Online identification of nonlinear time-variant systems using structurally adaptive radial basis function networks

This paper presents a new algorithm to train direct linear feedthrough radial basis function (RBF) networks, especially designed for online identification of time-variant nonlinear dynamical systems. The algorithm basically explores the network's input space and the model error to determine automatically the number of RBF neurons, and to adapt their center positions (adaptive error dependent clustering). The widths and the output layer weights are adapted using two in series connected recursive least squares algorithms. This lead to parsimonious models of SISO or MIMO dynamical systems, a primordial aim when solving nonlinear system identification problems. The effectiveness and the performance of the new method is demonstrated by the identification of two highly nonlinear systems (time-invariant and time-variant types, respectively).