Adaptive NeuroControl: How Black Box and Simple can it be

Jean-Michel Renders, Hugues Bersini, Marco Saerens

Abstract: In this paper three objectives are pursued. The first one is to connect two among the seminal neurocontrol methodologies due to Jordan and Narendra, and to make clear how the first one is obtained by a simplification of the second one. The next objective is to alleviate more and more the prior knowledge about the process required to achieve its control. In particular, we propose three extensions of Jordan's method making feasible the control of processes characterized by an arbitrary time delay, and which are powerful enough to be used for a large set of applications while preserving the simplicity inherent to neural network approaches in general. Finally, we show by simulations that these methods, although resulting from an approximation of Narendra's gradient descent approach, do not suffer from any degradation regarding the control performance while relaxing the necessary prior knowledge of the process and gaining in simplicity and computer economy.