Deep Graph Unfolding for Beamforming in MU-MIMO Interference Networks

Arindam Chowdhury, Gunjan Verma, Ananthram Swami, Santiago Segarra

TL;DR

Researchers have developed UWMMSE, a hybrid "unfolded" deep learning model that solves the complex problem of beamforming in multi-user MIMO wireless networks. By turning the classical WMMSE optimization algorithm into a neural network and enhancing it with Graph Neural Networks (GNNs), the authors achieved higher data rates than traditional methods while running 24 times faster.

What Problem Does This Paper Address?

In modern wireless ad-hoc networks (WANETs), multiple devices with multiple antennas (MU-MIMO) try to communicate simultaneously. The "holy grail" is to find the perfect beamforming strategy—shaping the radio signals so that each receiver gets its intended data while minimizing interference from everyone else.

Mathematically, maximizing the total network "sum-rate" is an NP-hard, non-convex optimization problem. The industry standard is an algorithm called WMMSE. While effective, WMMSE is computationally "heavy" (requiring many matrix inversions) and must be re-run from scratch every time the channel conditions change slightly. Existing "pure" deep learning models are faster but often fail to generalize to new network sizes or complex interference patterns.

Methods

The authors use a technique called Algorithm Unfolding. They take the iterative update rules of the classical WMMSE algorithm and "unfold" them into layers of a deep neural network.

Key Innovations: * UWMMSE (Unfolded WMMSE): Instead of just mimicking the old algorithm, they replace a key internal parameter (the receiver weight matrix) with a learnable functional transformation. * Graph Neural Networks (GNNs): Because wireless networks are essentially graphs (nodes are transceivers, edges are interference), the authors use a Complex-Valued GNN (CV-GNN) to learn the network's topology. This allows the model to understand how interference flows between specific pairs of nodes. * Complex-Valued MLP: They use a multi-layer perceptron that handles complex numbers (magnitude and phase) to refine the beamforming weights. * Parameter Sharing: The GNN parameters are shared across all "unfolded" layers. This makes the model incredibly lightweight (only ~3,300 parameters) and allows it to be trained with one number of layers but deployed with another.

Datasets

The study relies on randomly generated geometric channel realizations to simulate realistic wireless environments:

  • Rayleigh Fading: Simulates environments with many obstacles and no clear line-of-sight.
  • Rician Fading: Simulates environments with a dominant line-of-sight signal (using a 20 dB K-factor).
  • Geometric Graphs: Transceivers are dropped randomly in a 2D space, and path loss is calculated based on their physical distance.

Key Insights and Ideas

  • The "Best of Both Worlds": By keeping the structure of the WMMSE algorithm but using GNNs to optimize the steps, the model retains the mathematical logic of wireless physics while gaining the speed of deep learning.
  • Generalization is King: Unlike standard black-box neural networks, UWMMSE generalizes remarkably well. A model trained on a network of 20 users can be deployed on a network of 100 users without retraining.
  • Phase Matters: Many previous attempts at this problem ignored the "phase" of complex signals. This paper uses Cartesian non-linearities to process both real and imaginary components, which is vital for beamforming.
  • Any-time Inference: Because of parameter sharing, you can add more layers at inference time to squeeze out more performance without needing to retrain the underlying GNN.

Key Results

  • Performance: UWMMSE consistently achieved higher sum-rates than the full classical WMMSE algorithm.
  • Speed: In a Rayleigh fading scenario, UWMMSE took 54ms per sample, compared to 1,305ms for classical WMMSE—a 24x speedup.
  • Efficiency: It outperformed other state-of-the-art learning models like IAIDNN and GCN-WMMSE, particularly in "low-noise" regimes where interference management is most difficult.
  • Robustness: The model maintained superior performance even when the input Channel State Information (CSI) was "noisy" or distorted (up to 20% coefficient distortion).

Limitations and Future Work

  • Centralized Training: While the model can be deployed in a distributed way, it currently requires centralized training where a single entity has access to the entire network's channel data.
  • Static Channels: The current work focuses on a "snapshot" of the network. The authors suggest that future work should address time-varying channels where the movement of devices creates a temporal dimension to the interference.
  • QoS Constraints: The study focuses on sum-rate maximization; future iterations could incorporate more complex Quality-of-Service (QoS) requirements like latency or fairness between users.

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 variables 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},
}

Loading related content...

Figure