Parameter Estimation for Blackbox Stochastic Models
Layne T. Watson
Host: Homa Alemzadeh
Time and Location:
Friday, September 25, 2020
Abstract: A common problem in science and engineering is estimation of the parameters defining a simulation model, by minimizing the difference between empirical data and the model's predictions. This is especially difficult when the model is stochastic, so different simulations with the same defining parameters produce different outcomes. Typical approaches are to (1) treat the model outputs as deterministic, and make multiple runs with a deterministic optimization algorithm, or (2) make ensembles of model runs (runs with the same defining parameters), and then apply a deterministic optimization algorithm to the ensemble means. Both approaches "work", but are inefficient, more so when the error function f(x) being minimized is only defined by a blackbox. A new algorithm QNSTOP is proposed, which is particularly efficient for a blackbox error function f(x) using limited empirical data.