Abstract: The selection of learning rates to obtain satisfactory performances for neural network controllers is a challenging problem. In order to skip any time consuming experimentation for the choice of an appropriate value of the learning rate, this paper is concerned with an online adaptive learning rate algorithm derived from the convergence analysis of the usual gradient descent method. Based on the feedback linearization method, a multilayer neural network controller approximates online the unknown dynamics of the system including the non–linear behaviours. The proposed controller does not require any preliminary off-line training. A stability proof of this control scheme is given. Simulations and a comparison with a PD controller and several fixed learning rate neural controllers illustrate the effectiveness of the proposed algorithm in case of daptive control for robot trajectory tracking.
Noureddine Guersi , Messaoud Djeghaba , Dimitri Lefebvre , Fabrice Druaux and Edouard Leclercq , 2005. Tracking Control for Robot Manipulator Based on Neural Networks with Adaptive Learning Rate . Asian Journal of Information Technology, 4: 927-934.