Kernel-based system identification improving robustness to input disturbances: A preliminary study on its application to AR models

This paper discusses the kernel-based identification which takes account of robustness against input disturbances. To this end, ARX models are employed in the framework of kernel-based system identification. Since ARX models utilize the past output data, which reflects the actual input including the disturbances, as well as the input data, its output estimate is expected to be more accurate compared to conventional FIR models. We will show the Tuned and Correlated kernel, which has been widely used in FIR model identification, works well for ARX models through numerical case studies with real data. Furthermore, results show that the stability of the identified ARX models is eventually maintained and they are superior to FIR models in robustness against input disturbances.