pasley@SRI-KL (Christine Pasley) (06/04/86)
CS529 - AI In Design & Manufacturing Instructor: Dr. J. M. Tenenbaum Speaker: Sanjay Mittal From: Xerox Palo Alto Research Center Title: Pride: A Knowledge-Based Framework for Design Guest Speaker: Kenneth Forbus From: Qualitative Reasoning Group University of Illinois Title: Qualitative Process Theory: Selected Topics Date: Wednesday, June 4, 1986 Time: 4:00 - 5:30 Place: Terman 556 Sanjay Mittal's abstract: This talk will describe the Pride project at Xerox. The first part of the talk will be about an expert system for the design of paper transports inside copiers. A prototype version of the system has been in field test for a year. It has been successfully used on real copier projects inside Xerox - both for designing and for checking designs produced by engineers. From an applications point of view we have been motivated by the following observations: knowledge is often distributed among different experts; the process of generating designs is unnecessarily separated from their analysis, leading to long design cycles; and design is an evolutionary process, i.e., a process of exploration. The second part of the talk will describe the framework in Pride for representing design knowledge and using it to support the design process. In this framework, the process of designing an artifact is viewed as knowledge guided search in a multi-dimensional space of possible designs. The dimensions of such a space are the design parameters of the artifact. In this view, knowledge is used not only to search the space but also to define the space. Domain knowledge is organized in terms of design plans, which are organized around goals. Conceptually, goals decompose a problem into sub-problems and are the units for structuring knowledge. Design goals have design methods associated with them, which specify alternate ways to make decisions about the design parameters of the goal. The third major element of a plan are constraints on the design parameters. The framework provides a problem solver for executing these plans. The problem solver extends dependency-directed backtracking with an advice mechanism and a context mechanism for simultaneously maintaining multiple designs. Kenneth Forbus' abstract: Much of our commonsense knowledge of the physical world appears to be organized around a notion of physical processes. Qualitative Process theory provides a formal language for describing such processes, including a qualitative representation of differential equations and the conditions under which they apply. This talk will briefly review Qualitative Process theory and discuss two topics of current research: Interpreting measurements taken across time, and a new implementation, based on an assumption-based truth maintenance system, that provides roughly two orders of magnitude performance improvement.