Feedback scheduling of model predictive controllers

The paper presents some preliminary results on dynamic scheduling of model predictive controllers (MPCs). In an MPC, the control signal is obtained by on-line optimization of a cost function, and the MPC task may experience very large variations in execution time from sample to sample. Unique to this application, the cost function offers an explicit, on-line quality-of-service measure for the task. Based on this insight, a feedback scheduling strategy for multiple MPCs is proposed, where the scheduler allocates CPU time to the tasks according to the current values of the cost functions. Since the MPC algorithm is iterative, the feedback scheduler may also abort a task prematurely to avoid excessive input-output latency. A case study is presented, where the new approach is compared to conventional fixed-priority and earliest-deadline-first scheduling. General problems related to the real-time implementation of MPCs are also discussed.