book

Hidden Semi-Markov Models

by Shun-Zheng Yu

October 2015

Beginner to intermediate

208 pages

6h 3m

English

Elsevier

Read now

Unlock full access

Abstract1.1 Markov Renewal Process and Semi-Markov Process
1.3 Dynamic Bayesian Networks1.4 Conditional Random Fields1.5 Hidden Semi-Markov Models1.6 History of Hidden Semi-Markov Models
Abstract2.1 A General Definition of HSMM2.2 Forward–Backward Algorithm for HSMM
2.4 Forward-Only Algorithm for HSMM2.5 Viterbi Algorithm for HSMM2.6 Constrained-Path Algorithm for HSMM

Abstract3.1 EM Algorithm and Maximum-Likelihood Estimation3.2 Re-estimation Algorithms of Model Parameters
3.4 Online Update of Model Parameters
Abstract4.1 Heuristic Scaling4.2 Posterior Notation4.3 Logarithmic Form4.4 Practical Issues in Implementation
Abstract5.1 Explicit Duration HSMM5.2 Variable Transition HSMM5.3 Variable-Transition and Explicit-Duration Combined HSMM5.4 Residual Time HSMM
Abstract6.1 Exponential Family Distribution of Duration6.2 Discrete Coxian Distribution of Duration6.3 Duration Distributions for Viterbi HSMM Algorithms
Abstract7.1 Typical Parametric Distributions of Observations7.2 A Mixture of Distributions of Observations7.3 Multispace Probability Distributions7.4 Segmental Model7.5 Event Sequence Model
Abstract8.1 Switching HSMM8.2 Adaptive Factor HSMM8.3 Context-Dependent HSMM
8.5 Signal Model of HSMM8.6 Infinite HSMM and HDP-HSMM8.7 HSMM Versus HMM
Abstract9.1 Speech Synthesis9.2 Human Activity Recognition9.3 Network Traffic Characterization and Anomaly Detection9.4 fMRI/EEG/ECG Signal Analysis

Content preview from Hidden Semi-Markov Models

Chapter 3

Parameter Estimation of General HSMM

Abstract

This chapter discusses the maximum-likelihood estimation of model parameters for the general HSMM, and uses the theory associated with the well-known EM (expectation-maximization) algorithm to prove that the parameter estimation procedure for the general HSMM increases the likelihood function and converges to the maximum. The formulas and algorithms for unsupervised, supervised, or semi-supervised learning of model parameters are derived for the cases when there are one or multiple observation sequences. An order estimation method for the general HSMM is provided. Finally, approaches for online update of model parameters are yielded based on forward-only algorithm or maximization of the likelihood ...