Copy the page URI to the clipboard
Priest, Anthony G.
(1987).
DOI: https://doi.org/10.21954/ou.ro.0000f80b
Abstract
The motivation for this work has been the development of knowledge about the behaviour of human problem-solvers that would enable an intelligent machine tutor to be designed. In the domain of Newtonian Mechanics, this breaks down into two necessary sub-tasks; how do people decide what equation to generate; and what do they produce when they do try to generate an equation? Although these are psychologically separate questions, an automatic tutor for the domain would need to make use of both kinds of knowledge.
Therefore, strategies for controlling search in physics problem-solving are investigated, and a computational model of erroneous solutions is described. Experimental data is used to evaluate the model. Errors in the domain are classified, and the behaviour of problem-solvers predicted under certain circumstances.
Prediction of Novice errors is a crucial ability for an intelligent tutorial system, and the error analysis implemented in the NEWT program is the main contribution of this thesis.
The investigation has two principal aims:
(1) To develop a model that allows a student's future behaviour to be predicted from an analysis of his past actions. It is argued that this is a necessary prerequisite for the construction of an intelligent tutorial system.
(2) To identify the psychological mechanisms used by problem-solvers working in the domain.
The thesis attempts to achieve these aims in two main ways:
(1) A computer program called NEWT has been constructed, which solves problems of Newtonian Mechanics correctly, or in one of a number of erroneous ways. This allows human errors to be matched, classified, and in some cases predicted.
(2) An analysis of published data leads to the formulation of a control strategy termed "planstacking". This is compared to alternative control strategies, and shown to explain existing data more adequately.
The program is evaluated both as a psychological theory, and as a proposed student model for use in a computer-based tutorial system. The NEWT program was developed from the MECHO program written by Bundy, Byrd, Luger, Mellish and Palmer (1979), at the Department of Artificial Intelligence, Edinburgh University. This program was adapted to produce erroneous problem solutions by the inclusion of procedures to implement malrules observed in the domain.