An Original RBF Network Learning Algorithm

This paper proposes an original RBF network structure learning algorithm aiming at improving its generalization ability. The algorithm determines the initial network hidden structure by decayed radius clustering algorithm, then modifies the hidden structure by adjusting the impurity of generalized clusters containing the classification information of the training samples until the conditions that all clusters' impurities are less than the average impurity level and all variants are less than the average variant level are satisfied. Then we get the final hidden structure. After determining the hidden structure, the back-propagation algorithm is used to training the weights between the hidden layer and output layer. The experiment of two spirals problem proves that our algorithm has higher generalization ability indeed.