A Novel Fuzzy Clustering Algorithm for Radial Basis Function Neural Network

Abstract

A Fuzzy Radial basis function neural network (FRBFNN) classifier is proposed in the framework of Radial basis function neural network (RBFNN). This classifier is constructed using class-specific fuzzy clustering to form the clusters which represent the neurons i.e. fuzzy set hyperspheres (FSHs) in the hidden layer of FRBFNN. The creation of these FSHs is based on the maximum spread from inter-class information and intra-class fuzzy membership mechanism. The proposed approach is fast, independent of parameters, and shows good data visualization. The Least mean square training between the hidden layer to output layer in RBFNN is avoided, thus reduces the time complexity. The FRBFNN is trained quickly due to the fast converge of input data to form the FHSs in the hidden layer. The output is determined by the union operation of the FHSs outputs which are connected to the class nodes in the output layer. The performance of the proposed FRBFNN is compared with the other RBFNNs using ten benchmark datasets. The empirical findings demonstrate that the proposed FRBFNN is highly efficient classifier for pattern recognition.