[liberationtech] Probabilistic Models for Warning of National Security Crises

Yosem Companys companys at stanford.edu
Tue Oct 16 18:51:30 PDT 2012


From: David Blum <dmblum at stanford.edu>

You are cordially invited to my dissertation defense:

Date: Tuesday, Oct 23rd, 2012
Time: 1:30 pm
Location: Encina Hall, CISAC East Conference Room (2nd Floor)

Title: Probabilistic Models for Warning of National Security Crises
Adviser: Elisabeth Pate-Cornell

Abstract:

Intelligence analysts have experimented with static forms of Bayesian
inference of crises since the late 1960’s. However their static
probabilistic approaches have proven unsuitable for warning analysis. In
this dissertation I develop an analytic framework for crisis early warning
that addresses four shortcomings of earlier efforts: (i) it has the
flexibility to incorporate geographical variation; (ii) it reflects the
dynamics of a crisis in order to enable lead time estimation; (iii) it
incorporates conditional dependencies among signals and data; (iv) it
treats an analyst’s decision of when to warn, and type of warning to give,
in decision-theoretic terms. The framework is rooted in a general warning
system developed by Pate-Cornell.  The models comprising the framework are
illustrated using a historical example, the lead up to Japan’s attack on
Pearl Harbor in 1941. They are then demonstrated through a contemporary
case study, the warning of violence against civilians in Guatemala being
perpetrated by a transnational criminal organization.

>From the Pearl Harbor illustration, I find that beginning on November 27,
1941, at no time would US intelligence analysts have been expected to
believe that an attack on the US Pacific Fleet based in Oahu was more than
one one-hundredth as likely as an attack on the US Asiatic Fleet based in
Manila Bay. Yet, in the days preceding Japan’s December 7 attack, the
expected disutility of an attack on Oahu may have exceeded that of an
attack on Manila Bay by a factor upwards of ten, assuming risk neutrality
and a discount rate driven by planned military deployments. Using
parameters, some of which are entirely illustrative, the model would have
issued an alert for an attack on Oahu on December 2, 1941. Despite the
illustrative nature of the results, the exercise highlights the importance
of incorporating decision analytic techniques in the warning process. It
also demonstrates the process of entity tracking through signal inference,
which is broadly applicable.

>From the transnational criminal organization case study, I identify an
allocation of Guatemalan military forces that satisfies a minimax criterion
for interdicting drug traffic. I further identify three municipalities that
consistently have high likelihood of being optimal targets for coercion
between February 1 and August 31, 2012, for purposes of securing
trafficking routes into Mexico. Lastly I identify two routes spanning the
width of Guatemala, which, over the same period of time, are consistently
the most secure smuggling routes given an uncertain military force
deployment whose maximum likelihood satisfies a minimax criterion.


Thanks,
David Blum


David Blum
Ph.D candidate, Decision and Risk Analysis Group
Department of Management Science & Engineering
Predoctoral Science Fellow
Center for International Security and Cooperation
Stanford University

510-414-4450 (m)
415-230-0645 (skype)
815-301-3500 (fax)
dmblum at stanford.edu
http://cisac.stanford.edu/people/davidblum/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.stanford.edu/pipermail/liberationtech/attachments/20121016/1ce03522/attachment.html>


More information about the liberationtech mailing list