Searching large hypothesis spaces by asking questions

One way people deal with uncertainty is by asking questions. A showcase of this ability is the classic 20 questions game where a player asks questions in search of a secret object. Previous studies using variants of this task have found that people are effective question-askers according to normative Bayesian metrics such as expected information gain. However, so far, the studies amenable to mathematical modeling have used only small sets of possible hypotheses that were provided explicitly to participants, far from the unbounded hypothesis spaces people often grapple with. Here, we study how people evaluate the quality of questions in an unrestricted 20 Questions task. We present a Bayesian model that utilizes a large data set of object-question pairs and expected information gain to select questions. This model provides good predictions regarding people’s preferences and outperforms simpler alternatives.