Accelerating MPC by online detection of state space sets with common optimal feedback laws

Abstract

Model predictive control (MPC) samples a generally unknown and complicated feedback law point by point. The solution for the current state $x$ contains, however, more information than only the optimal signal $u$ for this particular state. In fact, it provides an optimal affine feedback law $x\rightarrow u(x)$ on a polytope $\Pi \subset \mathbb{R}^n$, i.e., on a full-dimensional state space set. It is an obvious idea to reuse this affine feedback law as long as possible. Reusing it on its polytope $\Pi$ is too conservative, however, because any $\Pi$ is a state space set with a common affine law $x\rightarrow (u_0^\prime (x), \dots, u_{N-1}^\prime (x))^\prime\in\mathbb{R}^{Nm}$ for the entire horizon $N$. We show a simple criterion exists for identifying the polytopes that have a common $x\rightarrow u_0(x)$, but may differ with respect to $u_1(x), \dots, u_{N-1}(x)$. Because this criterion is too computationally expensive for an online use, we introduce a simple heuristics for the fast construction of a subset of the polytopes of interest. Computational examples show (i) a considerable fraction of QPs can be avoided (10% to 40%) and (ii) the heuristics results in a reduction very close to the maximum one that could be achieved if the explicit solution was available. We stress the proposed approach is intended for use in online MPC and it does not require the explicit solution.