Mahdaviani, K, Mazyar, H, Majidi, S and Saraee, M 2008, A method to resolve the overfitting problem in recurrent neural networks for prediction of complex system's behavior , in: IEEE World Congress on Computational Intelligence / IEEE International Joint Conference on Neural Networks, 1-6 June 2008, Hong Kong, China.

PDF - Published Version Restricted to Repository staff only Download (309kB) | Request a copy

Abstract

In this paper a new method to resolve the overfitting problem for predicting complex systems' behavior has been proposed. This problem occurs when a neural network loses its generalization. The method is based on the training of recurrent neural networks and using simulated annealing for the optimization of their generalization. The major work is done based on the idea of ensemble neural networks. Finally the results of using this method on two sample datasets are presented and the effectiveness of this method is illustrated.

Actions (login required)