Scholarlink

<p>Multi-access edge computing (MEC) faces the challenge of efficiently offloading diverse tasks from user devices to MEC hosts to reduce network traffic and latency. Existing deep reinforcement learning (DRL) methods often require full retraining and are inefficient in new environments. To address this issue, we introduce the deep meta reinforcement learning computation (DMRLC) framework. This framework transforms running applications on user devices into input nodes of Directed Acyclic Graphs, and it employs a custom Sequence-to-Sequence neural network to develop the offloading strategy. The optimal offloading strategy is obtained after fully training the DMRLC model, which involves optimizing the inner and outer loop, using the Sequence-to-Sequence neural network and Trust Region Policy Optimization (TRPO). Simulation results across various topologies, task sizes, and transmission rates show that TRPO achieves significant latency reductions compared to existing algorithms such as Greedy, HEFT, and Fine-Tuning DRL. Fine-Tuning DRL outperforms Greedy and HEFT but TRPO surpasses Fine-Tuning DRL, achieving 5.65% latency reduction across topologies, 2.66% improvement in task numbers and 6.78% latency reduction across bandwidth ranges. Moreover, TRPO achieves an average energy consumption reduction of 20.89% over Fine-Tuning DRL, 36.06% over HEFT, and 16.53% over Greedy, demonstrating its superior energy efficiency.</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

Deep meta reinforcement learning for efficient task offloading in edge computing environments

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Priyadarshni,
  • Dhruvan Kadavala,
  • Shivani Tripathi,
  • Praveen Kumar,
  • Rajiv Misra

摘要

Multi-access edge computing (MEC) faces the challenge of efficiently offloading diverse tasks from user devices to MEC hosts to reduce network traffic and latency. Existing deep reinforcement learning (DRL) methods often require full retraining and are inefficient in new environments. To address this issue, we introduce the deep meta reinforcement learning computation (DMRLC) framework. This framework transforms running applications on user devices into input nodes of Directed Acyclic Graphs, and it employs a custom Sequence-to-Sequence neural network to develop the offloading strategy. The optimal offloading strategy is obtained after fully training the DMRLC model, which involves optimizing the inner and outer loop, using the Sequence-to-Sequence neural network and Trust Region Policy Optimization (TRPO). Simulation results across various topologies, task sizes, and transmission rates show that TRPO achieves significant latency reductions compared to existing algorithms such as Greedy, HEFT, and Fine-Tuning DRL. Fine-Tuning DRL outperforms Greedy and HEFT but TRPO surpasses Fine-Tuning DRL, achieving 5.65% latency reduction across topologies, 2.66% improvement in task numbers and 6.78% latency reduction across bandwidth ranges. Moreover, TRPO achieves an average energy consumption reduction of 20.89% over Fine-Tuning DRL, 36.06% over HEFT, and 16.53% over Greedy, demonstrating its superior energy efficiency.