Rate-splitting multiple access for 6G and beyond: a machine learning and model-based approach
File(s)
Author(s)
Cerna Loli, Rafael
Type
Thesis
Abstract
The research community is working towards the standardization of 6G mobile communications, and rate-splitting multiple access (RSMA) has emerged as a promising interference management framework to achieve its strict service requirements. However, most RSMA works in the literature have studied its performance only under idealistic assumptions. Therefore, this thesis focuses on addressing three practical challenges of RSMA for 6G and beyond.
First, a model-based deep learning (MBDL) receiver is designed by merging the low-complexity structure of the classical successive interference cancellation (SIC) receiver with the pure data-driven power of deep neural networks (DNNs) to bypass the use of any channel estimation at the receiver (CSIR). The MBDL receiver performance is assessed using link-level simulations (LLS), and it is demonstrated that it outperforms the SIC receiver with imperfect CSIR, and approaches the performance of the SIC receiver with perfect CSIR by increasing the training overhead.
Next, a specialized hybrid automatic repeat request (HARQ) for RSMA is designed to inherently exploit the message splitting and combining operations of RSMA to prioritize the scheduling of retransmissions through the common stream using a layered HARQ (L-HARQ) approach. Through LLS, it is demonstrated that, in the high SNR regime, the devised HARQ scheme achieves better throughput, reliability, and latency than RSMA with conventional HARQ, and RSMA using an adaptive modulation and coding (AMC) algorithm.
Finally, the design of a meta-learning based optimization framework for large-scale wireless systems is introduced. Motivated by the prohibitive complexity of conventional optimization algorithms, the proposed solution exploits the training phase of a DNN to achieve non-convex optimization at significantly reduced complexity. The optimization framework is then used to successfully optimize the performance and reveal unknown aspects of the operation in the large-scale regime of three different applications: hierarchical RSMA, integrated sensing and communication (ISAC), and beyond-diagonal reconfigurable intelligent surfaces (BD-RIS).
First, a model-based deep learning (MBDL) receiver is designed by merging the low-complexity structure of the classical successive interference cancellation (SIC) receiver with the pure data-driven power of deep neural networks (DNNs) to bypass the use of any channel estimation at the receiver (CSIR). The MBDL receiver performance is assessed using link-level simulations (LLS), and it is demonstrated that it outperforms the SIC receiver with imperfect CSIR, and approaches the performance of the SIC receiver with perfect CSIR by increasing the training overhead.
Next, a specialized hybrid automatic repeat request (HARQ) for RSMA is designed to inherently exploit the message splitting and combining operations of RSMA to prioritize the scheduling of retransmissions through the common stream using a layered HARQ (L-HARQ) approach. Through LLS, it is demonstrated that, in the high SNR regime, the devised HARQ scheme achieves better throughput, reliability, and latency than RSMA with conventional HARQ, and RSMA using an adaptive modulation and coding (AMC) algorithm.
Finally, the design of a meta-learning based optimization framework for large-scale wireless systems is introduced. Motivated by the prohibitive complexity of conventional optimization algorithms, the proposed solution exploits the training phase of a DNN to achieve non-convex optimization at significantly reduced complexity. The optimization framework is then used to successfully optimize the performance and reveal unknown aspects of the operation in the large-scale regime of three different applications: hierarchical RSMA, integrated sensing and communication (ISAC), and beyond-diagonal reconfigurable intelligent surfaces (BD-RIS).
Version
Open Access
Date Issued
2024-04-15
Date Awarded
01/07/2024
License URL
Advisor
Clerckx, Bruno
Sponsor
Defence Science and Technology Laboratory (Great Britain)
Publisher Department
Electrical and Electronic Engineering
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)