5G New Radio specifications demands support of highly dense wireless mobile networks. In these dense environments, power allocation is an important issue for maximizing channel rate and saving energy consumption of devices. The power control problem can be abstracted in to a weighted sum-rate maximization optimization, in which the task is to allocate the optimal power that maximizes the channel rate within a power constraint. However, optimizing this problem is not trivial as it is non-convex and NP-hard.

Power control in a device-to-device network with single-antenna transceivers has been widely analyzed with both classical methods and learning-based approaches. Classical algorithms guarantee a convergence to a local maximum, but its iterative nature and computational complexity makes...[ Read more ]

5G New Radio specifications demands support of highly dense wireless mobile networks. In these dense environments, power allocation is an important issue for maximizing channel rate and saving energy consumption of devices. The power control problem can be abstracted in to a weighted sum-rate maximization optimization, in which the task is to allocate the optimal power that maximizes the channel rate within a power constraint. However, optimizing this problem is not trivial as it is non-convex and NP-hard.

Power control in a device-to-device network with single-antenna transceivers has been widely analyzed with both classical methods and learning-based approaches. Classical algorithms guarantee a convergence to a local maximum, but its iterative nature and computational complexity makes them scale poorly for large network sizes. Although the learning-based methods, i.e., data-driven and model driven, offer performance improvement, the widely adopted graph neural network suffers from learning the heterophilous power distribution.

In this work, we propose a deep learning architecture in the family of graph transformer for wireless power control problems to circumvent the issue. Experiment results show that the proposed methods achieves the state-of-the-art performance across a wide range of untrained network configurations. While the proposed method perform better than available methods, we show there is a trade off between model complexity and generality.