Ben Parker, School of Mathematical Sciences, London

Location :

Salle de réunion de l'équipe TREC,
Escalier de direction
2ème étage, porte de gauche
ENS - 45 rue d'Ulm - 75005 Paris
Directions 

Abstract :

Suppose we have a system that we can measure a fixed number of times, but at
any chosen interval. Motivated by an example of probing data networks, we model
this as a black box system: we can either chose to open the box or not at any
time period, our aim to find out how the system evolves over time.

We use the statistical principles of design of experiments to model numerical
experiments that can be designed optimally. We demonstrate how to analyse the
evolution of a system as a Markov Chain, and deduce its likelihood function,
and hence the Fisher information matrix. From this, numerical results provide a
guide to the best design for the experiment for different values of input
parameters, and we show that we can find estimators whose variance is close to
the minimum variance possible. We further develop our ideas to show what
happens when we take into account the effect of the observations interfering
with the experiment, as would always be the case with packet probing. We
present examples, and demonstrate how this could be useful to many fields, with
particular reference to experiments on data networks.

Host :

Marc Lelarge et François Baccelli