Abstract: Radio environment map(REM) can assist cognitive users to accurately perceive and utilize spectrum holes,achieve interference coordination between network nodes,and improve the spectrum efficiency and robustness of wireless networks.However,when cognitive users utilize and share REM,there are problems of high computational complexity and high distribution delay overhead,which limit cognitive users' ability to perceive spatial spectrum situation in real time.To solve this problem,this paper proposes a reinforcement learning-based REM construction and distribution method in mobile edge intelligence networks.First,we employ a low-complexity construction technique that combines kriging interpolation and super-resolution for REM construction.Second,we model the computational offload strategy selection problem during REM construction and distribution as a mixed-integer nonlinear programming problem by using edge computing.Finally,we combine artificial intelligence technology and edge computing technology,and propose a centralized training,distributed execution reinforcement learning framework to learn REM construction and distribution strategies in different network scenarios.Simulation results show that the proposed method has good adaptability,and it can effectively reduce the energy consumption and delay of REM construction and distribution,and support the near real-time application of REM by cognitive users in mobile edge network scenarios.

Key words: Radio environment map, Edge intelligence, Computation migration, Reinforcement learning