Abstract
With the introduction of machine learning (ML) technologies, the development of wireless access networks has gained more significant momentum, enabling WiFi and radio access networks (RAN) to provide more diverse services and better service quality. This thesis investigates ML models’ design and deployment related schemes to address the practical challenges of marching to more intelligent wireless access networks.The first part of this thesis presents a representative example of intelligent WiFi - indoor WiFi localisation. It begins with the proposed three types of novel neural network (NN) architectures for localisation, which take the standard channel sounding packets as the training data. Then, the model’s performance is extensively evaluated for continuous tracking in different indoor scenarios, and the corresponding NN transfer strategies and data requirements are investigated.
In the remainder of this thesis, the focus is directed to the emerging RAN architecture - Open RAN (O-RAN). The RAN intelligent controller (RIC) defined in the O-RAN structure enables the deployment of deep learning (DL)/reinforcement learning (RL) models in the real wireless communication system, while digital twins (DT) are regarded as a critical venue for DL/RL model developing to ensure the reliability of the developed models. Based on the above concepts, the indoor cells transmit power control challenge is solved in a federated reinforcement learning (FRL) manner under the context of O-RAN, where a well-designed hierarchical DT plays the role of the O-RAN radio link simulation, environment generation, and FRL model training. This FRL scheme improves the overall throughput and provides a template NN model for new environments.
Furthermore, for the purpose of improving the fidelity of propagation channel modelling in the DT, a two-tier likelihood reference signals received power (RSRP) prediction DL model is proposed. This prediction model is able to utilise environmental information and historical records to predict the RSRP values of the given location. It is integrated into the aforementioned DT as an add-on module. Then an RL-based intelligent handover use case is proposed to validate the proposed prediction model’s effectiveness in mitigating the simulation-to-real gap for O-RAN deployment.
Lastly, considerations are given for the potential data, environment and agent issues in the completed stage of the DL/RL model development and deployment. A series of solutions and principles are coined as RLOps, to manage the RL agent life cycles in O-RAN. A network anomaly detection case is investigated as a practice of the RLOps concept, whereas the corresponding NN model has been successfully deployed in the real-world wireless network.
| Date of Award | 20 Jun 2023 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Angela Doufexi (Supervisor), Dimitris Agrafiotis (Supervisor) & Robert J Piechocki (Supervisor) |