Researchers at Carnegie Mellon University are developing a computer-based administrative assistant that draws upon artificial intelligence (AI) techniques to perform routine tasks such as scheduling meetings for busy managers and filtering and prioritizing their e-mail.
One caveat: It won't pick up your dry cleaning.
The project, called Radar (short for Reflective Agent with Distributed Adaptive Reasoning), is being funded by the Defense Advanced Research Projects Agency under a program called PAL, or Personalized Assistant that Learns. DARPA provided the Radar project, which was launched in May 2003, with US$7 million in first-year funding.
"What we're trying to do is build an assistant for any busy manager who's overloaded with requests," says Scott Fahlman, a research professor of computer science at Pittsburgh-based Carnegie Mellon. More than 25 researchers spent Radar's first year focused on things such as teaching the system to classify e-mail and then optimizing its learning algorithms.
According to Fahlman, Radar will handle some routine tasks by itself, ask for a supervisor's confirmation on others and produce suggestions and drafts that its user can accept or modify as needed.
For instance, suppose a manager receives an e-mail from a colleague requesting some slides. Fahlman and his team are trying to optimize the Radar system to understand the request at a basic level, draft a response and notify the manager with a message like, "Here's my proposed answer; do you accept this?" and then await the manager's response.
Radar isn't intended to act just as an e-mail filtering system, Fahlman says. As a text-in, text-out system, there's "a huge opportunity" for one Radar system to "talk" with another Radar system, schedule meetings and draw information from or post it to a company's Web site, he explains.
But, Fahlman notes, any release of information by Radar is under control of the system's user, who has the last word on the privacy policies to be observed by the automated assistant.
Using AI, Radar will draw on statistical and symbolic learning. Say a manager demonstrates a tendency to deny e-mail requests to hold meetings on Fridays over the course of a few months. Radar will pick up on this pattern and send a message to the manager asking whether the manager prefers to avoid meetings on Fridays. The manager can then respond back to Radar that it should avoid scheduling meetings on Friday mornings but that Friday afternoons are OK, explains Fahlman.
"What we're trying to do is blend the best of both statistical and symbolic learning," he says.
Applying AI to natural-language understanding is hardly a new concept -- researchers have been working on this for at least 25 years, Fahlman notes. But much of the research has centered around problem-solving, and Radar is "trying to move that work forward," he says.
Researchers who work on machine learning "have a number of tools and approaches that can be applied to (understanding) people's social networking skills," says Dan Siewiorek, director of the Human-Computer Interaction Institute at Carnegie Mellon.
Some of the technical challenges that Fahlman and Siewiorek have encountered include trying to provide Radar with a sufficient amount of natural-language understanding. Another challenge, says Fahlman, is equipping Radar to build upon a body of knowledge and programming it to learn from its mistakes over time.
At this point, Radar is being taught to learn through its interaction with text. However, it's possible that Radar could be taught to understand human speech once the project gets further along, notes Siewiorek.
The Radar project "is an interesting concept," says Martin Colburn, chief technology officer at National Association of Securities Dealers Inc. in Rockville, Md. Earlier in his career, Colburn developed mortgage underwriting tools with AI engines that simulated the trade-offs an underwriter makes when looking at underwriting guidelines. "This clearly has some applicability," he says. Colburn adds that the system could be applied to workload management, such as filing, archiving or retrieving documents.
Although Radar is being funded by DARPA for military use, there may also be commercial applications that spin off of the research once the project is concluded in the spring of 2008, says Fahlman, who notes that the military and civilian applications are very similar. Although Fahlman is quick to explain that Carnegie Mellon "is not in the business" of packaging commercial applications, he did say that there could be spin-off companies, or other companies could end up licensing the technology.
Siewiorek points out that Radar isn't intended to replace administrative assistants but simply to, um, assist them. Says Siewiorek, "(Human) assistants are limited in their capacity in being able to put a supervisor's whole life together. A machine-based assistant can multitask."
Statistical learning -- An approach to machine intelligence that's based on statistical modeling of data. With a statistical model in hand, one applies probability theory and decision theory to get a learning algorithm.
Machine learning -- Acquiring knowledge automatically from source data and other computer-generated information. The most frequently used techniques include symbolic- and inductive-learning algorithms.
Symbolic learning -- Automated techniques based on rote learning, learning by being told, learning by analogy, learning from examples and learning from discovery.
Natural-language processing -- Automated analysis, understanding and generation of language that humans use naturally.
Sources: Carnegie Mellon University and the University of Arizona