Deep learning-aided constellation design for downlink NOMA

Massive connectivity is one of the most challenging issues for Internet of Things (IoT) to achieve the quality of service provisions required by the numerous IoT devices. Non-orthogonal multiple access (NOMA) technology, where multiple users multiplex on the same radio resources, is a promising candidate for next generation wireless networks (the 5th Generation, 5G) and has been expected to meet the requirements of high spectral efficiency and massive connections of 5G mobile communication systems. However, conventional downlink NOMA simply superimposes several single-user constellations, which does not consider the interactions between multiple data streams. This paper proposes a novel deep learning-aided downlink NOMA scheme by parameterizing the bit-to-symbol mapping and multi-user detection with deep neural networks (DNN). The network is trained in an end-to-end fashion with synthetic data, and then the trained bit-to-symbol mapping is extracted to derive the multi-user constellation for downlink NOMA. Simulation results demonstrate that, with the proposed constellations, our scheme achieves significantly lower symbol error rate than conventional downlink NOMA.