A novel training algorithm for RBF neural network using a hybrid fuzzy clustering approach

This paper introduces a novel clustering-based algorithm to train Gaussian type radial basis function neural networks. In contrast to existing approaches, we develop a specialized learning strategy that combines the merits of fuzzy and crisp clustering. Crisp clustering is a fast process, yet very sensitive to initialization. On the other hand, fuzzy clustering reduces the dependency on initialization; however, it constitutes a slow learning process. The proposed strategy aims to search for a trade-off among these two potentially different effects. The produced clusters possess fuzzy and crisp areas and therefore, the final result is a hybrid partition, where the fuzzy and crisp conditions coexist. The hybrid clusters are directly involved in the estimation process of the neural network's parameters. Specifically, the center elements of the basis functions coincide with cluster centers, while the respective widths are calculated by taking into account the topology of the hybrid clusters. To this end, the network's design becomes a fast and efficient procedure. The proposed method is successfully applied to a number of experimental cases, where the produced networks prove to be highly accurate and compact in size. ► We linearly combine the c-means and fuzzy c-means algorithms to train RBF networks. ► Each training sample does not affect distant cluster centers. ► Each cluster possesses crisp and fuzzy areas. ► The algorithm provides an efficient estimation of the basis function centers.