Channel Charting-based Radio Resource Management

5th Generation (5G) cellular networks are designed to deliver unparalleled performance in mobile environments with three promises: i) increased capacity, ii) ultra-reliable and low-latency connections, and iii) a massive number of connected devices. Achieving these ambitious goals necessitates the integration of novel technologies into networks. Millimeter-wave (mmWave) communications stand out as a key enabler for achieving ultrahigh data rates and low latency, leveraging the substantial bandwidth available at these high frequencies. Beamforming techniques have been extensively employed in the mm-wave bands to alleviate the path loss of mm- wave radio links. However, several challenges must be overcome, primarily associated with the high overhead of finding suitable beams. This thesis addresses key challenges in beam management for 5G and mmWave communication systems through the application of Channel Charting (CC) and Machine Learning (ML) techniques. CC is a self-supervised method that maps the collected high dimensional Channel State Information (CSI) at a Base Station (BS) into a low dimensional space which represents pseudo positions of User Equipment (UEs) in the radio environment. The low dimensional space preserves the local geometry of the UEs meaning that nearby UEs in real space are close to each other on the CC. A CC-based framework is designed where in an offline training phase, CCs are constructed and annotated with Signal-to-Noise Ratio (SNR)s of neighboring cells/beams. ML algorithms are used to predict the SNR of a user at neighboring cells/beams from its transmission in a massive Multiple Input Multiple Output (mMIMO) cellular system. By predicting the signal quality of neighboring stations without UE assistance, the protocol overhead for handover decisions can be reduced. Both standalone and non-standalone 5G system deployments are considered and the best beam prediction is investigated. Beam tracking based on CC is investigated and results show that at a very low beam-search overhead one can leverage a CCto-SNR mapping in order to track strong beams between the UEs and the BS. As the fundamental building block of the framework proposed in this thesis, CC necessitates enhancements in its construction to enable versatile applications across different scenarios. To address this, a CSI feature has been devised aimed at mitigating the influence of small-scale fading. This improvement empowers the framework to yield robust predictions even with low spatial sampling density. Additionally, a low complexity Out-of-Sample (OOS) algorithm has been developed, which boasts reduced computational requirements compared to conventional OoS algorithms, making it a more efficient choice for practical implementations.