Deep Learning-Aided NOMA Codebook Design with Improved Performance

Nonorthogonal multiple access (NOMA) techniques have drawn significant interest in recent years as the demand for high data rate and spectral efficiency increases. Among the NOMA designs, sparse code multiple access (SCMA) has been shown to achieve outstanding performance. How-ever, the SCMA performance highly depends on codebook construction, and it is difficult to construct codebooks that are optimal for different application scenarios in a hand-crafted manner. To address this issue, some solutions aided by deep learning were proposed to automate the codebook design. One of the solutions is to construct the codebook and decoder with the autoencoder structure. This paper further investigates this structure for codebook construction by proposing ideas to improve the performance of the constructed codebook and reduce the learning complexity. In addition to improving the performance by modifying different aspects of the training setting, we also show that the deep neural network (DNN) based encoder can be simplified to only one linear layer without sacrificing the performance. This result suggests that the DNN structure does not well exploit the potential of a general (possibly nonlinear) encoder, and a better structure may be needed to suit the NOMA application.