Power allocation using spatio-temporal graph neural networks and reinforcement learning

This research investigates power allocation in wireless device-to-device (D2D) networks using spatio-temporal graph neural networks (STGNNs). Specifically, we address the challenge of sum-rate maximization in D2D networks by formulating it as a reinforcement learning problem. In our approach, STGNNs act as agents, generating optimal power allocations to maximize the reward function, which is the overall sum-rate of the network. Our study operates under the realistic assumption of delayed, local channel state information (CSI). Various user mobility patterns, including constant positions, velocities, and accelerations are simulated. The robustness of our proposed method is evaluated against delayed and noisy CSI, which are crucial factors in real-world scenarios. Furthermore, the fairness of our approach is compared to the well-established load-spillage algorithm, which is guaranteed to converge to the globally optimal solution of the alpha-fair utility maximization problem. Finally, the convergence behavior of our method is analyzed in comparison to the policy gradient approach. Our empirical results demonstrate that the proposed STGNN significantly outperforms both the WMMSE benchmark and memoryless graph neural networks (GNNs) across all simulated scenarios, and converges to the globally optimal solution of the load-spillage algorithm, with lower computational complexity. Specifically, it achieves a remarkable performance gap of over 400% compared to the WMMSE algorithm and approximately 10% improvement over the memoryless GNN. These findings underscore the efficacy of STGNNs in addressing power allocation challenges in wireless D2D networks.