Deep Learning-based Demapper for Multiuser Detection

This paper proposes a deep learning (DL)-based demapper for multiuser detection of non-orthogonal multiple access (NOMA) scheme. The DL-based demapper uses feedforward neural networks (FNN) to learn complex mapping functions to reduce the computational complexity of the current demapper by eliminating the need to check each constellation point individually. To keep the complexity low, we design a DL-based demapper using two hidden layers with 8 and 4 neurons. The hyperbolic tangent (tanh) function is used as an activation function in the hidden layers to map real-valued inputs to a range between –1 and +1. Key parameters such as superposition received signals, channel, and noise variance are used to train the DL-based demapper for both uncoded and coded NOMA schemes. We found that two hidden layers with 8 and 4 neuron structures are enough to achieve lower bit-error rates (BER) compared to the traditional iterative spatial demapping (ISM), particularly in the uncoded scheme. These findings suggest that deep learning has significant potential to efficiently manage interference and handle complex demapping tasks, laying a foundation for future enhancements in NOMA systems.