[Paper Review] Deep Learning-Based Rate-Splitting Multiple Access for Reconfigurable Intelligent Surface-Aided Tera-Hertz Massive MIMO

Minghui Wu, Z. Gao+4 more|arXiv (Cornell University)|Sep 18, 2022

Advanced Wireless Communication Technologies4 citations

TL;DR

This paper proposes a deep learning (DL)-based rate-splitting multiple access (RSMA) scheme for reconfigurable intelligent surface (RIS)-aided terahertz massive MIMO systems, integrating a hybrid data-model-driven precoding framework with a Transformer-based CSI acquisition network (CAN). The approach achieves robust spectral efficiency under imperfect channel state information (CSI) by leveraging DL for low-overhead CSI estimation and precoding, outperforming conventional methods in accuracy and complexity.

ABSTRACT

Reconfigurable intelligent surface (RIS) can significantly enhance the service coverage of Tera-Hertz massive multiple-input multiple-output (MIMO) communication systems. However, obtaining accurate high-dimensional channel state information (CSI) with limited pilot and feedback signaling overhead is challenging, severely degrading the performance of conventional spatial division multiple access. To improve the robustness against CSI imperfection, this paper proposes a deep learning (DL)-based rate-splitting multiple access (RSMA) scheme for RIS-aided Tera-Hertz multi-user MIMO systems. Specifically, we first propose a hybrid data-model driven DL-based RSMA precoding scheme, including the passive precoding at the RIS as well as the analog active precoding and the RSMA digital active precoding at the base station (BS). To realize the passive precoding at the RIS, we propose a Transformer-based data-driven RIS reflecting network (RRN). As for the analog active precoding at the BS, we propose a match-filter based analog precoding scheme considering that the BS and RIS adopt the LoS-MIMO antenna array architecture. As for the RSMA digital active precoding at the BS, we propose a low-complexity approximate weighted minimum mean square error (AWMMSE) digital precoding scheme. Furthermore, for better precoding performance as well as lower computational complexity, a model-driven deep unfolding active precoding network (DFAPN) is also designed by combining the proposed AWMMSE scheme with DL. Then, to acquire accurate CSI at the BS for the investigated RSMA precoding scheme to achieve higher spectral efficiency, we propose a CSI acquisition network (CAN) with low pilot and feedback signaling overhead, where the downlink pilot transmission, CSI feedback at the user equipments (UEs), and CSI reconstruction at the BS are modeled as an end-to-end neural network based on Transformer.

Motivation & Objective

• Address the challenge of acquiring accurate high-dimensional CSI in terahertz massive MIMO systems with limited pilot and feedback overhead.
• Overcome the performance degradation of conventional space-division multiple access (SDMA) due to CSI imperfection in terahertz bands.
• Design a robust precoding scheme that maintains high spectral efficiency despite imperfect CSI in RIS-aided terahertz MIMO systems.
• Develop a low-complexity, end-to-end trainable CSI acquisition network (CAN) using deep learning to reduce signaling overhead.
• Integrate model-driven and data-driven deep learning techniques to optimize RIS passive precoding, base station analog, and digital RSMA precoding.

Proposed method

• Proposes a hybrid data-model-driven RSMA precoding framework combining data-driven RIS reflecting network (RRN) for passive RIS precoding, model-based match-filter (MF) analog precoding at the base station (BS), and model-driven deep unfolding active precoding network (DFAPN) for digital RSMA precoding.
• Designs a Transformer-based RRN to learn optimal RIS phase shifts for passive beamforming, trained end-to-end with negative average weighted sum rate (ARWU) as loss.
• Introduces a low-complexity approximate weighted minimum mean square error (AWMMSE) algorithm for digital precoding, later unfolded into a deep learning network (DFAPN) for improved performance and reduced complexity.
• Develops a CSI acquisition network (CAN) using a Transformer-based architecture to model downlink pilot transmission, CSI feedback, and CSI reconstruction as an end-to-end neural network.
• Employs negative ARWU as the loss function for training RRN and DFAPN, and normalized mean square error (NMSE) for CAN training to enhance spectral efficiency and estimation accuracy.
• Uses a hybrid training strategy combining offline pre-training and online fine-tuning to optimize the entire system for robustness and low signaling overhead.

Experimental results

Research questions

• RQ1 How can deep learning be leveraged to enable robust rate-splitting multiple access (RSMA) in RIS-aided terahertz massive MIMO systems under imperfect CSI?
• RQ2 What is the optimal design for a low-overhead, high-accuracy CSI acquisition framework in high-dimensional terahertz MIMO systems with RIS?
• RQ3 How does the integration of model-driven and data-driven deep learning improve the performance and complexity trade-off in RIS-aided RSMA precoding?
• RQ4 To what extent can a Transformer-based neural network outperform conventional channel estimation and feedback methods in terms of accuracy and signaling overhead?
• RQ5 What is the performance gain of the proposed hybrid DL-based RSMA scheme in spectral efficiency and robustness compared to conventional SDMA and RSMA schemes under CSI imperfection?

Key findings

• The proposed RRN and DFAPN achieve significantly better average weighted sum rate (ARWU) performance than conventional schemes, especially under imperfect CSI conditions.
• The proposed CAN reduces CSI estimation error by achieving lower normalized mean square error (NMSE) than conventional methods, even with low pilot and feedback overhead.
• The running time of the proposed DL-based CAN, RRN, and DFAPN is 3.24–5.63 ms on GPU and 15.64–25.39 ms on CPU, which is significantly faster than model-based schemes like AWMMSE (70.14 s on CPU).
• The computational complexity of the proposed DL-based schemes is $ frac{1}{10}$ to $ frac{1}{100}$ of model-based counterparts, with CAN/RRN/DFAPN complexity at $ frac{1}{100}$ of AWMMSE.
• The proposed AWMMSE algorithm achieves better spectral efficiency than conventional methods but is computationally expensive (70.14 s on CPU), motivating the need for the DFAPN to reduce complexity.
• Numerical results confirm that the proposed end-to-end DL-based framework enables high spectral efficiency with low pilot and feedback overhead, while maintaining robustness to CSI imperfection.