Cognitive Radios (CRs) tackle the spectrum scarcity problem by allowing unlicensed access on already-licensed spectrum. Where primary user (PU) radio systems hold a privileged medium access, secondary users (SUs) try to utilize and use spare time-frequency resources to establish communication. At their core, CRs need to sense, detect and predict the medium access pattern of PUs in order to facilitate communication. A promising approach for inferring these predictions is the use of Machine Learning techniques and in particular Artifical Neural Networks (ANNs). ANNs try to learn a mapping from a set of inputs to a specific desired output and can therefore be directly applied to the problem of PU activity prediction. Especially Recurrent Neural Networks (RNNs) show adequate performance for noisy measurement data and prolonged training periods. In this work, the applicability of ANN-based channel state prediction is examined via a case study on the upcoming aeronautical communication technology L-band Digital Aeronautical Communications System (LDACS). LDACS is envisioned to reuse spectral resources currently primarily allocated to other aeronautical systems such as the Distance Measuring Equipment (DME) system. Primary and secondary users are modeled with respect to preliminary LDACS specifications and the performance of the proposed algorithms is evaluated via simulation. We show that RNNs can function as prediction agents for SU LDACS medium access and that stringent reliability and interference requirements can be met. A supervised learning problem, together with an incremental learning strategy is proposed to address time-varying PU channel access patterns. Finally, a brief discussion on the use of predictions in a distributed scheduling approach is given.
