Old problems for testing a new math.

Old problems for testing a new math 

Computers are very good at figuring out the solution to a physics problem, but the heart of intelligence lies in what physicists themselves do: look at a situation and decide what the problem is. This requires the kind of qualitative reasoning--the commonsense physics -- that is so important in artificial intelligence (AI) research. The "qualitative physics" that lets us evaluate whether a bathtub with a leaky plug and the water pouring in from a spout will eventually overflow is something we learn intuitively as children, but it is hard to program into a computer.

What AI specialists need to create programs that design their own problems to solve is a new form of math -- one less exact than the simple math we learn in school but not so general that all the predictive value of the equations is lost, says Brian Williams of Massachusetts Institute of Technology in Cambridge. Other researchers though ot this, but took a look at the task and decided the prospects of finding such a mathematics "didn't look promising," Williams says. However, he has succeeded in inventing just such a "new algebra," called Q1, which can work at a level better suited for posing a problem -- deciding what the problem is that one needs to solve.

Q1 may allow computers not only to solve design problems, but also to suggest new designs that humans hadn't imagined. To test this, Williams is giving a computer all the knowledge available about physics and oil lamps in 300 B.C. He wants to learn whether the computer can, on its own, come up with the oil lamp design that proved a radical invention of its time: a lamp with a self-regulating oil level that Williams says represents the first example of mechanical feedback in history.

"Research suggests [the computer] can," he says.