Graph Neural Network-Based Intelligent Resource Allocation in RIS-Assisted 6G Networks

Abstract

Reconfigurable intelligent surfaces are among the strongest candidate technologies for sixth-generation wireless networks because they shape the propagation environment using nearly passive elements rather than additional radio chains. Realising that promise requires solving a joint beamforming and phase-shift problem that is non-convex, tightly coupled and must be re-solved whenever the channel changes. Classical alternating optimisation needs hundreds of iterations per channel realisation, which is incompatible with millisecond coherence times, and neural networks tied to one array geometry must be retrained whenever the network size changes. This work proposes HRG-Net, a heterogeneous relational graph network that represents the downlink as a bipartite graph over user and surface-element nodes and performs message passing along cascaded-channel edges. Because aggregation is permutation equivariant in both node types, one trained model transfers across network sizes it never saw. A model-driven output layer anchors the solution on regularised zero-forcing directions and learns only the power split, the surface phases and a residual correction, and the whole network is trained without labels by maximising achievable sum rate directly. Simulations of a multi-user multiple-input single-output downlink show that HRG-Net attains 108.8% of the weighted minimum mean square error sum rate while requiring 16 times less computation per channel realisation.