January 01, 2004

HAL 9000 picks bowl games?

Which team--Louisiana State or USC--will play Oklahoma for the national championship? And who will be making that determination? Why, an astrophysicist, of course; and an immunologist and an MIT-trained mathematician from Indiana, plus a math professor from Virginia.

Each week, seven PC's--scattered in various bedrooms, living rooms and offices around the country--calculate BCS rankings. Each computer ranks the teams differently. Computer rankings have been a major component, along with the polls, in determining the matchups of the top four bowl games--and the split of about $90 million. Little is known about how these computers make their choices, where they are based and who is behind the programming.

The computers used for the BCS are not run by large institutions or organizations (except for the NY Times), but by mathematicians and statistics geeks.

The computers make their selections not just on victories and losses, but also on such peripheral matters as the records of a team's opponents. They are based on the same statistics, but use them differently to assess teams, assigning different priorities to different pieces of data.

Take Dr. Peter Wolfe, who has a system employed by the BCS that uses a "maximum likelihood estimate" comparing how teams performed against common opponents. His system gives no importance to whether a team plays at home or away and weighs all games equally, regardless of when in the season they are played. Mr. Wolfe is a Harvard-educated immunologist.

Jeff Sagarin, whose computer rankings are published in USA Today, is a statistician and an MIT math major, adapted his program from a system used for chess players. He takes home-away distinctions into consideration, "It's a huge advantage to play at home," he said. "Playing an away game is like running a sprint with a grand piano on your back."

The only computer rankings used in the bowl race whose formula is made public are those of Wesley Colley, a University of Virginia professor who works on missile technology for NASA. But unless you are a mathematician--or have a doctorate in astrophysical sciences, as Dr. Colley does--you might have some trouble figuring it out. "It's basically a linear equation with 117 dimensions," he said. "The program analyzes the data until it comes up with the self-consistent solution. Solving it by hand for one team would take a year."

Computer ranking "is as much an art as a science," said Jerry Palm, an independent computer analyst who runs a Web site which keeps tabs on the computer rankings and publishes standings. Palm says the computer formulas should be open to scrutiny. "As it is now, nobody knows the other six criteria," he said, "so how can you know if they make a mistake?"

Meanwhile, heading into the season's final weekend, Dr. Colley has a question of his own. If his laptop crashes, he said, "what'll happen to LSU?"

(excerpted from the New York Times, 12-4-03)