Artificial neural network-based nonlinear channel equalization: A soft-output perspective

Abstract

The artificial neural network (ANN) has been shown that, is an effect technique used to gain insight into channel equalizer design, to combat nonlinear distortion in wireless communication systems. Also, the joint design of channel equalizer and decoder can provides great advantages for system performance. However, research on the soft output of an ANN-based equalizer still remains largely open. Towards this end, this paper proposes an accurate soft information characterization for an ANN-based channel equalizer, which is crucial for the joint development of equalization and decoding. Particularly, we focus on the functional link ANN (FLANN)-based channel equalizer. By adopting the Kolmogorov-Smirnov test, we find that the error signal of a FLANN-based equalizer is not Gauss, which would pose a challenge to the calculation of the soft information. We use the mix-Gauss distribution to model the error signal, and accordingly the log-likelihood ratio (LLR) from a FLANN-based equalizer is derived. We also give insight into the mix-Gauss model that one component stands for the channel noise and another component stands for the noise caused by the equalizer, which may shed some lights on the optimization of a FLANN-based equalizer.