A Performance Analysis of Supervised Learning Classifiers for QoT Estimation in ROADM-based Networks

Abstract

Machine learning techniques for optimization purposes in the optical domain have been reviewed extensively in recent years. While several studies are pointing in the right direction towards building enhanced transport network control systems including estimation algorithms, the physical effects encountered in the optical domain raise several challenges that are hard to learn from and mitigate. In this paper, we provide a performance analysis of various supervised learning algorithms when predicting the Quality of Transmission (QoT), in terms of signal to noise ratio (OSNR), of light paths when erbium doped fiber amplifier (EDFA) power excursions and fiber nonlinearities are taken into account. The analysis considers F1-scores and computational training times as the main comparison metrics. A customized optical data network simulator was used for the generation of synthetic labeled data samples. Our results depict similar performance among groups of classifiers, and a correlation between the data sample size and the prediction accuracy.