Research On Radial Basis Function Neural Network Learning Algorithms

Radial basis function (RBF) neural network, which has a profound physiological basis, a simple network architecture, fast learning algorithms, good approximation capability, has many successful applications in fields, such as function approximation, pattern recognition, signal processing, system identification, and so on. At present, it still plays an important part in neural network research. The selection of RBF hidden node centers is one of the most important problems in RBF network learning; and the presence of outliers in the given training data set, will influence the performance of the network. In this thesis, the existing RBF learning algorithms are investigated, and the solutions are given to the two problems.The results are as follows:(1)Analyzes the virtues and disadvantages of the existing RBF learning algorithms, in the selection of hidden node centers, determination of the hidden node width, and the optimization of the weights.(2)By combining the differential evolution with orthogonal least squares algorithm, the differential evolution orthogonal least squares (DEOLS) learning algorithm is proposed, which chooses the hidden node centers rationally, and improves the network generation ability.(3)By combining the subtractive clustering method with scaled robust loss function, a fast robust learning algorithm for RBF network is proposed, which not only enhances the network robustness against outliers, but also shortens the network training time.(4)The differential evolution orthogonal least squares algorithm and fast robust learning algorithm are implemented, and are applied to function approximation and time series prediction problems. So, the effectiveness of the two algorithms are proved.Finally, the research work involved in the thesis is summarized and the future developments in RBF network learning are forecasted.