[Paper Review] Understanding the Performance of Learning Precoding Policy with GNN and CNNs

Bai‐Chuan Zhao, Jia Guo+1 more|arXiv (Cornell University)|Nov 27, 2022

Advanced MIMO Systems Optimization4 citations

TL;DR

This paper investigates why deep neural networks (DNNs) generalize well in learning precoding policies by analyzing the interplay between inductive bias and policy smoothness. It proposes an edge-update graph neural network (EdgeGNN) with inductive bias matching the permutation-equivariant structure of precoding policies, demonstrating superior sample efficiency and generalization over CNNs and fully connected networks, especially under high SNR, spatial correlation, or high user-to-antenna ratios.

ABSTRACT

Learning-based precoding has been shown able to be implemented in real-time, jointly optimized with channel acquisition, and robust to imperfect channels. Yet previous works rarely explain the design choices and learning performance, and existing methods either suffer from high training complexity or depend on problem-specific models. In this paper, we address these issues by analyzing the properties of precoding policy and inductive biases of neural networks, noticing that the learning performance can be decomposed into approximation and estimation errors where the former is related to the smoothness of the policy and both depend on the inductive biases of neural networks. To this end, we introduce a graph neural network (GNN) to learn precoding policy and analyze its connection with the commonly used convolutional neural networks (CNNs). By taking a sum rate maximization precoding policy as an example, we explain why the learned precoding policy performs well in the low signal-to-noise ratio regime, in spatially uncorrelated channels, and when the number of users is much fewer than the number of antennas, as well as why GNN is with higher learning efficiency than CNNs. Extensive simulations validate our analyses and evaluate the generalization ability of the GNN.

Motivation & Objective

• To understand when and why deep learning-based precoding policies generalize well in MIMO systems.
• To analyze the impact of inductive bias in DNNs—specifically GNNs and CNNs—on learning performance for precoding.
• To identify system conditions where precoding policies are harder to approximate due to low smoothness.
• To derive estimation error bounds for DNNs and validate them via simulations.
• To demonstrate that policy-matched inductive bias (e.g., in EdgeGNN) reduces sample complexity and improves generalization.

Proposed method

• Designs a graph neural network (EdgeGNN) with edge-update mechanism to match the permutation-equivariance property of precoding policies.
• Compares EdgeGNN with two CNN variants (F-LCNN and CCNN) and fully connected networks (FNN) in learning baseband precoding for sum rate maximization.
• Analyzes the decomposition of learning performance into approximation error (dependent on policy smoothness) and estimation error (dependent on DNN inductive bias).
• Derives theoretical estimation error bounds for EdgeGNN, CNN, and FNN based on their inductive biases and hypothesis space constraints.
• Uses synthetic channel data with varying SNR, spatial correlation (ρ), and user-antenna ratios (K/N) to evaluate generalization across distributions.
• Trains models on single-distribution and mixed-distribution data to assess out-of-distribution generalization performance.

Experimental results

Research questions

• RQ1 In which system settings is the precoding policy harder to approximate by a DNN due to low smoothness?
• RQ2 How does the inductive bias of a DNN (e.g., weight sharing in CNNs or permutation equivariance in GNNs) affect estimation error and sample complexity?
• RQ3 Why does EdgeGNN outperform CNNs and FNNs in terms of generalization and sample efficiency?
• RQ4 How does training on mixed-distribution data affect generalization across different channel statistics?
• RQ5 To what extent does the inductive bias of a DNN align with the structural properties of the optimal precoding policy?

Key findings

• The EdgeGNN achieves 98.2% average test accuracy across diverse channel statistics (ρ,γ) when trained on mixed data, outperforming FNN (89.4%) and F-LCNN (89.7%) under the same conditions.
• EdgeGNN trained on uncorrelated Rayleigh channels (ρ=0,γ=0) generalizes best to other channel types, achieving 97.6% average performance, due to high coverage of the channel distribution.
• At high SNR (10 dB), models trained on low-SNR data (0 dB) show significant performance drop (e.g., FNN drops from 97.8% to 75.3% on mixed test set), indicating poor generalization.
• EdgeGNN and CCNN show similar generalization performance, both outperforming FNN and F-LCNN, due to smaller hypothesis spaces and better inductive bias matching.
• The EdgeGNN trained on (ρ,γ)=(0.8,0) performs worst (89.4% average) due to high channel correlation, confirming that spatial correlation increases learning difficulty.
• Estimation error bounds derived for EdgeGNN, CNN, and FNN are validated by simulations, showing that inductive bias significantly reduces estimation error and sample complexity.