| With the rapid development of mobile Internet,the increasing number of users in the network and the diversification of services put forward higher requirements for the scheduling and allocation of limited resources in the network system.As for the scheduling technology of base station,how to improve the overall performance of the system through better scheduling algorithm has become the key research content of the mobile technology of the fifth generation mobile communication.Because MIMO technology makes full use of spatial multiplexing gain and can effectively improve system performance,it has been widely used.Therefore,the performance evaluation of user scheduling algorithms in 5G network systems using MIMO technology is the main content of this paper.At present,the traditional multi-user scheduling algorithm needs to calculate inter-user interference,which involves a lot of matrix operations,and the scheduling time increases sharply with the increase of the number of users.In this paper,the Deep Determining Policy Gradient(DDPG)algorithm is applied to the user scheduling.The scheduling user is selected directly according to the algorithm,which greatly reduces the user selection time.Firstly,this paper investigated the existing downlink user scheduling algorithms of MIMO system,compared the system throughput,user fairness and scheduling time under singleuser MIMO and multi-user MIMO scenarios,realized different user scheduling algorithms through the NS3 simulation platform,and analyzed the performance differences of each scheduling algorithm.Due to the lack of flexibility in scheduling of traditional scheduling algorithms,reinforcement learning is applied to user scheduling,NS3-AI framework is used to connect the NS3 network simulation environment and deep reinforcement learning algorithm,and user-related data generated from the simulation environment is transmitted to the machine algorithm side for training.Compared with ideal data,it has more reference value for practical application.The experimental results show that compared with the traditional MU-MIMO algorithm,the scheduling consumption time is greatly reduced.In addition,this paper improves the NS3-AI tool on the basis of the original one.A new NS3 simulation environment can be created by adding the Subprosess module during the running process.The improved NS3-AI can complete the co-simulation in a terminal window,simplifying the operation process.Considering that user scheduling needs to be completed within one scheduling cycle in actual network communication,this paper proposes a real-time DDPG scheduling algorithm,which can complete user selection within one scheduling cycle,and greatly reduces the time of algorithm training and decision selection of users.Based on the deep deterministic strategy gradient algorithm,the real-time DDPG algorithm modified the target network parameter update mode,trained the network model offline,and selected users through inference online.When the loss function tends to be stable,the evaluation network parameters will not be updated,which reduces the training and parameter updating time by 56% compared with the original algorithm.After the training,the average time of the model inference selection user is 0.25 ms,which can meet the actual network time requirements.In this paper,the performance of different traditional scheduling algorithms is compared,and reinforcement learning is applied to network scheduling,which reduces the time of user selection while ensuring system performance,and is of great significance to user scheduling in real network. |
