Rule-extraction from radial basis function networks

Radial basis neural (RBF) networks provide an excellent solution to many pattern recognition and classification problems. However, RBF networks are also a local representation technique that enables the easy conversion of the hidden units into symbolic rules. This paper examines rules extracted from RBF networks. We use the iris flower classification task and a vibration diagnosis classification task to illustrate the new knowledge extraction techniques. The rules are analyzed in order to gain knowledge and insight into the network representations. We argue that the local Gaussian representation in RBF networks is particularly useful for rule extraction.