Recurrent CMAC: a powerful neural network for system identification

This paper deals with the application of neural networks for non-linear dynamic system modelling. Neural networks are non-linear black-box model structures where usually some training methods are used to estimate the network parameters (weights) using only input and output measurement data. In this paper some network architectures are suggested, where a special static network-Cerebellar Model Articulation Controller (CMAC)-and linear (FIR) filters are combined in different ways. Two possibilities are presented: in the first case the weights of the trainable layer are replaced by FIR filters, in the second case the inputs of the network are produced using linear filters. Both versions can be applied in feedforward or feedback structures. The paper deals with the feedback structures, determines their modelling capabilities and derives the training equations. In the end some simulation results are given to illustrate the possibilities and the limitations of the suggested networks.