Dynamic GNN for Predictive Resource Allocation to Mobile Users

Predictive resource allocation (PRA) can provide remarkable gains over non-predictive counterparts in boosting the performance of wireless networks. However, existing learning-based PRA methods suffer from poor generalizability and scalability due to using fully-connected, convolutional, or recurrent neural networks. Graph neural networks (GNNs) have been introduced to wireless communications due to their potential in size generalizability and low training complexity. So far, most GNNs for learning wireless policies are static GNNs, whose performance degrades when learning over the dynamic graphs with time-varying topologies to optimize PRA policies. In this paper, we design a dynamic GNN, namely DGNN, to optimize resource allocation with implicit prediction over the dynamic graphs with time-varying vertex sets. We use the predictive bandwidth and power allocation to mobile users for video streaming as an example to demonstrate how to design the DGNN. The designed DGNN consists of several modules for learning the representations of historical graphs and for making the decisions by extracting temporal and spatial information. We show that the DGNN should exhibit permutation property in order to learn over such dynamic graphs. Considering that matching the permutation property of the target function to be learned by a GNN is essential for improving the learning efficiency of the GNN, we identify the desired permutation properties of the modules from the property of the target function of the DGNN, and design the modules to satisfy the desired properties. Simulation results show that the proposed DGNN is superior to relevant learning-based methods in terms of both learning performance and complexity, and is generalizable to the number of users.