NMAP-Net: Deep-Learning-Aided Near-Field Multibeamforming Design and Antenna Position Optimization for XL-MIMO Communications

Extremely large-scale multiple-input-multiple-output (XL-MIMO) is a candidate technology for 6G wireless networks and massive Internet of Things (IoT) communications. In this article, we consider an XL-MIMO system operating in the near-field communication range, where a base station equipped with multiple movable (i.e., adjustable-position) antennas serves multiple desired users in the presence of multiple undesired users. In this system, we investigate a new joint problem of multibeamforming design and antenna position optimization to maximize the minimum beamforming gain for the desired users with a constraint on the maximum interference leakage to the undesired users. To effectively and intelligently solve this challenging nonconvex problem, we propose a novel DL model, called NMAP-Net, which is composed of three main learnable modules, namely, DL blocks I-III, for feature extraction, antenna position optimization, and multibeamforming design, respectively. A novel training strategy for the proposed NMAP-Net is also devised in an elegant manner using a customized loss function, called adaptive loss function, to maximize the minimum beamforming gain while adaptively suppressing the maximum interference leakage. Furthermore, an effective inference mechanism for the proposed NMAP-Net is developed based on a Gaussian randomization technique to ensure the feasibility of the predicted solution. Extensive simulation results substantiate that the proposed NMAP-Net performs markedly better and more effective than the existing techniques while achieving almost the same performance as its upper limit.