An Adversarial Approach to Adaptive Model Predictive Control

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