Scholarlink

<p>Edge computing plays a key role in smart IoT systems by reducing network load and improving application responsiveness. This paper presents a reinforcement learning (RL)-driven orchestration framework that jointly optimizes task offloading and dynamic pricing to enhance both economic profitability and energy efficiency. Unlike existing methods that simply map traditional objectives into RL, we propose a closed-loop design based on profit–energy coupling, where pricing, idle power, and task granularity are unified into monetary cost to enable balanced decision-making. We also introduce a unified state–reward abstraction for centralized, decentralized, and hybrid topologies, allowing policy reuse across deployment settings. Implemented in an extended EISim simulator, our system features real-time dual-loop feedback between pricing (via DDPG) and offloading (via RL agents). The performance analysis reveals that reinforcement learning (RL)-based techniques can consistently adjust their offloading strategies to balance local execution and edge offloading, thereby enhancing overall network performance. Moreover, we demonstrate how the adaptive capabilities of RL in conjunction with the pricing strategy set by DDPG to significantly affect the profitability and computational efficiency of the edge service platforms. The findings highlight the nuanced relationship of pricing, task offloading, and system topology in edge 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

Reinforcement learning-based intelligent edge orchestration for IoT: insights into performance and profitability

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Qianhua Luo,
  • Bo Xie,
  • Haixia Cui

摘要

Edge computing plays a key role in smart IoT systems by reducing network load and improving application responsiveness. This paper presents a reinforcement learning (RL)-driven orchestration framework that jointly optimizes task offloading and dynamic pricing to enhance both economic profitability and energy efficiency. Unlike existing methods that simply map traditional objectives into RL, we propose a closed-loop design based on profit–energy coupling, where pricing, idle power, and task granularity are unified into monetary cost to enable balanced decision-making. We also introduce a unified state–reward abstraction for centralized, decentralized, and hybrid topologies, allowing policy reuse across deployment settings. Implemented in an extended EISim simulator, our system features real-time dual-loop feedback between pricing (via DDPG) and offloading (via RL agents). The performance analysis reveals that reinforcement learning (RL)-based techniques can consistently adjust their offloading strategies to balance local execution and edge offloading, thereby enhancing overall network performance. Moreover, we demonstrate how the adaptive capabilities of RL in conjunction with the pricing strategy set by DDPG to significantly affect the profitability and computational efficiency of the edge service platforms. The findings highlight the nuanced relationship of pricing, task offloading, and system topology in edge computing environments.