Intelligent Tutoring Systems:

An Historic Review in the Context of the Development of Artificial Intelligence and Educational Psychology

Mark Urban-Lurain

urban@msu.edu


Table of Contents  

List of Tables

List of Figures

Introduction

Motivation

Overview

History

Artificial Intelligence and Behaviorism

Generative CAI

Adaptive Drill & Practice

Crises in AI and Educational Psychology

AI & Intelligent Tutoring Systems

The role of cognitive science

Challenges to ITS

ITS Meets Educational Psychology: Knowledge Communication

Domain Expertise

Student Models

Pedagogical Expertise

Interface

Social-Constructivist Transition

Cognition Across the Life Span (and beyond)

A Possible Middle Ground

Conclusion

References

List of Tables

Table 1: ACT* assumptions and related principles for a computer-implemented tutor.

List of Figures

Figure 1: Basic issues in knowledge communication


Introduction

Motivation

My interest in Intelligent Tutoring Systems (ITS) emerged from my work developing and using various instructional materials -- textbooks, television, computer simulations, computer-assisted instruction -- for teaching introductory computer science to non-computer science students. The most recent incarnation of these efforts, Integrated Introduction to Computing (Weinshank, Urban-Lurain, Danieli & McCuaig, 1995), drew upon the literature and experience of a number of specialists (Mayer, 1988; Olson, Sheppard & Soloway, 1987; Soloway & Spohrer, 1989b) to create a course which was innovative, engaging, and effective. Comparing student performance and instructor ratings for this course with its predecessor (Weinshank, Urban-Lurain & Danieli, 1988), indicates that we met many of our goals: students in this course perform better on standardized tests and a range of programming problems than did students in the previous course.

Although the average performance and student ratings of the course have improved over its predecessor, the dispersion of these measures is as great as or greater than it had been in the past. This implies that we are having problems reaching students who are well below the mean, losing a great number of the very students who most need this material to be successful in an information-rich world. In addition, we are seeing an increasingly large number of students who bring a richer set of backgrounds to the course than did previous students. These students do not find the course challenging and consequently believe that computer science is uninteresting. Hence, we lose a number of students who might have had potential to become successful computer science majors.

With these factors in mind, we began to rethink the architecture of the course. We realized that a more modular, finer-grained curriculum for this course could have the potential to address many of these concerns by better targeting a wider range of student interests and abilities. Nonetheless, we had very real constraints on the size of the teaching staff. We knew that we would have to continue to use instructional technology to leverage faculty resources. Consequently, we became interested in the prospects for creating an Intelligent Tutoring System that would allow us to:

The first step in this process is to understand what others had done before us and the implications for future developments.

Overview

In this paper, I review the history of Intelligent Tutoring Systems and discuss their development in the context of Artificial Intelligence (AI) and educational theory. Until very recently, workers in the AI community have performed the majority of work on ITS with little interaction with educational researchers. Although there was great enthusiasm for the prospects of ITS throughout the 1970's and into the 1980's, this excitement has recently waned. A number of developments in both AI and educational psychology have caused many to forsake ITS. Some consider them an embarrassing reminder of the naive enthusiasm both disciplines had, preferring to concentrate on issues such as the use of standard computer software as cognitive tools in the classroom. Others suspect that ITS advocates want to replace teachers. However, it may be premature to dismiss ITS as an educational dead end.

History

Artificial Intelligence and Behaviorism

Intelligent Tutoring Systems have an interesting history, originating in the Artificial Intelligence (AI) movement of the late 1950's and early 1960's. Then, workers such as Alan Turing, Marvin Minsky, John McCarthy and Allen Newell thought that computers that could "think" as humans do were just around the corner. Many thought that the main constraint on this goal was the creation of faster, bigger computers. It seemed reasonable to assume that, once we created machines that could think, they could perform any task we associate with human thought, such as instruction.

Education in the United States during this period also provided fertile ground for the development of ITS. Education exhibited the same post-war "can do" enthusiasm that characterized AI research. As thousands of GIs returned from World War II and attended college on the GI Bill, educational institutions underwent unprecedented growth. In turn, this required that the education system ramp-up to meet the challenges of preparing larger proportions of their students to attend college. To do so, they sought ways to increase the efficiency of instruction. As Berliner notes:

The War did not require theoretical elegance from its psychologists. It required solving practical problems, not laboratory problems, such as the problem of rapidly teaching masses of men to reach acceptable levels of competency in hundreds of specialty areas. With the help of psychologists, the task was accomplished. (Berliner, 1992) (p. 47)

Berliner points out that these teaching strategies carried into the textbooks and public education of the 1960's that were primarily based on Thorndike's S-R associations. In this environment, it seemed only logical to use emerging computer technology to meet the demands of education.

Generative CAI

In the 1960's, researchers created a number of Computer Assisted Instructional (CAI) systems that were generative (Uhr, 1969). These programs generated sets of problems designed to enhance student performance in skill-based domains, primarily arithmetic and vocabulary recall. Essentially, these were automated flash card systems, designed to present the student with a problem, receive and record the student's response, and tabulate the student's overall performance on the task. Most of the system designers' efforts were devoted to grappling with the technical challenges of programming these systems on the expensive and relatively cumbersome mainframe computers of the time. These systems did not explicitly address the issues of how people learn, with an implicit behaviorist / transmission model of teaching and learning. They assumed that, if systems presented information to the learner, the learner would absorb it.

Adaptive Drill & Practice

By the late 1960's and early 1970's, many researchers had moved beyond merely presenting problems to learners while collecting and tabulating their responses, to considering the student a factor in the overall instructional system (Suppes, 1967). Many researchers developed systems that altered the presentation of new materials based on the history of a student's responses. While it was no longer sufficient to present information in a predetermined pattern to all students, system designers still had to anticipate all possible responses. The programmers had to know in advance what types of student responses were possible and decide what information the system would then present. These systems were the first to "model" students, although they only modeled the students' behavior and not on any attempt to model their knowledge states. Even though these systems were relatively simple, by constraining themselves to the development of skills and recall they were very effective. Students who used these systems improved on measures of the relevant skills and factual recall. The implicit learning theory assumed that students needed to first learn these basic skills and facts in order to prepare them for higher synthetic skills.

Crises in AI and Educational Psychology

During this period, there was great hope for rapid progress in AI, fueled in part by the rapid advances in computational power and the assumption that there would be concomitant advances in AI. However, it soon became apparent that the problems of AI were much more intractable than the relatively straightforward challenges of building faster computers. Nonetheless, AI researchers remained optimistic throughout the 1960's, 70's and into the 80's, even though the goal of thinking computers always seemed to be "just ten years away."

About this time, educational psychology was questioning the assumptions of behaviorism. Piaget's theories of learning and constructivism began to take hold. Chomsky, along with Newell and others, introduced the ideas of symbolic information processing (Greeno, Collins, Beranek & Resnick, 1994), ideas that dovetailed with the AI community's interests in linguistics and natural language processing. Information Processing (IP) was emerging as a dominant paradigm during the late 1970's and early 1980's. This work conceived human cognition as set of "black box" processes, rather than merely responses to external stimuli. The precision of IP models held the promise of being amenable to being described by computer programs.

AI & Intelligent Tutoring Systems

In 1982, Sleeman and Brown reviewed the state of the art in computer aided instruction and first coined the term Intelligent Tutoring Systems (ITS) to describe these evolving systems and distinguish them from the previous CAI systems. The implicit assumption about the learner now focused on learning-by-doing. They classified the existing ITS as being computer-based (1) problem-solving monitors, (2) coaches, (3) laboratory instructors, and (4) consultants. (Sleeman & Brown, 1982) The emphasis in these systems was still as research platforms for refining AI theories, but now researchers were thinking about representing student knowledge within these systems. Here we find the first use of the term student model to describe an abstract representation of the learner within the computer program. They classified student models as:

Overlay: a model of student knowledge as a subset of an expert's knowledge.

Differential: similar to the overlay model, this focuses on the differences between the student's knowledge and that of the expert.

Perturbation: representing student misconceptions as variants of the procedural structure of the experts' correct skill. These are sometimes known as mal-rules (bad rules) that describe student misconceptions in relationship to the experts' knowledge.

Early attempts to model student knowledge were based on a "buggy" model first proposed by Brown and Burton (Brown & Burton, 1978). "Bugs" are student errors in discrete skills, such as incorrect carrying during subtraction. Burton elaborated on this model with the DEBUGGY system (Burton, 1982). DEBUGGY identified 130 "bugs" designed to account for mistakes in subtraction. The challenge was to analyze the problem space represented by the student's answers and determine which bug or set of bugs best accounted for incorrect subtraction.

Sleeman and Brown mention some learning issues related to the problems involved in creating ITS. They acknowledge that much human tutor communication is implicit and express the hope that ITS will provide a venue for educational theorists to develop "more precise theories of teaching and learning." (p. 9) Their assumption is that such precision is possible, being necessary for implementation of these theories within computer software. They also discuss the need to construct environments that encourage collaborative learning, while acknowledging that researchers (at that time) knew little about how such cooperation takes place in natural learning settings.

The role of cognitive science

During the 1980's, computer scientists specializing in AI continued to focus on the problems of natural language, student models, and deduction. However, the field also attracted researchers from outside the computer science discipline, most notably John Anderson. Anderson was working in cognitive science, developing the Adaptive Control of Thought (ACT*) theory of cognition (Anderson, 1983). Table 1 summarizes the ACT* principles and their implications for ITS (Corbett & Anderson, 1992) (p. 83)

Table 1: ACT* assumptions and related principles for a computer-implemented tutor.

ACT* Assumptions

Corresponding Tutoring Principles

Problem-solving behavior is goal driven Communicate the goal structure underlying the problem-solving task.
Declarative and procedural knowledge are separate. The units of procedural knowledge are IF-THEN rules called productions. Represent the student's knowledge as a production set.
Initial performance of a task is accomplished by applying weak (general) procedures to declarative knowledge structures. Provide instruction in the problem-solving context; let student's knowledge develop through successive approximations to the target skill.
Task-specific productions arise by applying weaker productions to declarative knowledge. These task-specific productions underlie more efficient performance. Provide immediate feedback on efforts.
As a result of additional practice, productions can be chained together into larger-scale productions. Adjust the step size of instruction as learning progresses.
The student maintains the current state of the problem in a limited capacity working memory. Minimize working memory load.

 

Although Anderson and his colleagues created ACT* as a cognitive theory, they believed that it was rigorous enough to test by implementing the principles in computer software. Two of the best known examples are the Geometry Tutor (Koedinger & Anderson, 1993) and LISPITS (LISP Intelligent Tutoring System). The LISPITS system, a program for teaching LISP programming, was designed to implement these principles in the context of "model tracing." LISPITS attempts to model the steps needed to write a LISP program. The program then compares the actual steps that the student takes with this model. Corbett and Anderson call the monitoring and remediating process knowledge tracing. Their goal is a mastery model, where every student masters 95% of the rules for a given set of exercises before moving to the next section. Corbett and Anderson found that students using LISPITS completed the mastery model exercises considerably faster than students who worked alone, but not as fast as students who worked with human tutors.

Anderson's name has become synonymous with ITS work insofar as people often speak of "Anderson-style tutors" (Chipman, 1993) (p. 352) Perhaps this is because his systems are some of the few that have actually been used in classroom settings and were not solely research projects.

Challenges to ITS

By the mid-1980's, much of enthusiasm in AI for creating "thinking" computers had waned as the field began to mature. Researchers turned to the more prosaic tasks of building expert systems that could function well in constrained domains, such as troubleshooting and diagnostic systems. At the same time, as ITS began to move out of the AI laboratories into classrooms and other instructional settings, they began to attract critical reactions. Some shortcomings of ITS became apparent as researchers realized that the problems associated with creating ITS were more intractable than they had originally anticipated. Rosenberg notes that most papers about ITS make few references to the education literature; the majority are grounded in the computing literature. He asserts that much ITS work suffers from two major flaws:

The systems are not grounded in a substantiated model of learning. Model formulation should be preceded by protocol analysis, but very little analysis is done, almost none of it qualitative. ITS models should be validated by the teachers and students who will use the systems, but ITS researchers do not appear to consult these experts.

Testing is incomplete, inconclusive, or in some cases totally lacking. Data on computerized tutorials is, at best, mixed. The almost universally positive claims for ITS and other computerized instructional systems -- most notable in the education literature -- are based on results from severely flawed tests. (Rosenberg, 1987) (p. 11)

It was obvious that the basic premises of ITS research needed revision.

ITS Meets Educational Psychology: Knowledge Communication

During this same period, constructivism in its various forms was becoming a dominant theme in educational psychology. Constructivists approached cognition from a more holistic perspective than either the behaviorists or the information processing advocates, claiming that cognition could not be reduced to the interactions among a number of black boxes because doing so fails to account for the learner reflecting on his or her cognitive strategy. Constructivists claim that the underlying computer metaphor makes it impossible for IP models to incorporate reflection since computers are not self-aware. (Kuhn, 1992)

By 1987, a review of the field in by Wenger demonstrated how much it had evolved in the five years since Sleeman and Brown's synopsis (Wenger, 1987). Wenger examined the goals -- implicit and explicit -- of ITS designers. Perhaps most significantly, he considered ITS a part of what he labels "knowledge communication" and focused on cognitive and learning aspects of these systems, in addition to the AI issues. He proposed what might become the basis for a discipline that combines the work of researchers from AI, cognitive science, and education.

"... consider again the example of books: they have certainly outperformed people in the precision and permanence of their memory, and the reliability of their patience. For this reason, they have been invaluable to humankind. Now imagine active books that can interact with the reader to communicate knowledge at the appropriate level, selectively highlighting the interconnectedness and ramifications of items, recalling relevant information, probing understanding, explaining difficult areas in more depth, skipping over seemingly known material ... intelligent knowledge communication systems are indeed an attractive dream." (p. 6)

Wenger calls for a move towards a "cognitively oriented form of software engineering" in which cognition is central, rather than focusing on computational models of the domain and pedagogy. He points out that by the mid-1980's there were two opposing views of ITS: the traditional view of computers as instructional delivery devices and the emerging view of computers as a tool for exploratory learning. He claims that by viewing ITS as knowledge communication tools it is possible to merge these apparently opposing views of ITS. Wenger presents the following model of ITS. (p. 24)

Figure 1: Basic issues in knowledge communication

While these classifications are consistent with previous work on ITS, Wenger attempts to abstract the various components beyond their more traditional software engineering definitions.

Domain Expertise

In traditional CAI, the expert knowledge is contained in blocks or chunks called "frames." The system presents frames to the student, with the presentation order determined by the student's responses to previous frames. If the student answers a set of test questions correctly, the system presents the next frame in a sequence. If the student answers incorrectly, alternative frames are presented. However, these frames are static, they are not capable of "applying" the knowledge as would a human expert.

The key feature that distinguishes a knowledge communication system from standard ITS on the Domain Expertise dimension is that the representation of the subject matter is not merely a set of static frames, but actually is a dynamic model of the domain knowledge and a set of rules by which the system can "reason." These systems have their roots in expert systems research (such as medical diagnostic or electronic troubleshooting systems) and have the ability to generate multiple correct sets of solutions, rather than a single idealized expert solution.

Student Models

Student modeling remains at the core of ITS research (Holt, Dubs, Jones & Greer, 1994). What distinguishes ITS from CAI is the goal of being able to respond to the individual student's learning style to deliver customized instruction. Although some authors are questioning the goal of student modeling either because of technical limitations (McCalla, 1992) or larger philosophical grounds (Sack, Soloway & Weingrad, 1994), this is still an area of active research. Mitchell argues that an ITS must model the world, the learner, and the teacher-learner interaction (Mitchell & Grogono, 1993). A recent conference on ITS focused exclusively on different aspects of this problem (Greer & McCalla, 1994), demonstrating that this is still a fruitful and active area of research.

According to Wenger, student models have three tasks.

They must gather data from and about the learner. This data can be explicit -- asking the student to solve specific problems -- or implicit -- tracking the students navigation and other interactions and comparing them to information about similar learner responses.

They must use that data to create a representation of the student's knowledge and learning process. This often takes the form of "buggy" models that represent the student's knowledge in terms of deviations from an expert's knowledge. The system then uses this model to predict what type of response the student will make in subsequent situations, compares that prediction to the students' actual response, and uses that information to refine the model of the student.

The student model must account for the data by performing some type of diagnosis, both of the state of the student's knowledge and in terms of selecting optimal pedagogical strategies for presenting subsequent domain information to the student. One of the biggest challenges is to account for "noisy" data, the fact that students do not always respond consistently, particularly when their knowledge is fragile and they are uncertain about the correct responses.

Pedagogical Expertise

According to Wenger, when "learning is viewed as successive transitions between knowledge states, the purpose of teaching is accordingly to facilitate the student's traversal of the space of knowledge states." (p. 365) The ITS must model the student's current knowledge and support the transition to a new knowledge state. He claims that this requires that ITS alternate between diagnostic and didactic support.

Diagnosis

Diagnosis means that an ITS infers information about the learner's state on three levels.

At the behavioral level, ignoring the learner's knowledge and focusing only on the observable behavior.

At the epistemic level, dealing with the learner's knowledge state and attempting to infer that state based on observed behavior.

At the individual level, covering such areas as the learner's personality, motivational style, self-concept in relation to the domain in question, and conceptions the learner has of the ITS. Wenger notes that, up to now, ITS have not been concerned with the individual level. However, he advocates further research in this area as a prerequisite for viewing the student as an active learner, rather than as a passive recipient of knowledge.

Gathering diagnostic information presents a challenge for ITS designers. The only data readily available are the outcomes of individual tasks or problems presented and some fine-grained interface data, such as keystrokes and the amount of time between the keystrokes or mouse movements. There have been two approaches to using this data.

Model tracing attempts to map the external data collected by the system to the internal model representation of the problem space. This approach was developed in frame-based CAI programs and attempts to classify the student's knowledge state into one or more states that the system designers had to anticipate.

Reconstructive interpretation is based on problem-solving in the context of goals or plans. Rather than attempting to classify the data into one of a small set of monolithic models, reconstructive interpretation starts with a set of plans (sometimes known as methods, rules, or schemes) which are decomposed into goals and subgoals. As it collects additional data, the ITS can prune this list to reduce the search space and assign probabilities to various nodes in the search space. The ITS adjusts these probabilities dynamically, allowing the ITS to respond more "intelligently" to the type of erratic, noisy responses that students often display when learning new material.

Didactics

The second facet of pedagogical expertise Wenger identifies is didactic support, the "delivery" aspect of teaching. Generally, ITS have concentrated on the modeling and manipulation of the content or domain, with little attention being paid to didactics. For example, although the self-improving quadratic tutor (O'Shea, 1981) can modify its teaching strategy by adjusting the production rules, the rules and their modifications are very integrated with the specifics of the domain. There has been little research on abstracting more general didactic principles in a manner that could be used to construct ITS.

Wenger suggests that didactics can be organized around four principles that may be fruitful for future ITS research.

Plans of action: "mini curricula" that are used to lead the learner and provide the context for diagnostic operations.

Strategic contexts: in which the plans of action are implemented. Generally, these are planned (intended by the teacher) or opportunistic (the learner does something or asks a question that provides this opportunity for teaching).

Decision base: rules or guidelines for allocating the system resources in the context of constraints. This area has been one in which ITS designers have made a number of tradeoffs. For example, SOPHIE (Brown, Burton & deKleer, 1982) has domain-specific expertise for electronics troubleshooting but does so at the expense of explanatory capability; it can demonstrate a wide variety of procedures, but it cannot explain what it is doing in the context of an "understanding" of the student. On the other hand, attempts to build more general explanatory systems such as MENO (Duchastel, 1989) sacrifice domain expertise, sometimes to the point of triviality.

Target level of the student model: selecting the level at which the teaching takes place. At any moment, the student may be functioning at the behavioral, conceptual, or metacognitive level, or at multiple levels simultaneously. An expert human teacher can understand and respond appropriately to these levels. Up to now, most ITS focused on behavior, or attempted to infer the student's conceptual level from it. Rarely have systems attempted to determine at what level the student needs coaching at any particular moment and adjust the plan of action accordingly.

Interface

The interface allows communication between the student and the other aspects of the ITS. Here, research from the human factors and software design disciplines is applicable, but the pedagogical implications of an ITS interface must also be considered. Wenger suggests that the goal of knowledge communication requires that the interface contain a discourse model to resolve ambiguities in the student responses. Since the learner is most likely to provide incomplete or contradictory responses when stymied, providing a properly supportive response that can advance the diagnostic process is important. This helps the ITS avoid redundant presentations and enhances instruction. For example, SOPHIE is explicitly designed to search through its knowledge database and compare it with the student answers for a "close" match when attempting to recover from unexpected student responses.

The other facet of the interface is knowledge presentation. If a system merely makes knowledge available, it becomes a knowledge exploration environment. Such systems place all the responsibility for learning upon the learner, who must navigate through the knowledge using the interface provided. For example, this is the current situation on the World Wide Web. For knowledge communication to take place -- even in an exploratory environment -- an ITS must provide some coaching or guidance to prevent the student from foundering or missing important aspects of the domain. The desire for ITS to provide more active guidance or tutoring raises the specter of ITS replacing human teachers, a topic that always prompts impassioned discussion (Epstein & Hillegeist, 1990). However, Wenger points out that replacing human teachers is not be the issue:

The anthropomorphic view that more intelligence for systems means more humanlike capabilities can be as much of a distraction as it is an inspiration. Indeed, the communication environment created by two people and that created by a person and a machine are not likely to be the same. The terms of the cooperation required for successful communication may differ in fundamental ways. Hence, computational models of knowledge communication will require new theories of knowledge communication, as computer-based systems evolve and as research in artificial intelligence and related disciplines provides more powerful models. (p. 426)

Social-Constructivist Transition

Vygotsky and others influenced educational psychology by raising concerns about the context in which learning and development takes place and focusing on the effects of culture on the development of the individual. These ideas had an important impact on one of the eminent researchers in ITS. During the early 1980's Elliot Soloway and his colleagues at Yale (with the Cognition and Programming Project) took interest in issues surrounding teaching computer programming to novice students. They were curious about the transfer of programming skills to general problem-solving skills, since the conventional wisdom was that computer programming improved problem-solving in other domains (Soloway & Spohrer, 1989a). Their work carried them into ITS research, and they developed an extensive research program (Highly Interactive Computing Environments [HiCE] Group at University of Michigan) using ITS to tutor students who were learning to program. Their main focus was on student modeling, categorizing student bugs and creating systems that could help students identify and repair these bugs (Sack & Soloway, 1992).

However, after more than ten years of work in this area, the members of the HiCE group began to question some of their underlying assumptions. In a fascinating epilogue from a workshop on Student Modeling (Greer & McCalla, 1994), Sack, Soloway and Weingrad step back and critically review the history of their work (Sack et al., 1994). They note that they have come to question many of their assumptions about the nature of learning, after considering the work they have done with students who used their software.

Central to the change in their thinking is the transition from an objectivist to a constructivist perspective on student learning. In their early work, they viewed bugs as deviations between the "correct" expert's solution and the "incorrect" student's version. Their goal was bringing the student's conception into congruence with the expert's. However, their experiences using ITS in real classrooms caused them to reconsider their assumptions about learning. Rather than thinking of learning as a process of transfer of knowledge and correction of "buggy" knowledge, they began to view learning as a process of enculturation into a knowledge community. This, in turn, caused them to rethink their ideas about student models.

"Student models are also issues of community: a given student model is, in a certain sense, a record or an example of a student's reputation and abilities in a given community; student models, in general, are the sorts of values and attributes that one might impose on the ideal, or model student. . . . instead of trying to model students, we are now trying to provide students with the tools, facilities and communities they need to support the development of models for their own uses." (p. 373)

Cognition Across the Life Span (and beyond)

During this same period, other researchers began to think of development as continuing to occur beyond childhood, across the adult life span. This work pointed out that, contrary to previous thought, adults continue to remain flexible in their cognitive performance and development. The need to account for this plasticity is a major challenge facing cognitive theorists. (Kuhn, 1992) Issues of cognition that were germane to ITS were continuing to extend to a wider number of domains, the time was ripe for a gestalt shift. Clarke summed it up succinctly.

"Cognitive science seeks to understand how the mind emerges from the brain. Artificial intelligence seeks to implant a mind within the machine. To their detriment, both have largely ignored work on the natural history of the mind." (Clarke, 1993) (p. 754)

Along these lines, in 1991, Merlin Donald reviewed a broad range of works across paleontology, linguistics, anthropology, cognitive science and neuropsychology to propose a theory of the evolution of human cognition. He claims that during the past two million years humans have undergone three major cognitive changes (Donald, 1991):

Mimetic skill; the "ability to produce conscious, self-initiated, representational acts that are intentional but not linguistic." (p. 168)

The development of language.

The use of External Symbolic Storage systems; pictographs, graphics, and written language.

Each of these cognitive changes has supplemented, but not replaced, the previous ones. Most critical for the understanding the rapid pace of our cultural and cognitive evolution is the use of External Symbolic Storage (ESS). Donald claims that the plasticity of the human cognitive apparatus and the use of ESS are inseparable.

"As we develop new external symbolic configurations and modalities, we reconfigure our own mental architecture in nontrivial ways. The third transition has led to one of the greatest reconfigurations of cognitive structure in mammalian history, without major genetic change." (p. 382)

Our cognitive structures are now so dependent upon the ESS that we must conceive of human evolution in terms of the advances in the technologies with which we represent the ESS, and not in biological terms. Genetically, we are no different than our ancestors of one hundred thousand years ago, but we have evolved into what, by any measure, is a unique species that would be unrecognizable to early Homo Sapiens. By taking advantage of the plasticity of the human cognitive apparatus, we have evolved at a rate beyond our capacity for biological change.

A Possible Middle Ground

Considering Wenger's daunting framework, and considering Sack, Soloway and Weingrad, one might conclude that ITS are impossible to create. However, the prospects for ITS may not be so bleak. Donald and (De Kerckhove, 1995) claim that global computer and video networks are further accelerating the pace of change and that the role of the individual mind is changing in ways that we cannot yet predict. These changes are reflected in the evolving cognitive theories that we have seen in the past thirty years in educational psychology. Consider the impact of written word on our cognitive processes. Reading and writing are now such an integral part of how we learn and think that studying them is now a major avenue for understanding our cognitive process. Just as a better understanding of reading informs our cognitive theories and these theories in turn inform the ways in which we teach reading, so too will understanding the ways in which we interact with evolving knowledge communication systems inform both our theories of cognition and the creation of these systems. Our cognitive theories will need to evolve, not only to describe how we interact with these systems, but in order to accommodate the changes to our cognitive processes that these systems will bring. With an improved understanding of the evolution of our cognitive processes, we will be able to create better knowledge communication systems. In turn, these systems will be built upon the evolving cognitive theories, in addition to computer science theories of information processing.

Conclusion

Intelligent Tutoring Systems emerged from Artificial Intelligence at the very time that AI was struggling to transcend the goal of mimicking human intelligence by creating machines that could "think" like humans. As researchers came to grips with the intractable problems of this task, they realized that trying to emulate human cognition with computers was misguided because they assumed that people thought like computers. The resulting crisis provoked a reassessment of AI's goals, allowing researchers to begin making progress in areas such as expert systems. Expert systems research was productive because it concentrated on systems that were useful in their right, rather than attempting to create "thinking" machines. However, this shift in focus prompted many to lose interest in ITS.

At the same time, educational psychology was undergoing a paradigm shift from behaviorism towards cognition, constructivism, and socially situated learning. This revolution prompted many educators to question the practices that evolved during the post-war education boom. ITS technology, much of which was grounded in the behaviorism of CAI, lost favor.

It might appear that ITS are doomed to become a footnote in the history of both computer science and educational psychology. However, the prospect of applying the rapidly expanding power of computers not just to information management, but to knowledge communication, is too appealing to allow us to dismiss ITS research just yet. Combining Wenger's framework with a global perspective such as that suggested by Donald provides one possible avenue for developing the necessary interdisciplinary theories upon which new research and ITS can be developed. Moving towards a cognitive understanding of productive communication environments is likely to be fruitful for both ITS and educational researchers. In this way we may be able to create the theories and technology required to make the dream of intelligent knowledge communication systems a reality.


References

Anderson, J. R. (1983). The Architecture of Cognition. Cambridge, Massachusetts: Harvard University Press.

Berliner, D. C. (1992, August, 1992). The Science of Psychology and the Practice of Schooling: The One Hundred Year Journey of Educational Psychology from Interest, to Disdain, to respect for Practice. Paper presented at the American Psychological Association, Washington, D.C.

Brown, J. S., & Burton, R. R. (1978). Diagnostic models for procedural bugs in basic mathmatical skills. Cognitive Science, 2, 155-192.

Brown, J. S., Burton, R. R., & deKleer, J. (1982). Pedagogical, natural language and knowledge engineering techniques in SOPHIE I, II and III. In D. Sleeman & J. S. Brown (Eds.), Intelligent Tutoring Systems (pp. 227-282). New York: Academic Press.

Burton, R. R. (1982). Diagnosing bugs in a simple procedural skill. In D. Sleeman & J. S. Brown (Eds.), Intelligent Tutoring Systems (pp. 157-183). New York: Academic Press.

Chipman, S. F. (1993). Gazing Once More Into the Silicon Chip: Who's Revolutionary Now? In S. P. Lajoie, Ed. & S. J. Derry, Ed (Eds.), Computers as Cognitive Tools (pp. 341-367). Hillsdale, New Jersey: Lawrence Erlbaum Associates, Publishers.

Clarke, T. L. (1993). A natural history of the mind: A guide for cognitive science. Behavioral and Brain Sciences, 16(4), 754-755.

Corbett, A. T., & Anderson, J. R. (1992). LISP intelligent Tutoring System: Research in Skill Acquisition. In J. H. Larkin & R. W. Chabay (Eds.), Computer-Assisted Instruction and Intelligent Tutoring Systems: Shared Goals and Complementary Approaches (pp. 73-109). Hillsdale, New Jersey: Lawrence Erlbaum Associates, Publishers.

De Kerckhove, D. (1995). The Skin of Culture: Investigating the New Electronic Reality. Toronto: Somerville House Publishing.

Donald, M. (1991). Origins of the modern mind: three stages in the evolutions of culture and cognition. Cambridge, MA: Harvard University Press.

Duchastel, P. (1989). ICAI Systems: Issues in Computer Tutoring. Computers and Education, 13(1), 95-100.

Epstein, K., & Hillegeist, E. (1990). Intelligent Instructional Systems: Teachers and Computer-Based Intelligent Tutoring Systems. Educational Technology, 30(11), 13-19.

Greeno, J. G., Collins, A., Beranek, B., & Resnick, L. B. (1994). Cognition and Learning. In D. Berliner & R. Calfee (Eds.), Handbook of educational psychology (pp. 1-51). 7/31/94 draft: CEP 900 Course Pack.

Greer, J. E., & McCalla, G. I. (Eds.). (1994). Student Modelling: The Key to Individualized Knowledge-Based Instruction. (Vol. 125). Berlin: Springer-Verlag.

Holt, P., Dubs, S., Jones, M., & Greer, J. (1994). The State of Student Modeling. In J. E. Greer & G. I. McCalla (Eds.), Student Modelling: The Key to Individualized Knowledge-Based Instruction (NATO ASI Series ed., Vol. 125, pp. 3-35). Berlin: Springer-Verlag.

Koedinger, K. R., & Anderson, J. R. (1993). Reifying Implicit Planning in Geometry: Guidelines for Model-Based Intelligent Tutoring System Design. In S. P. Lajoie, Ed. & S. J. Derry, Ed (Eds.), Computers as Cognitive Tools (pp. 15-45). Hillsdale, New Jersey: Lawrence Erlbaum Associates, Publishers.

Kuhn, D. (1992). Cognitive development. In M. H. Bornstein & M. E. Lamb (Eds.), Developmental psychology: An advanced textbook (3rd ed., pp. 211-272). Hillsdale, NJ: Lawrence Erlbaum.

Mayer, R. E. (Ed.). (1988). Teaching and Learning Computer Programming: Multiple Research Perspectives. Hillsdale, New Jersey: Lawrence Erlbaum Associates.

McCalla, G. I. (1992). The Central Importance of Student Modelling to Intelligent Tutoring. In E. Costa (Ed.), New Directions for Intelligent Tutoring Systems (Vol. 91, pp. 107-131). Berlin: Springer-Verlag.

Mitchell, P. D., & Grogono, P. D. (1993). Modelling Techniques for Tutoring Systems. Computers and Education, 20(1), 55-61.

Olson, G. M., Sheppard, S., & Soloway, E. (Eds.). (1987). Empirical Studies of Programmers: Second Workshop. Norwood, New Jersey: Ablex Publishing Corporation.

O'Shea, T. (1981). A self-improving quadratic tutor. In D. Sleeman & J. S. Brown (Eds.), Intelligent Tutoring Systems (pp. 309-336). New York: Academic Press.

Rosenberg, R. (1987). A Critical Analysis of Research on Intelligent Tutoring Systems. Educational Technology, 27(11), 7-13.

Sack, W., & Soloway, E. (1992). From PROUST to CHIRON: ITS Design as Iterative Engineering; Intermediate Results are Important! In J. H. Larkin & R. W. Chabay (Eds.), Computer-Assisted Instruction and Intelligent Tutoring Systems: Shared Goals and Complementary Approaches (pp. 239-274). Hilsdale, New Jersey: Lawrence Erlbaum Associates, Publishers.

Sack, W., Soloway, E., & Weingrad, P. (1994). Re-Writing Cartesian Student Models. In J. E. Greer & G. I. McCalla (Eds.), Student Modelling: The Key to Individualized Knowledge-Based Instruction (NATO ASI Series ed., Vol. 125, pp. 355-376). Berlin: Springer-Verlag.

Sleeman, D., & Brown, J. S. (1982). Introduction: Intelligent Tutoring Systems. In D. Sleeman & J. S. Brown (Eds.), Intelligent Tutoring Systems (pp. 1-11). New York: Academic Press.

Soloway, E., & Spohrer, J. (1989a). Novice Mistakes: Are the Folk Wisdoms Correct? In E. Soloway & J. Spohrer (Eds.), Studying the Novice Programmer (pp. 401-418). Hillsdale, New Jersey: Lawrence Erlbaum Associates, Inc.

Soloway, E., & Spohrer, J. (Eds.). (1989b). Studying the Novice Programmer. Hillsdale, New Jersey: Lawrence Erlbaum Associates, Inc.

Suppes, P. (1967). Some theoretical models for mathematics learning. Journal of Research and Development in Education, 1, 5-22.

Uhr, L. (1969). Teaching machine programs that generate problems as a function of interaction with students. Proceedings of the 24th National Confernece, 125-134.

Weinshank, D. J., Urban-Lurain, M., & Danieli, T. (1988). Introduction to Computing: Telecourse with Structured BASIC. (2nd ed.). Dubuque, Iowa: Kendall/Hunt Publishing Company.

Weinshank, D. J., Urban-Lurain, M., Danieli, T., & McCuaig, G. (1995). Integrated Introduction to Computing. (updated first edition, revised and enlarged ed.). Dubuque, Iowa: Kendall/Hunt Publishing Company.

Wenger, E. (1987). Artificial Intelligence and Tutoring Systems: Computational and Cognitive Approaches to the Communication of Knowledge. Los Altos, CA: Morgan Kaufmann Publishers, Inc.