X
Business

DARPA's new plan for machine learning

The U.S. DARPA Agency is starting a new project to develop computer software able to learn and reason in complex military planning jobs by being shown how to perform a task only once. The goal is to replace some military decision by software technology within four years.
Written by Roland Piquepaille, Inactive

The U.S. Defense Advanced Research Projects Agency (DARPA) is starting a new project to develop "computer software able to learn and reason in complex military planning jobs by being shown how to perform a task only once," according to Military & Aerospace Electronics. The Integrated Learner will be able to combine external inputs with its own knowledge generated through reasoning. The goal is to replace military decision by software technology. The project, which is in the proposal stage by now, will last four years. And its possible applications include Air Tasking Order (ATO) planning, CAD for Mechanical Assembly Planning, Bioinformatics or eScience Workflow Management. Once the project is completed, the Integrated Learner should exceed human performance by at least 25%. Read more...

Here are some short quotes from the Military & Aerospace Electronics article.

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.

As the future Integrated Learner software will have to handle many types of information, including the ones it will produce, and to interact with humans to improve its own skills, it can even get confused.

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.

All companies or organizations -- providing they comply with DARPA rules -- can participate in this program which will consist of four one-year phases.

Below is a small diagram showing the preliminary program schedule of this project (Credit: DARPA).

The four phases of the Integrated Learner project

You'll find much more information on this project by reading the Integrated Learning Proposer Information Pamphlet produced by the DARPA's Information Processing Technology Office.

For example, here is a short description of the four phases with the performances required for the software.

  • Phase I: Integrated Learning; Go/No-go: 65% of human performance
  • Phase II: Learning Problem Solving; Go/No-go: 85% of human performance
  • Phase III: Open The Learner, Cost/Benefits, Credit/Blame; Go/No-go: 105% of human performance
  • Phase IV: Learn Meta Processes, Expectation Driven Learning, Sharing; Go/No-go: 125% of human performance

And this program will be focused on "processes in the cyberspace."

In contrast to the physical assembly example used here, the Integrated Learning program will focus on learning generalized plans for processes in the computational world or cyberspace. For this program, the exact choice and articulation of an application is the responsibility of the proposers. Military domains are of interest but civilian application ideas are also welcome.

And below is a list of some sample cyberspace application domains

  • Air Tasking Order (ATO) Planning
  • Delivery AtlasQuest
  • CAD for Mechanical Assembly Planning
  • Bioinformatics
  • Electro-Mechanical Devices Design (CAD)
  • eScience Workflow Management
  • Intelligent Travel Assistant (ITA)

Will this project be successful? You'll have to wait around 2010 to know it.

Sources: John Keller, Military & Aerospace Electronics, September 2005; and DARPA web site

You'll find related stories by following the links below.

Editorial standards