Multiobjective evolutionary optimization for Elman recurrent neural networks, applied to time series prediction

In previous works, we showed how Recurrent Neural Networks, trained with Genetic Algorithms, can achieve a good performance in Time Series Prediction problems. However, the problem of finding the optimal structure of the Network is still a tedious task, which requires several experiments and multiple tests. In this work, we propose a Multiobjective Evolutionary Algorithm to find the optimal topology of the Network. The algorithm also provides a set of optimal networks trained. In the experimental section, we apply the algorithm to obtain optimal Elman Recurrent Neural Networks, in order to predict the evolution of the population of the U.S. and the ECB reference exchange rate between the US dollar and the euro, between other time series.