The adjusted Viterbi training for hidden Markov models (Q1002581): Difference between revisions

The problem of effective parameter estimation of hidden Markov model \((X_n,Y_n)\), \(n=1,2,\dots\) is considered. Here \(Y_n\), \(n=1,2,\dots\) is unobservable time-homogeneous Markov chain with state space \(S=(1,\dots,K)\), transition matrix \(\mathbb{P}=(p_{ij})_{i,j \in S}\) and initial distribution \(\pi=\pi \mathbb{P}\). When \(Y_n\) is in state \(l\in S\) an observation \(x_n\) on \(X_n \in \mathbb{R}^D\) is emitted with distribution density \(f(x|\theta_l)\) which is known up to parametrization \(\theta_l \in \Theta \subseteq \mathbb{R}^d\) and independent of anything else. Given observations \(x_1,x_2,\dots,x_n\), the problem is to estimate the matrix \(\mathbb{P}\) and parameters \(\theta_1,\dots,\theta_K\). The principle tool is the EM procedure. Applications often use less computationally intensive Viterbi training procedure but it is biased and does not satisfy fixed point property (being initialized to the true parameters it can move away from initial points). In the article adjusted Viterbi training, a new method to restore the fixed point property is proposed. The advantages of this improved method are investigated analytically and by simulations.