LAUSR

As a promising solution for massive machine-type communication, grant-free non-orthogonal multiple access (GF-NOMA) has received considerable attention in recent years. However, the multidimensional constellation design (MCD) and multiuser detection (MUD) in GF-NOMA are usually optimized in a divide and conquer way, leading to local optima and performance degradation. To address this issue, we investigate the joint optimization of MCD and MUD for GF-NOMA. The formulated joint optimization is based on variational inference, which is intractable due to the signal superimposition that makes the optimization variables intricately coupled. Then, we resort to end-to-end deep learning (DL) to obtain the optimal solution. Specifically, we propose a DL-based multi-task variational autoencoder (Mul-VAE) that adopts a variational autoencoder network to optimize the distribution of the constellation points. We further derive the loss function of the proposed network and analyze it from an information-theoretic perspective. On this basis, multi-task learning is employed to deal with mutually conflicting yet related detection processes. Besides, taking heterogeneous transmission rates of users into account, a multi-task prioritizing strategy is designed to balance training performance. Simulation results reveal that the proposed method enables significant gains compared to state-of-the-art techniques.