Predicting communications activity in the radio spectrum

This work develops a method for predicting the activity of incumbent users of the radio spectrum so that an opportunistic user can schedule communications during anticipated periods of inactivity. The method uses a high-dimensional hidden Markov model with multivariate Gaussian observations to capture the rich temporal covariance of communication activity among multiple incumbent users. Care is taken to constrain the dimensionality of the model so that it uses only the significant states of the possible state space. Prediction performance on simulated and recorded traffic is seen to be significantly improved over that of previously explored occupancy prediction models. The method not only predicts channel occupancy but also predicts which of multiple transmitters will be active and the power level they will be observed at. The relationship between prediction error rate and the entropy rate of the incumbent users´ protocol is explored and the useable prediction horizon is characterized in terms of the model´s mixing time.