Path‐Based Deep Reinforcement Learning for On‐Board Routing in Satellite Constellation Networks

ABSTRACT Efficient usage of available network resources is a crucial factor for broadband services in interconnected satellite constellations. To meet required quality of service standards under heavy network loads, it is essential to optimize traffic distribution among the intersatellite links. To address this challenge, we propose an adaptive traffic engineering framework based on deep reinforcement learning. Our approach employs a path‐based decision‐making strategy, using a centralized agent to distribute incoming flow requests on a set of candidate paths. This method approximates optimal solutions to the multicommodity flow problem with relatively low computational complexity, making it suitable for in‐space network control despite on‐board processing limitations. The performance of the proposed scheme is evaluated against state‐of‐the‐art rule‐based benchmarks in various scenarios. We quantify the impact on performance of different candidate path sets and traffic patterns. Overall, the proposed solution presents a viable approach for optimizing flow distribution in satellite constellation networks, suitable for the integration into the controller logic of software‐defined networks.