MISO-NOMA Uplink Resource Optimization Based On Machine Learning

Nowadays,the fifth-generation mobile communication network(5G)has been vastly deployed all over the world.Many countries have also launched some pilot research on the sixth-generation mobile communication network(6G).With the development of wireless networks,the frequency resources become more and more limited,and can hardly support the massive demand from users and service providers.In the meantime,the energy consumption of the wireless networks increases dramatically with the increased data rate.Therefore,improving spectral efficiency and reducing energy consumption become the critical issues for the wireless networking technology.As the cornerstone of 5G,large-scale multiple-input multiple-output(MIMO)technology can achieve higher energy and spectrum efficiency and enable reliable signal transmission.The non-orthogonal multiple access(NOMA)technology,on the other hand,can support larger connections in the larger-scale wireless network.Therefore,this thesis researches the problem of MISO-NOMA uplink resource optimization using machine learning technique,and proposes a low-complexity sub-optimal power control scheme and an optimal power control scheme combining user clustering and power minimization in order to increase the number of accesses and reduce the energy consumption.Firstly,this thesis proposes a sub-optimal uplink power control algorithm based on K-means.In MISO-NOMA system,joint optimization of clustering and power control becomes too complex to solve.In this regard,this thesis proposes a low-complexity twostep power control scheme,which decouples the overall optimization problem into two sub-problems of user clustering and power minimization,respectively.This thesis proposes a user clustering algorithm using K-means,which employs a semi-orthogonal user selection(SUS)algorithm to obtain the optimal number of clusters and the initial cluster centers in a dynamical way,and then proposes the user clustering metrics by taking the channel similarity and channel gain difference into consideration.Based on the results of clustering,the thesis derives the closed-form solution of the intra-cluster minimum transmit power,and proposes an iterative inter-cluster power adjustment algorithm.The convergence of the iterative algorithm is also proved.The simulation results show that in scenarios where the channel similarity is taken into consideration,the proposed scheme achieves near-optimal performance in terms of power consumption,energy efficiency,and interruption probability with lower complexity.Secondly,this thesis proposes a joint optimization algorithm for user clustering and power control based on deep reinforcement learning.The mixed-integer non-convex problem of this kind is difficult to solve by the traditional mathematical tools.In this thesis,the joint resource control task is modeled as a Markov decision process,and the iterative optimization of user clustering and power control is modeled as the reinforcement learning process.With this basis,a power control algorithm based on deep reinforcement learning is proposed.It uses the artificial neural networks to fit the actionvalue function to solve the problem of excessively large action space in communication scenarios.This thesis designs a fully cooperative action learning mechanism for multiagents to avoid the combinatorial explosion in NOMA clustering,and proposes a collaborative centralized training-distributed execution framework to avoid the dimensional disaster of the joint action space.The thesis proposes a reward function based on the optimal power allocation strategy to optimize the transmit power of the whole system.Experimental results show that the proposed algorithm can achieve the performance close to the optimal solution obtained by the exhaustive search,and its power consumption is better than that of the two-step algorithm.