隐马尔可夫模型:状态序列估计问题(维特比算法)
维特比(Viterbi)算法用于求解HMM的第二个问题,即给定观察序列 O = O 1 O 2 ⋯ O T O = O_{1} O_{2} \cdots O_{T} O=O1O2⋯OT和模型 μ = ( A , B , π ) \mu = (\mathbf{A}, \mathbf{B}, \mathbf{\pi}) μ=(A,B,π),选择在一定意义下“最优”的状态序列 Q = q 1 q 2 ⋯ q T Q = q_{1} q_{2} \cdots q_{T} Q=q1q2⋯qT,使得该状态序列“最好地解释”观察序列。该问题的答案并不唯一,它取决于对“最优状态序列”的理解。一种理解是,使该状态序列中每一个状态都单独地具有最大概率,即使 γ t ( i ) = P ( q t = s i ∣ O ; μ ) \gamma_{t}(i) = P(q_{t} = s_{i} | O; \mu) γt(i)=P(qt=si∣O;μ)最大。
根据贝叶斯公式,
γ t ( i ) = P ( q t = s i ∣ O ; μ ) = P ( q t = s i , O ; μ ) P ( O ; μ ) = α t ( i ) β t ( i ) ∑ i = 1 N α t ( i ) β t ( i ) (6-19) \begin{aligned} \gamma_{t}(i) & = P(q_{t} = s_{i} | O; \mu) \\ & = \frac{P(q_{t} = s_{i}, O; \mu)}{P(O; \mu)} \\ & = \frac{\alpha_{t}(i) \beta_{t}(i)}{\sum_{i = 1}^{N} \alpha_{t}(i) \beta_{t}(i)} \\ \end{aligned} \tag {6-19} γt(i)=P(qt=si∣O;μ)=P(O;μ)P(qt=si,O;μ)