Quantifying Error due to Input Modelling in Computer Simulation Event

Event details

With recent advances in computational power simulation has become an important tool in many industries for carrying out experiments that are otherwise too expensive or time consuming. Simulation models aim to mimic the real world allowing a practitioner to experiment indefinitely with a system without any physical impact. But no simulation model is a perfect representation of its physical counterpart; all outputs from the model should be reported with knowledge of the error about them to avoid over confidence in the results and poor decisions being made. Input uncertainty (IU) and bias due to input modelling are the outcome of driving simulation models using input distributions estimated by finite amounts of real-world data. Often the estimated input distributions are assumed true, overlooking these sources of error, but it can be shown that in many cases input uncertainty overwhelms stochastic estimation error caused by finite simulation effort. Moreover, although bias due to input modelling is known to decrease faster than the variance due to input uncertainty this does not mean it is irrelevant when considering the error in a simulation performance measure. We therefore argue ignoring these sources of error is a mistake. This talk will discuss methods for quantifying input uncertainty and bias due to input modelling in computer simulation models. We first present a method for quantifying input uncertainty in simulation models with piecewise-constant non-stationary Poisson arrival processes. We then illustrate a response surface approach to bias estimation along with a diagnostic test for identifying, with controlled power, bias due to input modelling of a size that would be concerning to a practitioner.