Multiagent reinforcement learning-based dynamic controller placement algorithm for satellite network

Low Earth orbit (LEO) satellite networks have emerged as a research frontier and technological hot point in space communications, and software-defined network (SDN) decouples control and data planes through centralized control plane. However, SDN implementations in satellite network face critical limitations in control plane reliability and latency when handling exponentially increasing traffic. To address dynamic topology and burst traffic challenges, this paper proposes a multi-agent reinforcement learning-based dynamic controller placement algorithm for satellite network (MARL-DCP). The proposed MARL-DCP specifically targets challenges of traffic load volatility and topological uncertainty through adaptive controller placement and load balancing strategies, significantly enhancing control plane responsiveness while maintaining service continuity across time-varying satellite networks. Firstly, the dynamic controller placement problem (CPP) is divided into control handover and controller migration, which simplifies the complexity of dynamic CPP. Secondly, the dynamic CPP is modeled as a decentralized partially observable Markov decision process (DPOMDP), and positions of controllers are decided collaboratively by agents based on the network environment data. Finally, the MARL-DCP is proposed to dynamically hand over control rights and adjust positions of controllers to maximize the dynamic controller placement payoff value, improving control plane reliability. The simulation results show that, compared to the static controller placement method, the proposed MARL-DCP effectively reduces control latency and improves the reliability of the satellite network control plane over 10%.