Marcella.Zaragoza@ISL1.RI.CMU.EDU.UUCP (01/30/87)
AI SEMINAR TOPIC: Understanding How Devices Work: Functional Representation of Devices and Compilation of Diagnostic Knowledge SPEAKER: B. Chandrasekaran Department of Computer & Information Science The Ohio State University Columbus, OH 43210 PLACE: Wean Hall 4605 DATE: Wednesday, February 4, 1987 TIME: 10:00 a.m. ABSTRACT: Where does diagnostic knowledge -- knowledge about malfunctions and their relation to observations -- come from? One source of it is an agent's understanding of how devices work, what has been called a ``deep model.'' We distinguish between deep models in the sense of scientific first principles and deep cognitive models where the problem solver has a qualitative symbolic representation of the system or device that accounts qualtitatively for how the system ``works.'' We provide a typology of different knowledge structures and reasoning processes that play a role in qualitative or functional reasoning. We indicate where the work of Kuipers, de Kleer and Brown, Davis, Forbus, Bylander, Sembugamoorthy and Chandrasekaran fit in this typology and what types of information each of them can produce. We elaborate on functional representations as deep cognitive models for some aspects of causal reasoning in medicine. Causal reasoning about devices or physical systems involves multiple types of knowledge structures and reasoning mechanisms. Two broad types of approaches can be distinguished. In one, causal reasoning is viewed mainly as an ability to reason at different levels of detail: the work of Weiss and Kulikowski, Patil and Pople come to mind. Any hierarchies in this line of work have as organizing principle different levels of detail. In the other strand of work, causal reasoning is viewed as reasoning from @i(structure) of a device to its @i(behavior), from behavior to its @i(function), and from all this to diagnostic conclusions. In this approach, the hierarchical organization of the device or system naturally results in an ability to move into more or less levels of detail. We discuss the primitives of such a functional representation and show how it organizes an agent's understanding of how a systems functions result from the behavior of the device, and how such behavior results from the functions of the components and the structure of the device. We also indicate how device-independent compilers can process this representation and produce diagnostic knowledge organized in a hiererchy that mirrors the functional hierarchy. Sticklen, Chandrasekaran and Smith have work in progress that applies these notions to the medical domain. If you wish to meet with Dr. Chandrasekaran, please contact Marce at x8818, or send mail to mlz@d.