End-to-End Optimization of Multiuser MIMO Systems Through Autoencoder with Channel Estimation by Recurrent Network
Machine Learning, Multiple-Input Multiple-Output (MIMO), Signal Detection, Channel Estimation.
The spectral efficiency gains introduced by multiuser MIMO systems render them attractive schemes for the next generations of mobile communication networks. Due to the intrinsic complexity of the mathematical models of these systems under realistic conditions and the interdependence between the processing steps of the transmitters and receivers, machine learning is an option that allows the complete system to be designed by training a noisy autoencoder. This paper presents a proposed neural network architecture for end-to-end optimization of a multiuser MIMO system. The performance of the system, measured in terms of symbol error rate, was compared using different constraints on the energy of the constellations and using sequential and joint detection. Simulations were performed using a Rayleigh fading channel model, as well as using the realistic 3GPP TR 36.873 model, and a M-PSK modulation scheme with zero-forcing was used as a reference for comparison. An LSTM-based channel estimator was implemented to capture the temporal correlations of the realistic channel with a pilot transmission rate of only 10%. The results reveal that joint detection performs slightly better than sequential detection in the presence of estimation errors and that relaxing the constellation constraints can improve performance by 2 dB without any changes to the average transmitted power.