Beyond Explaining Black-box Models: An Interpretable DeepRMSA Reformation for EONs

The lack of model interpretability presents as one of the major obstacles to the practical deployment of machine learning (ML) applications for critical optical network tasks. This paper proposes an interpretable DeepRMSA reformation for elastic optical networks (EONs) exploiting fragmentation representations learned automatically through training blocking probability predictors and knowledge distillation. Results show that the proposed design can accelerate training by ∼12× while achieving a blocking reduction of ∼36% over the original design.