Intelligent agents getting practical
From laboratory curiosity to practical application
The Arizona Republic
March 27, 2001
You're on your way home from work. Your dashboard computer warns you that there is slowing traffic ahead. It suggests that you get off at the next exit and take an alternate route.
The computer's software application, called an intelligent agent, is able to do this because it's been monitoring traffic conditions and your driving habits. It has concluded that on a Wednesday night at 7:30, on this highway, you are probably headed home. It knows that you sometimes choose the less-direct route. When it gets word from a traffic monitoring service of the traffic jam, it directs you to the alternate route.
Intelligent agents, in their most common form, are like little personal assistants. They can swim into a computer network, track down information, make decisions and report back to you with the results. And they can learn from their experiences.
Agents have been a laboratory curiosity for years. And some simpler examples have escaped the lab and are helping us navigate the Web or teach us how to use complex computer programs.
Katia Sycara, a senior research scientist at Carnegie-Mellon University, thinks it may not be many years before full-fledged agents start picking up some slack in our daily regimens.
Sometime this fall, Sycara expects to meet with a major car manufacturer interested in her on-board navigation project along the lines of the above example.
Definitions vary widely as to what is or isn't an agent. Sycara says there are three key components to a true intelligent agent.
First, it can operate from a general goal and does not need to be told every little step to take. Conventional software follows specific directions - "in this scenario, do this." An intelligent agent makes the little decisions - "in this scenario, I'll do this" - to meet its goal.
Second, an agent is social. It is able to communicate with people, networks and other agents.
Third, it can adapt and learn.
"You don't tell it exactly what to do," she said. "You give it a high-level goal. It needs to reason in some way and make suggestions."
The navigator agent isn't smarter than its human owner. It doesn't do anything that a driver cannot do by listening to traffic reports on the radio, Sycara says. It just does it for you. Over time, it learns about rush-hour traffic. It learns that you drive straight home from work on most nights. When you set out on your regular route home, you don't have to tell the agent where you're going. It figures it out. If there are no problems, you don't hear from it.
"Maybe you'd rather listen to classical music than the traffic reports," the researcher said.
Sycara is also working on a demonstration to use agents to filter phone calls and electronic intrusions during a trip.
"If my phone is ringing all the time, I'm not concentrating on driving," she said. "The agent can decide whether a phone call is relevant to the current trip. It can filter my phone calls."
Anyone who has used a recent version of Microsoft Word has likely experienced a simple agent application. The animated paper clip "help" icon attempts to use intelligence to understand questions asked in sentence form instead of keyword queries.
Marcelo Hoffman, SRI Consulting's senior consultant, says most users describe the paper clip as ineffectual and obnoxious. But Microsoft will tell you the obnoxious icon saves the company "a tremendous amount of money" in questions deferred from the help desk, he said.
Another example is the message at Amazon.com that greets return customers, with "You may be interested in . . . " The agent checks previous purchases and bases suggestions for other products on the earlier choices.
"It's an agent, but not a very intelligent one," said Brock Hinzman, SRI Consulting's technology navigator. "Intelligence implies that the agent takes all the information it gathers and uses it at a later date for a different purpose."
Hinzman and Hoffman say that today's agents are generally relegated to simple, repetitive tasks, such as constantly checking several travel Web sites, shopping for the best price.
Researchers at universities such as Carnegie-Mellon and Massachusetts Institute of Technology are exploring a variety of agent uses. Many involve electronic commerce and mobile commerce. One at MIT helps renters find apartments. The agent offers several choices and learns from a person's responses. It then offers a more refined selection and repeats the process.
At Carnegie Mellon, Sycara has a demonstration agent project called Warren, named after investor Warren Buffett. Warren's agents acquire information from, and monitor changes to, stock-market databases. They interpret stock information, track and filter relevant financial news articles, and notify users of important changes.
Sycara says Warren could be developed into a consumer product before long. At this point, it remains a demonstration. She says she hasn't yet trusted it to manage her portfolio.
"It's just working with imaginary money right now," she said.
SRI Consulting's technologists say there are several hurdles before agents become significant forces in consumers' lives. Security is one big issue, they say. Before agents are set loose with credit-card numbers, security will have to be more hacker-proof than it is today.
Another issue is agents' inability to understand the content or meaning of the Web pages they explore.
Today's agents are not quite ready for prime time, Hinzman said.
"I subscribe to a particular newspaper because I know its editorial position and I know the type of stories it will run. They have a skilled team of human agents that are so much better than software agents," he said.
Reach the reporter at firstname.lastname@example.org or at (602) 444-8269.