Power Allocation for Device-to-Device Interference Channel Using Truncated Graph Transformers

Dohoon Kim, Shenghui SongMeditCom 2023 · DOI · OpenReview

This paper presents a novel approach to a classic problem in wireless communications: how to allocate power to different devices to maximize the total data rate while minimizing interference. While deep learning has recently taken over this field, the authors identify a fundamental flaw in the current "gold standard" (Graph Neural Networks) and propose a more efficient, "truncated" transformer-based solution.

TL;DR

Standard Graph Neural Networks (GNNs) struggle with power allocation because they assume neighboring nodes should behave similarly, whereas interference management often requires them to do the opposite. This paper introduces Truncated Graph Transformers (TGT), a lightweight architecture that uses attention mechanisms to outperform classical algorithms and standard GNNs with significantly fewer parameters.

What Problem Does This Paper Address?

In Device-to-Device (D2D) networks, multiple pairs of transmitters and receivers share the same frequency. If everyone transmits at maximum power, they drown each other out in a sea of interference. Finding the "sweet spot" for power levels is a non-convex, NP-hard optimization problem.

While classical iterative algorithms like WMMSE work well, they are slow and computationally heavy for large-scale networks. Recently, researchers turned to Graph Neural Networks (GNNs) because wireless networks naturally look like graphs (devices are nodes, interference links are edges). However, GNNs have a "homophily bias"—they assume connected nodes should have similar features. In an interference channel, the optimal solution is often heterophilous: if two transmitters are close together, one should likely be "loud" while the other stays "quiet" to avoid a collision. Standard GNNs struggle to learn this "opposites attract" behavior.

Methods

The authors propose the Truncated Graph Transformer (TGT). Unlike standard Transformers used in NLP (like GPT), which are massive, TGT is designed for extreme efficiency:

  1. Cross-Attention as Rewiring: Instead of relying on a fixed graph structure, TGT uses a multi-head attention mechanism. This allows the model to "dynamically rewire" the graph, deciding which interference links matter most for each specific node.
  2. Truncation and Weight Sharing: To keep the model lean, the authors share weights across all three layers of the transformer. This "truncation" reduces the parameter count from 173K (in a standard graph transformer) to just 13.1K, without sacrificing performance.
  3. Input Features: The model takes channel gains ($h_{ii}$ and $h_{ij}$) and user weights as inputs and outputs a power level between 0 and 1 via a sigmoid activation function.

Datasets

The researchers evaluated their model using a simulated Single-Input Single-Output (SISO) D2D network:

  • Training Set: 25,000 network instances with 50 pairs, using Rayleigh fading and Additive White Gaussian Noise (AWGN).
  • Test Set: 1,000 previously unseen network instances.
  • Generalization Tests: They tested the model on varying network sizes (20 to 100 pairs) and different fading coefficients to see how well the "trained-on-50" model handled different scenarios.

Key Insights and Ideas

  • The Dinner Party Analogy: The authors compare power control to a busy dinner party. If two people are close to their partners, it’s better for one to speak loudly and the other to stay silent than for both to shout over each other. This is the essence of heterophily.
  • Attention > Convolution: Standard GNNs act like low-pass filters, smoothing out differences between neighbors. TGT’s attention mechanism acts as an active graph re-wirer, allowing it to capture the sharp differences in power levels required for interference management.
  • Complexity vs. Generality: The paper reveals a fascinating trade-off: larger models (more parameters) generalize better to drastic changes in network density, while smaller, "truncated" models are faster but may require more frequent retraining if the network environment changes significantly.

Key Results

  • State-of-the-Art Performance: TGT consistently achieved higher sum-rates than the classical WMMSE algorithm and other leading deep learning models (like PCGNN and UWMMSE).
  • Efficiency: TGT outperformed a standard Graph Transformer while using 13 times fewer parameters (13.1K vs. 173K).
  • Robust Generalization: A model trained on 50 nodes performed excellently on networks ranging from 20 to 100 nodes, showing that the model learns the physics of interference rather than just memorizing a specific setup.

Limitations and Future Work

The primary limitation identified is network density. While TGT handles different numbers of users well, it struggles if the spatial distribution (how tightly packed the users are) deviates significantly from the training data. If the "room size" of the dinner party changes, the model's performance drops.

Future work could involve making the architecture more robust to these density shifts or developing "meta-learning" approaches that allow the model to adapt to new environments with minimal retraining.

Summary updated

Abstract

Power control for the device-to-device interference channel with single-antenna transceivers has been widely analyzed with both model-based methods and learning-based approaches. Although the learning-based approaches, i.e., data-driven and model-driven, offer performance improvement, the widely adopted graph neural network suffers from learning the heterophilous power distribution of the interference channel. In this paper, we propose a deep learning architecture in the family of graph transformers to circumvent the issue. Experiment results show that the proposed methods achieve the state-of-the-art performance across a wide range of untrained network configurations. Furthermore, we show there is a trade-off between model complexity and generality.
BibTeX
@inproceedings{kim2023power,
  author = {Dohoon Kim and Shenghui Song},
  title = {{Power Allocation for Device-to-Device Interference Channel Using Truncated Graph Transformers}},
  booktitle = {MeditCom},
  year = {2023},
  doi = {10.48550/arxiv.2307.10538},
}
Figure