Genetic algorithm design of complexity-controlled time-series predictors

Abstract

A genetic algorithm that designs artificial neural networks for time-series prediction encodes the structure and the weight magnitudes in a novel genome representation. This allows the genetic algorithm to perform training and complexity control simultaneously, thus directly addressing the problems of generalization and overfitting of data in the evolution of the network. Modified genetic crossover and modified mutation operations are introduced to increase population diversity and improve speed of convergence. Well performing neural networks were evolved automatically for time-series prediction of atmospherically-perturbed light waves in adaptive optics and the time series used in the 1998 Leuven predictor competition.