Khan, Abdul Manan (2026) XAI-guided optimizer distillation for real-time UAV relay positioning in Terahertz 6G low-altitude ad hoc networks. Ad Hoc Networks, 194. ISSN 1570-8705

Preview

PDF 1-s2.0-S1570870526002556-main.pdf - Published Version Available under License Creative Commons Attribution. Download (2MB) | Preview

Abstract

Intelligent and resilient networking for the low-altitude economy requires autonomous unmanned aerial vehicle (UAV) relay positioning that adapts in real time to changing user topologies. In terahertz (THz) sixth-generation (6G) low-altitude aerial ad hoc networks (LAAHNs), this demands neural surrogates combining millisecond-scale inference with near-optimal placement accuracy. This paper presents a thirteen-method comparison — spanning heuristics, deep reinforcement learning (DRL), DeepSets, a multilayer perceptron (MLP), a Set Transformer, a fine-tuned language model, and a differential-evolution oracle — evaluated on 200 shared scenarios at 300 GHz. To improve the best-performing architecture, this paper introduces an explainable artificial intelligence (XAI)-guided optimizer distillation pipeline: a Set Transformer is first trained on 10,000 globally optimal solutions, then three explainability methods (attention heatmaps, Integrated Gradients,
GradientShap) identify user spatial coordinates as the dominant features, and a streamlined variant, ST-Lite, is retrained on those features alone. ST-Lite achieves the highest surrogate throughput: 194.2 Mbps (96.1% of the evolutionary oracle) at 1.2 ms—surpassing both the MLP (182.1 Mbps) and the language model (185.4 Mbps at 822 ms). ST-Lite also leads under building obstructions (156.8 vs. 148.1 Mbps for the MLP) and vehicular mobility up to 120 km/h, where millisecond-scale latency enables 100 ms-cadence repositioning that outperforms every non-oracle method. Multi-seed validation across five random seeds confirms reproducibility. Under 3rd Generation Partnership Project (3GPP) TR 38.901 fading, all methods compress to 72–79 Mbps; retraining ST-Lite directly on 3GPP-optimized data yields only +0.9% improvement, confirming that stochastic fading — not training-channel mismatch — is the binding constraint. A two-relay extension further shows the
surrogate composes to multiple relays, recovering 96.2% of a joint two-relay oracle at millisecond latency, and robustness is corroborated on a structured 3D city-grid model. All models run as Robot Operating System 2 (ROS 2) Jazzy nodes on a consumer GPU.

Item Type:	Article
Identifier:	10.1016/j.adhoc.2026.104389
Keywords:	UAV relay positioning; Low-altitude aerial ad hoc network; Terahertz 6G communications; Intelligent networking; Explainable AI; Optimizer distillation; Set transformer; Resilient autonomous systems; Low-altitude economy
Subjects:	Computing > Intelligent systems
Date Deposited:	29 Sep 2026
Dates:	Date Publication status 9 August 2026 Accepted 19 August 2026 Published Online
School, department or research centre:	School of Computing and Engineering
Keywords:	UAV relay positioning; Low-altitude aerial ad hoc network; Terahertz 6G communications; Intelligent networking; Explainable AI; Optimizer distillation; Set transformer; Resilient autonomous systems; Low-altitude economy
URI:	https://repository.uwl.ac.uk/id/eprint/15440
Sustainable Development Goals:	Goal 9: Industry, Innovation, and Infrastructure

Downloads

Actions (admin access)