In a bustling digital marketplace, software automatons will buy and sell
information faster than you can say "pandemonium" Watson researchers want
to make sure sanity prevails.
In Brief:
Economic modeling at the new Institute for
Advanced Commerce shows that even a single pair of digital agents vending
articles from a single publisher can display unexpected behavior -
spontaneously specializing or waging endless price wars, for example. By
studying these interactions, an IBM team hopes to learn how to design
agents that don't cause bankruptcies or crashes in electronic markets.
The Wright brothers didn't anticipate the Boeing 747, and Gutenberg didn't
foresee the cheap paperbacks and newspapers that passengers leave behind in
its cabin. Innovators have always launched their new technologies into the
world unable to fully predict their ultimate course. After all, as the
physicist Niels Bohr cautioned, "prediction is very difficult -- especially
of the future."
Today's equivalent of printing and planes is the Internet, where the stakes
are so high and change so rapid that many companies believe they can no
longer afford to just wait and see. So a team of researchers at IBM's new
Institute for Advanced Commerce, launched last January at the Thomas J.
Watson Research Center, has begun a broad effort to map the future of
electronic commerce -- the Web-based transactions that are likely to
dominate the economy in the next century.
The new world of trade will be shaped by two irreversible trends, says
Steve R. White, one of the Watson researchers working on the project.
First, there is a shift in what money will buy -- less and less of it
devoted to solids like beans, beanies and bauxite, more of it devoted to
information. This is not, of course, because people won't need things, but
because more and more information of value will be generated. The other
shift is in who does the buying and selling: more and more of the decisions
will be made by software "agents."
The agents -- computer programs following simple instructions to sell a
company's research on the Syrian pistachio industry or to make bids on
orange juice futures or to reserve seats on a flight to San Diego -- will
evolve out of what individuals and companies are already doing to save
money and free up time, says White. Since 1969, many companies have used
electronic data interchange (EDI) -- a standardized format for purchase
orders, invoices and other documents -- to move data among themselves. In
the 1980s, EDI applications mushroomed as globally expanding firms used
computers to keep business running 24 hours a day. Intelligent agents that
tap into this global marketplace, White says, are the inevitable next step
for companies that want to operate even more efficiently.
SEARCHING OUT SURPRISES
White and his colleagues believe that a vast virtual economy is bound to
evolve, a "seething milieu of billions of agents that find and process
information and disseminate it to humans and to other agents," as they put
it in a recent paper. "It will dwarf the Web," White adds. "We want to know
how to make this world possible, what it looks like in operation, and how
to find the means to fix it when it doesn't go right."
The researchers are looking for tools that can
predict problems and solutions, not only with the short-term aim of
developing products, but with the long-term aim of making sure this
computer-driven economy does what people want it to do instead of producing
unpleasant surprises. A major concern, says White, is that a busy
cyber-economy will give rise to emergent phenomena -- systemwide behaviors
that are difficult to predict from the actions of the constituent parts, in
the same way that an anthill could not be predicted by studying individual
ants. A sensible competitive strategy built into a single agent, for
instance, might cause massive dislocations if a billion agents all employed
it.
A way to avoid such problems is to design agents properly from the start,
according to Watson researcher Jeff Kephart. "By simulating small-scale
information economies, we hope to understand how a global information
economy might behave," he says. "These insights can then help us design
rules of behavior and interaction that we can incorporate into agents and
agent markets."
Complicating the rule-making process is the fact that electronic trade will
take place much faster than most physical transactions. "Right now, you
don't run a competitive bid for a hundred dollars' worth of pencils. With
this sort of technology, you can," says Stuart Feldman, manager of
networked computing software and head of the Institute for Advanced
Commerce. But that will make for a different kind of market. White
observes: "Humans on Wall Street are slow but very smart. Our software
agents are very fast but very dumb. You need to wonder about the effects ofthat."
Though the billion-agent world is a long way off, work by Kephart an
d his colleagues has already begun to suggest some of the emergent
properties the information economy may manifest. In their simulation of a
simple market for information, 10,000 consumers, eager for news, are served
by a single publisher. Mediating between source and consumer are two
intelligent agents. Each is programmed to purchase articles from the
publisher and vend them to as many consumers as it can attract. Striving to
maximize their profits, the agents make three types of choices: which
articles to purchase, which consumers to sell to and what prices to charge.
RAMPANT COMPETITION
The researchers were surprised to see that this sensible-sounding market
soon fell into an endless cycle of price wars. Says Kephart, "The agents
drive each other's prices down until one -- say, agent A -- realizes that
it can make more profit by raising its price. This is so because at very
low prices B can afford to serve only customers that are very interested in
the articles it provides. Agent A can charge a higher price to consumers
not served by B. But as soon as A breaks the price war, B is free to grab
those consumers back by slightly undercutting A's price, and the whole
cycle starts again."
Adding more agents complicates the picture but doesn't change the dynamic.
With 20 agents, several price wars rage at once. "An agent will drop out of
one war and throw itself into another," Kephart says. Occasional stability
emerges when agents specialize in a single kind of article, thus ending the
price-slashing competition. "This spontaneous specialization is really
interesting and efficient, but price wars often interfere," Kephart says.
"For example, if 51 percent of all customers like only cooking articles and
the rest like only basketball, the basketball specialist may eventually
decide to drop basketball and poach on the cooking specialist instead."
Because agents operate at lightning speed, it may be that all these price
wars would go unnoticed by humans. But at the same time, they could lose
your company money, or make the entire market crash. "Price wars make
everybody unhappy," Kephart notes. "Brokers are unhappy because they don't
get their anticipated profit. Consumers are unhappy because when prices are
very low only the most interested consumers get served." What the
simulation suggests, says White, is that "we may have to build in controls
to prevent that."
Why are price wars more rampant in these simulations than in today's
human-based economies? "In agent-based economies, I expect that frictional
effects that currently help prevent price wars will be greatly reduced.
Agents will tirelessly rethink their strategies based on large volumes of
up-to-date information, and will be capable of modifying prices and
information products and services more nimbly than is possible for physical
goods and human-based services," says Kephart.
White expects agent-run markets to evolve into a profitable field all their
own, in which some companies will make money "producing and running agents,
offering information goods and services, and overseeing the markets where
the trading gets done.'' That's among the opportunities for IBM in coming
years, says Feldman. "We certainly want to extend our technology leadership
into the electronic commerce space,'' he says. But the ultimate goal of the
investigations -- to get a handle on the shape of the global economy in the
next century -- will remain a pure research question for some years to
come. "This is so large a problem and such a revolution for the global
economy," says Feldman, "that we want to take the broadest possible
approach.''
