ИСТИНА

This report presents an algorithm for construction of a hierarchical neural network complex for analysis of time series. At the first level of the complex, a specialized committee of artificial neural networks (ANN), trained on different sections of a time series, is used. At the second level, a generalizing ANN is used, that is trained to predict the total probability of an event or continuous value, based on local estimations obtained by the committee. Such an approach allows predicting an event or continuous value, and also automatically determining and taking into account the delay between a precursor phenomenon and the event. The algorithm has been applied to study Sun-Earth relations.