The N-step-ahead control method, called chain-back-propagation, is applied to articulated manipulators. The method is based on a pair of neural networks-the neural controller and the neural emulator-and focuses mainly in the appropriate design of the various global and chain-local penalty functions and the convergence control limitations, needed for the on-line training of the controller. No reference models or paths are needed for the implementation of the method, only set-points. The results, compared to conventional proportional-derivative (PD) control and to traditional one-step-ahead neural control, are quite satisfactory, indicating improved accuracy, faster response, and greater overall efficiency.
展开▼