Adaptive Task Offloading Under Dynamic Topology via Graph Attention-Enhanced Reinforcement Learning for Layered UAV-Enabled Digital Twin Channel Modeling
- 1School of Communications and Information Engineering, Nanjing University of Posts and Telecommunications, Nanjing 210003, China
Abstract
This study considers the timely delivery of channel measurements for digital twin (DT) channel modeling in dynamic wireless environments. Layered uncrewed aerial vehicle (UAV) networks can separate sensing, offloading, and edge processing, but the active UAVs, available edge servers (ESs), and feasible service links may vary over time. We formulate a dynamic-topology task-offloading problem involving edge-reception age of information (AoI), service delay, offloading failures, and local backlog, and propose DTGA-RL. Its coordinated design combines local graph-attention encoding, shared candidate-wise action-value evaluation, and feasibility-aware action masking for a variable set of candidate ESs. In the reported tests with 15–27 working UAVs, the policy trained at the default 15-UAV scale is evaluated without retraining; task data sizes of 0.35–0.95 Mbit/slot are also tested. At 0.95 Mbit/slot, the plotted DTGA-RL values are approximately 34.9 slots of AoI, 3.13 s of service delay, and an 0.106 offloading failure rate, versus 35.3 slots, 3.85 s, and 0.134 for FixedRL, corresponding to approximate reductions of 1.1%, 18.7%, and 20.7%, respectively. The performance gap narrows at higher UAV counts. AoI is measured at ES reception.
Related
- CodeTD: Topology of Attention Detects Hallucinations in Code LLMs
- Data-Driven On/Off Topology Optimization of Electromagnetic Devices Based on Graph Physics-Informed Neural Networks and Physics-Guided Reinforcement Learning
- A fast risk assessment method for power balance under extreme weather conditions based on deep reinforcement learning with a coupled attention mechanism
- Adaptive hybrid coupling with operator inference, the overlapping Schwarz alternating method and reinforcement learning
- Co-Evolving Zero-Day Jamming: Adaptive Attack Synthesis and Graph Attention-Based Online Detection
BibTeX
@article{wei2026adaptive,
title = {{Adaptive Task Offloading Under Dynamic Topology via Graph Attention-Enhanced Reinforcement Learning for Layered UAV-Enabled Digital Twin Channel Modeling}},
author = {Yifei Wei and Taiming Zhang and Miao Liu},
journal = {Mathematics},
year = {2026},
month = nov,
doi = {10.3390/math14214065},
url = {https://www.mdpi.com/2227-7390/14/21/4065},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Adaptive Task Offloading Under Dynamic Topology via Graph Attention-Enhanced Reinforcement Learning for Layered UAV-Enabled Digital Twin Channel Modeling
AU - Yifei Wei
AU - Taiming Zhang
AU - Miao Liu
PY - 2026
DA - 2026/11/05
JO - Mathematics
DO - 10.3390/math14214065
UR - https://www.mdpi.com/2227-7390/14/21/4065
AB - This study considers the timely delivery of channel measurements for digital twin (DT) channel modeling in dynamic wireless environments. Layered uncrewed aerial vehicle (UAV) networks can separate sensing, offloading, and edge processing, but the active UAVs, available edge servers (ESs), and feasible service links may vary over time. We formulate a dynamic-topology task-offloading problem involving edge-reception age of information (AoI), service delay, offloading failures, and local backlog, and propose DTGA-RL. Its coordinated design combines local graph-attention encoding, shared candidate-wise action-value evaluation, and feasibility-aware action masking for a variable set of candidate ESs. In the reported tests with 15–27 working UAVs, the policy trained at the default 15-UAV scale is evaluated without retraining; task data sizes of 0.35–0.95 Mbit/slot are also tested. At 0.95 Mbit/slot, the plotted DTGA-RL values are approximately 34.9 slots of AoI, 3.13 s of service delay, and an 0.106 offloading failure rate, versus 35.3 slots, 3.85 s, and 0.134 for FixedRL, corresponding to approximate reductions of 1.1%, 18.7%, and 20.7%, respectively. The performance gap narrows at higher UAV counts. AoI is measured at ES reception.
SN - 2227-7390
ER -