Estimating nonparametric IPA derivatives of loss functions in tandem fluid models

Abstract

This paper concerns infinitesimal perturbation analysis (IPA) for loss-related functions in tandem networks of continuous flow models (CFM). It considers the loss volume at a CFM node as function of the buffer size at an upstream node. The derived IPA estimators are shown to be unbiased and nonparametric in the sense that they do not require knowledge of the probability law of the underlying traffic processes. Consequently, such derivatives can be computed not only in a simulation setting but also in realistic network settings. In the latter case, real-time online computation of IPA derivatives from observations of real-life traffic processes holds the promise of applications to network provisioning and congestion management.