Neural-Network-Assisted Multi-Dimensional Constellation Design for Downlink NOMA Systems

Using the neural network (NN) as a means to solve model-based optimization problem, we propose the finite-alphabet constellation design which aims at optimizing the sumrate performance of downlink power-domain NOMA system when the base station is equipped with one (SISO) or multiple antennas (MISO). While traditional constellations such as QPSK or QAM are deemed suboptimal in NOMA, finding the optimal constellations using numerical searches becomes prohibitively complex, especially when the constellation size or the number of transmit antenna increases. We propose that, with the aid of NN, arbitrary multi-dimensional constellation improving the information-theoretic weighted sum-rate, which takes the form of a loss function in the NN, can be generated. This departs from the existing AI approaches mainly focusing on symbol-level performance and relying on end-to-end training. We consider both perfect CSI and imperfect CSI scenarios, where in the latter case the channel estimation errors can be accounted for by proper NN training. Simulations results show that the proposed NN-assisted approach can be effectively used to produce multidimensional constellation for MISO-NOMA system which outperforms the existing approaches.