Improving Network Delay Predictions Using GNNs

Abstract Autonomous network management is crucial for Fifth Generation (5G) and Beyond 5G (B5G) networks, where a constantly changing environment expected configuration must adapt accordingly. Modeling tools are required to predict the impact on performance (packet delay loss) when new traffic demands arrives changes in routing paths applied network. Mathematical analysis simulators techniques modeling networks but both have limitations, as former provides low accuracy latter requires high execution times. To overcome these machine learning (ML) algorithms, more specifically, graph neural (GNNs), proposed due their ability capture complex relationships from graph-like data while predicting properties with computational requirements. However, one of main issues using GNNs lack generalization capability larger i.e., trained small (in number nodes, length, links capacity), predictions poor. This paper addresses GNN problem by use fundamental networking concepts. Our solution built baseline model called RouteNet (developed Barcelona Neural Networking Center-Universitat Politècnica de Catalunya (BNN-UPC)) that predicts average paths, through simple additions significantly improves prediction networks. The improvement ratio compared 101, 187.28% 1.828%, measured Mean Average Percentage Error (MAPE). In addition, we propose closed-loop control context resulting could be potentially used different cases.