Adding diversity in ensembles of neural networks by reordering the training set

Abstract

When an ensemble of neural networks is designed, it is necessary to provide enough diversity to the different networks in the ensemble. We propose in this paper one new method of providing diversity which consists on reordering the training set patterns during the training. In this paper three different algorithms are applied in the process of building ensembles with Simple Ensemble and Cross-Validation. The first method consists in using the original training set, in the second one the training set is reordered before the training algorithm is applied and the third one consists in reordering the training set at the beginning of each iteration of BackPropagation. With the experiments proposed we want to empirically demonstrate that reordering patterns during the training is a valid source to provide diversity to the networks of an ensemble. The results show that the performance of the original ensemble methods can be improved by reordering the patterns during the training. Moreover, this new source of diversity can be extended to more complex ensemble methods. © Springer-Verlag Berlin Heidelberg 2008.