Abstract:
|
We discuss an inverse problem for stochastic simulation where, given only output data, we nonparametrically calibrate the input models and other related performance measures of interest. Such problems face the challenge of high-dimensionality (since the inputs and outputs can be regarded as probability distributions) and the contamination of stochastic or Monte Carlo noises. We propose an optimization-based framework to compute confidence bounds on input quantities, with constraints connecting the statistical information of the real-world outputs with the input-output relation via a simulable map. We analyze the statistical guarantees of our approach from the view of data-driven robust optimization and in relation to constraint complexities, and discuss numerical solution methods.
|