Gastvortrag: Robert Goldstone

22. Dezember 2015

Computational Models of Mental Models of Computational Models of the World
Robert Goldstone
(Reporting on a collaborative project with Francisco Lara-Dammer and Douglas Hofstadter)

In classroom and laboratory observations of students interacting with computer simulations to learn complex systems principles, we have observed systematic misinterpretations of these simulations.  Students (and scientists) often discover erroneous patterns in the simulations, and construct underlying rules for the interactions among simulation parts that diverge substantially from the actual rules underlying the simulations.   At the same time, students can also sometimes learn a considerable amount about the underlying causal mechanisms of a simulation by interacting with it.  To understand both the successes and failures of students’ interpretative efforts, we have developed a computational model of the process by which human learners discover patterns in natural phenomena.  Our approach to modeling how people learn about a system by interacting with it follows three core design principles: 1) perceptual grounding, 2) experimental intervention, and 3) cognitively plausible heuristics for determining relations between simulation elements.  In contrast to the majority of existing models of scientific discovery in which inputs are presented as symbolic, often numerically quantified, structured representations, our model takes as input perceptually grounded, spatio-temporal movies of simulated natural phenomena.  Given this relatively raw visual representation, instilling plausible (per/con)ceptual constraints is key to building apt and efficient relation detectors.  We will consider the recognition of relations such as: collide, attract, repel, change state, transfer state, excite, and inhibit.  An application of the model to the discovery of ideal gas laws will be described.