Learning Uplink Power Control with MMSE Receiver-based Graph Neural Network

Optimizing uplink resource allocation is of paramount importance to support the growing demands for high-rate uplink applications. Optimizing power control is an effective method to improve spectral efficiency (SE), but existing iterative algorithms are with unaffordable computational complexity, while existing learning-based approaches that optimize transmit power based on suboptimal receivers fail to maximize SE. In this paper, we learn power control policy based on graph neural networks (GNNs) given a minimum mean square error (MMSE) receiver. To improve both the learning performance and learning efficiency, we propose an MMSE-GNN, which incorporates the MMSE receiver model and iteratively learns the optimal power control from equivalent channel gains. Simulation results demonstrate that compared to the pure data-driven GNN, the MMSE-GNN can increase the SE of the learned policy, reduce the training complexity, and enhance the generalization ability to the numbers of antennas and users. Furthermore, even when the base station has partial channel information, the MMSE-GNN achieves favorable learning performance.

I. Introduction

With the increasing demand for high-rate uplink applications, uplink resource allocation has drawn attention recently [1]. Optimizing power control is one of the effective resource allocation techniques to enhance uplink spectral efficiency (SE). Several iterative algorithms, such as weighted minimum mean square error (WMMSE) [2] and fractional programming [3] can be used to solve power control problems. However, the computational complexity of these algorithms is high, making them hard to be implemented in real-time.