Semi-supervised learning of dynamical systems: a preliminary study

M. Mazzoleni, S. Formentin, M. Scandella, F. Previdi

System identification has, in recent years, drawn insightful inspirations from techniques and concepts of the statistical learning research area. Examples of this consist in the widely adoption of regularization and kernels methods, in order to better condition the identification problem. By pursuing the same purpose, we introduce the concept of semi-supervised learning to tackle the system identification challenge. The problem, casted into the framework of the Reproducing Kernel Hilbert Spaces, leads to a new regularization technique, called manifold regularization. An application to the identification of a NFIR model is carried out, and a comparison with the standard Tikhonov regularization technique is shown.