Educating the computer
Lockheed pushes the edge of machine learning for combat-ready systems
Observing a wine steward expertly using a
wine key to open a bottle of pinot noir provides
enough instruction for a novice to at
least roughly imitate the act. But if that same
novice were to observe a nuclear-plant technician
operating a power plant console, the
neophyte wouldn't be ready to perform that
task without a lot more instruction.
However, for a computer loaded with
machine-learning technology, mastering a
complex task after observing it just once
might be possible.
Technology that lets computers learn has
been around for years, but the first phase of a
Defense Advanced Research Projects Agency
project shows that the technology might be
poised for a breakthrough.
Defense Department officials, hoping
the technology can become a reality, are
funding research to help its development.
Lockheed Martin Corp. is working on
the Generalized Integrated Learning
Architecture (GILA), which is a type of
machine learning, said Ken Whitebread, the
company's program manager.
DARPA's goal for GILA is to create new
computer-learning capabilities that let systems
learn complex workflows by
observing warfighters performing
their regular duties. That capability isn't
The program is focused on tasks such as
air operations center planning and military
medical logistics. The learning technology
should make it possible to create many types
of military decision-support systems that
learn by watching experts rather than relying
on hand-encoded knowledge, which is
expensive and error-prone.
Successful testing of the first phase of the
technology led to the phase two award for
Lockheed Martin's Advanced Technology
Air operation centers use Airspace Control
Orders to help manage airspace. Improper
orders endanger pilots.
Under the second phase of the DARPA
project, the company is attempting to enable
a computer to learn to manage combat air space used by manned and unmanned
Lockheed Martin's technology is designed
to help create orders by automatically learning
flight planners' tasks from experts.
"This is fairly new ground in the sense that
it's an approach to using machine learning
that hasn't really been pursued to a great
degree by researchers," Whitebread said.
Some machine learning gives computers
huge amounts of data to master a task.
Another method focuses on learning that
doesn't require large datasets. Lockheed
Martin is working on the latter. The company
and its partners are trying to extend
and evolve that kind of machine learning so
that a computer could learn a task after a
Lockheed is focusing on what the military
calls deconfliction, or making airspace safe.
Deconfliction is one of several broad military
areas that DARPA is pursuing.
"When we picked through all that goes
on within air operations, planning and
management, which as you can imagine
gets very complicated, we felt that this was
a potentially good application to begin
with," Whitebread said.
Air operations work relies on Web-enabled
software tools. Because the tools are Webbased,
and Lockheed Martin already works
with the software, the company had an entry
point to collect information about the task.
Through the Web software, researchers
examine how an expert or competent air
space manager performs duties.
"The idea is we're supposed to learn from
a demonstration of a task," Whitebread said.
"You've got to think about how the computer
program is going to be able to observe the
Web-based software provides a natural
way for a computer to do that. Another
approach ? such as trying to get the technology
to interpret images from a camera ?
would have introduced many other technology
variables outside the machine-learning
The Web-enabled tools are a natural interface
with the task of observing the operator's
Initially, the machine-learning technology
likely will be an added layer of capability
available to the Pentagon, not a total replacement
for the way things are done today.
Following successful learning demonstrations,
researchers face the new and difficult
technical problem of how best to use it in
other military applications.
"One of the things that's part of our program
requirement is to begin interactions
with people within the Air Force involved in
air operations and that sort of thing, to have
an ongoing dialogue with them about what
the potential opportunities are for exploiting
this technology," Whitebread said.
The usability of the technology hinges on
how much background knowledge ? in the
form of hand coding ? a computer needs to
learn a task, Whitebread said.
"If we have to build too much background
knowledge into the computer program for it
to be able to learn, then we've defeated the
purpose," he said.
"The point is to be able to get the system
to learn the task on its own without having
to do extremely expensive development of
the program."Doug Beizer (firstname.lastname@example.org) is a staff
writer at Washington Technology.