Stochastic polynomial chaos expansions to emulate stochastic simulators

TL;DR

This paper introduces stochastic polynomial chaos expansions as a new surrogate modeling approach for emulating stochastic simulators, effectively capturing complex response distributions without requiring repeated simulations.

Contribution

The paper develops a novel stochastic surrogate model using polynomial chaos expansions with latent and noise variables, enabling efficient emulation of stochastic outputs.

Findings

Outperforms generalized lambda model and kernel estimator in accuracy

Effectively models bimodal and complex response distributions

Does not require repeated runs for the same input parameters

Abstract

In the context of uncertainty quantification, computational models are required to be repeatedly evaluated. This task is intractable for costly numerical models. Such a problem turns out to be even more severe for stochastic simulators, the output of which is a random variable for a given set of input parameters. To alleviate the computational burden, surrogate models are usually constructed and evaluated instead. However, due to the random nature of the model response, classical surrogate models cannot be applied directly to the emulation of stochastic simulators. To efficiently represent the probability distribution of the model output for any given input values, we develop a new stochastic surrogate model called stochastic polynomial chaos expansions. To this aim, we introduce a latent variable and an additional noise variable, on top of the well-defined input variables, to reproduce the stochasticity. As a result, for a given set of input parameters, the model output is given by a function of the latent variable with an additive noise, thus a random variable. In this paper, we propose an adaptive algorithm which does not require repeated runs of the simulator for the same input parameters. The performance of the proposed method is compared with the generalized lambda model and a state-of-the-art kernel estimator on two case studies in mathematical finance and epidemiology and on an analytical example whose response distribution is bimodal. The results show that the proposed method is able to accurately represent general response distributions, i.e., not only normal or unimodal ones. In terms of accuracy, it generally outperforms both the generalized lambda model and the kernel density estimator.