Magazine

Introduction


On the eve of Steven Spielberg's new movie, A.I., artificial intelligence researchers flocked to one of the field's newer events, the International MultiConference in Las Vegas. Senior Writer Port was on hand and filed this report. As microprocessors become ever faster and memory-chip capacities increase, computer-based systems such as automated teller machines acquire new features and grow more complex. The result is growing befuddlement for many users. So Japanese researchers at the University of Electro-Communications in Tokyo and Daikoku Denki Co. in Aichi are developing software that spots signs of confusion when someone is using ATMs or other computerized devices. The programs analyze the source of bewilderment and help the user solve the problem or back out of the predicament.

To do so, the programs first must store a detailed model of how the system should be used for various tasks, such as depositing, withdrawing, or transferring money. Then, to catch mix-ups, the software tracks the time the user takes to read each screen of instructions and the sequence in which buttons are tapped and compares these to the model. The software can even monitor a user's eye and arm movements to infer the intended activity, says Shun'ichi Tano, an associate professor of information science. Ultimately, ATMs and other electronic equipment could become so insightful that they'll seem to read a user's mind. Manufacturing has been a perennial challenge for AI researchers. This year, more than two dozen papers focused on improving factory operations--often with the help of programs known as intelligent agents.

Among the standouts at the meeting were agentlike critters known as Holons, pioneered over the past decade by the international Holonic Manufacturing Systems Consortium in Europe. Holons are designed to work together, and each is often linked to a particular shop-floor operation. Thus, an array of Holons can be brought together to represent all the activities in the factory.

But the Holon idea goes beyond simulation, as two teams from Canada's University of Calgary showed. These agents can actually "soft-wire" a factory by controlling devices such as assembly robots or transport dollies. Ultimately, a factory manager wishing to switch from one product to another for production runs of one could instruct the Holons, who would negotiate among themselves and make all the necessary changes on the floor. Techniques for outfoxing the stock market have been a staple of artificial intelligence research since the '80s. "Expert systems," for example, use if-then rules for spotting buying and selling opportunities. And neural networks that mimic the brain's circuitry can learn to detect patterns that usually precede stock-price turning points. In Vegas, it was clear that such approaches can offer far better odds than craps or blackjack.

A group of researchers from Malaysia's University of Kebangsaan, for instance, described a new trick for training a neural net using a special formula to adjust raw stock-price data based on historical patterns. This extra step, dubbed modified returns function, can substantially boost a neural net's accuracy in predicting whether a stock's price will rise or fall the following day. Running tests on canned market data, the Malaysian team claimed it had a 91% hit rate.

Similarly, Thomas Hellstrom, a computer scientist at Umea University in Sweden, reported on another supplemental neural-net training step that augments next-day forecasts by ranking buy/sell opportunities. In a simulation in which an investor acted on the software's top pick for each day, the virtual investor nearly doubled the money it invested every year throughout the five-year period from 1993 to 1997. Androids as intelligent as Gigolo Joe and Gigolo Jane, the pleasure robots in A.I., are still science fiction. But prosthetic devices may soon rival the robotic limbs in the film. The next generation of artificial hands and arms will be much closer to the real thing, say researchers at the University of Medicine & Dentistry of New Jersey.

The team has developed an ultrasensitive neural network that can pick up multiple nerve signals sent to muscles near the stump of an amputated arm. These so-called myoelectric signals are already used in some prostheses, but only for one function at a time. To pick up a glass, an amputee must concentrate first on reaching out, then rotating the wrist, then opening the hand, and finally grasping the glass. The process is mentally and physically taxing, says researcher S. Srinivasan. In contrast, he says, an amputee can train the New Jersey system to execute complex, multifunctional movements with relatively little stress. In tests, four amputees were able to program their artificial arms with sets of muscle signals for executing complex feats such as opening the door of a microwave oven.


Tim Cook's Reboot
LIMITED-TIME OFFER SUBSCRIBE NOW
 
blog comments powered by Disqus