Finite-Alphabet Signature Design for Grant-Free NOMA: A Quantized Deep Learning Approach

Abstract

Grant-free Non-Orthogonal Multiple Access (NOMA) is a promising solution to enable massive wireless access service for 5G systems and beyond. Conventional grant-free NOMA schemes directly apply the spreading signatures optimized for the grant-based scenarios, which, however, ignore the users’ diversified activation probabilities. In addition, the conventional grant-free NOMA schemes are not designed with finite-alphabet signatures, which hinders the encoder/decoder implementation using practical low-cost hardware. Therefore, to overcome these limitations, we propose a finite-alphabet signature design for the grant-free NOMA with random and nonuniform user activations. Herein, the NOMA signatures are optimized by the autoencoder-based transceivers with both transmitter and receiver being in the form of deep neural network. First, the quantized deep learning is employed in the NOMA signature training, which jointly optimizes the sequence generation and quantization. Moreover, in order to improve the training rate, we propose a specific neural network receiver, where the network structure resembles the successive interference cancellation procedures. The experiment results show that the obtained NOMA signatures commendably exploit the users’ activation profiles, and the proposed scheme outperforms the conventional ones especially when the users have unequal activation probabilities.