Machine learning for channel coding in low-resolution OFDM systems

This report investigates the quantization effects of low-resolution analog-to-digital converters in OFDM systems for channel coding. Channel codes require soft-bit information in order to perform effectively, but quantization inhibits the ability to derive optimal closed form expressions for these soft-bits. To overcome this issue, a novel deep learning architecture is proposed to perform joint OFDM detection and channel decoding. The joint network is comprised of a log-likelihood ratio detection network and a neural belief propagation network. The LLR detection network is shown to improve performance with respect to bit error rate and weighted mean square error under certain quantization conditions, compared to the the naive LLR computation. In this report, I provide relevant background information, describe existing machine learning methods for LLR estimation and channel decoding, and perform experiments on the proposed joint architectureElectrical and Computer Engineerin