Simulation Optimization

Modern stochastic systems often have a sophisticated structure. The system performance is generally not an analytical function of the decision variables, but a complex surface that can only be evaluated at discrete locations via noisy samples, usually from a simulation model. Since the sampling process is often expensive, we would like to identify the optimal decision with minimal samples. Simulation optimization is essentially a trade-off between exploitation, which tends to sample more at ‘‘promising’’ areas, and exploration, which tends to sample more at ‘‘uncharted’’ areas.