Adaptive estimation of hidden nearly completely decomposable Markov chains
Vikram Krishnamurthy · 2002
We propose maximum-likelihood (ML) estimation schemes for nearly completely decomposable Markov chains (NCDMC) in white Gaussian noise. Aggregation techniques based on stochastic complementation are applied to reduce the dimension of the resulting hidden Markov model (HMM) and hence substantially reduce the computational costs of the estimation algorithms. We then present an aggregation based expectation maximization (EM) algorithm for estimating the parameters and states of the HMM.>