A nonlinear system predictor from experimental data using neural networks

Abstract

A novel iterative technique is proposed by the authors in order to build a discrete-time nonlinear dynamical system predictor from experimental input-output pairs. The iterative technique is capable of representing a class of dynamical systems as static multidimensional mappings. The practical solution of the prediction problem is strongly related with the availability of suitable representations of multidimensional mappings. The proposed technique, belonging to the memory-based techniques, highly reduces the memory amount required to store the representation of the mapping. The iterative technique is very well suited to work in conjunction with an associative memory structure as the monodimensional CMAC and in presence of on-fly data. An application example to dynamical system output prediction is presented. Moreover, a convergence discussion for the proposed algorithm is provided. Finally, computer simulations verify the stated theory.