|
Title
|
|
|
|
Machine learning for wireless communication : from next-generation spectrum sharing frameworks to communication-aware learning agents
| |
|
Author
|
|
|
|
| |
|
Abstract
|
|
|
|
| As services and networks grow more complex, the need for network management automation rises, encompassing both services and network functions. The primary aim of these systems is to optimize network resources to achieve efficiency while meeting user requirements, leading to the concept of Autonomous Networks (ANs). Advances in Artificial Intelligence (AI) and Machine Learning (ML) enable these networks to manage resources autonomously, learning and adapting to complex environments to perform tasks like self-healing and self-provisioning. The radio access domain is evolving into a fully Autonomous Domain (AD), driven by the complexities of wireless networks and the demanding requirements of 5G and future networks. These challenges necessitate advanced management techniques and innovations in spectrum allocation, with Cognitive Radios (CRs) and Dynamic Spectrum Access (DSA) emerging as critical technologies. New AI techniques, particularly Deep Learning (DL), provide the tools to manage these networks intelligently. Successful deployment of radio access ADs requires significant innovation in AI/ML algorithms for resource management, addressing challenges in algorithm selection, deployment location, and lifecycle management. Tailored AI solutions must consider sustainability, reliability, scalability, resource awareness, training efficiency, communication overhead, and responsiveness. This dissertation explores these challenges from two perspectives: radio networks and ML. It proposes a novel spectrum-sharing framework for radio networks, using a two-tier architecture for efficient spectrum sharing and protection of incumbent transmissions. A general framework for developing Traffic Classification (TC) algorithms optimized for wireless networks is also presented, along with a DL-based Technology Recognition (TR) algorithm. On the ML side, the dissertation addresses training efficiency and scalability in DL and Reinforcement Learning (RL) algorithms for networking problems. It introduces a Semi-supervised Learning (SSL) based DL algorithm for label-efficient TR and a novel approach to minimize communication overhead in distributed Parallel Reinforcement Learning (PRL) algorithms. These contributions aim to advance the deployment and management of autonomous and intelligent network systems. |
| |
|
Language
|
|
|
|
English
| |
|
Publication
|
|
|
|
Antwerp
:
University of Antwerp, Faculty of Sciences, Department of Computer Science
,
2024
| |
|
DOI
|
|
|
|
10.63028/10067/2077280151162165141
| |
|
Volume/pages
|
|
|
|
xxix, 166 p.
| |
|
Note
|
|
|
|
:
Latré, Steven [Supervisor]
| |
|
Full text (open access)
|
|
|
|
| |
|