Cross Layer Power Allocation by Graph Neural Networks in Heterogeneous D2D Video Communications

Abstract

The massive connectivity trend was set to shape B5G/6G networks. Device-to-device (D2D) communications play a crucial role in massive connectivity in the context of Internet of Things (IoT) applications. Recently, a heterogeneous interference graph neural network (HIGNN) was proposed for resource allocation in heterogeneous networks. The HIGNN captured the spatial information hidden in heterogeneous network topology and was scalable. However, existing methods primarily focused on resource allocation at the physical layer only and did not adequately address the cross-layer issues involved in video transmission. Therefore, in this paper, we propose the video-optimized heterogeneous interference graph neural network (VD-HIGNN) as a cross-layer D2D resource allocation method for video transmission, which introduces the following contributions: 1) joint source encoder rate and beamforming/power control, 2) incorporating video rate distortion function parameters from the application layer into the node features, and 3) changing the loss function from data rate to Peak-Signal-to-Noise-Ratio (PSNR), a function of video rate distortion and a metric of video quality. Simulation results demonstrate that our proposed VD-HIGNN outperforms two physical layer baseline schemes: the iterative fractional programming method by 0.53 dB and HIGNN by approximately 2 dB for video transmission. Moreover, when scaled to larger problems with 2-12 times the number of nodes within a fixed area size, the VD-HIGNN achieves 94% or more of the performance of a retrained model, showcasing its scalability and generalization ability.