Research On 5G Downlink Scheduling Algorithm Based Reinforcement Learning

The fifth-generation mobile communication system puts forward higher demands on transmission rate,user mobility,and the number of terminals.Downlink wireless resource scheduling is an important part of the mobile communication field,and the traditional resource scheduling method has been difficult to adapt to new technologies and scenarios.At present,the commonly used proportional fairness scheduling algorithm lacks flexibility in scheduling,and it is difficult to ensure short-term fairness in the face of some changing scenarios.With the development of multiple input multiple output systems towards massive MIMO,user selection algorithms need to reduce the impact of inter-user interference on system performance.The traditional scheduling algorithm based on the model has been difficult to adapt to complex and changeable network scenarios.Reinforcement learning does not require external environment models,and has achieved good results in many complex decision-making problems.In this context,this thesis studies the 5G-oriented downlink radio resource scheduling algorithm,which aims to combine reinforcement learning methods,enable the downlink scheduling algorithm to be combined with new scenarios to improve system throughput and fairness performance.First,based on the application requirements and characteristics of downlink resource scheduling and reinforcement learning,this thesis designs a general framework for the application of reinforcement learning to downlink resource scheduling.It is composed of a user manager,a controller based on reinforcement learning and an action executor.For different scenarios,a suitable reinforcement learning algorithm can be selected based on this framework to solve the resource scheduling problem.Secondly,for the single-user MIMO scenario,this thesis analyzes the common proportional fairness algorithm and the flexibility and poor convergence of several improved algorithms in this scenario,and proposes an improved algorithm based on Deep Deterministic Policy Gradients(DDPG).According to the state of the system,the value of the parameter in the generalized proportional fair algorithm is obtained,and then the scheduling weight is dynamically adjusted.Simulation results show that the algorithm is more effective than proportional fair algorithm to improve throughput and fairness,in which throughput can be increased by 8%,fairness can be increased by 16%,and short-term fairness can be guaranteed Compared with the improved algorithm based on Q-Learning and Actor-Critic,the algorithm is easier to converge,and the convergence speed is nearly doubled.Finally,for multi-user MIMO scenarios,since multiple users occupy the same resource together,interference between users will affect system performance.Inter-user interference needs to be considered in resource scheduling,and a better user set is selected on a resource to achieve the goal of maximizing system throughput.In response to this problem,this thesis proposes a user selection algorithm based on improved DDPG,combined with a deep separable convolutional neural network,to obtain the user set on all resources according to the system state.Theory and simulation show that compared with the CUS algorithm based on user correlation,this algorithm has less computational complexity,the number of users increases,and the complexity only increases linearly,the total throughput of the system can be increased by a maximum of 6% on the premise of ensuring fairness.Wireless resource scheduling is one of the core technologies of communication systems,and scheduling algorithms have always been the focus of research.In this thesis,reinforcement learning is applied to wireless resource scheduling,which improves system performance and provides ideas for improving resource scheduling algorithms.