Learning and reachability analysis for stochastic hybrid systems using mixtures of Gaussian processes
Hamzah A. Abdel-Aziz, Xenofon Koutsoukos · 2016
Robust and efficient modeling and reachability analysis of stochastic hybrid systems for control and decision is very demanding and challenging task. In this paper, we develop a novel methodology which provides a model of stochastic hybrid systems based on Gaussian Processes. This model uses observed data to update the model in an online fashion. In addition, we provide an efficient reachability analysis methodology that utilizes mixtures of Gaussian Processes to predict the reachable states for a finite horizon. We demonstrate the efficiency of the proposed approach using a multi-room heating system. Despite dynamic changes in the system parameters, the results show that the model can adapt and efficiently predict the reachable states.