Ensembles of Neural Networks through crossover based pattern generation

The goal of an ensemble construction with several neural networks is to achieve better generalization than that of a single neural network. A Neural Network Ensemble (NNE) performs well when the component networks are diverse, so that failure of one is compensated for by others. Training data variation (i.e., different training sets for different networks) is a good source of diversity because the function that a network approximates is learned from its training data. We introduce a new approach to training data variation and propose the Ensemble based on Crossover based Pattern Generation (ECPG). ECPG generates some new training patterns for a particular network; a pair of pattern is generated interchanging some of input feature values in between a pair of selected original patterns. The effectiveness of ECPG was evaluated using several benchmark classification problems, and ECPG was found to achieve better or competitive performance with respect to related conventional methods. With several benefits over conventional methods, crossover based pattern generation appears to be a good technique for ensemble construction