arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Signal ProcessingEnglishPublished

Diffusion‑Based Graph Neural Policy Nearly Matches Optimal Throughput While Keeping Link Constraints in Check

November 8, 2026arXiv: 2611.08274v1

A new scheduling policy for wireless networks uses a diffusion model over graph signals to choose which links transmit each time slot. The model is trained by imitation learning, learning from sequences of optimal schedules produced by a dual‑gradient descent solver. The diffusion process is conditioned on the conflict graph of the network and on the minimum‑rate requirement, referred to as δ.

When tested, the policy produces sum‑rates that are 95–98 % of those achieved by the dual gradient descent (DGD) algorithm. Across all tested graph sizes from 10 to 5,000 links, the fraction of links that fail to meet their minimum transmission requirement stays below 1 % for δ ≤ 0.1. The same policy also generates schedules roughly 20 % faster per time slot than DGD, making it feasible for real‑time operation.

Figure 2 shows a comparison of the U‑GNN‑N diffusion model with the Dual Gradient Descent (DGD) heuristic. Panel (a) displays the objective value relative to DGD across various test graph sizes; panel (b) presents the proportion of links violating the minimum transmission requirement at δ=0.1, with outlined cells marking matched training‑testing sizes; panel (c) depicts the average inference time per schedule over 50 graphs, highlighting faster inference for U‑GNN.
Figure 2. Transferability and scalability across test graph sizes Ntest. U‑GNN‑N denotes a model trained on graphs containing N links. (a) Objective value as a percentage of the DGD objective. (b) Percentage of links violating the minimum transmission requirement at δ=0.1. Outlined cells indicate matching training and testing sizes. (c) Inference time per schedule averaged over 50 graphs.Shang, Teresa, Camargo, Romina Garcia, Uslu, Yiğit Berkay, Ribeiro, Alejandro · arXiv:2611.08274v1 · Figure 2: · CC BY 4.0

The diffusion policy can generalise to networks larger than those seen during training. When models were trained on 500‑link graphs, they still achieved superior performance on 5,000‑link graphs, suggesting good scalability.

However, the study found that if the required minimum transmission rate rises above δ = 0.1, the policy’s ability to satisfy all links declines sharply. At δ = 0.2, up to 20 % of links may violate their requirement, signalling that the current approach can struggle with very strict constraints.

These results show that diffusion‑based graph‑neural policies can come close to optimal throughput while maintaining link constraints, yet they also reveal a limit to how stringent those constraints may be before reliability drops.

[Figure 2(a)–(c) illustrates the objective performance, constraint violation rates, and inference time comparisons with DGD across graph sizes, supporting the comparison claim (c1) and revealing the diffusion model’s implementation via a U‑Graph Neural Network (m1).]

Source paper

Wireless Link Scheduling with Graph Signal Diffusion

Signal Processing
View on arXiv
Diffusion‑Based Graph Neural Policy Nearly Matches Optimal Throughput While Keeping Link Constraints in Check | arXiv News