机器学习模型LaTeX公式版：隐马尔科夫模型

最新推荐文章于 2023-03-15 15:40:40 发布

anch3or

最新推荐文章于 2023-03-15 15:40:40 发布

阅读量697

点赞数

分类专栏：机器学习文章标签：机器学习隐马尔科夫 hmm LaTex

本文链接：https://blog.csdn.net/anch3or/article/details/62046021

版权

机器学习专栏收录该内容

2 篇文章 0 订阅

订阅专栏

状态集合

Q = {q 1, q 2, \dots, q N} ∣ ∣ Q ∣ ∣ = N

$\begin{align*} & Q=\left\{q_{1},q_{2},\ldots ,q_{N}\right\} \quad \left| Q\right| =N \end{align*}$
观测集合

V = {v 1, v 2, \dots, v M} ∣ ∣ V ∣ ∣ = M

$\begin{align*} & V=\left\{v_{1},v_{2},\ldots ,v_{M}\right\} \quad \left| V\right| =M \end{align*}$
状态序列

I = {i 1, i 2, \dots, i t, \dots, i T} i t \in Q (t = 1, 2, \dots, T)

$\begin{align*} & I=\left\{i_{1},i_{2},\ldots ,i_{t},\ldots,i_{T}\right\} \quad i_{t}\in Q \quad \left(t=1,2,\ldots,T \right)\end{align*}$
观测序列

O = {o 1, o 2, \dots, o t, \dots, o T} o t \in V (t = 1, 2, \dots, T)

$\begin{align*} & O=\left\{o_{1},o_{2},\ldots ,o_{t},\ldots,o_{T}\right\} \quad o_{t}\in V \quad \left(t=1,2,\ldots,T \right)\end{align*}$
状态转移矩阵

A = [a i j] N \times N

$\begin{align*} & A=\left[a_{ij}\right]_{N\times N} \end{align*}$
在

t $t$ 时刻处于状态

qi $q_{i}$ 的条件下，在

t+1 $t+1$ 时刻转移到状态

qj $q_{j}$ 的概率

a i j = P (i t + 1 = q j | i t = q i) (i = 1, 2, \dots, N) (j = 1, 2, \dots, M)

$\begin{align*} & a_{ij}= P\left( i_{t+1}=q_{j}|i_{t}=q_{i}\right) \quad \left(i=1,2,\ldots,N \right) \quad \left(j=1,2,\ldots,M \right)\end{align*}$
观测概率矩阵

B = [b j (k)] N \times M

$\begin{align*} & B=\left[b_{j}\left(k\right)\right]_{N\times M} \end{align*}$
在

t $t$ 时刻处于状态

qi $q_{i}$ 的条件下，生成观测

vk $v_{k}$ 的概率

b j (k) = P (o t = v k | i t = q j) (k = 1, 2, \dots, M) (j = 1, 2, \dots, N)

$\begin{align*} & b_{j}\left(k\right)= P\left( o_{t}=v_{k}|i_{t}=q_{j}\right) \quad \left(k=1,2,\ldots,M \right) \quad \left(j=1,2,\ldots,N \right)\end{align*}$
初始概率向量

π = (π i)

$\begin{align*} & \pi =\left( \pi _{i}\right) \end{align*}$
在时刻

t=1 $t=1$ 处于状态

qi $q_{i}$ 的概率

π i = P (i 1 = q i) (i = 1, 2, \dots, N)

$\begin{align*} & \pi_{i} =P\left( i_{1}=q_{i}\right) \quad \left(i=1,2,\ldots,N \right) \end{align*}$
隐马尔科夫模型

λ = (A, B . π)

$\begin{align*} & \lambda =\left( A,B.\pi \right) \end{align*}$
隐马尔科夫模型基本假设：
1. 齐次马尔科夫性假设：在任意时刻

t $t$ 的状态只依赖于时刻

t−1 $t-1$ 的状态。

P (i t | i t - 1, o t - 1, \dots, i 1, o 1) = P (i t | i t - 1) (t = 1, 2, \dots, T)

$\begin{align*} & P\left( i_{t}|i_{t-1},o_{t-1},\ldots,i_{1},o_{1}\right)=P\left(i_{t}|i_{t-1}\right) \quad \left(t=1,2,\ldots,T\right) \end{align*}$
2. 观测独立性假设：任意时刻

t $t$ 的观测只依赖于时刻

t $t$ 的状态。

P (o t | i T, o T, i T - 1, o T - 1, \dots, i t + 1, o t + 1, i t, i t - 1, o t - 1, \dots, i 1, o 1) = P (o t | i t) (t = 1, 2, \dots, T)

$\begin{align*} & P\left( o_{t}|i_{T},o_{T},i_{T-1},o_{T-1},\ldots,i_{t+1},o_{t+1},i_{t},i_{t-1},o_{t-1},\ldots,i_{1},o_{1}\right)=P\left(o_{t}|i_{t}\right) \quad \left(t=1,2,\ldots,T\right) \end{align*}$
观测序列生成算法:
输入：隐马尔科夫模型

λ=(A,B.π) $\lambda =\left( A,B.\pi \right)$ ,观测序列长度

T $T$ ;
输出：观测序列

O={o1,o2,…,ot,…,oT} $O=\left\{o_{1},o_{2},\ldots ,o_{t},\ldots,o_{T}\right\}$ ；
1. 由初始概率向量

π $\pi$ 产生状态

i1 $i_{1}$ ；
2.

t=1 $t=1$ ；
3. 由状态

it $i_{t}$ 的观测概率分布

bj(k) $b_{j}\left(k\right)$ 生成

ot $o_{t}$ ；
4. 由状态

it $i_{t}$ 的状态转移概率分布

aitit+1 $a_{i_{t}i_{t+1}}$ 生成状态

it+1(it+1=1,2,…,N) $i_{t+1} \quad \left(i_{t+1}=1,2,\ldots,N\right)$ ；
5.

t=t+1 $t=t+1$ ；如果

t<T $t<T$ ，转至3.；否则，结束。

隐马尔科夫模型的3个基本问题：
1. 概率计算：已知 $\lambda =\left( A,B,\pi \right)$ 和 $O=\left\{o_{1},o_{2},\ldots ,o_{t},\ldots,o_{T}\right\}$ ，计算 $P\left(O| \lambda \right)$
2. 学习：已知 $O=\left\{o_{1},o_{2},\ldots ,o_{t},\ldots,o_{T}\right\}$ ，计算 $\lambda^* =\arg \max P\left( O|\lambda \right)$
3. 预测（编码）：已知 $\lambda =\left( A,B.\pi \right)$ 和 $O=\left\{o_{1},o_{2},\ldots ,o_{t},\ldots,o_{T}\right\}$ ，计算 $I^* =\arg \max P\left( I|O, \lambda \right)$

前向概率

α t (i) = P (o 1, o 2, \dots, o t, i t = q i | λ)

$\begin{align*} & \alpha _{t}\left( i\right) =P\left(o_{1},o_{2},\ldots ,o_{t}, i_{t}=q_{i}| \lambda \right) \end{align*}$
给定模型

λ $\lambda$ ，时刻

t $t$ 部分观测序列为

o1,o2,…,ot $o_{1},o_{2},\ldots ,o_{t}$ 且状态为

qi $q_{i}$ 的概率。
前向概率递推计算

α t (i) = P (o 1, o 2, \dots, o t, i t = q i | λ) ＝ P (i t = q i, o t 1) = \sum j = 1 N P (i t - 1 = q j, i t = q i, o t - 1 1, o t) = \sum j = 1 N P (i t = q i, o t | i t - 1 = q j, o t - 1 1) \cdot P (i t - 1 = q j, o t - 1 1) = \sum j = 1 N P (i t = q i, o t | i t - 1 = q j) \cdot α t - 1 (j) = \sum j = 1 N P (o t | i t = q i, i t - 1 = q j) \cdot P (i t = q i | i t - 1 = q j) \cdot α t - 1 (j) = \sum j = 1 N b i (o t) \cdot a j i \cdot α t - 1 (j)

$\begin{align*} & \alpha _{t}\left( i\right) =P\left(o_{1},o_{2},\ldots ,o_{t}, i_{t}=q_{i}| \lambda \right)＝P\left(i_{t}=q_{i},o_{1}^t \right) \\ & =\sum _{j=1}^{N}P\left(i_{t-1}=q_{j},i_{t}=q_{i},o_{1}^{t-1},o_{t}\right) \\ & =\sum _{j=1}^{N}P\left(i_{t}=q_{i},o_{t}|i_{t-1}=q_{j},o_{1}^{t-1}\right)\cdot P\left(i_{t-1}=q_{j},o_{1}^{t-1} \right) \\ & =\sum _{j=1}^{N}P\left(i_{t}=q_{i},o_{t}|i_{t-1}=q_{j}\right)\cdot \alpha _{t-1}\left( j\right)\\ & =\sum _{j=1}^{N}P\left(o_{t}|i_{t}=q_{i},i_{t-1}=q_{j}\right)\cdot P\left(i_{t}=q_{i}|i_{t-1}=q_{j}\right)\cdot \alpha _{t-1}\left( j\right) \\ & =\sum _{j=1}^{N}b_{i}\left(o_{t}\right)\cdot a_{ji}\cdot \alpha _{t-1}\left( j\right)\end{align*}$
概率计算

P (O | λ) = P (o T 1 | λ) = \sum i = 1 N P (o T 1, i T = q i) = \sum i = 1 N α T (i)

$\begin{align*} & P\left(O| \lambda \right) =P\left(o_{1}^{T}| \lambda\right) \\ & = \sum_{i=1}^{N}P\left(o_{1}^{T},i_{T}=q_{i}\right)\\ & = \sum _{i=1}^{N}\alpha _{T}\left( i\right)\end{align*}$
观测序列概率计算的前向算法：
输入：隐马尔科夫模型

λ $\lambda$ ,观测序列

O $O$ ;
输出：观测序列概率

P(O|λ) $P\left(O| \lambda \right)$ ；
1. 初值

α 1 (i) = π i b i (o 1) (t = 1, 2, \dots, N)

$\begin{align*} & \alpha _{1}\left( i\right)= \pi_{i}b_{i}\left(o_{1}\right) \quad \left(t=1,2,\ldots,N\right) \end{align*}$
2. 递推对

t=1,2,…,T−1 $t=1,2,\ldots,T-1$

α t + 1 (i) = \sum j = 1 N b i (o t + 1) \cdot a j i \cdot α t (j) (t = 1, 2, \dots, N)

$\begin{align*} & \alpha _{t+1}\left( i\right) =\sum _{j=1}^{N}b_{i}\left(o_{t+1}\right)\cdot a_{ji}\cdot \alpha _{t}\left( j\right) \quad \left(t=1,2,\ldots,N\right) \end{align*}$
3. 终止

P (O | λ) = \sum j = 1 N α T (i)

$\begin{align*} & P\left(O| \lambda \right)= \sum _{j=1}^{N}\alpha _{T}\left( i\right)\end{align*}$

后向概率

β t (i) = P (o t + 1, o t + 2, \dots, o T | i t = q i λ)

$\begin{align*} & \beta_{t}\left( i\right) =P\left(o_{t+1},o_{t+2},\ldots ,o_{T}| i_{t}=q_{i} \lambda \right) \end{align*}$
给定模型

λ $\lambda$ ，时刻

t $t$ 状态为

qi $q_{i}$ 的条件下，从时刻

t+1 $t+1$ 到时刻

T $T$ 的部分观测序列为

ot+1,ot+2,…,oT $o_{t+1},o_{t+2},\ldots ,o_{T}$ 的概率。
后向概率递推计算

β t (i) = P (o t + 1, o t + 2, \dots, o T | i t = q i, λ) ＝ P (o T t + 1 | i t = q i) = P ( o T t + 1 , i t = q i ) P ( i t = q i ) = \sum N j = 1 P ( o T t + 1 , i t = q i , i t + 1 = q j ) P ( i t = q i ) = \sum j = 1 N P ( o T t + 1 | i t = q i , i t + 1 = q j ) \cdot P ( i t = q i , i t + 1 = q j ) P ( i t = q i ) = \sum j = 1 N P (o T t + 1 | i t + 1 = q j) \cdot P ( i t + 1 = q j | i t = q i ) \cdot P ( i t = q i ) P ( i t = q i ) = \sum j = 1 N P (o N t + 2, o t + 1 | i t + 1 = q j) \cdot a i j = \sum j = 1 N P (o T t + 2 | i t + 1 = q j) \cdot P (o t + 1 | i t + 1 = q j) \cdot a i j = \sum j = 1 N β t + 1 (j) \cdot b j (o t + 1) \cdot a i j

$\begin{align*} & \beta _{t}\left( i\right) =P\left(o_{t+1},o_{t+2},\ldots ,o_{T}| i_{t}=q_{i}, \lambda \right)＝P\left(o_{t+1}^T |i_{t}=q_{i}\right) \\ & =\dfrac {P\left(o_{t+1}^{T}, i_{t}=q_{i}\right)} {P\left(i_{t}=q_{i}\right)}\\ & =\dfrac {\sum_{j=1}^{N} P\left(o_{t+1}^{T},i_{t}=q_{i},i_{t+1}=q_{j}\right)}{P\left(i_{t}=q_{i}\right)}\\ & =\sum_{j=1}^{N} \dfrac {P\left(o_{t+1}^{T}|i_{t}=q_{i},i_{t+1}=q_{j}\right) \cdot P\left(i_{t}=q_{i},i_{t+1}=q_{j} \right)}{P\left(i_{t}=q_{i}\right)} \\ & = \sum_{j=1}^{N} P\left(o_{t+1}^{T}|i_{t+1}=q_{j}\right) \cdot \dfrac {P\left(i_{t+1}=q_{j}|i_{t}=q_{i}\right) \cdot P\left(i_{t}=q_{i} \right)}{P\left(i_{t}=q_{i} \right)} \\ & = \sum_{j=1}^{N} P\left(o_{t+2}^{N},o_{t+1}|i_{t+1}=q_{j}\right) \cdot a_{ij} \\ & = \sum_{j=1}^{N} P\left(o_{t+2}^{T}|i_{t+1}=q_{j}\right) \cdot P\left(o_{t+1}|i_{t+1}=q_{j}\right) \cdot a_{ij} \\ & = \sum_{j=1}^{N} \beta_{t+1}\left(j\right) \cdot b_{j}\left(o_{t+1}\right) \cdot a_{ij}\end{align*}$
概率计算

P (O | λ) = P (o T 1 | λ) = \sum i = 1 N P (o T 1, i 1 = q i) = \sum i = 1 N P (i 1 = q i) \cdot P (o 1 | i 1 = q i) \cdot P (o T 2 | i 1 = q i) = \sum i = 1 N π i b i (o 1) β 1 (i)

$\begin{align*} & P\left(O| \lambda \right) =P\left(o_{1}^{T}| \lambda\right) \\ & = \sum_{i=1}^{N}P\left(o_{1}^{T},i_{1}=q_{i}\right)\\ & = \sum_{i=1}^{N}P\left(i_{1}=q_{i}\right) \cdot P\left(o_{1}|i_{1}=q_{i}\right)\cdot P\left(o_{2}^{T}|i_{1}=q_{i}\right) \\ & = \sum_{i=1}^{N} \pi_{i} b_{i}\left(o_{1}\right) \beta_{1}\left(i\right)\end{align*}$
观测序列概率计算的后向算法：
输入：隐马尔科夫模型

λ $\lambda$ ,观测序列

O $O$ ;
输出：观测序列概率

P(O|λ) $P\left(O| \lambda \right)$ ；
1. 初值

β T (i) = 1 (t = 1, 2, \dots, N)

$\begin{align*} & \beta_{T}\left( i\right)= 1 \quad \left(t=1,2,\ldots,N\right) \end{align*}$
2. 递推对

t=T−1,T−2,…,1 $t=T-1,T-2,\ldots,1$

β t (i) = \sum j = 1 N β t + 1 (j) \cdot b j (o t + 1) \cdot a i j (t = 1, 2, \dots, N)

$\begin{align*} & \beta_{t}\left( i\right) =\sum_{j=1}^{N} \beta_{t+1}\left(j\right) \cdot b_{j}\left(o_{t+1}\right) \cdot a_{ij} \quad \left(t=1,2,\ldots,N\right) \end{align*}$
3. 终止

P (O | λ) = \sum j = 1 N π i b i (o 1) β 1 (i)

$\begin{align*} & P\left(O| \lambda \right)= \sum _{j=1}^{N}\pi_{i} b_{i}\left(o_{1}\right)\beta _{1}\left( i\right) \end{align*}$

$P \left( O | \lambda \right)$ 的前向概率、后向概率的表示

P (O | λ) ＝ P (o T 1) ＝ \sum i = 1 N \sum j = 1 N P (o t 1, o T t + 1, i t = q i, i t + 1 = q j) ＝ \sum i = 1 N \sum j = 1 N P (o t 1, i t = q i, i t + 1 = q j) P (o T t + 1 | i t + 1 = q j) = \sum i = 1 N \sum j = 1 N P (o t 1, i t = q i) P (i t + 1 = q j | i t = q i) P (o T t + 1 | i t + 1 = q j) = \sum i = 1 N \sum j = 1 N P (o t 1, i t = q i) P (i t + 1 = q j | i t = q i) P (o t + 1 | i t + 1 = q j) P (o T t + 2 | i t + 1 = q j) = \sum i = 1 N \sum j = 1 N α t (i) a i j b j (o t + 1) β t + 1 (j) t = 1, 2, \dots, T - 1

$\begin{align*} \\ & P \left( O | \lambda \right) ＝ P \left( o_{1}^{T} \right) \\ & ＝ \sum_{i=1}^{N} \sum_{j=1}^{N} P \left( o_{1}^{t}, o_{t+1}^{T}, i_{t}=q_{i}, i_{t+1}=q_{j} \right) \\ & ＝ \sum_{i=1}^{N} \sum_{j=1}^{N} P \left( o_{1}^{t}, i_{t}=q_{i}, i_{t+1}=q_{j} \right) P \left( o_{t+1}^{T} | i_{t+1}=q_{j} \right) \\ & = \sum_{i=1}^{N} \sum_{j=1}^{N} P \left( o_{1}^{t}, i_{t}=q_{i} \right) P \left( i_{t+1}=q_{j} | i_{t}=q_{i} \right) P \left( o_{t+1}^{T} | i_{t+1}=q_{j} \right) \\ & = \sum_{i=1}^{N} \sum_{j=1}^{N} P \left( o_{1}^{t}, i_{t}=q_{i} \right) P \left( i_{t+1}=q_{j} | i_{t}=q_{i} \right) P \left( o_{t+1} | i_{t+1}=q_{j} \right) P \left( o_{t+2}^{T} | i_{t+1}=q_{j} \right) \\ & = \sum_{i=1}^{N} \sum_{j=1}^{N} \alpha_{t} \left( i \right) a_{ij} b_{j} \left( o_{t+1} \right) \beta_{t+1} \left( j \right) \quad \quad \quad t=1, 2, \cdots, T-1\end{align*}$
给定模型

λ $\lambda$ 和观测

O $O$ ，在时刻

t $t$ 处于状态

qi $q_{i}$ 的概率

γ t (i) = P (i t = q i | O, λ) = P ( i t = q i , O | λ ) P ( O | λ ) = P ( i t = q i , O | λ ) \sum N j = 1 ( i t = q i , O | λ ) = P ( o t 1 , i t = q i ) P ( o T t + 1 | i t = q i ) \sum N j = 1 P ( o t 1 , i t = q i ) P ( o T t + 1 | i t = q i ) = α t ( i ) β t ( i ) \sum N j = 1 α t ( i ) β t ( i )

$\begin{align*} \\ & \gamma_{t} \left( i \right) = P \left( i_{t}=q_{i} | O, \lambda \right) \\ & = \dfrac{ P \left( i_{t}=q_{i}, O | \lambda \right) } { P \left( O | \lambda \right) } \\ & = \dfrac{ P \left( i_{t}=q_{i}, O | \lambda \right) } { \sum_{j=1}^{N} \left( i_{t}=q_{i}, O | \lambda \right) } \\ & = \dfrac{ P \left( o_{1}^{t}, i_{t}=q_{i} \right) P \left( o_{t+1}^{T}| i_{t}=q_{i} \right) } { \sum_{j=1}^{N} P \left( o_{1}^{t}, i_{t}=q_{i} \right) P \left( o_{t+1}^{T}| i_{t}=q_{i} \right) } \\ & = \dfrac{ \alpha_{t} \left( i \right) \beta_{t} \left( i \right)} { \sum_{j=1}^{N} \alpha_{t} \left( i \right) \beta_{t} \left( i \right) }\end{align*}$
给定模型

λ $\lambda$ 和观测

O $O$ ，在时刻

t $t$ 处于状态

qi $q_{i}$ 且在时刻

t+1 $t+1$ 处于状态

qj $q_{j}$ 的概率

ξ t (i, j) = P (i t = q i, i t + 1 = q j | O, λ) = P ( i t = q i , i t + 1 = q j , O | λ ) P ( O | λ ) = P ( i t = q i , i t + 1 = q j , O | λ ) \sum N i = 1 \sum N j = 1 P ( i t = q i , i t + 1 = q j , O | λ ) = α t ( i ) a i j b j ( o t + 1 ) β t + 1 ( j ) \sum N i = 1 \sum N j = 1 α t ( i ) a i j b j ( o t + 1 ) β t + 1 ( j )

$\begin{align*} \\ & \xi_{t} \left( i,j \right) = P \left( i_{t}=q_{i}, i_{t+1}=q_{j} | O ,\lambda \right) \\ & = \dfrac{ P \left( i_{t}=q_{i}, i_{t+1}=q_{j},O | \lambda \right) } { P \left( O | \lambda \right) } \\ & = \dfrac{ P \left( i_{t}=q_{i}, i_{t+1}=q_{j}, O | \lambda \right) } { \sum_{i=1}^{N} \sum_{j=1}^{N} P \left( i_{t}=q_{i}, i_{t+1}=q_{j}, O|\lambda \right) } \\ & = \dfrac{ \alpha_{t} \left( i \right) a_{ij} b_{j} \left( o_{t+1} \right) \beta_{t+1} \left( j \right) } { \sum_{i=1}^{N} \sum_{j=1}^{N} \alpha_{t} \left( i \right) a_{ij} b_{j} \left( o_{t+1} \right) \beta_{t+1} \left( j \right)}\end{align*}$
在观测

O $O$ 下状态

i $i$ 出现的期望

\sum t = 1 T γ t (i) = \sum t = 1 T P (i t = q i | O, λ)

$\begin{align*} & \sum_{t=1}^{T} \gamma_{t} \left( i \right) = \sum_{t=1}^{T} P \left( i_{t}=q_{i} | O, \lambda \right) \end{align*}$
在观测

O $O$ 下由状态

i $i$ 转移的期望

\sum t = 1 T － 1 γ t (i) = \sum t = 1 T － 1 P (i t = q i | O, λ)

$\begin{align*} & \sum_{t=1}^{T－1} \gamma_{t} \left( i \right) = \sum_{t=1}^{T－1} P \left( i_{t}=q_{i} | O, \lambda \right) \end{align*}$
在观测

O $O$ 下由状态

i $i$ 转移到状态

j $j$ 的期望

\sum t = 1 T － 1 ξ t (i, j) = \sum t = 1 T － 1 P (i t = q i, i t + 1 = q j | O, λ)

$\begin{align*} & \sum_{t=1}^{T－1} \xi_{t} \left( i,j \right) = \sum_{t=1}^{T－1} P \left( i_{t}=q_{i}, i_{t+1}=q_{j} | O, \lambda \right) \end{align*}$
将观测序列作为观测数据

O $O$ ,将状态序列作为隐数据

I $I$ ，则应马尔科夫模型是含有隐变量的概率模型

P (O | λ) = \sum I P (O | I, λ) P (I | λ)

$\begin{align*} & P \left( O | \lambda \right) = \sum_{I} P \left( O | I, \lambda \right) P \left( I | \lambda \right)\end{align*}$
完全数据

Q(λ,λ⎯⎯) $Q \left( \lambda, \overline{\lambda} \right)$ 函数

Q (λ, λ ⎯ ⎯) = E I [log P (O, I | λ) | O, λ ⎯ ⎯] = \sum I log P (O, I | λ) P (I | O, λ ⎯ ⎯) = \sum I log P ( O , I | λ ) P ( O , I | λ ⎯ ⎯ ) P ( O | λ ⎯ ⎯ )

$\begin{align*} \\& Q \left( \lambda, \overline{\lambda} \right) = E_{I} \left[ \log P \left( O, I | \lambda \right) | O, \overline{\lambda} \right] \\ & = \sum_{I} \log P \left( O, I | \lambda \right) P \left( I | O, \overline{\lambda} \right) \\ & = \sum_{I} \log \dfrac{P \left( O, I | \lambda \right) P \left( O, I | \overline{\lambda} \right) }{P \left( O | \overline{\lambda} \right)}\end{align*}$
其中，

λ⎯⎯ $\overline{\lambda}$ 是隐马尔科夫模型参数的当前估计值，

λ $\lambda$ 是隐马尔科夫模型参数。
由于对最大化

Q(λ,λ⎯⎯) $Q \left( \lambda, \overline{\lambda} \right)$ 函数，

P(O|λ⎯⎯) $P \left( O | \overline{\lambda} \right)$ 为常数因子，
以及

P (O, I | λ) = π i 1 b i 1 (o 1) a i 1 i 2 b i 2 (o 2) \dots a i T - 1 i T b T (o T)

$\begin{align*} & P \left( O, I | \lambda \right) = \pi_{i_{1}} b_{i_{1}} \left( o_{1} \right) a_{i_{1}i_{2}} b_{i_{2}} \left( o_{2} \right) \cdots a_{i_{T-1}i_{T}}b_{T}\left( o_{T} \right)\end{align*}$
所以求

Q(λ,λ⎯⎯) $Q \left( \lambda, \overline{\lambda} \right)$ 函数对

λ $\lambda$ 的最大

λ = arg max Q (λ, λ ⎯ ⎯) \Leftrightarrow arg max \sum I log P (O, I | λ) P (O, I | λ ⎯ ⎯) = \sum I log π i 1 P (O, I | λ ⎯ ⎯) + \sum I (\sum t = 1 T - 1 log a i t i t + 1) P (O, I | λ ⎯ ⎯) + \sum I (\sum t = 1 T log b i t (o t)) P (O, I | λ ⎯ ⎯)

$\begin{align*} & \lambda = \arg \max{Q \left( \lambda, \overline{\lambda} \right) }\Leftrightarrow \arg\max \sum_{I} \log P \left( O, I | \lambda \right) P \left( O, I | \overline{\lambda} \right) \\ & = \sum_{I} \log \pi_{i_{1}} P \left( O, I | \overline{\lambda} \right) + \sum_{I} \left( \sum_{t=1}^{T-1} \log a_{i_{t}i_{t+1}} \right) P \left( O, I | \overline{\lambda} \right) + \sum_{I} \left( \sum_{t=1}^{T} \log b_{i_{t}} \left( o_{t} \right) \right) P \left( O, I | \overline{\lambda} \right)\end{align*}$
对三项分别进行极大化：
1.

max \sum I log π i 1 P (O, I | λ ⎯ ⎯) = \sum i = 1 N log π i 1 P (O, i 1 = i | λ ⎯ ⎯) s . t . \sum i = 1 N π i = 1

$\begin{align*} & \max \sum_{I} \log \pi_{i_{1}} P \left( O, I | \overline{\lambda} \right) = \sum_{i=1}^{N} \log \pi_{i_{1}} P \left( O, i_{1}=i | \overline{\lambda} \right) \\ & s.t. \sum_{i=1}^{N} \pi_{i} = 1 \end{align*}$
构造拉格朗日函数，对其求偏导，令结果为0

\partial \partial π i [\sum i = 1 N log π i 1 P (O, i 1 = i | λ ⎯ ⎯) + γ (\sum i = 1 N π i - 1)] = 0

$\begin{align*} & \dfrac{\partial}{\partial \pi_{i}} \left[ \sum_{i=1}^{N} \log \pi_{i_{1}} P \left( O, i_{1}=i | \overline{\lambda} \right) + \gamma \left( \sum_{i=1}^{N} \pi_{i} - 1 \right) \right] = 0\end{align*}$
得

P (O, i 1 = i | λ ⎯ ⎯) + γ π i = 0 \sum i = 1 N [P (O, i 1 = i | λ ⎯ ⎯) + γ π i] = 0 \sum i = 1 N P (O, i 1 = i | λ ⎯ ⎯) + γ \sum i = 1 N π i = 0 P (O | λ ⎯ ⎯) + γ = 0 γ = - P (O | λ ⎯ ⎯)

$\begin{align*} & P \left( O, i_{1} = i | \overline{\lambda} \right) + \gamma \pi_{i} = 0 \\ & \sum_{i=1}^{N} \left[ P \left( O, i_{1} = i | \overline{\lambda} \right) + \gamma \pi_{i} \right] = 0 \\ & \sum_{i=1}^{N} P \left( O, i_{1} = i | \overline{\lambda} \right) + \gamma \sum_{i=1}^{N} \pi_{i} = 0 \\ & P \left( O | \overline{\lambda} \right) + \gamma = 0 \\ & \gamma = - P \left( O | \overline{\lambda} \right)\end{align*}$
代入

P(O,i1=i|λ⎯⎯)+γπi=0 $P \left( O, i_{1} = i | \overline{\lambda} \right) + \gamma \pi_{i} = 0$ ，得

π i = P ( O , i 1 = i | λ ⎯ ⎯ ) P ( O | λ ⎯ ⎯ ) = γ 1 (i)

$\begin{align*} & \pi_{i} = \dfrac{P \left( O, i_{1} = i | \overline{\lambda} \right)}{P \left( O | \overline{\lambda} \right)} \\ & = \gamma_{1} \left( i \right) \end{align*}$
2.

max \sum I (\sum t = 1 T - 1 log a i t i t + 1) P (O, I | λ ⎯ ⎯) = \sum i = 1 N \sum j = 1 N \sum t = 1 T - 1 log a i j P (O, i t = i, i t + 1 = j | λ ⎯ ⎯) s . t . \sum j = 1 N a i j = 1

$\begin{align*} \\ & \max \sum_{I} \left( \sum_{t=1}^{T-1} \log a_{i_{t}i_{t+1}} \right) P \left( O, I | \overline{\lambda} \right) = \sum_{i=1}^{N} \sum_{j=1}^{N} \sum_{t=1}^{T-1} \log a_{ij} P \left( O, i_{t}=i, i_{t+1}=j | \overline{\lambda} \right) \\ & s.t. \sum_{j=1}^{N} a_{ij} = 1 \end{align*}$
得

a i j = \sum T - 1 t = 1 P ( O , i t = i , i t + 1 = j | λ ⎯ ⎯ ) \sum T - 1 t = 1 P ( O , i t = i | λ ⎯ ⎯ ) = \sum T - 1 t = 1 ξ t ( i , j ) \sum T - 1 t = 1 γ t ( i )

$\begin{align*} \\ & a_{ij} = \dfrac{\sum_{t=1}^{T-1} P \left( O, i_{t}=i, i_{t+1}=j | \overline{\lambda} \right)}{\sum_{t=1}^{T-1} P \left( O, i_{t}=i | \overline{\lambda} \right)} \\ & = \dfrac{\sum_{t=1}^{T-1} \xi_{t} \left( i,j \right) }{\sum_{t=1}^{T-1} \gamma_{t} \left( i \right)}\end{align*}$
3.

max \sum I (\sum t = 1 N log b i t (o t)) P (O, I | λ ⎯ ⎯) = \sum j = 1 N \sum t = 1 T log b j (o t) P (O, i t = j | λ ⎯ ⎯) s . t . \sum k = 1 M b j (k) = 1

$\begin{align*} \\ & \max \sum_{I} \left( \sum_{t=1}^{N} \log b_{i_{t}} \left( o_{t} \right) \right) P \left( O, I | \overline{\lambda} \right) = \sum_{j=1}^{N} \sum_{t=1}^{T} \log b_{j} \left( o_{t} \right) P \left( O, i_{t}=j | \overline{\lambda} \right) \\ & s.t. \sum_{k=1}^{M} b_{j} \left( k \right) = 1 \end{align*}$
得

b j (k) = \sum T t = 1 P ( O , i t = j | λ ⎯ ⎯ ) I ( o t = v k ) \sum T t = 1 P ( O , i t = j | λ ⎯ ⎯ ) = \sum T t = 1 , o t = v k γ t ( j ) \sum T t = 1 γ t ( j )

$\begin{align*} \\ & b_{j} \left( k \right) = \dfrac{\sum_{t=1}^{T} P \left( O, i_{t}=j | \overline{\lambda} \right) I \left( o_{t} = v_{k} \right)}{\sum_{t=1}^{T} P \left( O, i_{t}=j | \overline{\lambda} \right)} \\ & = \dfrac{ \sum_{t=1,o_{t}=v_{k}}^{T} \gamma_{t} \left( j \right)}{\sum_{t=1}^{T} \gamma_{t} \left( j \right)}\end{align*}$
Baum-Welch算法：
输入：观测数据

O=(o1,o2,⋯,oT) $O = \left( o_{1}, o_{2}, \cdots, o_{T} \right)$
输出：隐马尔科夫模型参数
1. 初始化
对

n=0 $n=0$ ，选取

a(0)ij,bj(k)(0),π(0)i $a_{ij}^{ \left( 0 \right) },b_{j} \left( k \right)^{\left( 0 \right)},\pi_{i}^{\left( 0 \right)}$ ，得到模型

λ(0)=(a(0)ij,bj(k)(0),π(0)i) $\lambda^{\left( 0 \right)} = \left( a_{ij}^{ \left( 0 \right) },b_{j} \left( k \right)^{\left( 0 \right)},\pi_{i}^{\left( 0 \right)} \right)$
2. 递推
对

n=1,2,⋯, $n=1,2, \cdots,$

a (n + 1) i j = \sum T - 1 t = 1 ξ t ( i , j ) \sum T - 1 t = 1 γ t ( i ) b j (k) (n + 1) = \sum T t = 1 , o t = v k γ t ( j ) \sum T t = 1 γ t ( j ) π (n + 1) i = P ( O , i 1 = i | λ ⎯ ⎯ ) P ( O | λ ⎯ ⎯ )

$\begin{align*} \\ & a_{ij}^{\left( n+1 \right)} = \dfrac{\sum_{t=1}^{T-1} \xi_{t} \left( i,j \right) }{\sum_{t=1}^{T-1} \gamma_{t} \left( i \right)} \\ & b_{j} \left( k \right)^{\left( n+1 \right)} = \dfrac{ \sum_{t=1,o_{t}=v_{k}}^{T} \gamma_{t} \left( j \right)}{\sum_{t=1}^{T} \gamma_{t} \left( j \right)} \\ & \pi_{i}^{\left( n+1 \right)} = \dfrac{P \left( O, i_{1} = i | \overline{\lambda} \right)}{P \left( O | \overline{\lambda} \right)} \end{align*}$
其中，右端各值按观测数据

O=(o1,o2,⋯,oT) $O = \left( o_{1}, o_{2}, \cdots, o_{T} \right)$ 和模型

λ(n)=(A(n),B(n),π(n)) $\lambda^{\left( n \right)} = \left( A^{\left( n \right)},B^{\left( n \right)},\pi^{\left( n \right)} \right)$ 计算。
3. 终止
得到模型

λ(n＋1)=(A(n+1),B(n+1),π(n+1)) $\lambda^{\left( n＋1 \right)} = \left( A^{\left( n+1 \right)},B^{\left( n+1 \right)},\pi^{\left( n+1 \right)} \right)$
在时刻

t $t$ 状态为

i $i$ 的所有单个路径

(i1,i2,⋯,it) $\left( i_{1}, i_{2}, \cdots, i_{t} \right)$ 中概率最大值

δ t (i) = max i 1, i 2, \dots, i t - 1 P (i t = i, i t - 1, \dots, i 1, o t, \dots, o 1 | λ) i = 1, 2, \dots, N

$\begin{align*} \\ & \delta_{t} \left( i \right) = \max_{i_{1}, i_{2}, \cdots, i_{t-1}} P \left(i_{t}=i, i_{t-1}, \cdots, i_{1}, o_{t}, \cdots, o_{1} | \lambda \right) \quad \quad \quad i = 1, 2, \cdots, N \end{align*}$
得递推公式

δ t + 1 (i) = max i 1, i 2, \dots, i t P (i t + 1 = i, i t, \dots, i 1, o t + 1, \dots, o 1 | λ) = max 1 \leq j \leq N [max i 1, i 2, \dots, i t - 1 P (i t + 1 = i, i t = j, i t - 1, \dots, i 1, o t + 1, o t, \dots, o 1 | λ)] = max 1 \leq j \leq N [max i 1, i 2, \dots, i t - 1 P (i t + 1 = i, i t = j, i t - 1, \dots, i 1, o t, o t - 1, \dots, o 1 | λ) P (o t + 1 | i t + 1 = i, λ)] = max 1 \leq j \leq N [max i 1, i 2, \dots, i t - 1 P (i t = j, i t - 1, \dots, i 1, o t, o t - 1, \dots, o 1 | λ) P (i t + 1 = i | i t = j, λ) P (o t + 1 | i t + 1 = i, λ)] = max 1 \leq j \leq N [δ t (j) a j i] b i (o t + 1) i = 1, 2, \dots, N

$\begin{align*} \\ & \delta_{t+1} \left( i \right) = \max_{i_{1}, i_{2}, \cdots, i_{t}} P \left(i_{t+1}=i, i_{t}, \cdots, i_{1}, o_{t+1}, \cdots, o_{1} | \lambda \right) \\ & = \max_{1 \leq j \leq N} \left[ \max_{i_{1}, i_{2}, \cdots, i_{t-1}} P \left( i_{t+1}=i, i_{t}=j, i_{t-1}, \cdots, i_{1}, o_{t+1}, o_{t}, \cdots, o_{1} | \lambda \right) \right] \\ & = \max_{1 \leq j \leq N} \left[ \max_{i_{1}, i_{2}, \cdots, i_{t-1}} P \left( i_{t+1}=i, i_{t}=j, i_{t-1}, \cdots, i_{1}, o_{t}, o_{t-1}, \cdots, o_{1} | \lambda \right) P \left( o_{t+1} | i_{t+1}=i, \lambda \right)\right] \\ & = \max_{1 \leq j \leq N} \left[ \max_{i_{1}, i_{2}, \cdots, i_{t-1}} P \left( i_{t}=j, i_{t-1}, \cdots, i_{1}, o_{t}, o_{t-1}, \cdots, o_{1} | \lambda \right) P \left( i_{t+1}=i | i_{t}=j, \lambda \right)P \left( o_{t+1} | i_{t+1}=i, \lambda \right)\right] \\ & = \max_{1 \leq j \leq N} \left[ \delta_{t} \left( j \right) a_{ji}\right] b_{i} \left( o_{t+1} \right)\quad \quad \quad i = 1, 2, \cdots, N \end{align*}$
在时刻

t $t$ 状态为

i $i$ 的所有单个路径

(i1,i2,⋯,it) $\left( i_{1}, i_{2}, \cdots, i_{t} \right)$ 中概率最大值的路径的第

t−1 $t-1$ 个结点

ψ t (i) = arg max 1 \leq j \leq N [δ t - 1 (j) a j i] i = 1, 2, \dots, N

$\begin{align*} \\ & \psi_{t} \left( i \right) = \arg \max_{1 \leq j \leq N} \left[ \delta_{t-1} \left( j \right) a_{ji} \right] \quad \quad \quad i = 1, 2, \cdots, N \end{align*}$
维特比算法：
输入：模型

λ=(A,B,π) $\lambda = \left( A, B, \pi \right)$ 和观测数据

O=(o1,o2,⋯,oT) $O = \left( o_{1}, o_{2}, \cdots, o_{T} \right)$
输出：最优路径

I∗=(i∗1,i∗2,⋯,i∗T) $I^{*} = \left( i_{1}^{*}, i_{2}^{*}, \cdots, i_{T}^{*} \right)$
1. 初始化

δ 1 (i) = π i b i (o 1) i = 1, 2, \dots, N ψ 1 (i) = 0

$\begin{align*} \\ & \delta_{1} \left( i \right) = \pi_{i} b_{i} \left( o_{1} \right) \quad \quad \quad i = 1, 2, \cdots, N \\ & \psi_{1} \left( i \right) = 0 \end{align*}$
2. 递推
对

t=2,3,⋯,T $t=2,3, \cdots, T$

δ t (i) = max 1 \leq j \leq N [δ t - 1 (j) a j i] b i (o t) i = 1, 2, \dots, N ψ t (i) = arg max 1 \leq j \leq N [δ t - 1 (j) a j i] i = 1, 2, \dots, N

$\begin{align*} \\ & \delta_{t} \left( i \right) = \max_{1 \leq j \leq N} \left[ \delta_{t-1} \left( j \right) a_{ji}\right] b_{i} \left( o_{t} \right)\quad \quad \quad i = 1, 2, \cdots, N \\ & \psi_{t} \left( i \right) = \arg \max_{1 \leq j \leq N} \left[ \delta_{t-1} \left( j \right) a_{ji} \right] \quad \quad \quad i = 1, 2, \cdots, N \end{align*}$
3. 终止

P * = max 1 \leq j \leq N δ T (i) i * T = arg max 1 \leq j \leq N [δ T (i)]

$\begin{align*} \\ & P^{*} = \max_{1 \leq j \leq N} \delta_{T} \left( i \right) \\ & i_{T}^{*} = \arg \max_{1 \leq j \leq N} \left[ \delta_{T} \left( i \right) \right] \end{align*}$
4. 最优路径回溯
对