AbstractBridges deteriorate over time due to various environmental and mechanical stressors. Deterioration is a significant risk to bridge owners (asset risk) and the traveling public (network risk). To tackle this issue, transportation agencies carry out bridge management under limited resources to preserve bridge conditions and control the risks of bridge failure. Nonetheless, existing network-level analysis for bridge management cannot explicitly consider the effects of preservation actions on network risk, measured directly by functionality indicators such as network capacity. In this paper, a novel method based on deep reinforcement learning is proposed to devise network-level preservation policies that can reflect bridge importance to network functionality. The proposed method is based on the proximal policy optimization algorithm adapted for bridge management problems and improved via distributed computing and architecture. The method is applied to an illustrative bridge network. The results indicate that the proposed method can produce significantly better preservation policies in terms of minimizing long-term costs that include asset and network risks. The devised policies are also investigated in depth to allow for transparent interpretation and easy integration with existing bridge management systems.