Neuron-level Transfer Learning for ANN-based QoT Estimation in Optical Networks

Quality of transmission (QoT) estimation of lightpaths is essential for their deployment in optical networks. Machine learning (ML) is a promising solution for QoT estimation, which needs a large set of training data to ensure the accuracy of ML models. However, due to the shortage of monitorable lightpaths and/or the absence of optical performance monitors, the acquisition of such a number of training data is unavailable. To solve this issue, we propose a neuron-level transfer learning (NL-TL) scheme for ANN-based QoT estimation, where a neuron importance-based transfer model search algorithm is proposed to determine the trainable neurons. Simulation results show that the NL-TL-based QoT estimation scheme improves estimation accuracy compared with conventional layer-level transfer learning-based QoT estimation schemes.

I. Introduction

Estimating the quality of transmission (QoT) of lightpaths prior to their establishment is capitally important for ensuring the reliability of lightpaths in optical networks. Recently, machine learning (ML), especially the artificial neural network (ANN), becomes a promising technology for QoT estimation. As for the ANN-based QoT estimator, a large set of training samples from lightpaths' observations are required to ensure its accuracy [1]. Unfortunately, the acquisition of the training samples is hindered by practical limitations, such as the shortage of monitorable lightpaths in the early stage of optical network deployment or/and the absence of optical monitors in partial network nodes. Therefore, how to obtain a high-precision QoT estimator with few training samples becomes a crucial problem.