Evolutionaty training for dynamical recurrent neural networks: an application in finantial time series prediction

Abstract

Theoretical and experimental studies have shown that traditional training algorithms for Dynamical Recurrent Neural Networks may suffer of local optima solutions, due to the error propagation across the recurrence. In the last years, many researchers have put forward different approaches to solve this problem, most of them being based on heuristic procedures. In this paper, the training capabilities of evolutionary techniques are studied, for Dynamical Recurrent Neural Networks. The performance of the models considered is compared in the experimental section, in real finantial time series prediction problems.