Rule Extraction from Radial Basis Function Networks by Using Support Vectors

In this work, a procedure for rule extraction from radial basis function (RBFN) networks is proposed. The algorithm is based on the use of a support vector machine (SVM) as a frontier pattern selector. By using geometric methods, centers of the RBF units are combined with support vectors in order to construct regions (ellipsoids or hyper-rectangles) in the input space, which are later translated to if-then rules. Additionally, the support vectors are used to determine overlapping between classes and to refine the rule base. The obtained experimental results indicate that a very high fidelity between RBF network and the extracted set of rules can be achieved with low overlapping between classes.