A closer look at the radial basis function (RBF) networks

This paper takes a closer look at the increasingly popular RBF networks by showing that not all the RBF networks are the same. They differ in training, in architecture and in the type of RBF used, and consequently they differ in performance and characteristics. Some of them are remarkably better than the others. With analysis and examples, this paper examines the issues of generalization, smoothness of interpolation, and compares different training methods. A robust modelling procedure and a novel fine-tuning procedure are among the recommended features for consistently good performance. Connections with fuzzy information processing and with spline interpolation are also discussed.<
>