Cooperative Beamforming With Nonlinear Power Amplifiers: A Deep Learning Approach for Distributed Networks

Abstract

Network densification is essential to enhance the areal capacity of future wireless networks, for which the user-centric network is a promising solution. In the perspective of network scalability, user association and cooperative beamforming (BF) among the base stations (BSs) should be designed elaborately to manage strong interference from nearby BSs. However, practical systems with hardware impairments including nonlinearity of power amplifiers may disturb these techniques and make the network difficult to be scalable. The nonlinearity of power amplifiers affects not only radio frequency (RF)-level signal but also performance according to the number of user equipments (UEs), resulting in necessity of re-designing network-level operations including BS-UE association. This paper proposes a deep learning-based cooperative BF framework for distributed networks with nonlinear power amplifiers. To optimize the cooperative BF for complicated nonlinear systems, we adopt unsupervised learning approach with constraints. To reduce the communication overhead of the network, we propose a novel neural network structure, from which the BSs can perform the cooperative BF in distributed manner. In particular, the information exchange between the central unit and the local BSs is reduced by designing neural network so that the local channel state information are utilized only in the local BSs while the central unit utilizes only covariance matrices. We show that the proposed scheme achieves a higher effective sum rate compared to the baseline schemes by adjusting the user association, BF, and power allocation to control interference and nonlinear distortion with reduced communication overhead.