Off-line identification of nonlinear systems using structurally adaptive radial basis function networks

This paper presents a novel off-line algorithm to train direct linear feedthrough radial basis function (DLF-RBF) networks. The algorithm basically explores the model error surfaces and combines an automatic determination of the number of RBF neurons with a hybrid optimization step to tune all parameters in the network. This leads to parsimonious models of SISO or MIMO dynamical systems, a primordial aim when solving nonlinear system identification problems. To demonstrate the effectiveness and the performance of the new method, it is applied to the identification of two highly nonlinear systems (one SISO and one MIMO system).