Importance Sampling for Deep System Identification

This paper revisits the methodology of system identification and shows how new paradigms from machine learning can be used to improve the model identification performance in the case of non-linear systems observed with a noisy and unbalanced dataset. We prove that using importance sampling schemes in system identification can provide a significant performance boost on a wide variety of systems, in particular when some of the system dynamic is only exhibited by relatively rare events. The performance of the approaches is evaluated on a real and simulated drone and two standard datasets from real, robotic systems. Our approach consistently outperforms baseline approaches on these datasets, all the more when the datasets are noisy and unbalanced.