Abstract
In this paper, a new method named orbital correction to adjust the parameters of the controller of nonlinear systems is presented. Control systems are described by the Universal Learning Network, in which the controller and the controlled object can be dealt with using the same network architecture. And training of the parameters of the controller is carried out by the gradient method. The basic idea of this method is to introduce a kind of a priori information on control performance into the learning process in order to enhance the learning ability. From the simulations of a nonlinear crane control system, it is shown that the proposed method can obtain the better controller than the conventional methods by selecting an appropriate time and volume of orbital correction. And the proposed method may be useful to overcome the local minimum problem always encountered in the gradient method.
