AI-assisted bridge management system using deep reinforcement learning with expert demonstrations

Bridge deterioration can lead to a reduction in structural capacity, affecting both serviceability and ultimate limit states, and may ultimately result in bridge failure. Therefore, bridge management plays a vital role in safety, operation, and resilience. However, traditional bridge management primarily relies on expert judgment, which often tends to be conservative and uneconomical. Ensuring the structural risk is within an acceptable level while minimizing maintenance costs within the service life has become a challenge. The Actor-Critic algorithm in deep reinforcement learning is proposed for the decision-making within the maintenance procedure. It can process structural information as input and generate inspection & maintenance actions as output. This paper develops hybrid Markov decision processes to satisfy different deteriorated patterns. Finally, to overcome the agent convergence issue in complex environments, this paper proposed a new training method that combines imitation learning with deep reinforcement learning.