Modeling of Key Quality Indicators for End-to-End Network Management: Preparing for 5G

Abstract

Thanks to evolving cellular telecommunication networks, providers can deploy a wide range of services. Soon, 5G mobile networks will be available to handle all types of services and applications for vast numbers of users through their mobile equipment. To effectively manage new 5G systems, end-to-end (E2E) performance analysis and optimization will be key features. However, estimating the end-user experience is not an easy task for network operators. The amount of end-user performance information operators can measure from the network is limited, complicating this approach. Here we explore the calculation of service metrics [known as key quality indicators (KQIs)] from classic low-layer measurements and parameters. We propose a complete machine-learning (ML) modeling framework. This system's low-layer metrics can be applied to measure service-layer performance. To assess the approach, we implemented and evaluated the proposed system on a real cellular network testbed.

Measuring Key Quality Indicators in Cloud Gaming: Framework and Assessment Over Wireless Networks

“…Moreover, these indicators reduce time and costs that are necessary for traditional QoE assessment. A previous use case for KQIs can be found in References [20,22], where these indicators were estimated and used for decision making in FTP and video streaming services.…”

“…Previous publications refer recommend KQIs in order to assess the QoE of different services [19]. Other works, such as Herrera-Garcia et al [20], present a KQI modeling Artificial Intelligence (AI)-based strategy for video streaming and FTP services, which enables a QoE-based network management in a real cellular network [21]. In the same context, the authors in Reference [22] introduce an application of KQI modeling for network slicing in video streaming, where KQIs are estimated using AI-based methods from application data, giving a provider-network operator perspective.…”

“…Nonetheless, due to its recent expansion, there are no works that analyze the impact of the network on Cloud Gaming KQIs, which enables a service-oriented network configuration as proposed in References [20,22].…”

Acting selfish for the good of all

“…Different research works (e.g., [10,19]) have analyzed clientbased data rate prediction for mobile networks based on network indicator measurements. An important observation is that Classification and Regression Tree (CART)-based methods such as Random Forests (RFs) [5] often achieve a better prediction accuracy than more complex methods such as deep learning which require a significantly higher amount of training data in order to overcome the curse of dimensionality [24].…”

Method for Artificial KPI Generation With Realistic Time-Dependent Behaviour

“…In order to overcome this data deficit in ML training datasets, in the literature the use of testbenches [3], simulations [4] and modelling [2], [5], [6] are proposed. Firstly, in the testbenches [3], commercial equipment originate the data, but they are only partially realistic as a result of nonnatural activity and traffic from the users. In case of real user monitoring, the cons derive from the high number of users and the extensive time-span requirements, which produce datasets with a very high dimensionality, apart from privacy concerns.…”