Loading...
HMM Parameters Optimization using Combine Genetic Algorithm and Iterative Training
Author(s)
Date Issued
December 1, 2003
Type
Conference Paper
Abstract
HMM have been used extensively for recognizing observation sequence especially in speech recognition. Iterative training procedure such as Baum-Weltch, or gradient techniques are normally used to find locally optimize HMM parameters. This paper presents genetic algorithm (GA) to perform global search for Hidden Markov Model (HMM) parameters that maximize probability of observation sequence given the model. In order to increase the convergence rate and parameters optimization, we combine iterative procedure with GA. The probability of observation sequence of the train model using iterative procedure, GA, and GA with iterative procedure will be compared along with their convergence rates. The test patterns are chain code sequences generated from 38 isolated on-line Thai handwritten characters. The recognition rate and the probability of the train observation sequences of GA were better than the iterative training. The recognition rate of HMM with iterative training 95.05%, GA 97.50% and GA with iterative training 98.41% on 3839 patterns.
Citation
Proceedings of the IASTED International Conference on Artificial Intelligence and Soft Computing, 7, 255-260, 2003