Constrained Learning for Dynamical Systems

Learning has shown great success in recent years in controlling complex dynamical systems. However, for the most part, when training a policy most of the algorithms only consider a single objective function. However physical systems are required to satisfy a set of operation constraints, such as safety constraints or minimum performance levels. Naturally, one can express these problems as constrained optimization problems. These problems are in general non-convex and thus challenging. In this talk, I will establish that solving Reinforcement Learning problems under constraints is in fact not harder than solving unconstrained Reinforcement Learning problems.