Reasearch And Implementation Of Virtual Network Mapping Problem Based On Reinforcement Learning

Network virtualization technology can abstract the under-lying physical resources into multiple virtual networks,enabling multi-tenant sharing of physical resources.Different tenants present personalized requirements for virtualized networks,including node requirements and link requirements,thus creating virtual network requests.Based on the virtualization of the underlying physical network,virtual network requests can be mapped to the physical network.This is the virtual network mapping technology.At present,virtual network mapping technology mostly uses heuristic algorithms to manually customize a series of rules and assumptions,and the experimental results are not convincing.This paper proposes two algorithms for virtual network mapping based on reinforcement learning:A Reinforcement Learning Based on Spectral Method for Virtual Network Embedding(SR-VNE)and matrix-based perturbation learning virtual A Reinforcement Learning Based on Matrix Perturbation for Virtual Network Embedding(PR-VNE).The innovations of this paper can be summarized into the following three points:(1)The node information of the traditional physical network is represented by the attribute matrix and the link information.The link information is represented by the adjacency matrix,but the two representations are incomplete and contain noise.The SR-VNE algorithm uses the spectrum analysis method of the physical network to jointly consider the attribute matrix and the adjacency matrix to obtain a robust consensus matrix that can represent the physical network.(2)After each virtual network request mapping is completed,the physical network changes,so the physical network characteristics are dynamically changed at high frequencies.The PR-VNE algorithm uses the matrix perturbation theory to capture the changes of the physical network of continuous time nodes,and completes an efficient update method of the feature representation of the physical network.(3)The SR-VNE algorithm and the PR-VNE algorithm use the reinforcement learning method to train the virtual network mapping model.The reinforcement learning agent can effectively discover the relationship between the physical network representation and the virtual network request,thus completing efficient virtual network mapping.algorithm.According to our existing knowledge,our SR-VNE algorithm and PR-VNE algorithm are the first algorithms to apply spectral analysis and matrix perturbation theory to virtual network embedding,which can efficiently apply reinforcement learning to virtual networks.In order to verify the effects of SR-VNE algorithm and PR-VNE algorithm,we compare SR-VNE algorithm and PR-VNE algorithm with other three commonly used virtual network mapping algorithms.The results show that compared with the two algorithms proposed in this paper.The other three commonly used virtual network mapping algorithms can obtain better results in the three evaluation indicators of long-term average income,long-term revenue consumption ratio and acceptance rate.