Learning Resilient Radio Resource Management Policies with Graph Neural Networks

Navid Naderializadeh, Mark Eisen, Alejandro Ribeiro

Abstract

As wireless networks evolve toward 6G, managing resources in ultra-dense environments becomes increasingly complex. Traditional Radio Resource Management (RRM) optimization often fails when strict fairness constraints meet volatile channel conditions. This paper introduces a resilient RRM framework that leverages Graph Neural Networks (GNNs) to jointly optimize power control and user selection. By integrating learnable slack variables into a primal-dual learning architecture, the proposed method adaptively relaxes constraints in "unlucky" scenarios, ensuring robust performance and enhanced fairness for cell-edge users.

Introduction

The transition to ultra-dense wireless networks brings the challenge of managing massive numbers of devices competing for limited spectrum. Radio Resource Management (RRM)—deciding which Access Point (AP) serves which User Equipment (UE) and at what power level—is a classic NP-hard problem. While traditional optimization methods use minimum-capacity constraints to ensure fairness, these constraints can become mathematically impossible to satisfy in real-world "dead zones," leading to system-wide performance degradation.

The Resilient Framework

To address the brittleness of strict constraints, the authors propose a resilient formulation. They introduce slack variables into the optimization objective. When a user's minimum rate requirement cannot be met due to extreme interference or poor signal, the slack variable increases to "loosen" the requirement. This creates a tunable tradeoff: the system strives for fairness but gracefully prioritizes aggregate network utility when individual constraints are infeasible.

Graph Neural Networks for RRM

Wireless networks are inherently structured as graphs, where nodes represent APs and UEs, and edges represent channel gains. The authors utilize Graph Neural Networks (GNNs) to parameterize RRM policies. GNNs offer two critical advantages for this domain:

  1. Permutation Equivariance: The policy's output remains consistent regardless of how users are indexed.
  2. Size Invariance: A model trained on a small-scale network can be deployed on much larger networks without retraining.
Graph-based representation of a wireless network with two Access Points and five User Equipments.

As shown in Figure 1, the network topology is mapped to a graph where interference patterns and associations are explicitly modeled, allowing the GNN to learn spatial dependencies.

Learning via Primal-Dual Optimization

The RRM problem is solved in the Lagrangian dual domain using an unsupervised primal-dual learning approach. This method is particularly powerful because it does not require "ground truth" optimal solutions, which are computationally prohibitive to obtain for large networks. Instead, the model learns by directly optimizing the Lagrangian function through stochastic gradient updates.

Algorithm 1: Primal-Dual Learning of Resilient RRM Policies.

Algorithm 1 outlines the iterative process where primal RRM policy parameters, ergodic average rates, and dual variables are updated concurrently to find a resilient equilibrium.

Key Insights and Results

The proposed GNN-based resilient framework yields several significant findings:

  • Fairness for Cell-Edge Users: The method significantly improves the 5th percentile rate, a key metric for users with the poorest service. It outperforms traditional baselines like WMMSE and ITLinQ by providing a more equitable resource distribution.
  • Slack as a Topology Sensor: The learned slack variables naturally identify difficult network regions. Users in low-SINR areas exhibit higher slack values, indicating that the GNN autonomously detects "dead zones" and adapts the local constraints accordingly.
  • Generalization and Scalability: Models trained on configurations with 4 APs and 40 UEs successfully generalized to networks with 10 APs and 100 UEs, maintaining performance gains and demonstrating the robustness of the GNN architecture.

Conclusion

By combining the flexibility of slack variables with the structural advantages of GNNs, this resilient RRM framework offers a promising path for 6G network management. Future work aims to relax the assumption of perfect Channel State Information (CSI) and move toward decentralized implementations where APs make decisions based on local feedback.

Summary updated

BibTeX
@article{naderializadeh2023learning,
  author = {Navid Naderializadeh and Mark Eisen and Alejandro Ribeiro},
  title = {{Learning Resilient Radio Resource Management Policies with Graph Neural Networks}},
  journal = {IEEE Trans. Signal Process.},
  year = {2023},
  doi = {10.1109/tsp.2023.3255547},
}
Figure