Human teachers are expected to improve with experience; they should learn how to handle different types of students, and learn which methods work for which learning difficulties. But how can a computerized teacher improve its skills over time after working with hundreds of students? How can it recognize that certain problems are too difficult for a student or identify which hints should follow other hints? How can it become more skillful over time?
These are the issues raised in a new award from the National Science Foundation’s Program for Research on Learning and Education (ROLE), to Beverly Woolf (PI), Andrew Barto, Sridhar Mahadevan, and Ivon Arroyo from Computer Science and Don Fisher from Mechanical and Industrial Engineering. The goal of this research is to use machine learning to model prior student behavior, to learn what is effective and to develop new and more effective pedagogical strategies. Optimization might be directed at reducing a student’s time to achieve mastery or advance through a curriculum.
This NSF research focuses on designing computer aided tutors to improve their own knowledge about individual students and pedagogy. This improvement is needed for many reasons. Tutors are let loose in a constantly changing environment under conditions that cannot be predicted. It is not feasible to build teaching systems that emulate master teachers given the variety of student interests, abilities and learning styles. Nor is it feasible to build a tutor for every new population of students, for example, those with low cognitive or spatial ability.
Imagine an intelligent tutor that teaches fractions to 8-10 year old children. Such a tutor typically assumes that all incoming students have a fairly standard set of mathematics skills and will acquire new knowledge in a fairly standard way. Yet, this is not always the case. Not all 8 - 10 year olds understand the concepts needed to work with fractions. Younger, more advanced students (such as mathematical "super stars" who are learning algebra earlier than their age peers) may simply not have had the opportunity to learn how to work with fractions. Older students returning to college may need to learn refresh basic skills, and need help with remedial mathematics.
We have used machine learning with an earlier tutor to predict how each student will react to a variety of teaching actions. Joseph Beck (Ph.D. ‘01), now at Carnegie Mellon’s Center for Automated Learning and Discovery, demonstrated the advantages of machine learning to form a useful student model from a corpus of data obtained from an earlier population of students. In the current research, machine learning techniques are being applied to Wayang Outpost, a tutor developed at UMass Amherst as part of an NSF funded project (Carole Beal, Department of Psychology, P.I.). This tutor for Scholastic Aptitude Test geometry problem solving uses graphics and animation to motivate students, and keep them engaged in the learning process.