Comparison of Statistical Signal Processing and Machine Learning Algorithms for Spectrum Sensing

In a cognitive radio network, what is the trade off between statistical signal processing and machine learning algorithms that perform the same task? When should a system use the former and when should it use the latter? In this paper we present an empirical comparison study of different techniques for two tasks: detecting multiple transmitters in the same time-frequency domain, and automatic modulation classification. We develop and improve upon an unsupervised learning technique for the former, based on the log-Rayleigh distribution. Results are based on data generated from GNU Radio Companion, and implementations of these algorithms in software. They show that there is a tradeoff between accuracy and computation/implementation complexity signal processing has a several orders of magnitude advantage over machine learning, but slightly lower accuracy. Thus, there is a need for an overarching framework that can meld machine learning and statistical signal processing.