Multi-Service Distributed Routing Algorithm Based on Multi-Agent Reinforcement Learning

Satellite routing faces challenges such as highly dynamic changes in network topology and various user demands. To adapt to this dynamic environment and diverse services, we propose a Multi-Service distributed routing algorithm based on Multi-Agent Reinforcement Learning (MS-MARL). We schedule the priority of different types of service based on priority queuing. Then, we model the routing process as a partially observable Markov decision process and optimize routing decisions based on spherical distance and the QMIX model. The simulation results show that MS-MARL can reduce the packet loss rate by 11% and 38% for delay-sensitive services, and 5.5% and 33% for bandwidth-sensitive services, while maintaining end-to-end delay and throughput.