Cognitive networking for next generation of cellular communication systems

Öztürk, Metin (2020) Cognitive networking for next generation of cellular communication systems. PhD thesis, University of Glasgow.

Full text available as:
[thumbnail of 2020OzturkPhD.pdf] PDF
Download (2MB)

Abstract

This thesis presents a comprehensive study of cognitive networking for cellular networks with contributions that enable them to be more dynamic, agile, and efficient. To achieve this, machine learning (ML) algorithms, a subset of artificial intelligence, are employed to bring such cognition to cellular networks. More specifically, three major branches of ML, namely supervised, unsupervised, and reinforcement learning (RL), are utilised for various purposes: unsupervised learning is used for data clustering, while supervised learning is employed for predictions on future behaviours of networks/users. RL, on the other hand, is utilised for optimisation purposes due to its inherent characteristics of adaptability and requiring minimal knowledge of the environment.

Energy optimisation, capacity enhancement, and spectrum access are identified as primary design challenges for cellular networks given that they are envisioned to play crucial roles for 5G and beyond due to the increased demand in the number of connected devices as well as data rates. Each design challenge and its corresponding proposed solution are discussed thoroughly in separate chapters.

Regarding energy optimisation, a user-side energy consumption is investigated by considering Internet of things (IoT) networks. An RL based intelligent model, which jointly optimises the wireless connection type and data processing entity, is proposed. In particular, a Q-learning algorithm is developed, through which the energy consumption of an IoT device is minimised while keeping the requirement of the applications--in terms of response time and security--satisfied. The proposed methodology manages to result in 0% normalised joint cost--where all the considered metrics are combined--while the benchmarks performed 54.84% on average. Next, the energy consumption of radio access networks (RANs) is targeted, and a traffic-aware cell switching algorithm is designed to reduce the energy consumption of a RAN without compromising on the user quality-of-service (QoS). The proposed technique employs a SARSA algorithm with value function approximation, since the conventional RL methods struggle with solving problems with huge state spaces. The results reveal that up to 52% gain on the total energy consumption is achieved with the proposed technique, and the gain is observed to reduce when the scenario becomes more realistic.

On the other hand, capacity enhancement is studied from two different perspectives, namely mobility management and unmanned aerial vehicle (UAV) assistance.
Towards that end, a predictive handover (HO) mechanism is designed for mobility management in cellular networks by identifying two major issues of Markov chains based HO predictions. First, revisits--which are defined as a situation whereby a user visits the same cell more than once within the same day--are diagnosed as causing similar transition probabilities, which in turn increases the likelihood of making incorrect predictions. This problem is addressed with a structural change; i.e., rather than storing 2-D transition matrix, it is proposed to store 3-D one that also includes HO orders. The obtained results show that 3-D transition matrix is capable of reducing the HO signalling cost by up to 25.37%, which is observed to drop with increasing randomness level in the data set. Second, making a HO prediction with insufficient criteria is identified as another issue with the conventional Markov chains based predictors. Thus, a prediction confidence level is derived, such that there should be a lower bound to perform HO predictions, which are not always advantageous owing to the HO signalling cost incurred from incorrect predictions. The outcomes of the simulations confirm that the derived confidence level mechanism helps in improving the prediction accuracy by up to 8.23%.

Furthermore, still considering capacity enhancement, a UAV assisted cellular networking is considered, and an unsupervised learning-based UAV positioning algorithm is presented. A comprehensive analysis is conducted on the impacts of the overlapping footprints of multiple UAVs, which are controlled by their altitudes. The developed k-means clustering based UAV positioning approach is shown to reduce the number of users in outage by up to 80.47% when compared to the benchmark symmetric deployment.

Lastly, a QoS-aware dynamic spectrum access approach is developed in order to tackle challenges related to spectrum access, wherein all the aforementioned types of ML methods are employed. More specifically, by leveraging future traffic load predictions of radio access technologies (RATs) and Q-learning algorithm, a novel proactive spectrum sensing technique is introduced. As such, two different sensing strategies are developed; the first one focuses solely on sensing latency reduction, while the second one jointly optimises sensing latency and user requirements. In particular, the proposed Q-learning algorithm takes the future load predictions of the RATs and the requirements of secondary users--in terms of mobility and bandwidth--as inputs and directs the users to the spectrum of the optimum RAT to perform sensing. The strategy to be employed can be selected based on the needs of the applications, such that if the latency is the only concern, the first strategy should be selected due to the fact that the second strategy is computationally more demanding. However, by employing the second strategy, sensing latency is reduced while satisfying other user requirements. The simulation results demonstrate that, compared to random sensing, the first strategy decays the sensing latency by 85.25%, while the second strategy enhances the full-satisfaction rate, where both mobility and bandwidth requirements of the user are simultaneously satisfied, by 95.7%.

Therefore, as it can be observed, three key design challenges of the next generation of cellular networks are identified and addressed via the concept of cognitive networking, providing a utilitarian tool for mobile network operators to plug into their systems. The proposed solutions can be generalised to various network scenarios owing to the sophisticated ML implementations, which renders the solutions both practical and sustainable.

Item Type: Thesis (PhD)
Qualification Level: Doctoral
Keywords: Cognitive networking, cellular networks, machine learning, intelligence, capacity enhancement, energy optimisation, dynamic spectrum access.
Subjects: T Technology > T Technology (General)
T Technology > TK Electrical engineering. Electronics Nuclear engineering
Colleges/Schools: College of Science and Engineering > School of Engineering > Electronics and Nanoscale Engineering
Supervisor's Name: Imran, Professor Muhammad and Hussain, Dr. Sajjad
Date of Award: 2020
Depositing User: Mr. Metin Öztürk
Unique ID: glathesis:2020-81368
Copyright: Copyright of this thesis is held by the author.
Date Deposited: 03 Jun 2020 10:22
Last Modified: 08 Sep 2022 16:33
Thesis DOI: 10.5525/gla.thesis.81368
URI: https://theses.gla.ac.uk/id/eprint/81368
Related URLs:

Actions (login required)

View Item View Item

Downloads

Downloads per month over past year