The qualitative problem solving system PHYSYS

Abstract

Any expert system designer should consider answering questions such as What knowledge must be available to the system so that it reaches the solution of a given problem, How should that knowledge be represented, and What Programming Language is most suited for implementation of such a It has been shown by research on various types of rule-based systems that representations of prototypical situations confronted in a domain are one kind of knowledge crucial to the performance competencies of a system. This paper presents a problem solving system, called PHYSYS, which uses knowledge about prototypical situations to guide its processing and to compute the solution of the problem at hand. PHYSYS represents its knowledge as frames, organized in a semantic net, and performs tasks in the domain of engineering mechanics. The frames are called prototypes because they represent typical situations which can be used as a basis for composing the actual situation. The overall idea of the PHYSYS system is simple. It is an interactive system, one in which the user communicates with the system on a one-to-one basis. Problems are posed to the system in terms of simple physical objects such as blocks, inclined planes, etc., and forces acting on these objects. Input is specified with icons representing the objects and forces. The system performs an analysis analogous to constructing vector diagrams for the forces. The results of this analysis allow the system to (1) derive the equations of motion and (2) reason qualitatively about the physical behavior of the problem elements. In contrast to finding the solution of the equations the system tries to show the user how to obtain the solution and using its knowledge of physics, the system attempts to make qualitative statements about the solution. The system decomposes the pictorial description of the problem supplied by the user, into primitive problem constituents. The system then searches its knowledge base looking for those specific primitive consituents of the problem using pattern matching. Finally, the system composes these primitive constituents back together, to complete the problem. Since the system can not solve differential equations it generates only qualitative predictions.We chose Prolog as the implementation language primarily because of its inherent ability to do backward chaining and its promise of shortened development time for such problems. A prototype version of PHYSYS has been implemented in Turbo Prolog.