ICML-98 Submission #185
Teaching an Agent to Test Students
Gheorghe Tecuci and Harry Keeling
Department of Computer Science
MSN 4A5, George Mason University
4400 University Dr., Fairfax, VA 22030-4444
Abstract
This paper presents an innovative application of the Disciple Learning
Agent Shell to the building of an educational agent that generates
history tests for middle school students, to assist in the assessment
of their understanding and use of higher-order thinking
skills. Disciple is an apprenticeship, multistrategy learning agent
that can be taught by an expert how to perform domain-specific tasks
in a way that resembles the way an apprentice would be taught by the
expert. Disciple has been taught by an educator to generate and answer
basic test questions and to explain the answers. From its interaction
with the educational expert, Disciple has learned general rules that
allow it to generate a large number of new test questions for
students, together with hints, answers, and explanations of the
answers. As a result, it can guide the students during their practice
of higher-order thinking skills as they would be directly guided by
the educator. The Disciple agent is also a useful tool for the
educator, being able to generate a different exam for each student in
the class. It has been experimentally evaluated by history experts,
students and teachers, with very promising results. The work on
developing this educational agent illustrates also an integration of
machine learning, knowledge acquisition, problem solving and
intelligent-tutoring systems.
Keywords: application of machine learning
apprenticeship and multistrategy learning
integration of learning,
problem solving and intelligent tutoring
building of educational agents
Email address of contact author: tecuci@gmu.edu
Phone number of contact author: (703) 993-1722