HMM(监督学习)

UTF8gbsn

The subject of today is the supervised learning of HMM
$\lambda =(A,B,\pi )$. The task of the supervised learning of HMM is to
estimate $\lambda =(A,B,\pi )$ from these known observation sequences $O$
and their corresponding state sequences $I$.

Information

• $(O_1,I_1),(O_2,I_2),\cdots ,(O_S,I_S)$

Quesiton: Estimating $\lambda =(A,B,\pi )$

1. Estimating transfer probability $a_{ij}$:Let $A_{ij}$ be the
   frequence of time $t$ with state $i$ and $t+1$ with state $j$, then
   $a_{ij}$ is estimated by the following formula.
   $${\hat{a}}_{ij}=\frac{A_{ij}}{{\sum }_{j=1}^N{A}_{ij}}$$

2. Estimating the observation probability $b_j(k)$: Let $B_{jk}$ be
   the state $j$ with observation $k$, then $b_j(k)$ is estimated by
   the following formula.

   $${\hat{b}}_j(k)=\frac{B_{jk}}{{\sum }_{j=1}^N{B}_{jk}}$$

3. Estimating the initial probability ${\pi }_i$: The estimation of
   ${\pi }_i$ is called ${\hat{\pi }}_i$. It comes from the frequence of
   state $q_i$ of the first state.