Minimum Bayesian Risk Based Robust Spectrum Prediction in the Presence of Sensing Errors

Abstract

Spectrum prediction is known as an effective technique complementary to spectrum sensing, which infers the spectrum evolution from historical spectrum data especially the historical sensing data. However, the prediction in the presence of sensing errors lead to the deterioration of prediction accuracy. To address this issue, we present a minimum Bayesian risk-based robust spectrum prediction scheme (MBR-RSP) in this paper. We prove that the spectrum prediction output follows normal distribution through distribution fitting test, and formulate the problem of robust spectrum prediction in the presence of sensing errors as a binary hypothesis testing. On the basis of these, we develop MBR-RSP, in which the optimal threshold is proved to minimize the Bayesian risk of threshold detection. The experimental results both on simulated data and measured data show that the prediction accuracy of MBR-RSP outperforms that of neural network-based spectrum prediction in the presence of sensing errors. Moreover, in prediction-based dynamic spectrum access, the secondary user with the proposed MBR-RSP also shows the significant improvement in spectrum efficiency and the interference to primary users.