Proceedings of the Annual Meeting of the Cognitive Science Society

Planning requires significant cognitive effort, as it involves constructing and evaluating numerous potential futures. A key method to simplify planning is the grouping of sequences of actions into cohesive units-often referred to as action chunking. While studies confirm the existence of action chunking in human planning, we explore chunking as a learning process across multiple tasks with different statistical properties. To that end, we ran a large-scale preregistered study using the Lightbot navigation paradigm, in which we manipulated the underlying statistical properties of the planning tasks. We found that the majority of participants adapted their chunking behaviour to match the environment's statistical properties by either learning to use the action chunk that is optimal given these statistical properties or a subsequence thereof. Together, our results suggest that statistical properties of the environment influence action chunking in human planning.