Abstract
This paper presents a novel approach to introducing adaptation in Model Predictive Control (MPC). Assuming limited a priori knowledge about the process, we consider a finite set of possible models (a dictionary), and use the theory of adversarial multi-armed bandits to develop an adaptive version of MPC called adversarial adaptive MPC (AAMPC). Under weak assumptions on the dictionary components, we then establish theoretical bounds on the performance of AAMPC and show its empirical behaviour via simulation examples.