Deep Learning Based Modified Message Passing Algorithm for Sparse Code Multiple Access

Abstract

Shuffled message passing algorithm (SMPA) is a serial variant of message passing algorithm (MPA) for sparse code multiple access (SCMA) signal detection, which accelerates the convergence rate of MPA. However, SMPA still achieves the near-optimal performance due to the effect of cycles in the factor graph. In the paper, we propose to optimize the weights assigned to the edges of the factor graph by unfolding SMPA as layers of deep neural network (DNN). We consider the weights as network parameters and then train the network offline to obtain weights which can minimize the loss function. With simulations, we show that DNN based SMPA (DNN-SMPA) outperforms SMPA in terms of bit-error-rate (BER) for the same level of computational complexity.