Learning to Slice: ML-Assisted Segmentation for Dynamic Resource Allocation in SDM-EONs

This paper proposes a machine learning-based bandwidth slicing approach for crosstalk-aware routing, modulation level, core, and spectrum allocation in space division multiplexed elastic optical networks, implemented within a software-defined elastic optical network controller. Our method aims to mitigate blocking probability due to spectrum fragmentation and optical transmission reach while improving the time efficiency of traditional slicing approaches. We implemented and evaluated three classification models, logistic regression, K-nearest neighbors, and decision tree, on a Pan-European topology. The logistic regression model, when deployed, performed the best, reducing the blocking probability by an average of 46% and 4.3% compared to baseline models and further improving time complexity post-training.