Structure discovery in hidden Markov models.

Abstract

The Baum-Welch algorithm for training hidden Markov models (HMMs) requires model topology and initial parameters to be specifed, and iteratively improves the model parameters. Sometimes prior knowledge of the process being modeled allows such specifcation, but often this knowledge is unavailable. Experimentation and guessing are resorted to. Techniques for discovering the model structure from observation data exist but their use is not commonplace. We propose a state split-ting approach to structure discovery, where states are split based on two heuristics: within-state autocorrelation and a measure of Markov violation in the state path. Statistical hypothesis testing is used to decide which states to split, providing a natural termination criterion and taking into account the number of observations assigned to each state, splitting states only when the data demands it.