Training of Deep Joint Transmitter-Receiver Optimized Communication System without Auxiliary Tools

Abstract

Deep Joint transmitter-receiver optimized communication system (Deep JTROCS) is a new physical layer communication system. It integrates the functions of various signal processing blocks into deep neural networks in the transmitter and receiver. Therefore, Deep JTROCS can approach the optimal state at the system level by the joint training of these neural networks. However, due to the non-differentiable feature of the channel, the back-propagation of Deep JTROCS training gradients is hindered which hinders the training of the neural networks in the transmitter. Although researchers have proposed methods to train transmitters using auxiliary tools such as channel models or feedback links, these tools are not available in many real-world communication scenarios, limiting the application of Deep JTROCS. In this paper, we propose a new method to use undertrained Deep JTROCS to transmit the training signals and use these signals to reconstruct the training gradient of the neural networks in the transmitter, thus avoiding the use of an additional reliable link. The experimental results show that the proposed method outperforms the additional link-based approach in different tasks and channels. In addition, experiments conducted on real wireless channels validate the practical feasibility of the method.