Deep reinforcement learning for hybrid beamforming in multi-user millimeter wave wireless systems

This paper proposes a Machine Learning (ML) algorithm for hybrid beamforming in millimeter-wave wireless systems with multiple users. The time-varying nature of the wireless channels is taken into account when training the ML agent, which identifies the most convenient hybrid beamforming matrix with the aid of an algorithm that keeps the amount of signaling information low, avoids sudden changes in the analog beamformers radiation patterns when scheduling different users (flashlight interference), and simplifies the hybrid beamformer update decisions by adjusting the phases of specific analog beamforming vectors. The proposed hybrid beamforming algorithm relies on Deep Reinforcement Learning (DRL), which represents a practical approach to embed the online adaptation feature of the hybrid beamforming matrix into the channel states of continuous nature in which the multiuser MIMO system can be. Achievable data rate curves are used to analyze performance results, which validate the advantages of DRL algorithms with respect to solutions relying on conventional/deterministic optimization tools.