Ultra-Reliable Risk-Aggregated Sum Rate Maximization via Model-Aided Deep Learning

Hassaan Hashmi, Spyridon Pougkakiotis, Dionysis KalogeriasarXiv 2025 · DOI · arXiv

Introduction

In modern wireless systems, the standard objective is to maximize the Weighted Sum Rate (WSR). Traditionally, this is approached ergodically, optimizing for the best performance on average. However, this "average-optimal" strategy is often channel-opportunistic. When a user experiences a "deep fade"—a temporary but severe drop in signal quality—opportunistic algorithms may stop serving that user entirely to boost total network throughput. This results in abrupt service drops and poor reliability, particularly for users at the cell edge.

This paper introduces a novel framework for "ultra-reliable" wireless communication by shifting the focus from average throughput to a risk-aggregated sum rate. By utilizing Conditional Value-at-Risk (CVaR) and a specialized "unfolded" Graph Neural Network ($\alpha$RGNN), the authors demonstrate a method to eliminate zero-rate events, ensuring a consistent user experience across the network.

Methodology: Risk-Aware Optimization

The authors propose replacing standard expectation-based metrics with Conditional Value-at-Risk (CVaR). Unlike standard metrics that look at the mean, CVaR focuses on the average of the "worst-case" outcomes—specifically the lower tail of the rate distribution.

A depiction of the CVaR metric, which focuses on the lower tail of the probability distribution to ensure reliability.

Theoretical Equivalence and the $\alpha$RGNN

A key contribution of this work is the mathematical proof that this risk-averse problem can be transformed into a weighted risk-averse Mean Squared Error (MSE) problem. This mirrors the structure of the well-known WMMSE algorithm.

Leveraging this structure, the authors designed the $\alpha$RGNN (Alpha-Robust Graph Neural Network). This is a "model-aided" deep learning architecture that "unfolds" the iterative steps of an optimization algorithm into layers of a Graph Neural Network. The architecture consists of three cascading GNNs (uGNN, wGNN, and vGNN) that update beamforming parameters. Because it is a GNN, it naturally scales with the interference topology of the wireless network and is more parameter-efficient than standard "black-box" neural networks.

Experimental Setup

The framework was evaluated using synthetic datasets simulating a MISO (Multiple-Input Single-Output) downlink cell:

  • Environment: 10 users and a base station with 6 antennas.
  • Channel Model: Rician fading ($\beta = -3$ dB) with standard power-law pathloss ($d^{-2.2}$).
  • Data: 200,000 channel realizations for training and 50,000 for testing.

Key Results and Insights

The parameter $\alpha$ serves as a tunable "knob" for reliability. Setting $\alpha=1$ recovers the standard risk-neutral case, while lower values (e.g., $\alpha=0.7$) make the system increasingly robust against deep fades through "risk sharing."

Elimination of Outages

The most significant result is the near-total elimination of "zero-rate" events. In standard WMMSE and risk-neutral models, distant users (Users 8, 9, and 10) frequently lose service entirely during fades. As shown in the rate density plots, the risk-aware $\alpha$RGNN shifts the probability mass away from zero.

User rate densities across 10 users. Note how the risk-aware $\alpha$RGNN ($\alpha=0.7$) prevents the 'zero-rate' spikes seen in WMMSE for distant users.

The Price of Reliability

There is an inherent trade-off between total throughput and individual reliability. While WMMSE achieved a higher average sum rate (1.11 bits/s/Hz), the risk-aware $\alpha$RGNN (0.74 bits/s/Hz) provided a much more equitable and stable service. This stability is reflected in the Sharpe Ratio (mean rate divided by standard deviation), which was significantly higher for the $\alpha$RGNN across most users.

Conclusion and Future Work

By incorporating domain knowledge into a GNN architecture, the authors provide a scalable solution for reliable beamforming. Future research may extend this framework to more complex topologies, such as cell-free massive MIMO, or apply risk-aggregation to other Quality of Service (QoS) metrics like latency and jitter.

Summary updated

BibTeX
@misc{hashmi2025ultra,
  author = {Hassaan Hashmi and Spyridon Pougkakiotis and Dionysis Kalogerias},
  title = {{Ultra-Reliable Risk-Aggregated Sum Rate Maximization via Model-Aided Deep Learning}},
  year = {2025},
  eprint = {2509.26311},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.26311},
}
Figure