Posted on Wed, Jun. 01, 2005

Pentagon envisions electronic office assistant for busy human bosses

Knight Ridder Newspapers

WASHINGTON - (KRT) - With a strong push from the Pentagon, computer scientists are trying to create an artificial "personal office assistant" that's smart enough to handle routine tasks for a human boss, military or civilian.

The researchers aim to build an electronic system that understands human language, takes and remembers instructions, learns from its experiences and copes with unexpected situations.

It won't make coffee, but it also won't grumble or demand a raise.

The automated aide-de-camp is supposed to be able to sort e-mail, schedule meetings, make plane reservations, collect information for reports and carry out other humdrum, time-consuming chores for busy human managers.

Although the duties seem routine, creating a software program that can handle them is one of the most difficult challenges in computer science. Artificial-intelligence experts have struggled for years to make machines perform functions that are simple for people but stump electronic devices.

Today's increasing computing speed and power, however, make things that were impossible five or 10 years ago more practical, researchers said. "Progress has been slow but steady," Eric Mathews, the associate director of the Institute for Intelligent Systems at the University of Memphis, said in an e-mail message.

The office assistant program is sponsored by the Defense Advanced Research Projects Agency, a Pentagon unit that pioneered such once blue-sky developments as the Internet, stealth aircraft and microelectronic machines.

DARPA Director Anthony Tether told the House Science Committee last month that his agency is moving into the field of "cognitive computing," meaning computer systems that "perceive, reason and learn," not just crunch numbers and manipulate data. The Pentagon project is called PAL, an acronym for "personalized assistant that learns."

"Cognitive systems that learn to adapt to their users could dramatically improve a wide range of military operations," said Ronald Brachman, the director of DARPA's Information Processing Technology Office. "They could learn and even improve on their own."

The PAL program aims to "make military decision-making more efficient and more effective at all levels, from the individual soldier to the high-level commander, and to reduce risk for humans," Brachman said. The system is supposed to "perform well in specific scenarios that are exactly like those that a human executive assistant would face."

For work on PAL, DARPA so far has granted $22 million to SRI International, a research organization in Menlo Park, Calif., and $7 million to Carnegie Mellon University in Pittsburgh. More than 20 universities and research laboratories are contributing to the effort, which was launched in 2003.

Researchers at these and other organizations already have produced numerous more or less successful artificial-intelligence programs that can carry out bits and pieces of DARPA's vision. There are many language translators, speech recognizers, e-mail sorters, report summarizers, calendar managers and the like on the market. The PAL project is an ambitious attempt to integrate these scattered systems into a whole.

"It forces researchers from different subfields to work together on problems we associate with human-level intelligence," John Laird, a computer science professor at the University of Michigan in Ann Arbor, said in an e-mail message.

Among other skills, DARPA wants its office assistant to be able to:

_Learn by observing a human partner or by being told something directly.

_Be aware of events as they happen.

_Decide what to do and act in real time.

_Remember its experiences and recall them when needed.

"We are making progress in all areas, but the predominant emphasis and progress is in (computer) learning," DARPA spokeswoman Jan Walker said.

Some computer scientists are skeptical that DARPA, despite its many technological breakthroughs, can make an artificial office assistant helpful in the real world.

The goal is "highly unrealistic," said Bram van Heuveln, the director of the Minds and Machines Program at Rensselaer Polytechnic Institute in Troy, N.Y.

"Trying to get a computer system to perform these kinds of mundane tasks has plagued AI (artificial intelligence) research from the very start," van Heuveln said. "Some say this is just a matter of producing faster and more powerful processors, but I myself think the problem is much deeper. Powerful and reusable cognitive-processing technology is not in the offing anytime soon, certainly not in the next few years."

Laird disagreed: "It is very realistic to expect that software using AI techniques will aid office management in the next few years where there is a real need, not just in the office but also the battlefield."

SRI's part of the project is called CALO, short for a "cognitive agent that learns and organizes." In Latin, "calo" means a soldier's servant. It runs on a personal computer, a laptop or, with limitations, a cell phone.

CALO already has learned how to suggest folders in which a human boss might like his or her incoming e-mail filed. It knows how to schedule meetings among four groups of four people each.

The system will have cameras that can keep track of who's attending a meeting and remember who was there for future reference, said Mark Drummond, an artificial-intelligence expert at SRI.

"This is a really cool machine-learning technique," Drummond said. "The system learns while it's being used."

Carnegie Mellon's personal assistant is called RADAR, after the young corporal on the old "M*A*S*H" TV show who always seemed to know what was going to happen before the officers did.

"Like any good assistant, RADAR must understand its human master's activities and preferences and how they change over time," said Scott Fahlman, a computer scientist at the university's Human-Computer Interaction Institute.

For example, Fahlman said in a Carnegie Mellon release: "RADAR must respond to specific instructions such as `Notify me as soon as the new budget numbers arrive by e-mail.' It must know when to interrupt its master with a question and when to defer."