Generative adversarial imitation learning assisted virtual network embedding algorithm for space-air-ground integrated network

Abstract

The space-air-ground integrated network (SAGIN) comprises a multitude of interconnected and integrated heterogeneous networks. Its network is large in scale, complex in structure, and highly dynamic. Virtual network embedding (VNE) is designed to efficiently allocate resources within the physical host to diverse virtual network requests (VNRs) with different constraints while improving the acceptance ratio of VNRs. However, in a heterogeneous SAGIN environment, improving the utilization of network resources while ensuring the performance of the VNE algorithm is a very challenging topic. To address the aforementioned issues, we first introduce a services diversion strategy (SDS) to select embedded nodes based on different service types and network state, thereby alleviating the uneven use of resources in different network domains. Subsequently, we propose a VNE algorithm (GAIL-VNE) based on generative adversarial imitation learning (GAIL). We construct a generator network based on the actor-critic architecture, which can generate the probability of physical nodes being embedded based on the observed network state. Secondly, we construct a discriminator network to distinguish between generator samples and expert samples, which aids in updating the generator network. After offline training, the generator and discriminator reach a Nash equilibrium through game confrontation. During the embedding process of VNRs, the output of the generator provides an effective basis for generating VNE solutions. Finally, we verify the effectiveness of this method through experiments involving offline training and online embedding.