Topology Aware Deep Learning for Wireless Network Optimization

Data-driven machine learning approaches have been proposed to facilitate wireless network optimization by learning latent knowledge from historical optimization instances. However, existing works use simplistic network representations that cannot properly encode the topological difference. They are often limited to fixed topology, and the performance is degraded because the learning target does not get sufficient information since the topological information is not well captured.To address this, we leverage the graphical neural network techniques and propose a two-stage topology-aware deep learning (TADL) framework, which trains a graph embedding unit and a link usage prediction module jointly to discover links likely to be used in optimal scheduling. By properly encoding the network structure, it makes input data with varying topology possible, and also provides more informative clues for the learning target.Important techniques are developed to ensure learning efficiency. The performance is evaluated on canonical multi-hop flow problems with diverse network structures, sizes and realistic deployment scenarios. It achieves close-to-optimum solution quality with a significant reduction in computation time without retraining.