Loss Function Analysis for DNN Iterative Decoder on Rayleigh Fading Massive Random Access

This work considers user activity detection and estimates the fading coefficient of massive random access on a Rayleigh block-fading AWGN channel. Only a fraction of them is active among a huge number of potential users on the massive random network at each time slot. The active user transmits a unique vector to the base station to identify and estimate the fading coefficient. We consider an iterative decoder for user activity and channel estimation and use two deep neural networks to implement them. A generalized loss function with loss coefficients is proposed for the DNN-based iterative decoder. The trend of the loss coefficients is analyzed from the point of view of NMSE performance, training convergence, and overfitting. The simulation results show that loss coefficients that satisfied the geometric progression with a common ratio of 0.5 can obtain better NMSE performance in all SNR regimes and reduce the overfitting problem in higher SNR.