The role of spatial structure in problem solving: analysis at an information type level of abstraction

Abstract

A big challenge when problem solving is choosing the right information for a task. Humans do this subconsciously and thus beyond the grasp of introspection. Modern AI systems are also often black boxes. I hypothesize that it should be possible to study problem solving at a level of abstraction that allows determining which information types a solver uses. In this thesis I built the theoretical framework and define the requirements and challenges for such an approach. I designed three experimental paradigms in the domain of spatial cognition addressing these challenges. (1) An analogy task to apply spatial information to a non-spatial domain. (2) A tic-tac-toe isomorph to determine whether problem solvers seek task-irrelevant spatial information to aid them and if this could also be detrimental. (3) A card-sorting task to test relative salience of information types. I ran pilot studies of them and built a computational model for the latter. The tasks were successful in showing that spatial information can be applied to a non-spatial domain, is sought if it aids a task but ignored otherwise, and no difference in salience of spatial vs non-spatial information was detected. This thesis discusses the merits of studying information type use and provides tools for doing so.