A Comparison of Classical Identification and Learning-Based Techniques for Cyber-Physical Systems

This paper proposes a comparative study on the predictive accuracy of some classical system identification techniques with respect to recent learning-based approaches. The comparison is made on three case studies in the paradigm of Cyber-Physical Systems: a bilinear model whose parameters are derived from a real building experimental setup; an F-16 aircraft benchmark based on ground vibration experimental test data (with unknown model); a climate control experimental setup in a real building of the University of L'Aquila (with unknown model). For each case study, we compare diverse approaches producing predictive models of heterogeneous classes, whose accuracy is tested over a fixed time horizon. A discussion on the trade-off between predictive accuracy and model complexity is proposed with control applications in mind.