ADMM-RMCBF-Net: A neural network decision for distributed robust multi-cell beamforming

Abstract

In this paper, to construct a promising deep-learning architecture for the distributed robust multi-cell beamforming (RMCBF) decision, we reconsider the existing typical power-minimization problem. By thoroughly unfolding the conventional algorithm of the Alternating Direction Method of Multipliers (ADMM), we establish a novel neural network for fast distributed RMCBF design, namely ADMM-RMCBF-Net. The most important step in unfolding lies in that we explicitly solve the key semi-definite programming sub-problems by resorting to the primal–dual inter-point method instead of the encapsulated convex solvers. It is worth stressing that all parameters of the ADMM algorithm can be learned from end-to-end data-driven training in the proposed ADMM-RMCBF-Net, rather than being predetermined empirically in the conventional ADMM method. Simulation results confirm the advantages of the proposed deep-learning approach over its conventional optimization-based counterpart in terms of these distributed RMCBF decisions’ performance. Specifically, in addition to accelerating the convergence of the distributed RMCBF design, more importantly, the proposed ADMM-RMCBF-Net could adapt to the practical propagation environment quickly through small-sample learning.