Statistical Emulators for Simulation Models

18 Oct 2018 20:28

---

Yet Another Inadequate Placeholder.

This is a technique for situations where simulation models are sufficiently computationally expensive that running them many times, under varying conditions, is infeasible. Instead, one wants to fit a computationally cheap statistical model to the output of a limited number of simulation runs, with some hope that the statistical model will extrapolate to what the simulation model would have done under other conditions. The cheap statistical model is the emulator.

(This is either a complement or an obverse to simulation-based inference methods like indirect inference; perhaps both.)

See also: Computational Statistics; Macroscopic Consequences of Microscopic Interactions; Monte Carlo; Simulations; Statistics;

To read (utterly inadequate):
• Cari G. Kaufman, Derek Bingham, Salman Habib, Katrin Heitmann, Joshua A. Frieman, "Efficient emulators of computer experiments using compactly supported correlation functions, with an application to cosmology", Annals of Applied Statistics 5 (2011): 2470--2492, arxiv:1107.0749