The Growing Radial Basis Function (RBF) Neural Network and Its Applications

This paper proposes a framework based on the cross-validation methods for constructing and training radial basis function (RBF) neural networks. The proposed growing RBF (GRBF) neural network begins with initial number of hidden units. In the process of training, the GRBF network adjusts the hidden neurons by eliminating some “small” hidden units and splitting one “large” hidden unit at the same cycle. If the prediction error in the system is not less than the pre-given threshold, the proposed method increases hidden units to re-estimate the parameters in the next process of training, until the stop criterion is satisfied. In practice, the proposed GRBF network are evaluated and tested on two real 3D seismic data sets with very favorable self-adaptive ability and satisfactory results.