隐马尔科夫模型HMM，这回理解了

很好的讲解：http://www.hankcs.com/ml/hidden-markov-model.html

以前一直迷惑，这回理解了

java 源码：

https://github.com/anusreeb/mahout-1/tree/master/mrlegacy/src/main/java/org/apache/mahout/classifier/sequencelearning/hmm

python源码：

https://github.com/hmmlearn/hmmlearn

补充些代码有助于理解：

alpha:

def _forward(states, observations, A, B, Pi):
    num_states = len(states)
    num_observ = len(observations)
    alpha = np.zeros((num_observ, num_states))
    for i in range(num_states):
        alpha[0, i] = Pi[i] * B[i, observations[0]]
    for t in range(1, num_observ):
        for i in range(num_states):
            sum = 0.0
            for j in range(num_states):
                sum += alpha[t - 1, j] * A[j, i]
            alpha[t, i] = sum * B[i, observations[t]]
    return alpha

 beta: 

def _backward(states, observations, A, B):
    num_states = len(states)
    num_observ = len(observations)
    beta = np.zeros((num_observ, num_states))
    for i in range(num_states):
        beta[num_observ - 1][i] = 1
 
    for t in range(num_observ - 2, 0):
        for i in range(num_states):
            sum = 0.0
            for j in range(num_states):
                sum += beta[t + 1][j] * A[i][j] * B[j][observations[t + 1]]
            beta[t][i] = sum
 
    return beta

HMM应用讲解：http://www2.coe.pku.edu.cn/tpic/2012121843059445.pdf

HMM实践：https://www.jianshu.com/p/942d1beb7fdd