The Flux by Epistemix
Welcome to The Flux - where we talk data, decisions, and stories of people asking the what-if questions to create an intentional impact on the future.
The Flux by Epistemix
The Future of Agent-Based Modeling: Insights from Josh Epstein
In the inaugural episode of The Flux, John Cordier, CEO of Epistemix, interviews Josh Epstein, Director of the Agent-Based Modeling Lab at NYU and a prominent figure at the Santa Fe Institute. The discussion revolves around the potential and progress of agent-based modeling (ABM), particularly in public health, economics, and beyond.
Epstein shares insights into how ABM creates simulations of artificial societies to generate real-world patterns like epidemics or wealth distributions. He highlights the generative explanatory standard, which uses individual agent behaviors to explain macroscopic phenomena, contrasting it with traditional models that assume rational actors. Epstein introduces Agent Zero, a framework offering an alternative to the rational actor model by incorporating emotional, cognitive, and social elements into agent behavior. His work emphasizes the need for cognitively plausible agents in simulations, moving beyond simplified assumptions in fields like epidemiology and economics.
This episode underscores the transformative potential of ABM in areas such as public policy, disease modeling, and economic forecasting, stressing the importance of integrating human behavior into models to improve decision-making. Epstein envisions ABM becoming a core tool in tackling complex, real-world problems, with the field poised for continued growth as computational power advances.
Timestamps
00:00 Introduction to The Flux Podcast
00:23 Meet Josh Epstein: Pioneer in Agent-Based Modeling
02:32 Exploring Agent-Based Models
04:59 Generative Social Science and Agent Zero
07:31 Inverse Generative Social Science
15:58 Behavioral Dynamics in Epidemiology
19:00 Leadership and Decision Making in Modeling
19:33 Future of Agent-Based Modeling
20:35 Global Adoption of Agent-Based Models
22:34 Visualization and Pandemic Modeling
33:46 Why Model?
37:34 Optimism for the Future
All right. Welcome to our inaugural episode of the flux. So on this podcast, we're going to be talking data decisions and the stories of people asking the what if questions to create an intentional impact on the future. In each of our episodes, we hope you can turn lessons from our guests into foresight so you can uncover how you can create an intentional impact on the future. Today, our first guest of the podcast is Josh Epstein.
Josh:Epstein.
John:He's director of the Agent Based Modeling Lab at NYU and a longstanding member of the Santa Fe Institute. Josh, welcome onto the show.
Josh:show. Thank you, John. Pleasure to be here. Honored to be the inaugural speaker.
John:Absolutely. It's, it'd be difficult to kick off an agent based modeling oriented podcast and not have you on. and we're glad to have you be the first guest. So why don't you tell our listeners a little bit about your current roles and some of the things you've been working on recently.
Josh:as you mentioned, I'm director of the agent based modeling lab at NYU, which includes people from all over the world and all over the university. And it's meant to be the kind of crossroads for agent based computational modeling. in public health, but in many other fields too. Economics, migration, violence, what have you. And I have appointments at the Courant Institute for Mathematical Sciences and the Department of Politics and, as I say, affiliates with people in neuroscience and psychology and economics and politics and external faculty. it's a kind of movable feast, with lots of participation. And I just completed a sabbatical where I was in residence, at a series of universities. With the intent of firming up these connections and make it a more formal constellation of leading centers. One of which was Doan Farmer and Eric Beinhacker's Institute for New Economic Thinking at Oxford. The Turing Institute in London, the Alan Turing Institute in London, the Vienna Complexity Hub. In Vienna and at the, called, superior for, social sciences in Paris and at the University of Milan and in Piza at the Superior. and the point was really to spread the word, show some work that I've been doing, and also talk about, how do we make a more concrete constellation? of places doing agent based computational modeling, really worldwide. And, and I'm very happy at the energy and, I think we're really rolling to be honest. so that's what I've been up to most recently. I guess over the longer term, I think what I've been up to really is five fold, and this is not the order in which I came to these things, but this is the, I think the main things I've really been up to are, first, to articulate a particular explanatory standard for the social, behavioral, and health sciences that I call the generative explanatory standard. And for people who are unfamiliar with agent based modeling, agent based models, Typical agent based model is an artificial society. of software individuals. The software individuals all interact on some landscape. It might be a real physical landscape, or a network, or inside an organization, or what have you. But individual agents are software people. They interact with one another, and these interactions change their state. It might change their wealth or their health status or their, beliefs. and it also might change the environment, might change the landscape itself. And when we have a macroscopic pattern, like a segregation pattern, a wealth distribution, epidemic trajectory, what I mean by an explanation is, an account. where individual agents generate that macroscopic regularity from the bottom up through agent interactions. Alright, and in that case, the micro world, the world of the agents and their interactions, is the generative, explanation of the macroscopic regularity. And I've written a lot on this idea of generative social science, and there's lots of important qualifications. to say that this micro specification generates segregation is not to say That it's the only way to generate segregation. And there might be others. And merely to generate it, you could generate it in a way that's really absurd and has nothing to do with people. So mere generation is not what I mean. And there might be multiple generators. But it's a very different idea. And I should say that the other thing is I want to generate the regularity from the bottom up in populations of cognitively plausible individuals. And, this is a whole other thrust of the work. which is to provide alternatives to the rational actor model, which assumes high levels of information, high cognitive functioning, mathematical abilities, dispassionate rationality, all sorts of other things. And one effort, called Agent Zero, is an attempt to give a formal mathematical alternative to the rational actor that generates some of the phenomena we care about, from genocide, to health behaviors, to opinion dynamics, to all sorts of things. So first, contribution, perhaps, is this generative explanation. what do we mean by an explanation? I want to change that in the social sciences. Second contribution would be, yes, of course, I've part, I've been one of the people implicated in the development of the scientific instrument of generative social science, which is the agent based computational model. So what's the epistemology, What's the scientific instrument? And then, many applications from the work, Don Burke and I did on smallpox, to civil violence, to, economic dynamics, a million applications, reconstruction of ancient civilizations, the Anasazi, lots and lots of different applications of the scientific instrument to give generative explanations of these phenomena. And as I say, A major thrust has been the development of a formal alternative to the rational actor. there's a lot of, evidence against the rational actor. There are a lot of experiments. Kahneman, Tversky, Paul Slovic, all sorts of people have shown experimentally that the rational actor, axioms are routinely violated by humans. But, there's Mere demonstrations that it doesn't work, doesn't really displace it as a method. You gotta have an alternative and the method, the alternative has to be formal. crude, provisional, what have you. Agent zero is a formal alternative. I think it can be extended. I think it can be all of these things. the strength is that it produces testable hypotheses. It could be falsified in principle, which is fine by me. the idea was that build an agent, a cognitively plausible agent, has emotions. Yes, it deliberates, but it's bounded rationality, and it's connected to other agents who are driven by emotions and have bounded rationality. And when you put a lot of these creatures together, you're able to generate collective dynamics that are very far from optimal. and very recognizable. so then, and then finally, the last thing I've been up to is this idea of inverse generative social science. And I'm co editor, lead editor, of a collection in the Journal of Artificial Societies and Social Simulation called Inverse Generative Social Science. And here, what we've typically done, and what I did with Agent Zero, and what we've done with many models is, you Design an agent with all their rules and internal regulations and all of these features, and you put a bunch of'em together and see if you can grow the target wealth, distribution, segregation, epidemic dynamics, what have you, and inverse generative social science really stands that on its head. Instead of designing agents and senior, if you can grow the data, you start with the target data itself and evolve agents. that are highly fit, meaning agents that have high generative capacity. So instead of designing agents, you just begin with a primordial soup of agent constituents. And combinators. The combinators could be mathematical, they could be logical, they could be if then statements, they could be, they could be nested. There's all sorts of things that, how nature would combine those, right? And the idea is you start with the primordial soup and you begin to have these agents, which are these constituents, which of course are chunks of code. You have them combine and mutate and have offspring. And eventually you get a population of agent architectures that are able to generate the phenomenon. And here I come back to this point about multiple generators. The inverse generative method produces a family of agents that generate the target dynamics. And then of course, that's an embarrassment of riches. And we need to sort those out by collecting appropriate data or designing appropriate experiments or doing what you do in any science. where you have competing hypotheses. So that is, not quite in a nutshell, but that's how I think what I've been up to.
John:Oh, absolutely. I think when you look at people who just get introduced to agent based modeling, one of your lines that everybody knows about, if you can't grow it, you don't understand it. And so all these different types of generators and what you're getting to with the inverse generative social science is you can do this goal seeking exercise. But then there is a person that comes in and says, you know what,
Josh:you know what,
John:this set of things, it doesn't totally make sense. Maybe this set does. Or we need to get some more information to understand it.
Josh:it. Absolutely. again, absolutely right. And we can and should discuss the phrase cognitive plausibility. I'm saying you've got to grow it in a population of cognitively plausible agents. And I depart from Hume, who said, Reason is a slave to the passions. And, Locke and Aristotle, Man is a social animal. So I thought, look, can we get something where there's a primitive emotional module, and I used fear. Because we know a lot about the neuroscience of fear. And I've consulted a lot of neuroscientists about this. And I thought, let's endow the agent with an emotional module based on what we know about fear, because it's a very, it's crucially important in areas like vaccine refusal, flight from disasters, violence, a million things, financial panics, and I used for agent zero, a very, A crude but very reputable model called the Raskorla Wagner model of fear learning. There are many alternatives and I mentioned many of them. And of course, by all means, let's play with alternatives and so forth. But this is a crude, credible representation of fear learning in vertebrates, including humans. Thank you. And then I said, okay, that's part of the story, but people do deliberate. They do take in information and they make some sort of appraisal of risk. And the fear is not necessarily conscious either. You don't necessarily know you're acquiring a fear association. but you do, and we can come back to all of this, but the first module is an emotional and affective module. And the second was a boundedly rational, deliberative module where agents, for example, might be moving around the space and they take a statistical estimate of, how many attackers are in my neighborhood. And they make an estimate based on the relative frequency of attackers over the population within their vision. And they add those and they form some overall disposition to retaliate. and, but they're also influenced by the fear and bounded deliberations of other actors. And we think of them as being in networks. Actually, they construct their own network and disassemble their own network. but that's the basic idea. And if their total disposition exceeds some level, then they retaliate on the attackers. and you can get wonderful results where. There's a region where agents are under attack and there's one agent who's never been attacked and by this process of contagion, dispositional contagion, he wipes out his village even though he's never been attacked and he can even lead. the massacre or the financial panic when he's never been subjected to any direct threat. it, it puts the whole idea of leadership in a funny context, which I think is productive. Anyway, in the book, there are many exercises, the Arab spring jury behavior. all sorts of, again, I call them computational parables because the book is a theoretical exposition and I don't purport to test the thing against data, although there's several lines of work underway to do that, including financial panics, actually
John:we're trying to get agent based modeling out there into the world. There's this tension Where looking at the world through cause cognitively plausible Agents that aren't just absolute rational actors, because we know that's not true,
Josh:right?
John:There's the field of economics. There's the field of actuarial sciences. Where do you feel like the tension comes from, And then where can you mitigate the tension to get something like agent based modeling adopted more easily?
Josh:I guess my first answer is I don't mind tension. if there's tension. I think the ultimate, the arbiter will be the data. The arbiter will be the scientific, productivity of these approaches in their different contexts. And in some contexts, the rational actor is fine. if you're interested in, consumer behavior, perhaps, or other spheres of economic life, then, it might be fine there. In some settings, game theory might be fine, and in some settings, other things might be fine. So we have to be, and, and it depends on the question. Am I trying to simply predict the next stock price? In which case, I might just use a black box, machine learning, LLM, where there's no account of what individuals are doing. It just gives me an idea. Next stock price and if I care about making money in the stock market, maybe I don't care about explanation generative explanation I just care about predicting so in that context, maybe Explanation is moot if I care about by what cognitive mechanisms do people end up voting, against their own interests, joining, genocides against innocent people? Then, I don't think the rational actor is a productive, a very fertile method. so it depends on the question, and where there's data, it depends on the data. I used to say, and still say, if you say you've got a better mousetrap, show me some mice. Mice are the real data, and if standard economics can't give a good explanation, an account of financial contagions, and this one can, then I say, look, the data will decide. It's not about tension. It's not about consensus. It's about the damn data. Are we doing science, or are we not? And if we're doing science, tension, schmention, it'll, the real selection pressure, is the data. scientific accuracy and results. And as, Planck said, Science progresses funeral by funeral. And, in some areas, I am occasionally reminded of that remark. and I'd be lying if I didn't say economics was one of those. but there's also tension with classical epidemiology. Where people routinely still assume perfect mixing, no behavioral adaptation. I think this is a huge topic in epidemiology and in public health. you have behaviors like, vaccine refusal. You have, people, understating, the level of, I think there's a lot of risk associated with we had COVID recently. distrust in the government. There are a million things that affect behavior and we need to try to include them and, again in a cognitively plausible way. and I think a lot of classical epidemiologist would say, look, behavior is too hard. we're not including it and I'm saying no, you are You're including it whether you like it or not. When you say that people keep mixing, despite a plague in their town, that's a strong behavioral assumption. So when you say we're not including it, yeah, you are including it in a way that's obviously implausible. so is there some simple way to include it? And I've done quite a lot of work, mathematical work, in a field called coupled contagion dynamics of fear and disease, where you have an epidemic, where you Spreading, but you also have fear of the epidemic spreading, and the fear of the epidemic can lead people to self isolate, get out of circulation, but then their fear dies off. What happens is, the epidemic comes along, there's a, that produces a spike of fear. Spike of fear leads people to get out of circulation. That reduces the growth rate of the epidemic, and people lose their fear, and they come out of isolation, but they come out while they're still infectives in the community, and it produces multiple waves. Of epidemics, a behavioral mechanism for multiple waves that is not available in the standard well mixed picture. You can't get multiple waves out of the famous SIR basic models. and then we extended that recently, Erez Hatna and I and others, to a triple contagion model where there's a contagious disease, there's contagious fear of the disease, and there's also contagious fear of the control, like vaccine. And we know that this produced historical cycles. of, of smallpox, for example, and many other diseases. So I think it's very, it's critical, in public health and in infectious disease modeling and in turn in control of infectious disease that we include behavior and we design messages that moderate behaviors that are injurious. And I don't think you can do that unless your models include behavior in the first place. Absolutely. So I think all of this is a huge watershed. And I think things like social media mining and, and the rest of it are going to be very useful in ferreting out what people are really feeling, what they're really saying, what they really believe, and all of this. So I think it's a watershed, the combination of agents and behavior and big data and the rest of it. I think we're on, it's tremendous advances are coming.
John:one of the things you mentioned earlier very briefly was leadership. and that infers decision making from the use of these types of models and getting to a deeper level of understanding. and you've probably seen plenty of decisions where, give me the prediction and I'm just going to press go on what we think is going to lead us to get there. Rather than let's try to get to this, a deeper level of understanding. Yeah. there's If there's a triple epidemic going on, under these outcomes that are going to impact public health agencies, hospitals, the pharmaceutical industry, the economy.
Josh:Yeah, all of
John:let's start with understanding to see how these pieces come together. How do you think agent based modeling, 20 years from now might change the way that public policy or leaders of large corporations are going to need to start thinking differently?
Josh:Honestly, I think it's, predictions are hard, especially about the future, as they say. But, I think there's a good chance that it'll really revolutionize a lot of this. And I also think it's very important what's happening with AI, and, and the effect of AI on labor forces, productivity, the displacement of labor, all sorts of things. but yeah, I think, look, my colleague, Rob Axton, my colleague and longtime friend and others have built, full scale AI. artificial economies. And, they've been used by the Bank of England. They've been, they're being used by the Canadian government. They are, taken with, taken seriously in Europe. I think it's already happening that large scale agent modeling is informing decision making at very high levels. And Don Burke and I, we had the, we, We had access to very high level decision makers and they were very interested in our agent based modeling for smallpox containment. And then we also, that group, the smallpox working group that, that Don and I founded with DA Henderson, that led to this huge MIDAS NIH network, worldwide network of modeling teams, including agent based modeling. I think in fields like epidemiology, it's already Had, a rather transformative effect in economics. I think it's having that effect, although not all economists recognize it, but economics is changing even if the economists aren't. And, And I think there's a whole, field of migration, all sorts of areas. Geographical information systems are completely changing the way we think about coupled behavioral and environmental dynamics. climate change is obviously an area where it's crucially important to understand the, climate denial and the sort of, what information are people getting? And what would make them believe it and act on it? I think these are like kind of life and death things. Pandemic behavior, financial crashes, climate change. we have to do better. And these are coupled systems. And I'm, it may be that there are very elegant, low dimensional equations that capture all the complexity we're interested in, but I wonder whether that's the case, and even though there's a predilection toward analytical models and beautiful mathematical expressions that I, of course, share, I think for challenges like this, agent based computational modeling with high performance computing and large data, is very likely to have, a more important effect on how people see these processes and react to them, and preventively also. And the visualization part is a huge deal, honestly. when we built a planetary scale agent based model years ago, John Parker was the real software engineer and, that had 6 billion agents moving around a planetary map over time and you could see how things might develop under various assumptions that people do nothing, that people do something, that flight restrictions are imposed and so forth. And we'll use these to think about Ebola and. Pandemic flu and all these other ideas. And again, they're highly imperfect. but they bring into view a coherent approach. That's just not possible without modeling. you can't just have a tape, a round table of doctors asking, where do I think Ebola might spread? Or when is it too late to close schools? Or when should I lift a quarantine? Or when should I impose international flight restrictions? you just have to have models. And agent based models are, have the advantage that you can see what's going on. Don and I, Don Burke and I had great success, with the smallpox working group because we could work with doctors and medical experts and ask them, we could say, look, once you are really in the late stages of smallpox, do you move around at all? And we had DA Henderson and other actual medical experts say, you're out, you're on your, you're on your back after day three. okay, that's a modeling assumption we'll use. And then when you present it to them, we say, look, this is informed by these domain experts and they're bought in. So you build the model with them, which is another wonderful thing about agent based modeling is that it's, it facilitates participatory modeling. And Don and I had this great experience with medical experts, but I've also had the same experience with archeologists doing reconstructions of ancient civilizations. And when they, when the experts help you build the model, and they understand everything that's going on, they might be willing to take risks based on counterintuitive results that they won't take. If you just have an impenetrable thousands of differential equations and you say, remarkably, they say you should do this unexpected thing. politicians are risk averse, policy makers are risk averse, and the agent based model cuts through that by its transparency and participatory potential.
John:That's a really good point. So one of the things we're gonna try to do on each of these episodes is talk about a, what the flux type moment. So using the riff from back to the future and the flux capacitor, if you could start in the past and then run a simulation forward, has there been a decision where agent-based monolith has not been used? but in your mind you're like, this was such an obvious example. if we were able to run a simulation, we might've been able to avoid. A negative downstream event or create a more positive upstream
Josh:upstream event. I think it might have been helpful in a million spheres over, over many years. I think understanding things like mask refusal, vaccine refusal, distrust of public health authorities, acceptance of crazy theories, drinking bleach, things like this. If we understand that, I think it would save lots of lives. we wouldn't underrate the threat and we wouldn't promote, dangerous, approaches, and we'd enforce what we know to be good approaches, so I think in a thing like COVID, it really, would have been, very instrumental in avoiding a lot of death and other, morbidity. It also helps you just envision possible futures in a realistic way that, that, might not be possible. without a model.
John:So Josh, if you go back to your beginning of getting an agent based modeling,
Josh:modeling,
John:what were you doing at the time? What was your aha moment? What did you
Josh:What did you do next? okay. so I was at, I was doing mathematical modeling, nonlinear dynamical systems. and, the way it all started was really that I had one of the early, people at Santa Fe and, Murray Gell Mann. my Nobel laureate in physics and the late Murray Gell Mann was a good friend and George Cowan and John Steinbruner and others of us, Murray had this idea. Let's take a crude look at the whole, which is out to 2050 and think about sustainability to the year 2050. This was probably in 19, I don't know what, early nineties or something. so I remember we all went to the MacArthur foundation in Chicago. I don't know. And we said, look, there's all this complexity, there's all this huge global dynamics and so forth. Let's take a crude look at the whole using models. And we were able to convince the MacArthur Foundation to go ahead and invest in, something called Project 2050, which was a collaboration between Santa Fe, the Brookings Institution, where I was, And the World Resources Institute. And we all merrily went back to our places and thought, Wow, this is, what a great opportunity this is. Nobody had the foggiest idea what we were actually going to do. And Murray called me, I remember this moment, and my wife remembers it well also. We were in our little place in Washington, and Murray asked me, Would you please direct, group. And I felt completely anointed and accepted with glee, still not knowing what the heck we were going to do, and Murray didn't know either. but I was, I gave a talk at Carnegie Mellon. It was the first time Herb Simon gave a talk and I gave a talk and it was the first time I met Simon and I gave this talk. It was on arms, race dynamics and collective security arrangements. And it was differential equations, difference equations. And I went back to Brookings and I got a call from this kid who'd been at the talk. The kid was Rob Axel, of course, and he said, I went to your talk. And I couldn't replicate the results. And I thought, great, here's a kid who's got the technical chops to implement the thing. And the chutzpah to tell me he couldn't replicate it. So I thought, why don't you come to Brookings and we'll sort it all out. He came to Brookings and we did sort it all out. And, And we also talked about a million other things. We were kindred spirits on a lot of things like economics and so forth. Computing, all of this stuff. And it was just a great, really exciting thing. And I raced into John Steinbrenner's office, our boss at Brookings. And I said, Hey, this guy's great. He's great. And, and John said, why don't you hire him? Hire him on the 2050 project. I offered Rob a, a position and he took it. And there we were, the theoretical group. And we still had no idea what we were going to do. But one day in the Brookings Cafeteria, everybody had left. And Rob and I were there. And on napkins in the Brookings Cafeteria, we sketched on these napkins. I can still see the ink dripping into the napkin. Rob has these napkins. A little hunter gatherer society, and we said, what's the simplest possible thing we could build, this way? And it was a simple landscape. With little agents running around harvesting this resource and we thought look there's a peak of some sort of resource And they can follow the gradient to upper to the higher levels of the peak and we decided later to call it a sugar scape and lo and behold the thing started producing more and more fascinating results And we were just really hooked and I to me and I may be for Rob to the aha moment was When we figured look agents are accumulating sugar You So at all times, there's a distribution of sugar wealth in this society. Does it look like anything real? So we built a histogram of wealth And it turned out to look it was a Pareto law just like the wealth distributions of all industrialized countries and I remember John Steinbrenner saying pointing to it and saying that's important and yeah That was the first time that We really thought you could do legitimate science with this thing. You could actually try to generate patterns and compare them to real world data. And if there was an aha moment about the empirical potential, the real scientific potential of the instrument, it was really that moment where we said, holy mackerel, that looks like a real thing. And then there have been all sorts of other real things that have been generated. But I think from a scientific standpoint, that might have been a really important threshold, that moment.
John:If you look at the backgrounds of people Who do really well, what would you recommend we'll look at, whether that's for an undergraduate program or other fields that you think, if you transitioned from one into agent
Josh:agent based modeling, people do really well? With the advent of programming languages, when we started there were no programming languages for agent based modeling at all. but NetLogo and a lot of other ones. And it's fantastic for teaching. It's fantastic to get students interested. but it's also, what I say is, it's made it very easy to do bad agent based modeling. you can easily do that LearnNet logo in a week and have fascinating landscapes of dot worlds and all sorts of other things. And I think that's great. And I think that level of modeling can be very illuminating, qualitatively. So I'm not disrespecting that. But I think to really be, a serious agent based computational scientist, you really have to know the relevant areas of mathematics. I think, agent based models, you're going to compare those to real world Outcomes. So you should know statistics and time series analysis and other things You should understand what do you know, you should understand equilibria and dynamics and Oscillations and other fundamental patterns in nature, right? And I think you should be, to the extent that we can get analytical results, formal results, and so forth. I think that's important to do. And I would say that many of the really leading practitioners, are from technical areas. Network science, engineering, mathematics, physics, mathematical social science, mathematical biology, and so forth. So I think really, the really scientific future. belongs to people who have all those skills, or teams that include people with those skills. which can also be fine, but but I think there's a, I think it's very important to, to command, enough mathematics to recognize that, hey, you know what, this might be idealized in this way, and analyzed in this way. And it could be statistics, it could be probability, it could be all kinds of things. But I think I think the leaders now and probably the leaders in the future, are going to be people who have very strong technical backgrounds. Not just in, in computing, by the way, but in other things. I think a lot of programming is going to be displaced but I don't think modeling is the same as programming. Programming is implementation of a model. The model is an act of imagination. And I don't see that being replaced.
John:I want to go to your one paper that everyone who gets an agent based modeling reads at one point in time. Why model? And when you look at inverse generative social science, advent of AI becoming more and more approachable for people, I think there's 17 points on the list. Are there any others that you
Josh:that you might want to weave in now? 16 reasons other than prediction to build a model. And again, that was meant to be liberating and encouraging and, give people the liberate people to just build models for all, there's all kinds of reasons. Don't be cowed by somebody who says, can you predict anything? It's maybe I can, maybe I can't, but there's a million other reasons to build a model, and I tried to go through lots of those, suggest dynamical analogies, call prevailing assumptions into question and mainly explain. and I tried to, that's the core distinction is between predicting. and explaining. And again, I, to me, I was most interested and remain most interested in this generative explanation and how agent based modeling can facilitate that. But to just go off on a tiny tangent about predict and explain, we can predict, I'm sorry, we can explain earthquakes and tsunamis. why do they happen? It's because of plate tectonics. But we can't predict either. Why are there new viruses? Because of mutation and selection and other understood processes. But we can't predict next year's flu strain. you might understand the mechanism. And not be able to predict. And I think agent based modeling makes it possible to identify generative mechanisms. And sometimes it's predictive, and sometimes they're not. But I, as an epistemological matter, I want to enforce this distinction between explaining, that is to say identifying why things happen, what are mechanisms, or even plausible mechanisms, and prediction, which might be mechanism free. that was one of the very kind of the main point of that article. And to give a, tough, hard nosed answer to some of the silly criticisms you get of English based modeling. So it was meant to be empowering to people, and I have the feeling it might just be for lots of people who read that paper. the other distinction I think is quite important is between, AI. I think there's a lot of AI, a lot of AI is displacing humans, and I think there's a lot of AI, a lot of AI is displacing humans. It's defeating humans in chess and other things, and it's emulating human behavior, but it isn't yet explaining human behavior. And my, I think the original sin is really Turing and the Turing test. so here's my Turing test. Behind the screen is a soprano and a perfect recording of the soprano. And let's say no human can tell the two apart. The recording is indistinguishable from the soprano. But the recording doesn't tell me anything about how humans vocalize, right? By what mechanism are humans producing vocalizations and so forth. it's because they, blowing wind across your vocal cords disturbs the medium, which produces waves that crash into somebody else's tympanum that you experience as sound, right? But mere emulation of the output Which the Soprano and the recording does emulate the output, but it doesn't identify the mechanism. And I think a lot of AI is focused on emulation and so forth without explanation. And I think inverse generative social sciences uses AI, evolutionary computing, genetic programming, all of these things, to produce explanatory generators. So I think it's an advance over Kind of minor bird work.
John:All right. final question. What gives you optimism about
Josh:about the field of agent based modeling into the future? I think all that's happened, it's exploded. the bibliometrics of it is decisive. it's been a huge increase, huge adoption, very successful applications, really good science. and I just think the future is incredibly bright for agent based modeling. The computing is awesome. the computing barriers are just disappearing, really, to very large scale modeling. And I think, we're going to be able to populate large scale models with cognitively plausible agents and compare the results to real data and do very foundational science that is, easily communicated to decision makers and to the relevant publics. So I think the future is dazzling for agent based modeling.