Recommendations for the tuning of rare event probability estimators

Mathieu Balesdent (1) , Jérôme Morio (2) , Julien Marzat (1, 3)

1 ONERA - The French Aerospace Lab [Palaiseau]
2 ONERA - The French Aerospace Lab [Châtillon]
3 GdR MASCOT-NUM - Méthodes d'Analyse Stochastique des Codes et Traitements Numériques

Being able to accurately estimate rare event probabilities is a challenging issue in order to improve the reliability of complex systems. Several powerful methods such as importance sampling, importance splitting or extreme value theory have been proposed in order to reduce the computational cost and to improve the accuracy of extreme probability estimation. However, the performance of these methods is highly correlated with the choice of tuning parameters, which are very difficult to determine. In order to highlight recommended tunings for such methods, an empirical campaign of automatic tuning on a set of representative test cases is conducted for splitting methods. This allows to provide a reduced set of tuning parameters that may lead to the reliable estimation of rare event probability for various problems. The relevance of the obtained result is assessed on a series of real-world aerospace problems.

Mots clés: Tuning of simulation parameters; Adaptive importance Splitting; rare event probability estimation; Kriging-based optimization

Domaines: Optimisation et contrôle [math.OC]; Probabilités [math.PR]

Mathieu Balesdent, Jérôme Morio, Julien Marzat. Recommendations for the tuning of rare event probability estimators. Reliability Engineering and System Safety, 2015, 133, pp.68-78. ⟨10.1016/j.ress.2014.09.001⟩. ⟨hal-01064029⟩