Debriefable Agents
The purpose of the Debriefable Agents project is to develop
technology that enables Soar-based problem solvers to describe and
explain their problem solving activities. This technology,
inplemented in a system called Debrief,
enables
problem solvers to engage in a question-answer dialog, in which
the user can ask the problem solver about the actions it took and
the reasoning behind those actions. This technology is being
employed in the
Intelligent Forces domain, and will be employed in the
Virtual Environments
for Training domain. There is also substantial sharing with the
I-Doc project.
In order to make agents debriefable, the following capabilities
are added to an existing Soar agent.
- An episodic memory of events and states. This is implemented
via productions and operators that monitor significant events and
state changes during the problem solving activity, and build
recognition chunks to aid in the subsequent recall of problem solving states.
- Analysis of previous decisions through mental simulation. In
order to determine the rationales for the agents' decisions, Debrief
recreates the problem solving state in which a decision was previously
made, and replays the decision making process. It then experiments
with modifying the problem solving state in order to determine which
aspects of it were significant to the decision. The technique relies
heavily upon Soar's chunking mechanism to recognize and recall the
factors leading to the decision.
- A multimedia presentation capability. Debrief determines what
information to present to the user, based upon a user model, and
allocates this information among various available media, including
natural language text and graphical depictions.
Relevant Publications
- Johnson, W. L. 1994.
Agents that explain their own actions.
Proceedings of the Fourth Conference on Computer Generated Forces and Behavioral Representation. Orlando, FL.
Two of the diagrams in the above paper do not print due to a bug in
Framemaker. They are available in a separate file by clicking
here.
The slides of the presentation at this conference can be viewed by
clicking here.
- Johnson, W. L. 1994.
Agents that learn to explain themselves.
Proceedings of the Twelfth National Conference on Artificial
Intelligence. Menlo Park, CA: AAAI.
Permission to release pending from AAAI.
- Johnson, W.L. and Tambe, M. Using Machine Learning to
Extend Autonomous Agent Capabilities. Proceedings of the
Summer Computer Simulation Conference, 1995.
Click here to see slides from a
seminar given at ISI in April 1994 on the topic
(provided that your viewer can handle the
fonts. If you have trouble, you should be able to print a hard copy
from your viewer).
Contact
Lewis Johnson
(johnson@isi.edu)
