Management of cellular broadband networks by means of machine learning techniques
Διαχείριση ετερογενών ευρυζωνικών δικτύων με χρήση μηχανισμών μηχανικής μάθησης
Doctoral Thesis
Author
Margaris, Aristotelis
Μάργαρης, Αριστοτέλης Γ.
Date
2021-02-05View/ Open
Keywords
3GPP ; Het-Net ; Ultra-Dense Networks ; Optimization ; Network management ; Knowledge-building ; High-Level Objectives ; Energy efficiency ; EMF reduction ; Quality of service ; Fault prevention ; Network KPI Forecasting ; Network element clustering ; Self-organizing maps ; Growing neural gas ; Neural networks ; Gradient boosted treesAbstract
Cellular networks are one of the most impactful technologies of today’s ICT industry. They provide wireless access to internet and services with very high availability and effectiveness. The evolution of this technology comes with the maturity of the 3GPP-based network and their upcoming releases that promise to deliver even higher quality of service, additional capabilities, and solutions to previous drawbacks. To achieve this, vendors of these technologies must analyze the complexity of these networks and their different deployment options and provide intelligent management software. Variations of cellular networks can be found in literature as Heterogeneous Cellular Networks (HetNets) or Ultra-Dense networks which are improved design flavors of the same system with increased complexity and configurations. The added capabilities of these networks must be used as a toolbox to improve various operational aspects of the networks such as energy efficiency, network performance and system fault prevention. The scope of this Doctorate Thesis is to analyze different approaches of optimizing HetNets in order to suggest plausible suggestions for extensions that will optimize all high-level objectives. Static management and configuration will be used in conjunction with knowledge-building to improve the energy efficiency of key simulation scenarios of 3GPP networks. Dynamic Resource allocation schemes will be used as a real time management algorithm to improve quality of service in a micro-scale. Predictive models based on acquired historical data will be used to predict network operational KPIs, evaluate the probability of network congestion and identification of unknown network element groups based on their behavior. These generated insights will help the infrastructure providers to impose countermeasures to prevent quality deterioration and enforce the technological standards. They will also lead to the reduction of the OPEX and the energy footprint of the system making technology investments sustainable and profitable for network operators. The framework for developing and testing these algorithms is a custom-designed software platform for HetNet simulations and algorithm experimentation. This system is designed according to standards and specifications in order to provide realistic results that will establish the suggested algorithms as strong candidates to be included in future 3GPP-based wireless networks.