This dissertation describes the development and evaluation of a procedural bug diagnostic tool and feedback device. It also describes an experiment on feedback. The domain of the tool and the study is children's subtraction.

The procedural bug diagnostic tool, MEADOW (Mathematics, Errors, and Automatic Debugging Of Written input), incorporates the content and sequence of handwritten marks made during the execution of a mathematics procedure into the knowledge base of an expert system. The rule-based system then utilizes these process data to analyze the procedure and detect procedural errors. The process data allow analysis of the intermediate steps of a procedure; insight into the problem-solving technique is therefore more finely grained than mere examination of the final result.

A computer input/output device consisting of a flat display and electronic pen are used to collect handwritten process data. MEADOW uses a neural network approach for handwriting recognition and a rule-based for analysis of subtraction procedures. MEADOW then provides interactive feedback incorporating voice narration, visual clues, and animation.

An experiment involving 76 third-grade students in four classrooms is also described. Two varying styles of feedback comprise the treatment groups; namely, student-based feedback, which provides feedback directed at exactly where the individual student went wrong, and domain-based feedback, which reteaches the correct procedure without regard for the individual's specific mistakes. In addition, one control group received Right/Wrong feedback and another control group received no feedback. Students competed a total of forty problems at three levels of complexity and were measured for accuracy both during the learning phase and on a post test.

The system evaluation included handwriting recognition and bug detection rates. The handwriting recognition rate was approximately.88. Full, partial, and unknown bug detection rates were.61,.25, and.13 respectively.

Psychological experiment results indicated that student-based feedback was more effective than domain-based feedback. During the course of the experiment, both experimental groups performed significantly better on problem-correctness than the control groups but the experimental groups did not differ significantly from each other. However, both experimental groups showed a linear trend on correctness within each complexity level; the student-based group's trend was significantly greater than the domain-based group at the first level of complexity. Student-based subjects retried problems significantly more than domain-based subjects. A delayed post test found that only student-based subjects performed significantly better than the control groups.