How the Science of Artificial Intelligence
Shapes Knowing
EDUC 800, Ways of Knowing, April 25, 2001

Introduction
Of all the ways the human mind has of knowing something – and by knowing, we mean a shared, valued process of getting to a collective truth – the science of Artificial Intelligence (AI) has proven to be the among the most enigmatic. Bursting onto the scene a few years after the end of World War II, AI quickly established itself as what Kuhn (1962) calls “normal science”, and for nearly a quarter of a century, defined the scientific process to a collective set of truths about the boundaries of human intelligence versus machine intelligence.
There is no single, agreed-upon definition of Artificial Intelligence. However, Finlay and Dix (1996) offer a working definition when they note that “AI is concerned with building machines that can and react appropriately, adapting their response to the demands of the situation, and displaying behavior comparable with that considered to require intelligence in humans.” (p. 2). At first, AI appeared to be an excellent illustration of Kuhn’s paradigm dialectic. After establishing itself in 1970 as “normal science”, AI encountered anomalies in the 1980s that sparked a crisis in the field that generated alternative theories of intelligence. Unlike Kuhn’s dialectic, however, no best alternative has ever appeared to replace the old AI paradigm and become the “new” normal science. Instead, those alternative theories generated their own crises, which in turn generated even more theories. The end result – at least, the result as it looks in 2001 – are three collective sets of truths, each with their own “normal science” status, recognized by the scientific community, co-existing and shaping knowing in their own way. Yet, each sits “comfortably” under the umbrella name of Artificial Intelligence. The purpose of this paper to describe ...

AI Shaping Knowing: The Road to “Normal Science”
Between 1945 and 1950, scientists who had worked with wartime calculating machines began to re-examine the use of computers for peaceful purposes. Two groups of innovators emerged, each with its own vision of the potential of computers and each with its own research program. The first group, led by Herbert Simon and Allan Newell, started from the premise that the human mind and the computer are two different instances of a single species of device: a device that generates intelligent behavior by manipulating symbols using a set of formal rules grounded in the mind’s representation of the world (Simon 1996). For Simon and Newell, problem solving was the paradigm of intelligence. As such, science should flesh out the mind’s representations, so that the computer can be programmed with the appropriate sets of rules to carry out the desired behavior.
The other group, led by Frank Rosenblatt, started from the premise that learning is the paradigm of intelligence. Drawing upon the principles of neuroscience, this group believed that science should attempt to automate the procedures by which a network of neurons learns to discriminate patterns and respond appropriately. Computers should be used as a medium for modeling the human brain. Dreyfus and Dreyfus (1988) summarize the differences between the two groups as follows:

By 1956, Simon and Newell had successfully programmed a computer using symbolic representations to solve simple puzzles and prove theorems. They also predicted that computer “intelligence” would exceed that of humans. In an article in the Jan.-Feb. 1958 issue of Operations Research (as cited in Dreyfus and Dreyfus 1988), Simon and Newell state that “intuition, insight, and learning are no longer exclusive possessions of humans: any large high-speed computer can be programmed to exhibit them also.” (p. 19). Rosenblatt had also put his ideas to work by building what he called a perceptron, an information processing automaton constructed as an analogue of the biological brain, and presented as empirical evidence of the feasibility of human cognitive functions in computers. At Dartmouth College that same year, John McCarthy called together the world’s foremost logicians, computer scientists and cybernetics technicians involved in the design of automata, to launch the idea of Artificial Intelligence (AI).

By 1970, the brain simulation research group began losing ground. Although both groups had discovered that as problems got more complex, the computation required by both approaches would grow exponentially and soon reach intractable proportions, funding for brain simulation research all but dried up. There have been a variety of reasons offered as to why the brain simulation group failed, ranging from conspiracy theories about MIT researchers Papert and Minsky trashing their competitors in a book, to flaws in the biology- and psychology-based neural nets that were the foundation of brain simulation. However, Dreyfus and Dreyfus (1988) have a very valid point when they note that Simon and Newell’s physical symbol system more closely resembled the rationalist/reductionist tradition in which all scientists were accustomed to working. In other words, the physical symbol system approach more closely resembled the “normal science” tradition than did the holism of the brain simulation approach.
Returning to Kuhn’s (1962) definition of “normal science” as “research firmly based upon one or more past scientific achievements that some particular scientific community acknowledge for a time as supplying the foundation for its further practice”, the paradigm of Artificial Intelligence defined by Simon and Newell is an excellent fit. It was sufficiently unprecedented as to attract an enduring group of adherents away from competing modes of scientific activity (brain simulation), and was sufficiently open-ended as to leave much to be resolved. Thus, from 1955-1965, Simon and Newell focused on the research themes of representation and search to show how a computer could solve a class of problems using the heuristic search principle of means-end analysis, then incorporated this principle into their General Problem Solver (GPS). The GPS was based on substantive rationality, an intelligent system’s adjustment to its outer environment, and procedural rationality, an intelligent system’s ability, through knowledge and computation, to discover appropriate adaptive behavior. Simon then applied the GPS to decision-making theory, in which the decision-makers’ way of knowing how to make the best decision is grounded in finding a good course of action through selective search. This approach to decision-making still informs most of corporate decision-making.

From 1965-1975, Papert and Minsky at MIT focused on what facts and rules to represent in order to develop methods for dealing systematically with “microworlds”. Computer programs such as SHRDLU, which could follow natural language commands (Winograd and Flores1986) were examples of the several stellar outcomes of AI. As we shall see in the next section, however, its “normal science” status was rather short lived.

Breakdown in the Science of AI: Anomaly Leads to Crisis
In the previous section, I mentioned Papert and Minsky’s pursuit of methods to deal systematically with knowledge in isolated domains called “microworlds”. The intent was to make these microworlds more realistic and combined in order to approach real-world understanding. However, efforts to build various data structures of representation yielded little result. Repeated failures would soon turn the anomaly of representation replication into a full-blown crisis. It appeared that the microworlds approach could not be extended to the real world, creating what Kuhn (1962) called the sense of “malfunction” that is a pre-requisite to crisis and to paradigm shift.
While AI was wrestling with the representation of knowledge, developments in biology, psychology and philosophy theories began to challenge the core concept upon which physical symbol systems-AI was based: namely, that the objective world of physical reality and the subjective mental world of an individual’s perceptions are two separate, but combinable domains. Noted AI researchers Terry Winograd (the author of the natural language SHRDLU program discussed in the previous section) and Fernando Flores began to questions AI assumptions openly. Just as Papert and Minsky’s book had torn down the foundations of Rosenblatt’s brain simulation research some twenty years earlier, Winograd and Flores’ book (1986) attacked the physical symbols system approach, to which they had been key contributors. However, they did so by drawing upon the research from the physical and social sciences to substantiate their attacks, particularly the work of Gadamer, Heidegger, and Maturana.

The common thread among Gadamer, Heidegger, and Maturana is the emphasis on the social and environmental context of human knowledge, with the “microworlds” merely instances of a shared whole of common human concerns. Winograd and Flores summarize Gadamer’s discourse on language and tradition (the hermeneutic circle) and knowing when they state, “what we understand is based on what we already know, and what we already know comes from being able to understand”. (p. 30). They go on to review Heidegger’s rejection of the concept that the objective physical world is the primary reality, and also of the concept that an individual’s perceptions are the primary reality. Instead, Heidegger argued that it is impossible for one to exist without the other, and that objects and properties are not inherent in the world, but arise only in an event of “breaking down” in which they become “present-at-hand”. They summarize Heidegger’s point, stating “what really is is not defined by an objective omniscient observer, nor is it defined by an individual but rather by a space of potential for human concern and action.” (p. 37). Winograd and Flores also draw upon Maturana’s account of linguistic domains - which emphasizes social structure rather than the mental activity of human individuals as the foundation of human action – as further evidence of the common sense knowledge problem of AI.

Clancey (1997) also cites Winograd and Flores’ book and the rise of the connectivist view of the perceptual and physical worlds as heralding the era of questioning of the underlying assumptions of AI. Although Winograd and Flores were not the only AI defectors, they were the best known. Thus, the defection of two authoritative sources severely challenged how AI had been shaping knowing. In an earlier work on AI and cognitive science, Clancey (1991) summed up the main error of AI as supposing that “the interpretation of a representation is known prior to its production. But the meaning of a representation is neither pre-definable nor static; it depends on the observer.” (p. 57). In short, the “normal science” of AI was experiencing what Heidegger called “breakdown”, an interruption to the shared, valued process of getting to the collective truth as defined by the physical symbols systems approach to AI.

The crisis in AI bubbled on well into the 1990s and still exists today, spurring a variety of alternatives to the physical symbols system paradigm. In a review of the prevailing philosophical and technical criticisms of AI, Sterling, Beer and Chiel (1991) note that the AI crisis included ideological perceptions ranging from total denunciation of the science of AI (Winograd and Flores), to a feeling of malaise (McDermott), to a denial of any problem (Hayes). Technical criticisms ranged from connectivism (Smolensky), to a new approach to problem definition (Brooks), to the focus on open systems (Hewitt). None of these alternatives has succeeded in acquiring the collective consensus required to adopt a “best alternative” and move the paradigm dialectic along to a replacement of physical symbol systems “normal science”. Moreover, there have been spectacular successes with high-speed symbol manipulation by knowledge-based expert systems. Clancey (1997) draws particular attention to two successes of the early 1990s. One such success is MYCIN, a program that assists physicians in diagnosing certain types of bacterial blood infections and to determine suitable treatments. The success of MYCIN and other expert systems lies in their ability to make decisions based on thousands of “if-then” logic rules grounded in knowledge gleaned from leading authorities in the field. Another success is Cohen’s “artist” AARON a robot that is able to draw pictures of humans based on this same type of logic pre-programming. These successes have spawned further research of situated robots, with the objective of developing a theory of spatial learning without pre-defining categories in the architecture of the robot. Aiming for gradual measurable success, robot research is now focused on insect and animal behavior. Clancey sees this “downsizing” in the focus of expert systems as a victory, not a defeat:

Nevertheless, there have been no successes where computers have matched human flexibility over wide domains or in tasks requiring “common sense” knowledge. Consequently, AI has fragmented. In addition to physical symbols systems AI has been joined by low-level, microscopic biological models of AI, each with their own approach to shaping knowing and some having found successful practical applications. Munakata (1998) describes this second category of AI as including the following paradigms: 1) neural networks, offering computational models of the brain; 2) genetic algorithms, offering computational models of genetics and evolution; 3) fuzzy systems, a technique of extending concepts to a continuous paradigm, especially for traditionally discrete disciplines; 4) rough sets, a technique for mapping data and discovering relationships in the data. These have served as the theoretical foundation of contemporary data mining software systems; 5) chaos – non-linear deterministic dynamical systems that exhibit sustained irregularity and extreme sensitivity to initial conditions. Fractal Geometry is the most well know practical application of Chaos.

Future Trends/Directions
Few if any in the field of Artificial Intelligence still adhere to the notion that computers will one day be superior to humans. Although there is the potential for the type of havoc wreaked by HAL in Kubrick’s 2001: A Space Odyssey, humans will do exactly what Astronaut Dave did to HAL. As stated so well by Gillies (1996):

Nevertheless, there is no consensus as to what the future of AI really is. Finlay and Dix (1996) focus on the differences between the human mind and the computer, noting that the computer performs poorly in those areas in which the human mind works well (creativity, flexibility, and learning), but perform well in those areas that exceed our mental capacity (memory, speed, and accuracy). Aiello and Nardi (1991) believe that AI is still in the process of becoming a well-defined science, while Flach (1991) contends that AI should now concentrate on the realization of intelligent skills on a computer.

In recognition of the finite nature of funding, Michalski and Lithman (1991) propose that future AI research address the following: 1) how to implement common sense and plausible reasoning; 2) how to develop powerful learning capabilities that are able to take advantage of all kinds of prior knowledge and to explain to humans what was learned; 3) how to introduce and organize large amounts of human knowledge in machines and how to update or improve that knowledge in the process of its normal use, and; 4) how to recognize objects and concepts from incomplete, variable and context-dependent cues. Karakash (1998) advocates using AI research to evolve and refine our understanding of the concept of intelligence, stating that “the notion of natural (human) intelligence has turned out to be dependent on AI. The proof is that progress in the latter constantly incites us to redefine the former. In the AI mirror, the concept of intelligence has become labile, as if the object of our study changed with our research.” (p. 23). In assessing the lack of consensus about the success of AI – other than in expert systems and computer games – Chapman (1991) concludes the following:

Implications for Learning and Knowing
AI as a discipline providing engineering techniques to solve difficult problems in concrete, well-defined application areas has had measurable success. AI as the study of the nature of intelligence and how to reproduce it has been a clear failure. Where the jury is still out is with AI as a means of gaining insight into the nature of human intelligence and learning, rather than as a replica of these. In this area, AI-based agent technology appears to offer a great deal of promise in dealing with the challenges of the 21st Century learning environment. Agent-based tutoring systems have had the most success to date. O’Riordan and Griffith (1999) see the rise of intelligent agents as an attempt to overcome the lack of support for peer-peer learning, the static nature of course content, and the lack of support for personalized learning in web-based education. In programming intelligent agents, Kurhila and Sutinen (1999) note the importance of leaving the responsibility of the pedagogics of the material to human experts creating the learning material. Albrecht, Koch and Tiller (2000) have developed a web-based teaching system that takes into account the spatial and mental abilities of learners by incorporating a great deal of user-system interactivity. Some AI applications to educational systems are constructivist in approach, all focus on collaborative learning.

Although finite funding resources remains an issue for AI educational applications, early successes have kept the funding faucet open. However, research in and applications of AI in education have fallen under the Instructional Technology umbrella, with some researchers refusing to associate themselves with the term Artificial Intelligence. However, it is not yet determined as to what extent these educational applications will lead to either new AI paradigm alternatives or to a “best alternative”. Until then, AI will remain what Ratsch and Stamatescu (1998) call “transdisciplinary, nomadic, if you will, its followers stretching out behind as it migrates from one territory to another, with conceptual or technological breakthroughs in one area providing feedback for all the others.” (p. 25).

References
Albrecht, Florian, Koch, Nora, & Tiller, Thomas. (2000). SmexWeb: An adaptive web-based hypermedia system. Journal of Interactive Learning Research, 11 (3/4), 367 – 388.

Baylor, Amy. (1999). Intelligent agents as cognitive tools for education. Educational Technology, 39 (2), 36 – 40.

Chapman, David. (1991). Vision, instruction and action. Cambridge, MA: The MIT Press.

Clancey, William J. (1997). Situated cognition: On human knowledge and computer representations. Cambridge, UK: Cambridge University Press.

Devedzic, Vladan, Jerinic, Ljubomir, & Radovic, Danijela. (2000). The GETS-BITS model of intelligent tutoring systems. Journal of Interactive Learning Research, 11 (3/4), 411 – 434.

Dreyfus, Hubert L. and Dreyfus, Stuart E. (1988). Making a mind versus modeling the brain: Artificial Intelligence back at a branchpoint. Daedalus Journal of the American Academy of Arts and Sciences, Winter 1988, 15 – 41.

Finlay, Janet & Dix, Alan. (1996). An introduction to artificial intelligence. London: UCL Press.

Flach, Peter A. & Meersman, Robert A. (Eds.). (1991). Future directions in artificial intelligence. Amsterdam, The Netherlands: Elsevier Science Publishers.

Gillies, Donald. (1996). Artificial intelligence and scientific method. New York: Oxford University Press, Inc.

Kurhila, Jaako & Sutinen, Erkki. (1999). Sharing an open learning space by individualizing agents. Journal of Interactive Learning Research, 10 (3/4), 287 – 300.

Munakata, Toshinori. (1998). Fundamentals of the new artificial intelligence: Beyond traditional paradigms. New York: Springer-Verlag New York, Inc.

Ranwez, Sylvie, Leidig, Torsten, & Crampes, Michel. (2000). Pedagogical ontology and teaching strategies: A new formalization to improve lifelong learning. Journal of Interactive Learning Research, 11 (3/4), 389 – 410.

Ratsch, U., Richter, M.M. & Stamatescu, I.-O. (Eds.). (1998). Intelligence and artificial intelligence: An interdisciplinary debate. Heidelberg, Germany: Springer.

Simon, Herbert A. (1996). The sciences of the artificial. Cambridge, MA: The MIT Press.