已知初始状态概率向量 pi,状态转移概率矩阵 P,发射概率矩阵 B,求观察序列为 O 的条件下状态序列为 Q 的概率。代码如下:
def HMM_ProCond(pi, P, B, Q, O):
Q, O = [q - 1 for q in Q], [o - 1 for o in O]
N, proNume, proDeno, proCond = len(pi[0]), - numpy.log(pi[0][Q[0]]) - numpy.log(B[Q[0]][O[0]]), 0, []
for i in range(N):
proCond.append(- numpy.log(pi[0][i]) - numpy.log(B[i][O[0]]))
for i in range(1, len(Q)):
proNume -= numpy.log(P[Q[i - 1]][Q[i]]) + numpy.log(B[Q[i]][O[i]])
proCondi = []
for j in range(N):
proCondi.append(0)
for k in range(N):
proCondi[-1] += numpy.exp(numpy.log(P[k][j]) - proCond[k])
proCondi[-1] = - numpy.log(proCondi[-1]) - numpy.log(B[j][O[i]])
proCond = proCondi
for i in range(N):
proDeno += numpy.exp(- proCond[i])
return numpy.exp(- proNume - numpy.log(proDeno))
已知 pi, P, B 如上所述,求观察序列为 O 的条件下概率最大的状态序列 Q。该问题用 Viterbi 算法解决,代码如下:
def HMM_Viterbi(pi, P, B, O):
N, Q, V = len(pi[0]), [], []
for i in range(N):
Q.append([i + 1])
V.append(- numpy.log(pi[0][i]) - numpy.log(B[i][O[0] - 1]))
for Oi in O[1 :]:
Qi, Vi = [], []
for j in range(N):
Qi.append([])
Vi.append(-1)
for k in range(N):
Vk = V[k] - numpy.log(P[k][j]) - numpy.log(B[j][Oi - 1])
if Vi[-1] == -1 or Vi[-1] > Vk:
Qi[-1], Vi[-1] = Q[k] + [j + 1], Vk
Q, V = Qi, Vi
m = 0
for i in range(N):
if V[m] > V[i]:
m = i
return Q[m]
Viterbi 算法本质上是动态规划,设 V[i, j] = Max{P(Q[1 … i] | O[1 … i] and Q[i] = j)},则有状态转移方程 V[i, j] = B[j, O[i]] * Max{V[i - 1, k] * P[k, j]},边界 V[1, j] = B[j, O[1]] * pi[j],于是 Max{V[len(O), j]} 即为所求 Max{P(Q, O)},由条件概率公式得 P(Q|O) 最大时 P(Q, O) 亦为最大,每次迭代时记录状态即可得到 Q。
给个例子,pi = [ 0.6 0.4 ] \begin{bmatrix} 0.6 & 0.4 \end{bmatrix} [0.60.4], P = [ 0.6 0.4 0.3 0.7 ] \begin{bmatrix} 0.6 & 0.4 \\ 0.3 & 0.7 \end{bmatrix} [0.60.30.40.7], B = [ 0.6 0.3 0.1 0.1 0.4 0.5 ] \begin{bmatrix} 0.6 & 0.3 & 0.1 \\ 0.1 & 0.4 & 0.5 \end{bmatrix} [0.60.10.30.40.10.5],当 O = [1, 2, 3], Q = [1, 2, 2] 时,求 P(Q|O);当 O = [2, 1, 3, 2, 3, 2, 2, 3, 3, 1, 2, 1, 1, 1, 2, 3, 3, 3, 3, 2] 时,求 Q 使 P(Q|O) 最大。代码如下:
import numpy
pi, P, B = [[0.6, 0.4]], [[0.6, 0.4], [0.3, 0.7]], [[0.6, 0.3, 0.1], [0.1, 0.4, 0.5]]
N = len(pi[0])
print(HMM_ProCond(pi, P, B, [1, 2, 2], [1, 2, 3]))
print(HMM_Viterbi(pi, P, B, [2, 1, 3, 2, 3, 2, 2, 3, 3, 1, 2, 1, 1, 1, 2, 3, 3, 3, 3, 2]))