Abstract
Reconfigurable Intelligent Surfaces (RIS) are a cornerstone technology for 5G and 6G, enabling programmable wireless environments through passive element arrays. However, optimizing these surfaces in real-time is hindered by the non-convex unit modulus constraint and the extreme computational complexity of traditional optimization methods. This paper introduces RISBFNN, an unsupervised deep learning framework that optimizes RIS configurations without labeled data. By leveraging structural channel priors and a customized loss function, RISBFNN achieves near-optimal beamforming performance while reducing computational latency by several orders of magnitude, making real-time deployment feasible.
The Challenge of Real-Time RIS Optimization
RIS technology consists of arrays of passive elements that "steer" radio waves by adjusting their phase shifts. While this improves signal strength and energy efficiency, finding the optimal configuration—known as passive beamforming—is mathematically difficult. Each reflecting element is subject to a unit modulus constraint, meaning it can only change a signal's phase, not its amplitude.
Standard optimization methods like Semi-Definite Relaxation (SDR) suffer from high computational complexity ($O(N^{6.5})$). In dynamic environments where the wireless channel changes every few milliseconds (the coherence time), SDR is too slow to be practical. While supervised deep learning offers faster inference, it requires expensive "optimal labels" for training, which are themselves difficult to calculate for large-scale surfaces.
System Model and Proposed Architecture
The authors propose a system where an Access Point (AP) communicates with a user through an RIS controlled by a dedicated unit. The goal is to maximize the received signal power by adjusting the RIS phase shifts based on the channel states.
To address the limitations of traditional methods, the authors developed RISBFNN, a five-layer fully-connected neural network designed for unsupervised learning. The framework incorporates three critical innovations:
- Feature Engineering: Instead of raw channel data, the network uses the product of channel elements (feature fusion). This exploits the known physics of signal reflection to reduce input dimensions and improve training efficiency.
- Unsupervised Loss Function: The network is trained to minimize the negative of the receive SNR. By optimizing for the end goal directly, the model learns to maximize signal strength without needing pre-calculated optimal solutions.
- Lambda Layer: A specialized layer at the end of the network converts real-valued predictions into complex phase shifts ($e^{j\theta}$), strictly enforcing the unit modulus constraint.
Key Results and Performance
The effectiveness of RISBFNN was evaluated using synthetic Rayleigh fading channel data. The results demonstrate that the unsupervised approach is not only accurate but also remarkably efficient.
Accuracy and Scalability
In single-antenna setups, RISBFNN achieved near-optimal performance. In multi-antenna configurations, it maintained 88% to 96% of the performance of the SDR-based approach. Crucially, the computational complexity was reduced from $O(N^{6.5})$ to $O(N^2)$, allowing the system to scale to much larger surfaces than previously possible.
Real-Time Feasibility
The most significant breakthrough is in processing speed. In an indoor environment with a channel coherence time of approximately 13.77 ms, SDR takes 200–700 ms to find a solution, making it obsolete before it finishes. In contrast, RISBFNN completes the same task in 0.04–0.11 ms, fitting comfortably within the stability window of the wireless channel.
Conclusion
RISBFNN demonstrates that unsupervised learning can bridge the gap between theoretical RIS potential and practical deployment. By bypassing the need for expensive labels and drastically reducing inference time, this framework enables real-time passive beamforming. Future research will focus on extending this model to multi-user scenarios and adapting it for hardware that supports only discrete phase shifts.
Summary updated