smadar@JARRE.RUTGERS.EDU (Cabelli) (08/03/87)
Ken Laws writes: >Semantic classification thus requires at least three viewpoints: >structure, intended function, and perceived or implemented function. There has been alot of research recently in machine learning on formulating concepts with these viewpoints in mind. I am amazed at the omittion of any relevant AI work in this discussion on natural kinds! For example, no mention was made of Winston's work on learning structural descriptions from functional definitions (AAAI-83), (I was surprised Minsky omitted that work). My work on "Formulating Concepts According to Purpose" (AAAI-87) presents a prototype system which formulates definitions of a "cup" based on the purpose for which an agent intends to use it (one specialized notion of intention). If the agent intends to use a cup to drink hot liquids from, one definition is automatically generated. If on the other hand, the cup has an ornamental purpose, a different definition can be formed. The key idea of the technique is to simulate the plan of actions the agent will go through in drinking hot liquids from a cup, (say POUR, GRASP, PICKUP, DRINK). Then, computing the (weakest) preconditions of this plan derives a functional description (must contain hot liquids, must be graspable by agent with hot liquid, must be liftable, and so on). A technique like Winston's is then used to compute the structural description from the functional one. Smadar Kedar-Cabelli Rutgers University