심층강화학습 기반 저궤도 위성 네트워크 라우팅을 위한 CNN과 GNN의 성능 비교

초록

Routing in Low-Earth-Orbit (LEO) satellite networks is more complex than terrestrial routing due to rapid topology changes, multi-hop load balancing, and real-time decision constraints. State representation is crucial for DRL-based routing, and the graph-structured nature of satellite topology has led most prior studies to adopt GNNs.. From the perspective of inference speed and embedded deployment, however, CNNs with fixed-size inputs can be more favorable.. We compare the two representations in an identical DQN environment. While GNNs achieved 15.2% higher throughput than CNNs, CNNs achieved an inference latency of 2.84 ms, which was approximately 56.5% lower than that of GNNs. This study provides design guidelines for applying CNNs and GNNs under the studied experimental conditions and confirms a trade-off between routing throughput and inference speed.

키워드