Stability-Oriented Multi-Objective Container Scheduling for UAV-Assisted Edge Networks via TD3
TL;DR
MORA (Multi-Objective Resource Allocation), a stability-oriented deep reinforcement learning scheduler based on Twin Delayed Deep Deterministic Policy Gradient (TD3), with a normalized multi-objective reward that jointly optimizes energy, latency, SLA adherence, and proactively minimizes container migrations is proposed.
Abstract
UAV-assisted fog computing places drone-mounted edge nodes between IoT devices and the cloud, enabling ultra-low-latency services where fixed infrastructure is unavailable. However, battery constraints, node heterogeneity, and bursty workloads make container scheduling particularly difficult: unnecessary migrations deplete drone battery and violate SLA deadlines. We propose MORA (Multi-Objective Resource Allocation), a stability-oriented deep reinforcement learning scheduler based on Twin Delayed Deep Deterministic Policy Gradient (TD3), with a normalized multi-objective reward that jointly optimizes energy, latency, SLA adherence, and proactively minimizes container migrations. Evaluation via COSCO under AzureFog traces over 200 intervals demonstrates that MORA achieves 79.6% task completion, reduces migrations by 94% versus DRLMOTS (27 vs. 182), cuts energy by 21–27%, and limits SLA violations to 47 over 200 intervals with a stable per-interval range of 9–14, outperforming TD3, DDPG, and DRLMOTS across all metrics.