A Novel Training Algorithm for Neuro-control of Discrete-time Nonlinear Systems

Guntur District, AP, India Guntur District, AP, India Physical systems are inherently nonlinear, and traditional methods of controlling such systems are based on a first-order linear approximation. Even though several nonlinear control design methods have been proposed in the past decades, most of these require analytical models. In contrast, artificial neural networks have been used as models for control since the paper published by Narendra and Parthasarathy in 1990. These recurrent networks are trained using the back-propagation algorithm, popular in other contexts. The principal issue was the notoriously slow convergence. More recently, an online sequential learning algorithm was proposed, which had better convergence properties. However, this algorithm applies to feedforward neural networks with a single hidden layer. We propose extending this algorithm’s applicability to networks with two hidden layers. The extension is achieved by incorporating error back-propagation. Further, we demonstrate that this novel algorithm has better convergence properties.