Towards Self-Driving Optical Networking with Reinforcement Learning and Knowledge Transferring

This paper presents a self-driving networking paradigm exploiting the state-of-the-art reinforcement learning and transfer learning algorithms for highly resource-efficient, scalable, and adaptive next-generation optical networks. The proposed design makes use of learning agents to constantly learn and optimize service provisioning policies through an observe-analyze-act cycle in dynamic operations. Case studies on routing and spectrum assignment tasks show significant advantages of the proposed design over traditional heuristic algorithms.