Scholarlink

<p>The internet-of-things applications are growing rapidly, which imposes stringent requirements for real-time, compute-intensive processing on resource-constrained user equipment (UE). Mobile edge computing addresses this challenge by enabling computation at the network edge and leveraging cloud resources when edge capacity is insufficient. However, most existing task offloading strategies rely on static or instantaneous resource states, limiting their adaptability to highly dynamic environments with fluctuating workloads and heterogeneous resources. To overcome this limitation, this study proposes a predictive and adaptive decision framework for task offloading that integrates LSTM-based resource prediction with a DDQN-based offloading policy. Specifically, the proposed TODRL integrates prediction-based resource allocation using long short-term memory with collaborative UE–edge–cloud offloading. This predictive capability enables the system to anticipate resource availability and optimally schedule delay-sensitive and computation-intensive tasks. The problem is presented as a dynamic optimization objective that minimizes a weighted system cost, jointly considering latency, energy consumption, service cost, and task acceptance ratio. Extensive simulations demonstrate that TODRL consistently outperforms state-of-the-art baselines by achieving lower energy consumption, reduced task execution delay, and improved system utility, validating the effectiveness of coupling resource prediction with reinforcement learning for adaptive edge computing.</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

Task offloading based on deep reinforcement learning in edge computing network

  • 期刊论文

  • 发表日期：

    2025年10月19日

查看全文
  • Shailja Kumari,
  • Divya Gupta

摘要

The internet-of-things applications are growing rapidly, which imposes stringent requirements for real-time, compute-intensive processing on resource-constrained user equipment (UE). Mobile edge computing addresses this challenge by enabling computation at the network edge and leveraging cloud resources when edge capacity is insufficient. However, most existing task offloading strategies rely on static or instantaneous resource states, limiting their adaptability to highly dynamic environments with fluctuating workloads and heterogeneous resources. To overcome this limitation, this study proposes a predictive and adaptive decision framework for task offloading that integrates LSTM-based resource prediction with a DDQN-based offloading policy. Specifically, the proposed TODRL integrates prediction-based resource allocation using long short-term memory with collaborative UE–edge–cloud offloading. This predictive capability enables the system to anticipate resource availability and optimally schedule delay-sensitive and computation-intensive tasks. The problem is presented as a dynamic optimization objective that minimizes a weighted system cost, jointly considering latency, energy consumption, service cost, and task acceptance ratio. Extensive simulations demonstrate that TODRL consistently outperforms state-of-the-art baselines by achieving lower energy consumption, reduced task execution delay, and improved system utility, validating the effectiveness of coupling resource prediction with reinforcement learning for adaptive edge computing.