UC Berkeley

Humans don't grow a mind by sitting in an armchair and reading flashcards containing world facts. Much of human knowledge comes from experimentation. For instance, do antidepressants affect both mood and thoughts directly, or do they affect thoughts via mood? To test competing hypotheses, one can intervene on a variable (e.g., changing mood using another method) to see how it affects other variables. Good interventions generate information that discriminates between hypotheses. However, informative interventions are hard to design. In many past studies, explaining why something occurs is used as a simple but powerful tool to help learners acquire generalizable abstractions useful for future scenarios. In this dissertation, I investigate whether asking causal learners to explain why they plan to carry out certain interventions helps them select more informative interventions.

Chapter 1 describes the task used throughout the dissertation: Three light bulbs are connected one way or another; learners intervene on one light bulb to find out their true structure. The optimal intervention maximizes the expected information gain (EIG) by generating distinct outcomes under different structures. The suboptimal positive test strategy (PTS) tests one hypothesis at a time and favors the intervention that can potentially affect the highest proportion of hypothesized connections. A Bayesian model captures how much a learner relies on EIG vs. PTS to choose interventions.

In Chapter 2 (Study 1), I examine intervention strategies that adults and 5- to 7-year-olds naturally use to select interventions in the Light Bulb Game described above. Adults mainly relied on the optimal strategy, EIG maximization, whereas children mostly used PTS. Following informative interventions, adults identified the correct structures most of the time, but children were at chance.

In Chapter 3 (Study 2), I prompt adults and 5- to 7-year-olds to explain their intervention choices ("Why do you wanna turn on X light bulb?'') and examine if it changes their intervention strategies. Explainers did not intervene differently. However, children who either explained or reported their choices performed above chance at identifying true causal structures from intervention outcomes.

In Chapter 4 (Study 3), I train 7- to 11-year-olds on the difference-making principle which underlies EIG maximization: That a light bulb is helpful if it makes different things happen in different structures, and unhelpful if it leads to the same outcome either way. Training led children to rely more on EIG. The effect was more pronounced in 9- to 11-year-olds than it was in 7- to 8-year-olds.

In Chapter 5, I synthesize findings in Studies 1--3 and propose directions for future research.