Hybrid-Driven Dynamic Neural Network for Adaptive User-Activity Detection in Massive Random Access

Grant-free random access (GF-RA) has recently emerged to support massive random access. Due to the absence of access grant in GF-RA, the base station (BS) has to first identify each active user. However, the state-of-the-art user-activity detection (UAD) solution, i.e., covariance-based maximum-likelihood detection (CB-MLD) is still subject to some critical deficiencies. Specifically, the update step size in each CB-MLD iteration relies on an asymptotically large antenna number, which may cause convergence issues in practice. In addition, the hard-decision threshold for UAD remains to be fine-tuned in complicated scenarios. Both deficiencies are hard to address via analytical methods. Thus, we propose a hybrid-driven UAD network (HyD-UADNet), where a model-driven network is constructed to learn the proper update step size, and a data-driven network is designed to learn the soft decision on user activity. Furthermore, we construct a dynamic configuration-adaptive mixture-of-expert network (CA-MoENet). This CA-MoENet can adaptively produce weighting coefficients for different expert HyD-UADNets, so as to enhance the UAD robustness against varying configurations. Finally, simulations show the superior UAD accuracy of the HyD-UADNet, and reveal the robustness of the CA-MoENet even if the testing configuration is never seen by any expert during training.