End-to-end cross-domain network slicing for LEO satellite Internet of Things using deep reinforcement learning

Abstract

Low earth orbit satellite Internet of Things (LEO SIoT) represents a pivotal infrastructure for enabling global, ubiquitous, and low-latency communications. Network slicing has emerged as a promising paradigm to address the core challenges of dynamic resource allocation and service differentiation in LEO SIoT. To ensure end-to-end Quality of Service (QoS) across slices, we propose an end-to-end cross-domain slicing framework. This framework comprises a centralized cross-domain coordinator and multiple domain-specific controllers. An adaptive cross-domain delay-balancing strategy is devised for the coordinator to allocate delay budgets across domains. Based on the allocated delay budgets, a radio access network (RAN) slicing policy is developed for the RAN controller using a Dueling Double Deep Q-Network (D3QN), enabling dynamic wireless resource allocation that guarantees low-latency for ultra-reliable low-latency communication (URLLC) services while optimizing throughput for enhanced mobile broadband (eMBB) users. To cope with the rapidly evolving LEO topology, a rollback mechanism-based authentic boundary Proximal Policy Optimization (RMABPPO) algorithm, enhanced with an integrated Graph Attention Network and Sequence-to-Sequence module (iGATSeq), is introduced for core network (CN) slicing. Simulation results demonstrate that the proposed end-to-end cross-domain slicing solution not only ensures Quality of Experience (QoE) for both eMBB and URLLC users, but also significantly improves resource utilization in LEO SIoT.