A Graph Convolutional Network Approach for AP Selection in User-Centric Cell-Free Massive MIMO

This paper studies the access point (AP) selection problem in user-centric cell-free (CF) massive multiple-input multiple-output (MIMO). AP selection is an essential issue in user-centric CF massive MIMO, it determines the scalability and per-user spectral efficiency(SE) in the network. We propose a graph convolutional network (GCN) to address the AP selection problem. The large-scale fading coefficients are described in a graphic data structure, which is used as the input of the proposed GCN. To avoid the labeled training samples generated by a computationally complicated algorithm, an unsupervised learning framework is used for training the GCN. More specifically, the average SE of the all users is denoted as the output and the training of the GCN is to maximize the average SE by the gradient descent algorithm. The numerical results , in a specific scenario with i.i.d. Rayleigh channel model, show that the proposed GCN approach achieves a slightly higher per-user SE than the benchmark iterative algorithm.