I had the title of this post sitting in “Drafts” for a couple months now, but Kris and Paul have given me good reason to actually write about it. These thoughts are largely off the cuff, but they do come from experience.
What is Agent-Based Modeling?
This is not actually as straight-forward question as one might think. If you define it broadly enough as, say, any model within which agents make decisions in accordance with pre-defined rules and assigned attributes, then the answer to the overarching question posed by this post becomes: well, actually, economics has been producing agent-based models for decades, but that answer is as annoying as it is useless.
Instead, let’s start with a minimal definition of an agent-based model:
- They are composed of n >3 agents making independent decisions
- Agents are individually realized within the model.
- Decisions are made in accordance with pre-defined rules. These rules may or may not evolve over time, but the manner in which they evolve are themselves governed by pre-defined rules (e.g. learning, mutation, reproduction under selective pressures, etc).
If we stop at this minimalist definition, then the answer becomes only marginally less trivial, as essentially any dynamic programming/optimal control model within macroeconomics would meet the definition. This leads to what I consider the minimal definiton of an agent-based model as a distinct subclass of computational model:
- Agents within the model are characterized by deep heterogeneity.
- Agents exist within a finite environment which serves as a constraint in at least one dimension (lattice, sphere, network, etc).
- Decisions are made sequentially and repeatedly over time
Now we’re getting farther into the weeds and beginning to differentiate from whole swaths of modern macroeconomics that either employ a “representative agent” or collapse agent attributes to the 1st and 2nd moments of distributions. But that doesn’t eliminate all of modern macro. If embracing heterogeneous agents in your models of macroeconomics, banking, etc, are of interest to you, there are scholars waiting to embrace you with open arms.
Which brings me to the final attribute that I believe fully distinguishes the bulk of the agent-based models and their advocates from modern economics:
- Agent-based models exist as permanently dynamic creations, absent any reliance on equilibria as a final outcome, characterization, or prediction.
The departure from general or partial equilibria as outcomes or predictions is where the schism actually occurs and, I suspect, is where many purveyors found themselves with a research product they had a hard time selling to economists. Economics, perhaps more than any other social science, demands that theoretic predictions be testable and falsifiable. Agent-based models (ABMs) don’t always produce particularly tidy predictions that lend themselves to immediate validation. Which doesn’t preclude them from making a scientific contribution, but it puts them on unsteady footing for economists who are used to having a clear path from the model to the data.
OK, but really, why didn’t agent-based modeling happen?
As much as big, irreconcilable differences in scientific philosophy would make for a satisfying explanation, I suspect the most salient reasons are less sexy and, in turn, less flattering of the day-to-day realities of grinding out research in the social sciences. Here are a few.
Economics was already a “model” social science
One of the reasons mobile phones caught on faster in Africa than North America was an absence of infrastructure. The value add of going from “no phones” to “mobile phones” is far larger than going from “reliable land lines in every edifice” to “mobile phones”, making it easier to justify both investments in relevant infrastructure and bearing of personal costs. Such a thing occurred across the social sciences with regards to ABMs.
Rational choice and mathematical sociology always had a limited following. Evolutionary biologists were often alone in their mathematical modeling, computational biology barely existed, and cultural anthropologists were more excited about Marx’s “exchange spheres” than they were about formal models of any kind. For a PhD student in these fields, the first time they saw a Netlogo demonstration of an agent based model, they were seeing something never previously available to their field: the ability to formalize their own theories in a way fully exogenous to themselves. There would be no fighting about what their words actually meant, whose ideas they were mischaracterizing, what they were actually predicting. Their critics, be it journal referees or thesis committee members, would have no choice but to confront their theory as an independent entity in the world.
This advantage of formality, of independent objectivity, in agent-based modeling was not something new to economics. While critics have many (often correct) complaints about modern economics, it’s rare to air concerns that economics is insufficiently formal or mathematized.
Too many “thought leaders”, not enough science
Axtell and Epstein wrote their landmark book “Growing Artificial Societies” in 1996. In it they produced a series of toy simulation models within which simple two-good economies emerged. This wasn’t revolutionary in it’s predictions by any means (whole swaths of macro models were able to make comparable predictions for two decades prior), but the elegance through which minimalist computer code could produce recognizable markets emergent from individual agent decisions was just incredible. The potential to readers was immediately obvious: if we can produce such things from 100 lines of code, what could we simulate the fully realized power of modern programming?
What came next was…still more people evangelizing and extolling the power of ABMs to revolutionize economics. What didn’t come were new models. Forget revolutionary, its hard to even find models that were useful or at least interesting. The ratio of “ABMs are gonna be great” books and articles to actual economic models is disappointing at best, catastrophic to the field at worst.
There were a couple early models that got attention (the artificial Anastazi comes to mind), but after a few years everyone noticed that same 2-3 models were still be brought up as examples by evangelists, and none of them had meaningful economic content. As for the new models that did end up floating out there, there was also an oversupply of “big models”, with millions (billions) of agents and gargantuan amounts of code that intended to make predictions about enormous chaotic systems. Models, such as the Santa Fe Artificial Stock Market, tried to broadly replicate the dynamics of actually stock markets across a large number of dimensions. Such ambitions were greeted with skepticism by economics for a variety of reasons, not least of which the “curse of dimensionality”, which limits what you can learn about underlying mechanisms when the number of modeler choices exceeds your ability to test them or, for that matter, verify their internal coherence. For better or worse, these models felt akin to amateurs trying to predict a town’s weather 30 days out.
Bad models drove out good
The problem of too few good models was closely followed by the over-supply of bad models. Agent-based modeling, for good and for ill, is not a technique with high entry costs. A successful macroeconomic theorist is effectively a Masters-level mathematician, bachelors level computer programmer, and PhD economist. Netlogo programming can be learned in a week. You can get really good at programming agent-based models in a dedicated summer.
This isn’t unto itself a problem, but I can tell you this: in my first 5 years as an assistant professor, I was asked to review at least 100 papers built around agent-based models. I’m not sure if any of them were any good. I am sure that many of them were extremely bad. Most concerning is that I don’t think I learned anything from any of them. The costs of producing bad ABM papers is much lower than the costs of producing bad theory papers based on pure math. Bad science is often evolutionarlily selected for in modern science, a dynamic that in the case of ABMs was only amplified by a lower cost supply curve.
Now, here’s the thing: there was probably huge selection effects into what I interacted with. I doubt I was getting the best papers sent to me for review given my status in the field. But the quantity of bad papers was astonishing. They were just too easy too churn out. I suspect that some decent papers were lost in haystack of ad hoc pseudoscience and, in turn, some decent scientific careers probably got lost in the shuffle. More than once I had the thought “Editors are going to start rolling their eyes every time they see the term agent-based modeling if this what keeps coming across their desks.” Combined with the fact that ABMs are tricky to evaluate because you really need to go through the code to know what is driving the results, I think a lot good modelers got lumped in with the dreck.
[Not for nothing, it wasn’t uncommon for ABM papers to spend the bulk of the paper describing model outputs, while having nearly nothing about model inputs (i.e. rules, code, math, etc). These models were essentially black boxes that expected you to take their coherence on faith. I should note here that I haven’t really kept up with the field in the past few years. Hopefully transparency norms have improved, particularly in biological, ecological, and anthropological modeling, where ABMs have thrived to a far greater extent.]
The empirical revolution took hold of economics
I’ve save the biggest reason for last, but honestly I think it dwarfs the others.
The same rise in cheap computational power that gave rise to other forms of computational modeling, including ABMs, came along with the plummeting cost of data creation, storage, analysis, and access. By 2010 it was already increasingly clear that theory was taking a backseat in economics. Not because we were becoming an a-theoretic discipline (far from it), but because the marginal contribution of theory against the body of broadly accepted economic framings was small compared to those made by empirically testing the predictions of the existing body of theories against real data. The questions were no longer “How do we mentally organize and make sense of the world”, but instead “What is the actual measured effect of X on Y?” Theory gave way to statistical identification. Modeling technique gave way to causal inference.
Agent-based models are hard to empirically evaluate and test
Which gives way to a sort of subsidiary problem. It is more difficult for agent-based models to take advantage of the new data-rich world we live in. They don’t produce neatly direct predictions the way that microeconomic theories do, nor do they lend themselves to measured empirical validation in the same way as general equilibrium predictions of macroeconomic models. Empirical validation is by no means impossible, but it requires the matching of observed dynamics or patterns, which is generally a taller order. In this way, agent-based computational models are a bit of a throwback to the days of “high theory”, making for interesting discussion but of secondary importance when it comes to the assigning of journal real estate that makes and breaks careers.
I once presented my ABM paper on emergent religious divides, only to have an audience member become extremely upset, closing with the denouncement that “This isn’t agent-based modeling, this is economics!” That was my first exposure to the theme of ABMs as “antidote” to the hegemony of economics and all of its false prophecies. The idea that the destiny of ABMs was to unseat economics as the queen of the social sciences was probably an effective marketing strategy in many hallways, but not so much in economics departments (well, maybe at The New School).
So whey should economists give agent-based modeling another shot?
That’s another post for another day. If you’re curious though, I did write about how and why ABMs are useful for economists interested in the study of religious groups and movements. The logic of that piece applies to anyone interested in studying the macroscopic dynamics characterizing social norms, group formation and decay cycles, and how social outliers can pull entire populations in interesting directions.
Lot of good points and good lines here:
“…The ratio of “ABMs are gonna be great” books and articles to actual economic models is disappointing at best,…”
“… The costs of producing bad ABM papers is much lower than the costs of producing bad theory papers based on pure math. Bad science is often evolutionarlily selected for in modern science…”
Hi, interesting and thoughtful. I’m just curious where you would put population games models (evolutionary game theory)? These models are not ABM, but allow for heterogeneous agents and dynamics over time
I’ve played around with building ABMs in my consulting business and the inputs vs. outputs thing described here is spot on. There isn’t necessarily a good theory-based way to model very complex interactions between several types of agents, but the agent models don’t have good predictive power either in most contexts because nailing down the full set of input parameters (i.e., not missing any important parameters *and* getting your various numbers right enough) is just a bear. And if it isn’t a bear, well… then you probably don’t need an ABM in the first place.
This is really interesting but I wonder if it is worth separating out the science issues from the non science “academic culture” issues (which are really just fad and shouldn’t be taken too seriously.) But at one point I’m afraid something struck me as a bit off. The economic passion for equilibrium really can’t be about data and testing. We really have no idea whether economic systems are actually in equilibrium and there’s no restriction (in ABM at least or econometrics come to that) on tracking a “moving” system. Both points are fair enough but I’m not sure they combine to make an argument. Incidentally, on good models driving out bad: https://rofasss.org/2022/03/09/keijzer-reply-to-chattoe-brown/ (but see also, though not economics, https://www.socresonline.org.uk/19/1/16.html)
Excellent post; well thought and laid out across every point. Two additional thoughts:
I often summarize this as “ABM works especially well when you have many theories and limited data, so that simulation can quickly rule in or out which theories could never match the limited data (and then we can focus on the others). When you have tons of data, you can often test a theory more directly with more classical regression or experimentation.” You’ve unpacked this idea with great rigor and detail.
Another issue is that **actually using** rich heterogeneity in business or policy applications typically should require accurate intercorrelations. Finding and managing these in agent simulation is super challenging in most applications, and doing it well makes the verification and validation challenges you discuss even harder.