Educating the computer

Observing a wine steward expertly using awine key to open a bottle of pinot noir providesenough instruction for a novice to atleast roughly imitate the act. But if that samenovice were to observe a nuclear-plant technicianoperating a power plant console, theneophyte wouldn't be ready to perform thattask without a lot more instruction.However, for a computer loaded withmachine-learning technology, mastering acomplex task after observing it just oncemight be possible.Technology that lets computers learn hasbeen around for years, but the first phase of aDefense Advanced Research Projects Agencyproject shows that the technology might bepoised for a breakthrough.Defense Department officials, hopingthe technology can become a reality, arefunding research to help its development.Lockheed Martin Corp. is working onthe Generalized Integrated LearningArchitecture (GILA), which is a type ofmachine learning, said Ken Whitebread, thecompany's program manager.DARPA's goal for GILA is to create newcomputer-learning capabilities that let systemslearn complex workflows byobserving warfighters performingtheir regular duties. That capability isn'tavailable today.The program is focused on tasks such asair operations center planning and militarymedical logistics. The learning technologyshould make it possible to create many typesof military decision-support systems thatlearn by watching experts rather than relyingon hand-encoded knowledge, which isexpensive and error-prone.Successful testing of the first phase of thetechnology led to the phase two award forLockheed Martin's Advanced TechnologyLaboratories.Air operation centers use Airspace ControlOrders to help manage airspace. Improperorders endanger pilots.Under the second phase of the DARPAproject, the company is attempting to enablea computer to learn to manage combat air space used by manned and unmannedaircraft.Lockheed Martin's technology is designedto help create orders by automatically learningflight planners' tasks from experts."This is fairly new ground in the sense thatit's an approach to using machine learningthat hasn't really been pursued to a greatdegree by researchers," Whitebread said.Some machine learning gives computershuge amounts of data to master a task.Another method focuses on learning thatdoesn't require large datasets. LockheedMartin is working on the latter. The companyand its partners are trying to extendand evolve that kind of machine learning sothat a computer could learn a task after asingle demonstration.Lockheed is focusing on what the militarycalls deconfliction, or making airspace safe.Deconfliction is one of several broad militaryareas that DARPA is pursuing."When we picked through all that goeson within air operations, planning andmanagement, which as you can imaginegets very complicated, we felt that this wasa potentially good application to beginwith," Whitebread said.Air operations work relies on Web-enabledsoftware tools. Because the tools are Webbased,and Lockheed Martin already workswith the software, the company had an entrypoint to collect information about the task.Through the Web software, researchersexamine how an expert or competent airspace manager performs duties."The idea is we're supposed to learn froma demonstration of a task," Whitebread said."You've got to think about how the computerprogram is going to be able to observe thetask."Web-based software provides a naturalway for a computer to do that. Anotherapproach ? such as trying to get the technologyto interpret images from a camera ?would have introduced many other technologyvariables outside the machine-learningproblem.The Web-enabled tools are a natural interfacewith the task of observing the operator'sactions.Initially, the machine-learning technologylikely will be an added layer of capabilityavailable to the Pentagon, not a total replacementfor the way things are done today.Following successful learning demonstrations,researchers face the new and difficulttechnical problem of how best to use it inother military applications."One of the things that's part of our programrequirement is to begin interactionswith people within the Air Force involved inair operations and that sort of thing, to havean ongoing dialogue with them about whatthe potential opportunities are for exploitingthis technology," Whitebread said.The usability of the technology hinges onhow much background knowledge ? in theform of hand coding ? a computer needs tolearn a task, Whitebread said."If we have to build too much backgroundknowledge into the computer program for itto be able to learn, then we've defeated thepurpose," he said."The point is to be able to get the systemto learn the task on its own without havingto do extremely expensive development ofthe program."