Improving the Generalization Properties of Radial Basis Function Neural Networks

An important feature of radial basis function neural networks is the existence of a fast, linear learning algorithm in a network capable of representing complex non-linear mappings. Satisfactory generalization in these networks requires that the network mapping be sufficiently smooth. We show that a modification to the error functional allows smoothing to be introduced explicitly without significantly affecting the speed of training. A simple example is used to demonstrate the resulting improvement in the generalization properties of the network.