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

Abstract

Machine learning is a highly promising tool to design the physicallayer of wireless communication systems, but it usually requiresthat a channel model is known. As data rates increase and wirelesstransceivers become more complex, the wireless channel, hard-ware imperfections, and their interactions become more difficult tomodel and compensate explicitly. New machine learning schemesfor the physical layer do not require an explicit model butimplic-itly learnthe end-to-end link including channel characteristics andnon-linearities of the system directly from the training data.In this paper, we present a novel neural network architecturethat provides anexplicitstochastic channel model, by learning theparameters of a Gaussian mixture distribution from real channelsamples. We use this channel model in conjunction with an au-toencoder for physical layer design to learn a suitable modulationscheme. Since our system learns an explicit model for the channel,we can use transfer learning to adapt more quickly to changes inthe environment. We apply our model to millimeter wave commu-nications with its challenges of phased arrays with a large numberof antennas, high carrier frequencies, wide bandwidth and complexchannel characteristics. We experimentally validate the systemusing a 60 GHz FPGA-based testbed and show that it is able toreproduce the channel characteristics with good accuracy.