An improved scheme for direct adaptive control of dynamical systems using backpropagation neural networks (Q1901484): Difference between revisions

The neural networks based direct control scheme for the learning control of dynamical systems is studied. Let \(y\) and \(y_d\) be a system state and desired system state respectively. The problem consists in on-line learning of the characteristics of the dynamics and generating the control signal \(u\) to drive the system from the state \(y\) to the state \(y_d\). To do it the objective function \(E_y= {1\over 2} (y_d- y)^2\) has to be minimized by the neural network controller. A section of the paper gives a brief review of the known methods proposed to derive the network error from the system error measurement. Problems usually arising in implementations of these methods are also discussed. The paper is devoted to describing the modified architecture in which the network interconnection strengths are updated using the system output error rather than the transformed error used in the other schemes. The ill effects of the controlled dynamics on the on-line updating of the network weights are moderated by including a gain layer (a single layer neural network with linear transformation) between the controller and the system. Both single input single output and multiple input multiple output dynamics are considered. Simulation studies conducted for application of the proposed scheme to control the nonlinear dynamics of an autonomous underwater vehicle are presented. The results of the simulation are compared with those for known direct control schemes.