End-to-End Learning of Communication Systems with Novel Data-Efficient IIR Channel Identification

In this paper, we introduce a novel end-to-end deep learning procedure for communication systems, which is data-efficient and capable of dealing with infinite memory length of communication channels. Therefore, as opposed to recent works, we utilize a low-complexity algorithm to identify the communication channel. The channel model is obtained purely from data and its output is differentiable with respect to its input, which is a basic requirement for gradient-based optimization of the auto-encoder neural network. We study the performance of the algorithm for a variety of challenging channels from different domains of communication engineering showing the broad applicability of the proposed approach.