|
Abstract:
|
In some application domains, e.g., climate modeling and stochastic optimal control, constructing a point estimator can be extremely computationally expensive. Statistical inference in these domains is difficult because closed-form analytic approximations are not available and resampling-based approaches are not computationally tractable without modification. However, in settings where an inexpensive surrogate for the estimator of interest is available, we show that high-quality confidence intervals can be constructed by bootstrapping the inexpensive surrogate and then calibrating using a small number of judiciously chosen resamples of the original, expensive estimator. We illustrate the proposed methodology using a multiresolution spatial prediction model.
|
