System Identification through RBF Neural Networks: Improving Accuracy by a Numerical Approximation Method for the Centroids and Widths Adjustment

Within the last two decades there has been an increasing need for the development of mathematical models out of observed data captured from a system, a process called empirical modelling or systems identification. Under this circumstance, many techniques and methodologies have been proposed, among them the use of Artificial Neural Networks. It is proposed herein a non-hybrid gradient-based learning algorithm for a Radial Basis Function Neural Network aimed at improving the accuracy of non-linear dynamical system modelling. A single-stage non-hybrid approach is employed for the learning process, where the free parameters of the network – the centroids positioning, the receptive fields width, and the weights – are updated through a supervised method. Accurate identification capability is examined by the use of two non-linear datasets and the performance of the proposed method is compared with traditional techniques. Results demonstrate that nonlinear system identification can be significantly improved with easy-toimplement gradient-based RBF learning strategy.