HAtt3EGNN: An Unsupervised Power Allocation Method for Cell-Free Massive MIMO

To resolve the problem of power allocation (PA) in cell-free massive MIMO systems, GNN stands out among various machine learning methods due to its utilization of network topology and strong generalization capabilities. In cell-free systems, there is a competitive relationship among UEs served by the same AP, a cooperative relationship among APs serving the same UE, and an interference relationship among different communication links involving different APs and UEs. However, existing GNN methods do not fully utilize this knowledge. In this paper, we model the channels as nodes, with the competition, cooperation, and interference relationships represented as three types of heterogeneous edges. The proposed HAtt3EGNN employs mechanisms of hierarchical attention and edge updating. The hierarchical attention mechanism takes into account both the importance of neighbors along the same meta path and the importance between different meta paths; The edge updating mechanism accelerates the propagation of heterogeneous edge information as we have defined. Benefit from the robust representational capacity of this modeling approach, our proposed graph neural network model demonstrates excellent stability and convergence when employed in an unsupervised learning context. Simulation results indicate that our proposed method outperforms existing GNN methods on the sum-SE performance and approaches the performance of optimal mathematical algorithms.