CLASSIFICATION STYLE REGRESSION FOR SPECTRAL OPENING PMF ESTIMATION

Abstract

Dynamic spectrum allocation (DSA) permits unlicensed users to access spectrum owned by a licensed user given they do so without interference to the primary user. To avoid interference with other users, the unlicensed user needs to be aware of channel availability. Spectrum sensing allows a radio to find spectrum holes, but costs energy and time. Predictive methods can be used to decrease the amount of spectrum sensing needed to find an available channel. We designed a novel neural network architecture for spectrum hole prediction. This neural network is capable of creating probability mass functions (PMF) estimates of the length of channel openings with no assumptions of the initial probability distribution or prior knowledge about the traffic. This architecture is shown to work through a mathematical proof, and its performance is measured through simulation.