Informed selection and use of training examples for knowledge refinement.

Knowledge refinement tools seek to correct faulty rule-based systems by identifying and repairing faults indicated by training examples that provide evidence of faults. This thesis proposes mechanisms that improve the effectiveness and efficiency of refinement tools by the best use and selection of training examples. The refinement task is sufficiently complex that the space of possible refinements demands a heuristic search. Refinement tools typically use hill-climbing search to identify suitable repairs but run the risk of getting caught in local optima. A novel contribution of this thesis is solving the local optima problem by converting the hill-climbing search into a best-first search that can backtrack to previous refinement states. The thesis explores how different backtracking heuristics and training example ordering heuristics affect refinement effectiveness and efficiency. Refinement tools rely on a representative set of training examples to identify faults and influence repair choices. In real environments it is often difficult to obtain a large set of training examples, since each problem-solving task must be labelled with the expert's solution. Another novel aspect introduced in this thesis is informed selection of examples for knowledge refinement, where suitable examples are selected from a set of unlabelled examples, so that only the subset requires to be labelled. Conversely, if a large set of labelled examples is available, it still makes sense to have mechanisms that can select a representative set of examples beneficial for the refinement task, thereby avoiding unnecessary example processing costs. Finally, an experimental evaluation of example utilisation and selection strategies on two artificial domains and one real application are presented. Informed backtracking is able to effectively deal with local optima by moving search to more promising areas, while informed ordering of training examples reduces search effort by ensuring that more pressing faults are dealt with early on in the search. Additionally, example selection methods achieve similar refinement accuracy with significantly fewer examples.

Name dc:type.qualificationname

PhD

Level dc:type.qualificationlevel

Doctoral

Grantor dc:publisher.institution

Robert Gordon University

Year dc:date.issued

2000

Author dc:creator

• Wiratunga, Nirmalie

Advisor dc:contributor.advisor

• Susan Craw

• Knowledge refinement
• Knowledge based systems
• Refinement tools
• Search technologies

Language dc:language

en

Identifier

oai:rgu-repository.worktribe.com:248888

Author Identifier

0000-0003-4040-2496

OAI identifier oai:identifier

oai:rgu-repository.worktribe.com:248888