Reinforcement Learning for MPC: Fundamentals and Current Challenges

Abstract

Recent publications have laid a solid theoretical foundation for the combination of Reinforcement Learning and Model Predictive Control, in view of obtaining high-performance data-driven MPC policies. Early practical results, both in simulation and in experiments, have shown the potential of this combination but have also revealed certain challenges. In addition, the technical complexity of these results makes it difficult for interested readers to gather the fundamental ideas and principles behind this combination. This paper aims to provide a coherent and more accessible picture of these results and to offer significantly deeper and more mature insights into their meaning than has been proposed before. It also aims at identifying the current challenges in the field.