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-

It is widely appreciated (e.g. Marr, 1982) that the difficulty of a particular computation varies according to how the input data are presented. What is less well understood is the effect of this computation/representation trade-off within familiar learning paradigms. We argue that existing learning algorithms are often poorly equipped to solve problems involving a certain type of important and widespread statistical regularity, which we call `type-2 regularity'. The solution in these cases is to trade achieved representation against computational search. We investigate several ways in which such a trade-off may be pursued including simple incremental learning, modular connectionism, and the developmental hypothesis of `representational redescription'. In addition, the most distinctive features of human cognition --- language and culture --- may themselves be viewed as adaptations enabling this representation/computation trade-off to be pursued on an even grander scale.

We outline the rationale and preliminary results of using the State Context Property (SCOP) formalism, originally developed as a generalization of quantum mechanics, to describe the contextual manner in which concepts are evoked, used, and combined to generate meaning. The quantum formalism was developed to cope with problems arising in the description of (1) the measurement process, and (2) the generation of new states with new properties when particles become entangled. Similar problems arising with concepts motivated the formal treatment introduced here. Concepts are viewed not as fixed representations, but entities existing in states of potentiality that require interaction with a context—a stimulus or another concept—to ‘collapse ’ to an instantiated form (e.g. exemplar, prototype, or other possibly imaginary instance). The stimulus situation plays the role of the measurement in physics, acting as context that induces a change of the cognitive state from superposition state to collapsed state. The collapsed state is more likely to consist of a conjunction of concepts for associative than analytic thought because more stimulus or concept properties take part in the collapse. We provide two contextual measures of conceptual distance—one using collapse probabilities and the other weighted properties—and show how they can be applied to conjunctions using the pet fish problem.

A basic claim of this paper is that the foundational theoretical problem of the social sciences- the possibility of unconscious, unplanned forms of cooperation and intelligence among intentional agents (the very hard issue of the ‘invisible hand’, of the ‘spontaneous social order ’ but also of ‘social functions’)- will eventually be clarified thanks to the contribution of AI (and in particular of cognitive Agent modelling, learning, and MAS) and its entering the social simulation domain. After introducing Multi-Agent-Based Social Simulation and its trends, the limits of the very popular notion of ‘emergence’ are discussed; Smith’s and Hayek’s view of ‘spontaneous social order ’ are critically introduced; and serious contradictions in the theory of ‘social functions ’ among intentional agents are pointed out. The problem is how to reconcile the ‘external’ teleology that orients the agent’s behaviour with the ‘internal ’ teleology governing it. In order to account for the functional character of intentional action, we need a somewhat sophisticated model of intention, and a different view of layered cognitive architectures combining explicit beliefs and goals with association and conditioning. On such a basis it is sketched a model of unknown functions impinging on intentional actions through a high level form of (MA)reinforcement learning. This model accounts for both eu-functions and dys-functions, autonomous and heteronomous functions. It is argued that in order to reproduce some behaviour, its effects should not necessarily be 'good ' i.e. useful for the goal of the agent or of some higher macro-system.

This is the pre-copy-edited draft. Emendations and corrections resulted from copy-editing. Please check the published version before you use material from this paper. All rights reserved. Not to be quoted, paraphrased, copied, or distributed in any fashion.

this paper we investigate the claim that complex causal interactions cause trouble for the notion of inner representational vehicles. We review some of the cases supposed to put pressure on a representational-vehicle based understanding and conclude that the threat, even in these ongoing, interactive cases, is more apparent than real. The main contribution of the present paper, however, is to go beyond this negative thesis

Standard conceptions of how the environment influences the person are constrained by the dominant view of representation - and, therefore, perception, cognition, and language - as fundamentally consisting of encodings. I argue that this encoding view is logically incoherent. An alternative view of representation is presented, interactivism, and shown to avoid the incoherencies of encodingism. The interactivist model of representation provides accounts for standard presumed encoding phenomena, and highlights processes and forms of influence of the environment on the person that are obscure or entirely absent from the encoding account. The multiplicity and complexity of the processes of environmental influence acquire a theoretically coherent organization and development from within the interactive perspective.

ample, may see a fly, and, therefore, have the potentiality of flicking its tongue in a certain way followed by eating. But it may simultaneously see the shadow of a hawk overhead, in which case it also has the selection option of jumping into the water. Both potentialities must be somehow indicated to or for the frog so that a selection between them can occur. Furthermore, if the hawk shadow is not present and the frog misses the fly, it may be advantageous to detect that failure of the tongue flicking action and, on the basis of that detection, to make a further selection of interaction. That further selection might be to try again, or might be to move to a different location where flies are perhaps more numerous or slower. It can be advantageous, in other words, to be able to detect failures of actions, as well as to be able to select among potential actions. A slight addition to the ability to indicate potential interactions suffices to allow such error detection. In pa

this paper on these three: representation, action, and motivation. In particular, I will argue that the standard view of representation as some kind of correspondence, as an encoding, is wrong. I outline an alternative model of representation that emerges naturally in agents, biological or designed, that actually engage the world (Beer, 1990, 1995, in press; Beer, Chiel, Stirling, 1990; Bickhard, 1980, 1993; Bickhard & Terveen, 1995; Brooks, 1991a, 1991b, 1991c; Cherian & Troxell, in press; Malcolm, Smithers, Hallam, 1989; Smithers, 1994). One primary consequence of this alternative model of representation --- called interactivism --- is that functions that are standardly taken to reside in separate modules, such as representation, action, and motivation, are inherently integrated as separate functional aspects of one single underlying ontology. They are not inherently distinct modules. If standard models that permit such modularization are in error, then so are such modularizations per se. 2 Encoding Models of Representation.

On behalf of: