The Thought Experiment Approach to Qualitative Physics

This paper discusses the application of the thought experiment methodology to qualita­ tive reasoning. Problem solving using this technique involves simplification of the origi­ nal problem, solution of the simplified problem, and generalization of the results obtained. Our emphasis in this work is to demonstrate the ef­ fectiveness of this approach in addressing com­ plexity and grain size issues that affect qual­ itative simulation. The thought experiment methodology is presented formally, the imple­ mentation of a problem solver called TEPS is briefly discussed, and the methodology is com­ pared with related techniques such as approxi­ mation, aggregation, and exaggeration. 1 Introduction The use of qualitative models in explaining the behav­ ior of physical systems is an area of continuing research among the AI community. This paper develops the thought experiment methodology in the qualitative rea­ soning framework. The technique developed can be ap­ plied to a wide variety of problem solving situations. Naive physics or common sense problem solving is characterized by the lack of precise and complete knowl­ edge of all the constraint relations among parameters rel­ evant to a physical situation under study. This results in analysis being performed on weakly constrained systems which causes two major problems: (i) a large multiplic­ ity of possible solutions many of which may be physically incorrect, and (ii) the corresponding computational com­ plexity in deriving the behavior of even simple physical systems. It might be argued that constraint relation­ ships may be more easily obtained for human-engineered systems that have been built for a specific purpose. (As­ suming nothing is broken, at least the system design­ ers should be able to specify the complete and precise constraints). However, this is true only for very simple mechanical devices. As devices become more complex, the loose coupling of the component subsystems makes the complete set of parameter relations harder to define. For example, in an automobile engine it is not difficult to derive the relations that specify the location of each piston relative to the others, however, if one considers other components, such as the fuel pump, water pump, generator, and battery, the proliferation in the number of components and their weak coupling makes constraints harder to generate. Conventional qualitative simulation methods provide no solutions for dealing with the multiplicity of solutions and the computational complexity that arise because of the underconstrained nature of the problems. In a recent paper, Falkenhainer and Forbus …