The score function approach for sensitivity analysis of computer simulation models (Q1090473)

The score function (SF) approach is used to estimate in computer simulation experiments the sensitivities of the performance measure l(\(\nu)\) defined as \(l(\nu)=EL(Y(\nu))\), where L(y) is a given function. Static and dynamic simulation are considered. The accuracy of the SF approach to derive sensitivities of l(\(\nu)\) is computed while using the method of independent replications, batch means and regenerative methods. A number of examples from queueing models and reliability illustrate the approach. In the second part of the paper, the author studies the so called ''what if'' problem (performance extrapolation). The problem deals with the performance measure l(\(\nu)\) when the vector \(\nu\) is perturbed by \(\Delta \nu^ s\). Again, independent replication, batch means and regenerative methods are used when studying ''what if'' problems. It is shown that the SF approach can be efficiently applied for simulation of stochastic models under light traffic. Finally, a control variate procedure is presented for variance reduction.