Customized Network Slicing Technology Based On Machine Learning

The next generation of mobile Internet needs to meet the requirements and key performance indicators of different application scenarios.However,it is impossible to support all the requirements by the same network.Therefore,under the open network architecture,it is necessary to use software defined network(SDN),network function virtualization(NFV)and network slicing(NS)technology to construct customized exclusive virtual subnet,and meet the needs of specific network services.In the process of instantiation of network slicing technology,the classification and identification of network services are the basis of establishing service-oriented network slicing and it needs to be solved first.Due to the continuous expansion of network scale,traditional mathematical tools have encountered unprecedented obstacles in solving this problem.With the development of machine learning,the related algorithms of machine learning can realize the recognition of network services more accurately.Therefore,the network slicing scheme based on machine learning is worth studying.At the same time,flexible and efficient network resource allocation scheme is also one of the decisive factors to ensure user experience.How to ensure quality of experience(Qo E)while improving network spectrum efficiency in network resource allocation still needs further study.In view of the above two problems,the following studies are carried out in this paper.In this paper,a network service identification method based on neural network is studied,and a network slicing design scheme based on machine learning is presented.Each layer of the network slicing system model is designed in detail.Aiming at the problem of network service identification,a network service identification scheme based on neural network is designed,and the principles and schemes for determining the parameters of neural network are given.Then,aiming at the SDN environment,a sliding time window method is designed to collect the network information and complete the sample annotation with the access logs of the server.Finally,a network slicing prototype platform based on SDN/NFV is built by combining software and hardware.The data set is built on the SDN experiment net based on the sliding time window.The feasibility of the neural network based on service identification scheme is verified based on the dataset of the experimental network and the open Moore dataset.Meanwhile,it is also proved that the neural network model trained by the SDN dataset is more capable of generalization when applied.Then,based on the result of service type recognition in the experimental network,the experiments of fast creation and deletion of network slices are carried out.In order to further ensure the performance of network slicing,a network resource allocation algorithm based on reinforcement learning is studied in this paper.The algorithm uses the type of network service and the amount of downlink data for each service as the state set of Q learning,and the resource allocation scheme as the action of Q learning.The value function of Q learning is defined by considering the spectrum efficiency and Qo E.Using Bellman Equation to generate cumulative rewards for solving the problem.Q learning considers the influence of the future network state on the current decision,therefore the efficiency of Q learning algorithm is improved effectively by increasing number of iterations.Finally,the simulation experiments of network bandwidth resource allocation for three kinds of services are carried out.The results show that,on the basis of satisfying the Qo E demand of multi-service,the resource allocation algorithm based on reinforcement learning in this paper can improve the spectrum efficiency of the network.