A rational model of human categorization behavior is presented that assumes that categorization reflects the derivation of optimal estimates of the probability of unseen features of objects. A Bayesian analysis is performed of what optimal estimations would be if categories formed a disjoint partitioning of the object space and if features were independently displayed within a category. This Bayesian analysis is placed within an incremental categorization algorithm. The resulting rational model accounts for effects of central tendency of categories, effects of specific instances, learning of linearly nonseparable categories, effects of category labels, extraction of basic level categories, base-rate effects, probability matching in categorization, and trial-by-trial learning functions. Al-though the rational model considers just I level of categorization, it is shown how predictions can be enhanced by considering higher and lower levels. Considering prediction at the lower, individual level allows integration of this rational analysis of categorization with the earlier rational analysis of memory (Anderson & Milson, 1989). Anderson (1990) presented a rational analysis ot 6 human cog-nition. The term rational derives from similar "rational-man" analyses in economics. Rational analyses in other fields are sometimes called adaptationist analyses. Basically, they are ef-forts to explain the behavior in some domain on the assump-tion that the behavior is optimized with respect to some criteria of adaptive importance. This article begins with a general char-acterization ofhow one develops a rational theory of a particu-lar cognitive phenomenon. Then I present the basic theory of categorization developed in Anderson (1990) and review the applications from that book. Since the writing of the book, the theory has been greatly extended and applied to many new phenomena. Most of this article describes these new develop-ments and applications. A Rational Analysis Several theorists have promoted the idea that psychologists might understand human behavior by assuming it is adapted to the environment (e.g., Brunswik, 1956; Campbell, 1974; Gib-

Early studies of intuitive judgment and decision making conducted with the late Amos Tversky are reviewed in the context of two related concepts: an analysis of accessibility, the ease with which thoughts come to mind; a distinction between effortless intuition and deliberate reasoning. Intuitive thoughts, like percepts, are highly accessible. Determinants and consequences of accessibility help explain the central results of prospect theory, framing effects, the heuristic process of attribute substitution, and the characteristic biases that result from the substitution of nonextensional for extensional attributes. Variations in the accessibility of rules explain the occasional corrections of intuitive judgments. The study of biases is compatible with a view of intuitive thinking and decision making as generally skilled and successful.

ents. A Bayesian model may be considered an optimal account of induction that, ideally, would make predictions that bear some resemblance to what people actually do in inductive reasoning tasks. Assumptions for Rational Analysis The Bayesian model presented here is meant to be a computational-level account (Marr, 1982), in that it is a description of the task that is performed in evaluating inductive arguments, rather than a detailed process-level account. In this way, the Bayesian account fulfils the first step of Anderson's (1990) scheme for rational analyses, specifying the goals of the system during a particular task. However, this account does not contain other elements of a rational analysis, such as a description of the environment. For inductive reasoning, the environment might be something as large as all properties of all objects, or all beliefs about properties of objects, and it is not clear how a description of the environment would be undertaken. The Bayesian model for in

Recent research on contingency judgment indicates that the judged predictiveness of a cue is dependent on the predictive strengths of other cues. Two classes of models correctly predict such cue interaction: associative models and statistical models. However, these models differ in their predictions about the effect of trial order on cue interaction. In five experiments reported here, college students viewed trial-by-trial data regarding several medical symptoms and a disease, judging the predictive strength of each symptom with respect to the disease. The results indicate that trial order influences the manner in which cues interact, but that neither the associative nor the statistical models can fully account for the data pattern. A possible variation of an associative account is discussed. The ability to detect predictive relationships among envi-ronmental events grants humans and other animals a distinct benefit. Therefore, the mechanisms underlying this ability are of considerable interest. Recent research with humans on judgments of contingencies has shed light on these mecha-nisms. It has suggested two classes of theoretical models that

Abstract not found

Findings in recent research on the `conjunction fallacy ' have been taken as evidence that our minds are not designed to work by the rules of probability. This conclusion springs from the idea that norms should be content-blind Ð in the present case, the assumption that sound reasoning requires following the conjunction rule of probability theory. But content-blind norms overlook some of the intelligent ways in which humans deal with uncertainty, for instance, when drawing semantic and pragmatic inferences. In a series of studies, we ®rst show that people infer nonmathematical meanings of the polysemous term `probability' in the classic Linda conjunction problem. We then demonstrate that one can design contexts in which people infer mathematical meanings of the term and are therefore more likely to conform to the conjunction rule. Finally, we report evidence that the term `frequency ' narrows the spectrum of possible interpretations of `probability ' down to its mathematical meanings, and that this fact Ð rather than the presence or absence of `extensional cues ' Ð accounts for the low proportion of violations of the conjunction rule when people are asked for

Inductive learning is impossible without overhypotheses, or constraints on the hypotheses considered by the learner. Some of these overhypotheses must be innate, but we suggest that hierarchical Bayesian models help explain how the rest can be acquired. To illustrate this claim, we develop models that acquire two kinds of overhypotheses — overhypotheses about feature variability (e.g. the shape bias in word learning) and overhypotheses about the grouping of categories into ontological kinds like objects and substances.

this paper is to show that the category inclusion principle has only limited descriptive validity

Human concept learning presents a version of the classic problem of induction, which is made particularly difficult by the combination of two requirements: the need to learn from a rich (i.e. nested and overlapping) vocabulary of possible concepts and the need to be able to generalize concepts reasonably from only a few positive examples. I begin this thesis by considering a simple number concept game as a concrete illustration of this ability. On this task, human learners can with reasonable confidence lock in on one out of a billion billion billion logically possible concepts, after seeing only four positive examples of the concept, and can generalize informatively after seeing just a single example. Neither of the two classic approaches to inductive inference -- hypothesis testing in a constrained space of possible rules and computing similarity to the observed examples -- can provide a complete picture of how people generalize concepts in even this simple setting. This thesis prop...

accuracy issues more directly. Moreover, this research attracts a great deal of attention because of what many take to be its dismal implications for the accuracy of human social reasoning. These implications are illusory, however, because an error is not the same thing as a "mistake. " An error is a judgment of an experimental stimulus that departs from a model of the judgment process. If this model is normative, then the error can be said to represent an incorrect judgment. A mistake, by contrast, is an incorrect judgment of a real-world stimulus and therefore more difficult to determine. Although errors can be highly informative about the process of judgment in general, they are not necessarily relevant to the content or accuracy of particular judgments, because errors in a laboratory may not be mistakes with respect to a broader, more realistic frame of reference and the processes that produce such errors might lead to correct decisions and adaptive outcomes in real life. Several examples are described in this article. Accuracy issues cannot be addressed by research that concentrates on demonstrating error in relation to artificial stimuli, but only by research that uses external, realistic criteria for accuracy. These criteria might include the degree to which judgments agree with each other and yield valid predictions of behavior. The accuracy of human social judgment is a topic of obvious