``What-if'' analysis in computer simulation models: A comparative survey with some extensions (Q805130)

The article joins to the efforts in reducing computing time in Monte Carlo simulation for expected performance measures. The four methods presented (likelihood ratio, exponential, linear Taylor approximation and two point based interpolation) use one or two generated sets of values of a random variable y whose probability density function f(y,v) depends on a parameter v. These sets are sufficient for estimation of an expected performance measure \(J(v)=E[L(y)]\) for a sufficient small interval of variation of the parameter v round the value \(v_ 0\) (``what-if'' analysis). The authors establish also expressions for the variance of the expected performance measure J in the cases of ``likelihood ratio'' method. Numerical examples show differences between the above methods and the crude Monte Carlo method (Monte Carlo used for every new value of v), regarding the relative absolute error (\(| true-estimate| /true)\) and the coefficient of variation (100 (standard error)/estimate). The methods are not compared from the computing effort point of view and there is no reference to that aspect. The authors suggest further researches connected especially to the interpolation method which, at least for the numerical example, shows good results. [Some typographical errors must be corrected.]