A learning-based uplink massive-MIMO decoder

This work presents a deep learning based iterative network for massive-MIMO decoding. The structure of DLNet is based upon a projected gradient descent-based iterative trainable network. This DLNet is a 15-layer deep iterative network structure whose parameters are optimized using DL training for better performance on Rayleigh as well as correlated M-MIMO channels. Due to rigorous training on time-varying channels, DLNet can work for time-varying channels with single-time training. Simulation shows that the proposed DLNet decoder performs better than other MIMO decoding techniques by at least 2 dB in symbol-error-rate (SER), at least 11 times faster than the baseline (OAMPNet), and 9 times less complex. It also converges fast compared to other available M-MIMO decoders.