Theory and Practice Agree in Theory but not in Practice

June 17, 2009 10:58 AM

In what cases is game theory a good model of actual human behavior?

I'm curious as to the extent that game theory has been empirically validated as being a good model of human behavior, in terms of theories and domains. If you know of any specific theories or domains that have been validated, or any good survey papers or introductory material, I'd really appreciate it.

For example, (I'm making these next statements up) "game theory turns out to be surprisingly effective for modeling dating behavior among high school students" or "the predicted results of Prisoner's Dilemma based on game theory highly correlates with our user studies".

I'm familiar with the basics of game theory (minimax, nash equilibria) and basics of behavioral economics (the ultimatum game, Kahneman and Tversky's Prospect Theory,
The Paradox of Choice, some of George Loewenstein's work).

posted by jasonhong to science & nature (9 answers total)

Not exactly what you are asking but, this web site might be an interesting read.
posted by caelumluna at 11:17 AM on June 17, 2009 [1 favorite]

I'm curious as to the extent that game theory has been empirically validated as being a good model of human behavior, in terms of theories and domains.

The way you phrase this doesn't fit with my experience of how game theory is applied (in the areas I work in). Game theory, in my experience, is used as a mathematical tool -- it isn't exactly a theory in the sense of a scientific theory. That is, it can't be empirically validated by being successfully used, in the same way that algebra or predicate logic can't be empirically validated by being successfully used in some actual theory, and correspondingly it can't be falsified by being unusable or failing in some domain). It may be that in economics the theory and the tool are conceptualized as being more conflated, though, I really don't know. Scientific theories that use some piece of game theory as a mathematical tool can, of course, be empirically validated or falsified, but you do not learn anything about game theory per se by doing this.

Game theory has been used successfully as a tool in linguistics to model communication dynamics, and to capture the idea that communicators try to be maximally informative and relevant, up to a point. Here is a survey article on the topic. It has also been used to model the interpretation of complex quantifiers in language and logic by researchers such as the philosopher/logician Jaakko Hintikka; here is a book on the topic. (In this case it is a little harder to judge the success as this isn't the mainstream theory in this domain, but I don't know of any successful arguments against it except that it is very hard to understand.)
posted by advil at 11:35 AM on June 17, 2009

there was recently an article in the New York Times that showed how game theory would help the US protect itself from terrorism
posted by Jason and Laszlo at 11:47 AM on June 17, 2009

Duverger's law is a principle which asserts that a plurality rule election system tends to favor a two-party system

Jus one example in voting. Voting is one of the few areas that can be and has been modified after game theory developed. Sometimes for good, sometimes for evil.

I'm not in the field so couldn't point to any papers that'd likely satisfy you.
posted by FuManchu at 11:55 AM on June 17, 2009

Advil is right. Game Theory is math. What you're asking is like asking how logic has been empirically validated. Do you mean to ask whether behavior based upon the assumption of substantive rationality has been demonstrated empirically? Even behavioral economics need not be inconsistent with economic rationality; you can incorporate our known cognitive deviations into the model, after all.
posted by B-squared at 2:17 PM on June 17, 2009

As I understand it, the ultimatum game is employed empirically to study societies' attitudes towards equality. As implemented the ultimatum game is hard to experimentally "validate." Whoever sets the terms of the deal needs an understanding of the other party's breaking point.

Another example of game theory might be Games With a Purpose. They goal is to design a game such that the results are useful for common knowledge/search/AI, and game theory is helpful in understanding the incentives to defect.

The magic that GWAP does is replay half the games with a live party. Before you'd have a narrow range of proposals, but now you can get a better idea across the whole range. Instead of party A rejecting some proposed solution (or whatever the game is), several dozen are involved.
posted by pwnguin at 3:05 PM on June 17, 2009

That's a really involved question. B-squared has the general answer right: game theory is a way of modeling a certain kind of multi-party form of reasoning to predict behavior. What you're probably really asking is "has rational choice theory been empirically validated" -- that is, to what extent do predictions based in the rational strategic actor model, and generated using game theory, come out right empirically?

And there, results are mixed. Part of the problem is that there are often a variety of possible game theoretic models capturing a given behavioral phenomenon. How you write the model makes a huge difference, such as what preferences you attribute to the players. Second (and relatedly) these models are consciously reductive: most get cetiris paribus predictions that are very non-robust to changes in exogenous factors. Third, there are a bunch of very well-known anomalies: voting is the biggest. But it's still used and useful in a lot of domains -- every issue of the American Political Science Review or the American Economic Review will have papers with statistically significant results from game theoretic models.

So perhaps that's your best overview -- skim over the last few APSRs. Also, search google scholar for a paper by John Conlisk called "Why Bounded Rationality?" from, IIRC, the Journal of Economic Perspectives (I forget what year). It gives a pretty good overview of the debates surrounding rat choice.

(Although, a slight quibble with B-squared: you can, kinda, ask to what extent logic has been empirically validated. See Quine, "Two Dogmas of Empiricism" for more.)
posted by paultopia at 4:22 PM on June 17, 2009

advil, b-squared, jasonhong specifically notes that he means game theory as a model of human behavior, not just as an abstract thing.

The usual answer I give is that game theory, or rational-choice theories more generally, do well when you can appeal to something like an evolutionary force that drives out entities that don't "play rationally," or at least act like they do. Something that kills or otherwise eliminates actors who don't do the optimal things that game theory says they should.

In the political world, game theory and rational-choice theories more generally do well at explaining the behavior of highly-invested elites such as drafters of constitutions, national-level legislators, or decisionmaking-level bureaucrats. If a member of Congress screws up a few votes or misjudges which votes people back home might be induced to care about, they're out of a job. If party leaders screw up their coalition-making, they're likely punished in the next election or voted out of leadership.

On the other hand, rational choice theories don't work nearly as well when they're applied to the mass public. If I cast my vote "wrong," it doesn't have any real external effect on me unless that vote makes or breaks a tie. One of the classic examples has been getting rational-choice based models of individual turnout to work -- you can sort of do it, but it requires some post-hoc cludgery and even then it only really works in the comparative statics and not more directly.*

One of the classic early examples where game theory worked was the Battle of the Bismarck Sea in WW2, in which an IJN commander had to decide which route to take while a USN commander tried to predict which route and position forces accordingly. Both sides had serious incentives to get it right (measured in terms of being bombed by the enemy or getting to bomb the enemy), and both sides used their equilibrium strategies.

Other areas where game theory will probably do well are sports, where you'll likely find that teams or players mix strategies in something close to the probabilities predicted by game theory. Same idea: failure to optimize is quickly and harshly punished by losses.

*Though even then, the exercise can still be illuminating in that you can demonstrate that mass behavior is not consistent with a simple cost-benefit calculation of turnout, so something else must be going on instead or in addition...
posted by ROU_Xenophobe at 10:26 PM on June 19, 2009

OK ROU_, but wouldn't you agree that these are successful modelers, not successes for game theory per se? The biggest problem for game theory and rational choice is that they don't have a theory of editing. In real world situations, options are rarely determined exogenously; the players arrive at them, and alter them, themselves. A smart modeler can often come up with an optimal story after the fact based upon the assumptions of his/her model. The Battle of Bismarck is a classic case. Sports, chess, TFT games have been extensively tested precisely because all of the parameters of the game are established a priori.
posted by B-squared at 7:26 AM on June 21, 2009