Dynamic AP Selection for User-Centric Cell-Free Massive MIMO Using Knowledge-Driven Deep Learning

In this letter, a knowledge-driven deep learning (DL) approach is proposed for dynamic access point (AP) selection in user-centric cell-free (CF) massive multiple-input multiple-output (MIMO). The proposed knowledge-driven DL exploits the prior knowledge for optimization framework, achieving good scalability and generalization. Specifically, each user equipment (UE) selects its served AP by an independent deep Q-network (DQN), which models the AP selection as an action-reward mapping. With the channel state information (CSI) as input, the DQN estimates the expected spectral efficiency (SE) for each potential subset of serving APs. To reduce the dimensionality of potential subset space, the DQNs consider only the dominant AP-UE associations determined by large-scale fading coefficient. The UEs dynamically select optimal serving AP subsets to maximize the overall SE. Numerical results demonstrate the effectiveness and efficiency of the knowledge-driven DL approach.