Adversarial Bandit Learning Assisted Online Optimization for Digital Twin Placement and Update in End-Edge-Cloud Collaboration

Le résumé fourni par la source

Digital twin (DT) is envisioned not only to perform the high-fidelity virtual representation of its corresponding physical entity (PE), but also to serve as an active agent delivering diverse types of sophisticated services. This paper studies an end-edge-cloud collaborative DT placement and update framework. Specifically, we consider that DTs are dynamically placed across edge servers (ESs) via migration following their paired PEs' potential mobility, while being supported by real-time data fetched from the cloud center and user ends. On top of this, we emphasize a unique feature that DTs should also be continually updated capturing the uncertain evolutions for both personalized service ability improvement and versatile service ability maintenance, where the personalization is improved by utilizing the experiential knowledge from the cloud center and their corresponding PEs, and the versatility is maintained by integrating pre-stored profiles. To maximize the long-term system-wide average weighted quality-of-service (QoS) in handling all types of PEs' service requests under the stringent system cost constraint, we formulate an online problem to jointly optimize DT migrations, service priorities towards various request types, and all related DT updating strategies. To address underlying difficulties, we propose a novel adversarial bandit learning assisted online optimization approach, called ARBOK. We first leverage the Lyapunov decomposition method to transform the long-term problem into multiple instant ones, each of which is further decoupled into two correlated subproblems. For solving one subproblem with a bilinear structure, we develop a McCormick envelopes based algorithm (MO-EL). Besides, we design an extended adversarial combinatorial multi-armed bandit algorithm (AC-BL) to tackle the other subproblem, which constructs a super arm set to resolve the issue of excessively large decision space and employs a robust scheme to handle the inherent uncertainty and non-stationarity in each super arm's loss function. We integrate both algorithms seamlessly into ARBOK and alternately execute them till the convergence. Theoretical analysis and extensive simulations show the effectiveness of the introduced dynamic DT placement and continual update framework, demonstrating that ARBOK can converge to the asymptotic optimum within a polynomial-time complexity while outperforming counterparts.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.