Active learning for system rare event estimation

Abstract
This master thesis investigates data-driven techniques for rare event estimation in systems of computational models. In particular, it proposes two algorithms with the goal of improving and extending the existing class of active learning methods for system rare event estimation (a.k.a. system reliability analysis). The first algorithm proposed addresses the problems of computational efficiency, complex system configurations and small failure probabilities. The second algorithm is an extension to the first one, accounting for model evaluation costs and computational budget limitations. In both the algorithms, several advanced modelling techniques are used including subset simulation, (PC-)Kriging surrogate modelling and Sobol’ indices. The algorithms are validated on a set of benchmark problems and tested on a complex engineering application. The results show that the proposed algorithms constitute an efficient and flexible framework for rare event estimation in complex systems of expensive computational models.

Keywords
System reliability - Active learning - Surrogate modelling - Subset simulation
- Kriging - PC-Kriging