News

Reinforcement Learning (RL) can be used to tune MPC schemes towards closed-loop optimality in the real world. That tuning is fundamentally different than using classical fitting approaches to adjust the model underlying MPC. These ideas are supported by a rich theory on how MPC is a universal approximation for Markov Decision Processes, provided that it is viewed in the right way. That theory proposes a bridge between the open-loop predictions of MPC and its closed-loop implementation. It also explains why using classical model fitting is sometimes good enough for achieving near-optimal closed-loop performance with MPC, and why it often falls short of achieving that. It finally opens a theoretical pathway on using (imperfect) simulation models to build optimal policies for the real-world, and on closing the “sim2real gap”. In this talk, we will provide an overview of these recent advances.