Learning and Reasoning with Qualitative Models of Physical Behavior

Building models of the physical world from examples is an important challenge for qualitative reasoning systems. We describe a system that can learn intuitive models of physical behaviors from a corpus of multimodal, multi-state stimuli, consisting of sketches and text. The system extracts and temporally encodes exemplars from the stimuli and uses analogical generalization to abstract prototypical behaviors. Using statistical analysis, the system parameterizes these abstractions into qualitative representations for reasoning. We show that the explanations the system provides for new situations are consistent with those given by naïve students.