Predictive Spectrum Sensing Strategy Based on Reinforcement Learning

In this paper, we consider a cognitive radio (CR) system with a single secondary user (SU) and multiple licensed channels. The SU requests a fixed number of licensed channels and must sense the licensed channels one by one before transmission. By leveraging prediction based on correlation between the licensed channels, we propose a novel spectrum sensing strategy, to decide which channel is the best choice to sense in order to reduce the sensing time overhead and further improve the SU's achievable throughput. Since the correlation coefficients between the licensed channels cannot be exactly known in advance, the spectrum sensing strategy is designed based on the model-free reinforcement learning (RL). The experimental results show that the proposed spectrum sensing strategy based on reinforcement learning converges and outperforms random sensing strategy in terms of long-term statistics.