Function approximation based on self-adaptive RBF neural network with combined clustering algorithm

Abstract

Due to the difficulty of determining node number of hidden layers and center, slow learning speed and weaken generalization ability of RBF neural network when input data is generous and complex. A new method based on combined clustering is presented here to determine node number of hidden layer and centers of RBF neural network self-adaptively. In this paper, subtractive clustering is used to cluster the data firstly, node number and initial value of data center of RBF network are achieved, then GK fuzzy clustering algorithm is adopted to evaluate cluster validity and obtain optimum data center by estimating the shape and direction of clustering. The normalized LMS algorithm is used to tune weights. Thus, a self-adaptive RBF neural network with combined clustering algorithm is obtained. The simulation results of function approximation show that the RBF neural network designed has better approximation performance.