Resisting the influence of outliers in radial basis function neural networks

Radial basis function (RBF) neural networks have been shown to be a promising network model in function approximation. Training an RBF network is usually approached using a least square criterion, accompanied with an adaptive growing technique to determine the optimal size of network. With this approach, two problems usually arise when the training patterns contain outliers. Firstly, the least square would cause the network to incorrectly interpolate the outliers. Secondly, because of the interference of outliers, the number of nodes determined by the traditional growing algorithm will stick at a certain number, meaning that the proper network size cannot be reached. In order to cope with the first problem, this paper proposes a method to construct a robust criterion function to replace the least square criterion. For solving the second problem, the paper introduces a memory mechanism into the adaptive growing technique to restrain the influence of outliers. Simulation results indicate that the robust criterion function obtained using our method can effectively reduce the influence of outliers. Furthermore, with the incorporation of the memory mechanism, a better size of network can be obtained