Save Article Instructions
Close 

DARPA takes another look at improving machine learning

By John Keller

ARLINGTON, Va. - Scientists at the U.S. Defense Advanced Research Projects Agency (DARPA) in Arlington, Va., are approaching industry about a research project to advance the state of the art in computer learning.

The DARPA Information Processing Technology Office has a broad agency announcement on the street, BAA 05-43 V, for the Integrated Learning project, which seeks to develop machine learning computer software able to learn and reason in complex military planning jobs by being shown how to perform a task only once.

The software will learn by assembling knowledge from different sources-including generating knowledge by reasoning. DARPA scientists intend the Integrated Learning software technology to enable inexpensive military decision- and planning-support systems.

Potential applications of this technology include learning air-tasking-order (ATO) planning or computer-aided design processes, DARPA officials say.

The software has to combine limited observations with subject expertise, general knowledge, reasoning, and by asking what-if questions. The Integrated Learner also will have explicit learning goals, keep track of what it does not know, what it needs to know, as well as track and reason about its uncertainties.

The Integrated Learner software will attempt to figure things out, as well as tolerate errors and missing information by using whatever information or reasoning is available.

Integrated Learners must be able to manipulate many different forms of information and even trade off different types of information and reasoning, as well as interact with humans to fill in information gaps.

Different parts of the software can produce opposite conclusions, DARPA officials point out. When this happens, the software must find information that proves or disproves one conclusion or the other. If the software is able to learn or be told which components are likely to be right or wrong in a given circumstance, it can better manage inconsistencies.

Companies whose experts want to participate in this program must specify what their software will learn, using models with features like sequencing, precedence, temporal interactions, conditional performance, utility-driven choice, hierarchical organization, and joint action.

Each proposal from industry must specify an Integrated Learning software framework, such as common knowledge structures; reasoning components in the Integrated Learner, such as domain planners, simulation components such as tools that enable the learner to ask what-if questions about the world, general world knowledge and domain knowledge the learner will use, the means of interacting with the human user, and evaluation plans and metrics.

DARPA officials are particularly interested in applying the Integrated Learning software to tasks such as air tasking order (ATO) planning by ATO generation.

Air operations campaign planning is a complicated time-sensitive process that involves hundreds of people using different software systems to plan the activities of hundreds or thousands of aircraft, crews, support staff, and support logistics, DARPA officials say.

The ATO provides detailed information about scheduled missions, mission plans, participants, goals, schedules, resource assignments, and contingencies for a 24-hour duty cycle.

When forming the ATO, the human planners interact with many different software tools that perform functions like scheduling analysis, target and resource analysis, reconnaissance goal and resource analysis, coordinating airspace, and logistics analysis.

The Integrated Learner must learn the human ATO planning process or a significant and well-defined portion of the planning process.

The program is expected to have four one-year phases-of which only phase one has funding at the moment. Nevertheless, those submitting proposals should address all four phases. DARPA will accept proposals through 11 July 2006.

More details about the Integrated Learning research program are in the proposer information pamphlet, which is available online at www.darpa.mil/ipto/solicitations/open/05-43_PIP.htm.

For more information on the program, contact Dr. Tom Wagner at DARPA by e-mail at BAA05-43@darpa.mil, or by fax at 703-741-7804, or by post at DARPA/IPTO, Attn: BAA 05-43, 3701 N. Fairfax Drive, Arlington, Va., 22203-1714.


To access this Article, go to:
http://www.militaryaerospace.com/content/mae/en/articles/print/volume-16/issue-9/news/darpa-takes-another-look-at-improving-machine-learning.html