Machine learning-based warm starting of active set methods in embedded model predictive control

We propose to apply artificial intelligence approaches in a warm-starting procedure to accelerate active set methods that are used to solve strictly convex quadratic programs in the context of embedded model predictive control (MPC). The proposed warm-starting is based on machine learning where a good initialization of the active set method is learned from training data. Two approaches to generate the training data set are discussed, one based on gridding the feasibility domain, and one based on closed-loop simulations with typical initial conditions. The training data are then processed by machine learning-based classification algorithms that yield a good estimate of the initial active set for the iterative active set algorithm. By means of extensive case studies we demonstrate that the proposed approach is superior to existing warm-starting procedures in that it considerably reduces the number of active set iterations, thus allowing embedded MPC to be implemented using less computational effort.