Genetic Algorithms as an Alternative Method of Parameter Estimation and Finding Most Likely Sequences of States of Hidden Markov Chains for HMMs and Hybrid HMM/ANN Models

Article type: Research Article

Authors: Bijak, Katarzyna

Affiliations: Scoring Department, Credit Information Bureau, Modzelewskiego 77A, 02-679 Warsaw, Poland. E-mail: [email protected]

Note: [] Address for correspondence: Kluczborska 7/8, 01-461 Warsaw, Poland

Abstract: In this paper genetic algorithms are used in estimation and decoding processes of a Hidden Markov Model (HMM) and a hybrid HMM/ANN model with conditional binomial distributions. The hybrid model combines a hidden Markov chain with a perceptron which is assumed to constitute a match network. Genetic algorithms are applied here instead of the traditional methods such as the EM algorithm and the Viterbi algorithm. The paper demonstrates performance of an HMM and a hybrid model in modeling the annual number of months, in which some seismic events are recorded. Parameters of the discrete-time two-state models are estimated using the maximum likelihood method, on the basis of data on seismic events that were recorded in Romania in years 1901¨C1990. Then, on the basis of the estimation results, the most likely sequences of states of the hidden Markov chains are found.

Keywords: HiddenMarkovModels, hybrid HMM/ANN models, Markov chains, neural networks, genetic algorithms, seismic events

Journal: Fundamenta Informaticae, vol. 86, no. 1-2, pp. 1-17, 2008