Machine Learning and Data Science for Low-Margin Optical Networks

The point of lowering optical network margins is to optimize the efficiency of current and future infrastructures. Margins are initially meant to guarantee a high reliability. Thus, the low-margin conundrum is to determine how and by how much margins may be reduced with as little impact as possible on the quality of service. Solving this inevitably requires to forecast then track the consequences of margin reductions. This cannot be achieved without large-scale collection of monitoring data and advanced data science tools such as machine learning. For true optimization, it is also essential to clarify the complex interdependence between capacity, quality of transmission and reliability. To that end, we start by proposing a generalization of the concept of margin to achieve a more holistic view of margin optimization. Then we leverage public data from a large backbone network, to quantify the trade-off between capacity and availability in the context of rate elasticity, and to test the performances of rate automation algorithms as margin optimization schemes. We finally discuss the use of machine learning in the broad context of margin optimization. In this regard, we propose several use cases where combined with expert knowledge, machine learning can be most efficiently leveraged.