Model-Free Learning of Optimal Two-Stage Beamformers for Passive IRS-Aided Network Design

Hassaan Hashmi, Spyridon Pougkakiotis, Dionysios S. Kalogerias

Introduction

As wireless communications transition toward 6G, Intelligent Reflecting Surfaces (IRS) have emerged as a transformative technology for managing high-frequency bands like mmWave and THz. By utilizing metasurfaces to "steer" signals, IRS can effectively bypass physical blockages and enhance coverage. However, the practical deployment of IRS is hindered by significant optimization challenges. Traditional methods rely heavily on precise mathematical models and high-resolution Channel State Information (CSI), which are difficult to obtain for passive surfaces that lack active radio frequency chains. This paper introduces ZoSGA (Zeroth-order Stochastic Gradient Ascent), a model-free framework designed to optimize IRS-aided networks without the need for complex channel estimation or idealized hardware assumptions.

The Bottlenecks of Traditional IRS Optimization

Existing optimization strategies for IRS-aided networks typically encounter three primary obstacles:

  1. Modeling Uncertainty: Most algorithms require an exact mathematical representation of the wireless environment. In real-world deployments, multi-path fading and environmental dynamics make such models difficult to maintain.
  2. Channel Sensing Overhead: To optimize a passive IRS, the system must ideally know the CSI for both the Access Point (AP)-to-IRS and IRS-to-user links. For passive surfaces, acquiring this information is resource-intensive and often consumes the very bandwidth the IRS is intended to save.
  3. Hardware Non-Linearities: Theoretical models often assume "ideal" phase shifters. In reality, hardware components like varactor diodes exhibit non-linear relationships between control parameters (e.g., capacitance) and the resulting signal phase and amplitude.

The ZoSGA Framework

To address these issues, the authors propose a model-free approach that treats the wireless environment as a "black box." The core of this approach is the ZoSGA algorithm, which optimizes network performance based solely on terminal-end signal observations.

Two-Stage Optimization

The problem is structured across two distinct timescales to balance efficiency and performance:

  • Long-Term Stage: This stage optimizes the IRS reflection parameters. Since the geometric orientation of the IRS relative to the AP and users changes slowly, these parameters are updated at a lower frequency.
  • Short-Term Stage: This stage optimizes the AP precoders to account for fast-fading effects. The authors utilize the standard Weighted Minimum Mean Square Error (WMMSE) algorithm for this task, ensuring the AP can react instantaneously to channel fluctuations.

Zeroth-Order Gradient Estimation

Instead of calculating gradients through backpropagation or complex modeling, ZoSGA employs zeroth-order optimization. The algorithm slightly perturbs the IRS settings—probing the network twice—and measures the resulting change in the utility function (e.g., sumrate). This allows the system to estimate the gradient of the performance metric directly from observed data.

Wirtinger Calculus and Hardware Agnosticism

Wireless channels involve complex-valued variables, which can complicate standard optimization. The authors leverage Wirtinger calculus to derive a compositional gradient that operates in a Euclidean setting. This allows ZoSGA to optimize physical hardware parameters, such as the capacitance of varactor diodes, directly. By doing so, the algorithm bypasses the difficult "unit modulus" constraints and non-linear mapping issues that plague model-based methods.

Experimental Evaluation and Results

The researchers evaluated ZoSGA using simulated Multi-Input Single-Output (MISO) downlink environments with spatially correlated Rician fading. They also tested the algorithm against a realistic physical electromagnetic (EM) model of a varactor-based IRS.

Key Findings

  • Superior Performance: ZoSGA consistently outperformed the model-based state-of-the-art baseline, TTS-SSCO (Two-Time Scale Stochastic Successive Convex Optimization). Remarkably, ZoSGA achieved higher sumrates even though the baseline had access to "perfect" channel knowledge that ZoSGA did not require.
  • Scalability: The algorithm demonstrated robust performance as the number of IRS elements increased, showing significant gains even with surfaces containing up to 400 elements.
  • Hardware Robustness: In scenarios where phase and amplitude were coupled (a common real-world hardware constraint), ZoSGA maintained its performance while model-based methods—relying on idealized assumptions—suffered significant degradation.

Discussion and Future Directions

ZoSGA represents a shift toward "blind" optimization in wireless networks, proving that effective CSI (the compound channel observed at the receiver) is sufficient for high-performance beamforming. However, several areas for future research remain:

  • Discrete Optimization: While ZoSGA assumes continuous parameter tuning, many practical IRS units use discrete (e.g., 1-bit or 2-bit) phase shifters. Adapting zeroth-order methods to discrete search spaces is a critical next step.
  • Probing Efficiency: Although the two-probe requirement is minimal, reducing the overhead of gradient estimation would further improve performance in highly dynamic, fast-fading environments.
  • Convergence Speed: While ZoSGA reaches a higher performance ceiling than model-based approaches, it typically requires more iterations to converge. Improving sample efficiency remains a priority for real-time implementation.

Conclusion

By moving away from rigid mathematical models and toward a data-driven, zeroth-order approach, ZoSGA provides a robust solution for the deployment of passive IRS in 6G networks. Its ability to handle hardware non-linearities and eliminate the need for explicit channel sensing makes it a highly practical candidate for future wireless infrastructure.

Summary updated

Abstract

Electronically tunable metasurfaces, or Intelligent Reflective Surfaces (IRSs), are a popular technology for achieving high spectral efficiency in modern wireless systems by shaping channels using a multitude of tunable passive reflective elements. Capitalizing on key practical limitations of IRS-aided beamforming pertaining to system modeling and channel sensing/estimation, we propose a novel, fully data-driven Zeroth-order Stochastic Gradient Ascent (ZoSGA) algorithm for general two-stage (i.e., short/long-term), fully-passive IRS-aided stochastic utility maximization. ZoSGA learns long-term optimal IRS beamformers jointly with short-term optimal precoders (e.g., WMMSE-based) via minimal zeroth-order reinforcement and in a strictly model-free fashion, relying solely on the \textit{effective} compound channels observed at the terminals, while being independent of channel models or network/IRS configurations. Another remarkable feature of ZoSGA is being amenable to analysis, enabling us to establish a state-of-the-art (SOTA) convergence rate of the order of $\boldsymbol{O}(\sqrt{S}\epsilon^{-4})$ under minimal assumptions, where $S$ is the total number of IRS elements, and $\epsilon$ is a desired suboptimality target. Our numerical results on a standard MISO downlink IRS-aided sumrate maximization setting establish SOTA empirical behavior of ZoSGA as well, consistently and substantially outperforming standard fully model-based baselines. Lastly, we demonstrate that ZoSGA can in fact operate \textit{in the field}, by directly optimizing the capacitances of a varactor-based electromagnetic IRS model (unknown to ZoSGA) on a multiple user/IRS, compute-heavy network setting, with essentially no computational overheads or performance degradation.
BibTeX
@article{hashmi2024model,
  author = {Hassaan Hashmi and Spyridon Pougkakiotis and Dionysios S. Kalogerias},
  title = {{Model-Free Learning of Optimal Two-Stage Beamformers for Passive IRS-Aided Network Design}},
  journal = {IEEE Trans. Signal Process.},
  year = {2024},
  doi = {10.1109/tsp.2023.3346182},
}
Figure