In this paper, we introduce a new algorithm for rare event estimation based on adaptive importance sampling. We consider a smoothed version of the optimal importance sampling density, which is approximated by an ensemble of interacting particles. The particle dynamics is governed by a McKean-Vlasov stochastic differential equation, which was introduced and analyzed in (Carrillo et al., Stud. Appl. Math. 148:1069-1140, 2022) for consensus-based sampling and optimization of posterior distributions arising in the context of Bayesian inverse problems. We develop automatic updates for the internal parameters of our algorithm. This includes a novel time step size controller for the exponential Euler method, which discretizes the particle dynamics. The behavior of all parameter updates depends on easy to interpret accuracy criteria specified by the user. We show in numerical experiments that our method is competitive to state-of-the-art adaptive importance sampling algorithms for rare event estimation, namely a sequential importance sampling method and the ensemble Kalman filter for rare event estimation.
翻译:在本文中,我们引入了一种基于自适应重要性抽样的罕见事件估计算法。我们考虑最优重要性抽样密度的平滑版本,该密度由一组相互作用的粒子逼近。粒子动力学由McKean-Vlasov随机微分方程控制,该方程在(Carrillo等人,Stud. Appl. Math. 148: 1069-1140,2022)中介绍和分析,用于在贝叶斯反问题的上下文中进行共识抽样和优化 posterior分布。 我们为算法的内部参数开发了自动更新。其中包括指数欧拉方法的时间步长控制器,该方法离散化粒子动力学。所有参数更新的行为都取决于用户指定的易于解释的准确度标准。我们在数值实验中展示了我们的方法与罕见事件估计的最先进的自适应重要性抽样算法相比具有竞争力,即连续重要性采样方法和罕见事件估计中的Ensemble Kalman滤波器。