Model predictive control with memory-based discrete search for switched linear systems

Controlling systems with both continuous and discrete actuators using model predictive control is often impractical, since mixed-integer optimization problems are too complex to solve sufficiently fast. This paper proposes a parallelizable method to control both the continuous input and the discrete switching signal for linear switched systems. The method uses ideas from Bayesian optimization to limit the computation to a predefined number of convex optimization problems. The recursive feasibility and stability of the method is guaranteed for initially feasible solutions. Results from simulated experiments show promising performances and computation times.