Fusion of Domain Knowledge and Data-Driven Power Allocation Optimization Methods

Abstract

Optimal power allocation in wireless interference networks is challenged by the high computational complexity of traditional methods and the lack of domain knowledge utilization, interpretability, and generalization in deep learning-based approaches. To address these challenges, this paper proposes a fully unfolded weighted minimum mean square error (FUWMMSE) method based on graph neural networks (GNNs). The proposed method integrates the iterative structure of WMMSE with the data-driven advantages of deep learning, incorporating domain knowledge to enhance system’s performance. Further, it enhances the extraction of network topology features and the adaptability to channel scenarios through the optimization of the GNN algorithm architecture. Additionally, the method leverages GNNs to automatically learn hyperparameters, which significantly improves the algorithm’s computational efficiency and expressiveness. Theoretical analysis demonstrates that the proposed approach maintains permutation equivariance, ensuring interpretability and generalization. Numerical experiments validate that FUWMMSE achieves superior performance and adaptability under various network densities, scales, and channel distributions, highlighting its potential for broad applications in wireless communication scenarios.