Deep Graph Unfolding for Beamforming in MU-MIMO Interference Networks

Arindam Chowdhury, Gunjan Verma, Ananthram Swami, Santiago Segarra

Abstract

We develop an efficient and near-optimal solution for beamforming in multi-user multiple-input-multiple-output (MU-MIMO) single-hop wireless ad-hoc interference networks. Inspired by the weighted minimum mean squared error (WMMSE) method and the principle of algorithm unfolding, we present Unfolded WMMSE (UWMMSE). This method learns a parameterized functional transformation of key WMMSE parameters using Graph Neural Networks (GNNs), where the channel and interference components constitute the underlying graph. Comprehensive experimental analyses illustrate the superiority of UWMMSE over classical WMMSE and state-of-the-art learning-based methods in terms of performance, generalizability, and robustness.

Introduction

In modern wireless ad-hoc networks (WANETs), achieving optimal beamforming—shaping radio signals to maximize data rates for intended receivers while minimizing interference—is a critical challenge. Mathematically, maximizing the network sum-rate is an NP-hard, non-convex optimization problem. The industry-standard Weighted Minimum Mean Squared Error (WMMSE) algorithm is effective but computationally expensive, requiring numerous matrix inversions and restarts for every change in channel conditions. While pure deep learning models offer speed, they often struggle to generalize across different network sizes or complex interference topologies.

Methodology: Unfolded WMMSE (UWMMSE)

UWMMSE bridges the gap between classical optimization and deep learning through Algorithm Unfolding. The iterative update rules of the WMMSE algorithm are mapped onto the layers of a deep neural network.

Architectural Innovations

Unlike standard black-box models, UWMMSE retains the mathematical structure of the WMMSE updates but replaces fixed internal parameters with learnable functional transformations. A Complex-Valued Graph Neural Network (CV-GNN) is employed to learn the network's topology, treating transceivers as nodes and interference paths as edges. This allows the model to capture the physical relationships between nodes effectively.

Architecture of the Unfolded WMMSE (UWMMSE) layer, showing the integration of GNN-based parameter learning within the iterative update framework.

Key features of the architecture include:

  • Complex-Valued Processing: The model utilizes a multi-layer perceptron (MLP) designed for complex numbers, processing both magnitude and phase via Cartesian non-linearities.
  • Parameter Sharing: GNN parameters are shared across all unfolded layers, resulting in a lightweight model (~3,300 parameters) that supports "any-time" inference—allowing more layers to be added at runtime for increased performance without retraining.

Experimental Setup

The model was evaluated using simulated wireless environments, including Rayleigh fading (multi-path interference), Rician fading (line-of-sight dominant), and geometric graphs where path loss is determined by physical distance in a 2D space.

Key Results

  • Superior Performance and Speed: UWMMSE consistently achieved higher sum-rates than the classical WMMSE algorithm. Crucially, it achieved a 24x speedup, taking only 54ms per sample compared to 1,305ms for WMMSE in Rayleigh fading scenarios.
  • Generalization: The model demonstrates remarkable scalability; a version trained on a 20-user network can be deployed on a 100-user network without performance degradation.
  • Robustness: UWMMSE maintains high performance even with noisy Channel State Information (CSI), showing resilience to up to 20% coefficient distortion.

Conclusion

UWMMSE represents a "best of both worlds" approach, combining the physical grounding of optimization algorithms with the computational efficiency of deep learning. While current training is centralized, the model's deployment is inherently distributed. Future research will focus on extending this framework to time-varying channels and incorporating diverse Quality-of-Service (QoS) constraints such as latency and fairness.

Summary updated

Abstract

We develop an efficient and near-optimal solution for beamforming in multi-user multiple-input-multiple-output single-hop wireless ad-hoc interference networks. Inspired by the weighted minimum mean squared error (WMMSE) method, a classical approach to solving this problem, and the principle of algorithm unfolding, we present unfolded WMMSE (UWMMSE) for MU-MIMO. This method learns a parameterized functional transformation of key WMMSE parameters using graph neural networks (GNNs), where the channel and interference components of a wireless network constitute the underlying graph. These GNNs are trained through gradient descent on a network utility metric using multiple instances of the beamforming problem. Comprehensive experimental analyses illustrate the superiority of UWMMSE over the classical WMMSE and state-of-the-art learning-based methods in terms of performance, generalizability, and robustness.
BibTeX
@article{chowdhury2024deep,
  author = {Arindam Chowdhury and Gunjan Verma and Ananthram Swami and Santiago Segarra},
  title = {{Deep Graph Unfolding for Beamforming in MU-MIMO Interference Networks}},
  journal = {IEEE Trans. Wireless Communications},
  year = {2024},
  url = {https://openreview.net/forum?id=Q4e5LhidTT},
}
Figure