Abstract
Infinitesimal perturbation analysis is a technique for estimating derivatives of performance indices from simulation or observation of discrete event systems. Such derivative estimates are useful in performing optimization and sensitivity analysis through simulation. A general formulation of finite-horizon perturbation analysis derivative estimates is given, and then sufficient conditions for their use is presented with a variety of queuing systems. In particular, the effect of such features is investigated as multiple customer classes, state-dependent routing, finite buffers and complex queuing disciplines. In several cases, our conditions impose restrictions on the topology of a network; in all cases, the conditions are easy to check. The results contained here are obtained by specializing conditions established in a general framework in earlier work, and should serve as a practical guide for possible applications of perturbation analysis.