Digital Twins Paradigm: A Systematic Review from the Reinforcement Learning Perspective

Abstract

The Digital Twins (DT) paradigm has emerged as a powerful tool for simulating and analyzing complex systems in various domains. A DT is a virtual representation of a real-world object(s) whose goal is to accurately emulate real systems, optimize processes, minimize synchronization delays, cut down on overhead, and automate decision-making. DT technology is moving at a faster than expected pace with advances in Artificial Intelligence (AI), Internet of Things (IoT), Distributed Computing, and 5/6G. Being a highly beneficial technology, DT still faces issues of - (1) limited adaptability, (2) incomplete model representation, (3) suboptimal decision making, (4) limited generalization, and (5) scalability and computational efficiency. Reinforcement Learning (RL) offers unsupervised decision-making and intelligence, which can be immensely beneficial in addressing the current challenges faced by DT. This study offers a thorough analysis of the DT paradigm from the standpoint of RL. The survey compares and contrasts existing reinforcement learning-based Digital Twin frameworks, assessing their advantages and disadvantages. Moreover, discussions of approaches highlighting the tradeoffs between simulation fidelity and computing complexity is also studied. Additionally, a thorough understanding of the Digital Twins paradigm from a reinforcement learning perspective, is presented as a helpful resource for academics and industry professionals in the field. Finally, future research directions in this developing field at the nexus of digital modeling, simulation, and artificial intelligence is discussed.