Real-Time Machine Learning Methods for Two-Way End-to-End Wireless Communication Systems

Abstract

In this article, we study a data-driven real-time machine learning method for end-to-end wireless systems for the Internet of Things (IoT). For a two-way communication link between two IoT devices, we propose an efficient learning algorithm that can train the autoencoder-based transmitter and receiver in each device without needing to know the channel between two devices. To this end, we adopt the conditional generative adversarial network (cGAN) that can learn an output distribution of the channel for a given conditioning signal. Our proposed training method consists of the link update stage and the self-update stage. In the link update stage, two devices transmit the training data and update their own receiver and the cGAN simultaneously. Subsequently, in the self-update stage, the two devices train their transmitters at the same time for the given receivers and cGANs. Our proposed real-time training method is updated without the knowledge of the channel models nor information feedback for training. Finally, we demonstrate that the proposed training method achieves significant performance gains over conventional schemes in various practical communication scenarios.