Chaotic Time Series Prediction with Feature Selection Evolution

Abstract

Chaotic time series have been successfully predicted with the EPNet algorithm through the evolution of artificial neural networks. However, the input feature selection problem has either not been fully explored before or has not been compared against other algorithms in the literature. This paper presents four algorithms derived from the classical EPNet algorithm to evolve the input feature selection in three different chaotic series: Logistic, Lorenz and Mackey-Glass. Additionally, some flaws in the prediction field that may be considered in future works are discussed. A comparison against previous work demonstrates that in most cases the specialization of the EPNet algorithm allows better solutions with a smaller number of generations.