Deep reinforcement learning meets graph neural networks: Exploring a routing optimization use case

Deep Reinforcement Learning (DRL) has shown a dramatic improvement in decision-making and automated control problems. Consequently, DRL represents promising technique to efficiently solve many relevant optimization problems (e.g., routing) self-driving networks. However, existing DRL-based solutions applied networking fail generalize, which means that they are not able operate properly when network topologies observed during training. This lack of generalization capability significantly hinders the deployment technologies production is because state-of-the-art use standard neural networks fully connected, convolutional), suited learn from information structured as graphs. In this paper, we integrate Graph Neural Networks (GNN) into agents design problem specific action space enable generalization. GNNs models inherently designed generalize over graphs different sizes structures. allows proposed GNN-based agent arbitrary topologies. We test our DRL+GNN routing case optical evaluate it on 180 232 unseen synthetic real-world respectively. The results show outperform never seen