Abstract
The most challenging aspect of Cognitive Radio Networks (CRNs) is designing a sensing policy which guarantees minimal interference to licensed users and as well as maximal discovery of spectral resources. The statistics of channel activity of licensed users may be unknown and/or time-variant. This requires the sensing policy to be able to adapt to these changes as needed. In this paper, we propose a learning based MAC-layer sensing protocol which acts on sensing results in the form of binary streams which are made available by the physical layer abstraction. The protocol makes use the learning algorithm we designed to learn and estimate parameters of the statistical distribution of licensed users' activities on various channels, and this in turn allows the sensing parameters to be adapted accordingly. We also use the inherent design heuristics to approximate the convergence speed of the algorithm on the estimation of utilization of licensed channels. Our simulation results reveal that the convergence of the algorithm can be predicted reasonably well if the sensing period is smaller than both the average busy and idle durations of channel dynamics. The proposed design discovers, on average, 94 percent of idle channels. Moreover, the predetermined convergence bound is more accurate as the difference between number of users in a CRN and the total number of primary channels is reduced.