Architecture for Digital Twin-Based Reinforcement Learning Optimization of Cyber-Physical Systems

Elias Modrakowski, Niklas Braun, Mehrnoush Hajnorouzi, Andreas Eich, Narges Javaheri, Richard Doornbos, Sebastian Moritz, Jan-Willem Bikker, Rutger van Beek

Abstract: The optimization of complex cyber-physical systems is a crucial task for their correct functioning, usability, and commercial viability. Due to their complexity, scale and resource intensiveness, conventional manual optimization is infeasible in many instances. We investigate the combination of the Digital Twin paradigm and Reinforcement Learning framework to address the long response times, limited availability of data, and the intractability of such systems. Here, the Digital Twin functions as the training environment in different development phases of the optimization. In this position paper we showcase our ongoing research on developing a reference architecture of a Digital Twin-Artificial Intelligence optimization system. This includes presenting the development process of the optimization system in terms of phases, an architecture from four viewpoints and an exemplary implementation.