This thesis proposes a deep learning-based demapper that utilizes feedforward neural networks to learn the complex mapping functions required for multiuser detection in non-orthogonal multiple access (NOMA) system. By utilizing neural networks, the proposed deep learning-based demapper eliminates the need for the system to check each constellation point individually, hence decreasing the computational complexity of the demapping process, while maintaining a good bit-error rate (BER) performances.
This thesis developed a deep learning-based demapper, trained using a dataset generated with iterative spatial demapping (ISM), to process the received signals from a two-user NOMA scheme. This thesis analyzes two NOMA scenarios: (i) an uncoded scheme that utilizes binary phase-shift keying (BPSK) modulation, and (ii) a coded scheme that employs repetition coding and interleaver to improve transmission reliability. The proposed demapper trained on essential features such as the superposition received sig