Genetic Selection of Training Sets for (Not Only) Artificial Neural Networks

Creating high-quality training sets is the first step in designing robust classifiers. However, it is fairly difficult in practice when the data quality is questionable (data is heterogeneous, noisy and/or massively large). In this paper, we show how to apply a genetic algorithm for evolving training sets from data corpora, and exploit it for artificial neural networks (ANNs) alongside other state-of-the-art models. ANNs have been proved very successful in tackling a wide range of pattern recognition tasks. However, they suffer from several drawbacks, with selection of appropriate network topology and training sets being one of the most challenging in practice, especially when ANNs are trained using time-consuming back-propagation. Our experimental study (coupled with statistical tests), performed for both real-life and benchmark datasets, proved the applicability of a genetic algorithm to select training data for various classifiers which then generalize well to unseen data.