A Mixture Density Channel Model for Deep Learning-Based Wireless Physical Layer Design

摘要: Machine learning is a highly promising tool to design the physical layer of wireless communication systems, but it usually requires that channel model known. As data rates increase and transceivers become more complex, channel, hardware imperfections, their interactions difficult compensate explicitly. New machine schemes for do not require an explicit implicitly learn end-to-end link including characteristics non-linearities system directly from training data. In this paper, we present novel neural network architecture provides stochastic model, by parameters Gaussian mixture distribution real samples. We use in conjunction with autoencoder suitable modulation scheme. Since our learns can transfer adapt quickly changes environment. apply millimeter wave communications its challenges phased arrays large number antennas, high carrier frequencies, wide bandwidth complex characteristics. experimentally validate using 60 GHz FPGA-based testbed show able reproduce good accuracy.