Framework for Bridge Maintenance Decision-Making Based on Improved Deep Q-Networks
TL;DR
This study proposes a deep reinforcement learning framework for bridge maintenance decision-making and indicates that Prioritized experience replay (PER) Deep Q-Networks (DQN) fits engineering data samples well and shows stronger cost-reduction performance than Double DQN.
Abstract
Balancing economy and safety is critical in bridge maintenance decision-making. This study proposes a deep reinforcement learning framework for bridge maintenance decision-making. An evaluation system using economic loss metrics is developed to resolve dimensional inconsistencies between economic and safety measurement units. Prioritized experience replay (PER) Deep Q-Networks (DQN) are first introduced to bridge maintenance decision-making. Furthermore, the PER DQN policy is compared with DQN, Double DQN, and Dueling DQN to analyze the characteristics of various algorithms. A case study of a highway bridge demonstrates the feasibility of the proposed methodology. The results indicate that PER DQN fits engineering data samples well. While Double DQN policy performance remains highly stable under different conditions, its average cost is higher. PER DQN shows stronger cost-reduction performance, Double DQN shows more stable but higher-cost behavior, and Dueling DQN shows intermediate performance.