HMM(隐马尔科夫模型的无监督学习方法)

UTF8gbsn

Induction

The main topic of this article is about Baum-Welch algorithm. We will
estimate $\lambda =(A,B,\pi )$ with { O 1 , O 2 , ⋯   , O S } \{O_1,O_2,\cdots,O_S\} {O1​,O2​,⋯,OS​} and the
lenght of $O_i$ is $T$. The object of our probability model is
$$P(O|\lambda )=\sum _{I}P(I|\lambda )P(O|I,\lambda )$$

The parameters of the upon formula can be estimated by EM algorithm.

• E step:
  $$Q(\lambda ,\overline{\lambda })=\sum _{I}P(O,I|\overline{\lambda })logP(O,I|\lambda )$$

  $$P(O,I|\lambda )={\pi }_{i_1}{b}_{i_1}(o_1)a_{i_1,i_2}{b}_{i_2}(o_2)\cdots a_{i_{T-1},i_T}{b}_{i_T}(o_T)$$

  The function $Q(\lambda ,\overline{\lambda })$ becomes:

  $$\begin{array}{rl}Q(\lambda ,\overline{\lambda })=& \sum _{I}log{\pi }_{i_1}P(O,I|\overline{\lambda })+\\ & \sum _I(\sum _{t=1}^{T-1}log{a}_{i_t,i_{t+1}})P(O,I|\overline{\lambda })+\\ & \sum _I(\sum _{t=1}^{T}log{b}_{i_t}(o_t))P(O,I|\overline{\lambda })\end{array}$$

• M step:

  1. ${\pi }_i$
     $$\sum _{I}log{\pi }_{i_1}P(O,I|\overline{\lambda })=\sum _{i=1}^{N}log{\pi }_{i}P(O,i_1=i|\overline{\lambda })$$
     There is a constraint ${\sum }_{i=1}^N{\pi }_i=1$ ,and using the
     lagrange multiplier will help us to form the lagrange function
     as follow:

     $$f(\pi ,\gamma )=\sum _{i=1}^{N}log{\pi }_{i}P(O,i_1=i|\overline{\lambda })+\gamma (\sum _{i=1}^N{\pi }_i=1)$$

     The partial derivative of $f(\pi ,\gamma )$ is:
     $$\frac{\partial }{\partial {\pi }_i}\left[\sum _{i=1}^N\log {\pi }_{i}P\left(O,i_1=i\mid \bar{\lambda }\right)+\gamma \left(\sum _{i=1}^N{\pi }_i-1\right)\right]=0$$
     We simplify the upon formula as :

     $$P\left(O,i_1=i\mid \bar{\lambda }\right)+\gamma {\pi }_i=0$$

     From
     ${\sum }_{i=1}^N[P\left(O,i_1=i\mid \bar{\lambda }\right)+\gamma {\pi }_i]=0$,
     we can get $\gamma =-P(O|\overline{\lambda })$

     Finally, we get the following formula which is:
     $${\pi }_i=\frac{P(O,i_1=i|\overline{\lambda })}{P(O|\overline{\lambda })}$$

  2. $a_{ij}$

     $$\sum _I(\sum _{t=1}^{T-1}log{a}_{i_t,i_{t+1}})P(O,I|\overline{\lambda })=\sum _{i=1}^N\sum _{j=1}^N\sum _{t=1}^{T-1}\log a_{ij}P\left(O,i_t=i,i_{t+1}=j\mid \bar{\lambda }\right)$$

     Constraints of this function are
     ${\sum }_{j=1}^N{a}_{1j}=1$,${\sum }_{j=1}^N{a}_{2j}=1$,$\cdots$,${\sum }_{j=1}^N{a}_{Nj}=1$
     ,and our final object function will be :
     $$f(A,\gamma )=\sum _{i=1}^N\sum _{j=1}^N\sum _{t=1}^{T-1}\log a_{ij}P\left(O,i_t=i,i_{t+1}=j\mid \bar{\lambda }\right)+\sum _{i=1}^N{\gamma }_i(\sum _{j=1}^N{a}_{ij}-1)$$

     $$\frac{\partial f(A,\gamma )}{\partial a_{ij}}=\sum _{t=1}^{T-1}P(O|i_t=i,i_{t+1}=j|\overline{\lambda })+a_{ij}{\gamma }_i=0$$

     $$\sum _{j=1}^N(\sum _{t=1}^{T-1}P(O|i_t=i,i_{t+1}=j|\overline{\lambda })+a_{ij}{\gamma }_i)\Rightarrow {\gamma }_i=-\sum _{t=1}^{T-1}P(O,i_t=i|\overline{\lambda })$$

     Finally, we get our $a_{ij}$:
     $$a_{ij}=\frac{{\sum }_{t=1}^{T-1}P\left(O,i_t=i,i_{t+1}=j\mid \bar{\lambda }\right)}{{\sum }_{t=1}^{T-1}P\left(O,i_t=i\mid \bar{\lambda }\right)}$$

  3. $b_j(k)$
     $$\sum _I\left(\sum _{t=1}^T\log b_{i_t}\left(o_t\right)\right)P(O,I\mid \bar{\lambda })=\sum _{j=1}^N\sum _{t=1}^T\log b_j\left(o_t\right)P\left(O,i_t=j\mid \bar{\lambda }\right)$$
     Contraints of our function are
     ${\sum }_{k=1}^M{b}_1(k)=1$,${\sum }_{k=1}^M{b}_2(k)=1$,$\cdots$,
     ${\sum }_{k=1}^M{b}_N(k)=1$, and our final object function is:
     $$f(B,\gamma )=\sum _{j=1}^N\sum _{t=1}^T\log b_j\left(o_t\right)P\left(O,i_t=j\mid \bar{\lambda }\right)+\sum _{j=1}^N{\gamma }_j(\sum _{k=1}^M{b}_j(k)-1)$$
     A critical point of this derivative is $o_t=v_k$.
     $$\frac{\partial f(B,\gamma )}{\partial b_j(k)}=\sum _{t=1}^{T}P\left(O,i_t=j\mid \bar{\lambda }\right)I\left(o_t=v_k\right)+b_j(k){\gamma }_j=0$$
     $$\sum _{k=1}^M\left[\sum _{t=1}^{T}P\left(O,i_t=j\mid \bar{\lambda }\right)I\left(o_t=v_k\right)+b_j(k){\gamma }_j\right]=0$$
     There is just one $I(o_t=v_k)=1$ which means
     $$\sum _{t=1}^{T}P\left(O,i_t=j\mid \bar{\lambda }\right)+{\gamma }_j=0\Rightarrow {\gamma }_j=-\sum _{t=1}^{T}P\left(O,i_t=j\mid \bar{\lambda }\right)$$

     Let’s substitute the uppon formula into the original derivative
     which will give us the final result:
     $$b_j(k)=\frac{{\sum }_{t=1}^{T}P\left(O,i_t=j\mid \bar{\lambda }\right)I\left(o_t=v_k\right)}{{\sum }_{t=1}^{T}P\left(O,i_t=j\mid \bar{\lambda }\right)}$$

Conclusions

1. $$a_{ij}=\frac{{\sum }_{t=1}^{T-1}{\xi }_t(i,j)}{{\sum }_{t=1}^{T-1}{\gamma }_t(i)}$$

2. $$b_j(k)=\frac{{\sum }_{t=1,o_t=v_k}^T{\gamma }_t(j)}{{\sum }_{t=1}^T{\gamma }_t(j)}$$

3. $${\pi }_i={\gamma }_1(i)$$

Baum-Welch Algorithm

• Input:$O=(o_1,o_2,\cdots ,o_T)$ is the observed sequence with length
  $T$.

• Output:$\lambda =(\mathbf{A},\mathbf{B},\pi )$

Alg:

1. Initialization: For $n=0$, we randomly choose
   $a_{ij}^{(0)},b_j(k{)}^{(0)},{\pi }_i^{(0)}$ to form our initial
   model ${\lambda }^{(0)}=\left(A^{(0)},B^{(0)},{\pi }^{(0)}\right)$

2. recursion: $n=1,2,\cdots$

   $$a_{ij}^{(n+1)}=\frac{{\sum }_{t=1}^{T-1}{\xi }_t(i,j)}{{\sum }_{t=1}^{T-1}{\gamma }_t(i)}$$

   $$b_j(k{)}^{(n+1)}=\frac{{\sum }_{t=1,o_t=v_k}^T{\gamma }_t(j)}{{\sum }_{t=1}^T{\gamma }_t(j)}$$

   $${\pi }_i^{(n+1)}={\gamma }_1(i)$$

3. stop: With some critical,we stop our algorithm and treat the last
   result
   ${\lambda }^{(n+1)}=\left(A^{(n+1)},B^{(n+1)},{\pi }^{(n+1)}\right)$ as
   our HMM model parameters.