Humans form rich causal models of the world that support prediction, explanation, planning and control. While Bayesian methods help formalize how such representations can be learned from data, they are only tractable in the simplest cases. Thus, a key question is how bounded human learners succeed in the face of the world’s formidable complexity. I will discuss a project aimed at unravelling this mystery. We investigate how people learn about probabilistic causal systems by performing interventions (actions that perturb a system of interest, like pushing a button, taking a medicine, or implementing a policy). Across a line of studies and extensive model comparison, I show that people adjust their causal representations in a piecemeal fashion, making small local changes rather than more extensive “Kuhnian” revisions. I formalize this with a model inspired by algorithms for approximating Bayesian inference, and use this model to explain how bounded learners can find high probability hypotheses even in complex learning domains. If there is time, I’ll mention several ways we are extending these ideas this to capture other aspects of human learning.

Key paper: https://psycnet.apa.org/record/2017-08702-001

Recent extension to concept learning: https://www.sciencedirect.com/science/article/pii/S0010028522000421

And incorporation of bootstrapping: https://www.nature.com/articles/s41562-023-01719-1