Proceedings of the Annual Meeting of the Cognitive Science Society

Humans possess knowledge of causal systems with deep compositional structures. For example, we know that a good soccer team needs players to fill different roles, with each role demanding a configuration of skills from the player. These causal systems operate on multiple object types (player roles) that are defined by features within objects (skills). This study explores how human learners perform on novel causal learning problems in which they need to infer multiple object types in a bottom-up manner, using empirical information as a cue for their existence. We model subjects' learning process with Bayesian models, drawing hypotheses from different spaces of logical expressions. We found that although subjects exhibited partial success on tasks that required learning one object type, they mostly failed at those that required learning multiple types. Our result identifies the learning of object types as a major obstacle for human acquisition of complex causal systems.