organizations’ behavior is constrained,
he says. “Large organizations, even if
they are terrorist organizations, may not
have the ability to make radical changes
in their behavior without taking time to
do so, and that makes them a little more
predictable,” he says.
A particularly rich source for data
on behavioral patterns is news articles,
Subrahmanian says. He and colleagues
have begun testing a new algorithm that
identifies and extracts keywords, such
as “abduction,” “suicide bombing” and
“hostage” from news databases. The
team aims to have the program ultimately extract over 700 variables accurately and automatically. These word
frequencies could then be combined
with other data to form “rules” about
groups’ behaviors. One such rule is that
Lebanon-based Hezbollah, regarded by
many, including the U.S. government,
as a terrorist group, attacks Israel less
often during Lebanon’s election years,
the team found, probably in an effort to
garner public support at home.
Subrahmanian is exploring how to not
just predict terrorist group behavior but
to actually change it. With a grant from
the Air Force Office of Scientific Research,
he and his colleagues have developed a
tool called the policy analytics genera-
tion engine. “The goal is to take the same
kinds of data and turn it backwards,” he
says. “Rather than saying, ‘Here’s my data
and here are the behaviors I’ve learned
and here are my forecasts for what this
group will do in the third quarter of 2010,’
can we go the other way?”
Changing one key factor might make
a big impact, Subrahmanian says. For
instance, the rule about Hezbollah sug-
gests that the political pressure of an elec-
tion year keeps attacks low. Finding ways
to dial up Hezbollah’s political involve-
ment in Lebanon might lead to fewer
attacks on Israel. The idea is to identify
the critical factors, and then figure out
the constellation of forces that influence
those factors, Subrahmanian says.
Calculated risks
In a more concrete example of counterterrorism, Subrahmanian and his team
Terrorist trails two suspicious ships sail between ports (numbered), forming trails that can
be analyzed for insight into terrorist activities. One ship’s trail (red line) has a very predictable
pattern. the other ship’s trail (black line) isn’t as tidy, but still offers some useful clues. every
time the ship leaves Port 1, it goes somewhere for a quick trip and eventually ends up at Port 5.
Ship travel between ports
1
2
3
4
Port
5
6
have devised a new algorithm that
predicts where caches of improvised
explosive devices might be hidden. The
algorithm was designed to discern a
fine line: Stockpiled weapons kept too
close to attack locations run a high risk
of being blown up or targeted in military sweeps. The weapons also couldn’t
be too far away, because transporting
explosives long distances heightens the
risk of being caught. These dueling considerations push weapons caches into
doughnut-shaped rings of ideal locales.
Researchers fed the program the locations of seven months of attack data and
weapons caches found in Baghdad. Then
the program was asked to find the best
cache locations for the next 14 months
of attack data. The algorithm pinpointed
the weapons caches’ locations to within
about 0.7 kilometers — less than half a
mile — of actual caches discovered by
U.S. forces. The work, done with graduate
student (and U.S. Army Captain) Paulo
Shakarian and Maria Luisa Sapino of the
University of Turin in Italy, will appear
in an upcoming ACM Transactions on
Intelligent Systems and Technology.
“We view this as not being a predictor of
where attacks will occur,” Subrahmanian
says, “but rather, preventing those attacks
from occurring.” U. S. Army officials have
begun using Subrahmanian’s program
in Baghdad.
Since these results will soon be published for everyone — including terrorists — to see, the researchers also figured
out how the ideal cache location would
change if terrorists know that computer
scientists can predict their hiding places.
In an unpublished follow-up study,
Subrahmanian and colleagues found
that this situation would probably push
the caches outside of the doughnut-shaped comfort zone, both farther from
and closer to the attack sites.
7
8
Building the perfect cell
Applying principles from game theory —
the mathematical study of strategic
interactions — is also useful for studying
the structure of terrorist cells. Using a
version called cooperative game theory,
Roy Lindelauf of Tilburg University in
