NeTS: Small: AutoSlicing: Safe Online Autonomous Network Orchestration Towards Pervasive Slicing-as-a-Service

Sharing network resources using "network slicing" is the key technique in cellular 5G and beyond wireless networks to flexibly and cost-efficiently support emerging applications with extremely diverse needs, such as augmented reality (AR), autonomous driving, and unmanned aerial vehicle (UAV). With the momentum of open network initiatives (e.g., O-RAN), the network orchestration problem--how to provide needed resources to a dynamic and diverse set of applications--becomes increasingly complex and challenging, due factors such as the high number of network states and the need for nearly real-time control (e.g., milliseconds). Existing coarse-grained solutions cannot tackle this fine-grained network orchestration problem in terms of responsiveness, cost-efficiency, and autonomy, which limits the wide adoption of network slicing by telecommunication network providers. The vision of this project is to achieve network slicing as a service (slicing-as-a-service) with autonomous network orchestration to agilely support arbitrary mobile applications with extremely low costs. This project would advance online network automation in next-generation mobile networks, in terms of autonomy, intelligence, adaptability, and assurance. This project outlines a new autonomous network orchestration framework (AutoSlicing) to achieve pervasive slicing-as-a-service towards next-generation mobile networks. The fundamental idea is to enable deep reinforcement learning (DRL) policies to continually learn to orchestrate and adapt to domain shifting with assured service-level agreement (SLA) of slices via interacting with real-world networks. Specifically, the following research thrusts will be investigated. First, new offline simulator augmentation techniques will be designed to reduce the sim-to-real gap by augmenting existing network simulators. Second, new online network orchestration techniques will be designed to prepare the Deep Reinforcement learning policy for potential domain shifting in real-world networks. The augmented simulator will be used for policy training and evaluation during online network orchestration. In addition, AutoSlicing will be implemented and deployed on an end-to-end mobile network at a site-scale testbed at UNL and a city-scale PAWR platform. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.