Learning-Based Robust Model Predictive Control with State-Dependent Uncertainty

A robust model predictive control (RMPC) approach for linear systems with bounded state-dependent uncertainties is proposed. Such uncertainties can arise from unmodeled non-linearities or external disturbances, for example. By explicitly considering the state dependency of the uncertainty sets in the RMPC approach, it is shown how closed-loop performance can be improved over existing approaches that consider worst-case uncertainty. Being able to handle state-dependent uncertainties is particularly relevant in learning-based MPC where the system model is learned from data and confidence in the model typically varies over the state space. The efficacy of the proposed approach for learning-based RMPC is illustrated with a numerical example, where uncertainty sets are obtained from data using Gaussian Process regression.