The Flux by Epistemix

The Growing Impact of Agent-Based Modeling with Matt Kohler

Daniel Marzullo Season 1 Episode 2

In this episode of The Flux, John Cordier interviews Matt Kohler, Applied Complexity Scientist at MITRE and President of the Computational Social Science Society of the Americas, about the transformative power of agent-based modeling (ABM). Kohler explains how ABM simulates complex human systems and helps decision-makers understand the ripple effects of their choices. From anthropology to economics, Kohler shares examples of how ABM is helping scientists and leaders explore new insights.

As technological advancements in data, hardware, and software drive ABM to new heights, Kohler discusses its expanding role in solving real-world challenges, from traffic modeling to financial crises. With its potential to revolutionize industries by providing a clearer view of how complex systems evolve, ABM is poised to become an indispensable tool for policy-makers, researchers, and organizations.

This episode is for anyone interested in the intersection of complexity science and decision-making, and how ABM is shaping the future of predictive modeling.

Timestamps:

01:00 Matt Kohler's Flux Moment

03:17 Agent-Based Modeling Insights

05:13 Applications and Future of Agent-Based Modeling

08:56 Challenges and Successes in Agent-Based Modeling

23:41 Economic Theories and Agent-Based Modeling

26:58 Getting Started with Agent-Based Modeling

29:10 Future of Computational Social Science


John:

Hey there, welcome to the next episode of the flux, where we talk data decisions and hear stories from the people asking the what if questions to better understand our world, to help others create an intentional impact on the future. In each of our episodes, we hope you can turn the lessons and hindsight from our guests. into foresight so you can uncover how to create an intentional impact for the people and the causes that you care about. Today we're on with Matt Kohler. Matt is an Applied Complexity Scientist and President of the Computational Social Society of the Americas, who happens to work at the MITRE Corporation, where his work involves understanding complex systems in finance, economics, defense. And I'm certain there are plenty of other directions that you've gotten pulled into over the years. So right now, Matt, welcome to the flux.

Matt:

Thank you. It's a pleasure to be here.

John:

So you're right now in a pretty important role for the field of atrium based modeling being president of the computational social society. Social science society of the Americas. And we'll talk about that a little bit later on. But before we dive into talking about agent based modeling decision making here about one of your favorite, what the flux moments as we like to call them. Why don't you tell some of the listeners about what you're working on,

Matt:

i, love the flux moment. That's fantastic. And actually the first flux moment I think I really had with respect to complex systems, actually had almost nothing to do with agent based modeling. I was at the Complex System Summer School at the, Santa Fe Institute runs each summer. It was almost 24 years ago now. And it had to do with ergodic and non ergodic systems. And it absolutely blew my mind. There were a couple of guys giving a lecture about Brownian motion and things like that. and they had a cylinder inside another cylinder and some gel in between the two cylinders, and they injected a little bit of dye in one little blob, and they started turning the interior cylinder, and it, the dye started to get, fuzzed and turned into a stripe, as, and they continued to discuss Brownian motion, things like that. Then they started turning it back the other way, and after the same number of turns in the reverse direction, The die, subject to some Brownian fuzz, was back into a circle. And suddenly, I had this sense of ergodic versus non ergodic systems, and time. And that, that launched me down the road of agent based modeling because that suddenly basically showed how important path dependency is, and how fundamentally different. Those sorts of systems are so human systems in this particular case, and that ruined my legal career. I happen to be at the S. F. I. during law school of all things. And I was also really into microeconomics and game theory and things like that. But that moment where I suddenly understood path dependency and how important that is, and particularly human systems, basically meant, I was ruined and it was time to become an agent based modeler.

John:

That's a pretty common thing for a lot of people in agent based modeling, having one of these aha moments. Was there was there something that was agent based modeling specific that you've seen unlocks the aha moment for a lot of other people?

Matt:

An example that I've seen used a lot, there are two. From an experiential perspective, there's that I forget what they call it, but the game where you, choose two other people in a group, and you try to keep one of them in between yourself and the other, and the second person you chose. And then everyone does that, and the whole group. Collapses. They think that, or maybe they spread out. I can't remember which is which. And then the other one is you try to stay in between them. And so the whole group collapses or spreads out. Yeah, I don't remember which. So from an experiential perspective, that's always a big one because, of course, the rules are super simple, and yet when you ask people ahead of time what's going to happen, almost no one guesses it correctly. And then the other one that's a great one from more of a computational perspective, is the canonical Schelling segregation model. And again, that's the rules are very simple, but I probably shouldn't say no one. I'm sure someone has guessed it correctly, but typically people don't appreciate the fact that despite the fact that everyone would actually prefer to live in an integrated neighborhood, the system as a whole generates segregation and actually makes the society less well off.

John:

Oh yeah. Even on our team we've had a few people who weren't familiar with agent based modeling. Come on. And the Schelling model seems to be the thing that makes them go, Oh my goodness, we can see these simple rules leading to these macro level outcomes. And one of the guys on our team, he's from Milwaukee and he's I actually see this I know that neighborhood and I know that neighborhood. And oh my gosh, that's actually how it plays out. From your maybe past, 10 years or so, have there been any new areas where you've seen agent based modeling applied

Matt:

It's interesting. It continues to just, to, to percolate., but anthropology and archaeology, I think, it's gaining more traction, because it seems like such a wonderful application, because now you've got a way of encoding in a very logically consistent way. these theories about how past and even present humans behave and make decisions. And so now instead of the rich narrative of an ethnography or whatever, now we can have an agent based model. We can simulate these ideas and see if they generate the right dynamics. And, we've just recently seen the first textbook in archaeology that has agent based modeling in it that was, that's been published. Shameless shout out to Stephanie Crabtree and her co authors. We've got the work of Tim Kohler in there, no relation, unfortunately and all the work he's done. So that's new and interesting. There's also been an N of two, but that's a lot more than zero. of using agent based modeling to understand moral systems, which I think is absolutely fascinating and I'm a huge fan of. Of course, I'm super biased because I also tried to use agent based modeling to study legal institutions and how they function and their design,

John:

so one of the people I came across in the last year he's into computational law, and so I guess the idea of that is. If we look at policymaking, some of these policies are written in, and they might be going after outcomes A, B, and C. And, at best, they might get A right, they might get a little bit of B, C never happens, but they also get D, E, and F, and have to figure that out. Across your career other than agent based modeling, what else helps people try to see through those possible futures, or understand? Different unintended consequences. What other techniques could people be using?

Matt:

I'm obviously super biased, and I think everything should be an agent based model, and if it isn't, probably would be better if it was. But there, there certainly are lots of other tools and techniques out there. I'm willing to concede that, the toolbox is not just an agent based model, even if it should be. So I'm also a big fan of behavioral game theory and experimental game theory and experimental economics and things like that, where it explicitly starts to bring in that behavioral component. Canonical, say microeconomic theory with the, again, the canonical rational actor, those are some strong assumptions about human behavior. And now there are ways you can configure those sorts of analyses that I think start to really do get at how humans would work in those environments. And then especially if you explicitly add humans to it, you can obviously get a lot of very specifically human insights of it. And of course, what you find is that humans don't universally violate all of the assumptions of a rational actor. But since they're using heuristics to get there, it's that, there's path dependency. It's going to take time for the humans to get there. And, so there are certainly other techniques and things other than agent based model that let you start to gain those sorts of insights and to get a sense of how long it might take for the system to come to equilibrium. But I still think an agent based model, subject to a few caveats, is still probably a more efficient way to go after a lot of those questions.

John:

What are some of the things around agent, the field of interface modeling that has enabled it to become more applicable today than it has? In the past, the other example, seeing it be applied in new fields what are some of the drivers of that?

Matt:

I think a big part of it is that hardware and software is catching up with data is obviously getting ridiculous these days too, but we've always had a lot of data out there, or maybe even. More than data, I would say we've had a lot of data and we've had a lot of what I'll call sort of informal formulations out there. Pick on anthropology again, an ethnography in many ways, you could think of as a formulation for an agent based model. You've got a whole bunch of agents that someone spent a lot of time thinking about and perhaps talking to and certainly writing down everything they saw and everything they could figure out. But it's just been the past, I don't know, 10 ish years, when hardware and software really caught up and allowed you to start to build agent based models at an appropriate scale. And scale, of course, is a loaded term, which we can talk about, but it's gotten much easier to build an agent based model. You don't have to be a C or Objective C or Java programmer anymore. You can use scripting languages. You don't need to, get right down on the metal and just, okay, it's going to be 7k per agent, and I've got, 12 megs of RAM I can devote to this. So now that we've got huge compute power and much easier to use software, now it can just explode. It's now we've got Excel. It's much easier to use and much easier to do math with.

John:

Do you ever see a point where agent based modeling becomes as common speak as people saying, Oh, we just use machine learning to solve this. And there's this almost blind trust and Oh it's a machine learning model. Okay, there must be something right in there and then move on.

Matt:

Yeah, that's the I'm pathologically optimistic. So I think, yeah, of course that's going to happen. And we have probably next week sometime. Let's do this again next week.

John:

What conditions might need to happen or play out to enable that to happen?

Matt:

I think we're on the cusp. So there's growing work on, I'll use the term analytic agent based modeling. Another umbrella term that Josh Epstein coined is inverse generative social science. And so these are techniques that are trying to go more directly from observational data. to an agent based model, or from a collection of potential behaviors to a whole collection of agent based models that behave in some manner that is appropriate, however that's been defined. And as those techniques mature, then I think that's what'll, that will enable the ubiquity of agent based modeling, because now we've got tons of sensors in the environment. That data can be gathered. We can start to infer an agent based model from the data, and then we can start to have kind of a Cambrian explosion of potential models that all, to some degree or another, explain the behavior that we're that explain that are potential generating mechanisms for the patterns that we're finding in the data. And then, that's where the humans can come in and say these 700 models, no one can think that quickly and, process that much. So those are not right. This one is too simple. We know people do more than just be phototaxic. So that whole family of models is wrong. But these, I don't know, we'll say 75 in the middle, these are all plausible. Now we can go talk to psychologists, social psychologists, whoever, anthropologists, and try to figure out, okay, what are the commonalities of these things, do, could we come up with an experiment to differentiate between these two things as well,. The point is, we can then go from observational data much more quickly to a collection of plausible agent based models.

John:

is that going to make building agent based models more or less fun, do you think?

Matt:

It'll make it different. It's gonna shift the trouble. Right now, a lot of time is spent, how do I articulate this rule for an agent? That A will run in a reasonable period of time, and B doesn't do too much violence to how humans actually work and so you puzzle through that, you finally get that done. But then of course, you're still subject to the okay, that's a fun rule, but what about the 7, 000 other rules that might possibly work in that environment? Okay, fine, fair enough. So we've got the IGSS business to try to help puzzle through that the fun part, though, I think would be The surprise, my favorite example would be like, Oh, your agent based models are too simple. There's lots of ways you can be surprised by seemingly simple agents when there's enough of them and in the right environment, Like traffic's a great one. Humans are doing all kinds of amazing three dimensional image processing and trajectory calculations, and they're spilling coffee on themselves and they're adjusting the radio as the car moves forward. You don't have to model all of that to get plausible traffic dynamics. You basically go the speed you want to go, don't crash into the thing that's in front of you, and try to get around it if it's going too slow. With just, simple heuristics like that you get very realistic traffic dynamics. But I think I think it's going to be exciting, the bottom line. I can't wait to see all the bizarre collections of simple behaviors that generate the same, the right structure in the data. That's going to be craziness.

John:

Yeah. Absolutely. This kind of gets into the purpose of building these models is to get to a point of understanding, and that understanding is usually to make You know, a more informed decision or as best informed as you can so has there been a really good application of agent based modeling that you've seen when applied does help us get there? And I'm going to start here and then I'm going to go to the, what the flux question after.

Matt:

All right. Yeah That is a super interesting question, actually. It shouldn't be. It should be. Oh yeah, of course. It was so and so Amsterdam whatever something but I'm really not aware of an actual sort of high profile killer app in the ABM space and I Arguably, that's one reason why it's taken so long to get traction, because it hasn't just hit a home run. It's an obvious home run. Most traffic models are now at least entity based, and we can quibble if they're actually agent based, but they're certainly entity based. And, and they work really well, and we're doing a much better job of designing roads, but That's not exciting. Everyone's still stuck in traffic in a commute. And so the idea that ABMs help there maybe that's a bit of a head scratcher. Not sure I'm really feeling the love here from the ABMs. I'm sitting in six lanes of stopped traffic. Know that I've really seen a, just a home run example. There have been a lot Of super exciting, in the nerdiest sense of the word examples that have been out there the early work at SFI on artificial stock markets, Blake LeBaron and the gang where you got the, a nice collection of the statistical sort of facts we'll say of a stock market, much more simply. when it was an agent based model, then using a, canonical sort of efficient market hypothesis, microeconomic model, with lots of weird assumptions about how news gets injected. It was way simpler with an ABM. That's cool. You've got the Axtell and Epstein example of retirement age and how that took a long time for society to actually re equilibrate to, not because Congress passed a dumb law. But because humans are messy, and it took a while for the social system to finally say click, okay, and now we're in a new regime and we've, endogenized the retirement age. And so there are lots of sort of simple examples like that. I shouldn't say simple, but lots of, sort of small examples like that in lots of different fields, but nothing's really been, I was hopeful that COVID actually would be the killer app for ABMs. And certainly they were used widely, but again, it never really percolated, I think, into the general psyche that, Oh, they're using a new modeling technique to understand what's going on in these dynamics. It got used widely, but it was under the radar. And honestly, maybe that's okay. It doesn't have to be. I don't know, big and splashy, in my opinion. As long as it's helpful that's my metric. That's what I really am going for. I don't know. I guess aiming low should hold myself to a higher standard.

John:

A lot of people look at the process of building database models and, even when we start looking at who's working on what, a lot of the times we ask the question, how do we know if a model is successful or not? And we get a lot of different answers. And if some people came with the answer, if it ends up being helpful, I would argue that's a pretty good one. Other times we get, if it's accurate. Or if it has an impact and all these things come down to ultimately helping making a decision or helping learn something. What other, when you, when you were interviewing people and others that you've worked with in the past, when you have asked the question, what makes this model useful? What are other answers that you've gotten?

Matt:

I bundle all of that stuff up as the insights generated. And, because it goes by lots of different, it's useful. It helps me make a decision better. But. Oftentimes, to your point, better? What does that mean? And how will we even know? You're going to make a decision one way or the other. So you chose eight instead of six. We're never going to go, we're never going to rerun the tape with the other option and see if it was better. I just I really like the way you phrased it. Just was it helpful? Because sometimes, You never even hit go on the model, but it could still be helpful because it forced a set of stakeholders to get in the room together and express ideas in a common framework. And so that has generated more understanding among those stakeholders. And so even if the model never gets run, I would argue that producing the ABM was useful.

John:

Oh, yeah. This is where Josh goes on. Josh Epstein goes on and says, the why model paper that we pretty much send that out to anybody that we're working with to your point, like we're going to get a lot of insights. We're going to learn things. We're going to get possibly closer to consensus on some stuff just by going through that process.

Matt:

We've had some nice examples, over the years of this group, they don't know if they want to play, and they don't want to get involved. I'm like, okay, no problem. We'll make a whole bunch of assumptions about their chunk of this, and then we'll send them the results. And then we typically get a phone call about, what are you doing? This is all wrong. Our stuff is more like X or Y. I'm like, okay, great. Everybody plays, everybody wins.

John:

I'm gonna go back to when you've said it's not like we're going to rewind the tape, but under the idea that and, are what the flux moment if there was a decision that you were involved with or you're familiar with. Or you think if people could rewind the tape and ask some, what if questions run those simulations, where do you think that a decision could have been made that had fewer unintended negative consequences by using simulation?

Matt:

History is full of things that you know, if only we had an ABM that would have solved this or that problem. Maybe in the more recent past that's been pretty well articulated by folks way smarter than me, would be like the 2008 financial crisis, hiccup, whatever. Especially Doan Farmer, Rob Axtell I'm blanking on his name, Gennacopoulos, John Gennacopoulos? I hope that's his first name. Okay. And many others. It's not just them either. There are also a whole slew of folks looking at network network structure during that whole period as well. That I think the way they articulated it was pretty good and I'm probably going to do a terrible job of expressing it. But, basically, Genocopilist does a great job of arguing, is this leverage cycle? And it was also, banks were starting to diversify in similar ways. And so now rather than a couple of banks being subject to, we'll say subprime mortgage backed securities, now everyone was, but no one really understood What would be interesting in that circumstance is if we had an ABM of the banking sector, we'll say, we're picking on them, but it's obviously much broader of the economy. How about that? If we had a good ABM of the economy, then we could actually, and by we, I mean that very broadly, we could be running these kinds of scenarios and starting to understand what the impact is of what Bear Stearns going under and things like that. And, but more importantly, individual banks, assuming they had access to these sorts of tools as well, they could also be, running these kinds of simulations and you're like, okay, it's not too big of a deal, right? Going bankrupt or whatever they call what happened to them. It's not a huge economy It's really not gonna make that big of an impact It's a huge problem for Iceland and they need help and all that jazz But in the grand scheme of things, you know There are bigger things that could be going on and so an ABM can help start to filter out This is noise, and this is a big deal that we really need to be concerned about. And so I think that could help dramatically.

John:

As you're going through that, there are two books that come to mind. One that's called the end of theory, by Richard Bookstaber. There's that. And then Dwayne just came out with a new book that has not dropped in the U S yet I think it's getting the UK and Europe get, about a quarter's worth of advantage over us in America on this one. and yeah, so it goes, but both are about economics being ripe for agent based modeling. So if you had to tie some of the. Either micro or macro economic theories and we're behavioral economics or what is the right field of economics that you've seen the most amenable or most open minded to starting to use agent based modeling?

Matt:

So I'm speaking a little bit out of turn because I'm not an economist. But I would say experimental and behavioral economics simply because they've already embraced the behavioral component. And in so doing, they're already trying to better understand some of the assumptions that underlie human behavior. And at least in my eyes, means they're relaxing some of the assumptions associated with a very canonical rational actor. And so again, I would argue based on like Rob Axtell's work back when he was at Brookings and the bus, Papa Demetrio, Stiklitz paper allegory to chaos. I think under those circumstances, an ABM is just the most logical. way to start to embrace that kind of an analytic framework.

John:

One of our, areas of thinking is being that, consumer behavior is changing more and more rapidly. That might be, one of the things that are, the end of some of the economic theories because, there's newer technologies today that are swaying sentiment and swaying our behaviors more quickly than ever before in the past. Therefore, trying to understand it from a bottom up perspective with an ABM is possibly one of the only ways to get ahead of some of those bigger trends.

Matt:

Yeah, I'm biased, but I certainly agree, especially, what's particularly fascinating now is that, is it that the theory is wrong? Is it that humans are changing the way they behave? Or is it just that we're being buffeted by different forces? And, that's the nice thing about an ABM is that it lets you start to think about all that and think it through and say, okay, if, humans behave the same way they did 10, 000 years ago, their brain isn't a whole lot different, but now it's subjected to huge amounts more information and it can be filtered in very specific ways. And because it's so broad now, we can really find some specific niches. What does that do to how humans work? Even if it's the exact same sort of rules and things that they're using. And anyway, these are the sort of fascinating questions that ABMs let you really get at that I think is just makes them super exciting.

John:

Understanding the conditions, that's how we talk about it with anyone that we're working with one set of conditions leads to, a set of outcomes. If those are changed or different. What might happen and, the ability to play out those scenarios of what we found get people curious. And then when they start actually seeing data and it's plausible, that's what then says, Oh my gosh, this is a new way of thinking I'm in. So Matt, being the optimist that you are two final questions. One. Curiosity seems to be a major driver of why a lot of people get into complex systems and agent based modeling. What are some of the easy on ramps into agent based modeling, if somebody has not been exposed to it, but is curious about it?

Matt:

There's a great set of lectures on Complexly Explorer that are they were put together by Bill Rand. Rand and Walensky also have a great textbook in this space. If you're coming more from the physic so if you're coming from the behavioral or physics side of things, then definitely I think the Rand Walensky textbook is probably a great place to start. If you're coming in more from biology and especially ecology, then you might want to pick up the Rails Back and Grimm book because that has more of a ecology sort of flavor to it. I'm also still, especially if you're coming from the behavioral sciences, I'm a huge fan of the Epstein Axtell Growing Artificial Societies. And the reason I'm such a fan of that is A, it's short. I don't have much of an attention span. And B, it goes through a little bit of the philosophy of using an ABM in the beginning, and then it goes through the whole process of building Sugarscape. And we can quibble about Sugarscape and the conclusions drawn from it and all that jazz. But just having that soup to nuts discussion of What led them to an ABM and then how they developed it, I think, is super insightful. And if you're looking for a very fast read, very basic introduction, then I'd go after Turtles, Termites, and Traffic Jams by at MIT. I'm not gonna remember. Anyway, and then of course, the ubiquitous shout out to NetLogo. That's a great. The general, the basic install of NetLogo comes with hundreds of sample models that are well documented and fun to play with. It's an easy scripting language to use. This is not an endorsement of NetLogo. I should probably say that. But it is awfully easy to use and it's free. It's a great way to start.

John:

Last question here in your role as president of The Computational Social Science Society of the Americas, anything that is on the leading edge that you've seen in the last couple of years, I know you have the conference coming up, you have some lecture series what should people be paying attention to C triple S a wise and, anything else that you'd want to shout out in the ABM community,

Matt:

I'm very excited about the whole field of computational social science right now. We, it's, it is, I would argue it is the time to jump in. The data is ubiquitous, the compute is big enough to keep up with it, and now we've got, agent based modeling software that can let you start to Express, hypothesize, generating mechanisms very easily and start to run and gain insight and see what works and what doesn't very quickly. I think the things to really pay attention to these days are the amazingly seamless integration of data with agent based modeling frameworks. Now be it geospatial or synthetic population, all that jazz, there are more and more platforms out there that can support it. And that, and there are more and more libraries and datasets out there that you can just suck right in and start using. And that helps, depending on your question, that can help with the creation of the model that you need to start answering those sorts of questions and generating those sorts of insights. I also, I'm pretty excited about the promise of inverse generative social science. Right now it's still in its infancy. But there's a lot of space for growth here. And if it, if that's a nut we can crack, that will dramatically change the utility of agent based modeling, I would argue. Because now we can actually start to get to robust policies. And we can relax this optimality assumption where we have to pick, okay, this is our population, and now we're going to do a parameter sweep and we're going to come up with the best combo of things. Now we can say these are all of the plausible populations and these are all of the plausible parameter space combinations for the simulation. And the this, we'll see, is a robust policy that improves things across All of it, as opposed to a fragile optimal policy that's premised on our assumption that we got the population right, which we know we didn't. I would, I'd be particularly paying attention to the integration of data and simulation, as well as the ability to broaden the scope of the simulation itself. I think that's, those are going to fundamentally alter. The utility of agent based models. I'm sure I'll be proven wrong, but you never know.

People on this episode