Machine learning for quality of transmission: a picture of the benefits fairness when planning WDM networks

Abstract

In the present day, the evaluation of machine learning (ML) as a candidate for substituting analytical quality of transmission (QoT) estimators is done in a compartmentalized way. The assessment is not produced from a global optical network design perspective and with accurate optical design metrics; on the contrary, the evaluation heavily focuses on the physical layer impairment precision capabilities while underemphasizing the effects at the network layer. In this paper, we recommend a suitable methodology for evaluating the QoT substitution based on the foundational idea that different QoT estimators should be examined on a comparative basis by analyzing network-relevant metrics at parity of availability performance. Pragmatically, we recommend comparing performance estimation solutions through the aggregate network throughput, i.e., capacity, at the equity of their overestimation likelihood, which drives system margins. To demonstrate the need for such a network viewpoint and illustrate the potential drawbacks of an inadequate assessment of the QoT substitution, we use the proposed method in several scenarios (altering network topologies, input parameter uncertainty conditions, and availability requirements), showing that we can achieve gains in QoT estimation error or design margins while observing notable losses in terms of network throughput. Considering the results were contrary to what one may expect, we decided to undergo a statistical analysis in order to investigate and grasp the consequences of the model error distribution in relation to the network capacity.