Intellectual issues in the history of artificial intelligence

This paper sketches the history of artificial intelligence in terms of intellectual issues. These are the usually d ichotomous opposit ions that disciplines seem to generate for themselves in profusion, such as heuristics versus algorithms or declarative versus procedural representations. S u c h a history is complementary to the standard history in terms of the evolving content of the sc ience and the scientists who were responsible for the discoveries and developments. It is certainly no substitute for it. Furthermore, the role of such issues, especially as generated by the field itself rather than by the historian, is not clear. Yet, the picture such a history provides proves interesting. Th is paper does not pretend to be an adequate historiography, but must be taken as a source document prepared by a participant in the field.

clearly. Additional frameworks are then useful. Obvious ones of general applicability are proposed theories and research methodologies. Neither is very satisfactory for AI. The theoretical ideas put forth have, especially when successful, been embedded in computer systems (usually just as programs but sometimes including special hardware). Often, the systems speak louder than the commentary. Indeed, a common complaint of outsiders (and some insiders) is that there is no theory in AI worthy of the name. Whether true or not, such a perception argues against taking theories as the unit in terms of which history is to be written.
As for research methodology, AI as a whole is founded on some striking methodological innovations, namely, using programs, program designs, and programming languages as experimental vehicles. However, little additional methodological innovation has occurred within the field since its inception, which makes for lean history.
Similarly, the more sophisticated units of historical analysis, such as the paradigms of Kuhn or the research programmes of Lakatos, provide too course a grain. [Kuhn, 1962;Lakatos, 1970.] It can be argued that AI has developed and maintained a single paradigm over its short lifetime, or at most two. Similarly, it has contained at most a small handful of research programmes. But units of analysis work best with enough instances for comparative analysis or for patterns to emerge. There are certainly too few paradigms for an internal history of AI. The same is probably still true of research programmes as well, though it would be of interest to attempt such a description of AI.
Useful frameworks for historical analysis can often be based on the organization of subject matter in a field.
AI proceeds in large part by tackling one task after another, initially with programs that can accomplish them crudely, followed gradually by successive refinements. Game playing, theorem proving, medical diagnosis -each provides a single developmental strand that can be tracked. Thus a history of AI as a whole could be written in terms of the geography of tasks successfully performed by AI systems. Almost orthogonal to this task-dimension is that of the intellectual functions necessary for an intelligent system -representation, problem-solving methods, recognition, knowledge acquisition, etc. -what can be termed the physiology of intelligent systems. All these functions are required in any intellectual endeavor of sufficient scope, though they can be realized in vastly different ways (i.e" by different anatomies) and tasks can be found that highlight a single function, especially for purposes of analysis. Thus, a history can also be written that follows the path of increased understanding of each function and how to mechanize it Both of these structural features of AI, and perhaps especially their matrix, provide potentially fruitful frameworks for a history. Their drawback is just the opposite from the ones mentioned earlier, namely, they lead to histories that are almost entirely internal, shedding little light on the connections between AI and neighboring disciplines.
I settle on another choice, which I will call intellectual issues. It is a sociological fact of life that community endeavors seem to polarize around issues -Fluoridation versus Ban Fluoridation, Liberal versus

PAGE 3
Conservative. Such polarizing issues are not limited to the purely political and social arena, but characterize scientific endeavors as well Heliocentrism versus Geocentrism, Nature versus Nurture. Intellectual issues are usually posed as dichotomies, though occasionally three or more positions manage to hold the stage as in the tussle between capitalism, socialism and communism. Intellectual issues are to be distinguished from issues in the real world of action. No matter how complex and ramifying the issues of individual freedom and state control that lie behind a floridation campaign, the passage or defeat of an ordinance banning floridation is a concrete act and is properly dichotomous. But with Nature versus Nurture, the dichotomy is all in the eye of the beholder and the real situation is much more complex (as is pointed out ad nauseum). The tendency to polarization arises from the way people prefer to formulate intellectual issues.
Scientifically, intellectual issues have a dubious status at best. This is true even when they do not have all the emotional overtones of the examples above. Almost always they are defined only vaguely, and their clarity seldom improves with time and discussion. Thus, they are often an annoyance to scientists just because of their sloganeering character. Some time ago, in a conference commentary entitled "You can't play twenty questions with Nature and win", I myself complained of the tendency of cognitive psychology to use dichotomies as substitutes for theories (e.g., serial versus parallel processing, single trial versus continuous learning). [Newell, 1973a.] Intellectual issues surely play a heuristic role in scientific activity. However, I do not know how to characterize it. Nor am I aware of any serious attempts to determine it, though some might exist. Of course, large numbers of scientists write about issues in one way or another, and almost all scientists of an era can recognize and comment on the issues of the day. Were this not true, they could hardly be the issues of the particular scientific day. From a historical and social standpoint, of course, intellectual issues have a perfectly objective reality. They are raised by the historical participants themselves and both their existence and the activity associated with them can be traced. They enter the historical stream at some point and eventually leave at some other.
Whether intellectual issues make a useful framework for a scientific history seems to me an entirely open question. Such a history does not at all substitute for histories based on events and discoveries, laid down within a framework drawn from the substantive structure of a field. Still, ever since that earlier paper in 1973, I have been fascinated with the role of intellectual issues. Recently, I even tried summarizing a conference entirely in terms of dichotomies. [Newell, 1980a.] Withal, I try it here.

The Intellectual Issues
I will actually do the following. I will identify, out of my own experience and acquaintance with the field, all of the intellectual issues that I believe have had some prominence at one time or another. Although I will take the field of AI as having its official start in the mid-1950s, the relevant intellectual issues extend back much earlier. We surely need to know what issues were extant at its birth. I will attempt to put a date both on the start of an issue and on its termination. Both dates will be highly approximate, if not downright speculative. However, bounding the issues in time is important; some issues have definitely gone away and some have come and gone more than once, though transformed each time. I will also discuss some of the major features of the scientific scene that are associated with a given issue. I will often talk as if an issue caused this or that. This is in general illegitimate. At best, an issue is a publicly available indicator of a complex of varying beliefs in many scientists that have led to some result. Still, the attribution of causation is too convenient a linguistic practice to forego. Table 1 lays out the entire list of intellectual issues. In addition to the short title of the issue, expressed as a dichotomy, there is an indication of an important consequence, although this latter statement is necessarily much abbreviated. Tne issues are ordered vertically by date of birth and within that by what makes historical sense. All those born at the same time are indented together, so time also moves from left to right across the figure; except that all the issues on hand when AI begins in 1955 are blocked together at the top. Issues that show up more than once are multiply represented in the table, according to the date of rebirth, and labeled -#L #2, etc. When the ending date is not shown (as in Reason versus Emotion and Feeling #1: 1870-), then the issue still continues into the present The issues will be discussed in historical order, i.e., according to their order in the table. This has the advantage of putting together all those issues that were animating a given period. It has the disadvantage of mixing up lots of different concepts. However, since one of the outcomes of this exercise is to reveal that many different conceptual issues were coexisting at any one time, it seems better to retain the purely historical order.
Mechanism versus Teleology: 1640-1945. We can start with the issue of whether mechanisms were essentially without purpose. This is of course the Cartesian split between mind and matter, so we can take Descartes, as the starting point It is an issue that can not be defined until the notion of mechanism is established. It is and remains a central issue for AI, for the background of disbelief in AI rests precisely with this issue. Nevertheless, I place the ending of the issue with the emergence of cybernetics in the late 1940s. If a specific event is needed, it is the paper by Rosenbleuth, Wiener and Bigelow, which puts forth the cybernetic thesis that purpose could be formed in machines by feedback. [ posit a separate issue of mechanism versus intelligence to contrast with mechanism versus teleology. No such distinction ever surfaced. Instead, there is an underlying concern about the aspects of mentality that can be exhibited by machines. This shows itself at each historical moment by denying to machines those mental abilities that seem problematical at the time. Thus, the argument moves from purpose in the 1940s to intelligence in the 1950s. With the initial progress primarily in problem-solving, one occasionally heard in the 1960s statements that machines might solve problems but they could never really learn. Thus, the basic issue simply continues, undergoing continuous transformation.
Natural Biology versus Vitalism: 1800-1920. A critical issue for AI that had come and gone long before AI really began is the issue of vitalism -do living things constitute a special category of entities in the world, inherently distinct from inanimate physical objects. As long as this issue was unsettled, the question of whether the mind of man was mechanical (i.e., non-special) was moot It is difficult to conceive of concluding that the animate world does not generally obey the laws of the physical world, but that the mind is an exception and is entirely mechanical. Thus, only if vitalism has been laid to rest for our bodies can the issue be joined about our minds.
The vitalist controversy has a long and well chronicled history. Retrospectively, it appears as an inexorable, losing battle to find something special about the living, though the issue was joined again and again. Organic matter was just a different kind of matter from inorganic matter -an issue laid to rest finally with the synthesis of urea, an indisputably organic material, from inorganic components in 1828 by Wohler.
Organisms had their own inherent internal heat an issue laid to rest in by the work of Bernard by the mid-1800s. For our purposes, the starting and ending dates of the issue are not critical. Vitalism's last champion may be taken to be the embryologist Hans Driesch at the turn of the century, who proposed that organisms develop only by virtue of nonmaterial vital principles, called entelechies. [Driesch, 1914.] Issues almost never die of course, as the continued existence of the Flat Earth Society should remind us.
Nevertheless, no substantial intellectual energy has been focused on vitalism in more than fifty years. That the human body is a physical machine, operating according to understood physical laws and mechanisms, sets the stage for the consideration of the mechanistic nature of thought and intelligence.
Reason versus Emotion and Feeling #1: 1870-. The basic separation of the heart from the head occurred long ago and is a fundamental part of Christian folk psychology. It is background. What concerns us is the ascription of reason ("cold logic") to machines and the belief that a machine could have no heart -no feelings or emotions -to even conflict with its reason. I do not seem to find any good way to fix the initiation of this issue. The striking characteristic of the golem of Rabbi Loew in 1580 seemed to have been literalmindedness, not heartiessness. And nineteenth-century "artificial humans" seemed to combine all the human attributes, as did, for instance, Frankenstein's constructed monster. [Shelley, 1818.] But by the twentieth century, certainly in "R.IJ.R. (Rostrum's Universal Robots)", we clearly have the intelligent robot, who is without soul, hence without emotions or independently felt wants. [Capek, 1923.] So I have split the latter two dates and taken 1870 as the start The relevance of this for AI is in providing a basis for separating machines from humans that is different from the issue of purpose. Although a birthright issue of AI, it does not play a major role. That the issue is there can be seen clearly enough in the paper on "Hot Cognition", by Abelson, which put forth some proposals for how to move machine intelligence in the directionjof having affect [Abelson, 1963.] The lack of prominence stems in part, no doubt from the strong engineering-orientation of AI, which emphasizes useful mental functions (e.g., problem-solving and learning). In agreement with this, Abelson is one of the few social psychologists associated with AI and the paper was given at a psychology conference. Thus, this issue remains in the background, waiting to become prominent at some future time.  [Watson, 1913.] Thus, this issue emerged and vanished before AI began. The residue was a continuing tradition in philosophy concerned with mind that was completely distinct from work in psychology and, even more so from technology. This issue ensured that when AI did emerge, which happened instantly upon computers becoming sufficiently powerful, 2 it would be without more than peripheral involvement of the philosophy of mind.
Logic versus Psychologic: 1910-45. We continue to lay out the issues -and their resolutions that were in effect at the birth of AI. This issue concerns whether symbolic logic was to be taken as revealing how humans think or whether humans use some sort of unique "psychologic". It surely started out with logic identified with thought, as Boole's classic monograph entitled "The Laws of Thought" testifies. [Boole, 1854.] But logic U^as rapidly transformed from an explication of the possible varieties of thinking to a device for probing the foundations of mathematics. We can take the Principia Mathematica of Whitehead and Russell as marking the completion of this. [Whitehead andRussell, 1910-1913.] The effect was to separate logic from psychology (and also from the philosophy of mind, although that is a more complex story).
Modern logic, of course, was integrally involved in the development of the digital computer and thus it enters into the history of AI. But logic did not enter AI at all as the logic of thought. That separation remained. Logic was part of the underlying technology of making mechanisms do things. In fact, it was precisely the split of logic from thought that set logic on die path of becoming a science of meaningless tokens manipulated according to formal rules, which in turn permitted the full mechanization of logic.
Thus the issue was really settled by 1910 and the status in the first half of the century was that psychologic was not a significant item on the agenda of any science. This, of course, was due to behaviorism's restriction of psychology's agenda. I have placed a date of 1945 for the ending of this issue. This is really an ending of the phase of separating logic from thought Thenerve-net model of McCulloch and Pitts can be used to mark this, along with the work of Turing on which it depended. (Turing, 1936;McCulloch and Pitts, 1943.] They attempted to show that physical systems that echo the structure of the brain could perform all computations, which is to say, all logical functions. Whether this is seen as saying more about the brain or more about logic can be argued; in either case it brought them back into intimate contact One might think that the ending of one phase of the issue (the stable separation of logic from thought) should initiate a new phase, namely, a new controversy over the exact nature of the connection. But it did not happen that way. Rather, the issue was not discussed and the basic questions about the mechanization of mind took the form of other issues. The reason that happened cannot be explored here. In part it comes from the shift, with AI, away from the characterization of the brain in computational terms, over to the digital computer, where logic played a completely technical and engineering role in describing sequential and combinational logic circuits.
'A case can be made that serious AI started as soon a computers attained 4K of random-access primary memory.
Analog versus Digital: 1940-70. When computers were first developed in the 1940s, they were divided into two large families. Analog computers represented quantities by continuous physical variables, such as current or voltage; they were fast, operated simultaneously, and had inherently limited accuracy. Digital computers represented quantities by discrete states; they were slow, operated serially, and had inherently unlimited accuracy. There was a certain amount of skirmishing about which type of computer was better for which type of job. But the technical opinion-leaders maintained a view of parity between the two families -each for its own proper niche. Inevitably, there arose hybrid computers, which claimed to have the best of both worlds: digital control and memory, coupled with analog speed and convenience.
It was all over by 1970. The field of computers came to mean exclusively digital computers. Analog systems faded to become a small subpart of electrical engineering. The finish was spelled not just by the increased speed and cost-efficiency of digital systems, but by the discovery of the Fast Fourier Transform, which created the field of digital signal processing and thus penetrated the major bastion of analog computation. The transformation of the field is so complete that many young computer scientists hardly know what analog computers are.
The main significance of this issue, with its resolution, was to help create the discipline of computer science and to separate it from electrical engineering. Its effect on AI lies mostly in the loss of an analytical point of view, in which the contrast between analog and digital computation is taken as the starting point for asking what sort of information processing the nervous system does. An admirable example of this point of view can be seen in the notes for von Neumann's Silliman Lectures, published posthumously, [von Neumann, 1958.] This style of analysis belongs to the world of cybernetics and not to that of AI. I doubt if many young AI scientists have read von Neumann's little book, though it was highly regarded at the time and von Neumann was one of the towering intellects of the computer field.
Symbols versus Numbers: 1955-65. We now come to the first of the issues that characterizes AI itself, as opposed to the background against which it emerged. The digital-computer field defined computers as machines that manipulated numbers. The great thing was, they said, that everything could be encoded into numbers, even instructions. In contrast, the scientists in AI saw computers as machines that manipulated symbols. The great thing was, they said, that everything could be encoded into symbols, even numbers. The standard measure of a computation at the time was the number of multiplications it required. Researchers in AI were proud of the fact that there were no multiplications at all in their programs, though these programs were complex enough to prove theorems or play games. The issue was actively pursued as a struggle over how. the computer was to be viewed. However, it was joined in an asymmetric way. The bulk of the computer field, and all its responsible opinion-leaders, simply adopted the view that computers are number manipulators. There was no attempt to argue against the view that computers are symbol manipulators. It was just ignored and the standard interpretation maintained. The AI researchers, on the other hand, were actively engaged in promoting the new view, considering the standard one to be a radical misreading of the nature of the computer and one that provided a significant barrier against the view that computers could be intelligent The result of this clash of views was to isolate AI within computer science. AI remained apart of computer science, but one with a special point of view that made it somewhat suspect, indeed somewhat radical. This isolation is important historically, for it has affected the professional and disciplinary organization of the two fields. It derives ultimately, no doubt, from a basic divergence of views about whether computers can or cannot exhibit intelligence. This overarching issue, of course, continued to be important one its own, as witnessed by the debates that occurred throughout the 1950s on whether machines could think. But the more specific issues that it spawned also had independent lives.
The issue of symbols versus numbers did not arise until after the first AI programs came into existence, circa 1955. Before that time, programs were classified as numerical versus nonnumerical. This latter class was a miscellany of all the things that processed data types other than numbers -expressions, images, text, etc. 3 This included the few game playing and logic programs, but much else as well. The symbols versus numbers issue emerged only when a positive alternative became formulated, i.e., symbolic manipulation. This was not a synonym for nonnumerical processing, for it laid the groundwork for the separation of image-and textprocessing from AI. Indeed, the work on machine j^anslation, which started in the early 1950s, was initially considered as one strand in the development of intelligence on machines. [Locke and Booth, 1955.] But that effort became concerned with text and not symbols, and developed its own identity as computational linguistics. (All of this, of course, was before text processing in its current meaning emerged -an event that bore no significant relation to the development of computational linguistics.) I have placed the ending of this issue at about 1965, although I do not have a significant marker event for its demise. The issue is certainly not alive now and has not been for a long time. In part, this is due to the prominence of many nonnumerical data types in computer science generally, such as text and graphics. These make the characterization of computers as number manipulators no longer ring true. In part, it is due to the shift in theoretical computer science to algebraic and logical formalisms, with the concurrent retreat of numerical analysis from its early dominant role. In part, of course, it is due to the success of AI itself and the demonstrations it brought forward of the symbolic character of computation. It is tempting to say that the cause was simply the growth of scientific understanding ~ but such reasons do not fare well in historical ^Thfi concept of data type did not arrive in clear form until much later.
accounts, in any event, my recollection is that the symbols-numbers issue was no longer prominent by the late 1960s,.though a little historical digging might place it five years later.
Symbolic versus Continuous Systems: 1955-. An important characterization of a science, or an approach within a science, is the class of systems it uses to construct its theories. Classical physics, for instance, viewed systems as being described by systems of differential equations. Given a new phenomenon to be explained, a physicist automatically, without a thought, used differential equations to construct his theory of that phenomenon. Mathematical psychology in the 1950s and 1960s could be characterized by its acceptance of Markov processes as the class of systems widiin which to seek theories of particular phenomena.
The issue is within what class of systems should a description of intelligent systems be sought On one side were those who, following the lead of physical science and engineering, adopted sets of continuous variables as the underlying state descriptions. They adopted a range of devices for expressing the laws -differential equations, excitatory and inhibitory networks, statistical and probabilistic systems. Although there were important differences between these types of laws, they all shared the use of continuous variables. The other side adopted the programming system itself as the way to describe intelligent systems. This has come to be better described as the class of symbolic systems, i.e., systems whose state is characterized by a set of symbols and their associated data structures. But initially it was simply the acceptance of the programs per se as the theoretical medium.
Adopting a class of systems has a profound influence on the course of a science. Alternative theories that are expressed within the same class are comparable in many ways. But theories expressed in different classes of systems are almost totally incomparable. Even more, the scientist's intuitions are tied strongly to the class of systems he adopts -what is important what problems can be solved, what possibilities exist for theoretical extension, etc. Thus, the major historical effect of this issue in the sixties was the rather complete separation of those who thought in terms of continuous systems from those who thought in terms of programming systems. The former were the cyberneticians and the engineers concerned with pattern recognition. The latter became the AI community. The separation has been strongly institutionalized. The continuous-system folk ended up in electrical engineering departments; the AI folk ended up in computer science departments.
(It must be remembered that initially computer science departments were almost exclusively focused on software systems and almost all concern with hardware systems was in electrical engineering departments.) I believe this issue largely explains one peculiar aspect of the organization of the science devoted to understanding intelligence. By almost any account pattern recognition and AI should be a single field, whereas they are almost entirely distinct. By now, in fact due to another important historical twist* many people in computer science work in pattern recognition. But if such persons also know traditional pattern recognition, they are seen as interdisciplinary.
Another interesting implication is buried here. The issue is not properly dichotomous, for there exist other classes of systems within which to search for intelligent systems. One obvious candidate is logic. 4 Were there not scientists who believed that logic was the appropriate class of systems? And if not, why not? First, by "logical systems" is meant the class of systems that do logical operations, such as AND, OR, NOT,etc. 5 This is the class corresponding to the logical level in the hierarchy of computer structures. The logic level is located between the circuit level and the program (symbol) level. All three levels are equally comprehensive and provide, three possibilities for ways to describe intelligent systems. Indeed, the circuit and program levels correspond exactly to the continuous and symbol positions of the issue under discussion. Now in fact, in the early days there were attempts to build "logic machines" and to discuss the behavior of systems directly in terms of logic circuits. The classical neural networks of McCulloch and Pitts were an effort at modeling the neural system at the logic level. [McCulloch and Pitts, 1943.] But all these efforts rapidly died out and were all but gone by the mid-1960s. My own guess about why this happened is that the hierarchy of computer levels indicated quite clearly what to do with a logic level -namely, compose a higher-level system. But this implied simply reproducing existing program-level systems, at least without some new organizational ideas at the program level. But the logic level provided no such ideas, nor could it. Thus, there was nowhere to go. In fact, the history of these efforts seems quite obscure, and tracing the demise of logic as a system language for intelligent systems would be a substantial, though rewarding, undertaking.
Problem-Solving versus Recognition #1: 1955-65. An interesting issue grew up in association with the continuous/symbolic split Those thinking within the framework of continuous systems concentrated on pattern recognition as the key type of task for machines to do -character recognition, speech recognition, and visual-pattern recognition. They also often concentrated on learning (as noted below), but it was almost always a recognition capability that was being learned. The PERCEPTRON of Rosenblatt can be taken as paradigmatic here. [Rosenblatt, 1956.] Contrariwise, those thinking within the framework of symbolic systems concentrated on problem-solving as the key type of task for machines to do --game playing, theorem proving and puzzle solving.
This separation of tasks reinforced the split between these groups. To the AI community the intellectual depth of the tasks performed by the pattern-recognition systems seemed relatively trivial compared with the problem-solving tasks done by the programming systems. But just because of that, a myth grew up that it was In fact, there are additional possibilities. [Newell, 1970.] 5 It might also mean the class of theorem-proving systems using logical calculi; but this is really a subclass of symbol systems.
relatively easy to automate man's higher reasoning functions, but very difficult to automate those functions man shared with the rest of the animal kingdom and did well automatically, to wit, recognition. Thus, work on recognition was at the foundation of the problem of intelligence, whereas work on problem-solving was an add-on.
The symbolic/continuous split and the problem-solving/recognition split are organically related. Each task is the one mostly easily approached in terms of the class of systems adopted. However, that does not make the two intellectual issues the same. Scientists can hold quite different attitudes about the two splits, and the two issues can become uncoupled in a different era under different conditions. Both these issues emerged in the late 1950s, concurrently with the birth of AI. By 1965 the two fields of AI and pattern recognition had separated rather completely and taken up distinct, relatively permanent institutional roles. The conflict could be considered to have reached a resolution. However, it was to become unstuck again almost immediately.
Psychology versus Neurophysiology #1: 1955-65. Strongly coordinated with the issues of symbolic versus continuous systems and of problem-solving versus recognition was another, -conceptually distinct issue, namely whether AI would look to psychology or to neurophysiology for inspiration. That human intelligence is XJO^ bejjgth guide and goad to the engineering of intelligent systems was clear. However this did not discriminate between psychology and neurophysiology. As is well known, these two disciplines speak with entirely separate, though not necessarily contradictory, voices. In general, those concerned with continuous systems and pattern recognition looked to neurophysiology; those concerned with symbolic systems and problem-solving (i.e., AI) looked to psychology. Evidence of the exclusive attention of early AI to psychology (in contradistinction to biology) is amply provided by the two major sets of readings of those years.
[ Feigenbaum and Feldman, 1963;Minsky, 1968.] By 1965 this issue was no longer a live one and the cast of AI was set The split between neurophysiology and psychology did not dictate the split between symbolic and continuous systems. If anything, it was the other way round. Neurophysiology, of course, was linked to continuous variables, with its signals, networks, and geometry. But experimental psychology was not linked at all to symbolic systems. The dominant class of systems in psychology at the time was that of stimulus/response (S/R) systems, an abstract form of inhibition-and-excitation network. The only I alternatives were the continuous fields of Gestalt theory or the pseudo-hydraulic systems of Freudian I psychology (both only vaguely defined, though that is irrelevant here). In fact, the class of symbolic systems was discovered within AI and imported into psychology. [Newell and Simon, 1976;Newell, 1980b.] Thus, the choice of psychology by AI was made because the class of systems that AI took to work with, i.e ; , programming systems, led to psychologically, not physiologically, revealing tasks.
Neurophysiology played a key role in keeping continuous systems from suffering the same fate as logic systems. Whereas with logic systems there was nowhere to go except towards program-like organizations, with continuous systems there was the brain to model. One need not demand an answer to what the higher organization would be, one could just take as guide the brain as revealed in current neurophysiological work.
It is true, of course, that in the late 1940s and early 1950s, the discrete approximation to the nervous system (neurons as digital threshold devices) promised to provide neurophysiological inspiration for the class of logic systems. But under a barrage of criticism, even the engineers came to accept the nervous system as too complex to be modeled by logic-level systems, which is to say, its continuities had to be taken seriously. Thus, without any source of inspiration, the logic-level systems faded away as a separate language for modeling intelligence, but the continuous systems remained.
Performance versus Learning #1: 1955-65. Yet another issue can be identified that coordinates with the issue of symbolic versus continuous systems. AI concentrated on creating performance systems, i.e., systems that performed some task demanding intelligence. Cybernetics and pattern-recognition research concentrated on creating systems that learned. Indeed, another subfield grew up that called itself self-organizing systems. [Yovits, Jacobi and Goldstein, 1962.] In practice, it largely overlapped with the work in pattern recognition and it had common roots in cybernetics. But self-organizing systems took the problem of learning as the central focus rather than the problem of recognition. For instance, within self-organizing systems there was considerable interest in embryology, even though it had little to do with recognition at the time.
Through the early 1960s all the researchers concerned with mechanistic approaches to mental functions knew about each other's work and attended the same conferences. It was one big, somewhat chaotic, scientific happening. The four issues we have identified ~ continuous versus symbolic systems, problem-solving versus recognition, psychology versus neurophysiology, and performance versus learning -provided a large space within which the total field sorted itself out. Workers of a wide combination of persuasions on these issues could be identified. Until the mid-1950s the central focus had been dominated by cybernetics, which had a position on two of the issues -using continuous systems and orienting toward neurophysiology ~ but had no strong position on the other two. For instance, it did not concern itself with problem-solving at all. programs that operated by means of heuristic rules of thumbs --approximate, partial knowledge that might aid the discovery of the solution, but could not guarantee to do so. The distinction implied that intelligent problem-solving could be attained by heuristic programs. For a short while, one name for the field of AI was "heuristic programming", reflecting in part a coordination with subfields such as linear and dynamic programming (which w$re also just then emerging).
An important effect of this issue was to isolate AI within computer science, but along a different dimension than the issue of symbols versus numbers. Heuristic programming indicates a commitment to a different course than finding the best engineering solution or mathematical analysis of a problem. According to the standard engineering ethos, the proper use of the computer requires the engineer or analyst to exert his best intellectual efforts studying the problem, to find the best solution possible, and then to program that solution.
Providing a program with some half-baked, unanalyzed rules seemed odd at best, and irrational, or even frivolous, at worst. A good example of this tension can be found in the work of Wang, whose theoremproving program performed much better than the LOGIC THEORIST. [Newell, Shaw and Simon, 1957; Wang, I960.] The thrust of Wang's position was that much better theorem-provers could be built if the appropriate results in mathematical logic were exploited. The defense by the AI community stressed finding how humans would solve such problems, in effect denying that the fullest analysis of the experimental tasks was the object of the investigation. Another important example was the MACSYMA project to construct an effective computer system for physicists and engineers to do symbolic manipulation of mathematical expressions. Although this work grew out of two prior efforts in AI, it was cast by its leaders as "not part of AI", but rather as part of an area of computer science called symbolic manipulation, which took a thoroughgoing engineering and analytical attitude. [Slagle, 1963;Moses, 1967.] I have put the demise of the issue at the mid-1960s. The issue simply gradually ceased to be discussed, though the distinction continues to be made in textbooks and introductory treatments. Once the field was underway, with lots of AI systems to provide examples, the point at issue becomes transparent. Moreover, the distinction has difficulty becoming transformed into a technical one, because it is tied to features external to the procedure itself, namely, to the problem that is supposed to be solved and to the state of knowledge of the user of the procedure.
. Interpretation versus Compilation: 1955-85. A third issue served to separate AI from the rest of computer science, in addition to the issues of symbols versus numbers and heuristics versus algorithms. AI programs were developed in list-processing languages, which were interpretive, whereas the mainsteam of language development was moving irrevocably toward the use of compilers. Prior to the mid-1950s, programming languages beyond assemblers, were interpretive. The major turning-point in compilers, FORTRAN, was developed in the mid-1950s 6 and it determined the direction of programming-language development (though of course not without some controversy). Speed of execution was the consideration uppermost in the minds of the programming fraternity. In contrast, AI took the interpretive character of its languages seriously and declared them to be necessary to attaining intelligent systems. This was epitomized by the use of full recursion, but it penetrated throughout the entire philosophy of language-design with the attractive idea of putting intelligence into the interpreter.
This separation of AI programming from main-line high-level language programming, which started immediately at the birth of AI, has persisted to the present Its effects go much deeper than might be imagined. It has played a major role in determining the heavy AI involvement in interactive programming, which contrasts with the minimal involvement of the central programming-languages, with their adherence to the compile-and-run operating philosophy. Just for fun I have indicated the end of this issue in 1985, on the assumption that the coming generation of powerful personal computers will finally force all languages to come to terms with full dynamic capabilities in order to permit interactive programming. But this is pure conjecture and the separation may now be wide enough to require a generation to heal.
The grounds for this issue can be traced to demands for efficiency on the one hand versus demands for flexibility on the other. Perhaps, the issue should have been so labeled. For instance, the main programming community in the late 1950s also had a strong negative reaction to list-processing, because of its giving up half the memory just to link the "actual" data together. But, although the general efficiency issue was always on the surface of the discussions, the total situation seems better described in terms of distinct structural alternatives, i.e., interpreters versus compilers, list structures versus arrays, recursion versus iteration. Simulation versus Engineering Analysis: 1955-. One issue that surfaced right from the start of AI was whether to make machines be intelligent by simulating human intelligence or by relying on engineering analysis of the task. Those who were primarily trying to understand human intelligence inclined naturally to the simulation view; those who were primarily engineers inclined to the pure task-analysis view. The principle was frequendy invoked that you do not build a flying machine by simulating bird flight On the simulation side there was more than one position. The majority took the view that casual observation and casual introspection was the appropriate approach -i.e., the human was a source of good ideas, not of detail.
A few, usually with strong psychological interests or affiliations, took the view that actual experimental data on humans should be examined.
This issue seems never to have produced any important crises or changes of direction in the field.
However, it probably has decreased the amount of mutual understanding. There seems to be little movement of a scientist's position on this issue. Each investigator finds his niche and stays there, understanding only superficially how those with different approaches operate. The position adopted probably reflects fairly deep attitudes, such as determine whedier a scientist goes into an engineering discipline or a social/behavioral discipline in the first place. This is to be contrasted with many fields where methods are effectively neutral means to ends, to be used by all scientists as the science demands. There is little indication of diminution of this issue over the years, although starting in the 1970s there has been some increase in the general use of protocols to aid the design of AI systems, even when there is no psychological interest This completes the set of new issues that arose coincident with the birth of AI. Five of them ~ symbolic versus continuous systems, problem-solving versus recognition, psychology versus neurophysiology, performance versus learning and serial versus parallel -separated AI from other endeavors to mechanize intelligence. But the goal of mechanizing intelligence bound all of these enterprises together and distinguished them from the greater part of computer science, whose goal was performing tasks in the service of man. Three issues ~ symbols versus numbers, heuristics versus algorithms, and interpreters versus compilers -cluster together to make AI into a relatively isolated and idiosyncratic part of computer science.
Finally one -simulation versus engineering -was purely internal to AI itself.
Replacing versus Helping Humans: 1960-. An issue that surfaced about five years after the beginning of AI was whether the proper objective was to construct systems that replace humans entirely or to augment the human use of computers. The fundamentally ethical dimension of this issue is evident. Yet it was not overtly presented as an issue of social ethics, but rather as a matter of individual preference. An investigator would simply go on record one way or another, in the prefaces of his papers, so to speak. Yet there was often an overtone, if not of ethical superiority, of concordance with the highest ideals in the field. Those whose inclinations were towards AI, did not so much meet this issue head on, as ignore it Indeed, it was perfecdy possible to take the view that work in AI constituted the necessary exploration for man-machine symbiosis. [Licklider,I960.] A relatively weak issue, such as this, could not really become established unless man-machine cooperation offered as exciting technical possibilities and challenges as constructing intelligent machines. Thus, the beginning of this issue coincides with the appearance of interesting interactive systems, such as SKETCHPAD, which had an immense influence on the field. [Sutherland, 1963.] AI scientists have had a relatively large involvement in the development of user-computer interaction throughout the history of computer science, for example, in time-sharing in the 1960s and 1970s, in making languages interactive in the 1970s, and in the development of personal machines in the early 1980s. One explicit justification given for this involvement was that AI itself needed much better programming tools to create intelligent programs. This reason is quite independent of the issue presented here. However, it is not possible to untangle the relations between them without some rather careful historical analysis.
Many of those who opted for working in user/computer cooperation tended not to become part of AI, as the latter gradually evolved into a field. However, as we have noted above, it was entirely possible to work both in AI and in user/computer cooperation. Still, the net result was an additional factor of separation between those in AI and those in neighboring parts of computer science.
Epistemology versus Heuristics: I960-. It is easy to distinguish the knowledge that an intelligent agent has from the procedures that might be necessary to put that knowledge to work to exhibit the intelligence in action. 7 The initial period in AI was devoted almost exclusively to bringing into existence modes of heuristic processing worthy of consideration. In 1959 John McCarthy initiated a research position that distinguished such study sharply from the study of appropriate logical formalisms to represent the full range of knowledge necessary for intelligent behavior. [McCarthy, 1959.] This study was clearly that of epistemology -the study of the nature of knowledge. It bore kinship with the sub field of philosophy by the same name, although, as with so many other potential connections of AI and philosophy, the orientation of the two fields is highly divergent, although the domain of interest is nominally the same.
There has been little controversy over this issue, although the two poles lead to radically different distributions of research effort Work on epistemology within AI has remained extremely limited throughout, although recendy there has been a substantial increase. [Bobrow, 1980.] Search versus Knowledge: 1965-80. In the first years of AI, through the early 1960s, AI programs were characterized simply as highly complex programs, without any particular notion of common structure. For instance, the field was also called "complex information processing" as well as "heuristic programming". By 1965, however, it had become clear that the main AI programs used the same fundamental technique, which became known as heuristic search. [Newell and Ernst, 1965.] This involves the formulation of the problem to be solved as combinatorial search, with the heuristics cast in specific roles to guide the search, such as the selection of which step to take next, the evaluation of a new state in the space, the comparison of the present state to the posited goal-state, and so on. As the scope of AI programs seemed to narrow, there arose a belief by some AI scientists that the essence of intelligence lay not in search, but in large amounts of highly specific knowledge, or expertise. This issue was well enough established by the mid-1970s to occasion the declaration that a paradigm shift in AI had already happened, the original paradigm having been heuristic search with little knowledge of the task domain, and the new paradigm being knowledge-intensive programs. [Goldstein and Papert, 1977.] It may be doubted that these changes amounted to an actual paradigm shift. What clearly did happen was a major expansion of AI research to explore systems that included substantial domain-specific knowledge. The subfield currently called "expert systems", which includes many of the attempts at constructing applied AI systems, emerged in the mid-1970s in part as a result of this emphasis. However, it became clear that heuristic search invariably continued to show up in these programs. Whenever it did not, the problems being solved by the AI system were extremely easy relative to the knowledge put into the system.
It is useful to see that two types of searches are involved in intelligence. The first is the search of the problem space, i.e., heuristic search, which is combinatorial. The second is the search of the system's memory for the knowledge to be used to guide the heuristic search. This memory search is through a pre-existing structure that has been constructed especially for the purpose of being searched rapidly; it need not be combinatorial. Both types pf searches are required of an intelligent system and die issue of search versus knowledge helped to move the field to a full consideration of both types. The net result was not so much a shift in the paradigm as a broadening of the whole field. This had become clear enough to die field so that by 1980 the issue can be declared moot Power versus Generality: 1965-75. Another way to characterize the major early AI programs is that they took a single well-defined difficult task requiring intelligence, and demonstrated that a machine could perform it. Theorem-proving, chess and checker playing, symbolic integration, management-science tasks such-as assembly-line balancing, IQ-analogy tasks -all these fit this description. Again, a reaction set in to this. Although AI could do these sorts of tasks, it could not do the wide range of presumably trivial tasks we refer to as "having common sense". The need was for generality in AI programs, not power.
This call had been issued early enough. [McCarthy, 1959.] However, it was really not until the mid-1960s that a significant shift occurred in the field toward the generality and commonsense side. This gave rise to using small constructed puzzles and artificial problems to illustrate various components of everyday reasoning. A typical example was the monkey-and-bananas task, patterned after the simple tasks solved by Kohler's chimpanzee, Sultan. Whereas such problems would have seemed insignificant in the early years, they now became useful because the goal of research was no longer power, but understanding how commonsense reasoning could occur.
By 1975 this shift had run its course and new concerns for working with relatively large-scale real problems took over, with the development of expert systems mentioned above. As could have been expected, the end of this period of emphasis did not mean a shift back to the original power pole of the issue. Expert systems, although they tackled real problems and hence were obviously "powerful", did not achieve their power by the heuristic-search techniques of the early years; instead they used large amounts of domain-specific knowledge (coupled, sometimes, with modest search).
However, as is usual in the history of science, work on powerful AI programs never stopped; it only diminished and moved out of the limelight By 1975 highly successful chess programs emerged, built on heuristic-search principles, with an emphasis on large amounts of search -a million positions per move in tournament play -and good engineering. Thus, intellectual issues shift the balance of what gets worked on, but rarely shut off alternative emphases entirely.
Competence versus Performance: 1965-. The Chomskian revolution in linguistics also started in the late 1950s. It was, along with AI, just one of many similar and interrelated developments in engineering, systems and operational analysis. Although each of these developments had a particularly intense significance for some particular field, e.g., for linguistics or computer science, they all formed a common interdisciplinary flux. Gradually, these activities sorted themselves out into separate subfields or disciplines, developing opposing positions on the issues laid out above, as we have seen for AI vis-a-vis cybernetics and pattern recognition.
In many ways linguistics was a special case. It was already a well-formed discipline and the revolution was at the heart of the discipline, not in some peripheral aspect that could have split off and aligned with other intellectual endeavors. Furthermore, only a very few linguists participated in the general flux that was occurring in the world of engineering and applied mathematics. Linguistics was culturally and organizationally quite distinct, having strong roots in the humanities. In fact, it probably made an immense difference that Chomsky located at the Massachusetts Institute of Technology (MIT).
It took until the mid-1960s before the issues emerged that determined the relations between linguistics and the other sub fields and disciplines. A principal one was the distinction between competence and performance, which was moved to a central position in the new linguistics by Chomsky. [Chomsky, 1965.] Linguistic competence was the general knowledge a speaker had of the language, in particular of the generative grammar of the language. Performance was the actual production of utterances, which could be affected by many additional factors, such as cognitive limits, states of stress, or even deliberate modifications for effect The distinction made useful operational sense for linguistics, because there were two sources of evidence about human-language capabilities, the actual utterance and the judgment of grammaticality -a sort of a recognition/recall difference, although that analogy was never exploited.
This distinction might seem innocuous from the standpoint of science history, i.e., purely technical. In fact it served to separate quite radically the sciences concerned primarily with performance, namely AI, computational linguistics, cognitive psychology, and psycholinguistics, from linguistics proper. Linguistics itself declared that it was not interested in performance. More cautiously said, competence issues were to have absolute priority on the research agenda. But the effect was the same: work in any of the performance fields was basically irrelevant to the development of linguistics. There could be a flow from linguistics outward to these other fields. And indeed there was an immense flow to psycholinguistics. But there could not be any significant flow in the other direction. 8 A more effective field-splitter would be hard to find. It has remained in effect ever since, with the competence/performance distinction being extended to other domains of mentality. This has certainly not been the only significant cause of the sepaiateness of AI from linguistics. There are important isolating differences.in method, in the style of research, and in the attitudes towards evidence. Many of these other issues share substance with the competence/performance distinction, and affect the separation between psychology and linguistics much more than that between AI and linguistics. Thus, perhaps they can be left to one side.

Memory versus Processing: 1965-75.
During the immediate postwar decades, the mainstream of individual human psychology was strongly influenced by the general ferment of engineering, system and operational ideas (as we termed it above). This involved human factors and information theory in the early 1950s; and signal detection theory, control theory, game theory and AI in the mid-1950s. As with linguistics, in the period of 1955-65, all these ideas and all these fields seemed to mix while matters sorted themselves out By the mid-1960s, psychology had focused on memory as the central construct in its view of man as ah ^is is not the whole story of the relations of linguistics with other fields; e.g., there have been important contacts with logic and with philosophy.
information-processor. Short-term memory and the visual iconic store combined to provide an exciting picture of the interior block-diagram of the human mental apparatus (what would now be called the architecture). This settled what the main lines of investigation would be for the field. The marker event for this conviction is Neisser's book on "Cognitive Psychology". [Neisser, 1967.] This is important for the history of AI, because AFs influence on psychology in the 1955-1965 period was primarily in the area of problem-solving and concept formation. With psychology opting for memory structure, psychology and AI went fundamentally separate ways. Although the work on problem-solving remained a common concern, it was a sufficiently minor area in psychology that it exerted only a modest integrating effect. AI itself during this period had little interest in memory structure at the block-diagram level. Psychologically relevant research on memory by AI researchers did exist, but moved out of AI into psychology, for example the work on EPAM (Elementary Perceiver and Memorizer). [Simon and Feigenbaum, 1964.] In the second half of the 1960s came another major advance in cognitive psychology, namely the discoveries of how to infer basic processes from reaction times. [Neisser, 1963;Sternberg, 1966.] This insight promised even greater ability to dissect human cognitive processes and confirmed the basic choice of psychology to analyze die block-diagram level of cognition. It also broadened the analysis from just memory structure to the stages of information-processing. In this respect, it might seem better to call the issue under discussion one of system levels: AI focusing on the symbolic level; and psychology focusing on the architecture, 9 i.e., the equivalent of the register-transfer level. However, the concern with memory so dominates the years prior to 1965, when this issue was being sorted out, that it seems preferable to label it memory versus processing.
Long-term memory has been absent from the above account. During this period, AI was certainly concerned about the structure of long-term memory-, under the rubric of semantic memory. This would seem to provide common ground with psychology, yet initially it did not to any great extent. Two factors seem to account for this. First, in psychology the new results, hence the excitement, all involved short-term memories.
The established theory of learning, interference theory, against which these new ideas about memory made headway, assumed a single memory, which was in essence a long-term memory. Second, the memory that psychology considered was episodic -learning what happened during an episode, such as learning which familiar items were presented at a trial. This stood in marked contrast with semantic memory, which appeared to be a timeless organization of knowledge. Only gradually did the psychologically relevant work ^Though the term "architecture" is just now coming into common use in psychology. on semantic memory by a few investigators capture any significant attention of cognitive psychology. The seminal publication of Anderson and Bowers Human Associative Memory can be taken as a marker of the beginning of this attention. [Anderson and Bower, 1973.] Problem-Solving versus Recognition #2: 1965-75. In 1965 AI moved to take back the problem of recognition that had become the intellectual property of the pattern-recognition community. This can be marked rather precisely by the work of Roberts on the recognition of three-dimensional polyhedra. [Roberts, 1965.1 The essential features were two. First, the recognition was articulated, i.e., the scene had to be decomposed or segmented into subparts, each of which might need to be recognized to be a different thing.
Thus, the result of recognition was a description of a scene, rather than just an identification of an object. But a description is a symbolic structure that has to be constructed, and such processes were quite outside the scope of the pattern-recognition techniques of the time, though exacdy of the sort provided by AI. Second, a major source of knowledge for making such recognitions came from adopting a model of the situation (e.g., it consists only of polyhedra). This made recognition processes strongly inferential, again fitting well with work in AI, but not with work in pattern recognition.
By the late 1960s, work on vision was going on throughout AI. But the transformation went further than just vision. Three laboratories (at MIT, Stanford and the Stanford Research Institute) started major efforts in robotics. Vision was to be coupled with arms and motion, and in at least one AI center (Stanford) with speech. The entire enterprise was radically different in its focus and problems from the research in pattern recognition that was still going on in parallel in departments and research centers of electrical engineering. In fact there was little actual controversy to speak of. Both groups simply did their thing. But likewise, there was no substantial rapprochement.
Syntax versus Semantics: 1965-75. The Chomskian revolution in linguistics was strongly based in theory.
Built around the notions of generative and transformational grammar, it posited three distinct components (or modules) for phonology, syntax and semantics, each with its own grammar. The initial emphasis was on syntax, with work on semantics much less well developed. 10 Despite cautions from the competence/performance distinction, the inference was clear from both the theory and the practice of linguistics -syntactic processing should occur in a separate module independently of semantic processing.
Indeed, what computational linguistics there was in association with the new linguistics involved the construction of programs for syntactic parsing.
In the late 1960s a reaction to linguistics arose from within the AI and computational linguistics There was work on phonology, but the domain lay outside the range of interest of AI and, in fact, of psychology as well.
communities. It took the form of denying the separation of syntax and semantics in the actual processing of language. The initial analysis of an utterance by the hearer was as much a question of semantics as of syntax.
Language required an integrated analysis by the hearer and, hence, by the theorist. This reaction can be marked by the work of Quillian, whose introduction of semantic nets was a device to show how semantic processing could occur directly on the surface structure of the utterance (though presumably in conjunction with syntax). [Quillian, 1968.] This reaction was grounded more broadly in the assertion of the importance of processing considerations in understanding language, the very thing denied by the competence/performance distinction. It sought to put processing considerations into the mainstream of linguistic studies, the latter being owned, so to speak, by the linguistics community. One result, as might have been expected, was to compound the separation between linguistics, on the one hand, and computational linguistics and AI, on the other. Another was to create a stronger independent stream of work on language in AI, with its own basis.
Theorem-Proving versus Problem-Solving: 1965-. Theorem-proving tasks have always been included in the zoo of tasks studied by AI, although the attention they received initially was sporadic. However, some logicians and mathematicians worked on theorem-proving in logic, not just as another task, but as the fundamental formalism for understanding reasoning and inference. In the last half of the 1960s, with the development of a logical formalism called "resolution", this work in theorem-proving took center stage in AI. [Robinson, 1965.] It seemed for a time that theorem-proving engines would sit at the heart of any general AI system. Not only was their power extended rapidly during this period, but a substantial amount of mathematical analysis was carried out on the nature of theorem proving in the predicate calculus. Even further, theorem-proving programs were extended to handle an increasing range of tasks, e.g., questionanswering, robot-planning and program-synthesis. A consequence of this success and viewpoint was that theorem-proving was taken to be a fundamental category of activity distinct from other problem-solving. It had its own methods and style of progress. A good indicator of this is Nilsson's 1971 AI textbook, which divides all problem-solving methods of AI into three parts: state-space search, problem-reduction (i.e., subgoals) and predicate-calculus theorem-proving. [Nilsson, 1971.] It is not clear whether this issue has been laid to rest by now or not As recounted below, under the procedural/declarative issue, theorem-proving has become much less central to AI since the mid-1970s. But theorem-proving and problem-solving still remain distinct research strands. involve just a design? And more. The start of this issue coincides with the creation of departments of computer science in the mid-1960s, which served to raise all these questions. Whether the issue will ever be laid to rest is unclear, but it is certainly unlikely while the whole field grows dynamically, with a continuing flood of new and destabilizing notions.
AI participates along with the rest of computer science in the uncertainties over whether it is an engineering or science discipline. However, the issue for AI has its own special flavor. AI participates with many disciplines outside computer science in the attempt to understand the nature of mind and intelligent behavior.
This is an externally grounded scientific and philosophic goal, which is clearly not engineering. Thus, the nature of the science for AI is not really in doubt, as it is for the rest of computer science. However, this does not end the matter, for interactions occur with other issues. For instance, to the extent that one orients towards helping humans rather than replacing them, one may not wish to accept understanding the nature of mind as a scientific goal, but only as a heuristic device.
The orientation toward engineering or science can have major consequences for how a field devotes its energies. Currently, for example, an important divergence exists in the subfield of computer vision. Should the nature of the environment be studied, to discover what can be inferred from the optic array (a scientific activity); or should experimental vision systems be constructed, to analyse the data they generate within the framework of the system (an engineering activity)? That both activities are legitimate is not in question; which activity gets the lion's share of attention is in dispute. And there is some indication that an important determiner is the basic engineering/science orientation of a given investigator.
Language versus Tasks: 1970-80. The 1970s saw the emergence of concerted efforts within AI to produce programs that understand natural language, amounting to the formation of a subfield, lying partly in AI and partly in computational linguistics. The key markers are the works of Woods and Winograd. [Woods, 1970;Winograd, 1971.] This had been building for some time, as we saw in the issue of syntax versus semantics.
The emergence of such a subfield is in itself not surprising. Natural language is clearly an important, even uniquely important, mental capability. In addition to AI there existed another relevant field, computational linguistics, concerned generally with the application of computers to linguistics. Neither is it surprising that this subfield had almost no representation from linguistics, although of course linguistics was of obvious central relevance. 11 The syntax/semantics issue, which had reinforced the separation of linguistics from AI, was a primary substantive plank in the programme of the new subfield.
What is interesting was the creation of another attitude within a part of AI, which can be captured by the issue of language versus tasks. The study of the understanding of language was seen as the sufficient context for investigating the nature of common sense. An important discovery was how much knowledge and inference appeared to be required to understand even the simplest of sentences or miniature stories. Thus, the very act of understanding such stories involved commonsense reasoning and, with it, the essence of general human intelligence. Programs could be interesting as AI research, so the attitude went, without doing any other task in addition to understanding the presented language input The effect of this strategic position was to separate the work in natural language processing from the tradition in AI of posing tasks for programs to do, where the difficulty could be assessed. The issue did not occasion much discussion, although its effects were real enough. The issue was masked by the fact that understanding, by itself, was a difficult enough task for AI research to make progress on. No one could object (and no one did) to not adding what seemed like an irrelevant second difficult task for the system, which would simply burden the research endeavor.

Procedural versus Declarative Representation #1: 1970-80.
Recall that resolution theorem-proving flourished in the late 1960s and bid fair to become the engine at the center of all reasoning. In fact, it took only a few years for the approach to come up against its limitations. Despite the increases in power, relative to prior efforts, the theorem provers were unable to handle any but trivial tasks. Getting from logic to real mathematics -seen always as a major necessary hurdle -seemed as far away as ever.
• / The reaction to this state of affairs became known as the procedural/declarative controversy. Theorem provers were organized as a large homogeneous database of declarative statements (clauses, in resolution), over which an inference engine worked to produce new true statements, to add to the database. This was the essence of a declarative representation of knowledge and its attractions were many. Its difficulty lay in the costs of processing. The inference engine treated all expressions in the database alike, or, more precisely, without regard to their semantics. There seemed no way for a theorem prover to be given information about how to solve problems. These two features added up to a major combinatorial explosion. The remedy -the procedural side of the issue -lay (so it was claimed) in encoding information about the task in procedures.
Then knowledge would be associated directly with the procedures that were to apply it, indeed, the procedures would embody the knowledge and thus not have to be interpreted by another inference engine.
This would permit the appropriate guidance to problem-solving and thus keep the combinatorial explosion under control.
There are irremediable flaws in both sides of the argument whether knowledge should be coded in procedural or declarative form, just as there are irremediable flaws in both sides of the argument whether a program is heuristic or algorithmic. Both procedural and declarative representations are necessary to make any computation at all happen. In consequence, the arguments over the issue were largely inconclusive, although it produced the closest thing to a public issue-controversy in AFs short history. However, the effect on the course of AI research was enormous. First, work on theorem-proving shrank to a trickle, with what remained mostly devoted to nonresolution theorem-proving. Second, the so-called planning languages emerged as a result -PLANNER, QA4, CONNIVER, POPLAR, etc. [Bobrow and Raphael, 1974.] These programming-language systems were intended to provide a vehicle for writing the sorts of domain-dependent, procedure-oriented theorem provers called for in the debate. While that did not quite happen, these languages in themselves provided a major conceptual advance in the field. The effects of this issue had about run their course by 1980.
Frames versus Atoms: 1970-80. In a paper that circulated widely before it was published in the mid-1970s, Marvin Minsky raised the issue about the size of the representational units in an intelligent system. [Minsky, 1975.] Knowledge should be represented in frames, which are substantial collections of integrated knowledge about the .world, rather than in small atoms or fragments. The basic issue is as old as the atomistic associationism of British empiricism and the countering complaints of the Gestaltists. How are the conflicting requirements for units of thought and for contextual dependence to be reconciled?
This issue had hardly surfaced at all in the first decade of AI. List structures, the basic representational medium, were in themselves neither atomistic nor wholistic, but were adaptable to whatever representational constructs the designer had in mind. 12 But the coming to prominence of resolution theorem proving in the late 1960s brought with it, as a side effect, the clause as the unit of representation. The clause was a primitive assertion that could not be broken down into a conjunction of other assertions -primitive predicates (P), negations of primitive predicates (~P), disjunctions (P or Q), implications (P implies Q), etc. The total knowledge of the system was to be represented as the conjunction of clauses -that is, to use the old Gestaltist ^TTiis is because list structures approximate general symbolic systems. The neutrality is easily confirmed in the continued and universal use of list-processing languages to realize systems of all kinds along the above dimension.
phrase, as an And-sum of separate bits of knowledge.
Thus, the issue of size of representational unit grew out of the same ground as the procedural versus declarative controversy and, indeed, it was articulated by the same group at MIT who had made most of the latter issue. As is always the case, the concern was in fact widespread, but had been subordinated to other concerns. [Abelson, 1973;Norman, 1973;"Schank, 1973.] Minsky was the first one to give clear voice to the concern. The effect of the paper was dramatic, despite die fact that the paper itself was entirely speculative and discursive. Throughout AI, the concept of the frame as the appropriate data structure was widely embraced. By 1980 frame systems were an established part of AI and a very substantial fraction of the work in knowledge representation was involved in such systems.
Much follows upon this development (in conjunction with the procedural/declarative issue) -the rise of substantial research effort in knowledge representation and the strengthening of renewed ties with philosophy. [Brachman and Smith, 1980.] These efforts conjoin with those of AI epistemology, discussed earlier. They raise some new issues, such as die relation of the philosophic work on meaning to directlyinspired computational models. But these have not yet jelled enough to be included in their own right Reason versus Emotion and Feeling #2: 1970-. Philosophy has a long-standing concern with the mechanization of mind. Indeed, under the aibric of the mind/body problem, it can be said almost to own the problem, it having been bequeathed to philosophy by Descartes. In its genesis, AI had very little involvement with philosophy, beyond the background awareness that comes from participation in the general intellectual culture. No philosophers of mind were involved and no technical philosophical issues were dealt with. A glance at the content of the two fields provides one obvious clue. The phenomena attended to in philosophy are sensations as subjective experiences ~ raw feels, to use a bit of philosophic jargon. A typical article is entitled "The Feelings of Robots". [Ziff, 1959.] Thus, though AI and philosophy of mind ostensibly deal with the same problem, in fact they go after largely distinct phenomena. 13 The issue has not been especially active, but it has been raised. [Gunderson, 1971.] It is argued that performance functions (i.e., those functions AI currently deals with, called program-receptive functions) can be mechanized; but that sentient functions (i.e., feelings, called program-resistant functions) cannot. Whether this ever grows to a substantial controversy is hard to tell at this point. It is certainly available as a fall-back position that can serve to separate AI from the philosophy of mind. It adds to the general background concern, discussed in the first occurrence of this issue, of the absence of emotion and feelings in the 13 Another example is the problem of induction, in which philosophy is concerned with the certainty of induction and AI is concerned with performing the inductions. [Newell, 1973b.] development of intelligent systems.
Toy versus Real Tasks: 1975-. As noted in the power/generality issue, the field took a shift in the mid-1960s away from seeking powerful programs and toward programs that could exhibit common sense.
Further, as noted in the language/tasks issue, this line further transmuted to being concerned with understanding via the understanding of natural language. Concomitantly, programs were often built to work on small simple illustrative tasks or environments, usually puzzles or made-up situations.
By the mid-1970s some systems had been developed that worked with real tasks that had substantial intellectual content, to judge from their role in the real world. The initial such system can be taken to be DENDRAL, which determined the structural formula for chemical molecules, given the data on the mass spectrogram. 14 [Lindsay, Buchanan, Feigenbaum and Lederberg, 1980.] DENDRAL began in the late 1960s and grew in power throughout the early 1970s. It was joined in the mid-1970s by several systems that performed competently in real medical-diagnosis tasks, of which MYCIN was the paradigm. [Shortliffe, 1974.] This was the immediate locus of expert systems, which, as we noted above, grew up as pan of the general emphasis on knowledge in contrast to search. With it grew an attitude tiiat AI in general should no longer work on small illustrative, artificial tasks, but that it was time to work on real tasks. The simple artificial tasks came to be called "toy" tasks, not just because the term conveys the contrast between childish and grown-up pursuits, but also because stacking children's blocks had become a favorite illustrative task environment The tension between basic research and application exists in all sciences at all times. Sciences sometimes build institutional structures to contain the tension. As we saw in the issue of science versus engineering, computer science has kept its basic and applied components mixed together in a single discipline, thus exacerbating the tension. The tension was in fact especially severe for AI during the decade of the 1970s. The climate in Washington was not benign for basic research in general and there was sustained pressure from AFs primary government funding agency (DARPA, Defense Advanced Research Projects Agency) to make AI pay off. That said, however, the distinction of toy versus real tasks is not solely the distinction between basic and applied research. Tasks taken from the real world, performed by intelligent humans as part of their working lives, carry a prima facie guarantee of demanding appropriate intelligent activity by systems that would perform them. It can be argued that such tasks are the appropriate ones for AI to work on, even if the goal is basic research. Thus, the toy-versus-real-tasks issue stands ambiguously for both meanings -basic versus applied and irrelevant versus relevant basic science.
The other system often mentioned similarly is MACSYMA, the highly sophisticated program at MIT for doing symbolic mathematics. As mentioned earlier, it had deliberately removed itself from being an AI program.
Serial versus Parallel #2: 1975-. By the mid-1970s computer science had for some time been seriously exploring multiprogramming and multiprocessing. These provided the groundwork for considering parallel systems for doing AI. A major instigation occurred with the development of the Hearsay-II model of speech understanding. [Lesser and Erman, 1977.] Hearsay-II comprised a number of knowledge sources (acoustic, phonetic, phonological, lexical, syntactic, semantic, and pragmatic), each working concurrently and independendy off a common blackboard that contained the current working state about the utterance, and each contributing their bit to the evolving recognition and reacting to the bits provided by the others.
The Hearsay-II structure was certainly a parallel one, but it was at a grain of parallelism quite different from the earlier network models, namely, a modest number (tens) of functionally specialized processes.
Furthermore, the individual processes remained fundamentally symbolic (even though lots of signalprocessing was inherent in the speech-recognition task). Hearsay-II was only one of several efforts to pursue the notion that an intelligent system should be thought of in terms of communicating subprocesses, rather than as an individual serial machine. A metaphor arose for thinking about an intelligent system -the "scientific community" metaphor -which took the operation of science, with its notion of cooperation, publication, experiment, criticism, education, etc, as the appropriate model for intelligent activity. Gradually a group of people emerged interested in working on distributed AI.
Performance versus Learning #2: 1975-. As noted earlier, learning was generally associated with work on pattern recognition. With the split between problem-solving and recognition, work on learning within AI declined. As always, it never stopped entirely. Indeed, such is the basic fascination with learning processes, and with the belief that they hold the key to intelligence, that each learning program that was constructed received substantial attention. 15 [Samuel, 1961;Waterman, 1970;Winston, 1970;Sussman, 1975.] However, each learning system was relatively idiosyncratic, with its own interesting lessons, so that the whole did not add up to a coherent effort for the field A reversal of this state of affairs developed by the late 1970s. It was triggered by the spread of a class of programming systems, called production, or rule-based, systems. These are used both for constructing expert systems and for analyzing human cognition. [Waterman and Hayes-Roth, 1978.] To appreciate their role in the resurgence of work on learning one must take a step back. To create a learning system requires solving two research problems. First, a space of potential performance programs must be created, in which learning will constitute moving from one program to another, searching for programs with better performance. If the space of programs is too vast and irregular, then learning is, in effect, automatic programming and it becomes 15 Some other systems were built, which might have been viewed as learning systems, but instead were taken simply to be performance programs in specialized task environments, e.g., induction programs. extremely difficult. If the space is too limited, then learning is easy, but the performance programs are of little significance. Determining the right space is thus a critical research activity. Second, given the space, it is still necessary to design an interesting learning system, for the space only lays out the possibilities. Thus, inventing the learning system is also a critical research activity. A major reason why early AI learning-systems seemed so idiosyncratic was that each made unique choices on both these dimensions. Most important, doing research on learning was doing a double task and taking a double risk.
A production system is composed entirely of a set of if-then rules {if 'such and such conditions hold, then execute such and such actions). At each instant the rules diat hold are recognized, and a single rule is selected to execute. In such a system, the natural space of performance programs consists of subsets of if-then rules and the primitive act of learning is to add a new rale to die existing set (or sometimes to modify an existing rule in some simple way, such as by adding another condition). This space of performance programs is neither too limited nor too open, since it is easy to restrict the rules to be learned to a special class. As a consequence, the first research choice is essentially made for the researcher, who can then concentrate on constructing an interesting learning program. Moreover, learning programs will have much in common, since they now utilize similar spaces of performance programs. Indeed, this is just what happened in the late 1970s, as researchers began to construct a wide variety of small learning systems, all built around variants of the production-system formalism. [Michalski, Carbonell and Mitchell, 1982.] It must be realized, of course, that such focusing of effort does not remove the collective risk. If production systems are the wrong program organization to be exploring, then the entire field is moving down an unproductive path.
Psychology versus Neuroscience #2:1975-. AI would appear to be at the mercy of the immense gulf that continues to separate psychology and the biology of the brain. As each field continues to progress -which both do dramatically -hopes continually spring up for new bridging connections. No doubt at some point the permanent bridge will get built So far, although each increment of progress seems real, the gap remains disappointingly large.
It is possible that AI has a major contribution to make to this by exploring basic computational structures at a level that makes contact with neural systems. In the early instance of psychology versus neurophysiology (which was before the term "neuroscience" had been coined) that possibility seemed quite remote. The theoretical structures that did make contact with neurophysiology were remote from the computational structures that preoccupied AI researchers. Then the split occurred, with pattern recognition all but moving out of computer science.
In the mid-1970s, a new attempt began to connect AI with neuroscience, initiated by the work of David Marr. [Marr, 1976.] The emphasis remained on vision, as it had been in the earlier period. But the new effort was explicitly computational, focusing on algorithms that could perform various low-level vision functions, such as stereopsis. Although Marr's effort was new in many ways, and based on specific technical achievements, most of the global issues of the earlier time reappeared. This work has now expanded to a larger group, which calls its work, among other things, the "New Connectionism", and which promises to be a substantial subfield again, this time within AI.
. Serial versus Parallel #3: 1980-. The new wave of neuroscience-inspired AI contains, of course, a commitment to highly parallel network-structures. The issue of serial versus parallel merits a separate entry here to maintain a clear contrast with the distributed AI effort, which defined the second wave of concern with parallel systems. In this third phase, the degree of parallelism is in the millions and the computing elements in the network have modest powers. In particular, they are not computers with their own local symbols. In the new structures, computation must be shared right down to the roots, so to speak. The interaction cannot be limited to communicating results of significant computations. Furthermore, the communication media between the elements are continuous signals, and not just bits. However, unlike the earlier work, these new computational systems are not to be viewed as neural nets, that is ? the nodes of the network are not to be put in one-to-one correspondence with neurons, but rather witii physiological subsystems of mostly unspecified character.
Problem-Solving versus Recognition #3:1980-. Robotics has returned to AI after having left it for most of the 1970s. Perhaps it is unfortunate to call the issue "problem-solving versus recognition", since recognition is only one aspect of robotics. The main sources of the new wave of effort are external to AI -industrial robotics, plus the concern of the decline in American productivity and the trade position of the United States vis-a-vis Japan and West Germany. The initial growth of industrial robotics took place largely outside of AI as a stricdy engineering endeavor. As a result, it tended to minimize the intelligence involved, e.g., the sensory-motor coordination. One component of the new association of robotics with AI is the coupling of significant amounts of vision with manipulators, reflecting the continued advance in vision capabilities in AI throughout the 1970s. (Touch and kinesthetic sensing is increasingly important too, but this does not build so strongly on prior progress in AI.) Importantiy, along with the industrially motivated aspects, there is also a revival of basic research into manipulation and movement in space and over real terrains.
It might seem that this is just another, purely technical, progression. But with it has returned, as night follows day, the question of the relation of AI and robotics as disciplines, just as the question was raised in the issue of problem-solving versus recognition during the late 1960s. Is robotics a central part of AI or only an applied domain? Do graduate students in AI have to understand the underlying science of mechanics and generalized coordinate systems that are inherent in understanding manipulation and motion? Or is that irrelevant to intelligence? Cases can be made either way. [Nilsson, 1982.] / Procedural versus Declarative Representation #2: 1980-. In the late 1970s a new programming system called PROLOG emerged. [Kowalski, 1979.] It is based on resolution-theorem-proving and constitutes, in effect, a continuation of the effort to show that declarative formulations can be effective. The effort is based primarily in Europe and it is a vigorous movement The attack is not occurring at the level of the planning languages, but at the level of LISP itself. Over the years, LISP has established itself as the lingua franca of the AI community. Even though various other programming systems exist, e.g., rule-based systems of various flavors, practically everyone builds systems within a LISP programming environment. The planning languages (PLANNER, CONNIVER, etc.), which showed how to effect another level of system organization above LISP, have not proved highly effective as a replacement, and they receive only modest use. As noted above, their contribution has been primarily conceptual. Thus, although the original attack on theoremproving was in terms of the planner languages, the modern counterattack is at the level of LISP. By being centered in Europe, with very little attention being paid currently to PROLOG in the major AI centers of the United States, die issue takes on additional coordinated dimensions. The outcome is far from clear at this juncture.

Discussion
It should be clear by now why I entered the caveats about historical accuracy at the beginning. Each of the issues raises serious problems of characterization and historical grounding. No attempt has been made to define an intellectual issue, so that some modesdy objective way could be found to generate a complete set of issues, for example, by placing a grid over the literature of the field. Several additional issues might well have emerged and some of those presented here might not have made the grade. Thus, the population of issues exhibited must be taken, not just with a pinch of salt but soaked in a barrel of brine. Similar concerns attend the dating of the issues and my interpretation of them. Nevertheless, some comments about the total picture seem worthwhile.
What is missing. I do know why some issues did not make it. Three examples will illustrate some reasons.
The first is the broad but fundamental issue of the ethical use of technology and the dehumanization of man by reduction to mechanism. This issue engages all of technology and science. It seems particular acute for AI, perhaps, because the nature of mind seems so close to the quick. But the history of science reminds us easily enough that at various stages astronomy, biology and physics have "seemed special targets for concern. There has been continued and explicit discussion of these issues in connection with AI. [Taube, 1961;Weizenbaum, 1976;McCorduck, 1979.] I have not included them in the list of intellectual issue because they do not in general seem to affect the course of the science. Where some aspect does seem to do so, as in the issue of helping humans or replacing them, it has been included. However, the broader issue certainly provides a thematic background against which all work goes on in field, increasing its ambiguity. It undoubtedly enters into individual decisions about whether to work in the field and what topics to select.
The second example involves Hubert Dreyfus, who has been a persistent and vocal critic of AI. [Dreyfus, 1972.] He has certainly become an issue for the field. However, this does not necessarily produce an intellectual issue. Dreyfiis's central intellectual objection, as I understand him, is that the analysis of the Context of human action into discrete 'elements is doomed to failure. This objection is grounded in phenomenological philosophy. Unfortunately, this appears to be a nonissue as far as AI is concerned. The answers, refutations, and analyses that have been forthcoming to Dreyfiis's writings have simply not engaged this issue -which indeed would be a novel issue if it were to come to the fore.
The third example involves the imagery controversy, which has been exceedingly lively in cognitive psychology. [Kosslyn, Pinker, Smith and Shwartz, 1979.] The controversy is over the nature of the representations used by humans in imagining scenes and reasoning about them. There is no doubt about its relevance to AI -the alternatives are a classical dichotomy between propositional (symbolic?) representations and analog ones. Thus, at heart it is a variant of the issue of analog-versus-digital representation, which has received mention. But for reasons that are quite obscure to me, the imagery issue has received hardly any interest in the AI community, except where that community also participates in cognitive psychology. As things stand at the moment, this would be an issue for cognitive science, but it is not one for AI.
Though enumerating intellectual issues exposes a certain amount of the history of a field, even if only from particular viewpoints, some important parts can be missed. These seem to be endeavors that were noncontroversial, or where the controversies were merely of the standard sort of what progress had been made, what subfields should get resources, etc. Thus work on program synthesis and verification goes unnoticed. Also, the major effort in the 1970s to construct speech-understanding system is barely noticed.
Perhaps, this is not a valid point about the basic historical scheme, but reflects only the unevenness of my process of generating issues. Certainly there were issues in speech-recognition research, both in the 1960s, when Bell Laboratories decided to abandon speech recognition as an inappropriate task, and in the 1970s, when a substantial effort sponsored by DARPA. to construct speech-understanding systems was dominated by AI considerations over speech-science considerations. Perhaps, intellectual issues are generated from all scientific efforts in proportion to the number of scientists involved in them (or to their square?); all we need to do is look for them.

Characteristics of the history.
Turning to what is revealed in Table 1, the most striking feature, to me at least, is how many issues there are. Looked at in any fashion ~ number active at one time (fifteen on average) or total number of issues during AFs quarter-century lifespan (about thirty) -it seems to me like a lot of issues. Unfortunately, similar profiles do not exist for other fields (or I do not know of them). Perhaps the situation in AI is typical, either of all fields at all times, or of all fields when they are getting started. In fact, I suspect it is due to the interdisciplinary soup out of which AI emerged. [Newell, 1983.] Many other related fields were being defined during the same post-World-War-II era ~ cybernetics, operations research, management science, information theory, control theory, pattern recognition, computer science, general systems theory. Even so, I do not see any easy way of pinning down a correct interpretation of why there are so many issues.
Issues are not independent. They come in clusters, which are coordinated. Researchers tend to fall into two classes, corresponding to one pole or another on all the issues in the cluster. A cluster might seem to define a single underlying issue, which can then replace the component issues.
However, just because issues are coordinated does not make them identical. Some scientists can always be found that are aligned in non-standard patterns. In fact, some of the clusters seem much more consistent than others. Thus, the multiplicity of issues keeps the scientific scene complex, even though, because of clustering, it appears diat it should be clear and simple. In fact, many of the groupings above are more easily labeled by how they separate fields than by any coherent underlying conceptual issue.
Clustering of issues does seem to be a common occurrence. For instance, a standard advanced text on learning in psychology begins with a list of seven dichotomous issues that characterize learning theories. [Hilgard and Bower, 1973, pp. 8-13.] The first three ~ peripheral versus central, habits versus cognitive structures, and trial-and-error versus insight -form a coordinated cluster that characterizes stimulus-response theories versus cognitive theories. (To which could even be added tough-minded versus tender-minded, the contrast William James used to distinguish the two main types of psychologists.) One possible source for such coordinated clusters is the attempt to find multiple reasons to distinguish one approach from another. The approach comes first and the issues follow afterward. Then the issues take on an autonomous intellectual life and what starts as rationalization ends up as analysis.
A major role of the issues here seems to be to carve up the total scientific field into disciplines. AI, computer science, logic, cybernetics, pattern recognition, linguistics, cognitive psychology ~ all these seem to be discriminated in part by their position on various of these issues. The issues, of course, only serve as intermediaries for intellectual positions that derive from many circumstances of history, methodological possibilities, and specific scientific and technical ideas. Still, they seem to summarize a good deal of what keeps the different fields apart, even though the fields have a common scientific domain.
Is the large burst of issues that occured at the birth of AI just an artifact of my intent to gather the issues for AI? If the period just before AI began, say from 1940-1955, were examined carefully, would many more issues be added? The relevant question should probably be taken with respect to some other field as a base.
Would a burst like this be found for cybernetics, which started in 1940-1945? My own suspicion is yes, but I have not tried to verify it Perhaps then the situation of AI could turn out to be typical. Wc would find a plethora of issues in any science, if we would but look and count. The list from Hilgard and Bower, above, might serve as a positive indicator. However, before rushing to embrace this view, some counterevidence should be examined. An interesting phenomenon in this same postwar period was the emergence of several "one-theorem" fields.
Game theory, information theory, linear programming, and (later) dynamic programming all had a single strong result around which the field grew. 16 Certainly, each also provided a novel formulation, which amounted to a class of systems to be used to theorize about some field. But initially there was only one striking theorem to justify the entire field. It gave them a curious flavor. My personal recollection is that all these fields, while exciting, profound, and (sometimes) controversial, had none of the complexity of issues that we find in Table 1.
Intellectual issues and progress. There is a natural temptation to use the history of intellectual issues to measure progress, once it has been explicitly laid out. It is true that some issues have vanished from the scene, such as symbols versus numbers. That seems, perhaps, like progress. It is also true that, other issues seem to recur, such as problem-solving versus recognition. That seems, perhaps, like lack of progress. Neither interpretation is correct I think. Rather, the progress of science is to be measured by the accumulation of theories, data, and techniques, along with the ability they provide to predict explain and control. This story is not to be told in terms of intellectual issues such as populate this paper. It requires attention to the detailed content assertions and practice of the science itself. True, at the more aggregate level of the paradigms of Kuhn or the programmes of Lakatos, whole bodies of theory and data can become irrelevant with a shift in paradigm or programme. But on the scale of the twenty-five years of AI research (1955 to 1980), the story is one of accumulation and assimilation, not one of shift and abandonment It is not even one of settling scientific questions permanendy, by and large.
What then is the role of intellectual issues in the progression of science? To echo my earlier disclaimer,-!/ Another field, general systems theory, also had a single idea around which to build -that there are common laws across all levels of systems from the atomic through cellular through societal through astronomical. But there was no central result available, only the system view, and this field has been markedly less successful than the others in its growth and health.

PAGE 38
can only conjecture. Intellectual issues seem to me more like generalized motivators. They evoke strong enough passions to provide the springs to action; but they are vague enough so that they do not get in the way of specific work. They can be used to convey a feeling of coherence on investigations in their early stages, before it is known exacdy what the investigations will yield.
One evidence for this is that issues do not really go away. They return, and return again. Repetition is abundandy in evidence in Table 1. The model that suggests itself immediately is the spiral -each return constitutes a refined version of the issue. Though the issues are certainly not identical each time, it seems difficult to construe the changes as any sort of progressive refinement. Some seem more like wandering (e.g., the serial/parallel issue). More plausible (to me) is that intellectual issues reflect perennial unanswerable questions about the structure of nature -continuity/discontinuity, stasis/change, essence/accident, autonomy/dependence, etc. Whenever in the course of science one of these can be recognized in the ongoing stream of work, an appropriate intellectual issue will be instantiated, to operate as a high-level organizing principle for awhile. To be sure, this picture does not capture all that seems represented in our population of intellectual issues. But it seems substantially better than viewing science as progressively resolving such issues.

Conclusion
Putting to one side the questions about the accuracy of the particular set of issues displayed in Table 1, of what use is a history of a scientific field in terms of intellectual issues? To repeat once more: it cannot substitute for a substantive history in terms of concepts, theories, and data. However, it does seem to capture some of the flavor of the field in an era. It is clearly a component of die paradigm of a field or of research programmes within a field. And, let us confess it, intellectual issues have a certain spiciness about them that makes them fun to talk and write about Perhaps it is the sense of touching fundamental issues. But perhaps it also echoes Bertrand Russell's famous aphorism that dealing with intellectual issues has all the advantages of theft over honest toil.