Characterization of gradient estimators for stochastic activity networks

Abstract

This thesis aims to characterize the statistical properties of Monte Carlo simulation-based gradient estimation techniques for performance measures in stochastic activity networks (SANs) using the estimators' variance as the comparison criterion. When analyzing SANs, both performance measures and their sensitivities (gradient, Hessian) are important. This thesis focuses on analyzing three direct gradient estimation techniques: infinitesimal perturbation analysis, the score function or likelihood ratio method, and weak derivatives.

To investigate how statistical properties of the different gradient estimation techniques depend on characteristics of the SAN, we carry out both theoretical analyses and numerical experiments. The objective of these studies is to provide guidelines for selecting which technique to use for particular classes of SANs based on features such as complexity, size, shape and interconnectivity. The results reveal that a specific weak derivatives-based method with common random numbers outperforms the other direct techniques in nearly every network configuration tested.