It was senior year, and I had just taken a semester off to work for the Clinton campaign in Philadelphia. So I figured it'd be the easiest A ever if I signed up for an urban politics class.
The professor, a pearl-wearing blond fresh out of grad school, confessed she had never actually lived in a city before. But that didn't stop her from having all kinds of theories about how urban politics really worked. And that included a formula --- a mathematical formula -- that she said described how mayors and aldermen made their decisions. I think I laughed out loud when she first wrote it on the blackboard.
This Navy proposal (scroll down) is way more serious, of course. And they claim that it's already worked before. But I couldn't help thinking of that professor back at Georgetown, when I read about the Navy's idea to use a computer program to predict insurgent attacks in places like Iraq.
In current U.S. operations, terrorist and insurgent forces enjoy a significant advantage by being able to launch surprise attacks, whether by small arms, mortar, or improvised explosive devices (IEDs), against weakly defended or undefended targets and disappearing before U.S. forces can concentrate for a counterstrike. Better prediction of where and when such attacks are most likely to occur would therefore be of great benefit, allowing smart allocation of defensive resources as well as preparation for quick counteroffensive operations in response to terrorist and insurgent attacks. This task is significantly complicated by the fact that modern terrorist groups demonstrate an ability to learn and adapt quickly, making it difficult to predict future actions on the basis of past actions.
Recent work has applied and extended discrete choice models originally developed for use in econometrics to predicting the spatial probability of criminal activity. These point-pattern based density models have also been applied to the military domain for prediction of terrorist strikes and IEDs. The result is that the geographical patterns established by past events can be used to build threat maps showing where future strikes are most likely to take place, with accuracies notably better than hot-spotting techniques. The same basic strategy seems likely to be applicable to prediction of the timing of such activities as well as their location.
The technique utilizes as inputs a series of IED incidents... The models typically contain large numbers of attributes, such as population density, proximity to a police station, distance to a mosque, etc. From case to case different attributes and different numbers of attributes are important. For example, when this technique was applied to bombings in greater Jerusalem, it was found that a single attribute, the distance to a controlled intersection, was an accurate predictor.
A fundamental limitation of the techniques as they stand, however, is that they do not model changes in the subjects' decision-making processes; they must currently assume that the subjects' preferences are static. This limits the time horizon over which predictions are of use, and can cause periods of very poor prediction performance when a significant change in strategy occurs. An extension of discrete choice models that allows for learning-directed evolution in the subjects' decision-making processes would greatly improve their applicability to dynamic military situations.
The program is part of a larger effort to address the "human element" of the IED problem, National Defense reports.
"I'd like to be able to pick the terrorist out. I'd like a detector 'tricorder' for intent or evil. I'd like to know ahead of time that this person is planning to hurt other people with the use of IEDs," Office of Naval Research chief scientist Starnes Walker told the magazine.
This project won't do that, of course. But getting it right "will not only contribute to defensive operations, saving lives of civilians and U.S. servicemen, but will also contribute to quick and effective counterstrikes to weaken and eliminate enemy forces," the Navy notes. "The same techniques can be applied to civilian law enforcement to counter gangs, organized crime, and other groups with the capacity to adapt their patterns of behavior through experience."
Maybe it could even predict politicians' behavior, too.
