Random sieve based on projections for RBF neural net structure selection

Abstract

Our aim is to propose a method for selecting a radial basis functions terms to be included into a neural net model. As it is frequently met in practice, we consider the case of a deficit in the admissible number of observations (learning sequence) in comparison with a much larger number of candidate terms. The proposed approach is based on a random sieve that aims at selecting only necessary RBF's by a hierarchy of a large number of random mixing of candidate RBF's and testing their significance. The results of simulations are also reported. © 2013 Springer-Verlag.