- 107 views
- 135 downloads
Machine Learning and Stochastic Geometry Techniques for Future Mobile Communications
-
- Author / Creator
- Banerjee, Bitan
-
Advancements in wireless communication are continuously evolving, and the progression towards the 6th generation (6G) and beyond of cellular architecture will heavily rely on the implementation of machine learning (ML) algorithms in both cellular devices and base stations (BSs), along with the deployment of highly dense networks. ML algorithms have the potential to grant devices the ability to autonomously adapt and modify themselves, while also facilitating decision-making processes involving non-deterministic polynomial-time (NP)-hard problems. Therefore, this thesis explores the potential of machine learning (ML) algorithms in shaping the future of wireless communication and cellular architecture. Specifically, it focuses on addressing the challenges faced by conventional architectures in meeting the data rate and reliability requirements of the anticipated 6G cellular architecture. The research investigates the application of machine learning and stochastic geometry techniques to propose novel approaches for enhancing performance and overcoming limitations.
The thesis presents a heterogeneous network (HetNet) model that correlates the locations of small cell base stations (SBSs) with macro base stations (MBSs) using a Poisson-Voronoi tessellation. Theoretical analysis of this deployment scheme is studied using the tools of stochastic geometry and the results indicate an improvement up to 21% in the coverage probability and up to 28% in the rate coverage. This thesis also introduces a conditional generative adversarial network (CGAN)-based algorithm for uplink (UL) to downlink (DL) channel covariance matrix (CCM) mapping and direct UL to DL channel state information (CSI) mapping in massive MIMO systems operating in a frequency division duplex (FDD) mode.
Additionally, the research explores multi-agent reinforcement learning (MARL) and multi-agent federated reinforcement learning (MAFRL) algorithms for access point (AP) selection and clustering in cell-free networks. The MARL and MAFRL algorithms provide a sub-optimal solution to an NP-hard problem and achieve up to 88.3% of the maximum possible sum spectral efficiency achievable if all APs were to serve all users using centralized precoding. Furthermore, the thesis investigates the combination of long short-term memory (LSTM) and CGAN for predicting downlink CSI from earlier uplink CSI estimates and estimating complete uplink CSI from incomplete information. This algorithm demonstrates the ability to provide reliable network service to users moving at vehicular speeds with limited available power.
Through these approaches, the thesis aims to contribute to the development of more efficient and reliable cellular systems for 6G and beyond. The research findings demonstrate the potential of ML algorithms and highlight the benefits of integrating stochastic geometry and machine learning techniques in wireless communication systems.
-
- Graduation date
- Fall 2023
-
- Type of Item
- Thesis
-
- Degree
- Doctor of Philosophy
-
- License
- This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.