COLLEGE PARK -- Army Capt. Paulo Shakarian took a crash course in the tactics of Iraqi insurgents during his second deployment to Iraq in 2006 and 2007. Now he's using that hard-won knowledge to develop a technology that might help fellow soldiers locate enemy bomb supplies before they can be used.
The West Point graduate worked with U.S. Special Forces and Iraqi security trainees near the city of Balad, north of Baghdad, to suppress waves of roadside bomb attacks on U.S. and Iraqi forces.
In one instance, he said, soldiers were trying to protect a road that was "constantly" experiencing attacks with improvised explosive devices, or IEDs.
"We knew the ... triggermen had to be able to see the vehicles," Shakarian said. "If you're waiting to blow up a guy, you've got to be hiding out, sitting there quietly for a long period of time."
But where? Patrols soon found evidence of the insurgents' overnight camps, he said, "so we focused on those areas for patrols, and we cut off the enemy from using those places for attacks."
Shakarian, 29, is now a doctoral candidate in computer science at the University of Maryland, College Park, relying on his combat experience to help UM computer scientist V.S. Subrahmanian develop a technology to help soldiers figure out where insurgents are hiding their bomb supplies.
If it works, lives might be saved. And initial experiments have been promising, he said.
Using data from a series of related bombings in Baghdad, their system was able to predict - within less than a third of a mile in eight out of 14 cases - where explosives caches were actually found.
By crunching the attack data, the scientists determined that bomb supplies were most likely stored roughly 1 to 2 kilometers from the site of an attack. The system refined its predictions even more by accounting for ethnic differences, and the fact that insurgents from one group would likely try to avoid neighborhoods populated by different sects or protected by the U.S. military.
"We were not expecting results this good," Shakarian said. The team's initial successes were reported this month in College Park at the Third International Conference on Computational Cultural Dynamics.
Shakarian and Subrahmanian are now in talks with Defense Department contractors. They're hoping to interest one of the Pentagon's research laboratories in running independent tests of the system, using real-time data on bomb attacks in Iraq and Afghanistan.
They're hopeful that the process will produce a valuable battlefield tool.
"If I were a military officer," Subrahmanian said, "I could direct additional surveillance on these areas, or send search teams. ... I could put additional pressure on intelligence assets I have in this region to better determine where these cache sites are."
Shakarian graduated from West Point in 2002. He is now enrolled in the West Point Instructors' Program, which is funding his master's and doctoral studies at College Park so that he can one day teach at the military academy.
Subrahmanian is director of the UM Institute for Advanced Computer Studies (UMIACS). His research focus is on using computerized reasoning to model past behaviors as a way to predict group behavior.
A year ago, a few months after they began working together, the two men realized they might be able to use the technology to find enemy bomb caches.
Currently, the Army deals with IED threats in two ways, Shakarian said. They can use information from past attacks as a way to predict and prevent future attacks; and they can use forensic and intelligence data to identify and eliminate the insurgent network.
The UM system takes a third approach. Called SCARE, (for Spatio-Cultural Abductive Reasoning Engine) it uses data on past attacks to predict the location of bomb supply caches. Shakarian said such caches typically store explosives and trigger devices for five to 10 attacks. They're sometimes found in abandoned buildings, on construction sites or in hand-dug "spider holes."
First, Shakarian needed a database of previous attacks, preferably attacks carried out by the same organization. While individual behaviors are highly unpredictable, the behavior of large groups is more consistent, slower to change and thus easier to predict.
Shakarian found an online database of IED attacks in Iraq, compiled by military historians. He focused on 73 attacks in Baghdad during 2007 and 2008, all of which involved relatively sophisticated Iranian-built devices.
"These devices were more effective than your normal roadside bombs," Shakarian said. "And they were also less common." As a result, the reports described these attacks in more detail. The data set also represented the work of one insurgent group.
The two men began entering the latitude and longitude from all 73 attacks into the SCARE program that Subrahmanian's lab developed. They also added data on where bomb caches had been found during the same period.
"We don't have any information on which caches caused which attacks," Subrahmanian said. "We just know that these attacks occurred here, and these caches were found there."
Finally, they programmed SCARE to use Google Maps overlaid with information on Baghdad's ethnic and religious neighborhoods.
Focusing on the first seven months of the attacks, the computer then calculated the shortest and longest distances between bomb caches and attack sites, and began inferring which caches were most likely related to which attacks. The reasoning was simple.
"IED attacks seem to be carried out by insurgents who want to maintain some physical separation from the attack itself," Subrahmanian said. They also don't want to keep the bomb-making materials in or near their homes and risk arrest. But they don't want to hide them too far from the target, either. And the Iranian-backed Shiite bombers can't safely hide their heavy, bulky supplies in hostile Sunni neighborhoods, or those controlled by U.S. forces.
The SCARE program digested the data and "learned" that the typical distances between caches and the most likely related attacks ranged from 1.1 kilometer to just under 2 kilometers. The computer's job then was to use abductive reasoning to predict where the explosives used in the remaining 14 months of attacks had most likely been stored before they were planted.
Unlike deductive reasoning, in which you try to predict what will follow from a set of facts, with abductive reasoning you try to find the best explanation for a set of observations.
"We're saying that I know I've had this spate of attacks in these locations," Shakarian said. "I know these distance constraints ... and we have the overlay for sociocultural information. Based on these attacks, I want to know the cache sites used for these attacks."
In seconds, the computer painted a scattering of 14 yellow bull's-eyes across the map of Baghdad neighborhoods, each linked to one or more attacks. Then, when the researchers compared those predictions with reports of actual cache discoveries, they found that eight - more than half - were less than a third of a mile from the real caches. Three more were within two-thirds of a mile.
That's the type of figure that the military can put to use, Shakarian said. "It's easier for you to deny [access to or from] the area in a circle 700 meters around these predictions" than the 1,200- to 1,800-meter radius band around each attack site - the territory from which an insurgent likely traveled to plant the bomb, he said.
The "acid test," Subrahmanian said, would be to use real-time attack data, run the SCARE predictions, and see where the caches turn up. "It is something we hope the military will consider doing, or at least testing."
A strength of the system, he said, is that it is adaptable to a variety of conflicts, on different human and geographical terrain, such as those in Afghanistan. SCARE could even be used in entirely different situations. "Certainly it could be used to model the behaviors of large institutional investors," he said. "We could use it to model the behaviors of political organizations."
For now, Shakarian said, he and Subrahmanian remain focused on SCARE's potential value in combat. But "this work does provide a solid foundation for future systems based on these ideas."