Adaptive Digital Twin Architecture Using Reinforcement Learning for Autonomous Cloud-Edge Resource Orchestration
Main Article Content
Abstract
Modern cloud-edge computing deployments host increasingly heterogeneous and volatile workloads, which are hard to satisfy with conventional rule based resource orchestration mechanisms, without incurring non-compliant service-level agreement (SLA) violations or unnecessary operational cost. This paper presents the Adaptive Digital Twin Reinforcement Learning (AdaDT-RL) framework that integrates a high fidelity, continuously synchronized cloud-edge digital twin with a proximal policy optimization agent with the aim of making autonomous and anticipatory resource scaling decisions. The digital twin continually creates a multiphysics copy of the distributed infrastructure that allows the RL agent to test various counterfactual allocation policies in the virtual environment without risking real-world resources. An LSTM based anomaly detector is used for twin-physical divergence detection, which enables selective re-synchronization of the clock with a fidelity over 0.968 and a synchronization overhead of 31% of edge CPUs at intervals of 200 ms. AdaDT-RL boosts a composite resource efficiency score of 0.912 (±0.018) when evaluated across simulated cloud-edge clusters of 4 to 64 edge nodes, 5 workload intensity profiles, and shows 79.3% of SLA violations reduction compared to static allocation, while it helps reduce cloud resource cost by 41.8% at 64% workload intensity. The results are statistically significant and highly superior across 30 independent runs compared to five competing baselines, with effect sizes of Cohen's d > 2.6 in all comparisons (rejected value F = 391.2, p < 0.001, η² = 0.93 and Friedman rank tests χ² = 161.4, W = 0.86). Sobol' first-order sensitivity indices reveal that the discount factor (γ) and the frequency at which synchronization occurs are the main drivers of the performance, which gives direction for operational deployment.