Donald Bren School of Information and Computer Sciences

This dissertation addresses the problem of automatically revising production system rulebases using input-output mappings as the main source of information to guide the revision process. The main contribution of this work, the ability to revise production system rulebases, is important because production system rulebases are used so widely in industry. An implemented approach, CR2, is shown analytically and empirically to be able to revise these rulebases.

One important facet of this approach is an explicitly defined model, the revision problem space model, that was used to design CR2, and is used to understand and empirically analyze it. This model is important because, relatively speaking, the production system revision task is very difficult. The model allows for an understanding of when the revision system should succeed and fail.

Another facet of the approach is a set of techniques, the RIO techniques that is used to identify revisions to the rulebase. These techniques were designed in the context of the revision problem space model. Unlike most Horn clause based revision systems, multiple techniques are needed to identify revisions independent of the problem with the rulebase and the information available to identify the problem.

In order to produce a revised rulebase that a domain expert would find comprehensible, a technique called rule structure filtering is used to avoid revisions that would produce "ill-structured" rules. This technique is shown analytically to produce more comprehensible revisions and is shown empirically to produce more accurate rulebases.