Application of smoothed perturbation analysis to probabilistic routing

The on-line estimation of derivatives is of fundamental importance in gradient-based routing algorithm for data networks and other applications. Smoothed perturbation analysis as proposed in this paper requires minimal knowledge about the system statistics. It is shown that smoothed perturbation analysis provides asymptotically unbiased estimates of derivatives. We determine bias and variance of the estimate experimentally and compare them to those of a likelihood ratio estimator.