A Deep Reinforcement Learning based Analog Beamforming Approach in Downlink MISO Systems

Analog beamforming with low-resolution phase shifters is a key technique for 5G networks due to its superior hardware complexity and power consumption advantages. However, the optimal beamforming coordination is an extremely challenging issue in a downlink multi-antenna base station and single-antenna user equipment scenario. To avoid using global channel state information and reduce the communication overhead, in this paper, we propose a deep reinforcement learning based distributed analog beamforming approach to improve the energy efficiency for a downlink multiple-input and single-output (MISO) system. Specifically, each base station trains a neural network to steer its beamformer by phase shifters according to its local and obtained neighbouring information, with the purpose of maximizing its own energy efficiency and minimizing the negative impacts to its neighbouring cells. We evaluate the performance of the proposed approach by simulations, and validate its superiority by comparing with baseline schemes.