Evolving basis functions with dynamic receptive fields

Neural networks using radial basis functions (RBFs) are a popular representation for inducing classification schemes. However, RBF neural networks often require a large number of hidden units (basis functions) in order to adequately model the class distinctions. This is due to the static nature of each basis function. This paper uses an evolutionary program to induce dynamic basis functions whose receptive fields are dependent on the input vector. This technique requires only a single basis function per class to perform on par with RBF networks.