DRL-Driven Hybrid Beamforming Architecture for THz-Band MIMO Communication Systems

Abstract

The expansion of wireless communication systems and the rapid development of wireless communication systems have led to the transition to the use of ultra high-frequency bands, the terahertz (THz)-band communication can be regarded as one of the basic supports of 6G networks. THz communication coupled with massive multiple-input multiple- output (MIMO) systems can deliver extremely high data rates, ultra-low latency as well as high spectral efficiency. There are however major challenges associated with these systems, including complex hardware, high path loss as well as channel sparsity especially during development of hybrid beamforming architecture, a system that integrates both analog and digital precoding. This paper suggests a brand-new hybrid beamforming structure driven by deep reinforcement learning (DRL), particularly adjusted to THz-band MIMO communication systems in order to resolve those drawbacks. The aim of the project is to come up with an adaptive, scalable, and energy-efficient beamforming solution that is capable of intelligently exploring the high dimensional and dynamic wireless environment. The DRL approach (proposed method) consists of actor-critic based DRL architecture, with an agent constantly participating in the environment as a way of maximizing both analog and digital beamforming matrices. The system model considers a wideband, spatially sparse, frequency selective THz channel whose structure is built based on a high-order hybrid structure with a partially connected structure due to the hardware overhead. The DRL agent is trained through a soft actor-critic (SAC) algorithm on a set of a large number of simulations, in which to balance exploration and exploitation and to maximize a reward function related to spectral efficiency, power consumption, and signal quality. A realistic THz channel model, including beam squint as well as user mobility, supports the training process. Performance assessment reveals that the suggested DRL-centered beamforming system provides BSI throughput (up to 25Percent), channel dynamics resilience, and energy economy are substantially higher than that of the traditional practices like orthogonal matching pursuit (OMP) and codebook-based systems. The findings confirm the practicality of incorporation of deep reinforcement learning in the process of designing hybrid beamforming to provide a scalable and intelligent approach to the next-generation THz-band wireless systems. In future, the model can be extended to multi-user and multi-cell systems, include the consideration of reconfigurable intelligent surfaces (RIS), and consider the application of real-time hardware.