Scholarlink

For the real-time computing issues of Internet of Things devices (IoTD) in urban scenarios, this paper proposes a multi-unmanned aerial vehicles (UAV) collaborative mobile edge computing (MEC) network framework. In this framework, the UAVs avoid obstacles while communicating with ground-based IoTDs via probabilistic LoS channels. Subsequently, aiming to minimize the average delay of IoTD, we formulate an optimization problem that jointly optimizes connection scheduling, resource allocation, and three-dimensional (3D) trajectories of UAVs. To avoid long training times, a lightweight DRL based on knowledge distillation is proposed to optimize the flight trajectory. Specifically, we train the obstacle avoidance teacher model and the service teacher model separately, then perform model distillation according to the teacher weights, and update the teacher weights by minimizing the loss. Additionally, a low-time complexity greedy algorithm is designed to solve the connection scheduling problem, and the computing resource allocation problem is addressed through optimization theory. Compared to benchmarks, our algorithm significantly reduces the average latency of IoTDs while compressing the network size by 93%, and effectively reduces training time.

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

Average Delay Minimization Strategy for Multi-UAV Assisted MEC: A Lightweight Deep Reinforcement Learning Approach

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Heng Liu,
  • Qiang Tang,
  • Kun Yang,
  • Lei Yang

摘要

For the real-time computing issues of Internet of Things devices (IoTD) in urban scenarios, this paper proposes a multi-unmanned aerial vehicles (UAV) collaborative mobile edge computing (MEC) network framework. In this framework, the UAVs avoid obstacles while communicating with ground-based IoTDs via probabilistic LoS channels. Subsequently, aiming to minimize the average delay of IoTD, we formulate an optimization problem that jointly optimizes connection scheduling, resource allocation, and three-dimensional (3D) trajectories of UAVs. To avoid long training times, a lightweight DRL based on knowledge distillation is proposed to optimize the flight trajectory. Specifically, we train the obstacle avoidance teacher model and the service teacher model separately, then perform model distillation according to the teacher weights, and update the teacher weights by minimizing the loss. Additionally, a low-time complexity greedy algorithm is designed to solve the connection scheduling problem, and the computing resource allocation problem is addressed through optimization theory. Compared to benchmarks, our algorithm significantly reduces the average latency of IoTDs while compressing the network size by 93%, and effectively reduces training time.