Transductive Transfer Learning-Based Spectrum Optimization for Resource Reservation in Seven-Core Elastic Optical Networks

Abstract

In space-division multiplexing elastic optical networks (SDM-EONs), it is important to handle the complex resource optimization (RO) problem due to the coexistence of the requests that require immediate reservation (IR) and those that can be reserved in advance (AR). The introduction of artificial intelligence, especially machine learning algorithms, has made it possible to design more effective RO strategies. However, the model trained by current machine learning algorithms is only suitable for specific scenarios, and cannot be applied to new scenarios, resulting in the need to retrain the models in the new scenario and increase the training costs. Transfer learning (TL) can solve this problem by collecting a small amount of training data in the new scenario and strengthening the existing related models. This paper proposes a RO strategy for resource reservation based on TL in SDM-EONs. If the AR requests fail to reserve resources, TL will be used to predict the spectrum defragmentation time to complete the resource optimization before their start time. For the IR requests, they will occupy the low-level AR requests with the latest start time. Simulation results indicate that the proposed algorithm can decrease the probability for resource reservation failure and improve the resource utilization.