Understand the predictability of wireless spectrum: A large-scale empirical study

To solve the scarcity of wireless spectrum, Cognitive Radio (CR) is proposed to let unlicensed wireless users (secondary users) dynamically find and access unused channels without interference to licensed users (primary users). The performance of the CR based Dynamic Spectrum Access (DSA) mechanism can be dramatically improved if the wireless spectrum is predictable, and many works has been done based on this assumption. To understand the predictability of realworld wireless spectrum, we make a large scale empirical study in this paper. The study is based on the spectrum data collected in a metro city in 7 days, ranging from 20MHz to 3GHz. Our study includes the analysis of kth-order Markov universal predictability, the experiment of kth-order Markov on-line predictor, and finally the seeking for specialized predictor for wireless spectrum. We find that 1) it's not efficient to improve prediction by increase Markov order, because on nearly half channels kth-order (k > 1) Markov methods make no improvement at all, and for the rest channels, 1st-order Markov method makes the largest improvement and higher orders make little further improvement; 2) We also find that a sliding window method can improve accuracy considerably meanwhile reduce complexity of prediction model significantly.