Projection pursuit learning networks for regression

Two types of learning networks for nonparametric regression problems are studied and compared: one is the parametric two-layer perceptron type neural network, which is well known in artificial neural network (ANN) literature; the other is the semiparametric projection pursuit network (PPN), which has emerged in recent years in the statistical estimation literature. From an algorithmic viewpoint, both the PPN and the ANN parametrically form projections of the data in directions determined from interconnection weights. However, unlike an ANN which uses a fixed set of nonlinear nodal functions to perform an explicit parametric estimate of a nonparametric model, the PPN nonparametrically estimates the nonlinear functions using a one-dimensional data smoother. From experimental simulations, ANNs and PPNs perform comparably in predicting independent test data but PPN training is much faster than that of an ANN. >