Machine Learning Applied to an RF Communication Channel

In this paper we introduce a model of a node-to-node communication channel as an auto-encoder neural network where the inputs and outputs represent the transmitted and received bits respectively. With some simple assumptions, the encoding layer of the auto-encoder can be interpreted as a phase-modulated radio frequency (RF) signal to be transmitted across the communication channel. When the auto-encoder is trained to minimize bit errors between the transmitter and receiver nodes, it effectively learns the best modulation scheme for that particular communication channel. We show that an auto-encoder trained with Additive White Gaussian Noise (AWGN) has promising bit error rates compared to traditional modulation schemes in a simulated environment. Implementing the auto-encoder generated modulation scheme in GNURadio, we send auto-encoder modulated messages between two Universal Software Radio Peripherals (USRPs) and experimentally calculate the corresponding bit error rate. Preliminary results show bit error rates that are comparable to those of traditional modulation schemes measured in similar experimental conditions. Finally, we develop the methodology for incorporating the GNURadio flowcharts with the auto-encoder training program so that the auto-encoder can be trained on the actual communication channel itself and discuss how that can improve the auto-encoder results.