Research And Implementation Of Network Virtualization Technology Based On Artificial Intelligence

With the popularization of fifth-generation communication technology and the continuous increase of Internet services,network virtualization technology will play a more important role in constructing a new generation of Internet architecture.Network virtualization allows multiple virtual network applications or architectures to share the same physical network,thereby providing network users with a variety of customized end-to-end services.Virtual network mapping technology is the main way to realize network virtualization.It uses virtual network mapping algorithms to rationally allocate the nodes and links of the virtual network to the physical network according to their respective resource requirements.With the continuous development of artificial intelligence,many virtual network mapping algorithms have begun to introduce deep learning models to optimize the mapping mechanism.They basically use Convolutional Neural Networks(CNN)or Recurrent Neural Networks(Recurrent Neural Network,RNN)to train the virtual network mapping mechanism,which significantly improves the resource utilization of physical networks compared to traditional heuristic methods.However,both of these algorithms extract several attribute information of nodes from the physical network of the graph structure to form a feature matrix to model the physical network.This method will undoubtedly lose a lot of node spatial characteristics for data in the form of a topological graph of the physical network.information.Graph Convolutional Network(GCN)It is a network specifically for feature extraction of graph structure data.Its powerful function of modeling the dependency relationship between graph nodes has made breakthroughs in the research field related to graph analysis.In this paper,we use GCN to extract and model physical network features,and use reinforcement learning algorithms to optimize the node mapping problem in the virtual network mapping process,and finally propose a two-stage virtual network mapping algorithm GCN-VNE.This article has the following three innovations:(1)The extraction of features when traditional deep learning models are applied to graph-structured data is incomplete,and high-dimensional spatial features are often ignored.In this article,we use the graph feature extraction capability of GCN to model the physical network.As far as we know,this is the first time that GCN has been introduced into the virtual network mapping problem.The multi-layer GCN network extracts network node information and spatial structure information through convolution kernels for end-to-end learning,thereby optimizing the results of node mapping.(2)We use the policy gradient algorithm in reinforcement learning to train and update the GCN network parameters.Reinforcement learning agents can learn node mapping schemes for the resource utilization efficiency of the entire physical network through continuous strategy exploration.(3)Finally,in order to confirm the advantages of our proposed GCN-VNE algorithm in virtual network mapping performance,we compared this algorithm with three mature virtual network mapping algorithms in the industry.It can be seen from the results of the comparison experiment that the GCN-VNE algorithm has certain advantages over the other three mapping algorithms in terms of average revenue,average revenue-cost ratio and request acceptance rate.