Deep Learning for Energy Efficient Beamforming in MU-MISO Networks: A GAT-Based Approach

Abstract

This letter investigates the deep learning enabled energy efficient beamforming design for multi-user (MU) multiple-input single-output (MISO) networks. An energy efficiency (EE) maximization problem is formulated and solved by a proposed block successive upper bound minimization (BSUM)-based algorithm. Then, the MU-MISO network is reformulated into a graph and input into a proposed graph attention network (GAT) which is utilized to embed the node features due to the inter-link interference. With the labeled datasets generated by the BSUM algorithm, the proposed GAT is trained to learn the mapping from the channel state information to the beamforming vectors with the aim of maximizing the EE. Some interpretable insights were presented in designing each layer of the GAT. Numerical results validate that the proposed GAT-based approach is able to derive near-optimal performance and adapt to the dynamic and scalable wireless networks with real-time computation.