What if we could know about a crime before it happens? We could avoid huge amounts of suffering — but even though that goal fuels stories like Minority Report and Person of Interest, it seems to be far away from reality. Until now.
Experts say the field of "predictive policing" has advanced by leaps and bounds. And soon enough, the cops may have the ability to tell the future. Already, pilot projects have been able to predict the rise of a new drug market in Pittsburgh — and that's just the beginning. Here's what the experts told us.
Predicting future crimes was the topic of discussion at a panel held at the recently concluded WorldFuture 2012 Conference held in Toronto, Ontario. The panel consisted of three experts in the relatively new field of "predictive policing": Andreas M. Olligschlaeger of TruNorth Data Systems Inc., Tom Dover of the Federal Bureau of Investigation, and John Jarvis, who works at the FBI's Behavioral Science Unit.
Together, the three walked us through the cutting-edge systems that are already in the pipeline. But while their field faces serious technical and logistical challenges, their efforts are very likely to yield profound results in the not too distant future. If their work is any indication, the Minority Report future may eventually await us.
Forecasting rather than predicting
As the WFS panel got under way, Olligschlaeger and the others made it clear from the outset that they are not fans of the term ‘prediction.'
"I'd much rather frame the discussion in terms of foresight," he said, "we're not trying to construct perfect predictions of the future, or to predict single instances of crime — but rather develop aggregate accounts of crime."
To that point, Dover added, "We're hoping to anticipate events in such a way so that they can be effectively managed before they occur." It's through intelligent forecasting, said Dover, that they're hoping to get at the explanations that underlie crime and criminality. "That's what's going to help us discover those things that we're not expecting," he added.
While they're largely looking to anticipate criminal trends and isolate geographical hotspots, the panelists also indicated that they're hoping to conduct risk assessments to prevent the kind of mass shootings that recently occurred in Colorado. "We want to be able to focus on an individual who starts to show red flags, and that's where law enforcement and social services gets hung up," said Dover. The challenge, he added, "is in figuring out a way to alter the landscape to prevent outcomes like this from happening."
And it's through the use of information technologies and complex analytics that they're hoping to accomplish this. Ultimately, they're hoping to make law enforcement more efficient, and to prevent crimes as much as possible through more informed actions by police. Moreover, predictive policing will help cash-strapped cities that are increasingly looking for ways to save money. "This would allow police departments to replace street level officers with white collar specialists — what will result in substantial savings," said Olligschlaeger.
A methodological approach
The formal field of forecasting, what's often called predictive analytics, has been around for decades — but it's only been recently that the police have started to take notice. Work first got started back in the 1980s and 1990s, most notably by the Jill Dando Institute of Crime Science in the UK, and then later by Olligschlaeger, Will Gorr, Jackie Cohen and Yvonne Thompson of Carnegie Mellon University. And while their methodologies have largely remained unchanged since those early days, the technologies to support their ideas are finally making it possible.
In order to do this kind of forecasting, criminologists have borrowed heavily from other fields that are far removed from policing, from the financial sector to geology.
And it makes sense — this isn't your everyday policing. This is a highly complex and technical field that involves deep data aggregation, statistical analysis, and cutting-edge algorithms.
Specifically, their predictive systems utilize both univariate and multivariate statistical methods — such things as random walk analysis (a kind of status quo forecasting), Bayesian autogression (which is probability driven), artificial neural networks (which mimics the inputs and outputs of the human brain), and genetic algorithms (a process in which a model undergoes mutation periods until hitting a stopping point).
Self-exciting point processes in particular are a good example of how the police can effectively use these statistical techniques. Olligschlaeger used the example of earthquakes. "After a quake you typically get aftershocks," he said, "and with crime it's exactly the same way." He noted that burglaries tend to spawn burglaries in the same area. "We're trying to predict where and when these ‘crime aftershocks' will occur." And based on preliminary results, Olligschlaeger is particularly optimistic about this approach.
Another promising technique is the leading indicator model. "We also need to be able to forecast new kinds of crime," noted Olligschlaeger, "so we're trying to detect those things that could lead to a better understanding of what's going on — some kind of new trend or criminal possibility that we need to be aware of." Further to this, they're also hoping to be able to forecast crime in those areas in which crime is not happening, and vice-versa.
But how do you know if it's working?
Typically, experts create a geographical map, that looks like something that was generated by a heat sensor — red indicates potential hotspots, while shades of yellow indicate more moderate risk areas. These maps can help police departments dispatch their officers in an efficient manner, while not wasting their efforts in areas where crimes aren't likely to occur.
This said, these models have had relatively little use in the real world and there's very little evidence of their potential efficacy. But recent work in Pittsburgh and Rochester has been extremely encouraging — and in fact, one of these systems was successfully able to predict the rise of a new drug crime market in Pittsburgh. It turns out that these methods lend themselves quite well to tracking trends in drug trafficking.
In addition to these cities, pilot projects are underway in Los Angeles, Santa Cruz, and Charleston. While very much in the beginning stages, developers and investigators are hoping for universal adoption.
A strange irony of all this, however, is the challenge of validating success. If a crime does not occur in a specific area after the dispatching of police, for example, how do they know that a crime was actually prevented? The panelists agreed that this will not be an easy task.
The validation problem has given rise to another crucial tool used by the predictive police: Computer simulations.
"We can't test these systems on neighborhoods, and then just sit back and do nothing," Dover told io9 after the session had concluded, "That would open us up to some serious liability issues." Instead, Dover wants to create simulations to test his models — to see what would happen given a large set of different variables and conditions — including both police action and inaction. "As strange as this might sound," he told us, "we're even having to create simulations in which virtual people actually get murdered."
Taking this idea even further, Dover wants to use other kinds of simulations to get inside the minds of criminals — a notion that could seriously take police departments one step closer to a Minority Report-like world
What the future holds
All three panelists were excited about the long term potential for predictive policing. They were confident that their ideas will continually increase in power and sophistication, resulting in an adoption rate much faster than crime mapping. They fully anticipate an influx of both open source and proprietary products, and the rise of best practices.
But their visions of what's in store in about 25 years from now were particularly fascinating — and even a little bit disturbing.
The panelists speculated that computers will largely take over the work of analysts, and that criminals will be tracked by sophisticated algorithms that monitor internet activity, GPS, personal digital assistants, and all communications in real time. It's thought that unmanned aerial vehicles (i.e. drones) will increasingly be used to track potential offenders in order to predict intent through their body movements and other visual clues. They're also hoping to take advantage of increased transparency and crowd sourcing in order to acquire more data and get better leading indicators.
As for 50 years from now, the panelists foresee the use of human equivalent intelligence in their systems, and the use of completely automated algorithms. And quite unbelievably, they're anticipating being able to conduct remote scanning of human brains to acquire an interpretation of intent — leading to real time arrests of potential crimes.
"While we may never experience a complete Minority Report world," said Olligschlaeger, "it's not entirely outside the realm of what we may be capable in about 70 to 100 years."
Images via Cinekatz, Eurotv.us, SingularityHub.