Control-oriented Denoising Autoencoder: Robustified Data-Driven Model Reduction

Abstract Controllability quantification and model reduction of complex systems play an important role in many scientific and engineering fields. For this problem, the authors proposed a data-driven method based on statistical learning using neural networks, which we refer to as the control-oriented autoencoder. The important feature is that it is applicable to nonlinear systems, and that a suitable nonlinear projection can be given. For some systems, however, the method is excessively sensitive to computational error. In this paper, we analyze these phenomena from an over-fitting viewpoint, and enhance the robustness via denoising technique.