Fast Beamforming for IRS Assisted Multi-User Communication Systems by Lightweight Unsupervised Learning

Iterative optimization algorithms excel in beamforming design but usually suffer from high computational latency. In this paper, we present a lightweight unsupervised learning model for fast beamforming in intelligent reflecting surface (IRS) assisted multi-user communication systems, considering the often-overlooked phase shift noise (PSN) from hardware impairments. To effectively address the PSN, we adopt the expectation of the sum data rate as our objective function and derive its closed-form expression. This derivation facilitates the design of a loss function suitable for unsupervised learning networks. We then propose a lightweight cascade convolutional neural network (LCCNN) for fast beamforming, featuring fewer trainable parameters for accelerated training and efficient implementation. Numerical results demonstrate that our framework achieves comparable data rates to parameter-intensive networks while significantly reducing computational latency compared to iterative algorithms.