Deep Learning-Based Beamforming Optimization for Intelligent Reflecting Surfaces in 6G Networks

Abstract

The implementation of Intelligent Reflecting Surfaces (IRS) on sixth-generation (6G) wireless networks is a paradigm-shaping method to improve the network coverage, energy efficiency, and spectral efficiency. IRS technology allows the active reprogramming of wireless in-air propagation by programmable passive structures that regulate the phase of incident signals. Nevertheless, mixed optimization of active beamforming (at base station) and passive beamforming (at IRS) is a prohibitively expensive task to be completed because of its non-convexity and high dimension. The paper suggests a system based on deep learning to enhance beamforming in 6G systems with IRS. We propose a supervised deep neural network (DNN) to input channel state information (CSI) and output near-optimal beamforming vectors and IRS phase shifts. Noise-like data are generated with conventional optimization algorithms and the model is trained to adapt to real-time channel situations (fine-tuning). Simulation analyses demonstrate that the proposed scheme is more spectral efficient, energy efficient and computationally fast than other conventional optimization schemes e.g. Alarating Optimization (AO), Semidefinite Relaxation (SDR). Our System is a low-latency, scalable System to which we can apply in real-time applications in a dynamic 6G environment. The work explained the feasibility of artificial intelligence and combination with reconfigurable wireless hardware addressing the growing performance requirements of the next-generation wireless networks.