Performance and safety of Bayesian model predictive control: Scalable model-based RL with guarantees

Despite the success of reinforcement learning (RL) in various research fields, relatively few algorithms have been applied to industrial control applications. The reason for this unexplored potential is partly related to the significant required tuning effort, large numbers of required learning episodes, i.e. experiments, and the limited availability of RL methods that can address high dimensional and safety-critical dynamical systems with continuous state and action spaces. By building on model predictive control (MPC) concepts, we propose a cautious model-based reinforcement learning algorithm to mitigate these limitations. While the underlying policy of the approach can be efficiently implemented in the form of a standard MPC controller, data-efficient learning is achieved through posterior sampling techniques. We provide a rigorous performance analysis of the resulting `Bayesian MPC' algorithm by establishing Lipschitz continuity of the corresponding future reward function and bound the expected number of unsafe learning episodes using an exact penalty soft-constrained MPC formulation. The efficiency and scalability of the method are illustrated using a 100-dimensional server cooling example and a nonlinear 10-dimensional drone example by comparing the performance against nominal posterior MPC, which is commonly used for data-driven control of constrained dynamical systems