Probability-matching in 10-month-old infants

Probability-matching in 10-month-old infants Sarah J. Davis (sdavis@bcs.rochester.edu) Department of Brain & Cognitive Sciences, University of Rochester, Rochester, NY 14607 USA Elissa L. Newport (newport@bcs.rochester.edu) Department of Brain & Cognitive Sciences, University of Rochester, Rochester, NY 14607 USA Richard N. Aslin (aslin@cvs.rochester.edu) Department of Brain & Cognitive Sciences, University of Rochester, Rochester, NY 14607 USA Abstract Evidence from the probability learning literature indicates that when presented with simple situations that require making predictions, adults tend to probability match whereas children are likely to show maximization (Stevenson & Weir, 1959; Weir, 1964). The reason for this developmental difference is not fully understood, but one possibility investigated here is that children have fewer resources available to differentiate among the probabilities of the competing alternatives. To investigate this hypothesis at its origin, we used an anticipatory eye movement paradigm to gather two- alternative choice responses from 10-month-old infants. In two experiments we presented infants with either an entirely predictable (100-0%) or a probabilistic (70-30%) series of visual events. Infants showed evidence of probability matching rather than maximizing. These results are discussed in the context of alternative explanations for maximizing and the utility of eye-tracking as a window on infants’ probability learning. Keywords: Probability learning; probability matching; eye- tracking; infants Introduction As we explore our world, we sample from our environment in order to make predictions about future events and to assess the likelihood of receiving rewards. For example, evidence from the statistical learning literature indicates that adults, infants, and animals can extract information about the distributional properties of visual and auditory stimuli in the absence of a task (Fiser & Aslin, 2001, 2002; Kirkham et al., 2002; Saffran et al., 1996a,b; Toro & Trabalon, 2005). Evidence from the causal learning literature indicates that young children are sensitive to event contingencies (Gopnik et al., 2004). This ability to track and store information about probabilities allows learners to adjust their behavior to maximize their predictions and their receipt of rewards, even when there is not a perfect correlation between events and their outcomes. When faced with the task of predicting future events in an uncertain environment a learner has two strategies. One is to make predictions that directly match the exposure probabilities observed in the environment, a pattern known as probability matching. The other is to always choose the more common outcome, a pattern known as maximization. In the context of reward prediction, an ideal learner should choose the action that maximizes the overall rate of reward. However, evidence from the probability learning literature indicates that adults tend to probability match rather than maximize, at least in simple situations (Gardner, 1957; Weir, 1964, 1972). In the classic probability learning experiment, Gardner (1957) presented participants with two light bulbs and asked them on each trial to predict which light would illuminate. After participants made a choice, one of the bulbs would turn on. One bulb turned on 70% of the time and the other bulb 30% of the time. If the participants were probability matching (i.e., picking the 70% light on 70% of the trials and the 30% light on 30% of the trials), then their overall accuracy would average 58% correct. If, on the other hand, learners chose the 70% light on every trial, their overall accuracy would be 70% correct. In this situation, maximizing on the more probable alternative is the better strategy because it leads to higher overall accuracy. Yet under most circumstances, adults typically probability match. Whether adults show probability matching or maximizing on a given task can be influenced by a number of factors, including the contingency of the feedback (Weir, 1972) and the number of response alternatives. For example, when the number of alternatives increases, participants are more likely to show maximizing behavior in both visual tasks (Gardner, 1957; Weir 1964) and auditory language learning experiments (Hudson Kam & Newport, 2009). The age of participants is also related to performance in these tasks, with the youngest children often exhibiting the highest rates of maximizing behavior (Austin & Newport, unpublished manuscript; Hudson Kam & Newport, 2009; Stevenson & Weir, 1959; Weir, 1964). When given access to the same input, why might children act differently than adults? It seems unlikely that they are better strategizers than adults. Rather this behavior could be based on their greater cognitive limitations, either in their representations of the world or their use of those representations. When a learner comes into an environment where there are two possible outcomes, such as the two light bulb task, it