Estimating network internal link loss behavior from end-to-end multi-cast measurements

We study the use of multi-cast probes to infer network internal loss behavior from losses observed in multi-cast receivers. First part of the paper analyzes the estimation problem based on the assumption that there is no temporal correlation between link losses for different probes. We have applied expectation-maximization (EM) algorithm to this problem. We compared the results of EM with a direct approach developed in R. Cacerer et al. (1999). The second part of the paper is based on the assumption of Markov temporal correlation between packet losses. We applied EM algorithm to this problem with E-step approximated by Gibbs sampling, completely factorized variational and structured variational approximation. We observed from simulation that the links closer to observed nodes give more accurate results and that structured variational approximation improves simple variational approximation results considerably