An objective method to find better RBF networks in classification

Abstract

RBF networks are good at prediction tasks of data mining, and k-means clustering algorithm is one of the mostly used clustering algorithms for basis functions of RBF networks. K-means clustering algorithm needs the number of clusters for initialization, and depending on the number of clusters, the accuracy of RBF networks change. But we cannot resort to increasing the number of clusters in the RBF networks in sequential manner, because we have limited computing resources. This paper suggests an objective and systematic approach using decision tree in determining a proper number of clusters to find good RBF networks with respect to accuracy. Experiments with two different data sets showed very promising results.