Research On Model And Algorithm Of Spectrum Resource Sharing Based On Deep Reinforcement Learning In Cognitive Wireless Network

With the rapid development of wireless communication technology,radio spectrum resources become increasingly scarce,and the fixed allocation of spectrum resources can no longer meet the actual needs of today.The spectrum sharing technology in cognitive wireless network provides a solution model for this problem.However,the traditional cognitive spectrum sharing technology needs to obtain the priori information of the environment,which is mainly applicable to the relatively simple and predictable radio environment,but it is not satisfactory in the actual complex wireless environment.Deep reinforcement learning can process complex environmental information and autonomously learn the optimal actions,which provides a great possibility for performance improvement and practical application of cognitive spectrum sharing technology.Therefore,this thesis discusses a dynamic spectrum sharing solution based on deep reinforcement learning in cognitive wireless network.Firstly,in view of the problem of spectrum resource shortage,this thesis considers a spectrum sharing scenario in a cognitive wireless network without cooperation between primary user and secondary user,the primary users and the secondary user update their power according to their own power policies,the goal of the secondary user is to learn the optimal power control policy according to the obtained information,to successfully share the spectrum resources of the primary users for communication,and to ensure the normal communication of the primary users.Then,this thesis carries on the mathematical modeling to the scene,quantifies the model and analyzes the variable and formulas in the model in detail.Aiming at this model,a spectrum sharing scheme based on Proximal Policy Optimization algorithm is proposed.Different from the deep reinforcement learning algorithm based on value function,this deep reinforcement learning algorithm based on policy can not only process the complex and continuous environmental information,but also effectively process the continuous action space.Experimental results show that the proposed algorithm has good performance under different environmental parameters,can help secondary user learn effective continuous power control policy,and shows speed advantage compared with Deep Q Network algorithm.Finally,in order to further improve the training speed of PPO algorithm,a Distributed PPO algorithm based on multi-threading is proposed.The experimental results show that the secondary user trained by this method can perform well in different parameter settings,and it is found in the comparative experiment with the PPO algorithm that the distributed PPO algorithm can train faster under the condition of achieving the same effect.