AI-Empowered Virtual Network Embedding: A Comprehensive Survey

Abstract

For the challenges posed by Internet rigidity, network virtualization (NV) technology emerges as a pivotal approach, imparting diversity, resilience, and scalability to the evolution of new Internet architecture. By abstraction, allocation, and isolation, the physical network is enabled to host multiple heterogeneous virtual networks (VNs), thereby facilitating the accommodation of user-customized requirements to share physical resources. Nevertheless, a critical challenge in NV implementation is the virtual network embedding (VNE) problem, which concerns the efficient allocation of physical network resources to VNs. In recent years, researchers have increasingly focused on the integration of artificial intelligence (AI) to augment VNE with heightened intelligence, efficiency, dynamics, and interactivity. Therefore, this survey offers a comprehensive overview of AI-empowered VNE algorithms, presenting insights into the general modeling, definition processes, and applications of the fundamental VNE paradigm. Furthermore, an exhaustive taxonomy is presented, encompassing categories such as single-domain/multi-domain, centralized/distributed, online/offline, coordinated/uncoordinated, dynamic/ static, and survivable/unsurvivable. Subsequently, for the prevailing mainstream methods of VNE, reinforcement learning (RL)-based and deep reinforcement learning (DRL)-based, a comprehensive review and comparative analysis of the latest works are conducted within the delineated taxonomy. Finally, the open issues, research challenges, and opportunities for VNE in future settings are identified. In particular, the significant role and key bottlenecks in the urgent vision of satellite-terrestrial integrated networks (STINs) for the 6th generation (6G) communications. This survey is expected to provide comprehensive information, guide scientific research, illuminate frontier trends, and establish the theoretical basis for further research.