Ray-Traced D2D Link Scheduling: Dataset, Code and Results for "Graph Neural Networks for D2D Link Scheduling: Centralized and Distributed Schedulers Evaluated with Sionna Ray-Traced Channels"

Abstract

Supporting dataset, source code, trained models and result files for the manuscript "Graph Neural Networks for D2D Link Scheduling: Centralized and Distributed Schedulers Evaluated with Sionna Ray-Traced Channels" (Sensors, MDPI; revised version). The archive contains 24,000 device-to-device network snapshots generated by Sionna ray tracing on the publicly available Munich scene at 3 GHz (complex 10x10 channel matrices with transmitter and receiver coordinates), the code for a centralized graph-attention scheduler and a fully distributed per-pair scheduler trained with policy-gradient reinforcement learning, the trained models, and the result files from which every table and figure of the revised manuscript can be reproduced. Changes in version 1.1.0: the link budget was corrected (the antenna gain, already contained in the ray-traced channel coefficients, had been added a second time in version 1.0.0); all learned schedulers were retrained at the corrected operating point (23 dBm) with five random seeds each, and a feature ablation was added (27 models in total); all results were recomputed, and results for CPU/GPU latency, further CSI impairments, and operating-point sensitivity were added. The ray-traced dataset is unchanged.