Self-Optimizing Control

Perry Y. Li Roberto Horowitz Department of Mechanical Engineering Department of Mechanical Engineering University of Minnesota University of California at Berkeley 111 Church St. SE Berkeley CA 94720-1740 Minneapolis MN 55455 [email protected] [email protected] Abstract Self-optimizing control problems arise in applications where the desired task is implicitly speci ed as an optimization of an objective function which could be a function of the unknown plant. The control objective therefore involves the explicit determination of both the optimal task and the control action to achieve that task. The di culty with this problem lies in the con ict between the need to identify the optimal task and to control the plant to achieve it. The proposed solution combines a reference generator and an adaptive controller. The reference generator, which provides the task to be executed by the adaptive controller, time multiplexes a training task to enable the system parameters to be identi ed, and an estimated optimal task. Switching is done in a manner that the training task becomes infrequently selected as the estimate for the optimal task improves.