Research On User Grouping And Power Allocation Algorithm Of NOMA System Based On DRL

By stacking multiple users,non-orthogonal multiple access(NOMA)significantly improves the system sum rate,where user grouping and power allocation are the key factors determining the system sum rate.Traditional iterative methods have high complexity;the method based on deep reinforcement learning can reduce the online execution time,but there is a problem that the uniform sampling of the sample pool makes the low efficiency;in addition,the residual error generated by successive interference cancellation at the receiver imposes further constraints on power allocation.In order to settle the above three problems,this paper proposed a NOMA user grouping and power allocation method based on priority sampling and a power distribution based on power difference threshold.In order to solve the problems of difficulty in obtaining training labels,high complexity and low sampling efficiency in traditional algorithms,NOMA user grouping and power allocation method based on priority sampling is proposed,in which DQN network is adopted to complete the user grouping and DDPG network is used for power allocation.In this method,the channel state information is taken as the input,the system sum rate as the goal,and the TD-error represents the priority of the sample.It is shown that the DDPG used in this algorithm can deal with the matter of performance loss caused by quantization error,and the system sum rate is improved by 5%for the uniform sampling-based DQN+DDPG compared with the multi-DQN algorithm;based on this,the algorithm in this paper sets priority to samples and samples valuable samples first,compared with the uniform sampling-based DQN+DDPG algorithm,the learning rate is increased,and the system sum rate is improved by 3%.The deep reinforcement learning network continuously interacts with the NOMA system,which can dynamically update the optimal learning strategy when the channel changes.Compared with the optimal exhaustive search algorithm and iterative power optimization algorithm,the system sum rate reach 93%and 94%of them respectively,the running time is reduced by 74%and80%,respectively.For the existing literature in power allocation,there is no analysis of the relationship between successive interference cancellation residual error and SNR,this paper investigates the effect of SNR on residual interference error,gives a cubic polynomial fitting expression for residual error versus SNR by data statistics,and applies it to the residual interference term setting of the power allocation optimization objective function.In addition,this paper proposed a power allocation program with a power difference threshold,which is mainly to settle the problem of serious residual interference caused by similar user power values of different users,in the algorithm,first determining a reasonable power difference threshold through simulation,and then adjusts the obtained optimal power based on the threshold to increase the power difference among users.As a result,the algorithm effectively reduces the residual interference error and improves the system sum data rate.Moreover,in the case of 2 users,because the proposed method reduces the interference between each other by increasing the power difference,make it easy to distinguish detection users by successive interference elimination,and increases the detection accuracy,thereby improving the BER at the near user side,when the BER is10^-4,the algorithm in this paper can provide a 1.8d B SNR gain compared to the original method without a power difference threshold.However,the proposed algorithm readjusts the power to the user based on the optimal power,which makes the system sum rate decrease slightly under low SNR.