Deep Learning for a Fair Distance-based SCMA Detector

As a scheme of non-orthogonal multiple access in the code-domain, sparse code multiple access (SCMA) is one of the promoting candidate for the upcoming generations of wireless communication systems, and it has been actively investigated in recent years with several challenges in designing low complexity and high accuracy decoding algorithms. Deep learning technologies are of significant potential in solving several problems of communication systems. Motivated by this, we explore new approaches to improve SCMA detection performance using deep learning methods. In this paper, we propose to jointly design and train a denoising auto-encoder (DAE) and deep neural network (DNN) to decode SCMA signals over an additive white Gaussian noise channel. Simulation results show that our proposed DAE-DNN detector outperforms existing deep learning ones. However, the performance of the above-mentioned method is slightly worse than that of the traditional message passing algorithm (MPA). Nevertheless, this comparison is not fair, since the DAE-DNN detector, contrary to MPA, assumes that the SCMA codebook is unknown at the receiver. That is why, we propose a new distance-based DNN detector under the assumption that the codebook is known. The proposed detector can be fairly compared to MPA, and simulations confirmed that its performance is better than that of MPA.