Accelerated Model Predictive Control Using Restricted Quadratic Programming ?

摘要

We present a method to reduce the computational burden of solving a sequence of convex quadratic programs (QPs). By determining offline what search space is most important, we can restrict our online problem to that subspace, reducing the dimension and computational cost of the QP solver. The process we present is very simple requiring surprisingly little data. Further, we present a modified sequential QP algorithm that leverages the restricted QP approach to solve nonlinear programming problems found in model predictive control. Lastly, we apply these to a benchmark MPC problem and demonstrate their effectiveness using a variety of established QP solvers. We demonstrate that QP problems can be solved faster with minimal MPC performance degradation and highlight future directions for this work.