GNN-based Optimization for Multi-BS Multi-RIS Pinching-Antenna Systems

A recent study available on arXiv (2605.01307) introduces a graph neural network (GNN) method aimed at enhancing coordinated downlink transmission in systems featuring multiple base stations (multi-BS) and multi-reconfigurable intelligent surfaces (multi-RIS) with pinching antennas (PA). This system incorporates movable PAs situated on parallel waveguides, with each user equipment (UE) linked to a specific BS. The researchers address the challenges of maximizing sum rate (SR) and energy efficiency (EE) by optimizing PA positioning, RIS phase adjustments, transmit beamforming, and BS-UE connections while adhering to constraints on inter-PA distance, power limits, and phase shift requirements. To tackle this complex mixed-variable issue, they create a three-stage GNN that combines both heterogeneous and homogeneous graph models, trained in an unsupervised fashion. Their numerical findings demonstrate that the proposed GNN consistently surpasses existing system and learning benchmarks.