Model-Free Learning of Optimal Beamformers for Passive IRS-Assisted Sumrate Maximization

Hassaan Hashmi, Spyridon Pougkakiotis, Dionysios S. KalogeriasICASSP 2022 · DOI · OpenReview

Abstract

Intelligent Reflective Surfaces (IRS) are a cornerstone technology for 6G networks, yet their control remains challenging due to hardware passivity and the complexity of real-world channel modeling. This paper introduces ZoSGA (Zeroth-order Stochastic Gradient Ascent), a model-free algorithm that optimizes IRS phase shifts through system probing rather than explicit mathematical modeling. Simulations demonstrate that ZoSGA not only achieves state-of-the-art sumrate performance but also outperforms traditional model-based methods in realistic, noisy environments while maintaining rigorous theoretical convergence guarantees.

Introduction

The deployment of Intelligent Reflective Surfaces (IRS) aims to eliminate wireless dead zones by acting as programmable "smart mirrors." However, two primary hurdles exist: the passive hardware of IRS units, which cannot sense the channel, and the extreme complexity of modeling dynamic environments in real-time. Existing deep learning or model-based optimization approaches often lack robustness or fail when physical realities deviate from mathematical assumptions. This work proposes a "middle way"—a data-driven, model-free approach that remains mathematically grounded through zeroth-order optimization.

Methods

The proposed ZoSGA framework utilizes a two-timescale optimization strategy to maximize the network sumrate. An inner loop quickly optimizes Access Point (AP) precoders using the Weighted Minimum Mean Square Error (WMMSE) algorithm to handle fast-fading channel components. The outer loop optimizes IRS phase shifts using zeroth-order optimization.

System model for the IRS-assisted MISO downlink network showing AP, IRS units, and user geometry.

Instead of requiring a full channel model to calculate gradients, ZoSGA "probes" the system by applying small random perturbations to IRS parameters in two directions. By measuring the resulting change in the "effective" channel, the algorithm estimates the gradient and iteratively improves the sumrate based solely on observed end-to-end performance. This approach requires no extra hardware or specialized sensing on the IRS.

Results and Discussion

Evaluation in simulated Rician fading channels shows that ZoSGA consistently outperforms state-of-the-art model-based baselines like TTS-SSCO. This superiority stems from ZoSGA's ability to optimize the actual system response directly, whereas model-based methods often rely on simplified "surrogate" models that may not capture environmental nuances.

Sumrate performance vs. Rician factor β, demonstrating ZoSGA's performance advantage over model-based TTS-SSCO.

Key findings from the experimental results include:

  • Robustness to Geometry: The algorithm is independent of specific spatial coordinates or channel physics, making it highly adaptable to various deployment scenarios.
  • Sample Efficiency: While model-based methods might converge in fewer iterations, they require significantly more internal channel samples per iteration. ZoSGA is more efficient in its interaction with the environment.
  • Scalability: The framework successfully scales to multi-IRS configurations, providing significant gains when optimizing multiple surfaces simultaneously compared to single-IRS or random-phase setups.
  • Theoretical Soundness: Unlike many deep learning methods, ZoSGA has a proven convergence rate of $O(\sqrt{S}\epsilon^{-4})$, where $S$ is the number of IRS parameters.

Conclusion

ZoSGA offers a practical, sample-efficient solution for IRS control in 6G networks. By bypassing the need for complex channel models, it provides a critical advantage in unpredictable real-world scenarios. Future research will explore extending this framework to more complex Quality-of-Service (QoS) metrics and further reducing the number of probes required for gradient estimation.

Summary updated

Abstract

Although Intelligent Reflective Surfaces (IRSs) are a cost-effective technology promising high spectral efficiency in future wireless networks, obtaining optimal IRS beamformers is a challenging problem with several practical limitations. Assuming fully-passive, sensing- free IRS operation, we introduce a new data-driven Zeroth-order Stochastic Gradient Ascent (ZoSGA) algorithm for sumrate optimization in an IRS-aided downlink setting. ZoSGA does not require access to channel model or network structure information, and enables learning of optimal long-term IRS beamformers jointly with standard short-term precoding, based only on conventional effective channel state information. Supported by state-of-the-art (SOTA) convergence analysis, detailed simulations confirm that ZoSGA exhibits SOTA empirical behavior as well, consistently outperforming standard fully model-based baselines, in a variety of scenarios.
BibTeX
@inproceedings{hashmi2022model,
  author = {Hassaan Hashmi and Spyridon Pougkakiotis and Dionysios S. Kalogerias},
  title = {{Model-Free Learning of Optimal Beamformers for Passive IRS-Assisted Sumrate Maximization}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2022},
  doi = {10.1109/icassp49357.2023.10095269},
}
Figure