Passive Beamforming Design for Double RIS-muMIMO System: A Machine Learning Perspective

Abstract

Multiple-input multiple-output (MIMO) and reconfigurable intelligent surfaces (RIS) enable network system designers to create intelligent, energy-efficient network systems. Higher-order beamforming gain through the deployment of multiple RISs ensures enhanced system performance. To achieve this, a cooperative design of passive beamforming is required. This paper presents RIS passive beamforming design for a double RIS-assisted multi-user multiple-input multiple-output (muMIMO) system. The authors explore machine learning (ML)-based techniques, specifically fully complex-multilayer perception (FC-MLP) and extreme learning machine (ELM), that do not require any prior knowledge of the mathematical formulations of the network system, for RIS passive beamforming design frameworks. The proposed ML-based frameworks are compared to the alternating optimization (AO) as well as to the deep deterministic policy gradient (DDPG)-based techniques in terms of average spectral efficiency (SE), as a function of transmit power, their configurations, RISs’ locations and number of RISs’ cells, as well as their computational complexities. The proposed ELM framework outperforms the AO and FC-MLP methods, and demonstrates descent performance when compared to the DDPG method.

Keywords

• Double intelligent surface
• extreme learning machine
• machine learning
• multilayer perceptron
• multi-user MIMO
• passive beamforming