Unsupervised Adaptive Neural-Network Control of Complex Mechanical Systems

Unsupervised adaptive control strategies based on neural-networks are presented. The tasks are performed by two independent networks which act as the plant identifier and the system controller. A new learning algorithm using information embedded in the identifier to modify the action of the controller has been developed. Simulation results are presented showing that this system can learn to stabilize a difficult benchmark control problem, the inverted pendulum, without requiring any external supervision.