Multi-agent deep reinforcement learning for service function chain deployment in software defined LEO satellite networks

Large-scale low earth orbit (LEO) satellite networks constitute a core component of future sixth-generation (6G) communication systems. To address the challenges of resource scarcity and highly dynamic topologies, the integration of software-defined networking (SDN) and network function virtualization (NFV) technologies into LEO satellite networks has become imperative. We proposes a hybrid centralized-distributed software-defined LEO satellite network architecture. Within this framework, This study focuses on the service function chain (SFC) deployment problem in LEO space-ground integrated networks. time-expanded graphs (TEGs) are employed to model satellite networks with dynamic topological variations, aiming to satisfy diverse user requirements while jointly optimizing resource consumption costs and service latency. The problem is formulated as a weighted sum minimization of resource consumption costs and service latency, and this problem is proven to be NP-complete. Subsequently, we integrate the twin delayed deep deterministic policy gradient method with multi-agent techniques to design a multi-agent deep reinforcement learning SFC deployment (MADRL-D) framework for optimizing our objectives. Experimental results demonstrate that the proposed MADRL-D framework outperforms existing alternatives in terms of resource utilization efficiency, resource consumption costs, and service latency. • A novel MADRL framework for SFC deployment in software-defined LEO networks. • Time-extended graphs (TEG) effectively model dynamic satellite topologies. • Joint optimization of resource consumption costs and service latency.