Machine Learning based Cognitive Communications in White as Well as the Gray Space

This paper describes new ideas and results on machine learning based cognitive communications in White as well as the Gray space. We combine the concepts of signal processing, communications, pattern classification and machine learning to make a dynamic use of the spectrum such that the emanated signals do not interfere with the existing ones. Unlike other programs such as the neXt Generation (XG) communications program of the Defense Advanced Research Projects Agency (DARPA), where radio scene analysis is carried out to find the spectrum holes also known as the White space, we make use of the White as well as the Gray space for non-interfering signal transmission. Our assumption is that a learning module will facilitate adaptation in the signal classification process, so that the presence of new types of waveforms can be detected, features that best facilitate classification of the previously and newly identified signals can be determined, and waveforms can be generated by using the basis-set orthogonal to the ones present in the environment. Incremental learning and prediction allow knowledge enhancement as more snap-shots of data are processed, resulting in improved decisions. Use of non-competitive policy set results in zero interference with the already existing signals with a modest increase in the White and Gray space utilization. On the other hand competitive policy set utilizes machine learning to predict the future behavior of the signals which results in more than 90% utilization of spectrum at an expense of some interference due to errors in prediction.