Opti-DeepRoute: A Topology-Adaptive Deep Reinforcement Learning Based Service Provisioning Framework for Elastic Optical Network

The satisfaction of AI computing demands relies on the robust computational foundation provided by efficient optical backbone network. Deep reinforcement learning (DRL) methods have been widely adopted recently to address elastic optical network (EON) service provisioning problems with complex spectrum constraints. However, previous frameworks utilizing path-level decision spaces face challenges as path-level features cannot be directly transferred when network topology changes, making it necessary to retrain the agent, hindering the practical deployment of the algorithm. Based on the above observations, this paper proposes a novel hop-by-hop DRL service provisioning framework for EON: Opti-DeepRoute. A node-level feature engineering is employed to construct solutions incrementally through the creation of node sequences. Utilizing the topological and spectral local receptive fields provided by the Graph Attention Network (GAT), the agent acquires local network states and partial solution information to achieve topological adaptability. Preliminary evaluations demonstrate the advantages of Opti-DeepRoute in terms of topological adaptability compared with SOTA DRL approaches.