Traffic Prediction- and Explainable Artificial Intelligence-Based Dynamic Routing in Software-Defined Elastic Optical Networks

The recent research reveals a high potential and efficiency of machine learning (ML) algorithms applied to aid the optimization of telecommunication networks. Moreover, thanks to the tools of explainable artificial intelligence (XAI), it is possible to better understand and interpret the decisions of ML models, identify the features influencing them the most, and use that knowledge for further targeted optimization improvements. In this paper, we use these approaches to aid the optimization of the well-known and essential problem of dynamic routing in software-defined elastic optical networks (EONS). We use the XAI to verify which transmission and network state parameters influence the performance the most (expressed as a bandwidth blocking probability (BBP)) of a dynamic EON. The analysis reveals that two of the most relevant features are related to characteristics of the incoming traffic requests - the source node and the traffic prediction for the demand's source-destination nodes. Next, we use that data to propose six traffic prediction-and XAI-based demands ordering policies, which are used to improve a benchmark allocation algorithm. Then, we perform simulations to evaluate the policies' efficiency and compare them with reference methods. The results show that our targeted proposal is exceptionally efficient - the best policy allowed to reduce the BBP up to about 6.1%, representing 4203 Gbps of more served traffic compared to the best reference policy.