Learning the MMSE channel predictor

Abstract

In this work a feed-forward neural network-based channel predictor is derived, where assumptions on a physical wave propagation channel model in a fading scenario are incorporated into the design procedure of the predictor. We start with the general expression of an approximated minimum mean squared error (MMSE) predictor and derive a predictor having the structure of a feed-forward neural network by making two key assumptions. By properly training this neural network it is possible to compensate the approximation errors due to these assumptions. It is further possible to outperform the linear MMSE (LMMSE) predictor with perfect knowledge of the statistical moments of second order based on the covariance function for specific channel model assumptions, especially for low SNR values.