From Diagrams to Models by Analogical Transfer

We present a method for constructing a teleological model of a drawing of a physical device through analogical transfer of the teleological model of the same device in an almost identical drawing. A source case, in this method, contains both a 2-D vector-graphics line drawing of a physical device and a teleological model of the device called a Drawing-Shape-Structure-Behavior-Function (DSSBF) model that relates shapes and spatial relations in the drawing to specifications of the structure, behavior and function of the device. Given an almost identical target 2-D vector-graphics line drawing as input, we describe how an agent may align the two drawings, and transfer the relevant structural, behavioral and functional elements over to the target drawing. We also describe how the DSSBF model of the source drawing guides the alignment of the two drawings. The Archytas system implements this method in domain of kinematic devices that convert translational motion into rotational motion, such as a piston and crankshaft device. 1 Background, Motivation and Goals Diagram understanding is a persistent and central issue in research on diagrams. Novak [20] explicitly viewed the task of understanding a diagram as one of constructing a model for it: his Beatrix program understood a textually-annotated schematic diagram kinematics problem by constructing a structural model. In cognitive science, Narayanan, Suwa and Motoda [19] describe a psychological study in which human subjects built behavioral models of a physical device from its labeled schematic diagram. The computational advantage of model construction is that the model enables inferences for higher-level tasks, e.g., analysis of the functions of the physical device in case of Narayanan, Suwa, and Motoda’s human subjects. The type of the output model in the task of diagram understanding depends in part on the input diagram and partly on the inferences desired of the model. Current AI methods for constructing models from diagrams, e.g., in the Beatrix and the GeoRep [9] programs, often apply domain-specific rules to extract the model from the diagrammatic representation. For example, GeoRep first uses a domain-independent diagrammatic reasoner that extracts shapes and spatial relations from a given diagram, and then applies domain-specific rules for extracting a behavioral model from the intermediate representation of shapes and spatial relations. Given the state of the art in In D. Barker-Plummer, R. Cox, & N. Swoboda (Eds.): Diagrams 2006, LNAI 4045, pp. 55–69, 2006. Springer. 2 Patrick W. Yaner and Ashok K. Goel