Evaluation And Optimization For Artificial Intelligence-empowered Radio Resource Management Algorithms

The fifth-generation(5G)mobile networks have been developed to change our lives by connecting everything.The sixth-generation mobile networks are expected to deliver faster,more reliable,more comprehensive and smarter services than 5G.Accordingly,opportunities and challenges come together.In order to deal with resource management problem,artificial intelligence has gradually become the research hotspot.However,the selection,evaluation and optimization methods for resource management algorithms in different scenarios need to be further studied.Therefore,this thesis aims at efficiently evaluating and optimizing the algorithms under different application scenarios.The main contributions of this thesis include:Firstly,as for resource management technologies,artificial intelligence(AI)-empowered cognitive radio and AI-empowered network slicing are introduced.Then,the mathematical model-based resource management method and the artificial intelligencebased method are illustrated along with a brief survey.Secondly,aiming at evaluating radio resource management algorithms,a unified evaluation system and evaluation method are designed.The key of the designed evaluation method is to improve the evaluation efficiency through classification of scenario challenge levels by proposing a new parameter(i.e.,challenge index).Thirdly,the problem formulations are carried out for two radio resource-sharing scenarios,i.e.,network slicing and radio links,and the challenge levels of the scenarios are divided according to the proposed parameter.In the network slicing scenario,the challenge level is determined mainly according to the number of base stations.In the radio links-oriented scenario,a theoretical analysis on the number of links that the scenario can accommodate is conducted.The resource sharing algorithms based on ant colony optimization(ACO)and deep Q network(DQN)are designed to solve the resource management problems.The performance of these two algorithms,such as utility,resource satisfaction,delay,data rate and execution time,is compared under scenarios of different difficulty levels.In the network slicing scenarios,ACO and DQN have similar performance except execution time.In the radio links-oriented scenario,the designed DQN fails in medium and difficult scenarios due to the increase in the number of actions.Fourthly,dealing with the problem of high complexity of ACO in dynamic scenarios,memory-inherited ACO is proposed.The comparison between the proposed optimization method and ACO algorithm shows that the execution time can be significantly reduced by the proposed memory-inherited ACO while the performance can still be maintained when partial links move.Fifthly,in order to tackle the failure of the designed DQN in the radio links-oriented radio resource management scenario,the domain knowledge-based DQN is developed.The proposed optimization method is compared with ACO and DQN algorithms.The simulation results show that,compared with the DQN algorithm,the proposed optimization method can provide the spectrum allocation scheme effectively in scenarios of different difficulty levels.In addition,the performance of the proposed optimization algorithm is similar to that of the ACO algorithm,and even exceeds the performance of the ACO algorithm in terms of data rate and latency under difficult scenarios.Finally,based on the idea of explainable artificial intelligence,visualization method is designed to explain and optimize the resource allocation scheme provided by the AI algorithm.The proposed method can be employed to review the algorithm decision rule,confirm the accuracy of the decisions,and identify the key factors which have great impact on the performance.