Reconfigurable RAN Slicing for Ultra-Dense LEO Satellite Networks via DRL

Abstract

Ultra-dense low earth orbit (LEO) satellite network (UD-LSN) is an emerging architecture in the sixth-generation communication system. Network slicing technology can build multiple virtual logical networks for services provided by UD-LSNs on the common physical network. The spatiotemporal variabilities of service requirements and available satellite resources make it necessary to perform reconfigurable resource slicing in UD-LSNs. In this paper, we present a reconfigurable radio access network (RAN) slicing architecture based on grouping and clustering in UD-LSNs. Time is separated into several slicing windows, each further separated into multiple time slots. We take into account the features of the rate-constrained and delay-constrained slices and formulate an optimization problem aiming at maximizing the long-term slicing revenue that involves resource utilization, the service level agreement satisfaction ratio (SSR), and reconfiguration revenues. The problem is tackled by a two-tier deep reinforcement learning (DRL)-based reconfigurable satellite RAN resource slicing and user access (TDRL-RSUA) algorithm. We decouple the original problem into the RAN resource slicing subproblem in slicing windows and user access subproblem at time slots. Specifically, the resource slicing subproblem is solved with the multi-discrete mask Proximal Policy Optimization (MDMPPO) algorithm, while the user access subproblem is solved with the many-to-one matching algorithm. Simulation results demonstrate that our TDRL-RSUA algorithm can improve resource utilization by more than 30% in comparison to the non-reconfigurable resource slicing strategy and achieves higher slicing revenue and SSR.