Genetic algorithms to create training data sets for artificial neural networks

For supervised training of an artificial neural network (ANN) the size and composition of the training data set (TDS) is one of the most important prerequisites for an ANN with good generalization abilities. The exisiting methods to construct training data sets (TDSs) vary from random sampling to incrementally increasing the number of samples in the TDS until a certain quality measure is reached. In our approach we employ a genetic algorithm (GA) for the parallel selection of appropriate input patterns for the TDS. The parallel netGEN system that uses a GA to generate problem-adapted generalized multi-layer perceptrons being trained by error-back-propagation has been extended to evolve (sub)optimal TDSs adapted to an (evolved or user-defined) ANN of fixed topology. Empirical results on a benchmark problem are presented.