Simulation study demonstrates improved QoS metrics and routing stability in programmable software-defined networks, highlighting a scalable path toward autonomous network management.
Key Points
To develop and evaluate a combined Graph Neural Network and Reinforcement Learning framework with a multi-objective reward function for adaptive, topology-aware traffic routing in programmable networks.
Integrated Graph Neural Networks (GNNs) for structural topology modeling with Reinforcement Learning (RL) for adaptive routing decisions in an automated cloud simulation environment.
Trained and evaluated the hybrid framework using two heterogeneous network flow datasets (NetBench and SDNFlow) under a multi-objective reward function balancing throughput, latency, packet loss, and congestion penalty.
Reduced average network latency from over 70.7 ms down to 17.7 ms, reaching 9.6 ms with reward optimization.
Decreased packet loss from 6.3% to 1.49% while maintaining throughput at 151–231 Mbps, outperforming standalone baselines (85–88 Mbps).
Improved training convergence rate and enhanced routing stability under dynamic and heterogeneous traffic conditions.