Abstract
We develop a method for the evaluation of extreme event statistics associatedwith nonlinear dynamical systems, using a small number of samples. From aninitial dataset of design points, we formulate a sequential strategy thatprovides the 'next-best' data point (set of parameters) that when evaluatedresults in improved estimates of the probability density function (pdf) for ascalar quantity of interest. The approach utilizes Gaussian process regressionto perform Bayesian inference on the parameter-to-observation map describingthe quantity of interest. We then approximate the desired pdf along withuncertainty bounds utilizing the posterior distribution of the inferred map.The 'next-best' design point is sequentially determined through an optimizationprocedure that selects the point in parameter space that maximally reducesuncertainty between the estimated bounds of the pdf prediction. Since theoptimization process utilizes only information from the inferred map it hasminimal computational cost. Moreover, the special form of the metric emphasizesthe tails of the pdf. The method is practical for systems where thedimensionality of the parameter space is of moderate size, i.e. order O(10). Weapply the method to estimate the extreme event statistics for a veryhigh-dimensional system with millions of degrees of freedom: an offshoreplatform subjected to three-dimensional irregular waves. It is demonstratedthat the developed approach can accurately determine the extreme eventstatistics using limited number of samples.