Scholarlink

<p>Fog computing has become a pivotal paradigm for supporting latency-critical and computation-intensive Internet of Things (IoT) applications by enabling decentralized processing closer to data sources. Nevertheless, efficient task offloading in fog environments remains a complex challenge due to dynamic network topologies, heterogeneous resource capabilities, and time-varying task arrival rates. Existing reinforcement learning and federated learning–based scheduling approaches often experience slow convergence, synchronization overhead, and limited adaptability under rapidly changing fog conditions. To overcome these limitations, this paper introduces a Meta-Learning Assisted Graph Reinforcement Learning (MetaGRL) framework for intelligent task offloading in fog computing systems. The fog infrastructure is modelled as a dynamic graph, where vertices represent IoT devices and fog nodes, and edges encode communication latency and resource constraints. A Graph Neural Network is utilized to capture spatial and topological dependencies within the fog network, while an Actor–Critic reinforcement learning architecture learns optimal offloading decisions. To further enhance adaptability and accelerate policy convergence, a Model-Agnostic Meta-Learning strategy is incorporated, enabling the learning agent to rapidly adapt to new and unseen network states using limited training samples. Unlike asynchronous federated learning-based solutions, the proposed MetaGRL framework eliminates synchronization delays and facilitates fast policy generalization across diverse fog environments. Extensive simulation results demonstrate that MetaGRL consistently outperforms state-of-the-art methods, including GRL without meta-learning, DROO, HEFT, and EG. Performance evaluations indicate that MetaGRL achieves up to 22% improvement in offloading gain, approximately 32% reduction in task response latency, and a 17% increase in task completion rate compared to baseline approaches. These results validate the effectiveness of MetaGRL as a scalable, adaptive, and high-performance solution for real-time task offloading in dynamic fog computing environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Topology-Aware IoT–Fog Task Offloading via Graph Neural Networks and Meta-enhanced Reinforcement Learning with Energy–Latency Optimization

  • 期刊论文

  • 发表日期：

    2026年5月30日

查看全文
  • A. Sivakumar,
  • G. Arthy,
  • S. Srithar,
  • K Saritha,
  • T. Jayachandran,
  • M. Premalatha

摘要

Fog computing has become a pivotal paradigm for supporting latency-critical and computation-intensive Internet of Things (IoT) applications by enabling decentralized processing closer to data sources. Nevertheless, efficient task offloading in fog environments remains a complex challenge due to dynamic network topologies, heterogeneous resource capabilities, and time-varying task arrival rates. Existing reinforcement learning and federated learning–based scheduling approaches often experience slow convergence, synchronization overhead, and limited adaptability under rapidly changing fog conditions. To overcome these limitations, this paper introduces a Meta-Learning Assisted Graph Reinforcement Learning (MetaGRL) framework for intelligent task offloading in fog computing systems. The fog infrastructure is modelled as a dynamic graph, where vertices represent IoT devices and fog nodes, and edges encode communication latency and resource constraints. A Graph Neural Network is utilized to capture spatial and topological dependencies within the fog network, while an Actor–Critic reinforcement learning architecture learns optimal offloading decisions. To further enhance adaptability and accelerate policy convergence, a Model-Agnostic Meta-Learning strategy is incorporated, enabling the learning agent to rapidly adapt to new and unseen network states using limited training samples. Unlike asynchronous federated learning-based solutions, the proposed MetaGRL framework eliminates synchronization delays and facilitates fast policy generalization across diverse fog environments. Extensive simulation results demonstrate that MetaGRL consistently outperforms state-of-the-art methods, including GRL without meta-learning, DROO, HEFT, and EG. Performance evaluations indicate that MetaGRL achieves up to 22% improvement in offloading gain, approximately 32% reduction in task response latency, and a 17% increase in task completion rate compared to baseline approaches. These results validate the effectiveness of MetaGRL as a scalable, adaptive, and high-performance solution for real-time task offloading in dynamic fog computing environments.