Optimizing Delay in Wireless Networks with Graph Reinforcement Learning

This paper introduces a novel power allocation approach utilizing graph neural networks (GNNs)-based reinforcement learning (RL) to optimize delay in device-to-device (D2D) communication. The method considers various components as state information for each D2D pair, including channel state information, delay, the number of backlogged packets, and the number of transmitted packets. The GNN is used to parameterize these components into a low-dimensional embedding, which is leveraged by a RL agent as its state to optimize power control strategies. Simulation results demonstrate that the proposed method effectively reduces delay, significantly outperforms baseline methods, and possesses strong generalization capability.