Meta-learning-aided QoT estimator provisioning for a dynamic VNT configuration in optical networks

Abstract

Machine learning (ML)-based quality-of-transmission (QoT) estimation tools will be desirable for operating virtual network topologies (VNTs) that disclose only abstracted views of connectivity and resource availability to tenants. Conventional ML-based solutions rely on laborious human effort on model selection, parameter tuning, and so forth, which can cause prolonged model building time. This paper exploits the learning-to-learn nature by meta learning to pursue automated provisioning of QoT estimators for a dynamic VNT configuration in optical networks. In particular, we first propose a graph neural network (GNN) design for network-wide QoT estimation. The proposed design learns global VNT representations by disseminating and merging features of virtual nodes (conveying transmitter-side configurations) and links (characterizing physical line systems) according to the routing schemes used. Consequently, the GNN is able to predict the QoT for all the end-to-end connections in a VNT concurrently. A distributed collaborative learning method is also applied for preserving data confidentiality. We train a meta GNN with meta learning to acquire knowledge generalizable across tasks and realize automated QoT estimator provisioning by fine tuning the meta model with a few new samples for each incoming VNT request. Simulation results using data from two realistic topologies show our proposal can generalize QoT estimation for VNTs of arbitrary structures and improves the estimation accuracy by up to 18.7% when compared with the baseline.