Evolutionary Induced Time Series Predictors Based on Radial Functions

摘要

This paper describes an evolutionary methodology capable of generating parsimonious single-step prediction models, calculated from experimental data, by the aggregation of Radial Basis Functions (RBF). The characteristics of the model (i.e. the number of terms, the functions and the parameters involved in each term) emerge as result of an evolutionary process without predefined initial structures. The proposed methodology differs notably from the standard RBF approach. In the latter case, a model is constructed with a predefined number of terms, using a single type of radial function. The prediction models induced by the proposed evolutionary methodology support different types of radial functions and they are initiated with no predefined number of terms. The proposed methodology combines advantages from evolution programs, found to be useful search tools in complex spaces, with the power of the RBFs as prediction models. The methodology was experimentally tested over synthetic and real data. The logistic equation and Henon mapping generated the synthetic data. The real data consisted of several time series: The exchange rate of the Swiss Franc, seismographic readings of the most recent earthquake in Kobe, intensity of a laser in chaotic state and annual average of sunspots number. The prediction models induced by the methodology show, in general, good results, comparable or better than those reported by other methods.