Think outside the box: Making up causal hypotheses from unreliable evidence

Human adults and children think of the natural world as orchestrated by rules, yet many of them are neither equally rigid nor clearly evident. Some are beset by exceptions, and others are not intuitive. What cognitive mechanisms underpin human learning of how the world works? We propose a computational model for formulating and testing hypotheses in naturalistic contexts, that combines Bayesian inference under uncertainty over self-generated and social evidence with causal hypotheses based on formal rules and perceptual heuristics. We validate our model experimentally, showing that it explains how 7- to 10-year-olds' solve a physical puzzle, including the distribution and the types of evidence children sampled. The proposed model outperforms both a purely rule-based Bayesian hypothesis search and a resource-rational random sampling approach. Our results suggest that children implement an internal mechanism for generating and testing a limited number of hypotheses, including formal programmatic rules and heuristics generated from salient problem features to seek more evidence when formal rule generation fails.