The Flux by Epistemix

Exploring Data-Driven Decision Making with Bill Rand

Epistemix Season 1 Episode 8

In this episode of 'The Flux,' we dive deep into data-driven decision making with Bill Rand from NC State University. Join us as Bill discusses his work in agent-based modeling, social media misinformation, and various innovative projects. Discover how agent-based modeling can provide insights into human behavior, business analytics, and even intergenerational trauma. Learn how different modeling approaches can help forecast outcomes, refine decision-making processes, and provide better policy and business strategies. Perfect for anyone interested in AI, data science, and the social sciences.

00:00 Introduction to the Flux Podcast

00:20 Meet Bill Rand: Data-Driven Decision Making

02:42 Applications of Agent-Based Modeling

31:21 What the Flux Moment

32:02 Reflections and Methodological Insights


John: [00:00:00] Welcome to The Flux, where we hear stories from people who have asked what if questions to better understand the world and talk about how data can help tell stories that impact decisions and create an intentional impact on the future. This is your host, John Courtier, CEO at Epistemics. In a world where the flux capacitor from Back to the Future does not yet exist, people have to make difficult decisions without always knowing how the future will play out.

Our guests are people who've taken risks, made decisions when uncertainty was high, and who have assisted decision makers. Transcripts provided by Transcription Outsourcing, LLC.

hey there. Welcome to the next episode of the flux, today, we're on with Bill Rand, whose work is currently mainly based out of NC State University, where he focuses on data driven decision making and the diffusion of information amongst a number of other things. Bills receive funding [00:01:00] from NSF, DARPA, Google, WPP, the Marketing Science Institute, and probably a number of others.

John: And it's published hundreds of articles in the field of agent based modeling. And for those who are already, indoctrinated with why that, agent based modeling should be, More and more of a thing for everyone, most people get familiar with bill, by reading the book or at least, in whatever course they were first exposed to agent based modeling through an introduction to agent based modeling that,you've worked on for a while and continue to do.

Make updates and improve upon. So Bill, welcome and glad to have you on 

Bill: Yeah. Yeah. Thanks for having me, john. For sure. I don't know if I have multiple hundreds of agent based modeling articles, but it's probably approaching 100 probably at this point. 

John: So to get started, why don't you tell,our listeners a little bit about your current role and some of the work that you're up to today. 

Bill: Yeah, sure. right now I, serve as the executive director for something called the Business Analytics Initiative at the Poole College of Management at NC State University. I'm also the [00:02:00] McLaughlin Distinguished Professor of Marketing and Analytics there as well. So a lot of my work goes into, figuring generally how to use data driven decision making to make more informed decisions for businesses and organizations.

The agent based modeling component, often plays a role there. primarily, one small area of my research deals a lot with social media, misinformation and disinformation. And I found the agent based modeling is a great platform for modeling that spread a lot of times. And that's right now, that's one of the major components, but I still also, Work is a little bit as just an agent based modeling expert.

And so I will partake in projects that aren't directly related to marketing and analytics a lot of times. And as I briefly mentioned before we started, I'm working on a project on intergenerational trauma and how that can be modeled using, ABM and also most recently working on a paper on loss aversion and the evolution [00:03:00] of loss aversion, using ABM as well.

spanning a big variety of interests, but generally focusing on information diffusion and, data driven decision making. 

 

John: on things like the diffusion of information, spread of information, disinformation, what are people doing before they switch over to start trying to understand these things through agent based modeling. And what's the value add of using an agent based model versus whatever they're trying to do right now? 

Bill: Yeah, that's a great question. And in fact, I think a lot of people don't even realize they're using agent based modeling in this space, but are actually using it. there's a lot of times computer scientists often approach a problem like information diffusion by just modeling as best they can.

And in fact, in many of those cases, it turns out that the approach they're using is what many people within ABM would call agent based modeling, where they're representing actual individuals making a decision about how they're going to pass on information or not. and Leaving that aside, what are they doing?

Other than that? a lot of times [00:04:00] they're looking at aggregate communications on social media platforms. So they're looking at what is the sentiment? What is the maybe polarization that's occurring in these platforms? And they're just tracing it in a time series at the aggregate level, but they're not really taking necessarily the individual into account.

now, When we start to think about how does agent based modeling help you with that, I think the answer becomes pretty obvious. a lot of social media spaces, a lot of social network spaces. There are certain individuals who may provide different roles, right? They may be,the broadcasters who reach a lot of different individuals and push content out there, or they may be the, the bridges, the connectors between two groups of individuals.

So it might be that one group is talking about one topic and other groups talking about another, and they help spread, the communication between those two groups. And sometimes it's a very positive ways, right? Like they are, bringing information. in the [00:05:00] academic community, they might be bringing information from, behavioral economics that helps inform, consumer decision making consumer behavior models, right?

but they can also play negative roles, right? And that some of the works we've done with, Military and others in terms of looking at,social media communications by adversaries. That's often the case we see. And then of course there's a third space that just like the general disinformation space as well.

so we find some interesting phenomena, especially around conspiracy theories, where people who believe in one conspiracy theory are more likely to believe in other ones. And we'll find. Bridge nodes in social media that like span. So for instance, going from flat earth as a, as your theory to QAnon as your theory, right?

There will often be nodes who believe in both and participate in both communication networks and using agent based modeling. We can better help understand those, the role of those people and how we could provide. Different interventions to [00:06:00] either accelerate spread through bridge nodes or to decelerate or even stop 

John: so one of the things that comes up sometimes is talking about looking at these questions from an aggregate level versus a really low level of granularity. So when you're doing this, are you representing like every Twitter handle or Facebook thing? Or what's the level of granularity that you get into when you start thinking about these questions?

Yeah, that's a great point. I would say that I almost never start out by trying to represent all of Twitter, like for sure. and when we first start, we almost always started at a much smaller scale than that and start to think about how to add onto it. in many cases, some of the cases we have moved to the order of, say, hundreds of thousands to millions of users interacting.

Bill: We did a nice collaboration with, some people at the University of Central Florida. Most notably, Yvonne Garibay, who was a co P. I. N. R. O. P. I. On that project. and some of his students were for DARPA. We actually did simulate Okay. [00:07:00] populations of individuals and how they would interact with each other with the hopes of trying to track the emergence of trends around certain hashtags or certain pieces of content that might occur.

and given the news that happened, one of the data spaces we looked at was a telegram, right? And given the recent news on telegram, that's interesting. but, In this case, we're looking at pump and dump schemes in Telegram, right? And so we were looking at schemes where people would basically go on other non private social media and pump up a particular cryptocurrency telling everyone, oh, this is the next hot cryptocurrency, but in the Telegram channels, they'd be having conversations about how, as soon as it gets to a certain price point, they're all going to dump it.

And everyone in the Telegram space would claim the. The gains from that, while all these people who had just been following them in other spaces would lose a bunch of money. And in that context, we are almost, we are modeling. at the room level, all the different communications happening within the room.

And we're also monitoring [00:08:00] at this modeling at the same time. The general effect that those room communications would have on the greater social media space, right? Um,one of the things From a scientific perspective, I often think that agent based modeling at the super granular level at the level I just described, for instance, doesn't really gain us any new insights.

It's helpful from a predictive standpoint and from a. An analyst standpoint to understand what will happen, but it doesn't necessarily tell us anything new about the way humans communicate with each other. And so I often like playing around with very abstract models because I think that they do better at helping us understand communication patterns when we tie them into these larger models.

and and this idea is not just mine. This is. been brought up by a number of other gibbous modelers. I remember, Josh Epstein, for instance, who does a lot of work with epidemiological modeling would often point out that even when they built [00:09:00] big, almost global scale models of say flights, right around the world, they That he would get the same insights from a small model that he built with a hundred agents, right?

It was just that when he built it out to that larger level, you were able to show that those same patterns still hold up at that level, right? so I tend to do both is the answer to your question, right? both at the abstract level and then also at the highly detailed, 

John: so one of the things that,we get some pushback on at times within the agent based modeling communities, this balance of trying to build a model to understand, to make a better decision, and then to build a model, to get to a better forecast or a point estimate. And so you've been doing this a lot longer than I have, than anyone on our team has.

How do you think about that question of agent based simulation as a forecasting tool or a. decision support tool.

Bill: that's a great question. And I think that. initially, a lot of times, I [00:10:00] think, agent based modeling was more in the space of decision support, right? almost all the original agent based models that were created were more for a decision support system. And in fact, I've, interestingly enough, there's two different people who've talked about computer simulation in general, but also, Agent based modeling in particular.

One was John Howland, who's my advisor. And another one was John Sturman, who was at MIT and did a lot of the work on system dynamics in addition to dabbling loosely in some agent based modeling related work. both of them expressed the idea that computer modeling in this context could be used as essentially what they called a policy flight simulator, right?

So the idea is that it's. Much like a flight simulator doesn't tell you exactly how to take off and land from one airport to another airport necessarily, but it does tell you the types of things you might encounter along the way and teaches you how to guide the input so that you don't crash this up and along the way, right?

that's what agent based modeling can do, right? It can help you [00:11:00] understand. It's not going to tell you that, if I implement this policy, this is exactly the outcome that's going to happen. But instead tells you loosely, what are the potential pitfalls and benefits that I might be looking out for as I implement this policy, right?

and to this day, I probably think that is a lot of times, a, one of the better uses of agent based models, let's put it that way. Now, that being said, predictive sense, just to the sense to which our point forecast sense. To the same extent to which almost any tool can be used in a point forecast or a predictive sense, right?

so if I, do a good job of making sure that my model has empirically valid input data, and that the, and that I've done my best to validate the model on past data showing similar outcomes have occurred in past situations,then I have a better chance of Making a good point forecast right in this [00:12:00] space, right?

and I would argue that agent based modeling also, if done well, has the ability to overcome some of the critiques in general forecasting that happens in this space, right? so in particular, there's something called the Lucas critique that's been around for a long time, which basically says that you can't really use make predictions about macroeconomic policy and their effects on what pop and make point forecast predictions because all of the studies that have been done about these types of policies have been done in worlds in which that policy was not implemented, right?

So to give a concrete example, if I'm thinking about exploring like a new tax regime, Okay. I can't really make a forecast about what that's going to happen in that space because every model I've ever built in the past was a world in which that tax for that tax regime didn't exist, right? and what the Lucas critique says is that it is [00:13:00] primarily aimed at aggregate type of modeling approaches, right?

But in this particular case, when we started to work on an agent based model, If we have modeled correctly the goals that the agent has and the way in which they are learning to take actions to meet those goals, then a lot of the Lucas critique no longer applies because we've still learned the way people make these decisions at an individual level.

And therefore. we theoretically should be able to explore beyond the bounds of the space that we have, validated the model in. the real question is how do you get other people to believe that then in the end? And I think that comes down to a matter of convincing them that the agent, as you've described it.

is an empirically valid agent, and that the goals and actions they're using to take to achieve their goals are the way that real human beings act and behave, and that takes some [00:14:00] effort to convince

John: So when talking about the construction of agents in that sense, do you think, like we'd epistemics, like we construct synthetic populations and give that to people as a starting point. what are other ways that, and we, we're doing that with All different types of data sets that we pull together and, make sure each individual is represented across all those data sets and do some fitting.

people aren't being counted twice and the but outside of using synthetic populations, how else is that challenge solved or how have you seen people attempt to solve that challenge to have these agents be really realistic?

Bill: I think, this is often where I start to draw a contrast between agent based modeling and other approaches in modeling that are happening right now, especially If we cast a sidewards glance towards a black box type of approaches, like neural nets or deep learning or generative AI approaches right now, right?

the difference between agent based model and any of those approaches beyond just get some, using a synthetic population has the right characteristics [00:15:00] is that we have to write the rules down for exactly how those agents will behave, like what actions they will take and how they will take that.

And so we can. Always compare those rules of behavior, those heuristics of behavior that the agents are using to what we know about real individuals, right? And so we can, so the answer to your question is, I think we start to embed, individual theories of behavior into those agents, right? And that's the other way we make sure that's happening.

If, agents are not using, rules of behavior that represent real humans, right? for instance, if they're getting access to all the information, the entire world, and then deciding their best possible out action, then people can question that's not a valid model for sure.

and so the closer we can get to Relating our agents behavior to theories of individual action, consumer theory, for instance, right? pulling on sociological [00:16:00] theories, things like that as much as possible. I think the better we will do at making good predictions and better decision support in this space at the same time.

and so to me, agent based modeling is fundamentally the When we're talking about large groups of individuals is one of the best approaches out there for building upon actual theory and not just using a black box standard to model

what's going on 

John: people might look at decision support or forecasting as different types of exercises. throughout your career, have you seen agent based modeling come into scenario planning and have you seen that be like just running through scenarios that the decision maker is trying to make throughout some counterfactuals throughout these different what ifs?

Show a range of results. is that the more common way to get people to buy into like agents are realistic? Look, we ran through these scenarios, different things happened as outcomes. What's your take on that?

Bill: for sure. This is, actually this,reminds me of one, when I was a young [00:17:00] graduate student having a discussion.with a, particular agent based modeling researcher named Steven Bankis, who has, done a lot of work building these tools up and then using them in consultation with government and industry and things like that.

And one of the things,it's a glib comment, but it applies. And one of the things Steven often said was, Put at least two accesses on every, prediction that you're making. Right? And what he meant by that was that you never tell a company one value and say, this is going to happen if you do X, right?

Instead, put some accesses on there that say, if you spend this much money, this is where you're gonna wind up. And if you think, if there's another variable in the system that you can't really tie down, right? So for instance, we had a project where we were working with understanding how, incentives that a multi massively multiplayer online game provided to its participants.

Would result in them talking positively about the game to other people in the game. So [00:18:00] how likely was it? And we didn't know, we didn't know what the company had never run an experiment where they handed out like virtual currency to individuals who became premium members and saw what the effect was on positive word of mouth.

So what we did instead was we just included an axis that went from zero to, I forget, that basically affected the probability of how likely. Given that I received this current virtual currency as a gift from the company, how positively was I going to talk to the other members of the game? And that allowed them to do exactly what we're talking about, scenario planning, right? Thinking about if I give out 10 of virtual currency versus a hundred dollars of virtual currency, what's going to be the effect there? How much conversion am I going to see if I give out t shirts instead of, you Bumper stickers, right?

Or virtual currency. What's the effect going to be there? And we could play around with this and tell them when the space of positive outcomes grew and when it diminished as a result of that, [00:19:00] right? but more importantly, we could also start to ask questions. about other accesses that we might look at, right?

So we might look at who are you targeting with these virtual gifts or these gifts, right? Are you targeting people who have friends who are more likely to convert or less likely to convert, right? You can make arguments both ways, right? That if you target people who have a lot of friends who are Really highly likely to convert already.

Then you're wasting your money because those people are likely to convert anyways, but you could also argue that this would tip them over and make them premium members and bring them in. Whereas we target people who are less likely to convert. Maybe this would be enough incentive to tip them over, but it might not be.

It might be that you're just throwing your money away because it doesn't provide enough incentive for the neighbors of that individual to convert. I do think scenario planning is a good way. And I think One of the ways to get buy in from stakeholders is by, either presenting multiple scenarios or even presenting, a range [00:20:00] of values that can affect your output.

And you can think of them as. a large number of scenarios, or you could think of that as, a continuous range and see what the

effect is

John: I'm going to zoom us out a little bit. before we dive into some other concrete examples of agent based modeling, the applied,pod world that you've been working in for a while. a lot of times we like to ask people who are on our podcast, what got you into modeling and complex social systems in the first place?

And do you remember having your initial aha moment that agent based modeling just clicked in and, help make sense of things and what did you do next? Yeah. Yeah. Yeah.

Bill: I, that's a great question. And my background in particular is unusual in this space. I am, I'm now a, a chaired professor in marketing. But I don't actually have a degree in marketing, right? I have a degree in artificial intelligence, right? From the University of Michigan, which is where I did my work, right?

And so how did I make that transition, is the question you're [00:21:00] asking and the question of interest here. And I would say it actually goes all the way back to my undergrad days, right? As an undergrad, when I first started, I would always knew I was going to do computer science. But then I, got to Michigan state, which is where I did my undergrad.

And, it turns out I really wanted to do actually, virtual reality is what I wanted to look at. I was fascinated by the idea of this, of cyberspace and the metaverse shortly after William Gibson's work had come out, some of the newer,stuff, like Neil Steffens Snow Crash wasn't even out yet, but I was fascinated by that idea.

but I went there and, I was assigned a research advisor because I, luckily as an undergrad, I had this nice position that allowed me to work as a research student. And I was working with a guy named Bill Punch and, Bill punch was like, no one really does virtual reality. It's really hard to do.

Cause this is like back in the nineties and it was just a difficult problem at the time. and he's but I do a lot of work in AI. Would you be interested in that? So I started working with him in [00:22:00] AI and there were a lot of more traditional AI problems that we solve that had to do with 

or that we worked on, I should say, we didn't necessarily solve them, that had to do with like engineering material compositions for things like flywheels and stuff like that. And there, and there were some like traditional, just how do you get a robot to drive through a maze kind of things that we worked on.

But, I became more and more fascinated by the idea that we could use the same techniques to model and help understand social sciences. And so my, senior thesis in this space was a project I did where I looked at the evolution of cooperation game and I used AI to evolve different solutions to evolution of cooperation, and explore that space and how.

Evolving agents in different kind of environments affected the result in outcomes. and that was the seed, but it didn't really blossom until [00:23:00] almost my postdoc, Where I realized more and more that computer science, when it came to AI at the time at which I was graduating was really more interested in using AI to solve Kind of the same set of problems they'd been looking at for 30 years.

and so I got bored with that idea. What I wanted to do instead was take computer modeling, agent based modeling, and understand other systems like information diffusion, which we mentioned, But I was having a hard time at times publishing that stuff in computer science, because.they were like, this is already a technique that's known.

This is already a in a I approach that we know about. It's not you knew are unique. You're just doing it in a new domain. And so therefore, it's not really of interest to a I or to computer science, right? So what I realized is that I really needed to go to a social science field in order to make this happen, right?

I needed to take age based modeling and explore the social sciences with it and [00:24:00] try and publish new insights in the social sciences. And so that was the transition. It was around, it was about, I would say towards the end of my postdoc at Northwestern that I started Considering social sciences.

And then when I got an offer to be a professor of marketing at the university of Maryland, that I really jumped full, both feet into that space is I made the transition from, more technical backgrounds to more social science

John: Cool. So as you've made that transition, from more of like the computer science side and the social sciences, we're seeing that to be a little bit more common today. So thank you for, setting the tone that's doable. Uh,we, in what fields of social science are you most excited about agent based modeling having more of an impact on going forward? 

Bill: Yeah, that's a great question. and I think there's a number of places. I think, my own area of marketing, I think is a really interesting area, right? Like it's using,the, as I've continued to work in agent based modeling, one thing I've realized is that agent based modeling is almost [00:25:00] its most powerful when interactions are a driving force of the system interactions between autonomous agents or individual agents, right?

And that's nowhere is that more clear than in marketing, right? marketing, if you look at the earliest studies on marketing, they basically show that Word of mouth has orders of magnitude more effect than advertising, right? You go back to some of the the Ryan and gross hybrid corn studies, and they show that like what their neighbors said to them was much more important than what they were seeing on TV or hearing in the news or anything like that.

and Agent based modeling provides us with a great way to help understand word of mouth, right? And there are very few other techniques out there that will let you model the interactions between individual agents. And so you can see scenarios where advertising affects one particular agent, that agent then talks to ten other agents about what they did, right?

And causes the spread of product adoption that I don't know how you [00:26:00] capture in just about any other modeling technique that we look at, right? So that's an area. And I think generally business systems, studying organization and businesses as a whole has become increasingly difficult to do, using a lot of traditional techniques because of the number of interactions that happen within organizations and businesses.

so I think that's a powerful area where agent based modeling could really play a role. I also really am very positive about,describe it, but in general, the space of what I would call, digital twin aspects. So these are areas where for whatever reason, there is a highly dynamic system that has a big social component to it.

Usually that you want to model, Because you want to see if something happens, how that's going to affect other aspects. So give you a great example. I happen to be at a conference in Hong Kong a couple of years ago, and the, I think it was the CEO president, something of the Hong Kong airport was [00:27:00] there presenting how they are building a digital twin of the entire airport.

But what they were doing was plot, was building a simulation of like where every airplane is, where every baggage cart is, all that kind of stuff. But what they were missing was the human component, right? Like where were the human resources that we're going to interact with those componentry, right? And the reason why they want this big digital twin is because they want to understand.

If a plane is four hours late, how is that going to affect their gate arrival times and everything else? How do they need to shift things around in order to work more efficiently? But without humans, without modeling the human behavior, where are the passengers? Where are the, where's the flight crew that has to be on that plane?

Where is the gate agent that has to open that gate four hours later, right? You're missing a lot of the complexity of what's going on in that system. So I foresee a merger of some of these kind of infrastructure models that have been built, [00:28:00] like airports, traffic patterns, In the U S or anywhere in the world, for that matter.

I know it's already being used by places like amusement parks, right? So Disney and others have employed agent based modeling to model what's going on in those spaces and building digital twins of all those spaces so that you can react quickly to disruptions in the physical environment or in the social environment is a critical area to look into in the 

future, I think, as well. 

John: is there an area where ABM has not yet ventured that you see,10 years from now, 20 years from now, it'll just be so obvious that Why weren't we doing this all along, but agent based modeling is going to come and sweep up a new wave of adoption for.

Bill: interesting enough, I feel like ABM has at least been applied by at least one or two researchers in just about every single field now. So it's hard to say there are areas where I think it could take on a greater role. Let's put it that way. So one One that [00:29:00] pops to mind is in the medical space, right?

of course, we've heard a lot about agent based modeling and epidemiology, right? Especially with COVID and the pandemic and all. I think, This last pandemic in particular really pushed a B. M. Up above compartmentalized models is the solution to understanding that policy effects in these spaces.

but I think there's other parts of the medical field that we're not looking at them in as much. And I know some good research being done. But for instance, hospital environments, right? Trying to model The hospital environment is interesting because you not only need to have the right technical resources, like if you need to have someone get an MRI, the MRI has to be free, but you also need to have the right knowledge resources, right?

You need to have someone who can look at the MRI data and make a decision based upon it. And so being able to include agent based models in some of these kind of Resource allocation [00:30:00] decision making or, things as simple as looking at like hand washing behavior and how hand washing behavior affects spread inside of a hospital environment, I think, is something that is missing and that we could probably do more with and we should be doing more with already.

Another one that I think is interesting is, law, right? We've thought a lot about,policy effects. So there's been a lot of exploration, individual domains in law. so like tax policy, and there's a lot of work on agent based modeling and tax policy, for instance, right? But generally thinking about how different legal, policies might affect Access to the law for individuals, I think, is something that hasn't been considered.

for instance, to give you, make this a little more concrete, right? there's been a lot of talk about how generative AI could replace lawyers for some basic legal systems, like generating your will, or, appealing a traffic citation, or something like that, right? [00:31:00] But what does that mean in terms of, all of the access that underrepresented individuals within the legal system have?

Does that mean that, that people are going to trust a computer to work in this space? Or is there a certain proportion that is going to get more access? Is there a better way to make these legal systems available? There are these legal tools, I should say, available, right? and I think agent based modeling.

More generally could play a larger role in complexity theory as well could play a larger role in Legal access and understanding how we can make the law more equitable for a larger group of individuals 

John: a topic that comes up time and time again when trying to understand both like domestic or also international policy in a couple areas. our, our hope is to also see that agent based modeling becomes more and more used in that space. Thanks. So we have time for two more questions.

the one question that we ask everybody who's on the podcast, we call it a what the flux moment. So these are moments that people,think of, you know what, if I could have gone [00:32:00] back. And if I had that insight ahead of time, maybe I could have made a better decision. Maybe I would have understood the outcome a little bit more or, avoided some unintended consequences.

So do you have a favorite example, whether it's historical in politics or in business or something you were a part of that if you had the ability to go back, run some simulations,Make a different decision or help guide a different decision.what would that be? What comes to mind for you?

Bill: so I think the moment or the time that I'm going to most point to is actually more in the methodology of ABM than it is in a substantive area. so there was a, when I first started working with agent based modeling, I should say, not even just when I first started, but into my agent based modeling career, there was a,I considered, I thought a lot of times that we needed to make models.

As veridical as possible to the real [00:33:00] world phenomenon we are trying to model, right? Making them as detailed as possible. And the point I often point the particular flux point that I point to is a case where we were trying to build a Large scale model of suburban sprawl in southeastern Michigan. This is part of what was called Project Sluice.

the P. I. Was Dan Brown on the project. and we worked a lot with landscape architects and other individuals who had knowledge about how people chose landscapes. particular places to live, particular locations and things like that. and I was trying to, as my, as a grad student, do my best to build a model that represented all these competing voices, of what was going on in it, and I wound up with a rule set for an individual agent that looked like a very large, complex decision tree, right?

It had, something like, I think. On the order of two dozen branching points in this tree, that [00:34:00] was controlling the agent behavior. And all this was doing was deciding where the agent would most like to live in a particular neighborhood that was presented to them. what I realized after I've been playing around with this model and after implementing it was that about, I'd probably say more than half, I'd probably say two thirds of those decision rules We're never going to come into play and we're never, or could just as easily have been approximated by a random variable, right?

Or something like that, right? and that the impact on the results of the model would have been very, would have been none, not, there would have been no impact. It would have produced almost the same output, right? And so the, what the flexibility here for me was that realization I had. Thank you. They, even though a lot of times we work with subject matter experts, because we're not, I'm not an expert in every single modeling domain that I build a model in, and we rely upon them to tell us what.

What's going on that their [00:35:00] knowledge is greater than what is needed to make an, to represent an individual agent or an individual consumer. A lot of times,that the way that when you talk to an expert about how people make decisions, they often overlook the fact that most individuals don't know as much as they do.

And so we need to pare down the knowledge of the expert. To the most generalizable individual agent that we could think about, right? Rather than getting a more detailed model that necessarily isn't going to actually do anything different than the simpler model. And the simpler model is better for a number of different reasons.

And this is despite the fact that, I was reading They keep it simple, stupid type of work and stuff like this all along the lines. I still have this knowledge, right? That we really need to pare these models down because once we have the simpler model, we can provide better explanations for why the model produces the outcome it does.

And we can help to, make [00:36:00] forecast into the future about how changing that model is going to result in different outputs right than we would otherwise. So the kind of the finally just getting it really hit home to me that you really need to start with the simplest possible model and add complexity as you need to, to, to make it more valid.

was the approach. That was my what

the flex moment. 

John: cool. so final question, anything going on that's new with net logo or any other folks you want to, can I give a shout out to before we wrap up today? Whose work that you feel is really pushing agent based modeling forward?or a student you want to call out or something like that.

So the 

floor is yours 

Bill: sure. 

John: up.

Bill: I would say a couple of things. First of all, with regards to net logo, they, what one thing that's exciting is they have received a large, NSF grant to kind of transition that logo to a foundation so that it could become its own autonomous entity, which I'm really excited to see what happens.

And they're making cool changes to net logo. As a result, I've [00:37:00] been lucky enough to play around with kind of an early prototype of some of the new versions, and it looks a lot more modern and a lot, a lot more elegant than it did in the past at times, right? So I'm excited about that. the other thing I would mention is that Uri and I, we mentioned the textbook at the top of the podcast and Uri and I are working on a second edition of that.

we're hoping to see that published sometime next year, ideally. and that will be a lot of fun with that specific aspect. in terms of other things that really excite me about agent based modeling, I think the biggest area that I am interested going forward, and I've alluded to this over the rest of this The idea of mixing machine learning with agent based modeling in various ways, right?

in particular, I got involved,and helped set up a group with Josh Epstein, and a number of others on an area that we called, inverse generative social science, right? and the idea behind inverse generative social science was that you can take, [00:38:00] you can use machine learning tools to better understand, how to create an agent based model to begin with, right?

So we can take machine learning, we can apply it to a data set and then infer rules of behavior from that data set in a way that it's still tied to theory because we set up The basic structure and format of that machine learning application, it outputs something that might look like a decision tree or associative rules or other stuff along that space.

But, but so it still has captures all the richness of the data, but still has a theoretical underlying to it, right? And then we could automatically. Put those agents that we derive from that data directly into a simulation and explore what's happening. in my case, I've done work in the past where we looked at Twitter behavior and for [00:39:00] directly how the agents are deciding to post and when to post from their past behavior and then recreated an entire Twitter network.

Based upon that machine learning in an agent based modeling context. And we talked about the digital twinning stuff earlier, things like that, they become a lot more, they become a lot easier to do if we have the ability to automatically. interpret rules from this data. And so we organized a couple of workshops in this space.

We also had a,a special section in jazz, the journal of artificial social sciences journal, artificial societies and social simulation. that had me, Josh Epstein, Yvonne Garibay, Rez Hotna and, Matt Kohler. as a guest editors on it, right? And we also help put together a lot of the IDSS workshops.

And so that idea of using machine learning to build better age of base models is something that I'm really interested in. I want to keep looking

John: Awesome. [00:40:00] Bill, I'm really grateful for, you making the time to jump on our podcast and, we'll be sharing this out pretty soon. So thank you.


People on this episode