Increasing performance of parametrizations for linear MPC via application of a data mining algorithm

A new type of parametrizations is proposed which allows to reduce the size of the online optimization in linear MPC. The parametrizations combine a first part ensuring feasibility and asymptotic stability of the closed loop and a second part promoting performance. The performance promoting part is determined a priori offline based on a data mining algorithm which has been introduced in our previous work. In comparison to this work, the new results provide full flexibility in the choice of training data and thereby allow application of the method to larger problems. This is verified in two numerical examples which illustrate the benefits of the new method.