completion of a complex agent training experiment at the US
In the Spring 2002 session of the “Military Applications of Artificial
Intelligence” course, we have conducted a complex agent training experiment.
Seven teams comprising 15 senior military officers used personal copies of the
Disciple-RKF/COG learning agent. Each team has taught its agent not only to
identify center of gravity candidates (as was the case in the previous, Spring
2001, experiment), but also to test these candidates and eliminate those which
are not the center of gravity. Testing of candidates is a much more complex
activity than identifying them, therefore this was a very challenging experiment
aimed at clarifying the main strengths and weaknesses of the Disciple approach.
Each team trained its agent using a single scenario, by explaining to the agent how
to identify and test the strategic COG candidates for the opposing forces in
that scenario. At the end of the training phase, the developed agents were
tested on new scenarios. For instance, the following figure shows how each of the
developed agents analyzed the Grenada
A yellow dot on a row corresponding to a COG candidate indicates that
the candidate was identified by the agent from the column of that dot. A red dot
indicates that after further analysis the candidate was eliminated. A green dot
indicates that the candidate passed all the testes and was not eliminated,
indicating that it could be the actual COG. In one example, the agent trained with the
scenario has identified “will of the people of Grenada”
as a strategic center of gravity candidate for the Grenadan coalition and then,
after further testing, it has eliminated it.
Application of partially trained agents
on the Grenada scenario.
Notice that although each agent was
trained with only one scenario (indicated at the top of each column), their
performance is reasonably good. This demonstrates the power of the
Disciple learning methods, which allow the agent to learn useful reasoning rules
from only one or two examples.
At the end of the experiment 9 out of
the 15 experts agreed, 2 strongly agreed and 4 were neutral with respect to the
following statement: "I think that a subject matter expert can use Disciple
to build an agent, with limited assistance from a knowledge engineer." This
is a very important result which confirms the one obtained in the previous year,
in spite of the fact that the teaching required this time was significantly more
following are some general assessments made by the students (military experts):