Interview conducted by Bo Parker. PricewaterhouseCoopers
Dr. Mark Paich is a principal at Decisio Consulting and a former student of Dr. Jay Forrester of the Massachusetts Institute of Technology (MIT), who pioneered the field of system dynamics. In this interview, Paich describes Decisio’s involvement in the development of the AnyLogic simulator, an XJ Technologies product designed to facilitate complex adaptive systems modeling approaches, including both system dynamics and the more recent agent-based modeling. Paich provides examples of Decisio’s success in using AnyLogic and other simulation tools to help executives at large enterprises make major transformation decisions.
PwC: How did you get involved with the development of AnyLogic?
MP: I’ve done a lot of work in what would be called the system dynamics tradition, which has many elements common to agent-based modeling but works at a higher level of aggregation. Some of the observations I can offer are system dynamics-based, but still apply to agent-based modeling.
We got involved in AnyLogic’s development because we saw some limitations to the system dynamics framework we wanted to try to address. Some of what worked out really well with system dynamics can be done better now with agent-based modeling.
PwC: How did you come to this conclusion?
Can you give some examples?
MP: We’re seeing an explosion in the availability of data. All kinds of data sets are now becoming available through corporate information systems and various systems that provide the kind of information you can now use to build agent-based models. Whereas before, we would really be scrounging for data in a lot of cases. Now, so much more is available. The primary challenge is to make sense of all of this very detailed, disaggregated data.
We do a lot of work in pharmaceuticals, and we build models to look at the dynamics. These are agentbased models in some cases, and system dynamics models in others, to look at the launch of new pharmaceutical products and who might adopt various projects. You can now get a tremendous amount of data from different kinds of databases dealing with pharmaceuticals that were just never available before. And that’s just one example.
PwC: So the data are available for a variety of reasons, but are reporting and regulatory the primary reasons?
MP: I think so. On the pharma side, we regularly use a massive database that came from General Electric. That database includes health-related information on how humans behave, what drugs they take, when they switch, and that kind of thing. Those data have been instrumental in calibrating a number of different models that we’ve had.
Not only that, but corporations do a much better job now of keeping track of information about specific products. I’m now able to get data on the profitability and sales of products that I couldn’t have gotten 10 years before.
There is a tremendous opportunity out there. Let me give you an example. The one that I’ve done that probably is best known was a system dynamics model—it could also have been an agent-based model—that helped with the design of the General Motors [GM] OnStar business.
We all know that GM has not done a lot of things right in the last few years, but this is one that they did right. We used a dynamic model to help design GM’s entry into the telematics business, which was really creating the telematics business. That telematics business was OnStar. A system dynamics–like model was behind a good bit of the original strategy and GM’s decision to expand it over a lot of vehicle lines. All of that is written up publicly¹. We also were a finalist for the 2001 Franz Edelman Award for Achievement in Operations Research and the Management Sciences².
OnStar is an example of where you essentially can use some of the dynamic modeling tools to practically design a business model from scratch. GM made a lot of changes and alterations to it over time, which is what you’d expect. But to originally convince the folks inside GM that this was a viable opportunity and that the basic concept was right, I think the model was pretty instrumental.
We could have done some things with agent-based technology if it existed, but it didn’t. Since then, we’ve learned how to integrate the market research that so many companies perform into these kinds of dynamic models. And, you can do that on an individual level with the agent-based models. AnyLogic has good tools for this. You can do it relatively easily.
I’m sure you’ve seen things like conjoint analysis choice models. That kind of data and information can be integrated directly into an agent-based model, and you can get the full range of heterogeneity between different kinds of consumers.
PwC: What is the value of adding an agentbased modeling [ABM] approach to a model you have already established with system dynamics?
MP: There are a couple of things. One is that you are able to get a higher level of granularity, which can be important in some cases. For example, in the OnStar case, you want to keep track of a lot of different pieces of information about individuals. You would like to know what kind of GM car they drive. You would like to know various demographic data. You would like to know a whole series of things. In a system dynamics model, keeping track of all of that for a lot of different market segments is really hard—you get a combinatorial explosion—but with an agent-based model, it’s relatively straightforward and intuitive. You can keep track of a lot more information about the individual actors in the system.
PwC: So when you aggregate and use the system dynamics approach, you get a combinatorial explosion. Is this because of the variety of factors that are relevant to the model?
MP: Yes. If you have a lot of demographic factors, a lot of descriptors in those individuals, you can hit the combinatorial explosion pretty quickly.
PwC: With an ABM approach, you can express that variety as individuals.
MP: Right. You just express it directly. You set the characteristics of individuals and just replicate them directly.
The other thing that agent-based models get you is the ability to get at what I call a social network, or the word-of-mouth effect. For a variety of products, and everybody knows this, the social network influences what people around you buy and do. It has a tremendous impact on what you decide you want to buy. Ultimately, the social network will be very, very important, but we’re just starting to develop the data.
You have consumers that are connected together in networks. You want to find a leverage point where you can influence key actors in that network, and then produce a tipping response that changes the attitude or buyer behavior. There are strong positive feedback loops running, so if a certain number of people adopt a product or a technology or change their attitude about a product or technology, they talk to others and influence others, and you can produce the cascade effect.
PwC: It’s one thing to understand that social networks have a strong impact. How would the act of modeling itself surface the key information that you would need to know?
MP: That is the state of the art. But let me tell you what we did for a major manufacturer looking to change the attitude toward its products very quickly, and specifically in the Los Angeles area in California. We had data on what products people had from the competitors and what people had products from this particular firm. And we also had some survey data about attitudes that people had toward the product. We were able to say something about what type of people, according to demographic characteristics, had different attitudes.
PwC: So you matched attitudes with the types of products they had?
MP: Exactly. We synthesized this information into an agent-based model. We calibrated the model on the basis of some fairly detailed geographic data to get a sense as to whose purchases influenced whose purchases. Now, there were some leaps of faith there because we didn’t have direct data that said, “I influence you.”
PwC: So the model provided a substitute for what a social network analysis approach that actually had that data would have told you directly?
MP: In part. We made some assumptions about what the network would look like, based on studies that have been done on who talks to whom. Birds of a feather flock together, so people in the same age groups who have other things in common tend to talk to each other. We got a decent approximation of what a network might look like, and then we were able to do some statistical analysis.
By the way, the statistical analysis of agent-based modeling is a big frontier. We performed some statistical analysis on the model, and what came out of it was a targeting strategy. It said that if you want to sell more of this product, here are the key neighborhoods. We identified the key neighborhood census tracts you want to target to best exploit the social network effect.
Our study said that if you did the targeting that way, it would be five times more effective than a random targeting, and the number of marketing messages and level of expenditure would be the same. The company has not done this targeting yet, but I understand they are getting close to having one of their major divisions actually do it.
I have to admit that this stuff is way out on the bleeding edge, but it’s going to get better. We were inventing the statistical techniques a lot as we went on, but I think they can get a lot better, and I think we learned a lot in the process. But the basic idea is really important: try to find a place to intervene that creates a tipping point and then a cascade.
1. Vince Barraba, Chet Huber, Fred Cooke, Nick Pudar, Jim Smith, and Mark Paich, “Multimethod Approach for Creating New Business Models: The General Motors OnStar Project,” Interfaces 32, no. 1 (January–February 2002): 20–34.
2. See “2001 Edelman Award Winner” at http://www.informs.org/article.php?id=1009, accessed November 14, 2009.