Classification using Networks of Normalized Radial Basis Functions

Normalized Radial Basis Function Networks (NRBF) were invented at the same time as standard RBF nets, in 1989, but went unnoticed until recently, when it was found that they constitute a very interesting tool, especially for pattern classification. NRBF classifiers behave as Nearest Neighbour classifiers and have a functionality similar to Fuzzy Inference Systems but without the Curse of Dimensionality problem. NRBF nets are easy to use, with performances largely insensitive to model parameters and with a very fast training. In this paper the functionality of NRBF nets is compared with those of standard RBF nets. The applications to classification will be illustrated with benchmark problems. For the IRIS classification problem, NRBF match the performances of the best published techniques.