After trying for years to design a thinking computer,
scientists working in artificial intelligence have a new
approach. The best way to get a computer to think and behave like
a human, a growing number of experts say, may be to let it evolve
Artificial intelligence specialists are designing computer
programs that apply principles of genetics and evolution to their
own development, enabling the programs to grow, adapt and even
fend off competition in a kind of virtual natural selection
process, or "artificial life."
"The contrast is between building a robot. . . and letting it
build itself," said Daniel C. Dennett, director of the Center for
Cognitive Studies at Tufts University.
Programmers provide the "nature": a mathematical framework
similar to the human genetic code. But in artificial life as in
human life, nurture _ interaction with the environment _ plays a
critical role in development.
In one program gaining attention in computer circles around
Cambridge, Mass., for example, virtual "creatures" are placed in
an environment where they must compete for symbolic "food" to
The programmer, Karl Sims, formerly of Thinking Machines,
gives the creatures several options, such as sprouting limbs to
fend off rivals. But the entities choose among such options
themselves, after testing to see what works best.
"It has to start the way a baby starts _ with a tremendous
amount of design," Dennett said. "That's a big head start. But
then, it's going to find out for itself all the things about how
the world works."
"What we're really talking about here is learning," said
Patrick H. Winston, director of MIT's Artificial Intelligence
In the past, artificial intelligence programs emphasized more
straightforward programming, where information on how to "react"
to any situation was fed to the computer. In artificial life,
computers are given some leeway to change the shape and
composition of their programs as they interact with the
And there's growing conviction in computer science circles
that this approach will produce machines with a rich, deep and
versatile intelligence _ and perhaps, eventually, the ability to
elude human control.
Because programmers will not dictate the step-by-step
development of such machines, they also won't be entirely
responsible for the ultimate product, said Dennett. "The smarter
they get, the more impossible it will be to control them."
Versatile and autonomous thinking machines have been predicted
before, of course. Artificial intelligence, first begun as a
field in the '50s, was all the rage in the '80s as companies
promised systems that would mimic the finest stock trader or the
most nimble surgeon.
But most of those companies, many of them based in an area
around Massachusetts Institute of Technology and Kendall Square
in Cambridge known as "AI Alley," went bust. Artificial
intelligence concepts could produce an excellent chess-playing
computer or credit-card checker, but failed to meet expectations
in more demanding applications.
These days, some regard artificial intelligence as a shopworn
concept, a symbol of the unrealistic promises of those who
breathlessly make predictions about technology.
Yet artificial intelligence is a quiet reality in the computer
world in modest ways. So-called "expert systems," a kind of
artificial intelligence, are embedded in personal computer
software, whether spreadsheets or interactive card games.
Scientists at MIT's Artificial Intelligence Laboratory and
elsewhere have not abandoned the goal of thinking machines. It is
the path to that goal that has markedly changed _ from top-down
programming to standing back and letting a hundred bits bloom.
"Artificial life is the next stage," said Harvey Newquist,
author of the 1994 book "The Brain Makers" (MacMillan), which
chronicled the rise and fall of AI companies. …