A multiple model approach for prediction using genetic algorithm

Many real-life time series cannot be accurately described by using a single dynamic model. A large amount of real world time series are composed of more than one underlying regimes switching along the time scale. In this paper, we propose using multiple nonlinear models for prediction. Based on a hidden Markov process, the proposed multiple model (MM) is able to capture the temporal relationship among the underlying regimes. A genetic algorithm (GA) is employed to train the multiple model and to obtain an optimal segmentation of the time series. Using real-life sea clutter data, this named GA MM predictor is shown to provide an accurate model for sea clutter in various sea state conditions.