Blending and choosing within one mind: Should judgments be based on exemplars, rules,or both?

Abstract

Blending and Choosing Within One Mind: Should Judgments Be Based on Exemplars, Rules, or Both? Stefan M. Herzog (herzog@mpib-berlin.mpg.de) Center for Adaptive Rationality, Max Planck Institute for Human Development, Lentzeallee 94 14195 Berlin, Germany Bettina von Helversen (bettina.vonhelversen@unibas.ch) Department of Psychology, University of Basel, Missionsstrasse 62a 4055 Basel, Switzerland Abstract Accurate judgments and decisions are crucial for success in many areas of human life. The accuracy of a judgment or decision depends largely on the cognitive process applied. In research on judgment, decision making, and categorization, two kinds of cognitive processes have often been contrasted: exemplar-based processes, which use similarity to previously encountered items to make judgments, decisions, and categorizations, and rule-based processes, which use abstracted cue knowledge. Although most cognitive models of judgment and decision processes assume that people rely on both processes, they differ in whether they assume that one process is selected or that both processes are blended into a single response. The present research takes a functional perspective and investigates what kind of interaction between the two processes leads to accurate responses. Based on cross- validated simulations in real-world domains, it shows that blending rule- and exemplar-based processes generally leads to better judgments than does choosing between them, suggesting that the default strategy should be a blend of both processes, which is abandoned only when feedback justifies it. Keywords: accuracy; multiple-cue judgments; decision making; categorization; exemplar models; rules; cognitive models; mixtures of experts; simulation. Introduction Judging quantities, making decisions, and categorizing items are crucial elements of successful human behavior. A vast and diverse literature in cognitive science and judgment and decision making has investigated how people achieve these tasks (e.g., Ashby & Maddox, 2005; Gigerenzer, Hertwig, & Pachur, 2011; Kruschke, 2008; Payne, Bettman, & Johnson, 1993). The many different models and strategies proposed can be broadly classified into two categories with reference to the cognitive processes they assume: exemplar- based processes, which use similarity to previously encountered items to make judgments, decisions, and categorizations, and rule-based processes, which use abstracted cue knowledge (Hahn & Chater, 1998). Extensive research has compared the proposed models’ ability to describe human behavior. Furthermore, the performance of judgment and decision making strategies in predicting real-world criteria has been thoroughly investigated (e.g., Gigerenzer et al., 2011; Todd, Gigerenzer, & the ABC Research Group, 2012). To our knowledge, however, research in cognitive science and judgment and decision making has not previously investigated what kind of interaction between exemplar- and rule-based processes leads to accurate judgments, decisions, and categorizations: relying on just one of the two processes or using both? If both are considered, is it better to choose between them depending on the structure of the task, for instance (Rieskamp & Otto, 2006), or to blend them into a joint response? This paper presents first answers to these questions. A functional perspective on the interaction between exemplar- and rule based processes may be useful for at least three reasons. First, examining cognitive models’ ability to predict external real-world criteria goes a step further than comparing their ability to describe human behavior in idealized laboratory tasks, by adding a further evaluation criterion. If one class of cognitive models were superior to another in terms of predictive performance, this would make them more attractive as plausible models of human behavior (Chater & Oaksford, 1999). Second, many cognitive models are inspired by or share similarities with models from research fields interested in predictive performance (such as statistics, artificial intelligence, computer science, and machine learning; see e.g., Jakel, Scholkopf, & Wichmann, 2009; Marling, Sqalli, Rissland, Munoz-Avila, & Aha, 2002), and a functional perspective provides a common ground that serves to re-connect cognitive models with such fields. Third, knowledge of how to profit from the complementary strengths of the two processes could offer prescriptions for improving human judgment, decision making, and categorization by instructing decision makers on when and how to use the two processes. Models of Judgment, Decision Making, and Categorization There are two general approaches to modeling human cognition. First, single general-purpose models have been proposed (e.g., Lee & Cummins, 2004). For instance, judgment and categorization models assume either only exemplar-based (e.g., Juslin & Persson, 2002; Kruschke, 1992) or only rule-based processes (e.g., Ashby & Gott, 1988; Brehmer, 1994). Second, toolbox approaches have been proposed. These assume that people draw on multiple,