Regularization of Neural

Neural networks are exible tools for nonlinear function approximation and by expanding the network any relevant target function can be approximated 6]. The risk of overrtting on noisy data is of major concern in neural network design 2]. By using regularization, overrtting is reduced, thereby improving generalization ability on future data. In this contribution we present a scheme for estimation of regularization parameters using a simple validation set approach. Further work on empirical methods for optimization of neural networks models can be found in 10]. The objective neural network modelling is to establish an estimate of the nonlinear relation among two vector variables: the input x and the output y. The neural net is estimated (trained) from a dataset of related input-output examples. The neural network implements a nonlinear function described by the vector function f(x;w) where w is the m-dimensional vector of all network weights. For a standard two-layer feed-forward neural net, the q'th component of f is given by (see e.g., 5]):