Research On Slice Resource Allocation Technology Based On Cognitive Wireless Network

With the deepening of the integration of industrialization and informatization,the contradiction between radio supply and demand has become increasingly prominent.There are two main reasons for the lack of spectrum resource.First,the new communication services increase rapidly with the arrival of the fifth-generation mobile communication,and the business covers a wider range of fields.Second,the utilization rate of authorized spectrum for power grid,water conservancy and other industries is low.According to statistics,the utilization rate of authorized spectrum is less than 20%.Therefore,it is necessary to use cognitive radio technology and NOMA(Non Othogonal Multiple Access,NOMA)technology to improve the utilization of spectrum resource.At the same time,the network slice is divided and the network function is virtualized according to different types of user services.In this way,the slice resource can be dynamically and efficiently configured according to the demand,which can not only reduce the cost for the user,but also improve the communication performance.In the complex cognitive wireless network resource allocation model,the environment is intricate,and the established problems are often non-convex and cannot be solved analytically.However,reinforcement learning can make the agent explore and learn independently in the dynamic environment,so as to generate better strategies and allocate resource reasonably.This paper proposes a variety of deep reinforcement learning algorithms for different resource allocation scenarios,so as to allocate resources reasonably and reduce network complexity.The main contributions of this paper are as follows:(1)Dynamic resource allocation algorithm of cognitive wireless network based on DQN(Deep Q-Learning)The primary users are divided into three types according to their interference tolerance ability.The objective is to maximize the spectrum efficiency of the secondary users of the system.Therefore,a dynamic resource allocation algorithm for cognitive wireless network based on DQN is proposed.All secondary users are regarded as an agent.The action,state and reward function of reinforcement learning are designed,and the mapping relationship with cognitive wireless network is accurately established.The deep neural network model is designed according to the simulation scene,and the reward standardization and regularization technology are introduced.Many experiments are carried out on hyper parameter selection and network design process.The model converges fast,which proves the superiority of the algorithm.(2)Slice resource allocation algorithm of cognitive wireless network based on Actor-CriticConsidering that the service requirements of primary users in cognitive scenarios are mainly high-speed and low-latency,all primary users are divided into two types of slice users,namely eMBB(Enhanced Mobile Broadband)and URLLC(Ultra-reliable and Low Latency Communications),and the goal of maximizing throughput of secondary users is proposed.Therefore,a cognitive wireless network slice resource allocation algorithm based on Actor-Critic is proposed,which can deal with continuous action space,and considers the maximum long-term return when the secondary user selects the channel,power and other actions.Experiments show that the algorithm can achieve better resource allocation performance in multi-service scenarios.(3)Multi-agent reinforcement learning algorithm based on cognitive NOMA network slice resource allocation scenarioIn order to further improve the utilization of spectrum resource,the technology of NOMA is introduced,so that multiple users can carry out power superposition to transmit signals in the same channel.Aiming at reducing the complexity of NOMA user serial interference elimination,the optimization target not only considers the throughput of all secondary users of the current system,but also sets the limit of the maximum number of overlapped users in NOMA.To solve this problem,a multi-agent reinforcement learning algorithm is proposed by combining graph convolutional neural network and DQN algorithm,which is suitable for cognitive NOMA network slice resource allocation scenarios.The experimental results show that the algorithm can improve the performance of the model.