Autoencoder based Robust Transceivers for Fading Channels using Deep Neural Networks

Abstract

In this paper, we design transceivers for fading channels using autoencoders and deep neural networks (DNN). Specifically, we consider the problem of finding (n, k) block codes such that the codewords are maximally separated in terms of their Hamming distance using autoencoders. We design an encoder and robust decoder for these block codes using DNNs. Towards this, we propose a novel training methodology for the DNN that attempts to maximize the minimum Hamming distance between codewords. We propose a loss function for this training which has stable weight updates during back propagation compared to other loss functions reported in the literature. The block codes learned using the proposed methodology are found to achieve the maximal Hamming distance separation that is known in theory. We also propose two different receiver architectures based on fully connected deep neural network (FCDNN) and bidirectional recurrent neural network (BRNN) that are suited for complex fading channels. The proposed DNN based receiver is shown to achieve significantly better error performance when compared to their classical counterparts in the presence of channel model mismatches. In the presence of model mismatches such as imperfect channel knowledge and noise correlation, the proposed DNN based transceiver is shown to offer increased reliability and robustness than the conventional transceiver.