Abstract : In cognitive radio (CR) networks, fast and accurate spectrumsensing plays a fundamental role in achieving high spectralefficiency. In this paper, a reinforcement learning (RL) enabled cooperativespectrum sensing scheme is proposed for the secondaryusers (SUs) to determine the scanning order of channels and selectthe partner for cooperative spectrum sensing. By applying Qlearningapproach, each SU learns the occupancy pattern of theprimary channels thus forming a dynamic scanning preference list,so as to reduce the scanning overhead and access delay. To improvethe detection efficiency in dynamic environment, a discounted upperconfidence bound (D-UCB) based cooperation partner selectionalgorithm is devised wherein each SU learns the time varying detectionprobability of its neighbors, and selects the one with the potentiallyhighest detection probability as the cooperation partner.Simulation results demonstrate that the proposed cooperative spectrumsensing scheme achieves significant performance gain overvarious reference algorithms in terms of scanning overhead, accessdelay, and detection efficiency.
Index terms : Cooperative sensing, multi-armed bandit, Qlearning, reinforcement learning, spectrum sensing