Cooperative Beamforming and RISs Association for Multi-RISs Aided Multi-Users MmWave MIMO Systems through Graph Neural Networks

Mengbing Liu, Chongwen Huang, Marco Di Renzo, Mérouane Debbah, Chau Yuen

TL;DR

This paper introduces a heterogeneous Graph Neural Network (GNN) to solve the complex challenge of managing multiple Reconfigurable Intelligent Surfaces (RISs) in a mmWave MIMO system. By modeling the network as a graph, the authors jointly optimize active beamforming at the base station, passive beamforming at the RISs, and—crucially—the dynamic association between users and RISs, achieving up to 10 times better performance than standard deep learning models.

What Problem Does This Paper Address?

Millimeter-wave (mmWave) communication is a cornerstone of 5G and 6G, but it has a major weakness: signals are easily blocked by buildings or even human bodies. Reconfigurable Intelligent Surfaces (RISs)—essentially "smart mirrors" for radio waves—can reflect signals around these obstacles.

However, when you move from a single RIS to a network of multiple RISs and multiple users, the math becomes a nightmare. You have to decide:

  1. Active Beamforming: How the Base Station (BS) shapes its signal.
  2. Passive Beamforming: How each RIS element shifts the phase of the reflected signal.
  3. RIS Association: Which user should "talk" to which RIS?

This is a non-convex, high-dimensional optimization problem. Traditional algorithms are often too slow for real-time use, and standard Deep Neural Networks (DNNs) struggle because they don't understand the underlying "topology" (the spatial relationships) of the wireless network.

Methods

The authors propose a Heterogeneous GNN architecture. Unlike a standard GNN where all nodes are the same, this model treats the Base Station, RISs, and Users as different types of nodes with specific attributes.

Key Innovations:

  • Graph Modeling: The system is represented as a graph where edges represent communication links and interference. This allows the model to "understand" how a change at one RIS affects a user across the room.
  • Message Passing Paradigm: The model uses iterative updates. RIS nodes "listen" to user requirements to update their features, and User nodes "listen" to RIS configurations to update theirs.
  • Normalization Layers: To ensure the model respects physical hardware limits, the authors added custom layers that enforce transmit power constraints at the BS and the "unit modulus" constraint (the phase-only reflection) at the RIS.
  • Hybrid Loss Function: Since choosing an RIS is a "hard" discrete choice (User A connects to RIS 1 or 2), the authors used a combination of the Weighted Sum Rate (WSR) and a categorical cross-entropy term to train the network to make smart association decisions.

Datasets

The study relies on synthetic data generated using the Saleh-Valenzuela channel model, which is standard for characterizing the sparse scattering nature of mmWave environments.

  • Scale: 110,000 total samples (100,000 for training, 10,000 for validation).
  • Setup: A system with 8 BS antennas, 2 RISs (each with 16 elements), and 2 users.
  • Benchmarks: The model was compared against standard DNNs and the WMMSE (Weighted Minimum Mean Square Error) algorithm with random phase configurations.

Key Insights and Ideas

  • Topology is King: Standard DNNs treat network data as a flat vector, losing spatial context. The GNN’s ability to exploit the graph structure is why it outperforms DNNs so dramatically as the network scales.
  • Dynamic Association Matters: Most existing research assumes a user is fixed to a specific RIS. This paper shows that dynamically selecting the best RIS based on the current environment provides a massive boost in service quality.
  • Generalization: Because GNNs learn how nodes interact rather than just memorizing a fixed input size, they are much better at handling changes in the number of users or RISs than traditional deep learning models.

Key Results

  • 10x Performance Boost: The proposed heterogeneous GNN performed roughly 10 times better than standard DNNs in terms of maximizing the Weighted Sum Rate.
  • 30% Efficiency Gain: Implementing the dynamic RIS association strategy improved the quality of service for users by 30% compared to fixed-association setups.
  • Robustness: The GNN maintained high performance across various transmit power levels (from 10 dBm to 30 dBm) and different numbers of RIS elements (9, 16, and 25).

Limitations and Future Work

While the results are promising, there are a few areas for further exploration:

  • Perfect CSI Assumption: The model assumes perfect Channel State Information (CSI). In the real world, estimating the channel for RIS systems is notoriously difficult and often imperfect.
  • Computational Complexity: While faster than iterative optimization, the message-passing iterations in a GNN still require computational overhead that must be minimized for ultra-low-latency 6G applications.
  • Scalability to Massive Arrays: The simulations used relatively small arrays (16-25 elements). Future work could test how these GNNs handle "extra-large" RISs with hundreds or thousands of elements.

Summary updated

Abstract

Reconfigurable intelligent surface (RIS) is considered as a promising solution for next-generation wireless communication networks due to a variety of merits, e.g., customizing the communication environment. Therefore, deploying multiple RISs helps overcome severe signal blocking between the base station (BS) and users, which is also a practical and effective solution to achieve better service coverage. However, reaping the full benefits of a multi-RISs aided communication system requires solving a non-convex, infinite-dimensional optimization problem, which motivates the use of learning-based methods to configure the optimal policy. This paper adopts a novel heterogeneous graph neural network (GNN) to effectively exploit the graph topology in the wireless communication optimization problem. First, we characterize all communication link features and interference relations in our system with a heterogeneous graph structure. Then, we endeavor to maximize the weighted sum rate (WSR) of all users by jointly optimizing the active beamforming at the BS, the passive beamforming vector of the RIS elements, as well as the RISs association strategy. Unlike most existing work, we consider a more general scenario where the cascaded link for each user is not fixed but dynamically selected by maximizing the WSR. Simulation results show that our proposed heterogeneous GNNs perform about 10 times better than other benchmarks, and a suitable RISs association strategy is also validated to be effective in improving the quality services of users by 30%.
BibTeX
@inproceedings{liu2023cooperative,
  author = {Mengbing Liu and Chongwen Huang and Marco Di Renzo and Mérouane Debbah and Chau Yuen},
  title = {{Cooperative Beamforming and RISs Association for Multi-RISs Aided Multi-Users MmWave MIMO Systems through Graph Neural Networks}},
  booktitle = {ICC},
  year = {2023},
  doi = {10.1109/icc45041.2023.10278986},
}
Figure