The Architecture of Cognition

John R. Anderson

• Editorial: Psychology Press
• Año: 2013
• ISBN:
  9781317759522
• Edición: 1
• Categoría: Psychologie
• Categoría: Histoire et théorie en psychologie

Now available in paper, The Architecture of Cognition is a classic work that remains relevant to theory and research in cognitive science. The new version of Anderson's theory of cognitive architecture -- Adaptive Control of Thought (ACT*) -- is a theory of the basic principles of operation built into the cognitive system and is the main focus of the book.

1 | Production Systems and ACT

A Unitary Theory of Mind and Production Systems

THEORISTS ARE STRONGLY influenced by their various preconceptions. The most deeply rooted preconception guiding my theorizing is a belief in the unity of human cognition, that is, that all the higher cognitive processes, such as memory, language, problem solving, imagery, deduction, and induction, are different manifestations of the same underlying system. This is not to deny that there are many powerful, special-purpose “peripheral” systems for processing perceptual information and coordinating motor performance. However, behind these lies a common cognitive system for higher-level processing. Moreover, the essence of what it is to be human lies in the principles of this core, higher-level system. We may not differ from the many mammalian species to which we are related in our peripheral perceptual and motor processes, but we assuredly do differ in our complex thought patterns and our intelligence.

The view that the mind is unitary is certainly not universally held; it may not even be a majority opinion. To quote the most noted proponent of the alternative view:

We may usefully think of the language faculty, the number faculty, and others as “mental organs/’ analogous to the heart or the visual system or the system of motor coordination and planning. There appears to be no clear demarcation line between physical organs, perceptual and motor systems, and cognitive faculties in the respects in question. In short, there seems little reason to insist that the brain is unique in the biological world, in that it is unstructured and undifferentiated, developing on the basis of uniform principles of growth or learning—say those of some learning theory, or of some yet-to-be conceived general-purpose learning strategy—that are common to all domains. (Chomsky, 1980, p. 3)

This faculty approach holds that distinct cognitive principles underlie the operation of distinct cognitive functions. The unitary approach holds that all higher-level cognitive functions can be explained by one set of principles. In some ways the faculty approach seems just plain common sense. Many cognitive theories, extending back at least to the phrenology of Gall (see Boring, 1950, for a discussion of this and the more “reputable” faculty theories) have held this view. Its truth might almost seem a tautology: clearly we perform different intellectual functions, so it might seem that we must have different faculties for these functions. The faculty proposals that have been advanced have always gotten into difficulties in their specifics, but it never has been clear whether there is anything fundamentally wrong with the faculty approach.

The early proposals for unitary systems (for example, stimulus-response, or S-R, theories) were also shown to be basically inadequate. However, the unitary theory found an important metaphor in the modern general-purpose computer and, perhaps more significantly, in symbolic programming languages, which showed how a single set of principles could span a broad range of computational tasks; It also became clear that the set of computational functions was unlimited, meaning that general processing principles were essential to span broad ranges of tasks. It made no sense to create a special system for each conceivable function.

A number of candidates for a general system have been offered, including general problem solvers (Fikes and Nilsson, 1971; Newell and Simon, 1972; Sacerdoti, 1977), general inference systems (Green and Raphael, 1969; McDermott and Doyle, 1980; Robinson, 1967), and general schema systems (Bobrow and Winograd, 1977; Minsky, 1975; Rumelhart and Ortony, 1976; Schank and Abelson, 1977). My research has been predicated on the hypothesis that production systems provide the right kind of general computational architecture for achieving a unitary mental system. The particular line of production system theories I have developed all go under the name ACT. This book will describe a special ACT instantiation called ACT* (to be read ACT-star). As will become clear, ACT* is not just a random member of the ACT series. It is the product I have been working toward for the past seven years. In ACT* the same core system if given one set of experiences develops a linguistic facility, if given another set of experiences develops a geometry facility, and if given another set of experiences develops a programming facility. Therefore ACT* is very much a unitary theory of mind.

ARGUMENTS FOR A UNITARY THEORY OF MIND

One thing that distinguishes us from other creatures is our ability to acquire complex skills. All distinctively human activities—such as mathematics, language, chess, computer programming, sculpture—are acquired skills. There may be a significant innate component to their successful acquisition, but with the possible exception of language it is totally implausible to suggest that we have evolved special faculties or “organs” for mathematics, chess, computer programming, or sculpture. People become expert at activities for which there was no possibility of anticipation in our evolutionary history, and the essence of the human genius is just this plasticity. It is unnecessary to propose special organs for special abilities when we can fashion articulate abilities where there is no possibility of a prior organ. If all these abilities are fashioned from the same initial system (which hardly need be a tabula rasa), then in an important sense the adult human mind is a unitary construction.

Language is an important special case that might be the exception to the rule. It is not totally implausible to propose that it has had a long evolutionary history in which various language-specific adaptations have occurred. However, it seems more plausible that the language-specific adaptations are few and minor, that the language faculty is really the whole cognitive system. In our evolution we may have developed or enhanced certain features to facilitate language, but once developed, these features were not confined to language and are now used in nonlinguistic activities. Thus the mind is a general pool of basic structures and processes, which has been added to under evolutionary pressure to facilitate language. The additions have been used in skills, for example, computer programming, that were not anticipated in the original evolutionary developments. Part of the evidence for this view are the remarkable communalities between language and other skills, which will be discussed later in this book.

There is a tendency to regard the existence of “language areas” and other localizations of function in the brain as strong evidence for faculties. However, there is nothing necessary about this inference, as shown by a computer analogy: two programs can occupy different areas of computer memory, much as two different cognitive abilities might lie in two separate regions of the brain. However, the two programs may have identical principles. For instance, I can have one ACT simulation doing language and another doing geometry. Thus, there need be no connection between distinct physical location and distinct cognitive principles. The real issue concerns the uniqueness of the structure and processes underlying cognitive functions, not their physical location.¹

Another major reason for not believing in an organ for language or for other cognitive activities is that the boundaries between these organs cannot be drawn a priori. It is pretty clear where the activity of the lung leaves off and that of the circulatory system takes over, but this cannot really be said for cognitive faculties. The lung and the heart are both involved in an activity such as running, but it is possible to identify their distinctive contributions. It has been proposed that there is a language faculty, a number faculty, a deduction faculty, and a problem-solving faculty, but if there are such faculties, their activities are terribly intertwined in a task like computer programming. When we look at an expert programmer creating a program, we cannot separate the contributions of the various faculties. Indeed, if we applied any reasonable criterion for individuating faculties, we would have to conclude that computer programming was a separate faculty. This is because some of the core principles for this skill organization, such as strategies for creating recursive programs, apply across the entire range of programming behaviors and are seldom if ever evoked elsewhere. Since it is nonsense to suggest a programming faculty, we should be more skeptical of other proposed faculties.

An expert’s execution of a skill is special in that a strong task-specific cognitive organization has developed through extensive experience. This is not the case with the novice, but analysis of novice behavior gives no more comfort to the faculty approach. The remarkable feature of novices is that they are able to put together so many different facets of knowledge to solve a task. A novice programmer brings together recent facts learned about programming and the programming language, facts from mathematics, real-world experiences as analogies, general problem-solving skills, deductive strategies, linguistic analyses —all to solve the problem. The novice’s attempts at synthesizing this knowledge can be terribly off target, but this is only for lack of the right knowledge, not because of a fundamental incompatibility of the knowledge categories. What is remarkable is the ease with which novices switch among categories and the sheer impossibility of identifying where one faculty might begin and another end. Compartmentalization is similarly impossible in the case of language use (see Schank and Birnbaum, in press).

In summary then, there are three lines of evidence for the unitary approach. One is the short evolutionary history of many of the higher human intellectual functions, such as those concerned with mathematical problem solving. The second is that humans display great plasticity in acquiring functions for which there was no possibility of evolutionary anticipation. The third is that the various cognitive activities have many features in common.

I would like to head off two possible misinterpretations of my position. First, the unitary position is not incompatible with the fact that there are distinct systems for vision, audition, walking, and so on. My claim is only that higher-level cognition involves a unitary system. Of course, the exact boundaries of higher-level cognition are a little uncertain, but its contents are not trivial; language, mathematics, reasoning, memory, and problem solving should certainly be included. Second, the unitary position should not be confused with the belief that the human mind is simple and can be explained by just one or two principles. An appropriate analogy would be to a programming language like INTERLISP (Teitleman, 1976), which is far from simple and which supports a great variety of data structures and functions. However, it is general-purpose, that is, one can use the same data structures and processes in programs for language and for problem solving. Individual programs can be created that do language and problem solving as special cases. In analogy to INTERLISP, I claim that a single set of principles underlies all of cognition and that there are no principled differences or separations of faculties. It is in this sense that the theory is unitary.

Production Systems: History and Status

Production system theories have gradually increased in prominence in psychology over the past decade. Their basic claim is that underlying human cognition is a set of condition-action pairs called productions. The condition specifies some data patterns, and if elements matching these patterns are in working memory, then the production can apply. The action specifies what to do in that state. The basic action is to add new data elements to working memory. Informally stated, a typical production rule mi...