An Efficient Architecture Search for Scalable Beamforming Design in Cell-Free Systems

Scalable beamforming design (SBD) is an effective way to improve the computation efficiency of downlink communications. To this end, this paper proposes a scalable unsupervised network (SUNet) for SBD in cell-free systems. According to the analysis of beamforming characteristics, we first derive two architectural conditions for the SUNet. Then we confirm that the proposed SUNet achieves SBD within the improved computation efficiency as long as these two architectural conditions are met. With respect to the SUNet, an effective architecture search (EAS) algorithm is proposed. By applying a subpopulation selection strategy, the EAS algorithm automatically searches the desired architectures to fulfill the two architectural conditions of the SUNet, which avoids the labor-intensive task of manually setting the architectures. Furthermore, when the EAS algorithm calculates the individual fitness value, a parameter mapping strategy is proposed to map the parameters of the trained SUNet to the targeted SUNet as initial parameters, which reduces the number of iterations in the training process of the SUNet. Experimental results show that the SUNet reaches 95% of the upper bound of the scalable sum rate, where the sizes of areas as well as the number of base stations (BSs) and users are 5 times of these in training.