Application of Deep Learning to Fairness-Based Power Allocation for 5G NOMA System with Imperfect SIC

Abstract

Non-orthogonal multiple access (NOMA) has been a promising technique for 5G communication system, which has higher spectrum efficiency, energy efficiency and fairness than that of orthogonal multiple access (OMA). NOMA serves more than one user by sharing the same time-frequency resource block and uses successive interference cancellation (SIC) to separate multiuser signal. However, the error propagation in the SIC procedure, which is called the imperfect SIC, can cause a severe performance loss. In this paper, we apply the deep learning to power allocation in order to mitigate the impact of imperfect SIC under fairness perspective for downlink NOMA system. Firstly, we formulate an optimization problem aiming to maximize the minimum user rate to provide fairness for all users. Secondly, exhaustive search method is used to solve the formulated problem and thus the optimal power allocation factor is obtained. Lastly, we train the deep neural network to predict the obtained power allocation factor. The simulation results show that our proposed scheme provides the performance close to that provided by exhaustive search. Furthermore, the proposed scheme has much lower complexity than the exhaustive search scheme.