Probabilistic Machine-Learning Methods for Performance Prediction of Structure and Infrastructures through Natural Gradient Boosting

Abstract

The capabilities of data-driven models based on machine learning (ML) algorithms in offering accurate predictions of structural responses efficiently have been demonstrated in numerous recent studies. However, efforts to date have relied on essentially deterministic approaches, and prediction confidence measures were either derived from verification data sets or completely ignored. This study examined the potential of a new algorithm—natural gradient boosting (NGBoost)—that directly produces probabilistic predictions. This type of output fits the reliability and performance analysis frameworks naturally, and also opens the pathways to utilization of self-learning algorithms and optimal design of experiments and field measurement campaigns in engineering applications. After introducing NGBoost’s fundamentals, two representative problems in structural engineering were investigated to examine NGBoost’s feasibility: (1) prediction of the strengths of squat shear walls, and (2) classification of the seismic damage levels in ordinary bridges. The results indicate that NGBoost attains mean prediction accuracy levels comparable to those of conventional ML algorithms while providing robust estimates of prediction uncertainties.