Algorithm for Determining the Maximum Likelihood Estimate of Hidden Markov Models Having Tied Parameters
Original Publication Date: 1989-Feb-01
Included in the Prior Art Database: 2005-Jan-27
A technique is described whereby the maximum likelihood algorithms for the tied parameters, based on complete data, are used to produce maximum likelihood algorithms based on incomplete data (output sequence only) and the EM algorithm. In the use of interpolated estimation of Markov source parameters from sparse data in a hidden Markov model [*], with multinomial outputs, the pooling of counts for tied probabilities defines a valid maximum likelihood algorithm for the tied model based on incomplete data. The algorithm is generally based on the Baum-Welch re-estimation algorithm, which is a particular instance of the general EM algorithm as is discussed in statistical literature.