Chaos, Inc.

By M. Mitchell Waldrop

Appeared in Red Herring on January 22, 2003

It’s taken awhile. But after years of nurturing in university labs and think tanks, complexity science—a set of theories of how complex adaptive systems such as stock markets, supply chains, and even rain forests work—is finally moving out into the real world. A small, but vigorous coterie of startup firms are making its theories tangible with a new breed of computer simulations. And their clients are discovering that these simulations can make a substantial difference to the bottom line.

In Santa Fe, for example, the oldest of the startups, BiosGroup, recently ran a simulation to help Proctor and Gamble achieve a 75% inventory reduction in its immense supply network, which includes some 250 product lines ranging from Tide to Pringles to Pampers. For another client, Southwest, BiosGroup did a computer model of the airline’s freight delivery operation and showed it how to save some $2 million per year.

The details of these kinds of complex, adaptive simulations can be tricky, which is why practical business applications didn’t appear until the latter half of the 1990s. But the principle is straightforward: All complex systems—like delivering freight—share some profound similarities. Each of them is massively parallel. They have many quasi-independent “agents” interacting at once. (An agent might be a single firm in an economy, or a baggage handler at Southwest.) These agents are adaptive: they are constantly responding to each other (the baggage handlers interacting with their bosses). And they are decentralized: no one agent is completely in charge (baggage handlers don’t always follow management’s directives). Like all complex systems, then, the overall behavior of Southwest’s freight operation emerges spontaneously from a myriad of low-level interactions.

The key to these simulations, known as agent-based modeling, is to correctly identify these low-level interactions. And that is precisely what a BiosGroup’s team did beginning in 1999. Led by Fred Seibel, now the company’s vice president for supply chain operations, the team interviewed all the relevant Southwest employees to find out how they did their jobs. By March of that year, says Seibel, “we’d developed agents for the guys in the freight house who accept a customer’s package and figure out what flight it’s supposed to go on. We had agents for the guys on the ramp who were loading the planes—and who, as it turned out, weren’t strictly following the plans that the freight house guys were giving them. We even had agents representing inanimate actors, like the planes and the packages. Then we told the simulation to move the freight through the system. The agents took orders from customers, assigned packages to flights according to the rules the actual people used, and flew a week’s worth of operations while we measured how often they had to load and unload cargo from the planes, and how often they had to store it overnight.”

The simulated operation was a close match with Southwest’s actual performance, says Seibel, which gave the BiosGroup team some confidence that their model was accurate. “Finally,” he says, “we tried some different rules for assigning packages and found that Southwest could handle its cargo with much more efficiency.” (In the quest for speed, he explains, airline employees at each airport had been unloading packages from incoming flights and putting them on the next direct flight to the destination. Paradoxically, however, it turned out to be faster on the average to let packages take the long way around: if an aircraft was going to get to the destination eventually, even after several intermediate stops, it was better just to leave the package on board and eliminate all that loading and unloading.) Southwest, not surprisingly, lost no time in implementing the new strategy. The airline found that in practice, the new strategy cut the freight transfer rate by up to 85% at the busiest airports, which greatly reduced the workload for the ramp employees, and freed up dollars that previously went to renting expensive cargo storage space at those airports. From a simulation that took just 6 weeks to create, at a cost of just $60,000, the company is now saving an estimated $10 million over five years.

As Chuck Thomas, Southwest’s financial analysis director, put it in an article that he and Seibel co-authored for Cap Gemini Ernst & Young’s online innovation newsletter, the new strategy was like “striking gold in the schedule…a surprisingly intuitive solution that yielded dramatic results.”

A similar simulation at Proctor and Gamble found, among other things, that if the company ran its delivery trucks more frequently but less full, it could drastically reduce the cost of inventory in the warehouses. “We built a model of Proctor and Gamble’s entire transportation network, from the factories to the retail outlets,” says Seibel. Then they began to play “what-if.” The result of those tests were good enough, says Jake Barr, P&G’s associate director for supply network innovation, “that we’ve begun to work with our partners and suppliers to implement the lessons across our supply networks.”

One of the advantages of these agent-based simulations is that they tend to be much easier for company executives to understand. Unlike the equations and abstractions that underlie a conventional computer simulation, a software agent will usually correspond quite closely to something a manager can recognize in the real world, say a baggage handler not following the rules. So executives don’t have to be an expert to tell whether the software’s behavior makes sense. In addition, says Seibel, agent-based modeling allows you to play “what-if” at a whole new level. Instead of simply tweaking a number here and there, as you might do with a spreadsheet, you can try out the effects of entirely different strategies. Indeed, the simulations represent a whole new approach to forecasting the future.

That’s what makes agent-based modeling so invaluable for the insurance industry. Assuratech, a Santa Fe-based startup founded in 2001, is using an agent-based simulation called Insurance World to help clients like Swiss Re and the State of California find better strategies for managing risk—particularly when it comes to extreme events like 1992’s Hurricane Andrew or the September 11 terrorist attacks. An avalanche of interconnected insurance claims for property, auto, life, accident, and other losses can overwhelm the industry’s ability to pay. “Several insurance companies went bankrupt because of Andrew, simply because they hadn’t considered the interdependencies,” says Terry Dunn, a 30-year industry veteran who now serves as Assuratech’s president and CEO.

But then, says Dunn, that’s the great power of all these agent-based simulations: they offer a whole new way to think about such interdependencies—as well as to quantify, and even to manage them. “Firms are used to dealing with one risk at a time,” he explains, whether it’s the risk of paying out on auto accidents, or the risk to the company’s own portfolio from a stock market downturn. “Then they model each risk using historical, statistical data to get the total effect on the bottom line.” But of course, he says, that approach assumes that all the risks are independent, which they often are not. An agent-based model builds the connections in from the beginning.

TK quote from Assuratech.

Despite its strengths, the nascent agent-based simulation industry is still struggling to establish itself. The downturn has taken its toll; BiosGroup, for one, is currently down to just 37 employees from a high of TK. And the technique itself can sometimes be a very tough sell. All too often, a pitch to potential clients produces “a dogs-watching-TV” response. They don’t have the slightest idea what you’re talking about, says Roger Jones, a former Los Alamos physicist and the CEO Complexica, a Santa Fe-based incubator for complexity-based businesses that has spun off Assuratech and two other companies since its founding in 2000.

“While the power of the tools is dramatic,” agrees Dunn, “they are leading edge, and most of the world is still catching up.”

Nonetheless, success stories like Southwest and P&G are a big help in crossing the conceptual barrier “It gets easier every time we have another client,” says Michael Neely, head of BiosGroup’s Washington, D.C., office. To help that process along, he adds, BiosGroup has recently taken over the development of Ascape: a toolkit originally created at the Brookings Institution to make it easy for nonexperts to produce agent-based models on their own.

In another positive development, the worst of the financial crunch appears to be over. In the aftermath of the September 11th terrorist attacks, in fact, there’s been a notable upsurge of interest from the U.S. intelligence community, where agencies have commissioned a number of agent-based models of things they would just as soon not talk about.

In Cambridge, Massachusetts, meanwhile, a new firm called Icosystem became profitable just nine months after its founding in January 2001. Its agent-based models have helped the Air Force and Navy do a better job of coordinating swarms of unmanned aerial vehicles—the UAVs that proved so useful for surveillance and combat work in Afghanistan. Icosystem has also helped pharmaceutical companies do a better job of coordinating partnerships for drug development, and oil companies coordinate operations on drilling rigs.

It’s important to be realistic about what these agent-based models can do for a client, cautions Icosystem founder Eric Bonabeau. “It can be a big investment,” he says, “because you may be modeling an entire organization that’s very complex. So you have to decide when a full-scale model is worth it.” In many situations, such as an insurance company estimating a straightforward mortality risk, simple statistics may work just fine.

Still, the business world is only going to get more complex with time, not less. And that’s why, in the future, companies like P&G may very well find themselves relying even more on agent-based simulations. After all, says P&G’s Barr, the recent explosion of e-commerce and B2B transactions over the Internet “has given us the capability to move much more data about consumer choices, at a much finer granularity.” Within three to five years, in fact, thanks to microchip-based “electronic product codes” imbedded in the packaging, the industry expects to be tracking every single box of detergent or toothpaste through every stage in the product lifecycle, from factory to warehouse to retailer’s shelf to checkout line. So the pressure is on, says Barr: Every manufacturer is scrambling to make its supply network as flexible and as responsive as possible, so that it can adapt to changes in consumer demand just as rapidly as the data comes in. Agent-based simulations could help them do that. Eventually, in fact, the supply networks are going to be so complex that they will have to be managed by software agents living in cyberspace itself.