Assessing the accuracy of estimation of Gaussian systems via sensitivity analysis, a mathematical urban legend busted
by Felisa Vasquez-Abad
Abstract: Infinitesimal Perturbation Analysis (IPA) is a methodology that uses a
stochastic derivative to estimate the gradient of the expected value of a performance function, with respect to some parameter of the model. It is unbiased only when the expectation and derivative operations can be
interchanged. For almost two decades, many papers on gradient estimation for discrete event driven systems have claimed that IPA is
the preferred method, if applicable, and it is broadly accepted today as the method with minimal variance.
Dealing with discontinuities motivated the study and development of alternative methods for gradient estimation, notably Score Function (SF) and Weak Derivative (WD) methods, both of which use a measure
theoretical approach to deal with the derivative of a probability measure with respect to some parameter of the model.
In this talk we will focus on a ubiquitous class of problems where the dynamics of the system follows a Gaussian model. They arise whenever
the processes driving uncertainties are described by "cumulative small errors", and are used in Physics,Chemistry, Economics, Finance,
For More Information: Dr Owen Jones: firstname.lastname@example.org