The maximal positive invariant (MPI) set results from a finite set recurrence instantiated by the intersection of input and state bounds (e.g., the stage constraints of the linear model predictive control problem). When these constraints take the form of hyper-rectangles, zonotopes or constrained zonotopes, the resulting polyhedral MPI set may be succinctly described as a constrained zonotope, eliminating the need for explicit enumeration of its halfspaces. In this paper we discuss the various MPI computation algorithms (with both exact and sufficient stop conditions), recasted in the framework of constrained zonotopes. We analyze one of these variations over a dynamical system whose dimension can be arbitrarily increased in order to asses changes in computation time and storage requirements with respect to the polyhedral case (under the simplifying assumption of closed-loop invertibility of the state matrix).
The maximal positively invariant (MPI) set is obtained through a backward reachability procedure involving the iterative computation and intersection of predecessor sets under state and input constraints. However, standard static feedback synthesis may place some of the closed-loop eigenvalues at zero, leading to rank-deficient dynamics. This affects the MPI computation by inducing projections onto lower-dimensional subspaces during intermediate steps. By exploiting the Schur decomposition, we explicitly address this singular case and propose a robust algorithm that computes the MPI set in both polyhedral and constrained-zonotope representations.
We applied and adapted the linear encodings from [1] to the feasible-reference tracking model predictive control (MPC) formulation from [2] to reduce its computational cost. The improvements come from avoiding to explicitly use the vertex-based representation of the variable terminal set in testing its inclusion in the constraint set. We considered both polytopic and zonotopic formulations. For the later we have also proposed a positive invariant (PI) zonotopic approximation of the maximal PI set.