Paolo Gaudiano is a mathematician and cognitive scientist who is often asked to demystify the practice of predictive modeling for businesses. A former professor at Boston University, he left academia to become chief scientist at Artificial Life, a startup focused on evolutionary algorithms, before joining Icosystem as its president. Founded in 2001, the company has built predictive modeling systems for clients as diverse as French phone company Orange, pharma giant Eli Lilly, and casino operator Harrah’s. The firm, based in Cambridge, Massachusetts, has also done work for the U.S. military, most recently building simulation software to model the infrastructure efforts in Afghanistan.
TR: To some people, predictive modeling sounds like magic, because it promises to tell you what will happen in the future.
Gaudiano: No, it’s not magic at all. It’s a way of taking advantage of computers to replicate the real world. But you don’t just want to replicate what happened—you want to see what will happen if the world changes around you. What if the economy collapses? What if I change my sales strategy?
You can take factors into account that are otherwise incredibly difficult to take into account. So we’re not predicting the future but just giving you a better understanding of how things work and a slightly better likelihood that the things you do will actually turn out the way you expect them to. It’s a decision support tool. It makes your intuition more quantitative. It gives you a way to test the validity of your intuition with data and come up with a better answer. That’s all it is.
Your workhorse technology at Icosystem is “agent-based modeling.” What is agent-based modeling, where did the technology come from, and how do you implement it?
Agent-based modeling started a long time ago as a tool in the social sciences, for understanding the behaviors of populations. It’s come of age in the past ten years. The core idea is whenever you have a complex organization or ecosystem, it’s easier to understand and simulate the behavior of individuals and how they interact with one another and their environment than it is to come up with some kind of a mathematical law that tells you how the population behaves.
What are the “agents” in these simulations?
Agents are replicas of whatever elements of the system we’re studying. Typically, they’re humans, so if we’re solving a marketing problem, they’re consumers. But they can also be personnel in a company, cars on a highway, or computers on a network.
You simulate things from the bottom up. You quite literally capture the details of how these elements work and how they connect with one another. That turns out to be a very powerful way to predict how the system as a whole will behave. I can run on a laptop a simulation of 100,000 consumers doing their shopping and looking at advertising, and it will take two minutes to run it. By looking at the results, it gives you a different way of thinking about your problem.
What do you mean by a different way of thinking?
For example, we’re doing a project for the Navy, helping them understand reconstruction in territories like Afghanistan and how you combine that with strategic communications. So we built a model that looks at Afghan citizens and how they’re being exposed to things around them, like the international teams, the Taliban.
The Navy asks us: How do you know the model is correct? But it’s less about being correct [about how people interact now] and more about understanding what assumptions [about future events could] lead to. I don’t know how often Afghan citizens talk to each other about the water. But I can run 20 different simulations with 20 different assumptions about that.
What is the result you’re trying to achieve?
If I am in charge of some troops in Afghanistan and I have resources, money—what do I do about medical treatment, safety, education systems? Do I build wells in this village—do I build one well here, or two or three there? Or do I put money in veterinary support? Am I better off advertising on radio rather than TV, should I drop leaflets from an airplane, should I go to places of worship—so that they’re hearing my message rather than my opponent’s message?
It’s less about predicting [whether] spending $5 changes opinions by 2 percent, and really more about: I have these five different courses of action. Which are the most likely to be successful, and why? It’s about: this is the range, help me understand which will work and which won’t and why. You can literally trace why.
What about examples from the business world? For instance, what did you do with Orange, the French phone company?
They were concerned about the spread of [computer] viruses over cell-phone networks. So we built a simulation of hundreds of thousands of users and how viruses can spread. We took data from real viruses and infection rates. We modeled the behavior of users. You can be on the subway, and someone else is using a phone with a virus next to you, and you’re using Bluetooth and it asks to connect your headset with their phone, and if you say yes, you can catch the virus from the other phone. Or you can catch it by sending data through SMS. We predicted infection rates and helped to design strategies to prevent the spread when a virus is injected into the system.
I understand that it takes several hundred thousand dollars and up to get a predictive model project like this launched. Is there any way to bring those costs down?
That’s where things are getting interesting. It’s true that when we do custom projects, it is virtually impossible to get one started for less than $300,000. That’s for version one working on your desktop, and it’s not completely functional. And it may turn into a multiyear, multimillion dollar project.
But now, we are able to repeat work in certain industries. For instance, in consumer behavior, we’ve developed an agent-based simulation for measuring the return on investment for brand advertising, and we can license the tool to a client for a few thousand dollars per month. That’s a spin-off company we’re incubating, called Concentric ROI. So we’ve lowered the entry threshold, and it’s much more appealing to be able to use the model for just a few months.
Become an MIT Technology Review Insider for in-depth analysis and unparalleled perspective.Subscribe today