Hybrid Genetic Learning of Hidden Markov Models for Time Series Prediction

Abstract

This paper presents how a hybrid genetic/gradient search can be used to learn hidden Markov models in the context of time series prediction. This learning algorithm called GHOSP uses a gradient search, namely the Baum Welch algorithm, as a local search operator in the main loop of a genetic algorithm, in conjunction with standard genetic operators adapted for hidden Markov models. GHOSP is able to learn efficiently the coefficients and the architecture of hidden Markov models in order to maximize the probability of generating an observation O. This observation is used to encode the recent past of a time series. Once an efficient stochastic model of the series has been learned, this model can be used to predict the next values of the series. We apply this framework to several standard series including economical ones.