About Scope    Editor’s Mailbox    Back Issues    Skidmore Home


Spring 2004

- - - - - - - - - -

Contents

Features

Letters

Books

Observations

Who, What, When

Centennial spotlight

On campus

Faculty focus

Arts on view

Sports

Advancement

Class notes

 
 
He got game

You know how computer scientists are either neatnik control-freaks with pocket protectors or else absent-minded zombies with permanent bed-head? You don’t know Tom O’Connell.

His office may be a tornadic disaster zone, but this guy has social skills, a good vocabulary—and a strong intellectual streak. Skidmore’s first computer science PhD hired into a tenure track, he was an IBM programmer until he “got tired of fixing the same problems in the same ways” and quit to enter grad school. He was intrigued by artificial intelligence: computer programs that can learn from experience and even use “neural networks” to figure out complicated tasks. Citing the US Postal Service’s work on computer recognition of handwritten ZIP codes, he explains, “Imagine you encounter a squiggle that’s going to turn out to be either a one or a seven—how do you get from seeing that squiggle to identifying the numeral?”

AI’s neural networks are also useful in game theory—the ideas behind computers that can sometimes defeat expert chess players and can always win or draw in tic-tac-toe. In his doctoral research on game theory, O’Connell focused on “strategic entropy.” Defined as “a measure of the cost of randomization,” it figures into how a player can maximize the gain made with each move but minimize the information given away to the opponent. While previous theorists had posited this kind of entropy, O’Connell and his faculty mentor confirmed and calculated its role in an exact formula. Here’s just a modest sample of some related calculations:

Somehow O’Connell is able to disentangle his brain from such abstruse mathematics and relate to undergraduates. Dave Cadwallader ’04 is researching a dots-and-boxes game pitting a reinforcement-learning player against a standard fixed strategy. “It’s very easy to get excited in this field,” he says, but he credits O’Connell with “helping me focus my interests to design a project that’s practical.”

Relatively practical. For even a simple game, a learning program may need to run for a solid week, playing 20,000 games or more, to compile enough statistics about which moves work best in which situations. “You can see when it learns a strategy and improves its win percentage,” says Cadwallader, “but then it plateaus. So you adjust the program, to tell it more about what to look for in assessing the state of the board.” His dots-and-boxes player now has a 20 percent win rate, and he’s aiming higher. Meanwhile O’Connell is curious about games in which players don’t see the whole board —poker, for example, with its hole cards and bluffing strategies.

O’Connell comes to gaming naturally: his father, a horseracing fan, hired him as a teenager to research speed charts and other stats. “When I saw my dad winning at least 15 percent of the time on 6-to-1 or longer odds, I realized, Hey, this could work.” The pair even entered tournaments: “They’d start you with a $1,000 bankroll,” O’Connell recalls. “Our method was to bet in just three races and try to win twice. First we put 30 percent of our bankroll on one horse, on the nose; if we won, we’d put 30 percent of our now-larger bankroll on our next horse. One year we hit two early races and had about $10,000—usually enough to win the tournament—but a few other bettors hit big in the last races and beat us.”
That’s OK. For O’Connell, it really is “how you play the game”—exactly how—that’s most important. —SR





 

 


© 2004 Skidmore College