Selection of training data for neural networks by a genetic algorithm

In applications of artificial neural networks (ANNs), it is common to partition the available data into (at least) two sets. One is then used to train the net, while the other is used as a 'test set' to measure the generalization capability of the trained net. The partition is generally almost completely arbitrary, and little research has been done on the question of what constitutes a good training set, or on how it could be achieved. In this paper, we use a genetic algorithm (GA) to identify a training set for fitting radial basis function (RBF) networks, and test the methodology on two classification problems-one an artificial problem, and the other using real-world data on credit applications for mortgage loans. In the process, we also exhibit an interesting application of Radcliffe's RAR operator, and present results that suggest the methodology tested here is a viable means of increasing ANN performance.