Scholarlink

<p>Fog computing supports low-latency coordination among distributed edge nodes, yet real deployments must cope with uncertain workloads, partial observability, and intermittent links. This study develops a federated reinforcement learning (FRL) framework designed to remain stable under such non-ideal conditions. The approach combines three complementary mechanisms: (i) multi-kernel fuzzy preprocessing to capture task semantics, (ii) eligibility-trace–based policy initialization for faster adaptation, and (iii) topology-aware dissemination that prioritizes structurally influential nodes during aggregation. Together, these components enable faster convergence, lower communication load, and improved allocation accuracy in heterogeneous fog settings. Simulations on dynamic fog topologies show consistent performance gains: convergence time shortened by 40.3% relative to classical RL, energy use reduced by 28.0% compared with baseline FRL, latency lowered by 18.8% versus GNN-FL, and accuracy improved by 11.2% over FedAvg. Communication volume decreased by roughly 41.5% compared with full-broadcast aggregation. Statistical significance was verified using one-way ANOVA and Mann–Whitney U tests <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\((p&lt;0.05)\)</EquationSource> </InlineEquation>. Beyond numerical results, the framework illustrates a cross-domain learning paradigm that links cooperative multi-agent behaviors to fog-level task orchestration in the Internet of Everything (IoE). It offers a practical, interpretable, and resource-efficient architecture, supported by a lightweight and reproducible implementation that can be readily deployed on federated fog infrastructures.</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

Enhanced federated reinforcement learning framework for task management in fog computing using eligibility traces and targeted model dissemination

  • 期刊论文

  • 发表日期：

    2026年2月28日

查看全文
  • Seyed Omid Azarkasb,
  • Seyed Hossein Khasteh

摘要

Fog computing supports low-latency coordination among distributed edge nodes, yet real deployments must cope with uncertain workloads, partial observability, and intermittent links. This study develops a federated reinforcement learning (FRL) framework designed to remain stable under such non-ideal conditions. The approach combines three complementary mechanisms: (i) multi-kernel fuzzy preprocessing to capture task semantics, (ii) eligibility-trace–based policy initialization for faster adaptation, and (iii) topology-aware dissemination that prioritizes structurally influential nodes during aggregation. Together, these components enable faster convergence, lower communication load, and improved allocation accuracy in heterogeneous fog settings. Simulations on dynamic fog topologies show consistent performance gains: convergence time shortened by 40.3% relative to classical RL, energy use reduced by 28.0% compared with baseline FRL, latency lowered by 18.8% versus GNN-FL, and accuracy improved by 11.2% over FedAvg. Communication volume decreased by roughly 41.5% compared with full-broadcast aggregation. Statistical significance was verified using one-way ANOVA and Mann–Whitney U tests \((p<0.05)\) . Beyond numerical results, the framework illustrates a cross-domain learning paradigm that links cooperative multi-agent behaviors to fog-level task orchestration in the Internet of Everything (IoE). It offers a practical, interpretable, and resource-efficient architecture, supported by a lightweight and reproducible implementation that can be readily deployed on federated fog infrastructures.