Here's a little exercise in scientific thinking. What's wrong the approach to science described in the following passage? (This passage, about applying network analysis to counterterrorism, is taken from the complex systems special feature in the July 24th issue of Science.)
McCulloh and Carley used metanetwork analysis to analyze 1500 videos made by insurgents in Iraq. "The insurgents would videotape most of their attacks as propaganda," says McCulloh. "As of March 2006, we had something like almost three out of every four U.S. deaths [on tape]." Carley extracted data from these videos, he says, "made a big network out of it, and ran a fragmentation algorithm which clustered them into little groups. And when you go back and look at the videos in those groups, you see forensic clues that identify who some of the insurgent cells were." The details extracted from the videos are classified, "because we worry that the insurgents will learn what we're using," McCulloh says. He and Carley worked with the U.S. military to "operationalize" the technique in Iraq. U.S. commanders there are faced with too much information and too little time to act on it. McCulloh says that Carley's metanetwork software helps them find clues and patterns—boosting the chances of catching or killing insurgents.
McCulloh claims that the technique has yielded dramatic results. "Sniper activity in Iraq is down by 70%," he says, and he's confident that IED deaths also dropped because of the insights provided by Carley's programs, although he can't cite data. "It's a simple application of metanetwork analysis," he says.
But Sageman is skeptical that military progress in Iraq can be chalked up to network analysis. "I'm not convinced [metanetworks] have helped at all," he says. "An easier explanation [for the drop in sniper attacks] might be the tribal uprising" against the insurgency in Iraq. "There's no way to know, and that's a big problem with this field in general." Carley counters that Sageman "doesn't understand the methods."
This passage captures what is wrong with much of econophysics, systems biology, sociophysics, and almost any field that been tackled by heavily computational complex systems approaches. Many of these researchers don't understand what it means to test a theory. They build these complex models, which involves making important assumptions that could easily be wrong, and then if their models fit existing data, they think the model is right.
Hence you get this McColloh guy claiming that his network analysis model was responsible for a big drop in sniper attacks, ignoring the much more obvious and plausible causes for the drop in violence: the addition of 30,000 troops and the US Military's major new approach to counterinsurgency implemented by Petraeus. The network researchers can't justify ruling out the more obvious explanation; their only retort is to say that their critics don't understand their fancy methods. (Which is not true in many cases - there are plenty physicists, biologists, and economists who understand the mathematical/statistical/computational techniques, who are bothered by the scientific culture of complex systems research.)
This a dangerous mindset to have in science. What these researchers are doing is practicing a sham form of science called by Feynman Cargo Cult Science:
There is also a more subtle problem. When you have put a lot of ideas together to make an elaborate theory, you want to make sure, when explaining what it fits, that those things it fits are not just the things that gave you the idea for the theory; but that the finished theory makes something else come out right, in addition.
And no, that does not mean simply training your model on half of your data set and showing that you can effectively explain the other half of your data. You need to be proactive about probing your model for problems. So your model explains one situation well, now go find another very different situation and see how well you can explain that. Look for predictions made by your model that have not been noted in the real-world system before, and see if your prediction is borne out.
Another example from the same issue of Science:
Complex-systems experts have also made contributions in epidemiology. In 2001, Vespignani and colleagues showed that in certain types of highly connected networks called "scale-free," it's impossible to stop the spread of an epidemic no matter how many people are inoculated. Conversely, in 2003, Shlomo Havlin, a physicist at Bar-Ilan University in Ramat Gan, Israel, and colleagues found a simple strategy for inoculating against a disease that beats picking random individuals. By going a step further and picking randomly chosen friends of those individuals, health officials can, on average, inoculate people with more social ties through which to spread the disease.
Nowhere in the article do you read about any real-world tests of this model. And you don't see any real world tests in the actual research paper either. But simply having an untested model is a 'contribution in epidemiology' apparently worth writing about in a news feature.
At the heart of scientific thinking has to be a strong desire not to fool yourself, coupled by an understanding of how to actually put that desire into practice. Complex systems are an important and relevant topic, but they've been so difficult to tackle because they are messy and hard to study. It's difficult to find the right simplifying assumptions, and to make sure that you've considered all of the important factors that go into the behavior of the system. It's so easy to be wrong.
And so it's sad to see this emerging scientific culture that bizarrely believes that if you can produce a model that fits the data that inspired you to build the model, you've actually shown that your model accurately captures the system. This culture floods the scientific literature with zero-impact papers, dazzles the computationally naïve, captures a lot of air time in the news.
Here's a prediction of my own, one that I'm willing to put to the test: if complex systems researchers don't get serious about the scientific method, their field is going to fizzle out, if not crash and burn. Because in the end you have to move the field forward. The computer models can be dazzling, but unless they produce a demonstrated string of successes that end up changing the way everyone in the field thinks - the molecular biologists, the sociologists, the economists, then the sciences of complexity will be dismissed as unfruitful. In the end, your model has to inspire a someone to pick up a pipette and design an experiment.