A Graph Neural Network Power Allocation Algorithm Based on Fully Unrolled WMMSE

For the optimal power allocation problem in interference networks, traditional methods exhibit high computational complexity, while deep learning-based methods often underutilize domain knowledge. To address these challenges, we propose a novel approach termed Weighted Minimum Mean Square Error (WMMSE) full unfolding-based graph neural networks (FUWMMSE). This method integrates key modeling elements from traditional approaches with data-driven components of deep learning, employing the concept of deep unfolding-based methods. Specifically, by unfolding the iterative WMMSE method, optimizing hyperparameters, designing a graph neural networks (GNNs) network structure, augmenting modeling elements with GNNs, and leveraging the strengths of both model-based and data-based methods, our proposed neural network architecture is formulated. This design aims to overcome the limitations of insufficient domain knowledge utilization, complex computational structures, and poor generalization abilities observed in current unfolding-type methods. Theoretical analysis demonstrates that our method possesses the desirable properties of permutation equivariance, ensuring both interpretability and generalizability. Extensive numerical experiments validate that the FUWMMSE method has better performance and generalization capability than other unfolding class methods while maintaining a fast convergence speed.