Robust Restoration of IP Traffic from Optical Failures by Deep Reinforcement Learning and Graph Neural Networks

This paper applies Deep Reinforcement Learning (DRL) and Graph Neural Networks (GNN) for robust restoration of IP traffic from optical failures in IP-Optical transport networks. For that, DRL-GNN algorithms are formulated to optimize optically-aware IP restoration policies that minimize traffic losses. This is achieved by training DRL-GNN agents that learn IP routing policies in preselected sets of optical failures. Once trained, the agents are able to efficiently restore IP traffic from foreseen and unforeseen optical failures. The performance of the algorithms is assessed in selected network scenarios, where it is shown that - in the worst case - the algorithms reduce the traffic losses by a factor of 34.86% w.r.t the traffic losses obtained by exact robust Linear Program formulations.