I recently looked at Al Roth’s web page on “Game Theory, Experimental Economics, and Market Design” which is probably the best link collection on the web on these issues. What caught my eye was the nonchalant unusual combination of “Game theory” and “experimental economics”. From a CS point of view, we rarely see someone who is both a theoretician and also does “experimental-anything”. Al Roth’s reasoning for this combination is very simple: “My research is in game theory, experimental economics, and market design (for which game theory, experimentation, and computation are complementary tools)”.
So, should “algorithmic game theory” (in the broad sense) also put significant energy into experimentation? Should we specifically study the interaction of humans with our systems? My original point of view was that luckily we are spared the difficulty of dealing with these pesky humans. As we study rational behavior of fully controlled computer systems, we get to design their behavior and we will certainly design them to be perfectly rational according to our best analysis (the buzzword is “hyper-rational”). Thus if our analysis suggests that the rational behavior when working under protocol X is to do Y, then we will just program our software to do so — why shouldn’t we? Of course, there could be various technical challenges (computational limitation, partial information, etc) but these we can analyse and quantify and handle as well as is possible. “Rational Choice” debates about the relevance of game theory to human behavior concern economists — not us.
The problem is of course at the boundary of the computerized system: how are the utilities specified? A well-designed computerized system can act rationally to optimize a given goal, but it must be given this goal by humans. This is not just a conceptual difficulty, but actually the main challenge in most systems that I know. Elicitation of user preferences via some kind of a user interface is a main technical challenge. Consider software that is used for participating in some (not too simple) auction such as the adwords auction. While there are usually many interesting algorithmic and game-theoretic considerations in designing the bidding software (and the auction itself), I would say that the most important design bottleneck is still the user-interface: how do we let the user specify what he wants to get in the auction in an intuitive yet powerful way. Part of the problem is that “what the user wants” has different meanings: one of them is imprecise and sits in his head, while the other is formal and sits in the software. All too often optimizing for the latter does not do so for the former. This brings the framing effect into play in a direct technical sense: what users will seem to want depends on how things are presented to them. While I can see some companies designing their software as to “push” people towards behavior that optimizes the company’s revenue, it is not clear to me what should our normative goals be in the design of such user interfaces.
I don’t have a conclusion to end this post, so instead here are two related comments:
- I never quite understood the difference between experimental economics and behavioral economics. My impression is that it is simply a question of your attitude towards game theory: experimental-pro, behavioral-con. The anti-game-theory bias is probably not that of its founders Kahneman and Tversky. Shortly after winning the Nobel prize in economics in 2002, I heard Kahneman talk at a dinner held by the Hebrew University’s rationality center. His point was that while most psychologists would immediately dismiss the economic assmptions of rationality as absurd regarding human beings, the greatness of his co-author Tversky was to realize that there was an important underlying truth to this point of view, and hence that it was interesting to carefully study its limitations.
- The experimental vs. theoretical question is part of a much broader set of issues of the theory-vs-practice family. It seems that algorithmic game theory sits in an interesting cross-roads where we get many perspectives on these issues: from CS, from economics, the wider rational choice debate, the actual Internet, etc.
In a sense both the game-theory revolution which meant adding strategic considerations to otherwise basic and quite simple models of economics, and behavioral economics which replaces “rationality” (In the sense of agents optimizing their utility) with more complex models of individual behavior, had a similar effect – of making it impossible to analyze large economics scenarios. (Large can sometimes mean anything with more than one agent and sometimes even one agent.)
(BTW “behavioral economics” has a large body of theory and model building. It is not only experimental work.)
The internet can serve (and probably already does) as a good place for experimental economics, and the larger endeavor of empirical economics. Probably the most fruitful empirical and experimental issues will involve the study of fairly basic aspects of economics – perhaps those dating before both the game-theory and behavioral (somewhat competing) revolutions.
Here is a gross, but hopefully useful simplification.
Generally, all economics is the study of behavior. In order to model, we make certain simplifying assumptions. The economic assumptions center around formal notions of expected utility.
Behavioral economics, from the 1950’s through to 1980, meant one of two things: bounded rationality, or the careful experiments designed to show one or more axioms required for an expected utility function failed. Allais, Ellsberg and numerous other contributed to this field.
Tversky and Kahneman changed the focus in two ways. One by introducing framing, which challenged not just a particular axiom set, but the very idea of building a model of preference assuming invariance. Tversky, in particular, has a number of experiments showing how mathematically equivalent decision scenarios are treated differently – as if the decision maker could not deduce a=a.
This is a very powerful, if true, attack on the very idea of abstract model building.
Current academics who style their work as “behavioral economics” are largely adding to the scope of framing effects.
On the other hand, experimental economics as it relates to Vernon’s Smith’s work is aimed at exploring the circumstances in which markets clear. The original experiment in the 1950’s showed that a simple auction device could allow markets to clear at the price level and transactions that the simple supply and demand curve theory predicted.
Smith and others have followed this up with careful experiment showing how the creation of bubbles are possible, even in a world where most of the information is transparent.
Thanks for the info. Also see the Market design post: http://marketdesigner.blogspot.com/2009/04/market-design-and-experimental.html
[…] Turk-based behavioral experimentation has the immense appeal of being cheap, fast, and easy to administer. There are obviously pitfalls such as getting a good grasp on the population, but so does any experimental setup. Such a platform may be especially appropriate for Internet-related behavioral experiments such as figuring out bidding behavior in online auctions, or how to best frame a commercial situation on a web-page. Could this be a tool for the yet not-quite-existent experimental AGT? […]
[…] social networks, to cloud computing) and the growing roles of information economics as well as experimental economics within the fields of economics and game […]