Regarding the I in ITS: Student Modeling.
Valerie J. Shute · 1994
For an intelligent tutoring system (ITS) to earn its I, it must be able to (1) accurately diagnose students' knowledge structures, skills, and/or learning styles using principles, rather than pre-programmed responses, to decide what to do next; and (2) adapt instruction accordingly. While some maintain that remediation actually comprises the T in ITS, this paper takes the position that the two components (diagnosis and remediation), working in concert, make up the intelligence in an ITS. A framework for developing and assessing student models is presented, followed by a description of an attempt to apply the framework in the development of a student model incorporated within a non-intelligent computer tutor. The two systems (with and without a student model) are compared in terms of outcome and efficiency measures. The framework is an adaptation of Dillenbourg and Self's (1992) two-dimensional framework and notation for student modeling, which was modified to represent specific knowledge and skill types required during the learning process, procedural skills, conceptual knowledge rather than overt behaviors, and cognitive process measures. The horizontal axis remains basically the same as the original: learner's representation of the knowledge or the skill, system's representation of the learner's knowledge, and the system's/expert's representation of the knowledge or skill. This modified framework represents the standard microadaptive approach to student modeling. The intelligent and non-intelligent version of Stat Lady, an experiential learning environment and curriculum that teaches statistical concepts and skills, are described. The Stat Lady versions provide the basis for a planned experiment testing the degree to which inclusion of a student model may enhance learning outcome measures and/or improve learning efficiency. (Contains 16 references.) (Author/MAS) *********************************************************************** Reproductions supplied by EDRS are the best that can be made from the original document. *********************************************************************** Regarding the I in ITS: Student Modeling VALERIE J. SHUTE Armstrong Laboratory Brooks Air Force Base, Texas 78235-5352 USA [email protected] U.S. DEPARTMENT OF EDUCATION ()Mc. ot Educattonat Research and improvement EDUCATIONAL RESOURCES INFORMATION CENTER (ERIC) C This document has Peen reproduced as recetved from the person or organtzatton origmattng .1 C !Amor changes have been made to improve reproductron guallty Points of mew or oprhons Mateo tn INS docu. ment do not necessanly represent ofhcsal OERI posmon or poltcy Abstract: For an intelligent tutoring system (ITS) to earn its I, it must be able to (a) accurately diagnose students' knowledge structures, skills, and/or learning styles using principles, rather than pre-programmed responses, to decide what to do next, and then (b) adapt instruction accordingly. While some maintain that remediation actually comprises the T in intelligent tutoring systems, my position is that the two components (diagnosis and remediation), working in concert, make up the intelligence in an ITS. A framework for developing and assessing student models is presented, followed by a description of an attempt to apply the framework in the development of a student model incorporated within a non-intelligent computer tutor. The two systems (with and without a student model) will be compared in terms of learning outcome and efficiency measures. For an intelligent tutoring system (ITS) to earn its I, it must be able to (a) accurately diagnose students' knowledge structures, skills, and/or learning styles using principles, rather than pre-programmed responses, to decide what to do next, and then (b) adapt instruction accordingly. While some maintain that remediation actually comprises the T in intelligent tutoring systems, my position is that the two components (diagnosis and remediation), working in concert, make up the intelligence in an ITS. A framework for developing and assessing student models is presented, followed by a description of an attempt to apply the framework in the development of a student model incorporated within a non-intelligent computer tutor. The two systems (with and without a student model) will be compared in terms of learning outcome and efficiency measures. Jeremy (age 10) arrives at his math lab where he sits in front of a computer that is going to help him learn to solve algebra word problems. Today's focus is on those troublesome distance-rate-time problems. After stating his name, the computer accesses Jeremy's records, flagging his relevant strengths and weaknesses (i.e., not only his higher-level aptitudes from his computerized school records, but also the low-level rules that he's acquired and not yet acquired in this module). Beginning with an animated review of concepts and skills that he learned the day before, the computer generates a problem which is just a little bit beyond his grasp. The system then works out the correct solution to the problem, along with some alternative solutions that Jeremy is likely to come up with based on its student model of him. In fact, he incorrectly solves the problem like the tutor predicted. As part of its student model of him, the computer knows to instruct Jeremy with an emphasis on a graphical representation of the problem to clarify the discrepancy between the correct and incorrect solutions and facilitate the formation of a firnctional mental model (conceptual knowledge). Thus, the tutor presents two animated trains appearing on opposite sides of the screen, and converging at a point almost in the middle of the screen. They travel at different rates of speed. The problem statement stays up at the top of the screen, and the tutor points out, as it periodically pauses the simulation, what elements should be attended to and when. After Jeremy states that he understands the mapping between the explicated conceptual knowledge, the appropriate equation, and the relevant parts of the word problem, the computer presents an isomorphic word problem. This time he solves it correctly, without any supplemental graphics. The computer allows him to play around with some trains, missiles and boats on his own for a while to test his emerging understanding. He views his score of curricular elements acquired, and instruction and learning continue. The above scenario could describe events in a class-lab 15 to 50 Years from now, or just remain a figment of our imaginations. To achieve this future, w; need to conduct more systematic research on student modeling, focusing on increased system flexibility and diagnostic accuracy. We also need to adopt some framework and formalism for a more precise specification of models. All this presumes that the student model is the correct focus for developing more intelligent tutoring systems, so we additionally need controlled evaluations testing which modeling techniques are better for which kinds of domains, and even whether or not a student model, in general, is worth the research and development costs. The ability to diagnose student errors and tailor rernediation based on the diagnosis represents the critical difference between intelligent and merely clever computer-assisted instruction. The working definition of computer-tutor intelligence that I'll be using in this paper is that a system must behave intelligently, not actually be intelligent, like a human. Specifically, an intelligent system must be able to accurately diagnose students' -PERMISSION TO REPRODUCE THIS