Robustness of Neural-Network-based Nonlinear Iterative Learning Control

Abstract

The purpose of this work is to develop a robust iterative learning control for nonlinear systems based on neural networks. In order to introduce the robustness to the control scheme the problem of accurate estimation of uncertainties associated with the black-box type model is concerned. An uncertainty of the system is derived in terms of the variance of the model output prediction using a concept of Fisher information matrix well-known in the optimum experimental design theory. Once the bounds of the system response are estimated, they can be directly applied during training of the learning controller by a rigorous definition of the penalty cost function. Then, a neural controller is suitably adopted to the effective design of iterative learning control for nonlinear systems. The proposed approach is experimentally verified on the example of a magnetic levitation system.