Graph-based multi-agent reinforcement learning for railway infrastructure decision support

Data-driven methods have advanced infrastructure management and policy planning, but often overlook system-level interactions critical to networked systems like railways.In railway maintenance planning, for example, the topology of the network plays a central role in system dynamics and optimal decision-making.In this work, we propose a hierarchical Bayesian model that leverages a latent Gaussian Process on Graph kernel to embed topological features and enable accurate inference of a deteriorating railway network environment from real-world data.To address the complexity of network-level optimization, we integrate multi-agent reinforcement learning with graph-based deep learning, enabling topology-aware intelligent agents to collaboratively optimize maintenance strategies across the network.