Inference of Simulation Models in Digital Twins by Reinforcement Learning

The typical optimization and control activities of Digital Twins are driven by high-performance simulators. Due to the significant complexity of systems subject to digital twinning, constructing simulators of appropriate details is a costly and error-prone endeavor. To alleviate these problems, we propose an approach for inferring simulation models of Digital Twins by machine learning. Instead of learning the simulation model of one specific simulator, we aim at learning their construction process. This generality enables reusing the inferred knowledge in different (but congruent) Digital Twin settings. To achieve this level of generality, we propose the Discrete Event System Specification (DEVS) formalism for capturing simulation models; and reinforcement learning (RL) for inferring DEVS models. In this paper, we explore the opportunities and challenges in combining these two techniques.