ReViNE: Reinforcement Learning-Based Virtual Network Embedding in Satellite-Terrestrial Networks

This paper addresses the virtual network embedding (VNE) problem in integrated satellite-terrestrial networks (STNs). VNE consists of mapping virtual network functions (VNFs) in a service function chain (SFC) to physical nodes and mapping virtual links connecting the VNFs to physical links. Compared to terrestrial networks, VNE in STNs is challenging due to the movement of the non-geostationary orbit (NGSO) satellites and limited onboard processing resources. A static VNE strategy fails to efficiently address the diverse requirements of heterogeneous service requests in such a highly dynamic topology. In addition, existing solutions do not consider the capacity limitation and connectivity duration of inter-satellite links (ISLs) and ground-to-satellite links, which are essential parameters for deploying a VNE strategy in STNs. This work proposes a heuristic solution and reinforcement learning (RL)-based improved solution for the VNE scheme that can dynamically modify the existing VNF deployment strategy to maximize the average service acceptance rate and revenue. The RL agent selects a suitable VNE strategy for each service request considering the time-varying network topology and service's requirements. The proposed scheme increases the service acceptance ratio by 19.95% compared to the benchmark TS-MAPSCH.