Lucid Representations

This paper criticizes the widespread idea that knowledge bases in AI systems should be complete and that representations should be \model-like." The arguments in favor of such representations are less cogent and more ambiguous than they appear at rst. Levesque's suggestion that representations should be \vivid" is extremely restrictive, particularly in its uniform imposition of a closed-world assumption. Spatial representations that are adequate for reasoning about a wide range of physical phenomena must ultimately either use complex symbolic reasoning or deal with partial and imperfect approximations. Requiring that temporal representations be fully detailed simulations will often be extremely ineecient. Finally, a exible intelligent system must necessarily deal with partial information of all kinds, and the techniques for carrying out reasoning about partial information using complete representations are very limited in their application. There is a recurrent idea in the AI research community that representations should resemble models, rather than being an arbitrary collection of formulas or constraints. Like a model, a representation should correspond to one particular way the world could be, certain and complete. Like a model, the representation should characterize the world directly and perspicuously, not give a collection of interacting conditions on it. Answering questions from a model is a matter of checking and measuring rather than of reasoning; so it should be from a representation. This argument has appeared in a variety of forms and contexts. Levesque 1986, 1989] proposes that the core of a knowledge base should be a \vivid" representation, a collection of atomic formulas satisfying the closed world assumption. Halpern and Vardi 1991] propose that reasoning about knowledge should be performed by explicitly constructing Kripke systems of possible worlds. Forbus, Nielsen, and Faltings 1987] argue that spatial representations for spatial or physical reasoning programs should contain a \metric diagram", a representation that gives full geometric details. The vast majority of physical reasoning programs that do spatial reasoning do use such a representation (e. well as others, has suggested that planning programs should always use construct instantiated plans, to avoid the complexities involved in reasoning about partial plans. In this paper, I will generically call all such representations \lucid." (\Vivid" would be a better word, but Levesque has appropriated it for his own theory.) Representations that are not lucid are \indirect". The aim of this paper is to argue that, despite their attractiveness, lucid representations are often inappropriate or inadequate for AI systems.