A self-structuring algorithm for artificial neural networks

A new self-structuring algorithm is presented which determines the optimal set of terms to be included in a generalized single layer network. The terms are chosen from an initial large set of nonlinear basis functions. The subset of terms chosen produces the network model which best approximates the discriminant function required by the training data. Redundant terms are pruned from the network in parallel. Since overfitting is minimized, the resulting network has greater generalization abilities than the full-size original. Theoretical and experimental results are given