Children's Rational Exploration

Very young children have remarkably sophisticated causal knowledge about the world, yet relatively little is known about the process of causal learning. In this paper we provide a Bayesian model of how the interaction of prior theories and evidence can lead to ambiguity in competing causal hypotheses; we suggest that children seek to resolve such ambiguities through active exploration. In Experiment 1, we look at the model with respect to children’s causal judgments. In Experiments 2 and 3, we show that children selectively engage in exploration when evidence is formally ambiguous with respect to their prior theories. We suggest that children’s play is rational with respect to this model and that children’s active exploration of causal ambiguities might generate evidence that could support theory formation and theory change. Children’s Causal Reasoning Very young children have remarkably sophisticated causal knowledge about the world in a variety of domains. Children reason about the causes of mental states such as beliefs and desires (e.g. see Wellman, 1990), in the domain of physics, they reason about object properties and forces (e.g. Bullock, Gelman, & Baillargeon, 1982; Shultz, 1982) and with respect to their naïve biology, they reason about causes pertaining to illness and growth (Gelman & Wellman, 1991; Kalish, 1996). Many researchers have suggested that children’s causal knowledge can be characterized as intuitive theories: abstract, coherent, defeasible representations of causal structure (Carey, 1985; Gopnik & Meltzoff, 1997; Wellman, 1990; Keil, 1989). However, despite the importance of theories for children’s predictions, explanations, counterfactual reasoning, and exploration, relatively little is known about the processes responsible for this kind of causal learning. Some researchers have suggested that children’s naive theories might be instantiated in domain-specific modules, or innate concepts in core domains, (Carey & Spelke, 1994; Keil, 1995). For example, some researchers have argued that we have core knowledge about objects, agents, and number (Carey & Spelke, 1994). However, other researchers have emphasized the role of domain-general learning mechanisms, such as sensitivity to patterns of statistical evidence. Of the few studies that have directly compared domain-specific and domain-general causal learning, some have suggested that both adults and children privilege domain-specific information over domain-general evidence, (Ahn, Kalish, Medin, & Gelman, 1995; Bullock, Gelman & Baillargeon, 1982; Shultz, 1982). By contrast, other research suggests that children can use domain general learning mechanisms (such as the conditional probability of events) to override domain boundaries (e.g. Schulz.& Gopnik, 2004). Because previous research has focused on cases when evidence either overwhelmingly favored a domain inappropriate cause (suggesting the strength of domain general arguments), or cases when theories were strongly instantiated and little counter-evidence was available (suggesting the strength of prior theories), little work has demonstrated a graded interaction between the two. However, in previous research (Bonawitz, Griffiths, & Schulz, 2006; Schulz, Bonawitz, & Griffiths, in press), we proposed a formal model that suggested how prior knowledge and statistical updating may interact in cases when children are presented with ambiguous evidence. Our model is applicable both to cases when children have strong prior knowledge and cases when they do not. In this paper, we will argue that both domain-specific theories and statistical evidence play an important role in children’s causal inferences. In particular, we will describe scenarios in which the interaction of prior knowledge and evidence leads to ambiguity between two potential candidate causes. First we will describe how ambiguity arises and is reflected in children’s causal inferences. Then, we will suggest that in these cases of ambiguity, it is more optimal to explore than in cases when a single likely causal hypothesis is strongly favored. We will suggest that, although children may not construct carefully controlled experiments, their spontaneous exploration reflects sensitivity to these formal instances of ambiguity and is thus rational with respect to our model. We suggest that this sophistication in exploratory play is one mechanism that can allow children to ‘construct’ new knowledge and support the processes involved in theory change.