Content reference: Reasoning about rules

In a companion paper [1] we described the concept of saturation, the situation in a program in which so many knowledge sources (KSs) are potentially useful at each invocation cycle that it is unrealistic to consider unguided, exhaustive invocation. We argued that saturation appears almost inevitable in large AI programs. We suggested that the process of invocation can be viewed as occurring in three steps: retrieval (selecting some subset of KSs from the knowledge base), refinement (pruning or reordering that subset), and execution (executing one or more of the KSs). We then argued that one useful approach to dealing with saturation is by embedding intelligence in the refinement phase, and described meta-rules, a means of encoding knowledge used to effect refinement. In this paper we consider a more detailed, ‘engineering’ issue, but one with a number of interesting implications: While there are many ways to implement refinement, we suggest that one particular technique—which we call content reference—offers a number of advantages, including giving the system some ability to reason about the content of its knowledge.

Authors

• Randall Davis

Keywords

No keywords are indexed for this paper.