Optimal control and reinforcement learning share a common objective: the design of decision-making policies that optimize long-term performance objectives in dynamical systems. Despite their different historical origins – with optimal control rooted in mathematics and physics, and reinforcement learning in machine learning and artificial intelligence - their foundations are deeply interconnected.

In this talk, starting from the optimal-control formulation, I will introduce the building blocks of reinforcement learning and explore the similarities and differences of the two approaches. Then, I will present a few approaches that I have recently developed to mitigate limitations of state-of-the-art reinforcement learning algorithms, namely (I) data inefficiency, (II) poor generalization to variations of the system dynamics, (III) limited robustness in presence of uncertainties and (IV) scalability to high-dimensional control problems. I will show applications of these reinforcement learning-based techniques to control problems frequently arising in engineering and applied science such as, for example, flow and density control, and navigation problems in unsteady flows. Eventually, I will provide an overview future research activities.


Contatto:
andrea1.manzoni@polimi.it