© [2019 Optical Society of America]. Users may use, reuse, and build upon the article, or use the article for text or data mining, so long as such uses are for non-commercial purposes and appropriate attribution is maintained. All other rights are reserved.

Autonomous network operation realized by means of control loops, where prediction from machine learning (ML) models is used as input to proactively reconfigure individual optical devices or the whole optical network, has been recently proposed to minimize human intervention. A general issue in this approach is the limited accuracy of ML models due to the lack of real data for training the models. Although the training dataset can be complemented with data from lab experiments and simulation, it is probable that once in operation, events not considered during the training phase appear and thus lead to model inaccuracies. A feasible solution is to implement self-learning approaches, where model inaccuracies are used to re-train the models in the field and to spread such data for training models being used for devices of the same type in other nodes in the network. In this paper, we develop the concept of collective self-learning aiming at improving the model’s error convergence time as well as at minimizing the amount of data being shared and stored. To this end, we propose a knowledge management (KM) process and an architecture to support it. Besides knowledge usage, the KM process entails knowledge discovery, knowledge sharing, and knowledge assimilation. Specifically, knowledge sharing and assimilation are based on distributing and combining ML models, so specific methods are proposed for combining models. Two use cases are used to evaluate the proposed KM architecture and methods. Exhaustive simulation results show that model-based KM provides the best error convergence time with reduced data being shared.

Peer Reviewed