Arnold Kling  

Agent-Based Modeling: Promises and Pitfalls

Rescued from the Comments... The Dimensionality, Evolution,...

Which of the following impediments to economic adjustment do you believe to be the most important?

a) the cost of establishing a new enterprise
b) the cost of integrating new workers and equipment into an existing enterprise
c) the cost of adapting physical and human capital to new circumstances
d) the cost of whiting out an old price list (menu) and updating it with new prices

If you answered (d), then congratulations--you have shown your New Keynesian bona fides. If you answered anything else, then congratulations--you have shown common sense.

Which brings me to agent-based modeling. In the paradigm of Patterns of Sustainable Specialization and Trade, an important element of economic activity is what I call "discovery." Entrepreneurs and workers constantly have to discover how best to adapt to changing circumstances.

I have suggested that this is not well described by a Walrasian system of equations. The Walrasian system, or the "monetary Walrasian" system that adds a money demand function, is like the proverbial lamppost that the proverbial drunk looks under to find a watch that he dropped somewhere else.

So, is agent-based modeling, in which you set up a computer simulation of individuals in the economy, a way of shining light on the place where the watch is likely to be? Below the fold, I will describe what excites me about ABM, what concerns me about it, and how I would recommend going about it.

What excites me about ABM is that you are not constrained to model only those dynamic elements that yield close-form solutions using standard mathematical techniques. In some sense, ABM still boils down to a set of equations, but the equations can be ones that cannot be solved without the help of a computer.

What concerns me is that simulated results can be difficult to communicate and to understand. When you create a simulation, you can pick arbitrary rules. For example, you could leave out prices altogether. That is what the "Club of Rome" did in the 1970's, when they simulated a doomsday scenario based on resource exhaustion.

I suppose that when you do standard formal mathematical modeling, the rules are also arbitrary. But because the assumptions are more visible, I think that you understand them better.

My concern with ABM is getting a result and not really knowing why you got it. Alternatively, if you really understand how the ABM is getting its result, you should be able to show how to get that results in a simple model, perhaps with a numerical example.

My dissertation adviser, Robert Solow, would recommend always starting with a simple numerical example before trying to generalize. Of course, back when I was in graduate school, you generalized by doing formal math, not by doing computer simulations. But I think that the same advice applies. Get the result in a form where you clearly understand it. Then go about trying to generalize it.

My intuition is that a combination of (a) - (c) is what accounts for fluctuations in the aggregate amount of economic activity. A computer simulation using agent-based modeling is probably necessary to bring all three into the picture. But first, I would play with simple numerical examples.

Comments and Sharing

COMMENTS (20 to date)
B writes:

This is why I intend to take a class on search and matching theory in the fall.

I'm also interested in any current research on origin and nature of shocks, whatever that means. Just making TFP stochastic and fiddling with its distribution annoys the hell out of me.

Silas Barta writes:

Your a-d point was clever, and I got a laugh out of it. In fairness, though, New Keynesians and quasi-monetarists would say that the problematic cost related to price adjustment is *not* the menu cost, but the fact that, at new prices, businesses can no longer pay fixed-rate loans set in the old environment.

Where I differ from them is that, rather than believing businesses have an inalienable right to high revenues, my reaction is, "Well, sorry you took out a loan that didn't adjust for these things, use a less brittle business plan next time. Here's the email of some QMs who can shill for you."

Blake R writes:

Computational agent-based modeling isn't going to replace formal modeling, but it seems like a very powerful supplement. Because of the inexact nature of economic science, simple mathematical models do most of our conceptual work. Simulations give results without the same sort of understanding, as Arnold notes. The output looks more like experimental data than a model. As such, I think simulations should the thought of as replacements for infeasible, economy-wide experiments rather than for intractable models. Artificial economies can be a third source of model validation in addition to lab experiments and real-world data.

(Heavily influence by Computing the Perfect Model)

Lord writes:

Count me mystified. I can't see why the costs of a-c should change over time. Now I can see that small changes in technology would lead small changes in the economy and large changes in technology could lead to large changes in the economy but those aren't a-c. I can see how higher energy costs could have broad and sizable effects on the economy but those aren't a-c. I can see how if the Fed raised rates to 20% all these would all become much more difficult if not impossible, but this isn't a-c. So how can, how does, this cause a recession, slow or prevent the adjustment process without something causing or requiring more or less of them.

gabriel rossman writes:

I'm more interested in simulations over formal modeling for several reasons but the simplest one is that simulations have much more in common with empirical work in terms of the required skill set. Coming as I do from a discipline that generally doesn't do formal modeling (but loves statistics), this is a nontrivial advantage, both in terms of writing the simulation myself and in terms of explaining it to colleagues.

Dave writes:

I'm so excited to see agent based modeling receiving thoughtful consideration here. I'll add some thoughts:

Any model requires simplification and assumptions. Agent based modeling allows those assumptions to be made clearly at the micro-level. If you believe that macrobehavior is driven by micromotives (to borrow from Schelling and the Austrians), this is what you want. Micro-level assumptions are also easier to test empirically in real world experiments than are macro-level assumptions.

The agents don't have to be perfectly rational calculators, nor do they have to all use similar reasoning. It's easy enough to give them various heuristics that we observe in real world decision makers who satisfice. Information asymmetry, a distribution of skills and ability, the combination of ideas leading to technological progress (via genetic algorithms), geographic distances, changing resource constraints, etc, are all possible. If you can replace arguments about macroeconomic philosophy, which is inevitably politically driven, with arguments about micro-level assumptions in the models that are easier to verify, I think you've made real progress.

If you believe that the economy is somewhat akin to evolutionary biology, this is much better done with agent based modeling, where some industries, skill sets, etc, flourish under one set of circumstances but not in others. Creative destruction can be modeled as a result of an ever-changing fitness landscape, in which preferences, technologies and skill sets evolve together. You would likely see slowly changing patterns of sustainable specialization and trade emerge slowly, with occasional tipping points through disruptive technologies or overinvestment that throw the system into disarray. You would want to probably work in credit markets too if you wanted insight into the recent bust.

I see the benefits of agent based modeling to be demonstrative of what happens under a certain set of micro-level assumptions. In traditional macro, these assumptions are either not done at a micro-level, or they have to be too simplified, too abstracted, or too limited to make the math work. Of course simple numerical examples are always a great way to start to organize your thoughts, but simulations can show you unexpected results when you scale your examples and add in more and more about what we know of economic activity and decisions at the micro-level.

Again, I'd recommend The Origin of Wealth to see how clearly agent based modeling's assumptions, results, and relevance can be communicated.

db writes:

Not an economist, but an engineer with an interest in economics speaking with practically no knowledge of mathematical economic models.

How can these models be implemented? Would it be possible to use a tool such as OpenFOAM (free software used mostly for CFD modeling) to model an economy as a set of finite elements (agents) operating in fields? I'm not sure what form the equations describing most economic models take, but OpenFOAM is very useful for solving sets of partial or ordinary differential equations for large numbers of individual cells.

Maybe my ignorance is making me ask stupid questions but I thought there might be an interesting answer.

Dave writes:


Here is a list of agent-based modeling software. I don't see OpenFOAM on it, and I don't think that agent-based modeling would typically involve solving differential equations, unless that was incorporated into strategies used by some of the agents.

James Oswald writes:

This whole debate reminds of the 19th century debate over what caused prices - supply or demand. Yes, real factors are important. Lower taxes and reduced red tape would go a long way to helping recovery. But nominal factors are also important. Unstable prices/NGDP does not make recalculation easier; it makes adjustment harder.

Julien Couvreur writes:

As a programmer and amateur economist, I've spent some time thinking about this before. There are a few things that make it difficult, if not impossible:

  • individual preferences are important to such a model, but personalities and preferences are not measurable. An agent-based model would have to use faked preferences.
  • planning is one of the most difficult tasks for agents, as they involve discovering and forecasting other agents preferences and behaviors.
  • innovation is also very difficult to simulate as it means the agents can imagine something that was never programmed into the simulation.

In short, I like the idea, and think that it is useful for pedagogy, but see no way we could implement anything like this to a degree that would be useful. It's an artificial intelligence problem.

Original comment

Dave writes:

Julien Couvreur,

Good comments. Certainly an agent-based model would have to make some assumptions about how preferences are represented, but it seems to me that you could create some sort of indifference curve representation, or some other representation that fits with experimental data.

I agree that planning for agents could be difficult and complex. That's why agent-based models probably need to start off simple, with just a few, short term choices that the agents are concerned with. Complexity must be added carefully.

Innovation could occur with evolutionary algorithms in which agents' solutions to a problem could combine and mutate with each other in search for a better solution through trial and error.

I don't think some magical agent-based model is about to answer all of our questions about economics, but I find it promising and consistent with Arnold's PSST and recalculation ideas.

Scott Sumner writes:

No, that is not the position of the new Keynesians. To make that claim is to engage in the fallacy of composition.

No one is claiming that individual firms and workers can avoid the effects of recession by making their individual wages and prices more flexible.

And I mean no one.

Jason Collins writes:

Julien, are your criticisms equally applicable to neoclassical economic models? Neoclassical economic models require "faked" preferences and agents to have perfect foresight. Innovation (technological progress) is usually exogenous or a simple function of the internal state.

Personally, I consider part of the beauty of agent-based models to be that agents don't have to be perfect forecasting machines.

Troy Camplin writes:

One of the beauties of ABMs is that one can in fact model complex interactions. The economy isn't simple. The continuous attempts to simplify it through math have created much of the silliness that passes for economics nowadays. This isn't to say that simple models don't have their place. You start, as Mises pointed out, with a simple model that bears no resemblance to reality in the least, then as complexifying factors to see what happens. This is how ABM's could and should be used, so that you can understand what is going on with the dynamics.

Jeff Hallman writes:

Agent based modeling (ABM), like nonlinear modeling in general, has too many degrees of freedom. You set up your agents with some heuristics and simulate to get a result. Then I modify one of those heuristics a bit or introduce an additional agent type and get an entirely different result. Who's right? Most likely neither. There are an infinite number of possible plausible heuristics, and no reason to believe the results of any of them unless they all end up with very similar simulation results.

Of course, this is also a problem with DSGE modeling, as we all know that people really do use heuristics, are heterogeneous in ways that aren't modeled, don't have expectations consistent with a particular DSGE model, and so on. So there's not really any reason to believe the results of a DSGE model either, but at least you can get published.

ABM might be useful in evaluating some kinds of policy rules. For example, is a Taylor rule really the best you can do across a wide variety of AB model economies? Or is there another simple rule that does better? Is there some simple, yet plausible, tweak to "standard" agent behaviors that make the rule perform badly?

Lots of this stuff looks like it could be fun. But don't expect the current gatekeepers (journal editors, referees, etc.) to embrace it. They've spent years mastering a different skill set, and they're not about to depreciate themselves.

granite26 writes:

E - Costs of regulatory compliance and protection from hidden liabilities.

Its what has stopped me and mine in the past.

RebelEconomist writes:

I would not normally look at this blog, but I followed the link from FTAlphaville. To me this post exemplifies what is wrong with modern academic economics. The only useful comment is the last one, because it draws on someone's experience of business. The rest is about theorising and modelling. Why not just get out and ask entrepeneurs questions like this?

Dave writes:

Jeff Hallman,

I both agree and disagree with your comment. The degrees of freedom comment is irrelevant in this case. You are conflating it with the purpose of regression. But it can merely demonstrate the macro results of a number of micro-level assumptions. Micro assumptions are easier to verify in field experiments than are macro assumptions.

I agree with the idea that the gatekeepers will be closed minded to this type of modeling though. Econ PhDs have learned a certain skill set, and this probably incorporates too much computer science to change their attitudes quickly. That's no reason not to pursue it though, if you are interested in progress. Sadly though, it seems that today's economics PhDs are just people who are good at math, not people who are good at economic reasoning.

Andrew Montgomery writes:

Which would you rather do?
(a) Put off hiring one teaching assistant you otherwise would have hired, or
(b) talk to every teaching assistant, janitor, secretary, etc. in you department and tell each of them that they're getting a pay cut and are going to need to cut back on the cable bill?

Prices are easy to change. Wages are much more difficult (especially in a unionised or regulated environment).

Jeff Hallman writes:


I wasn't using the phrase "degrees of freedom" in the statistical sense. I meant it in the sense that the rest of that paragraph describes.

Every school teacher knows that it's a bad idea to tell students to "write a 10 page paper". You get much better results if you tell them "write ten pages about a Civil War figure you admire". The open-endedness of the first instruction will have people wandering about in a fog trying to think of something to write about.

If I tell you to just simulate an economy with ABM, how do you decide what heuristics your agents should use? How do they learn? What information do they have access to? How hard can they think about what they're doing? How much and how fast can they change their plans? What is a plan, anyway? And on and on. The choices you make in answering the questions are, outside of the rational optimizing agents framework, entirely arbitrary. Why should anyone believe the results of arbitrary simulations?

Comments for this entry have been closed
Return to top