Divide and conquer identification using Gaussian process priors

We investigate the reconstruction of nonlinear systems from locally identified linear models. It is well known that the equilibrium linearisations of a system do not uniquely specify the global dynamics. Information about the dynamics near to equilibrium provided by the equilibrium linearisations is therefore combined with other information about the dynamics away from equilibrium provided by suitable measured data. That is, a hybrid local/global modelling approach is considered. A non-parametric Gaussian process prior approach is proposed for combining in a consistent manner these two distinct types of data. This approach seems to provide a framework that is both elegant and powerful, and which is potentially in good accord with engineering practice.