Gain scheduling control of nonlinear systems based on neural state space models

This paper presents a novel method for gain scheduling control of nonlinear systems based on extraction of local linear state space models from neural networks with direct application to robust control. A neural state space model of the system is first trained based on in- and output training samples from the system, after which linearized state space models are extracted from the neural network in a number of operating points according to a simple and computationally cheap scheme. Robust observer-based controllers can then be designed in each of these operating points, and gain scheduling control can be achieved by interpolating between each controller. In this paper, we propose to use the Youla-Jabr-Bongiorno-Kucera parameterization to achieve a smooth scheduling between the operating points with certain stability guarantees.