Analysis of Sequential Data Using HMMs
Stephen N. Winters-Hilt · 2021
Generalized Hidden Markov model (HMM) methods are described for both signal feature extraction and structure identification. An HMM is the central method in all of these approaches because it is the simplest modeling approach that is obtained when students combine a Bayesian statistical foundation for Markovian stochastic sequential analysis with the efficient dynamic programming table constructions possible on a computer. The chapter presents Markov Models and HMMs in terms of graphical models. The generalized clique HMM begins by enlarging the primitive hidden states associated with the individual base labels to substrings of primitive hidden states or footprint states. All of the HMM generalizations and feature extraction methods can be optimized for speed with binned durations and thoroughly distributed table-chunking. AdaBoost learns from a collection of weak classifiers and boosts them by a linear combination into a single strong classifier.