Machine-Learning-Based Lightpath QoT Estimation and Forecasting

Abstract

Machine learning (ML) is more and more used to address the challenges of managing the physical layer of increasingly heterogeneous and complex optical networks. In this tutorial, we illustrate how simple and more sophisticated machine learning methods can be used in lightpath quality of transmission (QoT) estimation and forecast tasks. We also discuss data processing strategies with the aim to determine relevant features to feed the ML classifiers and predictors. We then introduce a preliminary study on the application of transfer learning to try to overcome the scarcity of field data.

Analysis and Mitigation of Unwanted Biases in ML-based QoT Classification Tasks

“…In (1) all samples have the same weight, not accounting for any (possible) imbalance or bias in the data. To mitigate bias in the data, we can adopt a variation of the binary cross-entropy loss shown in (2), where each sample n is associated with a weight w n . When adopting (2), the challenge becomes how to compute w n such that biases can be accounted for during training.…”

“…Estimating the quality-of-transmission (QoT) of unestablished lightpaths is one of the most important tasks to enable dynamic and efficient optical networking. Artificial intelligence/machine learning (AI/ML) models have been extensively studied for QoT estimation tasks, usually showing strong performance in terms of classification accuracy and estimation error [1,2]. Such models, when correctly applied, enable substantial benefits for the planning and operation of optical networks [3].…”

Forecasting Lightpath Quality of Transmission and Implementing Uncertainty in the Forecast Models

“…Allogba et al [14] implemented SVM and NN models for the real-time estimation of lightpath QoT. Univariate and multivariate LSTM and GRU models were also compared for forecasting tasks.…”

Experimental investigation of physics- and ML-based QoT estimation for WDM systems

“…Both the analytical approach as well as machine learning (ML) based approach has been investigated for QoT estimation. For analytical approaches, usually, the Gaussian noise (GN) model is used [1] whereas various ML approaches such as neural network (NN), support vector machine (SVM), K-nearest neighbour (KNN), random forest (RF) and so on [2][3][4] has been demonstrated. The analytical model suffers from parameter uncertainty in a real system whereas data-driven ML techniques rely on large datasets and provides less explainability.…”