Subset Simulation is a Markov chain Monte Carlo method, initially conceived to compute small failure probabilities in structural reliability problems. This is done by iteratively sampling from nested subsets in the input space of a performance function. Subset Simulation has since been adapted as a sampler in other realms such as optimisation, Bayesian updating and history matching. In all of these contexts, it is not uncommon that either the geometry of the input domain or the nature of the corresponding performance function cause Subset Simulation to suffer from ergodicity problems. To address these problems, this paper proposes Branching Subset Simulation. The proposed framework dynamically partitions the input space, and recursively begins Branching Subset Simulation anew in each partition. It is shown that Branching Subset Simulation is less likely than Subset Simulation to suffer from ergodicity problems and has improved sampling efficiency in the presence of multi-modality.
翻译:暂无翻译