Machine Learning for Adaptive Beamforming in Mimo Systems for Improved Network Throughput and Energy Efficiency

Abstract

Adaptive beamforming in Multiple-Input Multiple-Output (MIMO) systems is now an important technology in 5G and 6G networks because of the growing need of high-speed, steady, and energy-saving wireless communications. In this paper, an adaptive beamforming framework through Reinforcement Learning (RL) is proposed which uses a Deep Q-Network (DQN) to train the best beam selection policies with respect to channel state variables, such as Signal-to-Noise Ratio (SNR) and beam index. The model is trained on the 5G Adaptive Beamforming with SNR dataset and an ε-greedy exploration-exploitation strategy and experience replay are used to guarantee a steady convergence. The results of simulation and tests prove that the RL-based model has the significant benefit compared to the traditional approaches. Through the proposed approach, a mean throughput was 6.41bps/Hz, energy efficiency 3.54bits/Joule, and SINR 22.8dB, though both the MRT and ZF performed poorly. The model also demonstrated to be reliable in a range of SNR and remained the same in terms of learning behaviour as cumulative rewards were stabilised at the 250 training episodes. These results show that convergence of reinforcement learning and adaptive beamforming can be effectively used in dynamic MIMO to enhance spectral and energy efficiency. The proposed design offers a scalable and data-driven technology to the next-generation wireless networks and can be implemented to select intelligent beams that can modify to the varying channel conditions and maximises the overall network throughput and power usage.