Multiobjective Evolutionary Optimization of Training and Topology of Recurrent Neural Networks for Time-Series Prediction

This paper provides a new evolutionary multiobjective optimization method for automatically optimizing the network topology of recurrent neural networks. The feature of the proposed method is that it involves a procedure of intensively exploring a region including solutions with small training errors in the Pareto frontier, instead of finding a whole set of the Pareto optimal solutions. Several numerical experiments are executed in order to show the advantage of the proposed method over the existing effective algorithm by Delgado et al. with respect to the capability of time-series prediction.