Does the utility of information influence sampling behavior?

Does the utility of information in uence sampling behavior? Doug Markant (doug.markant@nyu.edu) Todd M. Gureckis (todd.gureckis@nyu.edu) New York University Department of Psychology 6 Washington Place, New York, NY 10003 USA Abstract reasoning task where the predictions of these measures could be readily distinguished. Learners who could query different stimulus features before making classi cation decisions were found to prefer to learn about features that maximized prob- ability gain, a measure of how well a potential observation is expected to improve classi cation accuracy. In these studies, however, costs were not explicitly manip- ulated or controlled. Taking costs into account can alter the optimal strategy in a given task, but it is unclear whether peo- ple adjust their behavior in a similar way. e goal of the present paper is to explore the impact of costs on sampling de- cisions. We begin by evaluating two alternative objectives that people may adopt when deciding what information to gather. Like the models reviewed by Nelson (2005), the rst ignores the implications of task-speci c costs and casts information sampling strictly in terms of uncertainty reduction (i.e., in- formation gain). e second approach balances the costs and expected bene ts of information in the context of the task. We then describe the results of two experiments that manip- ulate the concordance between these two approaches, in one case creating an environment where the goals of minimizing uncertainty and maximizing utility predict different patterns of information sampling. Our results show that people tend to value information (in terms of the number of hypotheses ruled out by a new observation) over situation-speci c costs and bene ts. e implications of these results for theories of information sampling are discussed. A critical aspect of human cognition is the ability to actively query the environment for information. One important (but oen overlooked) factor in the decision to gather information is the cost associated with accessing different sources of informa- tion. Using a simple sequential information search task, we ex- plore the degree to which human learners are sensitive to vari- ations in the amount of utility related to different potential ob- servations. Across two experiments we nd greater support for the idea that people gather information to reduce their uncer- tainty about the current state of the environment (a “disinter- ested”, or cost-insenstive, sampling strategy). Implications for theories of rational information collection are discussed. Keywords: information sampling, active learning, information access costs Introduction From controlling the movement of our eyes to determining which sources of news to consult, judging the quality of al- ternative sources of information is a critical part of adaptive behavior. Research exploring how people make information gathering (or “sampling”) decisions has shown that people can discern subtle differences in the potential information value of various aspects of the environment. For example, measure- ments of eye movements during object categorization show that people preferentially allocate attention to object features that are most useful for making subsequent classi cation deci- sions (c.f., Nelson & Cottrell, 2007; Rehder & Hoffman, 2005). One aspect that typically complicates the analysis of infor- mation sampling behavior is that information rarely comes for free. All natural tasks involve information access costs, even if the only cost is the time required to gather information (Fu, 2011). In addition, different pieces of information may be more useful depending on how one will be tested in the fu- ture. Optimal search behavior must weigh the costs of collect- ing particular bits of information against the bene t it is ex- pected to convey (Edwards, 1965; Juni, Gureckis, & Maloney, 2011; Tversky & Edwards, 1966), a point frequently made in research on animal foraging (Stephens & Krebs, 1986). Despite its importance, previous work on information sam- pling has oen failed to test whether people take into account costs related to different sources of information. For exam- ple, Nelson (2005) provides a comprehensive review of various ways an optimal Bayesian agent might value potential infor- mation sources in the absence of task-speci c costs (see also Nelson et al., 2010). One conclusion from this line of work is that people make information search decisions that are con- sistent with normative measures of information value (many of which oen make similar predictions). For example, Nel- son et al. (2010) studied information sampling in a diagnostic e rational analysis of information sampling: comparing “interested” and “disinterested” search How should a rational agent make information sampling deci- sions? Existing proposals fall into two broad categories which, borrowing from Chater, Crocker, and Pickering (1998), we call “interested” and “disinterested.” Unlike the distinctions explored by Nelson (2005), these two proposals differ signi - cantly in terms of the overall goal of information sampling. Interested (or cost sensitive) sampling e rst approach represents a decision-theoretic approach to information sam- pling. In particular, the agent considers the cost for collecting a piece of evidence and weighs this against the expected bene t it should convey with respect to the goals of the task. For ex- ample, a car shopper might decide if the possible savings avail- able from obtaining information contained in a vehicle history report is worth the cost of the report. Similarly, preferentially xating the features of an object that are diagnostic of its cat- egory membership may be a cost-sensitive strategy under the