Genetic modification of a neural networks training data

A major problem associated with artificial neural networks (ANNs) is that of overgeneralization. Exceptions in the training data are effectively ignored as they are few in number compared to the vast majority of training examples. Modification of the training data has the potential to alleviate this problem. Genetic algorithms are used to guide the search for an optimal set of training data, with the genotypic representation being the frequency of each training example in the training set. The authors investigate the combination of genetic algorithm and a neural network to provide a technique capable of handling exceptions.<<ETX>>