gQINN: Gradient-Enhanced Queue-Informed Neural Network for Estimating Queueing Delay in an Aggregation Network
IEEE Access · 2026 · 14 · 110171-110195
DOI: 10.1109/access.2026.3714448 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Kyota Hattori und 2 weiteren Autor:innen. Die Arbeit nennt 37 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Future carrier networks for industrial and logistics services must satisfy queueing delay targets under strict cost constraints. In aggregation networks, node-level delay estimation supports offline screening of candidate operating conditions before deployment, but dense delay labels are costly to obtain. Under sparse supervision, data-only estimators may fit observed delay values while producing unstable load-to-delay responses between labeled operating points. This paper proposes the Gradient-Enhanced Queue-Informed Neural Network (gQINN), a sparse-label estimator for queueing delay in aggregation networks. The main contribution is a sensitivity-aware queue-informed training formulation that regularizes both estimated delay values and local load sensitivity. The sensitivity term aligns the estimator derivative with respect to offered load with the analytical slope of a single-server M/G/1 reference queue. This gradient alignment constrains the load-to-delay response between sparse labels and supports bottleneck assessment under limited supervision. We evaluate gQINN on ns-3 datasets for passive optical network aggregation with deterministic multilayer perceptron and probabilistic neural process backbones. The results show that gQINN reduces queueing delay estimation error relative to data-only and queue-informed (QINN) baselines in most evaluated settings. In the representative fixed-weight sparse-label benchmark, gQINN with the multilayer perceptron backbone gives the lowest observed test mean squared error among the compared analytical, tree-based, data-only, and queue-informed baselines, while the additional gain over QINN remains regime dependent. Cost analysis using a floating-point operation model and measured wall-clock time shows that the added computation is confined to training, while inference uses the same forward computation as the corresponding fixed-backbone baseline.
Zitieren
APA
Hattori, K., Korikawa, T., & Takasaki, C. (2026). gQINN: Gradient-Enhanced Queue-Informed Neural Network for Estimating Queueing Delay in an Aggregation Network. IEEE Access, 14, 110171-110195. https://doi.org/10.1109/access.2026.3714448
Vancouver
Hattori K, Korikawa T, Takasaki C. gQINN: Gradient-Enhanced Queue-Informed Neural Network for Estimating Queueing Delay in an Aggregation Network. IEEE Access. 2026;14:110171-110195. doi:10.1109/access.2026.3714448
BibTeX
@article{Hattori2026,
author = {Hattori, Kyota and Korikawa, Tomohiro and Takasaki, Chikako},
title = {gQINN: Gradient-Enhanced Queue-Informed Neural Network for Estimating Queueing Delay in an Aggregation Network},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {110171--110195},
doi = {10.1109/access.2026.3714448}
}
EUVIMED