Nonparametric identification of linearizations and uncertainty using Gaussian process models – application to robust wheel slip control

Gaussian process prior models offer a nonparametric approach to modelling unknown nonlinear systems from experimental data. These are flexible models which automatically adapt their model complexity to the available data, and which give not only mean predictions but also the variance of these predictions. A further advantage is the analytical derivation of derivatives of the model with respect to inputs, with their variance, providing a direct estimate of the locally linearized model with its corresponding parameter variance. We show how this can be used to tune a controller based on the linearized models, taking into account their uncertainty. The approach is applied to a simulated wheel slip control task illustrating controller development based on a nonparametric model of the unknown friction nonlinearity. Local stability and robustness of the controllers are tuned based on the uncertainty of the nonlinear models’ derivatives.