An Empirical Evaluation of Robust Gaussian Process Models for System Identification

System identification comprises a number of linear and non-linear tools for black-box modeling of dynamical systems, with applications in several areas of engineering, control, biology and economy. However, the usual Gaussian noise assumption is not always satisfied, specially if data is corrupted by impulsive noise or outliers. Bearing this in mind, the present paper aims at evaluating how Gaussian Process (GP) models perform in system identification tasks in the presence of outliers. More specifically, we compare the performances of two existing robust GP-based regression models in experiments involving five bench-marking datasets with controlled outlier inclusion. The results indicate that, although still sensitive in some degree to the presence of outliers, the robust models are indeed able to achieve lower prediction errors in corrupted scenarios when compared to conventional GP-based approach.