Training Generalized Hidden Markov Model with Interval Probability Parameters
Yan Wang · 2014
Recently, generalized interval probability was proposed as a new mathematical formalism of imprecise probability. It provides a simplified probabilistic calculus based on its definitions of conditional probability and independence. The Markov property can be described in a form similar to classical probability. In this paper, an expectation-maximization approach is developed to train generalized hidden Markov models with generalized interval probabilities. With the consideration of systematic error in measurement, the training process provides a robust learning mechanism where data quality requirement is not as restrictive as the traditional hidden Markov model.