Using ANNs to predict the evolution of spectrum occupancy in cognitive-radio systems.

This paper analyzes the use of Artificial Neural Networks (ANNs) to identify and predict the evolution of vacant portions or frequency holes of the radio spectrum in Cognitive Radio (CR) systems. The operating frequency of CR transceivers can be modified over the air according to the information provided by the ANN in order to establish the communication in the least occupied band. To this end, ANNs are trained with time-series datasets sensed from the electromagnetic environment. Several network architectures are considered in the study, including Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks and hybrid combinations of them. These ANNs are modeled and compared in terms of their complexity, speed and accuracy of the prediction. Both simulations and experimental results are shown to validate the approach presented in this work.