FuguReport

Spectral- and Energy-efficient Multi-BS Multi-RIS Pinching-antenna Systems: A GNN-based Approach

Authors Changpeng He, Yang Lu, Wei Chen, Bo Ai, Arumugam Nallanathan, Zhiguo Ding
Affiliations Kyung Hee University / Queen Mary University of London / Beijing Jiaotong University / Nanyang Technological University
Categories Method / Graph Neural Networks / GNN for wireless system optimization, Application / Wireless Communication / Multi-BS multi-RIS downlink coordinated transmission, Evaluation / Energy Efficiency Evaluation / Spectral and energy efficiency joint optimization
License CC BY 4.0

Abstract Overview

This paper investigates coordinated downlink transmission in a multi-base-station (multi-BS), multi-RIS-assisted pinching-antenna (PA) system where each user is associated with one BS and each BS uses movable PAs on parallel waveguides. The authors formulate both sum-rate (SR) and energy-efficiency (EE) maximization problems by jointly optimizing PA placement, RIS phase shifts, transmit beamforming, and BS-UE association under inter-PA spacing, power budget, and unit-modulus constraints. To address the resulting mixed-variable, tightly coupled optimization problem, they propose a three-stage graph neural network composed of ChanGNN, BeamGNN, and AssocGNN. The model combines heterogeneous and homogeneous graph representations, is trained end-to-end without supervision, and incorporates feasibility-preserving output mappings for all constrained variables.

Novelty

The paper's main novelty is a unified three-stage GNN architecture for multi-BS, multi-RIS pinching-antenna systems that jointly handles transmitter-side antenna repositioning, RIS configuration, beamforming, and BS-UE association—a combination not previously addressed in the open literature. It is also distinctive in integrating heterogeneous graph attention (for BS-RIS-UE interactions) with homogeneous graph attention (for inter-BS and intra-BS interference modeling), along with explicit feasibility-preserving mechanisms for placement spacing, unit-modulus phase, per-BS power, and binary association constraints.

Results

Numerical results show that the proposed method consistently outperforms reported system baselines (No-RIS, Fixed-PA, Random-U) and model baselines (MLP, Single HAN, HAN, GAT) in both SR and EE across multiple problem sizes. The model generalizes to unseen numbers of UEs, BSs, and RISs with moderate degradation under larger topology mismatch, while maintaining millisecond-level inference time. Ablation studies confirm the importance of message passing, residual connections, and complex-valued fully-connected layers, and experiments show that movable PA and RIS gains are complementary with performance improvements increasing as the number of PAs grows before saturating.

Key Points

  1. A three-stage unsupervised GNN (ChanGNN, BeamGNN, AssocGNN) jointly optimizes PA placement, RIS phase shifts, beamforming via HZM decomposition, and BS-UE association for multi-BS multi-RIS downlink coordinated transmission.
  2. The architecture integrates heterogeneous graph attention layers (for BS-RIS-UE relational features) and homogeneous graph attention layers (for inter-BS and intra-BS interference), with feasibility-preserving mappings that enforce spacing, unit-modulus, power budget, and one-hot association constraints by design.
  3. Experiments demonstrate consistent EE and SR gains over system and model baselines across varying problem sizes, scalability to unseen network configurations with millisecond-level inference, and complementary benefits of RIS assistance and PA mobility.

References

This page was created using generative AI such as GPT-5, Claude Opus 4, Gemini 3, Gemini 3.1 Flash Image, and their higher-end successor versions. No guarantee can be made regarding its contents.