Sampling Without Probabilistic Model

In this paper a novel technique for random vector sampling starting from rare data is presented. This model-free sampling technique is developed to operate without a probabilistic model. Instead of estimating a distribution function, the information contained in a given small sample is extracted directly to produce the sampling result as a second sample of considerably larger size that completely reflects the properties of the original small sample. As a further enhancement, the new sampling technique is extended to processing imprecise data. Model-free sampling can be coupled to stochastic structural analysis and safety assessment by application to input data or to result data. In the case of limited data, for instance, due to a high numerical cost of the underlying computational model, the novel technique can be applied to generate a proper estimation of stochastic structural responses and, thanks to a sound reproduction of distribution tails, of structural reliability. In this context it can provide a basis for increasing the numerical efficiency of Monte Carlo simulations in computational stochastic mechanics. The usefulness of the model-free sampling technique is underlined by means of numerical examples.