Componentwise Holder inference for robust learning-based MPC

Abstract

This article presents a novel learning method based on componentwise Holder continuity, which allows one to consider independently the contribution of each input to each output of the function to be learned. The method provides a bounded prediction error, and its learning property is proven. It can be used to obtain a predictor for a nonlinear robust learning-based predictive controller for constrained systems. The resulting controller achieves better closed loop performance and larger domains of attraction than learning methods that only consider nonlinear set membership, as illustrated by a case study.