Feature selection for RBF networks

Radial basis function networks (RBFN) can be used for data classification. We present an a-posteriori feature selection method for trained RBFN that is not related to the specific learning procedure as long as every neuron belongs to exactly one class. Considering the center positions, the radii, the weights and the class labels of the neurons, we can easily calculate an index for feature selection that is based on one-dimensional projections. We present examples on different data sets by using Berthold and Diamond´s RBFN.