The qualitative analysis of physical systems is a central problem in artificial intelligence, but current techniques are successful only for simple models of small systems. Exponential time complexity and the ubiquity of ambiguous conclusions limit the usefulness of qualitative physics reasoning techniques when applied to large models. Effective reasoning about complex systems will require multiple models of the system that embody different simplifying assumptions appropriate for different tasks. It is therefore proposed to develop a theory of both qualitative models and the assumption on which they are based, and to test that theory by constructing a reasoning program. This reasoner will automatically determine which assumptions are appropriate, build and use the corresponding model, and finally check its conclusions. This approach should significantly speed up reasoning, decrease ambiguity, and enable automated analysis of more complex systems.
