A Sequential Radial Basis Function Neural Network Modeling Method Based on Partial Cross Validation Error Estimation

Abstract

Radial Basis Function Neural Network (RBFNN) is widely used in approximating high nonlinear functions. The network complexity and approximation accuracy are directly dominated by the training data. So how to sample data and obtain target system information in design space effectively is one of the key issues in improving RBFNN approximation capability. In this paper, a sequential RBFNN modeling method based on partial cross validation error estimation (PCVEE) criterion is proposed. This method can utilize the sample data as the validation data to test the approximation model accuracy, and expand the sample set purposively and refine the model sequentially according to the error estimation, so as to improve the approximation accuracy effectively. Two mathematical examples are tested to verify the efficiency of this method.