karp@SUMEX-AIM.STANFORD.EDU (Peter Karp) (05/25/89)
Let me complement Dan's answer to Chris's question about qualitative versus quantitative reasoning with a short passage from my thesis. I hope it isn't too much out of context. (You get a free Latex lesson in addition.) \section{Terminology} So many terms have been used of late to describe the research problems with which this chapter is concerned that it is worthwhile to attempt to establish some meaningful terminology before we are suffocated by inappropriate names for this field. I propose the term {\em declarative device modeling.} This phrase captures the viewpoint that this subfield of AI takes of a variety of systems. Be they electrical, mechanical, biological, economic, chemical, sociological, or ecological, we wish to view them as {\em devices} --- to describe their structure and function so that we may predict their future behavior, explain their past behavior, diagnose their current behavior, and design new devices that replace or incorporate them. Having a device {\em model} is key because the description of the structure and function of the device is what is common to all these tasks. The model must be in a {\em declarative} form so that the entities that perform the preceding tasks are able to reason about the model. Among the existing terms for this subfield are {\em qualitative reasoning,} {\em qualitative physics,} {\em qualitative simulation,} {\em causal reasoning,} {\em causal modeling,} and {\em commonsense reasoning.} Each has specific drawbacks. {\em Simulation} is troublesome because technically it means prediction, which is a subset of the tasks in which we are interested (some researchers use this word inaccurately to describe all the preceding tasks). The word {\em causal} is enticing, but its use is often presumptuous because we have yet to see declarative device models that include a rich, declarative, comprehensive, meaningful description of causal as opposed to other types of relationships, and that are based on a precise theory of causality. The term {\em qualitative physics} is overly specific because even though some researchers are immediately concerned with the domain of physics, their techniques are usually applicable to other domains, and modeling techniques developed from other fields are often relevant to qualitative physics. The term {\em qualitative} is not as general as some authors' use of the word implies: We can build declarative device models that are not qualitative, and qualitative models that are not declarative. The word {\em qualitative} usually indicates that a model incorporates a technique for abstracting mathematical relationships or state variables, such as the use of confluences, limit analysis, or aggregation. A declarative device model may or may not include such abstractions. Similarly, a model that does include such abstractions may or may not be in a declarative form. For example, Davis and Genesereth \cite{Davis84,Genesereth84} have constructed declarative device models that are not qualitative; they do not use special representations that abstract either the state variables or the component interactions in the devices they model.