This talk will consider the problem of optimizing prediction dynamics for robust linear MPC. By making connections with output feedback design problems, it will be shown that the optimal parameterization of degrees of freedom in predictions so as to maximize an ellipsoidal stabilizable set can be formulated as the solution of a convex problem. This involves a transformation of variables that reveals some fundamental properties of MPC laws based on ellipsoidal constraint approximations. Firstly the maximal stabilizable set is equal to the maximal invariant set under any linear feedback law. Secondly the maximal stabilizable set is only achieved if the prediction dynamics are allowed to vary over the prediction horizon depending on the evolution of the uncertain plant model. Online implementation and closed-loop properties will be described, and the talk will conclude with a discussion of the implications for dynamic feedback and MPC laws.
Back to Control Seminars Page