AI/ML for multi-technology RAN automation with full and limited infrastructure support
(English) This thesis studies and proposes solutions to some of the most relevant challenges in the Radio Access Network (RAN) management arising from its evolution beyond 5G and towards 6G. The tackled problems are selected due to the increasing inherent complexity with which these technologies com...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis doctoral |
| Fecha de publicación: | 2023 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/404655 |
| Acceso en línea: | https://hdl.handle.net/2117/404655 https://dx.doi.org/10.5821/dissertation-2117-404655 |
| Access Level: | acceso abierto |
| Palabra clave: | Àrees temàtiques de la UPC::Enginyeria de la telecomunicació |
| Sumario: | (English) This thesis studies and proposes solutions to some of the most relevant challenges in the Radio Access Network (RAN) management arising from its evolution beyond 5G and towards 6G. The tackled problems are selected due to the increasing inherent complexity with which these technologies come along, which justifies the need for Artificial Intelligence and Machine Learning (AI/ML) techniques. In particular, we identify two axes of complexity: the infrastructure support complexity axis (x-axis), where the complexity varies based on the level of infrastructure support, and the technology complexity axis (y-axis), which captures the complexity variation based on the number of technologies to be operated in a coordinated way. Based on these axes, we define three RAN scenarios: infrastructure-based single-technology, infrastructure-based multi-technology, and limited infrastructure-based single-technology. The main objective is to study these scenarios in depth and identify a set of representative use cases along these axes that can be addressed with AI/ML solutions to automate RAN. In this line, our methodology consists of a three-step approach, using existing and implementing new high-fidelity and standard-compliant simulation models of the open-source ns-3 and 5G-LENA system-level simulators coupled with the proposed AI/ML frameworks. In the first step, we focus on the two use cases, 1) handover (HO) management and 2) initial Modulation and Coding Scheme (MCS) selection in the infrastructure-based single-technology RAN scenarios. The traditional HO schemes have the drawback of considering only the quality of signals from the serving and the target BS to make a HO decision, which can impact users´ QOE. Also, the initial MCS at the start of the session is usually handled conservatively, i.e., the lowest MCS is assigned to a mobile device that connects to a new BS, impacting its initial throughput. To address these drawbacks, we propose AI/ML solutions prioritizing QoE for HO decisions and optimizing initial MCS allocation using network data. First, we design single-task AI/ML models for each use case, then propose a multi-task framework for addressing multiple use cases concurrently, reducing training costs. In the second step, we deal with the infrastructure-based multi-access technology scenarios by focusing on the coexistence of License Assisted Access (LAA) and LTE-Unlicensed (LTE-U) with WiFi. These technologies must ensure fair coexistence to operate in the unlicensed spectrum. However, this thesis discovers that the inherent delay in receiving HARQ feedback in the LAA enables it to monopolize the channel, which then degrades neighboring WiFi networks' performance. To solve this, we propose an AI/ML-based scheme that infers feedback without delay. Our scheme achieves a favorable trade-off between WiFi fairness and LAA performance in terms of throughput and latency compared to benchmark approaches. Additionally, we propose a statistical framework to evaluate fairness in LAA and LTE-U coexistence with WiFi, confirming LAA's better fairness over LTE-U's. Finally, in the third step, we focus on the limited infrastructure-based single-technology RAN scenarios. Without base stations, in these scenarios, operations like resource selection and scheduling are uncoordinated, introducing another level of complexity. Specifically, this thesis focuses on vehicle-to-vehicle communication with limited infrastructure, where a roadside unit broadcasts basic information using 3GPP NR-V2X technology. Nevertheless, vehicle resource selection in NR-V2X involves continuous channel sensing, but it consumes more energy. Alternatively, not employing sensing saves energy but increases interference. Hence, an energy-performance trade-off arises. To address this, we propose an AI/ML-based partial sensing mechanism to dynamically balance V2X user performance and energy consumption, surpassing the manual configuration of standard sensing parameters. |
|---|