This grant provides funding for the development of a comprehensive framework for stochastic simulation input modeling that can accomplish the following: (1) represent a wide range of steady-state simulation input models, including independent univariate processes, finite-dimensional random vectors, stationary univariate time-series processes, and stationary vector-time-series processes; (2) fit these input models to data via automated algorithms while enabling intuitive, direct modification of the fitted models via the user's subjective judgment or partial information such as bounds, percentiles, or moments; (3) generate realizations of these input processes quickly and accurately in order to drive large-scale computer simulations; and (4) facilitate sensitivity analysis of simulation outputs with respect to simulation inputs by making the input models readily adjustable in terms of easily understood parameters. The framework will be based on the ability to represent, fit, and generate observations from a stationary multivariate vector time series in which each individual component can have either a Johnson, Bezier, or discrete marginal distribution; moreover, the dependence structure is specified via product-moment correlations between pairs of components that are separated by selected time lags. Such an input process will be constructed by an appropriate transformation of a Gaussian vector autoregressive process. The primary benefit of this research is that it will take reliable input modeling out of the domain of statistical specialists and put it into the hands of everyday simulation users. Simulation inputs form the core of every stochastic simulation model, so this will substantially improve the fidelity of practical simulation models, leading to more accurate results and better decisions. Since simulation analysts use what they find in software, the software developed in this research and made available to commercial vendors should speed the technology transfer.
