :

$${\Psi }_t(i)=\arg \underset{1⩽j⩽N}{\max }[{\delta }_{t-1}(j)a_{ji}],i=1,2,\cdots ,N$$

At last, we can develop the Viterbi algorithm as bellow:

• Input: $\lambda =(A,B,\pi )$ and the observation sequence
  $O=\left(o_1,o_2,\cdots ,o_T\right)$;

• Output: The optimised path
  $I^{\ast }=\left(i_1^{\ast },i_2^{\ast },\cdots ,i_T^{\ast }\right)$.

Alg:

1. Initialization:

   $${\delta }_1(i)={\pi }_i{b}_i(o_1).i=1,2,\cdots ,N$$

   $${\Psi }_1(i)=0,i=1,2,\cdots ,N$$

2. Recursion, $t=2,3,\cdots ,T$

   $${\delta }_t(i)=\underset{1\le i\le N}{\max }\left[{\delta }_{t-1}(j)a_{ji}\right]b_i\left(o_t\right),i=1,2,\cdots ,N$$

   $${\Psi }_t(i)=\arg \underset{1⩽j⩽N}{\max }\left[{\delta }_{t-1}(j)a_{ji}\right],i=1,2,\cdots ,N$$

3. Stop:

   $$P^{\ast }=\underset{1⩽i⩽N}{\max }{\delta }_T(i)$$

   $$i_T^{\ast }=\arg \underset{1⩽i⩽N}{\max }\left[{\delta }_T(i)\right]$$

4. Backtracking

   $$i_t^{\ast }={\Psi }_{t+1}\left(i_{t+1}^{\ast }\right)$$

Finally, we will get the best path
$I^{\ast }=\left(i_1^{\ast },i_2^{\ast },\cdots ,i_T^{\ast }\right)$

Example

For example, $\lambda =(A,B,\pi )$

$$A=\left[\begin{array}{ccc}0.5& 0.2& 0.3\\ 0.3& 0.5& 0.2\\ 0.2& 0.3& 0.5\end{array}\right],B=\left[\begin{array}{cc}0.5& 0.5\\ 0.4& 0.6\\ 0.7& 0.3\end{array}\right],\pi =\left[\begin{array}{c}0.2\\ 0.4\\ 0.4\end{array}\right]$$

The observation $O=(red,white,red)$, what is the best path
$I^{\ast }=\left(i_1^{\ast },i_2^{\ast },i_3^{\ast }\right)$.

1. Initialization:
   $${\delta }_1(1)=0.4\times 0.5=0.10,{\delta }_1(2)=0.4\times 0.4=0.16,{\delta }_1(3)=0.4\times 0.7=0.28$$
   $${\Psi }_1(i)=0,i=1,2,3$$

2. $t=2$ δ 2 ( 1 ) = max ⁡ 1 ⩽ j ⩽ 3 [ δ 1 ( j ) a j 1 ] b 1 ( o 2 ) = max ⁡ j { 0.10 × 0.5 , 0.16 × 0.3 , 0.28 × 0.2 } × 0.5 = 0.028 Ψ 2 ( 1 ) = 3 δ 2 ( 2 ) = 0.0504 Ψ 2 ( 2 ) = 3 δ 2 ( 3 ) = 0.042 Ψ 2 ( 3 ) = 3 \begin{aligned} \delta_{2}(1) &=\max _{1 \leqslant j \leqslant 3}\left[\delta_{1}(j) a_{j 1}\right] b_{1}\left(o_{2}\right) \\ &=\max _{j}\{0.10 \times 0.5,0.16 \times 0.3,0.28 \times 0.2\} \times 0.5 \\ &=0.028 \\ \Psi_{2}(1) &=3 \\ \delta_{2}(2) &=0.0504 \\ \Psi_{2}(2) &=3 \\ \delta_{2}(3) &=0.042 \\ \Psi_{2}(3) &=3 \end{aligned} δ2​(1)Ψ2​(1)δ2​(2)Ψ2​(2)δ2​(3)Ψ2​(3)​=1⩽j⩽3max​[δ1​(j)aj1​]b1​(o2​)=jmax​{0.10×0.5,0.16×0.3,0.28×0.2}×0.5=0.028=3=0.0504=3=0.042=3​

3. $t=3$ $$\begin{array}{l}{\delta }_3(i)={\max }_{1⩽j⩽3}\left[{\delta }_2(j)a_{ji}\right]b_i\left(o_3\right)\\ {\Psi }_3(i)=\arg {\max }_{1⩽j⩽3}\left[{\delta }_2(j)a_{ji}\right]\\ {\delta }_3(1)=0.00756\\ {\Psi }_3(1)=2\\ {\delta }_3(2)=0.01008\\ {\Psi }_3(2)=2\\ {\delta }_3(3)=0.0147\\ {\Psi }_3(3)=3\end{array}$$

4. Stop:
   $$P^{\ast }=\underset{1⩽i⩽3}{\max }{\delta }_3(i)=0.0147$$
   $$i_3^{\ast }=\arg \underset{i}{\max }\left[{\delta }_3(i)\right]=3$$

5. Backtracking: $$\begin{array}{r}{i}_2^{\ast }={\Psi }_3\left(i_3^{\ast }\right)={\Psi }_3(3)=3\\ i_1^{\ast }={\Psi }_2\left(i_2^{\ast }\right)={\Psi }_2(3)=3\end{array}$$

At last, we get the best path
$I^{\ast }=\left(i_1^{\ast },i_2^{\ast },i_3^{\ast }\right)=(3,3,3)$