This letter presents a novel approach for spectrum sensing in cognitive satellite-terrestrial networks. The approach uses multi-agent deep reinforcement learning (DRL) and reconfigurable intelligent surface to address the problem of under-utilization of terrestrial network spectrum in remote areas. Unlike previous studies that rely only on current sensing data, this approach utilizes historical data to improve spectrum detection accuracy and post-decision state to accelerate agent learning speed. Simulation results show that it outperforms existing DRL methods in terms of faster agent learning convergence and more effective detection of primary network spectrum occupancy.
Read full Publication