| Abstract [eng] |
Online discrete-state Hidden Markov Model (HMM) parameter estimation is explored in this dissertation. In this case, continuous multivariate observations are processed sequentially (i.e. online, recursively) in time rather than after being stored in computer memory and then processed as data batch. The aim of this work is to investigate and apply proposed HMM multivariate parameter estimation algorithms to online analysis of sequence data. In this work, two online algorithms for online HMM parameter estimation are proposed – online Gaussian HMM parameter estimation algorithm and online Dirichlet HMM parameter estimation algorithm. The proposed online HMM parameter estimation algorithms save computational time and computer memory by eliminating the need to store all training data in computer memory by using sequential data analysis. The algorithms proposed in the work consist of two parts: the first part is dedicated to the training of HMM parameters and the second part is used to classify the observations and update the HMM parameters. Calculation of HMM state transition probabilities by changing the classic modified Forward-Backward procedure, is proposed. Detailed comparative analysis of the proposed algorithms was performed, solving the tasks of classification of multivariate observations. |
