Active learning for structural reliability: Survey, general framework and benchmark

Active learning methods have recently surged in the literature due to their ability solve complex structural reliability problems within an affordable computational cost. These are designed by adaptively building inexpensive surrogate of original limit-state function. Examples such surrogates include Gaussian process models which been adopted many contributions, most popular ones being efficient global analysis (EGRA) and active Kriging Monte Carlo simulation (AK-MCS), two milestone contributions field. In this paper, we first conduct a survey recent literature, showing that proposed actually span from modifying one or more aspects aforementioned methods. We then propose generalized modular framework build on-the-fly strategies combining following four ingredients modules: model, estimation algorithm, function stopping criterion. Using framework, devise 39 for solution $20$ benchmark problems. The results extensive (more than $12,000$ solved) analyzed under various criteria leading synthesized set recommendations practitioners. may be refined with priori knowledge about feature problem solve, i.e. dimensionality magnitude failure probability. This has eventually highlighted importance using conjunction sophisticated algorithms as way enhance efficiency latter.