Output-error model training for gaussian process models

Abstract

The training of a regression model depends on the purpose of the model. When a black-box model of dynamic systems is trained, two purposes are particularly common: prediction and simulation. The purpose of this paper is to highlight the differences between the learning of a dynamic-system model for prediction and for simulation in the presence of noise for Gaussian process models. Gaussian process models are probabilistic, nonparametric models that recently generated interest in the machine-learning community. This method can also be used also for the modelling of dynamic systems, which is the main interest of the engineering community. The paper elaborates the differences between prediction- and simulation-purposed modelling in the presence of noise, which is more difficult in the case when we train the model for simulation. An example is given to illustrate the described differences. © 2011 Springer-Verlag.