Recursive Router Metrics Prediction Using Machine Learning-Based Node Modeling for Network Digital Replica

Future network infrastructures need to safely and rapidly provide network services in complex conditions that include many devices and multiple access lines, such as 5th-generation (5G) and 6th-generation (6G) mobile systems supported by multiple carriers. Additionally, future telecommunications networks will utilize network disaggregation techniques to take advantage of the highest quality technology from various vendors to meet service requirements. Therefore, it is necessary to enhance the verification of combinations of various network equipment and components that constitute network infrastructure. Our motivation is to investigate the potential to enable the verification of network node performance digitally to support future network infrastructures. This study concentrates on improving the accuracy of the metric inference of black-boxed network nodes when only the network node configurations and traffic conditions are available as external conditions. Our main contribution is as follows: We provide a novel method of machine learning based on network node modeling to improve the accuracy of network node metric inference for throughput, packet loss rate, and packet delay by recursively appending inferred other node metrics to the training datasets in accordance with feature importance; we demonstrate the application of the proposed method to 14 baseline machine learning algorithms for evaluating the accuracy of inferred network node metrics; finally, we show improvement in utilization of network resources for accommodating traffic on a fixed network with a traffic policer, whose parameters are set using the proposed method. Additionally, we investigate the impact of appending inferred network node metrics to the training datasets, which is a key feature of the proposed method, on computational time and the possibility of overfitting.