Making Machines Think-and Act--Like Us
In the world of artificial intelligence, computer scientists have excelled at making computers do what they do best--compute. Faster, more complex, more comprehensive and less error-prone than humans, computers can calculate, connect and retrieve information in the twitch of a whisker. Yet despite their remarkable abilities, there continues to be a push to make computers more like us. And we're getting closer, for better or worse.
Jim Hendler is a professor in the Department of Computer Science currently working with the Defense Advanced Research Projects Agency, or DARPA, on the development of "intelligent agents." Such agents, says Hendler, "can be significantly more capable than other software approaches in performing tasks based on human goals and reacting to changes in the environment based on human needs."
Take, for example, travel planning, a popular application for computer users who book flights, rent cars and obtain hotel rates online. Sites such as Travelocity and Expedia are immensely useful, especially for booking basic travel services. But where the sites don't perform as well is in the realm of unstated preferences. These could include one's preference for non-smoking rooms, the ability to drive a stick-shift car, or a need for an aisle seat. What Hendler and his colleagues seek is the computer equivalent of a human travel agent--software that can search for your preferences, make decisions and then act on your behalf.
"A good Internet agent," Hendler explains, "like a good travel agent, needs to be communicative--able to understand your goals, preferences and constraints. It must be capable--able to take options. It must be autonomous--able to act on its own; and it should be adaptive--able to learn from experience."
While current search engines can recognize key words and even word relationships, they are limited because they have no knowledge of the subject they are searching. This is especially frustrating for scientists using the Web to stay current on research in their fields. "It would help if an engine searching a physics site had some knowledge of how experiments are perform-ed, which papers are theoretical and which empirical, and so on," Hendler says. After all, "not all squids are superconducting quantum interference devices."
This problem can be resolved, he continues, by an ontology--a formal definition of a body of knowledge--which employs class and subclass categories coupled with definitions about the relationships between things. A structure for creating onotologies developed at the University of Maryland, called SHOE, for Simple HTML Ontology Extensions, helps Web site developers annotate their pages to include knowledge that intelligent agents can actually read. The technology is still developing, and the early ontologies are "relatively shallow," Hendler says. "But the upside is that it can be written without specialized computer science training." As such, ontologies are being developed and applied at a rapid rate at universities and research centers around the world.
While there is some debate over how much autonomy computers should have over human decisions, the race to create such intelligent agents is advancing rapidly. Applications range from e-business (as anyone who has clicked an Amazon.com "recommendations" link has discovered) and to scientific research and information dissemination to the kind of personal assistant Hendler uses in his travel agent analogy. Such agents will soon be showing up in Palm Pilots and cell phones, he believes.
Hendler foresees the day when intelligent agent computer programs will simplify logistics for complicated military operations, or help a space station crew get more work done. Any complex planning task, like manufacturing supply chains, can be made more efficient with these agents. "If you're not using agent-based technology now," he says, "don't worry. You soon will be." --JB
To learn more about SHOE and the structure of ontologies, visit
www.cs.umd.edu/ projects/plus/SHOE/
