An Efficient Task Offloading in Fog Computing Using Hybrid RL

Abstract

Fog computing brings computations closer to edge devices, which reduces the latency and energy consumption of tasks. However, when operating in a fog environment, task offloading decisions are exacerbated by the dynamic nature of network conditions and the diversity of available resources. In this paper, we propose an adaptive task-offloading framework to ensure that system reliability is maintained and deadlines are met while reducing latency and energy consumption. In addition to optimisation based methods and reinforcement learning approaches such as Q-learning and Deep Double Q-Networks (DDQN), many existing solutions struggle to adapt effectively.DDQN and Particle Swarm Optimization (PSO) are combined in this study to create a hybrid framework that addresses these challenges by combining their strengths of adaptive learning and efficient global search.Several key performance metrics, including latency, makespan, and energy consumption, are assessed in simulations and prototype implementations.This work extends the commonly used Google Cloud Jobs (GoCJ) dataset to include arrival times, CPU and memory requirements, bandwidth, deadlines, and task priorities, unlike previous studies that used simplified workloads.The latency and energy consumption are minimized under realistic fog workloads simulated through the extended dataset.