Towards explainable artificial intelligence in optical networks: the use case of lightpath QoT estimation

Abstract

Artificial intelligence (AI) and machine learning (ML) continue to demonstrate substantial capabilities in solving a wide range of optical-network-related tasks such as fault management, resource allocation, and lightpath quality of transmission (QoT) estimation. However, the focus of the research community has been centered on ML models’ predictive capabilities, neglecting aspects related to models’ understanding, i.e., to interpret how the model reasons and makes its predictions. This lack of transparency hinders the understanding of a model’s behavior and prevents operators from judging, and hence trusting, the model’s decisions. To mitigate the lack of transparency and trust in ML, explainable AI (XAI) frameworks can be leveraged to explain how a model correlates input features to its outputs. In this paper, we focus on the application of XAI to lightpath QoT estimation. In particular, we exploit Shapley additive explanations (SHAP) as the XAI framework. Before presenting our analysis, we provide a brief overview of XAI and SHAP, then discuss the benefits of the application of XAI in networking and survey studies that apply XAI to networking tasks. Then, we model the lightpath QoT estimation problem as a supervised binary classification task to predict whether the value of the bit error rate associated with a lightpath is below or above a reference acceptability threshold and train an ML extreme gradient boosting model as the classifier. Finally, we demonstrate how to apply SHAP to extract insights about the model and to inspect misclassifications.

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

“…To assess the effectiveness of our approach, we trained an ANN with the same specification, but adopting the sample weights computed as described in the previous session. We selected two specific features due to their criticality for the operation of optical networks, and their relative importance for the output [6]: modulation format and number of spans. The distribution of these two features is shown in Fig.…”

Non-technological barriers: the last frontier towards AI-powered intelligent optical networks

“…A prime example of legacy issues is the ML-aided QoT estimation 15 – 17 operation in optical networks, which has traditionally been dominated by non-ML approaches such as Gaussian noise (GN) model 18 and its variants 19 . ML-based QoT estimation methods, despite offering significant advantage in scenarios involving certain uncertainties about link parameters values 20 , 21 , have not been successful yet in achieving broad adoption in current fiber-optic networks and it may take a while before these techniques are deemed suitable substitutes for their legacy counterparts.…”

“…A relevant example of lack of standardization and regulatory frameworks is the ML-aided lightpaths’ QoT estimation, where the ML algorithms are trained to learn the complex mapping between the feature vectors, comprising of few selected parameters of the link/signal, and the lightpath’s chosen QoT metric 15 – 17 . However, there is presently no standardization of the used feature vectors and various proposed solutions apply dissimilar parameter sets, leading to divergent QoT estimation performances.…”

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

“…In [8], Ayassi et al studied various ML models to estimate the QoT of lightpaths and assessed the feasibility of lightpath establishment in terms of the contributing parameter uncertainties. Ayoub et al [9] proposed an approach based on Exploiting Explainable Artificial Intelligence (XAI) to help understand the behavior of models for lightpath QoT estimation. A long short-term memory (LSTM) deep neural network (DNN) architecture was employed to forecast SNR for one lightpath over a 24-hour horizon based on 13-month historical field data collected in a production network [12].…”