Towards Transparent Deep Learning: Neural Precoder Design for Downlink RSMA

The non-convexity of rate-splitting precoder design precludes the direct use of efficient convex optimization algorithms. Instead, successive convex approximation (SCA)-based methods have emerged as a promising approach for precoder design in rate-splitting multiple access (RSMA) systems. Although S...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on communications p. 1
Main Authors: Wang, Chao, Li, Guo, Li, Zan, Zhang, Haibin, Jin, Liang
Format: Journal Article
Language:English
Published: IEEE 2025
Subjects:
alternating direction method of multipliers
Approximation algorithms
Computational complexity
Computational modeling
Covariance matrices
Deep learning
deep unfolding
Downlink
Iterative algorithms
Mathematical models
Optimization
precoder design
rate splitting multiple access
Spectral efficiency
ISSN:0090-6778, 1558-0857
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The non-convexity of rate-splitting precoder design precludes the direct use of efficient convex optimization algorithms. Instead, successive convex approximation (SCA)-based methods have emerged as a promising approach for precoder design in rate-splitting multiple access (RSMA) systems. Although SCA-based algorithms deliver satisfactory performance, their lengthy optimization process and high computational complexity-due to repeatedly solving approximate problems-hinder real-time implementation. To address this challenge, we propose an alternating direction method of multipliers (ADMM)-induced deep learning network (AIDLN)-based precoder design for multiuser downlink communications, leveraging the strengths of both model-based optimization and data-driven deep learning. Specifically, we first develop an iterative precoding algorithm based on ADMM, where each subproblem admits a closed-form solution. Next, we unfold the iterations of this ADMM-based algorithm into the layers of a neural network, constructing a deep-unfolding architecture. To accelerate convergence, we introduce trainable parameters alongside the ADMM hyperparameters. Numerical results demonstrate that our AIDLN-based precoder achieves superior performance compared to traditional model-based optimization while maintaining low computational complexity, making it suitable for real-time applications.
ISSN:0090-6778
1558-0857
DOI:10.1109/TCOMM.2025.3626023