Massive MIMO Demodulation Aided by NN

Abstract

In this work, we propose a demodulator aided by a neural network (NN) for massive multiple-input multiple-output (M-MIMO) systems. In particular, we consider the uplink (UL) phase of an M-MIMO system in which users transmit utilizing a quadrature amplitude modulation (QAM) and soft-estimates are obtained via the application of the zero-forcing (ZF) combiner. Based on the ZF soft-estimates, we propose suitable features that are used in the input layer of an NN, whose task is to learn how to output hard-estimates, that is, demodulate the ZF soft-estimates. We then adopt a supervised learning perspective by performing a regression analysis and training the NN with simulated data. The performance and complexity of our NN-aided demodulator is numerically compared to those of the hard-decisor (HD) scheme used as a benchmark for a 4-QAM. Through this comparison, we show that our NN-aided demodulator is 17.3% more computationally efficient with tolerable performance losses. We argue that demodulators assisted by NNs can be a promising alternative to cheaply demodulate high-order OAMs.