Generalizations of the surrogate estimation approach for sensitivity analysis

Gradient-based optimization of stochastic dynamic systems require “good” estimators of the sensitivity to the control variables. Numerous methods have been proposed for gradient estimation of discrete event systems, usually requiring an impractical amount of information and/or computational effort. Alternative solutions have been proposed in the literature using other estimates to drive the learning algorithm, but up to date such surrogates have been constructed case by case with ad-hoc arguments. In this paper we identify three main tools for the systematic construction of surrogate estimators: changes in time scale, changes of variables with the mean flow approach, and changes of variables through unification. We introduce each method with a detailed example of application. Surrogate estimation is also generalized with the introduction of correction factors, illustrated in the examples