A deep reinforcement learning method for structural dominant failure modes searching based on self-play strategy. (May 2023)

• Record Type: Journal Article
• Authors: Guan, Xiaoshu; Sun, Huabin; Hou, Rongrong; Xu, Yang; Bao, Yuequan; Li, Hui
• Is Part Of: Reliability engineering & system safety. Volume 233(2023)
• Publication Date: 2023-05
• Subjects: Structural reliability analysis -- Dominant failure modes -- Deep reinforcement learning -- Self-play strategy -- Monte Carlo tree search; Reliability (Engineering) -- Periodicals; System safety -- Periodicals; Industrial safety -- Periodicals; Fiabilité -- Périodiques; Sécurité des systèmes -- Périodiques; Sécurité du travail -- Périodiques; 620.00452
• DOI: 10.1016/j.ress.2023.109093
• Languages: English
• ISSNs: 0951-8320

Abstract

Highlights: An intelligent deep reinforcement learning-based method for dominant failure modes searching is proposed. A self-play strategy for optimization of the dominant failure modes searching process is designed. A scoring function for the self-play process and improved Monte Carlo tree search process is designed. Abstract: In the research area of structural reliability analysis (SRA), the dominant failure modes (DFMs) of a structural system make significant contributions to life-span failure prediction and safety assessment. However, the high computational cost caused by the combinatorial explosion is the main problem in DFMs searching that hinders its application and further development. Recently, many successful applications have proved that the self-play deep reinforcement learning (DRL) has a strong ability to obtain action policy in the face of combinatorial explosion problems. Inspired by this, a self-play strategy is designed to optimize the DRL-based DFMs searching process and reduce the computational effort. A scoring function is designed and used as the refereeing standard of the self-play games and helps improve the efficiency of Monte Carlo tree search (MCTS) in an asynchronous training process. In comparison with the β -unzipping method and exploration-based DFMs searching method, the proposed method significantly improved training efficiency with an accuracy of over 95% and a lower requirement of the number of finite element analysis (FEA), both of which contribute to the policy learning of failure component selection. In summary, the method shows potential applications for actual structures and makes valuable contributions to the problem with high computing costs.