Predict the Next War
New software allows practically anybody to try forecasting the future
Thanks to new software, even rank amateurs can engage in making predictions about the threat of biological warfare or the spread of epidemic disease. They can discover why ancient civilizations fell, what causes traffic jams or what makes an illegal drug popular.
These innovative computer programs are also becoming key toolkits for top researchers exploring the toughest societal questions. For instance, during the swine flu outbreak in 2009, scientists at the Brookings Institution used such a program to project that the disease could infect 30 percent of the world’s population if not controlled.
Now amateurs can use them, too.
Creating a Virtual World
The programs use a method of simulation known as agent-based modeling, or ABM. ABM allows users to create a virtual world in which interactions between virtual people shape the system. Early versions of the programs have been around since the early 1990s. Now many competing toolkits have emerged.
To be sure, these toolkits will not resolve every major question of our era. Both critics and seasoned users point out that ABM doesn’t recreate reality, but represents it. Like any other model, it cannot predict the future and will sometimes produce unrealistic results.
Nevertheless, advocates say ABM can give new insight into problems where other methods have failed.
Until about two years ago, the serious application of ABM in research was restricted either to those who already knew computer programming, or those willing to hire a programmer to do it for them. The latest toolkits are far more user-friendly. Cynthia Nikolai, a Notre Dame University doctoral candidate who has studied ABM software, told me that until recently, the need for programming expertise had slowed down the adoption of ABM in research.
But now, she said, “more and more people who are using the toolkits aren’t computer scientists. The field is in the stages of changing.” Her oft-cited study of ABM toolkits, co-authored by Notre Dame professor Gregory Madey, was published in March 2009 in the Journal of Artificial Societies and Social Simulations.
Many models now incorporate “graphical programming,” which allows users to design models by dragging-and-dropping buttons and graphs onto a virtual canvas. Other toolkits incorporate more sophisticated graphics into their simulations, and innovative ways to involve real people directly in the models.
To run a simulation, a modeler first designs a virtual world, populates it with agents, proscribes their behavior, determines that world’s initial conditions, and then sits back to watch the interaction. The essential idea is that many phenomena can best be understood as networks of relatively simple, independent agents who follow certain rules of engagement. Stephen Guerin, founder and president of the ABM contracting firm RedfishGroup, said that the interesting information comes from the interactions.
“It almost would have been better if they called it interaction modeling,” he said.
Andrei Borshchev is the CEO and managing director of a Russian company Name the company that developed the ABM software “AnyLogic,” which markets to non-scientific commercial users. He thinks the ABM approach has enormous potential for practical applications.
“Many people still think that ABM is a purely academic thing,” he said, “which is fun to look at but not more useful than watching a game. Agent-based models are very, very diverse. Agents can represent anything, from people to companies, to R&D [research and development] projects and political parties. They may live in a physical space or they may not. This enormous diversity that has a real and practical application.”
Leigh Tesfatsion and Robert Axelrod, two of the original ABM researchers, also think that ABM has the potential to increase our understanding of political and economic systems. Tesfatsion referred me to a paper she co-wrote with Axelrod in 2006, in which they explained that understanding a political or economic system means understanding how all the interactions between individuals can have a synergy-like effect on the system’s inherent complexity, and how past experiences shape the way people interact. ABM, they wrote, is the right approach for this job.
How it Works
ABM recreates instances in which simple, unorganized interactions lead to complex and ordered phenomena—a process known as emergence. Emergence occurs everywhere. It happens when shifts in the wind carve uniform ripple patterns into sand dunes, or when changing atmospheric conditions construct perfectly symmetrical snowflakes. It happens when ants react independently to chemical stimuli, and yet unconsciously form a working colony. It happens when millions of investors around the world, trying to maximize their own profits, consequently drive a stock market that regulates the asset prices of companies. Emergence exists whenever a pattern seems to “emerge” from chaos, and the modeling software tries to recreate this process, to study how it occurs.
The reduction in the need for programming expertise that the new toolkits make possible has meant that researchers using ABM can work more efficiently. Nikolai, for example, is currently working with biologists to model effective ways to reduce the spread of malaria, “In my research I found that one of the major points that people hated about ABM was the programming part. I don’t understand why,” she said with a chuckle, “but I am a computer programmer myself so I might be biased.”
Computer programming is akin to making a ceramic vase. Just as a small air bubble may cause the vase to shatter in the kiln, a single misplaced semicolon might drive the entire program to malfunction. Seasoned practitioners are used to this precision. Beginners are not.
Franziska Klügl was on the development team for the University of Wurzburg’s toolkit called SeSAm, or Shell for Simulated Agent Systems. She said it still takes at least a little bit of training in programming to understand the abstract thinking behind it.
SeSAm is one of the visually programmed toolkits that Nikolai and Madey examined. Originally designed in 2004 for researching traffic, the software has since become available for free to the public in a user-friendlier format. Unlike most toolkits, which require you to switch to a computer language as your model becomes more advanced, said Klügl, with SeSAm “you have a user-interface builder that is completely visual. You can simply click.” Researchers have used SeSAm for understanding insect behavior to improving a factory’s productivity.
Repast Simphony is another free, point-and-click toolkit specifically targeted for social science. Developed in 2007 at the Argonne National Laboratory in Chicago, RepastS—as it is frequently called—responded to the need for a powerful visually-programmable toolkit that could also be programmed normally, in case any computer scientists wanted to modify it for another use. The toolkit was used in many simulations, such as Nikolai’s malaria study.
I decided to try some agent-based modeling myself. I downloaded NetLogo, a toolkit designed at Northwestern University in 1999 that has a range of users, from elementary schoolchildren to advanced modelers like the RedFish group. While NetLogo doesn’t rely entirely on drag-and-drop modeling as RepastS and SeSAm do, it uses an extraordinarily manageable programming language called Logo.
I don’t have a shred of programming ability, but after perusing NetLogo’s online manual for about an hour, I was able to reconstruct a segregation simulation that came pre-loaded with the toolkit. Soon my labors lay sprawled across my screen: 2,500 pixilated turtles, half of them red, half green. I could tell the program at least how many neighbors of the same color a turtle needed to feel comfortable; using that criteria it then calculated what percent of turtles were “happy.” I picked a reasonable 45 percent as comfort point for my turtles—so no turtle would oppose life in a well-integrated world.
With the parameters set, the model came alive. It tracked how the turtles were likely to move over time, and stopped when 100 percent were happy. Two and a half thousand flickering dots crawled around my screen for about a minute, and arranged themselves into clean, distinct patches of red and green that resembled holly more than any integrated society.
Good Lord, did I just create virtual bigoted turtles?
Not quite. I had actually constructed a very crude version of Thomas Schelling’s famous Tipping Game.
Schelling, using pennies and nickels instead of red and green turtles, showed in 1969 that even small racial preferences for one’s neighbors could lead to completely segregated neighborhoods. Using Milwaukee as a test case, it explained the phenomenon of self-segregating communities. Even though each individual turtle—or coin—is perfectly fine with the majority of its neighbors being a different color, the unhappy turtles gravitated toward their own kind. On the other hand, the turtles that were very picky about their neighbors often wound up in integrated neighborhoods. Schelling, who is widely considered a pioneer of ABM, won the 2005 Nobel Prize in Economics for his work.
Agent-based models have similarly added important insight to all fields of research. Klugl’s team, for example, used SeSAm to model traffic patterns for cars; Guerin did the same for boats in Venice. The RedFishGroup also used modeling to study how “shocks” in the drug communities—e.g. large shipments or big arrests—affect patterns of drug use. In 2009, researchers from Michigan State, Brigham Young, and the University of Illinois used ABM to determine how violence-causing rumors can be avoided. Ultimately, the researchers seek to model human social behavior. ABM, after all, is merely a computerized attempt to understand the final outcome of our interactions.
Despite its promising uses, and all the instances in which it has served as a helpful research tool, ABM has its limits. The toughest problem, Nikolai told me, is validating the models, or proving that they accurately depict reality. “There are just so many variables and different parameters in a model,” she said, and ensuring that they are all accurate at once is wildly complicated.
Doug Samuelson is a scientist who helped found the North American Association for Computational Social and Organizational Sciences. He said that ensuring a model “makes sense in real life” is often a challenge, because simulations like disaster scenarios can only be tested virtually. You can’t set fire to a stadium, for example, just to check if your model of a stampede is accurate. More realistic models aren’t necessarily the remedy, Samuelson added, because “the closer a model approaches the complexity of reality, the harder it is to differentiate between programming errors and real results.”
In the end, we humans can be irrational and subjective—qualities difficult to quantify in the neat, formulaic world of computers. Thus no software can make agents into exact replicas of what they represent, limiting the scope of any model. Moreover, critics of complexity science note that a lot of the surprising behavior exhibited by agents—who, by definition, are extremely sensitive to initial conditions and small variations in their climate—is seldom found in the real world. This means that modelers, in the words of one critic, “are never likely to enjoy the intellectual comfort of laws.”
Perhaps this is why academic researchers and professional modelers alike are careful not to reject good old techniques such as non-computational modeling and empirical research. Jason Barr is a Rutgers economist and an organizer of the NYC Computational Economics and Complexity Workshop. He said that the economics field would benefit most if economists complemented their standard methods with ABM.
Still, ABM is a powerful tool with the potential to provide insight beyond what traditional models make possible. Further down the road, ABM software developers plan to integrate ABM with computing reality. RedFish is working on ambient computing, which uses projectors to convert an entire room into an interactive multi-agent simulation. Guerin has already worked on a project for the Venetian government that projected an interactive agent-based model of the city’s canal onto a sand table.
The biggest advantage of ambient computing, he said, was its power to involve everyone in the room in the simulation.
In a similar vein, the SeSAm development group and the University of Örebro in Sweden are working on integrated agent-based modeling with virtual reality chambers. “The human modeler has an avatar in the model,” explained Klugl, “he can participate in the running simulation and see what the avatar sees.” This also helps with validation, she said, because the modeler, from the perspective of an agent, can check if the other agents are behaving plausibly.
“Your computer is just as much about communication as it is about computation,” said Guerin. “ABM can be a strong communication device, more than it can be—or an equal amount to how much it can be—a predictive device.”