Theses & Reports
Extraction and Generalization of Expert Advice (Learning, Representation, Induction)
Benjamin, David Paul
Title: Extraction and Generalization of Expert Advice (Learning, Representation, Induction)
Candidate: Benjamin, David Paul
This work describes a method for representing knowledge in production systems which makes use of the conflict set. This permits a rich description of task situations, and allows the use of control productions to effect conflict resolution. A set of extensions to the OPS5 production system is described which facilitates the implementation of this approach within OPS5. This extended system is then used to implement a multi-level, goal-directed production system for the construction of expert systems, CAMERA, in which control information is automatically built from the actions of an expert trainer. This control information consists of sequencing and goal information which is interactively extracted from the trainer by CAMERA, and generalized by DISC, which models generalization as the process of finding 'discriminating' features, which are those features of a situation that cause a particular method to be chosen, and then constructing a description of those features. When solving a task, CAMERA examines only the discriminating features specified in the generalized control rules. Thus, instead of matching all the productions against the working memory, CAMERA considers only the relevant rules. Experiments with the system are described.