seminario E. Borgonovo

[Antonio Lijoi <lijoi@unipv.it>]

=========================================================
                  AVVISO DI SEMINARIO
=========================================================

Martedì 7 marzo alle ore 14:00, presso il Dipartimento di Economia
Politica e  Metodi Quantitativi dell'Università di Pavia, via S. Felice
7, aula L (primo piano)

                   EMANUELE BORGONOVO
            IMQ - Universita' Bocconi, Milano

terra' il seminario


            PROBABILISTIC SENSITIVITY ANALYSIS


ABSTRACT. According to the literature, sensitivity analysis techniques
can be divided in the categories of local methods [Borgonovo et al
(2003), screening methods [Kleijnen and Van Groenendall (2002)],
non-parametric methods [Saltelli and Marivoet (1990)], variance based
methods [Saltelli et al (2000), Sobol’ (2001)] and moment independent
approaches [Chun et al (2000), Borgonovo (2005) and (2006).] Besides
local methods, that aim at quantifying input influence at a fixed input
space point, all other methods aim at identifying the contribution of
uncertain inputs to output uncertainty. Their application ranges from
food-safety [Frey and Patil (2002)] to hurricane losses [Iman et al
(2005)]. Results and indications an analyst derives, however, depend on
the method selected for the study. After investigating the assumptions
at the basis of various indicator families we discuss the information
they convey to the analyst/decision maker. We start with nonparametric
techniques, and then present variance-based methods. By means of an
example we show that contradictory results may be obtained if one lets
uncertainty coincide with variance. We then examine moment-independent
approaches to global sensitivity analysis, i.e. techniques that look at
the entire output distribution without a particular reference to one of
its moments. We detail the definition and property of a new indicator
[Borgonovo (2005)], and then compare it to results that are obtained
with the Chun-Ahn-Tuck indicator and the Kullback-Leibler entropy
measure [Chun et al (2000)]. In particular, we show that the
Kullback-Leibler information can fail in identifying parameter
importance if the support of the conditional induced distribution and
the unconditional induced distributions differ. We then present
numerical results. They demonstrate that both moment-independent and
variance based indicators agree in identifying non-influential
parameters. However, differences in the ranking of the most relevant
factors evidence that inputs that influence variance the most are not
necessarily the ones that influence the output uncertainty distribution
the most. Perspectives of future research will be discussed.


Tutti gli interessati sono invitati a partecipare.




=============================================================================
NOTA:
Le norme per utilizzare il forum SIS e le istruzioni per iscrizione
e cancellazione sono disponibili all'indirizzo
.
.   http://w3.uniroma1.it/sis/forum.asp
.
L'archivio di tutti i messaggi (aggiornato al giorno precedente)
e' disponibile all'indirizzo
.
.   http://www.stat.unipg.it/cgi-bin/wilma/sis/
.
=============================================================================