An optimizing geo-distributed edge layering with double deep Q-networks for predictive mobility-aware offloading in mobile edge computing

Abstract

In Mobile Edge Computing (MEC), the exponential growth of connected devices and user mobility presents significant challenges in optimizing task offloading, reducing latency, and energy usage. Predictive and adaptive task offloading mechanisms are essential as devices become more mobile and generate demanding tasks. Current methods, such as local computing and random scheduling, struggle to efficiently manage resources and maintain Quality of Service (QoS) in dynamic environments. This paper proposes an optimized Geographic Distributed Edge Layering (GDEL) architecture integrated with Double Deep Q-Networks (DDQN) to enable predictive, mobility-aware offloading. Our model leverages reinforcement learning through a Markov Decision Process (MDP) framework to dynamically allocate resources across distributed edge nodes, making optimal decisions on whether to offload or process tasks locally based on real-time conditions. Simulations show that our model outperforms other methods in key performance metrics, reducing task completion time by up to 48 %, lowering offloading decision latency by 49.3 %, and decreasing energy consumption by 26.5 % compared to traditional models.