Acquiring (ir)relevance Knowledge for Problem Solving

A major drawback of artiicial intelligence systems that rely on declarative representations is that the eeciency of reasoning degrades quickly as the size of the knowledge base increases. To address this problem when building a system, we need to acquire not only knowledge about the domain, but also knowledge about control of reasoning. In this paper, we discuss one type of such control knowledge, namely, relevance of our knowledge to speciic problem solving goals. We show how this knowledge can be used by the problem solver either to ignore part of its knowledge or to automatically create abstractions and how the system can guide the acquisition of such knowledge. We ground our discussion in a framework in which knowledge about relevance can be stated, reasoned with and analyzed. We apply the framework to the problem of modeling physical devices, where creating abstractions for a given task is crucial in order to perform eeective problem solving.