MIMO-NOMA-DAE: A Deep Learning based Downlink MIMO-NOMA Scheme for Low-Power Applications with Imperfect CSI

Non-orthogonal multiple access (NOMA) has been posed as one of the future drivers for the massive Internet of Things and machine-to-machine communication. However, these applications typically operate with a low signal-to-noise ratio (SNR) due to the limited device power and long transmit distances. Conventional NOMA design approaches based on linear analytical methods fail to provide optimal results for these situations. This paper explores deep learning-based solutions to the end-to-end optimisation of downlink NOMA combined with multiple-input multiple-output (MIMO). Of interest to this work is the BER performance of MIMO-NOMA at low SNR ranges, given imperfect channel state information(CSI). MIMO - NOMA - DAE is proposed, which consists of parallel denoising auto-encoders trained at negative SNR ranges without power allocation coefficients or CSI knowledge. It is shown that MIMO-NOMA-DAE’s operation is non-linear and outperforms the BER performance of traditional NOMA and OMA by a factor of 20. Furthermore, the same performance for all modulation has been obtained which orders up to order 256.