Any notion that awarding a new thoroughbred-racing license is a simple matter surely has been disabused by now.

	Robin Rombach/Post-Gazette
	Tuomas Sandholm, a Carnegie Mellon University computer scientist, is this year's recipient of the Computers and Thought Award.

Competing applications for a new track in Western Pennsylvania this year have been linked with proposals to legalize slot-machine gambling as well as promises to help fund a new hockey arena, contribute to local charities and develop residential housing, not to mention urban strip mining.

Figuring out how to weigh all of those factors and make a decision that benefits the most people would daunt most people, but not Tuomas Sandholm.

A computer scientist at Carnegie Mellon University, Sandholm, 34, specializes in using artificial intelligence to solve this sort of complex problem. It's not that a computer can spit out the ultimate solution, but the programs he has developed are adept at picking out the best rules to govern the decision-making process.

"There are millions of different settings in the world where a game is in place," he explained, and the right set of rules will allow all players to act in their own self-interest and nevertheless come up with a solution that benefits all. His rule-setting programs could apply to any number of such "games" -- whether it's deciding whether to build a bridge, divvying up assets in a divorce settlement, dispatching delivery trucks from different terminals, electing a president, or auctioning goods and services.

If all this seems a bit grandiose and hard to grasp, consider that companies such as Bayer, H.J. Heinz and Procter & Gamble happily pay for Sandholm's patented "combinatorial optimization technology." In the last 18 months, a startup company he founded, CombineNet, has handled electronic auctions totalling $3 billion to $4 billion, saving his Fortune 200 clients a total of $300 million, or 11-12 percent on average, in procurement costs, he said.

And the broad applicability of these methods has not gone unnoticed within the artificial intelligence community. Tomorrow, at the International Joint Conference on Artificial Intelligence in Acapulco, Mexico, Sandholm will accept the Computers and Thought Award, which is presented every two years and is considered the premier award for artificial intelligence researchers under the age of 35.

"He's just done some terrific work," said Tom Mitchell, president of the American Association for Artificial Intelligence and director of the Center for Automated Learning and Discovery at CMU.

One thrust of artificial intelligence research, Mitchell explained, is developing mechanisms for making intelligent decisions. Sandholm has displayed particular insight in demonstrating how groups -- be they people, software programs, or other entities -- manage to make collective decisions.

This group dynamic is evident among the nerve cells in the brain.

"Even though you feel like you're one person, you're actually 10 billion neurons," said Mitchell, who won the Computers and Thought Award in 1983. "Somehow, those 10 billion neurons, each acting on their own, make a decision for you."

Sandholm was the top-ranked windsurfer in his native Finland and ranked among the best worldwide in the late 1980s. Tall and lanky, he still competes, but the only sign of his hobby in his Wean Hall office is a windsurfing calendar. His complexion suggests he has spent far more time lately in overcast Pittsburgh than sun-drenched Maui.

His thoughts migrated from wind-and-waves to the intersection of computer science and economics in 1990, while he was working on a master's degree in industrial engineering and management science at Helsinki University of Technology.

For his thesis, Sandholm designed a system for automatically reallocating trucks from different trucking companies' dispatch centers to complete shipping orders -- essentially allowing trucking companies to subcontract with each other on an order-by-order basis.

The idea of linking the centers by computer and using computer programs, called software agents, to automatically decide which trucks to use for different orders was revolutionary at a time when the Internet was in its infancy. Though the system was never fielded commercially, Sandholm showed it could cut trucking costs by 17 percent.

When he began to wonder if his software agents could be taught to strategize, "all hell broke loose," he said. "What followed was 13 years of research."

He went on to earn a master's and a doctorate in computer science at the University of Massachusetts, Amherst, where he also met and married his wife, Christina Fong, now a CMU economist. In 1996, he joined the computer science faculty of the Washington University in St. Louis. He launched CombineNet in St. Louis in 2000 and moved it to Oakland the next year, when he joined the faculty at Carnegie Mellon.

One of his company's main services is business-to-business auctions that allow companies to bid on packages of goods. Bidders may bid for the package, on part of the package, or might offer one price for the package and a second price for a portion of it. Someone who was seeking umbrellas, for instance, might submit one bid for a package of umbrellas and raincoats, and a side bid on just the umbrellas.

Sellers might add their own conditions, such as limiting any one bidder to a certain percentage of the total transaction.

Trying to mix and match the various bids and choose the combination that satisfies both the buyers and sellers is a task that can easily outstrip human capabilities as the number of bidders grows, he noted.

"Clearing the market -- deciding who should win -- is a very hard optimization problem," Sandholm said. It's akin to solving a Traveling Salesman problem -- the famously difficult task of calculating the shortest distance a salesman can travel to visit a large set of cities.

The Traveling Salesman is the best known of a class of problems known to mathematicians as NP-complete problems. Sandholm developed what he calls the world's fastest method for solving such problems and won a U.S. patent, No. 6,272,473, two years ago.

A larger issue, both for these software agents and for humans, is developing rules that allow each to pursue their own self-interest while choosing a desirable outcome. This, too, is an NP-complete problem that can best be solved by computer.

"Once the rules are designed, it doesn't have to be run by computer," Sandholm said, noting that these machine-produced rules could be used to govern such human endeavors as elections and divorce settlement negotiations.

Sandholm begged off making any prediction about whether politicians, who now set the rules, would cede that authority to a computer in deciding who should be elected, or who should receive a gambling license.

"I'm not political," he said, shrugging.