Dynamic systems identification using RBF neural networks

The identification of complex and non-linear plants plays an important role in the overall architecture of neurocontrol techniques as for example inverse control, direct and indirect neural adaptive control, etc. It is common within those approaches to use a Feedforward Neural Network (FNN) with Tapped Delay Line (TDL) or recurrent networks (Elman o Jordan) trained off-line to capture the system's dynamics (direct or inverse) and use it in the control loop. In this paper, we present an identification schema based on Radial Basis Function (RBF) neural networks that is trained on-line and dynamically modify his number of nodes in the hidden layer, allowing a real-time implementation of the identifier in the control loop. Copyright (C) 2007 CEA-IFAC.