Kephart says his team's findings could have a much greater impact than the famous victory of IBM's Deep Blue supercomputer over chess supremo Gary Kasparov. "The impact might be measured in billions of dollars annually," he says.
A commodities trader buys and sells goods, which are usually agricultural or mineral-based, like pork bellies or gold. Their aim is to buy low and sell high. While robotic trading agents have been entered in competitions before, they had never competed against people.
In IBM's test, software-based robotic trading agents-known as "bots"-made seven per cent more cash than people. Both bots and people had the same set-up, allowing them to trade through an unbiased software-based auctioneer. The auction was designed to mimic the kind of commodities market where buyers and sellers have a fixed amount of time to trade in a single commodity.
Six bots and six people traded against each other. Half were buyers and half were sellers. Buyers were given an upper spending limit, while sellers had a minimum sale price. Their goal was to maximise their profit at the end of trading.
The bots used very basic strategies. Some tried to make better offers incrementally, in the hope that they could strike a deal, while a more successful version tried to work out the best price to trade at, by calculating its probability of success based on the form.
Back in Britain, Dave Cliff, a researcher at Hewlett-Packard's research labs in Bristol who developed one of the trading algorithms used in IBM's experiment, described their finding as "significant". "I never designed it to outperform humans, so for me it's a very pleasant surprise that it does," says Cliff.
"If it were up against the Gary Kasparov of trade I would guess that that person would beat these agents," Kephart says. But he points out that these early robotic agents were very simple indeed.
Kephart believes that in the future billions of economic robotic agents will replace people in making these sorts of financial decisions on behalf of businesses. "We see agents being down there in the frenzy of the trading pit while humans are elevated to a more managerial role," he says.
Author : Duncan Graham-Rowe reports from Seattle on this week's International Joint Conference on Artificial Intelligence.
New Scientist issue: 11 August 2001
Please mention new scientist as the source of this story and, if publishing online, please carry a hyperlink to: http://www.newscientist.com