Relations between Model Predictive Control and Reinforcement Learning

Abstract

In this paper relations between model predictive control and reinforcement learning are studied for discrete-time linear time-invariant systems with state and input constraints and a quadratic value function. The principles of model predictive control and reinforcement learning are reviewed in a tutorial manner. From model predictive control theory it is inferred that the optimal value function is piecewise quadratic on polyhedra and that the optimal policy is piecewise affine on polyhedra. Various ideas for exploiting the knowledge on the structure and the properties of the optimal value function and the optimal policy in reinforcement learning theory and practice are presented. The ideas can be used for deriving stability and feasibility criteria and for accelerating the learning process which can facilitate reinforcement learning for systems with high order, fast dynamics, and strict safety requirements.