Neural Network-Driven Active-Passive Beamforming with Imperfect CSI for RIS-Assisted Communication System

For the reconfigurable intelligent surface (RIS)-assisted communication system, this work investigates a neural network-driven joint active-passive beamforming (APBF-NN) scheme with imperfect channel state information (CSI). To begin with, subsequently constructing a system model that includes dynamic interference, RIS reflection routes, and multi-antenna transmission with CSI estimation errors, a joint optimization problem for maximizing spectral efficiency (SE) under constant modulus constraint is formulated. Then, an APBF-NN-based architecture is developed to simultaneously optimize the active beamforming at the transmitter and the passive phase shifts at the RIS. Fully connected layers incorporate estimated channel information, interference statistics, and Signal-to-Noise Ratio (SNR) to construct interference-aware configurations with constant modulus, resulting in superior signal strength. Finally, experimental comparisons with benchmark methods, including maximum ratio transmission (MRT), zero-forcing (ZF), and random beamforming, validate the proposed scheme's superior performance in both SE and bit error rate (BER).