.png){kind=icon}
.png){kind=icon}
Return
Multi-agent DRL-based task offloading and trajectory optimization for low altitude UAV IoT systems
.png){kind=icon}
PRE
AI
DOI：10.1016/j.adhoc.2026.104164.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
In low-altitude Internet of Things (IoT) networks, the unmanned aerial vehicle (UAV) is employed as a mobile edge node to provide computational services for task processing. However, the spatio-temporal dynamics of user devices (UDs) and the heterogeneity of task prioritization exacerbate the multidimensional resource competition encountered during the processing of tasks. This significantly affects energy consumption, service delay, and task completion rate, degrading user Quality of Service (QoS). To address these challenges, we propose a collaborative Multi-Agent Deep Reinforcement Learning (MADRL) algorithm to improve user QoS through the joint optimization of UAV three-dimensional (3D) trajectories, resource allocation, and task offloading strategies. Specifically, we design a Graph Convolutional Network (GCN)-based UAV actor network to optimize the dynamic trajectory by modeling user distribution in a topology-aware manner. In addition, we construct a centralized critic network based on a multi-head attention mechanism, wherein attention scaling is utilized to quantify differences in task demands and guide resource decision-making. These two components are jointly optimized through a ”topology association–demand difference” cooperative evaluation mechanism, enabling a multi-dimensional coupling of spatio-temporal characteristics and task demand decision-making. Experimental results demonstrate that the proposed algorithm reduces system energy consumption and delay by approximately 18.5% and 22.7%, respectively, while improving the task completion rate by about 16.2%.
Journal
IF:
4.8
Papers:
486
Citations:
6.2K