Assessing Hypercube Sampling Techniques for Risk Assessment

Two variations of hypercube sampling techniques are introduced and computationally tested using benchmark problems. The methods are variations of the Latin hypercube sampling (LHS) and incremental-fractional LHS scheme. Both can be described as stratified sampling with one sample per strata. Because they ensure uniform marginals, they are observed to have computational advantages for linear problems where weighted response statistics are sought. Advantages are less pronounced for non-linear responses and sorted statistics, which is often the case for risk analysis. The complementary cumulative distribution is identified as being helpful in assessing a methods performance. Both methods are applied to an application problem having multiple responses of interest and 48 uncertain inputs. The hypercube methods are noted to produce estimates of the mean with orders-of-magnitude lower variance than that of simple random sampling.