Skip to main content

Hypothesis generation, evaluation and testing: An organic, developmental perspective

Final Report Summary - H-GET (Hypothesis generation, evaluation and testing: An organic, developmental perspective)

How do young children learn so much about the world, and so efficiently? Despite the centrality of this question, we do not appear to have very clear answers. My research investigates theoretically and empirically how children actively seek information in their physical and social environments as evidence to test and dynamically revise their hypotheses and theories over time. My program explores the development of active learning across the life span, analyzing the effectiveness of childrens information search and hypothesis testing strategies, such as question asking and selective exploration, and identifying potential sources of developmental change. In particular, my work inaugurates the developmental investigation of ecological learning, defined as the ability to flexibly and dynamically select those active learning strategies that maximize learning efficiency in different learning environments. Finally, I am interested in developing an approach to classroom learning that leverages childrens active learning strategies and theory-building abilities and harnesses them to inform education policy.
By bringing together methods and insights from developmental and cognitive psychology, philosophy, education, Bayesian reasoning, information theory and computational modeling, my approach offers a unique, multidisciplinary perspective to shed light on the cognitive, social and cultural mechanisms underlying active and ecological learning.

Developmental change in the effectiveness of active learning strategies
Deciding what evidence is most valuable to obtain is a basic challenge faced by learners of any age. Research investigating childrens active information search has used variants of the 20-questions game, where the task is to identify an unknown target object by asking as few yes-or-no questions as possible, either generating the questions from scratch or selecting them from a list of given alternatives. In my work, I take a novel developmental and computational approach to explore the efficiency of information search across the life span, by designing original tasks and developing formal, finer grained analyses of performance modeled within a Bayesian framework and based on expected information gain.

How does the ability to efficiently search for information develop across the life span? In a series of experiments using variants of the 20-questions game, I have found that the ability to efficiently ask questions and explore the environment undergoes a large developmental change from age 4 to adulthood (Ruggeri & Lombrozo, 2015; Ruggeri, Lombrozo, Griffiths, & Xu, 2016). For example, although 4- to 6-year-olds are able to spontaneously generate questions in a 20-questions game, those questions are often not the most effective (Ruggeri, Walker, Lombrozo, & Gopnik, in prep.). However, despite having difficulty generating informative questions from scratch, 5-year-olds can already identify the most effective among given questions (Ruggeri, Sim, & Xu, 2017). These results suggest that preschoolers have the computational foundations for developing successful question-asking strategies.

What are the mechanisms underlying developmental differences? Across several studies, I have found that the observed developmental change can be partially explained by childrens increasing ability to generate higher-order features that can be used to cluster similar objects into categories (e.g. quadrupeds vs. nonquadrupeds; see Ruggeri & Feufel, 2015) and by the development of more general verbal abilities and vocabulary (Ruggeri, Walker, Lombrozo, & Gopnik, in prep.). Additionally, my computational findings (Ruggeri, Lombrozo, Griffiths, Xu, 2016) provide compelling evidence of developmental differences in the implementation of stopping rules in information search: Children are significantly more likely than adults to continue their search for information beyond the point at which a single hypothesis remains and therefore to ask questions and select objects that are not informative. Seeking confirming evidence even when it is not strictly informative could make sense when there is uncertainty about the hypothesis space: As novice learners in a noisy world, children might do well to err on the side of obtaining extra feedback.

Ecological learning: Selecting the most efficient active learning strategies
Active learning strategies cannot be defined as optimal tout court. Instead, their efficiency depends on childrens prior knowledge and expectations, as well as on the tasks information structure, that is, the number of hypotheses available and their likelihood (Ruggeri & Lombrozo, 2015). For example, traditionally, the quality of a question has been tied to whether it is a constraint-seeking or hypothesis-scanning question (Mosher & Hornsby, 1966). Constraint-seeking questions target a feature shared by multiple hypotheses (e.g. Was the boy late because of something related to means of transportation?), whereas hypothesis-scanning questions target a single hypothesis (e.g. Was the boy late because his bike was broken?). Because constraint-seeking questions are able to rule out multiple hypotheses at each step of the search process, they are usually considered superior to hypothesis-scanning questions; however, this is not always the case. In my work, I consider for the first time how the traditional distinction between different question types maps onto the more formal distinction between more and less informative questions, as measured by their expected information gain, depending on the information structure of the problem being considered.

With this ecological learning perspective, I have provided the first evidence to demonstrate that 7- to 10-year-olds and young adults change the types of questions they ask in response to the information structure of the task (Ruggeri & Lombrozo, 2015). Across several experiments, I have also demonstrated that children as young as 5 years old successfully rely on different types of questions depending on their efficiency in the given hypothesis space, selecting the question type associated with higher information gain (Ruggeri, Sim, & Xu, 2017).

How does ecological learning develop across the life span? Despite the general developmental increase in performance, my work shows that adults do not adapt their active learning strategies more promptly than children do (Ruggeri & Lombrozo, 2015). On the contrary, children seem to be sometimes even more sensitive than adults to the information structure of the task at hand. For example, when presented with an open causal inference task (Yesterday, a man was late to work. Why?), 9-year-old children, but not adults, asked different types of questions depending on the a priori likelihood of the solution (Ruggeri & Lombrozo, in prep.).

Benefits of active learning and its potential to inform education
Despite widespread consensus that active learning leads to better outcomes than comparatively passive forms of instruction, it is often unclear when and why active learning benefits arise. For example, my previous work shows that in decision-making tasks children focus on more informative cues and are more accurate when they generate their own questions, as compared to when they are given a set of questions to choose among (Ruggeri & Katsikopoulos, 2013; Ruggeri, Olsson, & Katsikopoulos, 2015). On the same line, I found that actively controlling the flow of information during learning, as compared to merely observing the learning experience of another participant, leads to advantages in episodic memory for 7- to 9-year-olds, and that these effects last and increase after 1 week (Ruggeri, Markant, Gureckis, & Xu, 2016a). However, other work shows that the information search of 5- to 10-year-olds is more efficient when they select among given questions than when they generate questions from scratch, in both categorization (Ruggeri & Feufel, 2015) and causal inference tasks (Ruggeri, Walker, Lombrozo, & Gopnik, in prep.). More research is needed to identify what kinds of self-directed activities are beneficial in particular learning environments and how such benefits depend on underlying cognitive processes and more general developmental factors (Markant, Ruggeri, Gureckis, & Xu, 2016b)