Knowledge Acquisition and Use in Higher-Order Cognition

Giyoo Hatano & Kayoko Inagaki. The International Handbook of Psychology. Editor: Kurt Pawlik & Mark R Rosenzweig. Sage Publications. 2000.

This chapter reviews what psychologists know about knowledge acquisition and knowledge use in higher-order cognition or thinking as goal-directed activities involving transformations of mental representations. It focuses on problem solving and comprehension. These are, we believe, two major manifestations of higherorder cognition, because they have clear adaptational values that are complementary: whereas comprehension is a coherence-seeking process through which people can grasp what the world is like, problem solving is search in a space of alternatives which enables people to change the world as they want. We incorporate findings from research on cognitive development (including expertise), because cognitive development is almost equivalent to the knowledge acquisition through problem solving and comprehension, and shows what a long-term consequence of these higher-order cognitive activities is like. We also refer very briefly to experimental cognitive psychology of the constituent processes of higher-order cognition, concepts and deduction as examples.

This chapter emphasizes the active, constructive, and creative nature of the mind, that is, that humans have a tendency to interpret what they observe and to find problems to solve, not to absorb information presented to them or to try to solve externally imposed problems. At the same time, it emphasizes roles of cognitive, biological, and sociocultural constraints (enabling and restricting factors and conditions) that are imposed on the working of the mind.

History and Basic Concepts

Brief History of the Study of Higher-Order Cognition

Studies of higher-order cognition, thinking, knowledge, or reasoning date back at least to Aristotle. Even within the framework of experimental investigation, such studies have a long history of more than 120 years, as a number of psychologists in the nineteenth century engaged in empirical, though rather introspective, studies of thought. However, higher-order cognition was not a popular topic in psychology until recently. This was because, among others, behaviorism, the dominant paradigm in psychology from the 1920s to the early 1950s, barred the studies that relied on such mentalistic notions as representation, hypothesis, and interpretation. These studies were almost completely suppressed in North America, where behaviorism was very powerful. Studies of higher-order cognition survived slightly better on the European continent, where behaviorism was less salient, and were led by scholars who were interested in human development and education as well as in intelligence and creativity—the German Gestalt school represented by M. Wertheimer and K. Duncker (known for studies on productive thinking and insight), L. S. Vygotsky (higher mental functioning), and J. Piaget (logical operations as internalized actions).

Higher-order cognition resurfaced as a legitimate topic in psychology after the cognitive revolution in the 1950s, which is now often called its first wave. However, studies of higherorder cognition were not in the mainstream until the mid-1970s (with notable exceptions, such as Bruner, Goodnow, and Austin’ 1956 study on hypothesis testing, Newell and Simon’ 1972 study on information-processing models of problem solving, and a few others). Most of the successful cognitive studies in the early years were concerned with vision and language, and to a lesser extent, with memory, probably because these functions could be researched by manipulating stimuli and observing the corresponding response changes. Post-Piagetian cognitive developmental research was established toward the end of the 1970s.

Thus, studies on higher-order cognition have become an indispensable and integral part of psychology just in the last quarter of the twentieth century. Although still far fewer than studies of memory or language, the number of publications on thinking in the wide sense has increased in such general cognitive psychological journals as Cognitive Psychology and Cognition, and even a new journal specifically devoted to thinking, Thinking and Reasoning, began publication in 1995. Such topics as discourse understanding, concepts and categories, analogy and other forms of informal reasoning, and conceptual development seem to draw the particular attention of cognitive psychologists. Moreover, cognitive studies on thought processes constitute almost a majority in such related areas as instructional and social psychology. Last but not least, studies on higher-order cognition are encouraged and supported by partner sciences in cognitive science.

Active, Constructive, and Creative Mind

As the history of psychology indicates, research on higher-order cognition is not compatible with behaviorism in many aspects. This is obvious in terms of methodology. Research on higher-order cognition has to use, almost inevitably, mentalistic notions and terms, even though the studies themselves are performed using rigorous scientific experimental methods.

Equally important is that behaviorists and higher-order cognition investigators tend to be radically different in characterizing human beings. Whereas behaviorists generally assume that organisms including humans are passive and idle, proponents of higher-order cognition assume them to be active. According to behaviorists (especially C. L. Hull and his followers), organisms become active only when they are exposed to painful stimulation, physiological deficits like hunger and thirst, and other stimuli conditioned on the preceding two (Hunt, 1965). In contrast, most investigators of higher-order cognition believe that humans, and animals with advanced traits (traits that have appeared relatively recently in evolution) to some extent, are active, curious, and inclined to interact with physical and social environments if and only if they are given such opportunities.

Also, unlike behaviorists who assume that learning occurs only by external reinforcement, investigators of higher-order cognition regard human minds as constructive in the sense that they spontaneously coordinate pieces of information to build more and more coherent systems of knowledge. The latter investigators assume rewards or punishments work through being mediated by the interpretation of the learning organism. They may even assume that learning or the construction of knowledge is based on internal feedback, such as consistency within pieces of presented information or with prior knowledge (Hunt, 1965).

Moreover, most investigators of higher-order cognition assume that humans are creative in the sense that they generate new ideas that transcend what they have experienced and what they already know. Although whether the new ideas are significant or novel for other people depends on context, the investigators believe that creativity itself is ubiquitous (Ward, Smith, & Vaid, 1997). This view is in opposition to the views that creative thinking is distinct from other types of thinking or that there is only a small number of creative people. Such views are apparently supported by some psychologists as well as by many lay people.

Key Constituent Notions: Forms of Knowledge

Before discussing what has been learned about knowledge acquisition and use in higher-order cognition, we first present some of the basic notions in the field and their definitions. We will start with the term ‘mental representation.’ This term means pieces of information in the mind that represent features of the external environment. Biologically, it is often claimed, mental representation has evolved for the sake of the survival of individual organisms and species. Because the survival for animals with advanced traits presupposes being able to eat without being eaten, they have to mentally represent the current state of the world and use it when they decide how to behave. More or less stable mental representations, which are based on accumulated experiences, articulated into objects or phenomena in the world and actions on them, are called ‘knowledge’ or ‘beliefs’ (often used interchangeably by psychologists, though philosophers reserve the former term only when it is believed to be veridical).

Human minds abstract empirical rules from the regularities of experiences (Holland, Holy-oak, Nisbett, & Thagard, 1986). Each empirical rule can be described as a condition—action pair (called a production rule), like ‘If X, then do Y.’ Such empirical rules can readily represent the knowledge we have that indicates how to deal with objects and phenomena, like ‘If it is too warm and the windows are shut, open them.’ The rules can represent our factual knowledge by extending the action part to include the sending of an internal message, e.g., ‘If you have a one-dollar bill, then you can change it for four quarters.’ Thus, it does not seem totally impossible to approximate all human knowledge in terms of a set of a large number of these rules. Such a set is often called a ‘production system.’

However, other elaborate forms of knowledge are also involved in higher-order cognition. We will mention just some of them here. First, a set of empirical rules are integrated into representations of what specific objects and phenomena are like; these are called concepts or schemas. They appropriately summarize our general and static knowledge about entities; for instance, our concept of dogs represents what we expect for any dog unless specified otherwise (e.g., has four legs, wags its tail when pleased). Concepts or schemas serve as building blocks in higher-order cognition, because (a) they neatly mobilize related clusters of rules, (b) they are interrelated and constitute a conceptual structure, and (c) they can easily be combined directly (e.g., ‘dog house’ meaning ‘kennel’) or by predicates (e.g., dogs are animals). Some schemas are used to describe a sequence of events (e.g., dining at a restaurant) instead of entities, and these are called scripts.

Second, people may construct, also based primarily on empirical rules, mental models. Unlike concepts or schemas, mental models are dynamic, embodied, and often temporary representations of a particular object or situation. They change in response to mentally exerted actions and thus can be used for mental simulation. How vividly mental models indicate a concrete object or situation varies from context to context and from individual to individual.

Third, people are supposed to have theories in selected domains. The term, ‘theories’ means coherent bodies of knowledge that involve causal explanatory understanding. How similar ordinary people’ naive or intuitive theories are to theories scientists have is still an open question, but the former certainly have something more than a collection of facts and/or procedures to obtain desired results. Explanatory mechanisms distinguish theories from other forms of knowledge, such as the restaurant script.

Fourth, procedural knowledge includes strategies, that is, chains of actions to perform a given task or classes of such action chains sharing performance goals and resource allocations. Because our ability to retain or manipulate several pieces of information at a time is limited, we often rely on strategies that can reduce the processing load, even though they may not be the most efficient way to perform the task.

Finally, humans possess knowledge about knowledge, which is called ‘metacognitive knowledge.’ Metacognitive knowledge plays an important role in higher-order cognition, especially when it is regulated consciously and intentionally.

Concepts and Categorization

Concepts are the most frequently studied and probably the best understood elaborate forms of knowledge, because they are the building blocks of higher-order cognition, as mentioned above, and are small enough for experimentally controlled investigations. Studies of concepts exemplify how complicated the relationships are between individual pieces of experience and the form of knowledge accumulating them. We will review these studies focusing on concepts’ most salient function, categorization.

How can people classify new instances into the proper categories? In other words, in what form are concepts represented in the human mind? There are three major models (Medin, 1989): the classic model, the probabilistic model, and the theory-based model. (There is another prominent model of categorization, the connectionist network model, but we will not discuss it in this chapter.) The classic model, which is sometimes called the rule model, assumes that a concept is defined by common, criterial characteristics. Those entities and events that possess these characteristics, and only those, are classified into that concept. Early experiments on concept attainment (e.g., Bruner, Goodnow, & Austin, 1956) were all based on this model. They investigated how humans could identify the defining attributes based on the exemplars of a given concept.

The probabilistic model was proposed as an alternative to the classic model, because the classic model failed to explain some important phenomena (Medin, 1989). For example, there are everyday concepts (e.g., furniture) for which it is difficult to create a set of defining characteristics; there are marginal instances that we cannot classify with confidence (e.g., whether a television set belongs to the category of furniture); and some instances are more representative than others (e.g., a sofa is a better exemplar of furniture than a hat-rack). A ‘prototype’ version of this model (e.g., Rosch, 1975) assumes that a concept is indicated by a summary representation of many instances belonging to that concept. Common characteristics included in the summary representation are not defining, but serve as probabilistic cues for the category membership. Proponents of this version assert that a concept is best represented by a prototype that has modal characteristics of all instances, and that other instances are graded in terms of their similarities to the prototype.

Another version of this model assumes that a concept in the mind is represented by a set of instances that have been presented and stored (e.g., Nosofsky, 1988), although it shares the assumption that there are no defining characteristics. The summation of a measure of similarity between the new instance and each of the stored instances of a given category determines whether it belongs to this category. In either version, categorization is based on a similarity metric; what differs between the versions is whether the summary representation of properties shared by a majority of members is in the form of a prototype. The process of concept acquisition is, according to this model, to accumulate exemplars (and to build a prototype representing them).

The third, the theory-based model, is against the models that are based on similarity. It points out that similarity alone cannot explain why we have a particular concept (Murphy & Medin, 1985); for instance, why do we possess a category combining animals and plants in spite of the many perceptual differences, but not a concept of humans and dolls? It asserts that the properties that are shared by a majority of exemplars are not independent but correlated, or more accurately, causally connected (e.g., birds have wings, and are light in body weight, in order to fly).

However, as Medin (1989) hints, concepts are both formed based on similarities of instances and embedded in larger networks. Sometimes their integrated representation is salient, and other times they are taken as a collection of constituent instances. These features seem to be shared by all abstract forms of knowledge.

Reasoning: Deduction and Induction

By using forms of knowledge, human beings solve a variety of problems, that is, change the initial state into the goal state under an imposed set of constraints. Similarly, they understand or interpret what the target is like, again relying on an assumed set of constraints. Such processes inevitably involve reasoning. Reasoning (or inference, which is used as the synonym) can be defined as a goal-directed and constrained step-by-step transformation of mental representation or knowledge. It is not limited to logically valid or plausible transformations, but, unlike daydreaming, it has to be controlled by the reasoner’ reasoning rules as well as by domain-specific knowledge. It is obvious that reasoning is critically important in higher-order cognition, because knowledge can seldom be applied to new situations as it is. In other words, knowledge is useful only when appropriately expanded by reasoning. However, reasoning is almost never a salient component in problem solving or comprehension. Most investigators in this field assume that basic deductive and inductive reasoning abilities are shared by participants in their studies. We will briefly examine whether ordinary people can solve reasoning tasks and if they can, how they do so.

Deduction

There have been many laboratory studies on deductive reasoning. Initially, the investigators were interested in examining whether children as well as ordinary adults can reason in accordance with ways prescribed by logicians. However, as research progressed, the interest has shifted to clarifying reasoning strategies and the knowledge used in reasoning. Now we know, with certainty, that people often employ content-specific reasoning procedures, based on the substantive interpretation, even when they are presented with logical reasoning problems that can (and should) be solved purely syntactically.

Wason’ (1966) selection task is one that has been extensively studied. This task gives participating students a conditional rule in the form of ‘if p then q.’ They are also given four cards, each of which showsp, not-p, q, or not-q on one side or the other. They are required to decide which of the cards must be turned over to determine whether the rule is true or false. The logically correct choice is to select the p card (which may have not-q on its back) and the not-q card (which may have p on its back) only. Participants seldom make this correct choice when the conditional rule has arbitrary content (e.g., If a card has a vowel on one side, then it has an even number on the other). Rather, they tend to make various errors, of which the most common is to select the cards corresponding to p and q.

However, soon after the original publication by Wason it was shown that for certain formally equivalent rules indicating realistic content, the p and not-q cards are selected much more often (Johnson-Laird, Legrenzi, & Legrenzi, 1972). For example, for rules such as ‘If a person is to drink alcohol, then he must be at least 20 years old,’ the p (drinking alcohol) and not-q (not yet 20 years old) cases are chosen frequently. This finding could be interpreted as showing that people can handle realistic conditional rules only. However, Cheng and Holyoak (1985) proposed an alternative interpretation: rules that can be interpreted as expressing relations of permission or obligation tend to induce the correct choice. In other words, conditional regulations are understood in terms of permission or obligation schema which are pragmatic reasoning schema based on obligatory relations. To support their proposal, Cheng and Holyoak demonstrated that facilitation could be obtained for an abstract permission rule, ‘If one is to take action A, then one must satisfy precondition P.’

Moreover, more recent findings indicate that a rule may be interpreted differently depending on one’ perspective, or put differently, one’ selection is focused on the cases that are related most directly to one’ own goals (Holyoak & Cheng, 1995). For example, ‘If an employee works on the weekend, then that person gets a day off during the week,’ may be interpreted either as ‘If an employee works on the weekend, then the employer must grant a day off during the week’ (from an employee’ perspective indicating the obligation of an employer), or ‘If an employee works on the weekend, then that person may take a day off during the week’ (from an employer’ perspective indicating the obligation of an employee, because it is equivalent to ‘If an employee takes a day off during the week, then that person must have worked on the weekend’).

To summarize, people do not solve logical reasoning problems as abstract, logical ones. Although some in the field have proposed formal rules allegedly explaining everyday reasoning performance (e.g., Rips, 1994), none of the proposals were firmly supported empirically. People may often use, as claimed by Cheng and Holyoak, pragmatic rules which are in specificity in between formal rules and domain-specific rules. If so, it would be interesting to research how such inference rules are acquired and how they are triggered.

Another prominent paradigm in deductive reasoning research is the mental models theory proposed by Johnson-Laird (1983). This theory assumes that human reasoning consists of constructing a series of mental models, each representing a true possible situation of a given premise. Once these mental models are constructed, one can ascertain whether a conclusion holds in them—the conclusion is possible if it holds in at least one of the models; necessary if it holds in all the models; and impossible if it never holds. In other words, the theory treats deduction as an extension of discourse comprehension (see Section 10.4). Johnson-Laird and Byrne (1995) claim that it can explain the findings not only by Cheng and Holyoak but also by others. Although all its details have not yet been specified, it is a highly promising alternative in the study of human reasoning.

Induction

In contrast to the large number of studies on deduction, induction has not been a popular topic in experimental cognitive psychology (Holland et al., 1986). This is probably because there is no obviously correct answer for induction, and whether one’ answer is appropriate depends on the reasoner’ goal, prior knowledge, and context.

However, this makes induction tasks highly useful for assessing reasoners’ knowledge in research on knowledge acquisition. For example, how far reasoners inductively project a property observed in an entity depends on their understanding of the concept the entity belongs to and the conceptual structure including the concept (e.g., if a goldfish has a heart, it is likely that all fish have a heart).

The observed individual differences in inductive reasoning tasks are usually explained in terms of different prior knowledge, though logically it is possible to attribute them to different levels and modes of reasoning. A great majority of psychologists assume, though implicitly, that humans are not as different in basic reasoning ability as in domain-specific knowledge.

Knowledge Acquisition

People become increasingly competent with the accumulation of experiences over the course of years. Almost all people acquire competence in some domains rapidly and easily, whereas in other domains the acquisition of increased competence takes a long time and requires laborious effort. The former domains, which are called privileged or core domains, concern naive theories or coherent bodies of knowledge, the acquisition of which has been vital for the survival of our species. Good examples include those that deal with objects in space and their movement (naive physics), the human mind (naive psychology), and the human body as well as other animals and plants (naive biology). The latter domains, called non-privileged or peripheral domains, represent organized bodies of knowledge (and sets of skills for using the knowledge) that an individual acquires through his or her profession or hobby. Thus people often vary widely in the extent of mastery in these domains. Chess, cooking, and making diagnosis by reading X-ray films exemplify these domains.

Researchers studying conceptual development have dealt primarily with the former domains and those studying expertise have been concerned with the latter. However, these two lines of study share a focus on the construction and elaboration of domain-specific knowledge under constraints, and their findings are complementary. In what follows we define the term ‘domain’ in a wide sense to deal with both the issue of conceptual development and of expertise; it refers to a range of problem-solving and understanding circumscribed by a common set of constraints or a system of knowledge as produced by these constraints. Here the term ‘constraints’ refers to conditions or factors that facilitate the process of problem solving, comprehension, or aquisition of knowledge by restricting its possible range.

Characterizing Knowledge Acquisition

How is the human knowledge system acquired? We can derive from studies on the long-term acquisition of knowledge by humans four interrelated characterizations.

Knowledge is Acquired by Construction

The first characterization indicates that knowledge is acquired by construction; it is not acquired by transmission alone (Resnick, 1987). That knowledge is constructed is a corollary of the ‘Zeitgeist’ among contemporary cognitive researchers, that is, that human beings are active agents of information processing and action. Humans interact with the environment, find regularities, and construct condition—action rules and more elaborate forms of knowledge. Humans acquire knowledge richer than the knowledge they are presented with, or even invent knowledge which has never been presented, often as a byproduct of their problem solving and/or comprehension activity. Knowledge can be transmitted to some extent, but transmitted knowledge becomes usable in a variety of situations only after it is reconstructed, that is, interpreted and connected to the prior knowledge of the learner.

Knowledge Acquisition Involves Restructuring

The second characterization of knowledge acquisition is that it involves restructuring; that is, not only does the amount of knowledge increase but also one’ body of knowledge is reorganized as more and more pieces of knowledge are acquired (Rumelhart & Norman, 1978). Conceptual change or a fundamental change of a concept is the best known example of the restructuring of knowledge (Carey, 1985). Knowledge systems before and after the conceptual change may sometimes be locally incommensurable; that is, some pieces of knowledge in one system cannot properly be translated into the other, as exemplified by the shift from the children’ undifferentiated heat/temperature concept to the adults’ two separate concepts of heat and temperature.

However, restructuring includes milder and more subtle forms and other levels than concepts. Concepts can be substantially modified, but the old and new concepts are not incommensurable (e.g., the concept of animals is expanded to include birds, fish, and insects); relationships between concepts can also change; the same phenomenon may be explained differently as causal devices change; and so forth. Condition—action rules may also be differentiated or amalgamated. In any case, the occurrence of restructuring can be attributed to the creative mind of humans, because it necessarily involves the deconstruction of knowledge that has had some adaptive value.

Process of Knowledge Acquisition is Constrained

The third characterization denotes that the process of knowledge acquisition is constrained (Gelman, 1990; Keil, 1981). As a consequence the acquisition of some forms of knowledge are easy, but others are difficult. The construction and successive revision of knowledge in a domain takes place under the same set of constraints so that the acquired knowledge in the domain is often similar, if not identical, between different individuals. Many cognitive theorists agree that the process of construction is constrained both by innate tendencies and by acquired prior knowledge (i.e., by cognitive constraints).

In actual studies, however, investigators differ widely in the importance they assign to each of the constraints; investigators who focus on conceptual development consider innate constraints to be critical, whereas those studying expertise emphasize prior knowledge as constraints in the target domain. Therefore, more detailed discussion on constraints will be given below separately for conceptual development and expertise.

Knowledge is Acquired Domain by Domain

The fourth characterization indicates that most knowledge is acquired domain by domain. Recent cognitive studies have demonstrated that individual competence varies considerably from domain to domain. In research on conceptual development, Piaget’ stage theory, which posited that an individual’ competence depended on his or her logico-mathematical structures applicable across domains, has been challenged or even rejected by many current researchers (Siegler, 1978). Many of these researchers believe that, because each domain is under a different set of constraints, the course and process of development vary from domain to domain; in other words, knowledge is acquired, in part, in a unique fashion in each domain.

Investigators of expertise (e.g., Chi, Glaser, & Rees, 1982) have also asserted that the most critical determinant of problem-solving competence is not the general ability, but the relevant domain-specific knowledge. A scientific inquiry or an everyday attempt to understand the world usually takes place within a particular domain. Knowledge produced through such an activity is also incorporated into the relevant domain only. Thus, knowledge is usually acquired separately for each domain, although a small number of pieces of knowledge are shared by a number of domains (e.g., knowledge about literacy or measurement), and analogical transfer or generalization of knowledge based on the recognized isomorphism across domains may sometimes occur (Holyoak & Thagard, 1995).

Conceptual Development

Here we will concentrate on the emergence of and later development in three core domains. Results obtained in recent studies generally show that the privileged domains are differentiated early, due to innate constraints, but the knowledge system in the domains is restructured, probably a few times, before it becomes an adult version.

Emergence of Core Domains of Thought

A growing number of researchers of conceptual development have argued that even young children possess more or less coherent bodies of knowledge about important aspects of the world. They assume that such bodies of knowledge often constitute naive ‘theories’ containing ontological distinctions, coherent organizations among pieces of knowledge, and causal devices or explanations as essential components (Well-man, 1990). A majority of researchers now agree that by about six years of age children have acquired autonomous core domains of thought: naive physics, naive psychology (or a theory of mind), and naive biology (Wellman & Gelman, 1998). Considering that human beings are a distinctly social species, use objects as tools as well as inhabit the physical world, and possess physiological bodies that must be fed and kept healthy, it is plausible that children are endowed with capabilities to construct naive theories about psychological, physical, and biological phenomena early in life. We present below a few supporting findings from recent studies.

Current studies concerning the early acquisition of physical knowledge have revealed that infants expect an inanimate object to move in accord with the physical principles—action on contact, no action at a distance, continuity (objects exist continuously and move on connected paths) and so on—and are surprised if its motion apparently violates them (Spelke, 1991). In addition, Spelke, Phillips, and Woodward (1995) reported that by the end of the first year infants differentiate between an inanimate object motion and human action; for human actions they consider person’ direction of gaze or expression of emotion to be important, while they reason the object motion in accord with the physical principles. This suggests that infants’ reasoning already depends upon domain-specific systems of knowledge as far as physics and psychology are concerned.

Simons and Keil (1995) reported that before acquiring detailed factual knowledge, children possess ‘abstract’ knowledge, for instance, assume different insides for animals and artifacts. When asked to choose one from three photographs of an animal inside, a machine inside and an aggregate substance (e.g., a pile of rocks, or blocks), preschool children consistently picked different insides for animals and artifacts. In addition, when they erred, they tended to assign natural kind inside (e.g., rocks) to the animals and artifact inside (e.g., blocks) to the artifacts.

Wellman, Hickling and Schult (1997) indicated, through both experimental studies and natural language analyses, that young preschool children appropriately apply causal reasoning to psychological, physical, and biological phenomena. Children of 3 and 4 years of age were presented with scenarios consisting of three different types of desired actions and asked if the protagonist could do the desired action, and then required to justify their judgments. The three types of desired actions were: (a) voluntary actions, which the protagonist can do if he wants to perform them (e.g., jump on the floor); (b) physically impossible actions (e.g., floating in the air without any support); and (c) biologically impossible actions (e.g., staying the same size forever). These authors found that not only the 4-year-olds but also the 3-year-olds generated appropriate causal explanations differentially for the possibility of these three types of actions, although offering biological explanations for biological phenomena was somewhat difficult for the 3-year-olds. Moreover, analyses of younger children’ speech in everyday life supported the findings from the experimental studies.

Conceptual Change and Elaboration

That children have naive theories about the important aspects of the world does not mean that their theories are the same as intuitive theories that lay adults have. What changes occur in naive theories or their constituent concepts is a popular topic in conceptual development research.

With regard to naive psychology or theory of mind, Wellman (1990) claims that the first conceptual change (or theory change) is from a desire psychology to a belief—desire psychology at around age 3; the desire psychology that 2-year-old children are supposed to possess is based only on an understanding of simple desire as the determinant of human behavior, whereas the belief—desire psychology involves understanding that people possess belief as well as desire. From age 3 to 6 a second conceptual change occurs within this belief—desire psychology; it is a change from a copy-container theory of mind, i.e., a static mind, to an interpretive-homuncular theory, i.e., an active and constructive mind. After 6 years of age the change involves an increase in the proficiency in children’ belief—desire reasoning. Children elaborate and consolidate an interactive view of representation and an active-homuncular theory of mind.

Hatano and Inagaki (1996) characterize the initial form of young children’ biology as personifying and vitalistic in nature, and describe what needs to be incorporated and/or modified for this biology to become the lay adult’ intuitive biology. Weaknesses in young children’ biology are (a) limited factual knowledge; (b) a lack of inference based on complex, hierarchically organized biological categories; (c) a lack of mechanistic causality; and (d) a lack of conceptual devices in biology. During the early elementary school years, children gradually overcome weaknesses (a) to (c) through conceptual change. This change is almost universal, at least among children growing up in highly technological societies. However, the authors claim, the acquisition of conceptual devices of truly biological nature (e.g., ‘evolution,’ ‘photosynthesis’) requires systematic teaching, and only a limited portion of older children and adults may learn them even with good biological instruction.

To generalize, later conceptual development in privileged domains does not proceed as easily as earlier development. More specifically, whereas earlier conceptual changes are universal and spontaneous, later changes in naive theories which occur by incorporating scientific concepts and correcting ‘misconceptions’ are difficult to induce. Clement (1982), for example, reported that even after taking a mechanics course, more than 70% of the college students who majored in engineering continued to give erroneous answers for everyday mechanics problems (e.g., what direction of force acts on a coin that is tossed straight up into the air?); they still maintained the misconception, ‘motion implies a force.’

Constraints in Conceptual Development

Innate constraints, which can best be described as preferences and biases, are presumed to play a significant role, especially in the initial acquisition of naive theories, whereas prior knowledge as constraints operate to a greater extent later in conceptual development. Innate constraints should serve to direct attention, and restrict the range of hypothesis space to be searched. A unique set of constraints in each privileged domain draws attention to relevant aspects of the target objects or phenomena so that even young children can distinguish those which should be interpreted within the domain from those which should not. In the case of naive biology, for example, there seems to be a set of constraints that direct attention to those aspects of living things that distinguish them from non-living things, such as possessing the capacity for self-initiated movement (Gelman, 1990).

Another set of innate constraints serves to eliminate in advance a large number of logically possible interpretations or hypotheses. Again, an example from naive biology might clarify this point. As demonstrated by Garcia (1981), even rats tend to attribute their bodily disturbance to eating a novel food. It seems likely that humans are endowed with similar constraints—trying to find a cause among a variety of foods for diarrhea, or seeking a physical cause for a cut. In other words, it is assumed that, because those tendencies and biases in the search enable humans to explore possible interpretations highly selectively, they can reach, in most cases, a reasonable interpretation promptly, and thus can accumulate pieces of knowledge constituting a core domain of thought.

How innate constraints are represented in the brain is yet to be known, but recent studies on autism suggest that naive psychology may have a dedicated cognitive mechanism, which autistic individuals lack. In fact, autistic individuals reveal a clear dissociation between psychological and physical understanding. For example, Baron-Cohen (1995) reported that they show disproportionately lower performance on the theory of mind task, such as a false belief task, while their performance concerning physical causation is often excellent. The false belief task was designed to examine whether children can understand that a person’ behavior is caused by his or her false belief, and normal children 4 years of age and older succeed in this task without difficulty (e.g., Wimmer & Perner, 1983). Moreover, it has been pointed out that, even if some autistic individuals acquire theory of mind, they do so with great difficulty (see Section 10.5).

The notion of innate constraints does not exclude roles of experience, however. Peterson and Siegal (1997) found that deaf children raised by hearing parents, and thus without fluent signing conversational interaction, had selective deficits in naive-psychological reasoning, whereas deaf children raised in families with a fluent signer did not show such deficits. This suggests that conversational experience with a significant other contributes to the development of naive psychological understanding.

To generalize, even in privileged domains, knowledge is acquired through problem solving and comprehension within each domain, but that process is greatly enhanced by innate constraints.

Expertise

As mentioned earlier, there are a large number of domains in which people can gain expertise. These domains vary with respect to several important dimensions: some are knowledge-rich, and others are knowledge-poor; some require speed, while others do not; novel problems are continuously given in some, whereas the same set of problems is repeatedly presented in others; and some involve interpersonal competition or collaboration, while others are solitary. What is required to be qualified as an expert tends to vary accordingly. Therefore, the only characterization that is applied to all domains of expertise, in addition to the general features of knowledge acquisition mentioned at the beginning of this section, is as follows. Experts have acquired their domain-specific competence through experience in the domain over thousands of hours, which often takes the form of training called ‘deliberate practice’ (Ericsson, Krampe, & Tesch-Römer, 1993). In contrast to the early and easy acquisition of knowledge in the privileged domains, the process of gaining expertise in non-privileged domains is time-consuming and requires much effort, mainly because the latter is not helped by innate constraints, and this makes knowledge acquisition much harder, especially at the initial phase (see below).

Let us describe a few characteristics of experts, more specifically, their patterns of competence and kinds of underlying knowledge, observed in a number of domains. A cautionary note is needed here. In knowledge-rich domains experts are characterized primarily by their exceptional abilities for solving problems and understanding the relevant entities in the domain, but we will not discuss these aspects here, because they will be reviewed in Sections 10.3 and 10.4. All of the studies reported below used memory or simple performance tasks.

Knowledge Organized into Perceptual Chunks

Experts not only possess a great amount of domain knowledge but also have the knowledge organized into useful chunks (higher-order units involving a number of individual pieces) that can be triggered readily by perceptual cues. This finding was obtained first in the domain of chess. Chase and Simon (1973) demonstrated that chess experts recalled chessboard position patterns taken from actual games, which were presented for 5 s in each trial, more accurately and in a smaller number of trials than novices, although the difference disappeared for random board patterns. Moreover, the experts reported a cluster containing more pieces in the pattern than the novices. In other words, the experts’ chunks were larger than the novices.’ This perceptual superiority that was found among expert chess players has been replicated in other domains, such as in the game of Go, in reading circuit diagrams, in memorizing melodies, and in interpreting X-ray plates.

Knowledge as chunks cannot be applied across domain boundaries. Good evidence for this was provided by Oura and Hatano (1988). Musically experienced college students and elementary school children, both of whom had had about 5 years of piano training in classic music, and inexperienced college students, were presented a melody (an unfamiliar commercial song having clear tonality) auditorily and were asked to reproduce it. The results indicated that the musically experienced students were superior to the inexperienced ones, regardless of age, in the speed of acquisition as well as the eventual level of mastery of the melody. Two types of control experiments revealed that the superiority of the experienced students was not because they had a better memory in general. When asked to remember a short poem, the college students, regardless of their music experience, learned it faster than the children. Moreover, if the presented melody was a non-tonal (modal) Japanese folk song, the superiority of the musically experienced students disappeared; all three groups showed a comparably poor performance.

Knowledge Needed to Connect Components Meaningfully

Mckeithen, Reitman, Rueter, and Hirtle (1981), using computer programmers differing in skill levels (experts, intermediates, and beginners), examined the details of knowledge organization that these programmers possessed. The participants were presented with a computer program in either a normal or scrambled form for five 2-min study trials, and immediately after each trial, they were asked to write as much of the program as they could recall on a blank sheet. The results indicated that for the normal version of the program, the experts recalled the largest amount correctly, the intermediates could recall less, and the beginners recalled the least, but these differences disappeared for the scrambled version. The skill-related differences in correct recall for the normal version increased in later trials, which was different from similar studies in other domains, such as chess or Go, in which the experts’ superiority was revealed only at the initial trials. This was because, the authors interpret, the programmers needed several trials to understand the functions and structures of the program. This finding implies that programmers’ knowledge does not consist of perceptual chunks. Instead, it is to organize programming components meaningfully, which is similar to what enables us to comprehend discourse (see Section 10.4).

In another experiment the participants were presented with a series of words taken from a program and written on cards (one word per card), and asked to learn at their own pace so that they could recall them later. The recall orders for each subject were analyzed and the details of individual programmers’ clusters of key programming concepts were inferred. The beginners, whose programming knowledge was poor, seemed to use very general mnemonic techniques for memorizing and recalling the words; their patterns of recall showed a variety of common-language associations to the programming concepts, such as clusters of words with common first letters and lengths, or words alphabetized by first letters. In contrast, the experts recalled key programming language words in meaningful clusters; their organizations contained words grouped according to their function in the program. The intermediate programmers showed mixtures of programming and common-language associations.

Speed, Accuracy and Automaticity

It is not surprising that experts can solve problems and perform tasks in the target domain more quickly and accurately than novices, but the expert—novice difference is sometimes impressive. For example, abacus learners acquire, as they gain expertise, fast, accurate, and automated skills for operating the instrument (Hatano, 1997 for a review). Because people can learn to operate an abacus in a few hours, subsequent training is geared almost entirely to accelerating the speed of the operation. As a result of extensive training, abacus operation tends to be gradually interiorized to such a degree that most abacus masters can solve calculation problems by mental simulation. Because such mental operation is no longer constrained by the speed of muscle movement, it is much quicker than operating a real abacus.

Experts’ mental calculation, unlike real abacus operation, is not entirely free from errors, but the accuracy is still at a respectable level. It is reported that a 4th-grade (10-year-old) girl, who became a junior national champion, could solve 30 printed multiplication problems, 3 digits by 3 digits (e.g., 148×395) or 4 by 2 (3519×42), in 58 s. This was surprising, but her net calculation time was even shorter – she needed this amount of time for writing the answers down. Thus, she could solve the same number of problems requiring the inverse operation (e.g., 277836 ÷ 78) in 31 s, because the number of digits in the answers was nearly a half. She made just one error on these 60 problems.

Abacus experts’ calculation is highly automatic. Experienced abacus operators can converse during calculation, even without the instrument. Needless to say, the conversation cannot be very serious – usually just a short and simple factual or preferential question-answer exchange, but this is still impressive, because most ordinary people cannot speak even a word while performing a mental calculation. Therefore, abacus experts are characterized in terms of three types of knowledge: (a) a cluster of specific condition-action rules for adding and subtracting numbers that enable them to run the four operations very fast, (b) a mental model of an abacus for mental abacus operation, and (c) metacognitive knowledge that monitors the execution of operations with a minimal processing load. Evidence suggests that in general, experts in operating devices or instruments seem to develop similar types of knowledge.

Constraints in Expertise

Unlike in conceptual development, innate constraints do not work in gaining expertise. Instead, prior knowledge as cognitive constraints plays an important role. The acquisition of knowledge proceeds slowly at the early stages of expertise. After a certain amount of domain-specific knowledge accumulates, the acquisition of new pieces of knowledge is accelerated, because prior domain-specific knowledge starts to work as constraints. Domain-specific knowledge serves to direct one’ attention to relevant aspects. Moreover, it serves to eliminate in advance a large number of logically possible interpretations. Therefore, such domain-specific knowledge not only helps problem-solving and understanding, but also enhances the acquisition of new pieces of knowledge in the domain.

Problem Solving

This section is concerned with the solution of problems that we do not know how to solve. Our lives are full of problems in the wide sense, that is, we continuously strive to produce changes in the environment or in ourselves, but many of these challenges are not taken as problems in the real sense, because we already know how to solve them; executing a solution routine usually requires some effort, but no search. For example, washing one’ face or cleaning one’ teeth may be regarded as problem solving only when the water supply is suspended or toothpaste is unavailable.

To study problem solving experimentally, we need a problem that the participant, most often a college student, does not know how to solve. Moreover, for the experimenter’ convenience, it is desirable that the problem can be solved in most cases in a short period of time. Thus, earlier studies on problem solving often used puzzles.

Process of Problem Solving

Problem Representation

Problems must be represented before their solution processes start. According to Newell and Simon (1972), the representation of a problem includes four components: (a) the initial state to start with, (b) the goal state to be reached, (c) a set of operators or actions that can be taken to change the current state, and (d) path constraints that restrict the range of successful paths to the solution. With these four components, puzzles such as the Tower of Hanoi (a favorite task among cognitive psychologists) can be neatly defined (assuming that there are tests to determine whether the goal has been achieved without violating the constraints). Let us take the three-disk version of the Tower of Hanoi as an example: there are three pegs, A, B, and C on a board; the initial state has all three disks of decreasing size on peg A, and the goal state is to have all three on peg C. This puzzle’ only operation is to move disks, with two constraints: (a) only one disk can be moved at a time, and (b) a smaller disk must always rest on a larger disk when two or more disks are placed on the same peg. The search involves finding a path from the initial to the goal state in a problem space that includes 27 possible states. Because the components of this problem representation are clearly defined, the Tower of Hanoi and similar puzzles are often called well-defined problems.

Other problems are ill-defined. For instance, writing an essay is a typical ill-defined problem, because in this case it is not clear what the goal state is, what operators are available, etc. Still other problems fall in between. For example, many of the textbook problems in such subjects as mathematics and physics are well-defined for experts, but not so for students, because students may not have access to all possible operators, or know what constraints are imposed. As we will see below, these problems in knowledge-rich domains may be represented differently according to the solvers’ domain knowledge.

Search Strategies

Problem solving can be defined as the search in a problem space (consisting of all states that can potentially be reached by applying available operators). The problem space may be too large for an exhaustive or ‘brute-force’ search, because its size increases exponentially with the number of steps needed to reach the goal state. Chess has a problem space of about 1020 states; other games (e.g., Go) have an even larger problem space. Thus, human problem solvers often rely on a heuristic search, trying to find a reasonably good solution with a reasonable amount of time and effort.

Heuristic search often presupposes domain-specific knowledge relevant to the target problem (e.g., knowledge that suggests a promising part of the problem space). However, as proposed by Newell and Simon (1972), there are several general heuristic search methods, among which means—ends analysis is the best known. This involves searching for an operator that effectively reduces the difference between the goal state and the current state. When no such operators can be applied, it tries to create the needed condition for the application of one of them. Means—ends analysis is usually much more efficient than relying on either a forward or backward search alone.

Decomposition and Planning

Complex problems that we often encounter in our daily lives can be divided into relatively independent subproblems. We can find a solution for the entire problem much more efficiently if we can aptly decompose the problem. For example, the problem of reducing one’ weight can effectively be solved by eating food containing fewer calories and burning more calories through physical exercise. As this example suggests, another way to accelerate problem solving is by planning before executing the steps of a solution. It would be almost inconceivable to build a house without drawing its blueprint or write a novel without outlining its plot beforehand.

How Knowledge Operates in Problem Solving

Puzzles as typical examples of well-defined problems can be solved without prior knowledge (except for understanding the instructions), although experiences of solving the same or similar puzzles may make finding the solution more efficient. Interestingly, puzzles often require the solver to perform a move that seems to increase the distance from the goal state.

Most other problems in our lives, many of which are ill-defined, presuppose prior domain-specific knowledge to solve them successfully. Because the acquisition of such prior knowledge takes a long time, studies examining whether such knowledge makes a difference in problem solving have to compare experts and novices. Thus research on problem solving (as well as on comprehension) is often concerned with the expert—novice difference and expertise.

It is not easy to choose appropriate problems both for novices and experts; good problems for novices may not be problems for experts, because experts often possess ‘problem schemas’ by which they can recognize classes of problems and apply ready-made routines. However, some compromise may be possible. When the same problems are presented, experts are less likely than novices to make errors or take unnecessary steps, but a more interesting finding is that experts and novices use different heuristics in solving problems.

Simon and Simon (1978) compared in detail the processes of solving simple kinematics problems by an experienced and a less-experienced participant. These participants were asked, for example, to find the average speed of a bullet within a gun barrel half a meter long, assuming that it accelerates uniformly and leaves the muzzle at a speed of 400 m/s, how long the bullet remains in the gun after it is fired, etc. Both participants evoked an equation, inserted the known quantities into the equation, and found an unknown quantity by solving the equation. This cycle might be applied repetitively or recursively. The authors found that the novice searched for an equation that contained the quantity sought. If the equation could not be applied, then he turned his attention to one of the unknowns in that equation and proceeded to look for an equation that contained the new target. In other words, the novice relied on means—ends analysis or backward chaining. In contrast, the expert searched for an equation that would immediately yield a new quantity, and continued to test equations until the quantity sought was generated. In short, the expert used forward-chaining heuristics. The authors interpreted these results to mean that the expert’ apparently more primitive approach was based on his physical intuition, that is, his ability to construct a mental representation of the situation in which the components described in the problems were causally connected. If the forward generation of new quantities is totally unguided, it must lead the problem solver to unnecessary steps or blind alleys. Therefore, experts must be, though unconsciously, guided by the quantity sought.

Moreover, even when the same problem is given, experts and novices may represent it differently. Chi, Feltovich, and Glaser (1981) indicate that experts, unlike novices, categorize physics problems by the laws of physics. These categories elicit relevant knowledge structures to represent the problems, and these structures indicate potentially useful solution methods.

Oura and Hatano (1998) indicate that novices and junior experts represent the task of playing a given piece of music quite differently. For novices, the problem is to shift from a slow and error-prone performance to a smooth and accurate one. Thus, they adopt repetitive exercise as an action that will probably lead them to the goal state. In contrast, for junior experts, the problem is to transform their structural and hierarchical understanding of a piece into an actual performance that conveys this understanding to an audience. For this purpose, they search among a variety of dynamic, temporary, and phrasing parameters. In fact, these authors found that a junior expert player, in her exercise, kept trying to coordinate two perspectives in her mind, that is, the perspective of herself as the player and the perspective of listeners. When she was allowed as much practice as she wanted, she repeated the piece 45 times before finishing, during which the tempo of the performance was slightly varied from one rendition to another.

How Knowledge is Acquired in Problem Solving

As indicated in Section 10.2, gaining expertise often takes years. This implies that solvers cannot acquire much domain knowledge by solving a single problem. However, it has been shown that repeatedly solving well-defined problems like puzzles enables solvers to find the solution path promptly and to acquire strategies that can be applied to similar problems. Such learning or knowledge acquisition probably takes the form of combining old condition—action rules to produce a more efficient new rule, replacing a general rule that requires computation by a set of specific rules, and/or adding a new rule that makes the solution process better guided (Anderson, 1983; Anzai, 1987). For example, if there are two condition—action rules for transforming state A into state B, and state B into state C, a new, combined rule transforms state A directly into state C (called composition). Solvers may find bad moves that lead them to a blind alley or to a previous state, and also recognize good moves that lead closer to the goal state, avoiding the bad moves. New condition—action rules avoiding bad moves and choosing good moves are thus added. All these condition—action rules enable solvers to find the correct solution promptly for the type of problems they have solved in the past using less efficient procedures, but these rules cannot be applied to other types of problems, because they are situation-specific. In other words, through solving the same kinds of problems repeatedly, solvers become routine experts, that is, experts who are distinguished in terms of their speed, accuracy, and automaticity.

Solvers acquire richer knowledge through their attempts to solve problems in knowledge-rich domains. For example, solvers are likely to acquire knowledge that enables them to represent problems aptly and to apply an appropriate solution routine (e.g., a law in physics). Solvers may learn about goal states, operators, and path constraints by solving an ill-defined problems (e.g., writing a novel). And through the solution of a series of novel problems, solvers may learn about the target entities that are dealt with in problem solving (e.g., raising animals may lead to a more accurate and elaborate concept or schema of what the raised animals are like). As a result, they may become able to build mental models of the target when needed, in other words, to gain adaptive expertise. Unlike routine experts, adaptive experts are expected to be able to solve novel problems in the domain, because they have grasped the meaning of solution routines in terms of changes in the target entities (Hatano & Inagaki, 1986).

Analogy and Transfer

As mentioned above, domain-specific knowledge is critical in solving problems; solvers represent a problem and search for a solution using this knowledge. Does this mean that almost no transfer occurs across domains, in other words, that the knowledge gained in one domain is not used in solving problems in other domains? This question has been examined under the name of analogical transfer, because transfer across domains requires the recognition of relational or structural similarity between the original and the target problems, which is the essence of analogy (Gentner & Markman, 1997). If the problem spaces of the target and the source are mapped, one can choose a path for the target that is structurally similar to the successful path in the source.

Gick and Holyoak (1980) examined whether college students could solve a problem through cross-domain transfer. The transfer problem to be solved was what is known as the ‘tumor problem.’ This problem concerns a doctor who has to figure out how to use radiation to destroy a tumor deep inside the body without injuring the patient’ healthy tissues in the process. Because if rays are used that are intense enough to destroy the tumor, they will, apparently necessarily, injure the healthy tissues, this problem is known to be difficult for many people. A solution that can be used in this and similar cases is to give low-intensity rays from different directions at the same time so that they can converge on the tumor.

When Gick and Holyoak presented this tumor problem without any preceding problem, only about 10% of college students successfully solved it. However, of those who had previously read another story (called a ‘fortress story’) and were instructed to make use of it, 75% could find a satisfactory solution. The fortress story is apparently not similar to the tumor problem at all. In it, a general is trying to capture a fortress and needs to get his army to the fortress at full strength. Since the entire army cannot pass safely along any single road, the general sends his soldiers in small groups through several roads simultaneously, so that they can arrive at the fortress at the same time. What is common between the two situations is that it is necessary to divide the whole into smaller components so that each component can go through, and to recombine the components so that the whole can be reconstructed at the critical place.

Interestingly, when another group of college students read the fortress story but were not told to make use of it, only 20% could find a satisfactory solution for the tumor problem. Gick and Holyoak interpreted this result as consistent with their model of analogy; that is, before two domains are mapped, an appropriate source domain must be selected, and this process is the hardest one in analogy. Although creative human minds often try to apply to a novel problem seemingly relevant knowledge from another domain, finding the really applicable knowledge in other domains remote from the target appears to be very difficult.

However, in everyday life we do not choose a domain that best fits the problem in question. Rather, we examine only a very limited number of familiar domains as potentially useful source analogs. Because of this, even children can sometimes make use of analogies. In other words, we may have underestimated people’ ability to transfer by focusing on laboratory studies in which a stringent criterion of successful transfer is adopted.

Comprehension

Comprehension or understanding (terms we use interchangeably) are defined differently by different authors, but here we adopt the following, comprehensive definition: comprehension is a process through which a coherent and plausible interpretation is adopted for the pieces of information gathered from both outside (through perception) and inside (through memory) about a target object or phenomenon. The above definition is based on Piaget (1978), who defined the term as the solution of the problem of ‘the “how” and “why” of the connections observed and applied in action,’ but includes implicit forms of understanding. In other words, we assume that a coherent and plausible interpretation produces differentiated and appropriate predictions about the target but may not always lead to a verbal explanation for them. Our definition is also based on Kintsch’ (1998) conceptualization, but unlike him, we assume that a process of coherence-seeking or multiple constraint-satisfaction can involve mechanisms other than spreading activation, that is automatic and effortless.

Understanding enables us to build enriched, stable, coherent, and usable representations of the world, which often serve as the basis for solving problems, that is, for changing the world as one desires. In this sense, understanding is an investment in an unknown future, and the human tendency to seek understanding can be considered adaptive. Thus the coherence-seeking process may take the form of a comprehension activity that is necessarily a time-consuming and effort-intensive process (Hatano, 1998). Comprehension activity often offers multiple interpretations at one time, and compares their plausibility carefully by deriving and testing predictions from each of the interpretations being considered. It is not a process in which one holds onto the first idea that comes to mind or one waits for a single interpretation to emerge through spreading activation. Comprehenders may feel uncertainty or puzzlement. Scientific inquiry is prototypical of this activity (e.g., Dunbar, 1995).

Categorization as Understanding

Recognizing something that is ambiguous as an instance of a concept or schema constitutes understanding, because it is adopting a coherent and plausible interpretation of a set of observed features of the target based on relevant concepts or schemas. Once we understand the target object or phenomenon, we can make sense of the observations and can remember them much more easily (e.g., Bransford & Johnson, 1972). Moreover, we can decide reasonably how to treat the target.

The diagnosis of disease by reading X-ray films is a good example of such tasks of categorization or identification. The same disease does not always produce the same appearance on the film, and there can be several diseases that have similar appearances. Therefore, reading X-ray films involves many inferences and requires a constraint-satisfaction process to reach a diagnosis. This is undoubtedly a process heavily dependent on prior knowledge, in which experts and novices would be expected to differ.

Lesgold et al. (1988) compared the ability to make accurate diagnosis by reading X-ray films of expert radiologists with novice ones. The participants were asked to report a diagnosis after being shown chest X-ray films for 2 s, and their protocols were analyzed. It was found that compared with the novices, the experts reported more varied findings (i.e., attributed specific properties to the film or the patient), their reasoning chains were longer, and a greater number of their findings were connected to at least another finding. The authors interpreted these results as follows: ‘The data support a view of the expert as doing more inferential thinking and ending up with a more coherent model of the patient shown in the film’ (p. 317).

However, experts and novices do not make the same observations and then interpret them differently based on their different amounts of knowledge in natural settings. In a further study the participants were required to examine the film as long as desired, while thinking aloud, and to dictate their formal diagnostic reports. Qualitative analysis of the protocols revealed that the experts exhibited more flexibility and tuned better their schemas in the perception of features in the film to the specific case with which they were working. They could discount most of the obvious film features that the novices were likely to be misled to judge as evidence of disease symptoms, by using a mixture of technical knowledge about how the films were made and a better-developed perception of the features.

Diagnosing films apparently requires simple outputs, i.e., whether there are pathological changes, and what kind of disease the changes are due to. However, the process of understanding is more dynamic and interactive than the top-down process of applying the relevant schema stored in memory or the bottom-up process of summarizing and abstracting collected pieces of information. Perception, inference, and memory retrieval are inextricably connected. We assume that this is true for other processes of understanding.

Discourse Comprehension

Comprehension of discourse (including both narrative and expository text) is one of the most popular topics in the study of higher-order cognition. Here, the input information is in the form of a spoken or written sentence, and the compre-hender’ job is to form a coherent whole from individual sentences. Versions of schema theory (e.g., Schank & Abelson, 1977) were predominant in earlier studies. Schemas, each describing a sequence of actions or a set of constituent elements (e.g., a restaurant schema composed of the actions of being seated, ordering, being served, and paying) were supposed to be triggered by key words called headers (e.g., dinner) and then to control the comprehension process in a top-down fashion. However, as pointed out by Kintsch (1998), such top-down views have limitations; discourse comprehension is basically bottom-up, and it is much more flexible and context-sensitive than predicted from the schema theory.

Most contemporary investigators assume that to understand a story (i.e., a narrative text) is to build a micro-world in which these events described in it are likely to occur. Kintsch (1998) calls this micro-world the ‘situation model.’ In addition to the textbase that consists of those elements and relations that are explicitly described in the text itself, the situation model involves pieces of information that make the textbase coherent, complete, and interpretable in relation to prior knowledge. These pieces of information are inferred from both the text and prior knowledge. Multiple situation models may be constructed from one and the same text (e.g., differing in the coverage or the mode of representation).

When people understand a story, they will make inferences that are not given in the text. They sometimes infer that an event occurred, even though it is not explicitly written (e.g., from the sentence, ‘He took a train for Tokyo,’ the reader infers that the subject bought a train ticket); other times, the reader infers how an event occurred, although the text does not give any concrete details. (From the sentence, ‘He bought a ticket with coins,’ one infers, ‘He used a vending machine.’) Still other times, readers connect two adjacent propositions. (From the two sentences, ‘Taro wanted to buy a CD player at a discounted price. He went to a shop in Akihabara,’ one infers, ‘Taro went to Akihabara to buy a CD player there.’) How many inferences are spontaneously generated may depend on the readers, texts, and modes of reading. Graesser, Singer, and Trabasso (1994) claim that the inferences that are needed to explain why given events occur and to establish coherence of the text tend to be induced spontaneously as the text is being processed. Kintsch (1998) proposes that a coherent situation model is built based on these presented and inferred propositions through constraint-satisfaction, more specifically, spreading activation within the network of interrelated propositions.

Local situation models for a short sequence of events are further integrated into a global model in seeking overall coherence or a good story. This step may involve intentional processing. Let us take Spiro’ classic experiment (1977) as an example. Participating students in his experiment read the following story: Bob and Maggie were an engaged couple. Maggie wanted to have babies, whereas Bob did not. They had a serious dispute about having a baby. When the students read it, they would predict, based on their prior knowledge, that the couple’ marriage would fail. However, the experimenter told the participants that the story was true and Bob and Maggie were married and living happily. This information was inconsistent with the read story, and thus the students often tried to recover coherence by generating additional information that could reduce the incongruity they sensed; for example, the couple reached an agreement after a long discussion, the couple agreed to have just one baby, and so on. In fact, when the participants were asked to recall what had been explicitly indicated in the story, they tended to insert such additional pieces of information, especially long after the presentation of the story. Such an apparent inconsistency is seldom included in a well-organized actual text, but studies using experimentally designed defective texts (e.g., in Collins, Brown, & Larkin, 1980) have revealed that text comprehension may require a number of effortful attempts to instantiate, coordinate, and even insert pieces of information.

Almost all people are very experienced in discourse comprehension because it is a major medium of human communication. However, knowledge about the topic of the discourse still makes a difference, especially in the generation of inferences. As a result, more knowledgeable people can build a richer situation model than less knowledgeable ones. They learn more, especially when the text is less coherent and thus requires comprehension activity on the part of comprehenders. Schneider, Korkel, and Weinert (1989) demonstrated that students who knew a lot about the soccer game not only remembered details of a given story about soccer better, but also made more inferences and recognized contradictions in the text more often than their contemporaries who knew little about soccer.

Comprehension of Procedures and Targets

Just knowing the procedures to achieve goals differs from knowing how and why they work. By possessing knowledge to build mental models of the procedure—target combination (what the target is like and how it is changed by various procedures), we can grasp how and why these procedures work, modify the procedures flexibly, and even invent new procedures. Building such models may be through a process of comprehension.

Kieras and Bovair (1984) showed that having a ‘device model’ (a mechanical version of the procedure—target model) that describes the internal mechanism of a piece of equipment enhances not only the participants’ learning how to operate the equipment but also their inventing more efficient procedures than the ones they have been taught. Their first experiment compared two groups, i.e., a model group and a rote group. The participants in the model group were presented with the device model in a fantasy context before being given procedural instruction, whereas those of the rote group learned the identical operating procedures of the equipment by rote. It was observed that the model group learned the procedures faster, retained them more accurately, and executed them more efficiently than the rote group. More interestingly, the model group could simplify inefficient procedures much more often than the rote group, indicating the benefit of their deeper understanding. The second experiment showed that the model group could infer the procedures much more easily than the rote group, again showing the significance of learning with understanding. The final experiment showed that what was critical was the knowledge about the internal structures of the components of the equipment, not the context of fantasy.

In the experiments by Kieras and Bovair (1984) described above, an externalized form of the standard device model was presented and explained. Active human minds often ask the questions ‘how’ and ‘why,’ even when they know that a given procedure works, and may construct the procedure—target model through comprehension activity. For example, the recipe for bonito sashimi sometimes induces comprehension activity, because it is not just the slicing of raw fish, but involves steps of roasting the skin-covered surface quickly at a high temperature and putting the roasted side into ice-water (Hatano, 1998). People may wonder why such a recipe is required, and if they comprehend it, they can modify it flexibly when they have to meet a different set of constraints, e.g., when there is no ice or no source of searing heat. To understand a given procedure is, as in this example, to find a plausible interpretation for how and why it works, in other words, what kind of effect each step of the procedure has on the target entity of the procedure.

We sometimes engage in an even more ambitious attempt to reconstruct a procedure from its product in our daily lives, which is similar to the scientific inquiry that tries to model a process from its observed outcomes. For example, we may want to reproduce an unfamiliar but tasty dish (Hatano, 1998). In an experiment, college students who were given a slice of perch mousse and asked to find how it had been made by tasting it (more specifically, to indicate its recipe) seemed to engage in comprehension activity. They often proposed ideas that they eventually discarded themselves. For example, they offered a number of possible ingredients, but rejected some of them both from their taste and knowledge about cooking. They expressed their feelings of uncertainty or puzzlement, as well as insight.

Although there have been few studies, prior knowledge seems critical for the understanding of procedures. First, prior knowledge enables comprehenders to think of hypotheses. Second, it serves as constraints in the pursuit of coherence and plausibility.

Understanding by Analogy

Plausible interpretations can be found in a variety of ways. When people encounter a novel object or phenomenon, they may try to understand it by likening it to a highly familiar object or phenomenon. For example, at the time of Iraq’ invasion into Kuwait, many Americans (partly through the campaign by the government) likened Saddam Hussein to Hitler, and thus justified their bombing of Iraq. Interestingly, some objected to the military action by analogizing the Persian Gulf war to the Vietnam war. This illustrates how people try to understand a novel phenomenon by referring to their past experience that they think is most similar, and how their reactions to the phenomenon are mediated by their analogy (Holyoak & Thagard, 1995).

Using analogy, humans map their knowledge about the source to the present case (target), so that they can make a coherent interpretation of the set of observations for the target. Mapping involves search and thus goes beyond the retrieval of prior knowledge, often resulting in the production of new pieces of knowledge. In this sense, analogies are the key to creativity; they are often very useful in everyday cognition (Hatano, 1998) as well as in scientific reasoning (Dunbar, 1995) to generate hypotheses.

It should be noted, however, that this enterprise of understanding by analogy is entirely dependent on the analogist’ possessing highly familiar source that is subjectively similar. In children, because such familiar sources are limited, we often observe their heavy reliance on a few particular sources. Anthropomorphism or personification can be regarded as one such attempt. Although young children are able to classify entities into ontological categories, they apply their knowledge about humans to other animate objects or even to inanimate objects, when they have to infer or interpret an object’ unknown attributes or reactions (Carey, 1985; Hatano & Inagaki, 1996).

The analogies young children make may involve structurally inaccurate mapping (e.g., mapping the relationship between humans and food to that between plants and water), and induce biased reasoning (e.g., neglecting the roles of nutrients in the soil and photosynthesis). Young children may carry analogy beyond its proper limits and produce false inferences, as revealed in typical examples of animistic reasoning. However, they can generate ‘educated guesses’ by analogies, relying on their only familiar source analog of a person (Holyoak & Thagard, 1995), and animistic errors and/or overattribution of human characteristics to nonhuman animate objects should be regarded as negative byproducts unluckily produced by this process of reasoning. Even a highly productive scientist like Kepler was misled by analogies at the beginning of his career (Gentner et al., 1997).

Comprehension Monitoring

To comprehend well, it is necessary to assess one’ extent of understanding, and to take appropriate action based on that assessment. However, whether one understands is not obvious (whereas whether one has solved a problem can be recognized more clearly). To put it differently, every task of comprehension is inevitably ill-defined. Therefore, how well people can monitor their own understanding and how this monitoring ability develops are interesting questions in the study of comprehension.

Research on comprehension monitoring has shown that younger children tend to miss the insufficiency or inconsistency of a given message more often than older children or adults, but another line of research on metacomprehension has revealed that even college students tend to have this ‘illusion of comprehension’ (Glenberg & Epstein, 1985). College students often believe that they understand a given text, though in fact they do not, at least as assessed by a multiple choice test.

As suggested by Schneider et al. (1989), it is likely that the more and less knowledgeable people differ in their ability to accurately estimate the degree of coherence among pieces of information and the adequacy of their comprehension. Humans seem to have a general tendency to make sense of their observations and actions, in other words, to derive coherent and plausible interpretations. However, judging whether the achieved understanding is adequate is a totally different issue. The criterion of adequacy may vary from context to context. Therefore, only those who have performed many similar ill-defined tasks of understanding in the target domain can assess it accurately.

The Brain and Higher-Order Cognition

Unlike with vision and motion, there have been very few successful studies connecting a specific aspect of higher-order cognition to specific areas of brain activity until quite recently. This was because the major experimental findings of brain science came from studies of monkeys and other nonhuman mammals, whose higher-order cognition, if any, is very limited. These experimental findings were supplemented by clinical reports on human patients who suffered from local brain damage. However, such damage is seldom restricted to a particular area of the brain. Moreover, as time passes, the damage may be compensated for by newly formed connections that are not available in the normal brain.

New, non-invasive imaging technologies, such as PET and fMRI, have started to change this situation, because we can now observe patterns of activation in the normal brain as a person performs specified experimental cognitive tasks. Although it is highly unlikely for us to find any dedicated cognitive mechanisms in the brain for problem solving or comprehension—because higher-order cognition is a compound process consisting of a number of constituent sub-processes—we may be able to identify those areas that are critically important for specific types of problem solving or comprehension. These technologies provide a powerful means to investigate higher-order cognition, if they are skillfully combined with the manipulation of stimuli and the assessment of behavioral responses that psychological studies have developed.

However, it should be noted that the current imaging technologies still have limitations for investigating higher-order cognition. First, whereas their spatial resolution is high (they can locate the areas of activation with an error of millimeters), their temporal resolution is not high enough to specify the time course of rapid mental processes. This is especially true for PET; presently we can measure patterns of activation by PET only after a participant has been performing the same type of task for a few minutes. Researchers of higher-order cognition need more advanced technologies with high temporal resolution than are currently available, such as event-related fMRI. Second, because so many areas in the brain are usually activated by a complex cognitive task, it is necessary to set up a control condition, and the unique pattern of activation for a specific task is estimated by subtracting the pattern for the control task from the observed pattern for the target task. However, we do not have a good rationale for choosing the appropriate control condition, in other words, to equate other parameters than the critical one between the target and control tasks, either for problem solving or comprehension. For example, since almost every task of higherorder cognition requires a high information-processing load, its control task must also be cognitively demanding, but we do not yet know how best to choose such a control task.

Executive Control Function of the Frontal Lobe

One form of thinking can be regarded as interiorized action. More specifically, instead of exerting a series of motor actions on physical objects or moving one’ own body parts in sequence, it is mentally operating on and changing mental representations of the targets. In other words, this form of thinking is controlled manipulation of mental models (i.e., mental representations that can be modified in response to mental actions). Many investigators assume that such thinking has the control devices and targets of control, and that the frontal lobe operates as the control system, which regulates the condition—action rules, concepts or schemas, and mental models stored in the parietal and temporal lobes (e.g., Holyoak, 1990).

The frontal lobe plays an important role in a variety of cognitive functions including thinking. Cabeza and Nyberg (1997) conclude, based on their review of PET studies of higher-order cognitive processes (although very few of them dealt with thinking), that the frontal lobe was involved ‘in almost all cognitive processes’ (except for bottom-up perceptual processes). This is probably due to its assumed role in the executive control of cognition. Stuss and Benson (1986), based on an extensive review of frontal lesion studies, propose several functions of the frontal lobe, such as planning and ordering sequential behavior, establishing, maintaining, and changing a mental set, and monitoring one’ behavior. As Benson (1994) declares, reported clinical cases clearly indicate that damage to the prefrontal association cortex impairs the executive control of higher mental activities, although it seldom affects intelligence measured by conventional tests. It seems reasonable to assume that the frontal lobe is the control system for thinking defined as the goal-directed manipulations of mental representations.

It is far less certain where rules, concepts and other forms of elaborate knowledge are stored or how they are represented in the brain. However, imaging studies have begun to offer some interesting observations.

Category-Specific Imapairment and Differential Activation

According to clinical reports on human patients who have had a focal brain lesion, there are cases in which performances on highly similar tasks are differentially damaged. Some of these reports are highly relevant here, especially when combined with imaging studies revealing differential activation. Although the experimental tasks used were as simple as identification or naming tasks, the results suggest that knowledge about different types of entities—e.g., living things (represented by animals) and nonliving things (represented by human-made physical tools)—are stored in different locations in the brain. Such results reinforce views that entities belonging to different ontological categories (humans, nonhuman animals, plants and nonliving things), which are considered to be distinguished based upon schemes of mind (Atran, 1998), have specialized neural mechanisms acquired through evolution (Caramazza & Shelton, 1998).

Damasio, Grabowski, Tranel, Hichwa, and Damasio (1996) made one such pioneering attempt to combine clinical observation and imaging to examine whether knowledge about different classes of entities can be segregated. They prepared three naming tasks for different categories of entities, that is, for proper names of famous persons, basic category names of animals, and names of tools. They first administered these tasks to a large number of adults who had single and stable lesions in various parts of the brain. They found that all but one of the 30 participants who showed poor naming performances had damage in the left hemisphere. More interestingly, they found that poor naming performances for the three categories were associated with damage in different areas of the brain: poor identification of person names was associated with damage in the left temporal pole (TP), animal names with that in the left inferotemporal (IT) (mostly anterior), and tool names with that in the posterolateral IT. None of the participants who showed normal naming performances had damage in these areas.

They then gave the same three naming tasks to normal right-handed participants while their regional cerebral blood flow (rCBF) was measured by PET. The control task was to say ‘up’ or ‘down’ when presented with unfamiliar faces oriented correctly or upside down. As expected, a statistically significant increase in rCBF was observed in the left TP/IT for each of the naming tasks. Moreover, person naming activated the left TP but not the left IT, whereas animal and tool naming activated the left IT and a restricted portion of the left TP. Tool naming activated the area more posterior and more lateral than an area for animal naming. To summarize, Damasio et al. (1996) obtained nicely converging evidence for the involvement of multiple regions of the left cerebral hemisphere, located outside the classic language area for word naming, and also for differential activation in anatomically separable regions of the brain in processing words for distinct kinds of items, i.e., humans, animals, and tools.

However, not all studies using PET identified the same sectors for seemingly the same categories (Caramazza & Shelton, 1998). Although we are sure that damage in different areas of the brain tends to produce correspondingly differentiated patterns of disorder, we are not confident that these areas are mainly responsible for cognitive tasks that tap into different onto-logical categories.

Functional Localization and Plasticity in the Human Brain

An important finding for knowledge acquisition obtained by recent imaging studies is that each part of the human brain is genetically assigned some particular functions that can be performed very well, but each can do other things if needed. This bounded plasticity is probably the key for us to understand how fast and easy learning in privileged domains can take place on the one hand and slow and laborious learning in non-privileged domains on the other.

We can see an exciting example of bounded plasticity in research on theory of mind, in which we can have three sets of data: (a) brain activation patterns among healthy people while solving theory of mind tasks, (b) cognitive and neural investigations of autistic individuals who are supposed to have genetically determined deficits in their theory of mind, and (c) the investigations of brain-damaged patients who have selective impairments in their theory of mind. We will first discuss a functional neuroimaging study by Fletcher et al. (1995) as an example of (a). They found by using PET that when compared with a ‘physical’ sequence of events, the comprehension of mentally-caused sequences of events (i.e., so-called ‘theory of mind’ stories, understanding of which requires a sophisticated attribution of mental states) activated a specific area of the left medial frontal cortex (Broad-mann’ area 8).

With regard to (b), as described in Section 10.2, autistic individuals show inferior performance in ‘theory of mind’ tasks, though their performance in tasks concerning physical causation is intact. Moreover, Happé et al. (1996), using patients with Asperger syndrome (to simplify, a mild form of autism), provided data suggesting that autistic individuals are impaired in the development of normal brain systems supporting naive psychological reasoning and understanding. More specifically, whereas among the normal individuals, Broadmann’ Area 8 was highly active during reading theory of mind stories and answering a question, this area was not active among the individuals with Asperger syndrome. Instead, the stories activated an adjacent area (Broadmann Area 9/10 in left medial frontal cortex). This area was activated in the normal people, but to a much less extent than the above area responsible for attributing mental states to others.

These results are consistent with the idea that there is a special area in the brain for attributing mental states, that this area does not function well among people with Asperger syndrome, and that a neighboring area may develop capacities to compensate for the mentalizing function to some extent. Probably due to this substitution, some children with autism can develop an understanding of theory of mind, although it occurs much later than in the normal people. Moreover, they still reveal some limitations, for example, that fail to distinguish shouting in a noisy environment from shouting because of anger.

More recently, as for (c), Happé, Brownell, and Winner (1999) examined ‘theory of mind’ in adults who had had a stroke in the right hemisphere, a group of people known to reveal pragmatic and social difficulties. The participants showed inferior performance in understanding materials requiring the attribution of mental states to non-mental materials, which were of the same difficulty among healthy elderly individuals and left hemispheric patients. The authors claim that their results support the notion of a dedicated cognitive system for theory of mind, and suggest that a healthy right hemisphere is needed for properly attributing mental states. However, it should be pointed out at this stage that, even if there is a dedicated system for theory of mind, defining its location in the brain will require many further studies, because exact locations of activation and/or damage differ even among carefully performed studies.

Socializing Higher-Order Cognition Research

Until the late 1980s, after the first wave of the cognitive revolution, most cognitive researchers concentrated on the study of symbol manipulation within the individual head, ignoring the sociocultural context that surrounds the individual (Gardner, 1985). As a result, it was often overlooked that our competence in daily life is heavily dependent on our continuous interaction with other people and tools.

Through the second wave of the cognitive revolution, however, many cognitive researchers were prepared to accept that the cognitive process, among others, higher-order cognition, can be greatly influenced by sociocultural situations. Within the cognitive science community, those views that emphasize the role of the environment, the context, the social and cultural setting in cognition have gained favor (Norman, 1993). The views are referred to by several different names, but here we use ‘sociocultural’ to represent all those views that assume that cognition occurs not just in the head but as a continuous interplay between an acting person’ mind and the external environment. In this final section we discuss contributions from these sociocultural views to knowledge acquisition and use.

Acquisition of Knowledge through Participation in Practice

Sociocultural views indicate that knowledge is acquired usually through participating in culturally organized practices that involve a variety of sociocultural constraints and that are interesting and/or significant to the participants (Goodnow, Miller, & Kessel, 1995). Although the constraints in practices are to support participants’ performance, rather than knowledge acquisition, repeated participation enhances the skills and knowledge needed to perform well in these practices.

Sociocultural constraints in practices include (a) artifacts which are shared by a majority of people of the community or its subgroup, and (b) the behavior of other people, interactions with them, and social contexts created by them. Because the artifacts, such as physical facilities and tools, social institutions and organizations, documented pieces of knowledge, and common sense and beliefs, direct people’ attention and eliminate a great number of possible interpretations in advance, people can usually identify quite easily what they should do in practices, and also acquire knowledge and skills they need rather promptly. Likewise, most of what we do and acquire in practices is affected by other people. Our attention is often directed to what other people look at and we tend to take into account ideas proposed by others.

This analysis applies to everyday activities as well as to culturally valued practices related to production, ritual, etc. When the mastery of effective knowledge and skills requires extensive experience, the practice often takes the form of apprenticeship. In apprenticeship, the novice is initially assigned to relatively simple and peripheral parts of the practice, and comes to fulfill gradually more difficult and central parts as he or she gains better local knowledge and skills, and the understanding of the total practice. For example, Lave (1988) reported that apprentice tailors in Liberia are given such jobs as fixing buttons, which are simple and easy to repair but constitute authentic and essential parts of tailoring. Socioculturalists claim that participation in practices (including apprenticeships) tends to successfully support the knowledge acquisition of a great majority of ordinary people.

Knowledge is Situated

Human knowledge acquisition is situated in the contexts in which experience occurs and cannot help but be influenced by various features of these contexts. Because the most important contexts for knowledge acquisition and use are the practices people participate in repeatedly, it is heavily influenced by the characteristics of these practices (Greeno & MMAP, 1998). For example, when a practice is oriented toward solving a fixed class of problems skillfully, participants tend to become experts distinguished in terms of speed, accuracy, and automaticity. In contrast, when successful participation in practice requires flexibility and adaptiveness, they may learn to adapt their knowledge and skills to new situational demands.

Acquired knowledge is also situated in the sense that it reflects the history of its acquisition and use. Although textbooks of a given domain summarize a body of knowledge as a set of propositions, the knowledge that individuals ‘possess’ in their head includes representations of a more personal, concrete nature. For example, knowledge in the form of a formula or law is often accompanied by some preferred examples (e.g., those that were used when the formula was first introduced and those that the learner was able to solve for the first time by applying the formula). It may also be accompanied by the social context of its acquisition (e.g., how the teacher explained it and how other students reacted). These personal representations often serve as clues for retrieving the formula when it is appropriate. Proponents of the socio-cultural views emphasize that much of our acquired knowledge is sociocultural in origin, and as such involves the internalization of socio-cultural constraints so that they can work within the mind (as mental models of tools and other people).

Distributed Cognition

Individuals engaging in any collective activity often constitute a distributed system of cognition. None of them possess all the pieces of knowledge needed for successfully conducting the activity. In some cases, each of the members knows well what he or she has to do, but knows only partially what is being done by other members and/or by the system as a whole. In other cases, individuals behave based on their own motivation without an agreed division of labor. Even without a ‘director’ who monitors individual actions so that they can be coordinated, the system seems to function smoothly and productively. This smooth collective activity of a group whose members’ knowledge is limited is attributed again to sociocultural constraints working in the activity. More specifically, these individuals select an alternative from their behavioral and conceptual repertories that is in accordance with other individuals’ behaviors (social constraints), and with shared artifacts (cultural constraints). Thanks to these constraints, the coordination of individual actions can be achieved without conscious planning or monitoring.

A good example of a distributed system of cognition is a team navigating a ship (Hutchins, 1995). Ship navigation is based on local problem solving. Although there is a captain, he does not control everything or know all that goes on. In principle, each member of the team has to perform his job only when a particular condition occurs. However, mutual aid is given fairly often, and is appreciated, because, if there were no mutual aid, one member’ failure to do his job could destroy the entire operation. To make this possible, there must be some overlap among the pieces of knowledge each member has. Since members move from one job to another as they gain expertise, more expert members usually know about less expert members’ jobs. Moreover, members may know other members’ jobs if they interact often.

Division of labor in distributed cognition does not have to be agreed upon in advance. In group problem solving and collective comprehension activity, different views are offered, based on participants’ cognitive and social positions. When a team consists of members of varied backgrounds, it is likely that different problem representations or interpretations are offered, and different forms of reasoning are induced. Agents of higher-order cognition are more often people constituting a team, at least temporarily, than isolated individuals.