Machine Learning-based Reconfigurable Intelligent Surface-aided MIMO Systems

TL;DR

This paper introduces LPSNet, a low-complexity neural network approach for passive beamforming in RIS-assisted MIMO systems, achieving near-optimal spectral efficiency with significantly reduced computational cost.

Contribution

It presents a novel unsupervised neural network model, LPSNet, designed specifically for efficient phase shift optimization in RIS-aided MIMO systems, outperforming traditional methods in complexity.

Findings

LPSNet achieves 97.25% of AO's spectral efficiency.

LPSNet reduces computational complexity by over 95%.

The method is effective for 16x2 MIMO systems with 40 RIS elements.

Abstract

Reconfigurable intelligent surface (RIS) technology has recently emerged as a spectral- and cost-efficient approach for wireless communications systems. However, existing hand-engineered schemes for passive beamforming design and optimization of RIS, such as the alternating optimization (AO) approaches, require a high computational complexity, especially for multiple-input-multiple-output (MIMO) systems. To overcome this challenge, we propose a low-complexity unsupervised learning scheme, referred to as learning-phase-shift neural network (LPSNet), to efficiently find the solution to the spectral efficiency maximization problem in RIS-aided MIMO systems. In particular, the proposed LPSNet has an optimized input structure and requires a small number of layers and nodes to produce efficient phase shifts for the RIS. Simulation results for a 16x2 MIMO system assisted by an RIS with 40 elements show that the LPSNet achieves 97.25% of the SE provided by the AO counterpart with more than a 95% reduction in complexity.