A MPGNN-Based Unsupervised Learning Framework for Power Control in D2D Networks

The Internet of Things (IoT) allows physical devices to be connected over wireless networks, and while device-to-device (D2D) communication has become a promising technology for IoT, traditional solutions for D2D resource allocation are often computationally complex and time-consuming. In order to improve the computational efficiency and improve the generalization performance of the algorithm, a large-scale power control unsupervised learning framework algorithm based on Message Passing Graph Neural Network (MPGNN) was proposed, which aims to maximize the weighting and rate of all D2D links. To achieve the above goals, we first model the topology of the wireless network as a graph model, define the channel matrix as the graph structure, and then model the objective function of the system as a loss function and train the model parameters in the MPGNN with the help of the unsupervised learning model. After offline training, each D2D link can obtain the optimal power control strategy in a distributed manner according to the channel state information (CSI). Simulation shows that compared with the classic algorithm based on optimization theory, Weighted Minimum Mean Square Error (WMMSE), the computational efficiency of the proposed algorithm is increased by more than 50 times when the number of users is 100 pairs. In addition, compared with the traditional neural network algorithm, deep neural network (DNN), the proposed algorithm improves the generalization ability of the system under the same parameter settings.