Stochastic Hybrid Model Predictive Control using Gaussian Processes for Systems with Piecewise Residual Dynamics

Abstract

Data-driven control methods have been used to provide performance and safety benefits for systems where lower fidelity nominal dynamics models are insufficient when operating systems at their limits. These methods typically make the implicit assumption that the underlying model is unimodal and does not vary at different points in the workspace. However, when dealing with systems operating under the effect of piecewise unmodelled dynamics, approximating these by unimodal learnt models leads to inaccuracies in prediction over a horizon affecting performance and safety. In contrast, this thesis proposes the learning of hybrid models for use in a hybrid Model Predictive Control (MPC) framework to address these issues. An algorithm to help with improving the computational tractability of such a controller is also developed. Finally, a methodology is demonstrated that allows for efficiently identifying the active mode (component of unmodelled dynamics) in effect at different points of the workspace by leveraging the information contained in the hybrid model.