An affine Gaussian process approach for nonlinear system identification

The traditional Gaussian Process model is not analytically invertible. In order to use the Gaussian Process model for Internal Model Control, numerical approaches have to be used to find the inverse of the model. The numerical search for the inverse of each sample increases the already large computational load. To reduce the computation load an Affine Local Gaussian Process Model Network, as a combination of traditional Local Model Network and non-parametrical Gaussian Process Prior approach, is proposed in this paper. A novel algorithm for structure optimisation is introduced and exact inverse of the proposed network is derived. An Affine Local Gaussian Process Model Network and its inverse are illustrated on a simulated example.