An escape method from local minimum by orbital correction method for controller learning

In this paper, an escape method from local minimum, which can be applied to nonlinear control systems based on universal learning network is presented. In the gradient optimization problems, a number of methods such as annealing, random search and multi-start method have been presented in order to escape from local minimum. The proposed method has different features from the conventional methods in which the escaping from local minimum can be achieved by changing the nonlinear system dynamics instead of changing parameters. Changing of the dynamics can be realized by adding a special term to the evaluation function of the universal learning network. In simulations which study the control of a nonlinear crane system by neural networks, comparisons between the proposed method and the conventional multi-start method are studied, and it is shown that the proposed method is superior in performance to the conventional method