Dynamic safe active learning with NARX Gaussian processes

Abstract

Black-box modelling using Gaussian Processes has been widely and successfully studied and applied to model complex dynamic systems. So far, however, very little attention has been paid to the processes of obtaining the necessary data to train such systems in an efficient and safe manner. Zimmer et al. [ZMN18] proposed a Safe Active Learning framework for Time-Series Modeling with Gaussian Processes, which can be used to learn a Nonlinear Exogenous (NX) representation of a dynamic system in an efficient manner while considering safety constraints. In this masters’ thesis, the problem of efficiently and safely learning a Nonlinear Autoregressive Exogenous (NARX) representation of a dynamic system is addressed. With this purpose, an extension of the framework by Zimmer et al. was designed and implemented. Finally, the developed framework was evaluated in a real-world application. The results show an improvement on the original framework performance, as well as the suitability of the approach for real-world dynamic system modelling.