Deep Reinforcement Learning for Structural Model Updating Using Transfer Learning Mechanism

Structural simulation models using pre-defined assumptions and values of material properties usually produce results that differ from the real structures with varying degree of accuracy. This is commonly attributed to two broad types of uncertainties, namely aleatory (related to inherent randomness) and epistemic (related to lack of knowledge). Sources of such uncertainties include material properties, construction techniques, aging, and natural or man-made hazard-induced damage. Accurate computational models with on-time model updating capabilities are important goals in engineering research and practice for monitoring the structural health during the operation stage and for making rapid and well-informed decisions following extreme events, for example, major earthquakes. Moreover, the advances and recent adoption of artificial intelligence technologies bring effective and innovative solutions for the structural model updating endeavors. In this paper, a novel model updating method is proposed using two deep reinforcement learning algorithms, namely, Advantage Actor-Critic and Asynchronous Advantage Actor-Critic. In addition, transfer learning is adopted, which generalizes the trained model to various scenarios and enhances the computational efficiency. Through several computer experiments, the results demonstrate the high accuracy and computational efficiency of the proposed approach, which brings about its promising potential for practical engineering applications.