Risk-Aware Bandits for Digital Twin Placement in Non-Stationary Mobile Edge Computing

In the context of the metaverse, digital twins (DTs) enable the convergence of the virtual space and the physical world. To this aim, in this work, we investigate DT placement in Mobile Edge Computing (MEC) systems, i.e., the selection of an edge server (ES) to host the DT of a physical system (PS). DT placement in MEC is a challenging task. Firstly, the dynamic characteristics of MEC systems, e.g., wireless channels and the loads of the ESs, are unknown and can exhibit a statistically non-stationary behavior. Secondly, the accuracy of the DT relies on periodically synchronizing the PS with its DT. However, in order for the synchronization to be successful, the incurred latency must be below a predefined deadline. Thirdly, the synchronization should be energy efficient as many PSs are battery powered. Lastly, switching between ESs causes additional overhead as it requires the migration of the DT to a new ES. In this work, we investigate the DT placement problem in a dynamic and statistically non-stationary MEC system. Our goal is to jointly minimize the synchronization latency and energy consumed by the PS while accounting for the overhead caused by switching between ESs. Furthermore, we aim for reducing the risk of failed synchronization events. To this end, we propose a novel risk-aware piece-wise stationary Multi-Armed Bandit (MAB) algorithm. Our simulations verify that our proposed algorithm outperforms state-of-the-art schemes by 50% and 58% in terms of the percentage of failed synchronizations and the number of DT migrations, respectively.