aalto1 untyped-item.component.html
Modeling and optimization of unit load behavior for cloud radio network controller
Loading...
URL
Journal Title
Journal ISSN
Volume Title
Sähkötekniikan korkeakoulu |
Master's thesis
Electronic archive copy is available via Aalto Thesis Database.
Authors
Date
Department
Major/Subject
Mcode
ELEC3025
Language
en
Pages
91+10
Series
Abstract
With the telecommunications industry having transitioned to cloud computing and virtualization since the advent of 3G, we are now seeing the benefits of these technologies with regards to dynamic scalability for the effective and timely management and allocation of network resources for managing evolving demands in customer needs. Nokia's AirScale Radio Network Controller uses cloud computing and network function virtualization to yield a scalable solution for managing customer traffic. In order to completely avail the scalability of the AirScale Radio Network Controller from the very start, a revamped dimensioning solution is needed that can provide a reasonable approximation of the virtual network function configuration needed in order to meet the traffic and load threshold demands defined by the customer. This study explores the possiblity of a data driven Machine Learning approach to learn models that can approximate the relationship between virtual network function configuration, telecommunications traffic, and virtual network function load thresholds. Several features representing traffic are included in the dataset and tested for redundancy, and each type of virtual network function component is assigned its own regression target. Insights about the dataset are gained through exploratory data analysis, which reveals strong collinearity between many features, and the need for a more extensive data generation / collection process with respect to regression target values. Linear as well as non-linear machine learning models are evaluated, and the best performing model selected using a two-pronged selection process. The study concludes that non-linear methods yield the best generalization performance, in particular, Artificial Neural Networks and Kernel Ridge Regression. The study also shows that linear models, which are explainable, cannot keep up with non-linear models in terms of generalization performance, so further study involving more sophisticated explainable models e.g. decision trees, needs to be carried out in the future.