Enhanced Gaussian Process Regression for Active Learning Model-based Predictive Control

Model predictive control relies on suitable and sufficiently accurate model of the system dynamics. With the increasing amount of available system data, learning-based MFC considers the automatic adjustment of the system model during operation using machine learning methods. However, most learning approaches only passively leverage the available system data, which results in a slow learning with lacking of informative data. In this paper, we apply gaussian process regression model to assess the residual model uncertainty and reward the system probing by introducing an information content cost in the optimization problem. Based on this, we propose an active learning-based MFC scheme that actively seek informative system data. Finally, we experiment with a Van der pol oscillator and show the effect of our algorithm.