Intelligent agent modeling

I am more optimistic about intelligent agent modeling than is Tyler. For one we already have an important, convincing, and Nobel-bestowed variant of intelligent agent modeling, namely experimental economics. Experimental economics uses one particular type of intelligent agent, the type based on…genetic algorithms. True, the intelligent agents used in I-A models are typically not as sophisticated as the agents used in experimental economics but they are rapidly improving. (Moreover, such agents are already important economic actors in their own right in limited areas, e.g. portfolio insurance, and they will continue to become more important as time continues.)

I see bringing experimental economics and I-A modeling closer as an important goal with potentially very large payoffs. Here, for example, is my model for a ground-breaking paper.

1) Experiment
2) I-A replication of experiment (parameterization)
3) I-A simulation under new conditions
4) Experiment under the same conditions as 3 demonstrating accuracy of simulation
5) I-A simulation under conditions that cannot be tested using experiments.

Now that would be a great paper. I-A agent modeling is already very useful for modeling contagion, peer effects, and highly non-linear environments. It will become even more useful when combined with experimental economics in a way that demonstrates the equivalence of the two types of intelligent agents.

Comments

Hmm.. see chapters 3 and 4 of Douglas North's Understanding the process of economic change. People are not computers and human cognition does not work with 'genetic algorithms'. Pretending people can be accurately characterized as fleshy number chrunching machines will I believe add very little to the understanding of human behavior. The reason why IA modelling is becoming popular is not scientific but status potential: it's a apparent substitute for math skills cq analytical intelligence. IA is getting popular at our sociology department, for example.

On complexity/suitability, it depends on what you're trying to model. There's lots of intelligent agent modeling and simulation going on now in wireless networks in a field called cognitive radio and we're modeling exactly what will be fielded.

There intelligent agents control the operation of say WiFi routers or cellular base stations and implicitly or explicitly compete for resources (e.g., spectrum or power budget).

Beyond Tyler's equilbria, we also end up caring alot about both parameterization effects (as Alex alludes to, but parameterization also includes variations in reasoning capacity and information availability), but also convergence and stability. As we pass through a recession and experience commodity price spikes and collapses, it should be clear that even in traditional economics we should care at least as much about how we get to (if we get to) an equilbrium as the existence of an equilibrium.

So what Alex proposes is a) being done, b) being done within a game theoretic framework, and c) looking at more than he supposes.

However, it's being applied to model intelligent agents in interacting machines (wireless devices) as opposed to models of interacting humans, which has the benefit that the modeled agent can be the real agent (just move software from simulation to the machine). There's a longish tutorial on the topic at this link with some simulation examples / discussion as well.

This, of course, assumes that experimental economics doesn't suffer from the same or worse flaws as I-A modeling for anything more complicated than a dollar auction.

Another variation on what is suggested here by
Alex T. is being done in some places already is
to have experimental subjects interact with
computer simulation programs in various ways.
One place this has been done, and which has
produced papers published recently in JEBO,
is CeNDEF at the University of Amsterdam, founded
and directed by Cars Hommes, who has coauthored
important papers with that old complexity theorist,
William A. Brock.

Robin Hanson is right, of course, that prediction
remains the name of the game, and we shall just
have to wait and see. These sorts of models have
been able to replicate real world phenomena that
more standard traditional models have not, as I
argued in the other thread on this. But, prediction
remains as elusive and difficult as ever (and the
simple models touted by Tyler, do not necessarily
do all that much better, hence so much of the
economist-bashing we see going on right now all
over the place).

Happy new year, everybody.

"Agent models may be rapidly improving, but the are still a long long way away from capturing enough human complexity to predict well."

I think Tyler's criticism wasn't that agent models are insufficiently complex. Rather his point was that he doesn't really think complex models are necessarily better because in all complex models you have to make a large number of assumptions. And as he put it you generally get out what you put in. In other words models just end up reflecting the intuitions and biases of the modeller.

To Tylers criticism there are several other related criticisms I would add:

1) more complex models tend to overfit and thus are often worse at prediction since they have higher variance. This is the reason why many in econometrics prefer parsimonious models.

2) more complex models result in more possible points of failure. Every assumption in the model is another point of failure.

3) The ugly truth that nobody will admit is that computer simulations strongly encourage "curve-fitting" by practitioners. Imagine a grad student. He has a paper due in 2 weeks. He runs his IA simulation and he gets garbage results. So he tweaks some parameters and reruns the simulation. Again crap. Tweaks and reruns. Now he gets plausible results but not what he wants. So he tweaks and reruns. After a large amount of tweaking he publishes his paper with some interesting results but he never mentions all the tweaking he did. Some of his parameters are embedded in his code and he doesn't even report them. For other parameters he comes up with some plausible reason why he chose them but after the fact. His code is unpublished.

3) is the main reason I don't believe in global warming GCM models.

"Does the GMU Econ program interact much with the Social Complexity program at GMU's Krasnow Institute? I would think the faculty of the two departments could produce a lot of interesting research along the IA topic."

There is some interaction but not enough yet. Richard Wagner and Rob Axtell, along with a number of graduate students, have helped to bridge the two departments. Look for more crossover in the coming years. I have been trying to get the Austrians and the complexity folks to see each others viewpoints, and there are a number of other heterodox-complexity economists who interact with both camps.

DEAR JSK,

IA is also getting popular in REAL scientific fields such as biology, physics, chemistry, etc.

Analytical intelligence in economics is something akin to cornering the market on 1950's mathematics

Keep perfecting that buggy whip---

I prefer Economics 2.0

租車

Comments for this post are closed