A statistical inference based growth criterion for the RBF network

In this paper, a growth criterion is derived using statistical inference for model sufficiency. This criterion is developed for recursive estimation or sequential learning with neural networks. A growing Gaussian radial basis function (GaRBF) network trained by the extended Kalman filter (EKF) algorithm on-line, called incremental network is developed. Incremental network is similar to the resource allocating network (RAN). The criterion for growth is based on the network prediction error and the expected uncertainty in the network output. The criterion is computed within the EKF estimation end hence no additional computations are required. This is in contrast to the need for search in the RAN formulation. The incremental network performance on a function interpolation problem is shown to be superior in convergence speed and approximation accuracy than the RAN networks and a fixed size RBF network