Researchers Push for Computers That Think

Remember Stanley Kubrick's epic 2001: A Space Odyssey? More specifically, remember the malevolent supercomputer HAL 9000?

It may still be decades before computers become "self-aware" and smart enough to do something evil on their own, like turn against humans. But the Defense Advanced Research Projects Agency is hoping that renewed research and funding could within the next few years lead to machines that act more like human assistants.

For years, DARPA has funded research into so-called artificial intelligence systems — computers that approach tasks using advanced software and schemes that mimic how humans solve problems. But under a new, five-year project called Perceptive Assistant that Learns, or PAL, researchers are hoping to take "cognitive systems" to a new level.

The goal of the project is to develop a computer system that would help decision-makers — business managers and battlefield commanders, for instance — automatically manage the flood of daily mundane chores and allow them to concentrate on more important tasks.

Just like human administrative assistants, these developing "cognitive systems" would answer e-mail, arrange meetings, and write reports. But in handling these routine tasks, the computer would remain alert for information and events that are vital to its human manager's chores.

An Astute RADAR

One team of researchers at Carnegie Mellon University in Pittsburgh is tackling the PAL program challenge with a project called Reflective Agents with Distributed Adaptive Reasoning — RADAR. The project, funded for the initial year with $7 million from DARPA, brings together several "expert systems" that have been developed over the years for very specific purposes.

For example, voice recognition systems used in automated call centers to give information about credit cards or bank accounts use specially developed software that helps a computer understand a very limited number of spoken commands and phrases.

But researchers at the university hope that further study and development of these types of specific expert systems will lead to a common approach and methodology. By working from a common "base" of knowledge and structure, different parts of RADAR will know, or learn, how to interact with others.

According to Scott Fuhlman, one of the computer scientists working on the CMU project, the RADAR component that handles e-mail could be taught how to handle certain e-mails, looking for key phrases like "meeting" and times associated with them. When it finds such messages, it would then alert the RADAR piece that handles the person's scheduling, which in turn may directly interact with the e-mail sender's agent to determine a mutually agreeable meeting time.

"Everything I've described has a simple version that people have worked on and used in isolation in the past," says Fuhlman. "But by pulling them all together, we can create a new capability."

A Soldier’s Servant

SRI International, a nonprofit research organization in Menlo Park, Calif., is also tackling DARPA's PAL program with its own project called Cognitive Agent that Learns and Observes, or CALO.

Like the CMU project, SRI's CALO program will pull together different bits of "expert software" and technology that have been researched and developed in 20 universities and other research institutions, such as Boeing Phantom Works in Seattle. With $22 million in funding from DARPA, SRI hopes to develop a cognitive system specifically geared toward managing the flood of information and tasks that arise when the military faces an important decision-making process.

William Mark, vice president of the information and computing sciences division of SRI, says the military has developed specialized expert systems over the years. But CALO researchers hope that over the next five years they can develop a software system that can link all of these separate systems together.

"Right now, [expert] systems are designed as isolated engines, optimized for specific tasks and performances," says Mark. "We're turning that on its head. We're designing each piece of the [CALO] system so that it's part of the whole."

Speeding Along?

Researchers at both SRI and Carnegie Mellon say it will take years of work before computer programs are capable of handling informational tasks as aptly as a human assistant. But they are confident that the rapid advances in computing power will — eventually — help bring about smarter, more adaptive computers.

"Today's computer systems run faster, so computation-intensive tasks — speech processing, reasoning — can be performed in time frames we [normally] react in," says Mark. "Things that weren't tractable five years ago — speech recognition or natural language dialog systems, for example — are now capable."

Still, others note that we're a long way from the creation of a HAL-like computer.

"There's a large amount of [software] programming, knowledge engineering — getting knowledge into a format [computers] can digest," says CMU's Fuhlman. "There's an enormous amount of engineering that has to happen to make this all mesh."