GBLinks: GNN-Based Beam Selection and Link Activation for Ultra-dense D2D mmWave Networks

Shiwen He, Shaowen Xiong, Wei Zhang, Yiting Yang, Ju Ren, Yongming Huang

Abstract

In this paper, we consider the problem of joint beam selection and link activation across a set of communication pairs to effectively control the interference between communication pairs via inactivating part communication pairs in ultra-dense device-to-device (D2D) mmWave communication networks. The resulting optimization problem is formulated as an integer programming problem that is nonconvex and NP-hard. Consequently, the global optimal solution, even the local optimal solution, cannot be generally obtained. To overcome this challenge, this paper resorts to design a deep learning architecture based on graph neural network to finish the joint beam selection and link activation, with taking the network topology information into account. Meanwhile, we present an unsupervised Lagrangian dual learning framework to train the parameters of the GBLinks model. Numerical results show that the proposed GBLinks model can converges to a stable point with the number of iterations increases, in terms of the weighted sum rate. Furthermore, the GBLinks model can reach near-optimal solution through comparing with the exhaustive search scheme in small-scale ultra-dense D2D mmWave communication networks and outperforms GreedyNoSched and the SCA-based method. It also shows that the GBLinks model can generalize to varying densities and coverage regions of ultra-dense D2D mmWave communication networks.

Introduction: The mmWave Interference Puzzle

Millimeter-wave (mmWave) communication is a cornerstone of high-speed 5G and future 6G networks. However, its high-frequency signals are directional and susceptible to blockage. In ultra-dense device-to-device (D2D) environments—such as crowded stadiums or smart factories—hundreds of devices attempt to beam signals simultaneously, creating a chaotic interference landscape. To maximize the weighted sum rate (throughput), the network must solve two critical problems simultaneously:

  1. Beam Selection: Identifying the optimal direction (codeword) for each transmitter and receiver.
  2. Link Activation: Deciding which devices should remain silent to minimize noise for the rest of the network.

Mathematically, this joint optimization is a non-convex, NP-hard integer programming problem. Traditional algorithms are often too slow for real-time deployment or fail to scale as network density increases.

System Model and Problem Formulation

The proposed system involves a central control center that manages multiple D2D pairs. Each pair consists of a transmitter and a receiver equipped with antenna arrays for directional communication.

System architecture of an ultra-dense D2D mmWave network. The control center collects Channel State Information (CSI) and distributes beam selection policies to manage interference among multiple transmitter-receiver pairs.

As shown in the figure above, the control center collects CSI and determines the beam selection policy. The challenge lies in the "interference puzzle": signals from one transmitter can interfere with unintended receivers, necessitating a sophisticated coordination strategy that includes the possibility of link inactivation to maintain network quality.

To address these challenges, the authors introduce GBLinks, a Graph Neural Network (GNN) based approach. GBLinks treats the wireless network as a graph where nodes represent transmitter-receiver pairs and edges represent the interference relationships between them. By using graph convolutions, each node learns from its neighbors' interference profiles. The architecture employs symmetric functions (MAX and MEAN) to ensure permutation invariance, meaning the model's performance is independent of how devices are indexed.

Unsupervised Lagrangian Dual Learning (LDLF)

A key innovation in GBLinks is the training method. Since finding "ground truth" optimal solutions for large networks is computationally prohibitive, the authors use an unsupervised Lagrangian dual learning framework. This approach uses the optimization constraints themselves (e.g., "only one beam per device") to construct a loss function, allowing the model to learn optimal strategies autonomously without labeled data.

Experimental Setup

The study evaluates GBLinks using synthetically generated datasets that simulate realistic mmWave environments. The training set consists of 8,000 samples, with 300 samples reserved for testing. Devices are randomly distributed in regions ranging from 5m x 5m to 50m x 50m. The channel model follows a narrowband clustered Saleh-Valenzuela model, capturing the sparse scattering characteristic of mmWave signals, while standard Discrete Fourier Transform (DFT) codebooks are used for beamforming.

Key Insights and Results

  • Near-Optimal Performance: In small-scale scenarios where exhaustive search is possible, GBLinks achieves approximately 90% of the optimal throughput.
  • Efficiency and Scalability: GBLinks is significantly faster than traditional iterative methods like Successive Convex Approximation (SCA). For a 30-user network, GBLinks reduces the computational burden from 10^22 operations to just 10^7. Furthermore, a model trained on a small network can generalize to larger networks without retraining.
  • Strategic Silence: The model effectively learns that inactivating certain links (link activation) is often the best way to boost overall network performance by reducing aggregate interference.

Conclusion

GBLinks demonstrates that GNNs, combined with unsupervised learning, can effectively manage the complexities of ultra-dense mmWave networks. By capturing network topology and bypassing the need for labeled data, it provides a scalable and efficient solution for real-time interference management. Future work may explore distributed learning and "Deep Unfolding" to further enhance interpretability and local decision-making.

Summary updated

Abstract

In this paper, we consider the problem of joint beam selection and link activation across a set of communication pairs to effectively control the interference between communication pairs via inactivating part communication pairs in ultra-dense device-to-device (D2D) mmWave communication networks. The resulting optimization problem is formulated as an integer programming problem that is nonconvex and NP-hard. Consequently, the global optimal solution, even the local optimal solution, cannot be generally obtained. To overcome this challenge, this paper resorts to design a deep learning architecture based on graph neural network to finish the joint beam selection and link activation, with taking the network topology information into account. Meanwhile, we present an unsupervised Lagrangian dual learning framework to train the parameters of the GBLinks model. Numerical results show that the proposed GBLinks model can converges to a stable point with the number of iterations increases, in terms of the weighted sum rate. Furthermore, the GBLinks model can reach near-optimal solution through comparing with the exhaustive search scheme in small-scale ultra-dense D2D mmWave communication networks and outperforms GreedyNoSched and the SCA-based method. It also shows that the GBLinks model can generalize to varying densities and coverage regions of ultra-dense D2D mmWave communication networks.
BibTeX
@article{he2022gblinks,
  author = {Shiwen He and Shaowen Xiong and Wei Zhang and Yiting Yang and Ju Ren and Yongming Huang},
  title = {{GBLinks: GNN-Based Beam Selection and Link Activation for Ultra-dense D2D mmWave Networks}},
  journal = {IEEE Trans. Communications},
  year = {2022},
  url = {https://openreview.net/forum?id=AG-wtcduzSs},
}
Figure