Research On Deep Reinforcement Learning Based Resource Allocation For NOMA Systems

The fifth-generation wireless communication system(5G)needs to accommodate the exponential growth of wireless data traffic and support the high density of user devices.With the application of superposition coding and successive interference cancellation,the non-orthogonal multiple access(NOMA)overlays the information of multiple users on the same resource block,thereby realizing the sharing of communication resources,which improves user connection,spectrum efficiency and energy efficiency of the system,and meets the development requirements of 5G and future wireless communication systems.In the meantime,the performance of NOMA system depends critically on the resource allocation strategies.Thus,facing the application scenario of 5G,this thesis studies the deep reinforcement learning(DRL)based resource allocation algorithm for NOMA system while considering the quality of service,mobility of users and the energy efficiency(EE)of the system.The main contributions of this thesis are summarized as follows.Firstly,the problem of joint subchannel allocation and power allocation in an uplink multi-user NOMA system with EE as the objective is studied.To tackle the problem of modeling difficulty and high computation complexity that traditional optimization algorithms face when solving such joint resource allocation problem,a discrete and distributed DRL based method is proposed to reduce the computational complexity.With channel state information,this method generates the subchannel allocation policy first,then derive the allocated power of all users,and finally update its resource allocation policy based on the returned rewards to maximize the EE of the system.Simulation results show that this method can improve the EE of NOMA system while meeting user requirements.Secondly,on the basis of the first method,a continuous deep-deterministic-policy-gradient-based resource allocation method is proposed,which solves the quantization error and increased action space of the discrete method,thereby improving the performance of the power allocation part.Simulation results show that the performance of the continuous method is better than that of the discrete method.Thirdly,combining the idea of event-trigger learning,an enhanced resource allocation method is proposed based on the discrete method and the continuous method.This enhanced method monitors the channel state information and triggers the learning process only when it is needed,thereby reducing the unnecessary computation cost.Simulation results show that this enhanced method can improve the EE of the system while reducing average time cost,thereby further optimizing the resource allocation efficiency of the proposed methods.