An active learning approach for radial basis function neural networks

Abstract

This paper presents a new Active Learning algorithm to train Radial Basis Function (RBF) Artificial Neural Networks (ANN) for model reduction problems. The new approach is based on the assumption that the unobserved training data y at input x, lies within a set F x y f x y f x ( ) : ( ) ( ) = ! ! " # where F(x) is known from experience or past simulations. The new approach finds the location of the new sample such that the worst case error between the output of the resulting RBF ANN and the bounds of the unknown data as specified by F(x) is minimized. This paper illustrates the new approach for the case when x " R1. It was found that it is possible to find a good location for the new data sample by using the suggested approach in certain cases. A comparative study was also done indicating that the new experiment design approach is a good complement to the existing ones such as cross validation design and maximum minimum design.