A Multi-waveform Radio Receiver, an Example of Machine Learning Enabled Radio Architecture and Design

Abstract

We describe a machine learning enabled architecture and design for a multi-waveform radio receiver in the pursuit of a truly cognitive radio with more functionality and adaptability than current software defined radio implementations. This machine learning approach brings closer to reality the vision of cognitive radios proposed by Joseph Mitola III and Gerald Q. Maguire, Jr. Cognitive radios make decisions about their communications regime about where (in spectrum), and how (waveform parameters) to transmit and receive information [1]. And, such radios should be able to self-optimize their communications to most efficiently maximize data capacity in power and spectrum constrained environments. To achieve these goals, the software in the radio must control more of the functionality, including functions in the physical layer. Building on Tim O'Shea's and Jakob Hoydis' ideas [2], we have developed a generalized architecture, in which the physical layer functions: Frequency correction, timing correction, demodulation, and bit error correction are performed by an artificial neural network capable of processing several signal types and waveforms.