Virtual Network Embedding with Virtual Nodes Ranking and Multi Points Sampling

Network virtualization provides a flexible and dynamic way to allocate physical resources to multiple virtual networks. Deploying virtual network requests onto the substrate network with appropriate allocation of resources to achieve more efficiency and accept more requests is of great importance. In this paper, we solve the virtual network embedding problem with neural network. We design a trainable virtual network node ranking method leveraging a graph neural network to provide a more reasonable virtual node mapping sequence. Then, a multi-record point sampling strategy that can collect samples from multiple record points is designed to reduce the correlation of samples in the training set and obtain the global optimal solution in the embedding process. Simulation results show that the improved strategy can greatly improve the learning effect of agents, and the results have increased from 74.4% to 85.9% on the long-term acceptance, from 72% to 85% on revenue to cost ratio respectively.