The application of focusing in Intelligent Tutoring
A number of tutoring systems have been developed which attempt to learn concepts by constructing a method of distinguishing members of a concept from non-members. The implementation of concept learning in intelligent tutoring systems has been considered as the basis for building dynamic student models and also for a scheme to allow collaboration between student and tutor.
Self (1987) proposed a new type of student model which would function as a collaborative partner. The main task of such a partner or internal learner is to offer advice and suggestions about the teaching material and the learning process.
The internal learner would use the focusing algorithm to generate information about the student by examining instances examined by the student. Consequently the tutor could use the information obtained by the internal learner to provide guidance for the student when necessary. Self also indicated that to have a type of internal learner will be necessary to monitor any open-ended tutoring system.
There have been a number of attempts to apply similarity-based concept learning algorithms, of which focusing is an example, to build dynamic student models in tutoring systems. Examples of systems include MULTI (Gilmore 1986), and IMPART (Elsom-Cook 1988) which both employ variants of the focusing algorithm.
MULTI is a an enhanced version of the focusing program which was used in collaborative learning systems. This program combines new information with the current understanding of concepts to form a revised understanding.
IMPART is a tutoring system which uses guided discovery methods to teach programming in LISP. It teaches the semantics and syntax of LISP via experiments with the interpreter.
IMPART employs a new framework for user modelling called bounded user modelling. It is an alternative to the widely used overlay and perturbation student modelling methods. It describes the state of understanding of the learner in terms of upper and lower bounds of the possible states of the learner. The way the upper and the lower bounds are set is related to focusing. However, the goal of focussing is to use events to squeeze the upper and lower bounds of the version space onto a concept description. In IMPART an event may generate a completely new set of upper and lower bounds requiring the system to maintain a number of version spaces. The algorithm to handle this is greatly complicated.
LEX is a learning program which attempts to learn a set of heuristics for indefinite integration. The initial knowledge in LEX is a set of basic facts and rules. The machine learning technique employed by LEX is the ‘version space’ (Mitchell, Keller & Kedar-Cabelli, 1986) which is a similarity-based concept learning algorithm very similar to focusing. LEX had a number of problems mainly due to the way that it represented algebraic expressions.
Authored by Serengul Smith
E-mail to:
serengul1@mdx.ac.uk
School of Computing Science Middlesex University
Revised: September 1998