Multi-armed bandit-based federated reinforcement learning for dynamic task offloading in a containerized MEC

The development of mobile applications that are latency-sensitive, immersive, and compute intensive, notably mobile virtual reality (VR), augmented reality (AR), real-time video analytics, and IoT processing, has strained the edges of traditional cloud architectures. To reduce latency and improve responsiveness, mobile edge computing (MEC) has evolved to offload computation near the end users. However, the dynamism and resource constraints of edge environments and diverse task demands make efficient offloading a challenging multi-objective problem. This paper proposes a hybrid task offloading framework that leverages multi-armed bandit (MAB)-based algorithms for lightweight, real-time decision-making at the device layer and federated reinforcement learning (FRL) agents trained collaboratively across edge servers. The MAB component allows for adaptable performance under diverse workloads, whereas the FRL component enables scalable, privacy-conscious policy optimization. The frame work is deployed into a containerized MEC environment that provides both modularization and portability, allowing for better orchestration. Extensive simulations show that the proposed MAB–FRL framework achieves up to 22% higher QoE and 45%–47% reductions in latency and energy consumption compared to conventional local and greedy offloading schemes, while still outperforming centralized DRL by approximately 10% in QoE and 11% in latency and 19% in energy consumption. MAB as part of FRL provides a comprehensive adaptive and intelligent solution for dynamic task scheduling in forthcoming distributed edge ecosystems.