Stochastic Channel Modeling for Deep Neural Network-aided Sparse Code Multiple Access Communications

Sparse code multiple access (SCMA) has excellent application prospects due to its high spectral efficiency and accsess capacity. However, due to the nonorthogonal characteristic of SCMA in code domain, the codebook needs to be manually designed for all communication scenarios, and the receiver has high computational complexity. To address this issue, deep neural network-aided SCMA (DNN-SCMA) is proposed, but it is difficult to capture channel state information (CSI) in the dynamic and time-varying communication scenarios, which hinders the overall learning and optimization fot end-to-end communications. This paper proposes a stochastic channel model with conditional generative adversarial network (CGAN) for DNN-aided SCMA in a data-driven way. Particularly, a model-free learning method is adopted to accurately learn different types of random channel models, which realizes effective acquisition of dynamic channel information. Finally, the end-to-end training is achieved through the use of back propagation (BP), and then by an iterative training of the composed networks, the end-to-end loss can be optimized in a supervised manner. Results show the feasibility of CGAN-based channel modeling in end-to-end DNN-SCMA.