June 22, 2009 (Computerworld) Bruce Bueno de Mesquita teaches international politics at New York University and consults for the Pentagon and the private sector. He applies his predictive computer models to national security issues, public policy debates, mergers and acquisitions, legal proceedings and questions regarding regulation, corporate fraud and more.
Can you tell me a little bit about how you got to where you are today? How did you come to combine computer and political science? In the late 1970s, I was doing research for a book on war. That book attempted to build a general model of decision-making related to conflict. While I was working on that project, I was asked by an official at the Department of State to give my view of what was likely to happen in India following the collapse of its government.
It occurred to me that the political infighting between parties was not much different -- except for the stakes -- than the tensions and stresses that characterize choices about war and peace. I decided to construct a data set based on my knowledge of India's political parties and their leaders and feed it into the primitive model I was working on at the time for my war project. To my surprise, the model's output led to a prediction entirely different from my personal judgment. In any event, my personal opinion turned out to be wrong and the model's prediction turned out to be correct.
That led me to build a larger sample of applications to see whether I had been lucky or whether I had found a model with some predictive reliability. As it turned out, the model got right the next 17 cases it was applied to, and at that point I concluded that probably there was some genuine explanatory power behind the model's logic. Although prediction has never been the main activity in my research and teaching life, it has continued to occupy a niche in my research and teaching.
Your computer model has been used by the Department of Defense to help predict political outcomes across the globe. How does it work? The model starts by assuming that everyone cares about two dimensions on any policy issue: getting an outcome as close to what they want as possible, and getting credit for being essential in putting a deal together -- or preventing a deal. The model estimates the way in which individual decision-makers trade off between credit and policy outcomes. Some are prepared to go down in a blaze of glory seeking the outcome they want, knowing they will lose. Others have their finger in the wind, trying to figure out what position is likely to win and then attach themselves to that position in the hope of getting credit for promoting the final agreement.
How has the political science world taken to the introduction of mathematics and computers? The response has been divided. Those more oriented toward quantitative modeling are generally supportive, sometimes enthusiastic and sometimes well-informed critics. Those whose approach tends toward area studies or historical case study analysis tend to be dismissive of technical approaches to studying politics.
Your model is based on game theory. What is this exactly, and how does it pertain to your model? Game theory is a mathematical structure for examining how people, for instance, interact strategically, each taking into account the expected responses of others and recognizing that others are taking into account that their responses are being taken into account, etc. Game theory just involves people who are assumed to be pursuing their interests and who do what they believe -- perhaps mistakenly -- will give them the best available outcome, given the constraints under which they must make choices. These constraints can involve limitations to their resources, their beliefs about the intentions of others, and many other sources of uncertainty and of risk.
Computers have made huge technological leaps since you created the model 25 years ago. How has your model changed? Although the logic behind models such as mine could have been developed 100 years ago, it would have been very difficult and costly to produce useful analyses in a timely fashion. Even when I started doing forecasting 27 years ago, computer power was so limited that I could only examine problems with relatively limited numbers of players. Today there essentially is no memory constraint and no processing constraint for the sort of modeling I do.
How do you keep personal viewpoints and bias out of the initial input? I cannot look at a data set and anticipate what the outcome will be from the model. This is true even with a data set with as few as three or four players. The problems I look at with my model typically involve many dozens of players, sometimes more than 200. There is no way to construct biased data to produce a desired outcome except to make the data appear transparently wrong to anyone looking at the data. For instance, we could produce peace between the Israelis and Palestinians by assuming they agree on everything, but that would not be taken seriously. Realistic data need to reflect the current status quo. Then how things evolve is driven by the model's logic. I teach students how to construct data sets, and they always check, for instance, that the initial round's outcome looks like the current status quo, because that is what the data should reflect if they are realistic.
A robot can only do what it's told. At any point, do you have to step in and account for the fact that computers don't have emotions to help polish an outcome? When I report results, I leave it to experts to dig into the interpretive nuances of the analysis. I limit myself to reporting only what I can show in the model's output. Part of the point behind a model such as this is for users to realize it is just a model and not a replication of reality in all its subtlety. Instead, the model should help push people to think about contingencies they had not considered but that emerge from the modeling exercise.
The question of ethics is different from stepping in and substituting personal judgments about emotions, feelings, etc. I impose limits on the things I am willing to examine with the model based on my sense of propriety, but I do not presume answers to questions and choose on that basis. I will not address a question that I view as immoral or improper and have turned down requests on those grounds.
Predicting the future of international conflicts is stressful business. What do you enjoy outside of work? I don't find doing forecasts in the least stressful. It is a great pleasure to see basic research translated into a practical tool for assisting policymakers. And this is far from the main strand of my research. Outside of work I enjoy visiting my children and grandchildren, travel, playing squash and hiking.
Forrest is a freelance photographer and writer in New York (studio@saraforrestphoto.com).