Is Machine Learning the Best Option for Network Routing?

Liou Tang, Prashant Krishnamurthy, Mai Abdelhakim

Abstract: Machine Learning (ML)-based algorithms have been widely adopted in communication networking optimization problems However, many studies that utilize ML-based approaches often overlook the comparison between ML algorithms and traditional, heuristic algorithms. In this paper, we study the merits and downsides of ML-based algorithms in a Software Defined Networking (SDN) routing scenario by analyzing a Deep Reinforcement Learning (DRL) routing algorithm assisted by a Graph Neural Network (GNN). The performances of the ML and traditional routing algorithms are evaluated in different network topologies. We consider a novel network reliability metric as well. We observe that traditional routing algorithms provide comparable performance to ML.