FUNDING A REVOLUTION
Government Support for Computing Research
Committee on Innovations in Computing and Communications: Lessons from History,
National Research Council
Developments in Artificial Intelligence
Artificial intelligence (AI) has been one of the most controversial domains of inquiry in computer science since it was first proposed in the 1950s. Defined as the part of computer science concerned with designing systems that exhibit the characteristics associated with human intelligence--understanding language, learning, reasoning, solving problems, and so on (Barr and Feigenbaum, 1981)--the field has attracted researchers because of its ambitious goals and enormous underlying intellectual challenges. The field has been controversial because of its social, ethical, and philosophical implications. Such controversy has affected the funding environment for AI and the objectives of many research programs.
AI research is conducted by a range of scientists and technologists with varying perspectives, interests, and motivations. Scientists tend to be interested in understanding the underlying basis of intelligence and cognition, some with an emphasis on unraveling the mysteries of human thought and others examining intelligence more broadly. Engineering-oriented researchers, by contrast, are interested in building systems that behave intelligently. Some attempt to build systems using techniques analogous to those used by humans, whereas others apply a range of techniques adopted from fields such as information theory, electrical engineering, statistics, and pattern recognition. Those in the latter category often do not necessarily consider themselves AI researchers, but rather fall into a broader category of researchers interested in machine intelligence.
The concept of AI originated in the private sector, but the growth of the field, both intellectually and in the size of the research community, has depended largely on public investments. Public monies have been invested in a range of AI programs, from fundamental, long-term research into cognition to shorter-term efforts to develop operational systems. Most of the federal support has come from the Defense Advanced Research Projects Agency (DARPA, known during certain periods as ARPA) and other units of the Department of Defense (DOD). Other funding agencies have included the National Institutes of Health, National Science Foundation, and National Aeronautics and Space Administration (NASA), which have pursued AI applications of particular relevance to their missions--health care, scientific research, and space exploration.
This chapter highlights key trends in the development of the field of AI and the important role of federal investments. The sections of this chapter, presented in roughly chronological order, cover the launching of the AI field, the government's initial participation, the pivotal role played by DARPA, the success of speech recognition research, the shift from basic to applied research, and AI in the 1990s. The final section summarizes the lessons to be learned from history. This case study is based largely on published accounts, the scientific and technical literature, reports by the major AI research centers, and interviews conducted with several leaders of AI research centers. (Little information was drawn from the records of the participants in the field, funding agencies, editors and publishers, and other primary sources most valued by professional historians.)1
THE PRIVATE SECTOR LAUNCHES THE FIELD
The origins of AI research are intimately linked with two landmark papers on chess playing by machine.2 They were written in 1950 by Claude E. Shannon, a mathematician at Bell Laboratories who is widely acknowledged as a principal creator of information theory. In the late 1930s, while still a graduate student, he developed a method for symbolic analysis of switching systems and networks (Shannon, 1938), which provided scientists and engineers with much-improved analytical and conceptual tools. After working at Bell Labs for half a decade, Shannon published a paper on information theory (Shannon, 1948). Shortly thereafter, he published two articles outlining the construction or programming of a computer for playing chess (Shannon, 1950a,b).
Shannon's work inspired a young mathematician, John McCarthy, who, while a research instructor in mathematics at Princeton University, joined Shannon in 1952 in organizing a conference on automata studies, largely to promote symbolic modeling and work on the theory of machine intelligence.3 A year later, Shannon arranged for McCarthy and another future pioneer in AI, Marvin Minsky, then a graduate student in mathematics at Princeton and a participant in the 1952 conference, to work with him at Bell Laboratories during 1953.4
By 1955, McCarthy believed that the theory of machine intelligence was sufficiently advanced, and that related work involved such a critical mass of researchers, that rapid progress could be promoted by a concentrated summer seminar at Dartmouth University, where he was then an assistant professor of mathematics. He approached the Rockefeller Foundation's Warren Weaver, also a mathematician and a promoter of cutting-edge science, as well as Shannon's collaborator on information theory. Weaver and his colleague Robert S. Morison, director for Biological and Medical Research, were initially skeptical (Weaver, 1955). Morison pushed McCarthy and Shannon to widen the range of participants and made other suggestions. McCarthy and Shannon responded with a widened proposal that heeded much of Morison's advice. They brought in Minsky and a well-known industrial researcher, Nathaniel Rochester5 of IBM, as co-principal investigators for the proposal, submitted in September 1955.6
In the proposal, the four researchers declared that the summer study was "to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." They sought to bring a number of U.S. scholars to Dartmouth to create a research agenda for AI and begin actual work on it. In spite of Morison's skepticism, the Rockefeller Foundation agreed to fund this summer project with a grant of $7,500 (Rhind, 1955), primarily to cover summer salaries and expenses of the academic participants. Researchers from industry would be compensated by their respective firms.
Although most accounts of AI history focus on McCarthy's entrepreneurship, the role of Shannon--an intellectual leader from industry--is also critical. Without his participation, McCarthy would not have commanded the attention he received from the Rockefeller Foundation. Shannon also had considerable influence on Marvin Minsky. The title of Minsky's 1954 doctoral dissertation was "Neural Nets and the Brain Model Problem."
The role of IBM is similarly important. Nathan Rochester was a strong supporter of the AI concept, and he and his IBM colleagues who attended the 1956 Dartmouth workshop contributed to the early research in the field. After the workshop IBM welcomed McCarthy to its research laboratories, in large part because of IBM's previous work in AI and because "IBM looked like a good bet to pursue artificial intelligence research vigorously" in the future.7 Rochester was a visiting professor at the Massachusetts Institute of Technology (MIT) during 1958-1959, and he unquestionably helped McCarthy with the development of LISP, an important list-processing language (see Box 9.1).8 Rochester also apparently lent his support to the creation in 1958 of the MIT Artificial Intelligence Project (Rochester and Gelertner, 1958).9 Yet, in spite of the early activity of Rochester and other IBM researchers, the corporation's interest in AI cooled. Although work continued on computer-based checkers and chess, an internal report prepared about 1960 took a strong position against broad support for AI.
Thus, the activities surrounding the Dartmouth workshop were, at the outset, linked with the cutting-edge research at a leading private research laboratory (AT&T Bell Laboratories) and a rapidly emerging industrial giant (IBM). Researchers at Bell Laboratories and IBM nurtured the earliest work in AI and gave young academic researchers like McCarthy and Minsky credibility that might otherwise have been lacking. Moreover, the Dartmouth summer research project in AI was funded by private philanthropy and by industry, not by government. The same is true for much of the research that led up to the summer project.
THE GOVERNMENT STEPS IN
The federal government's initial involvement in AI research was manifested in the work of Herbert Simon and Allen Newell, who attended the 1956 Dartmouth workshop to report on "complex information processing." Trained in political science and economics at the University of Chicago, Simon had moved to Carnegie Institute of Technology in 1946 and was instrumental in the founding and early research of the Graduate School of Industrial Administration (GSIA). Funded heavily by the Ford Foundation and the Office of Naval Research (ONR), and the Air Force, GSIA was the pioneer in bringing quantitative behavioral social sciences research (including operations research) into graduate management education.10 Because of his innovative work in human decision making, Simon became, in March 1951, a consultant to the RAND Corporation, the pioneering think tank established by the Air Force shortly after World War II.11
At RAND, where he spent several summers carrying out collaborative research, Simon encountered Newell, a mathematician who helped to conceive and develop the Systems Research Laboratory, which was spun out of RAND as the System Development Corporation in 1957. In 1955, Simon and Newell began a long collaboration on the simulation of human thought, which by the summer of 1956 had resulted in their fundamental work (with RAND computer programmer J.C. Shaw) on the Logic Theorist, a computer program capable of proving theorems found in the Principia of Bertrand Russell and Alfred North Whitehead (Newell and Simon, 1956).12
This program is regarded by many as the first successful AI program, and the language it used, IPL2, is recognized as the first significant list-processing language. As programmed by Simon, Newell, and Shaw, a computer simulated human intelligence, solving a problem in logic in much the same way as would a skilled logician. In this sense, the machine demonstrated artificial intelligence. The project was funded almost entirely by the Air Force through Project RAND, and much of the computer programming was done at RAND on an Air Force-funded computer (the Johnniac, named after RAND consultant John von Neumann, the creator of the basic architecture for digital electronic computers).13
Newell's collaboration with Simon took him to Carnegie Tech, where, in 1957, he completed the institution's first doctoral dissertation in AI, "Information Processing: A New Technique for the Behavioral Sciences." Its thrust was clearly driven by the agenda laid out by the architects of GSIA. As Newell later stressed, his work with Simon (and that of Simon's several other AI students at GSIA) reflected the larger agenda of GSIA, even though most of this work was funded by the Air Force and ONR until the early 1960s. All of this work concentrated on the formal modeling of decision making and problem solving.
Simon and Newell developed another well-known AI program as a sequel to Logic Theorist--the General Problem Solver (GPS), first run in 1957 and developed further in subsequent years. Their work on GPS, like that on Logic Theorist, was characterized by its use of heuristics (i.e., efficient but fallible rules of thumb) as the means to simulate human cognitive processes (Newell et al., 1959). The GPS was capable of solving an array of problems that challenge human intelligence (an important accomplishment in and of itself), but, most significantly, it solved these problems by simulating the way a human being would solve them. These overall research efforts at GSIA, including the doctoral research of Simon's students--all funded principally by Air Force and ONR money--remained modest in scale compared to those at Carnegie Tech after 1962.14
Also modest were the efforts at MIT, where McCarthy and Minsky established the Artificial Intelligence Project in September 1957. This effort was funded principally through a word-of-mouth agreement with Jerome Wiesner, then director of MIT's military-funded Research Laboratory in Electronics (RLE). In exchange for "a room, two programmers, a secretary and a keypunch [machine]," the two assistant professors of mathematics agreed, according to McCarthy, to "undertake the supervision of some of the six mathematics graduate students that RLE had undertaken to support."15
The research efforts at Carnegie Tech (which became Carnegie Mellon University [CMU] in 1967), RAND, and MIT, although limited, yielded outstanding results in a short time. Simon and Newell showed that computers could demonstrate human-like behavior in certain well-defined tasks.16 Substantial progress was also made by McCarthy, with his pioneering development of LISP, and Minsky, who formalized heuristic processes and other means of reasoning, including pattern recognition. Previously, computers had been used principally to crunch numbers, and the tools for such tasks were primitive. The AI researchers found ways to represent logical formulas, carry out proofs, conduct plans, and manipulate such objects. Buoyed by their successes, researchers at both institutions projected bold visions--which, as the research was communicated to the public, became magnified into excessive claims--about the future of the new field of AI and what computers might ultimately achieve.17
DARPA'S PIVOTAL ROLE
The establishment in 1962 of ARPA's Information Processing Techniques Office (IPTO) radically changed the scale of research in AI, propelling it from a collection of small projects into a large-scale, high-profile domain. From the 1960s through the 1990s, DARPA provided the bulk of the nation's support for AI research and thus helped to legitimize AI as an important field of inquiry and influence the scope of related research. Over time, the nature of DARPA's support changed radically--from an emphasis on fundamental research at a limited number of centers of excellence to more broad-based support for applied research tied to military applications--both reflecting and motivating changes in the field of AI itself.
The early academic centers were MIT and Carnegie Tech. Following John McCarthy's move to Stanford in 1963 to create the Stanford Artificial Intelligence Laboratory (SAIL), IPTO worked a similar transformation of AI research at Stanford by making it the third center of excellence in AI. Indeed, the IPTO increased Stanford's allocation in 1965, allowing it to upgrade its computing capabilities and to launch five major team projects in AI research. Commenting in 1984 about how AI-related research at Carnegie Tech migrated out of GSIA into what became an autonomous department (and later a college) of CMU, Newell (1984) captured the transformation wrought by IPTO:
. . . the DARPA support of AI and computer science is a remarkable story of the nurturing of a new scientific field. Not only with MIT, Stanford and CMU, which are now seen as the main DARPA-supported university computer-science research environments, but with other universities as well . . . DARPA began to build excellence in information processing in whatever fashion we thought best. . . . The DARPA effort, or anything similar, had not been in our wildest imaginings. . . .
Another center of excellence--the Stanford Research Institute's (SRI's) Artificial Intelligence Center--emerged a bit later (in 1966), with Charles Rosen at the command. It focused on developing "automatons capable of gathering, processing, and transmitting information in a hostile environment" (Nilsson, 1984). Soon, SRI committed itself to the development of an AI-driven robot, Shakey, as a means to achieve its objective. Shakey's development necessitated extensive basic research in several domains, including planning, natural-language processing, and machine vision. SRI's achievements in these areas (e.g., the STRIPS planning system and work in machine vision) have endured, but changes in the funder's expectations for this research exposed SRI's AI program to substantial criticism in spite of these real achievements.
Under J.C.R. Licklider, Ivan Sutherland, and Robert Taylor, DARPA continued to invest in AI research at CMU, MIT, Stanford, and SRI and, to a lesser extent, other institutions.18 Licklider (1964) asserted that AI was central to DARPA's mission because it was a key to the development of advanced command-and-control systems. Artificial intelligence was a broad category for Licklider (and his immediate successors), who "supported work in problem solving, natural language processing, pattern recognition, heuristic programming, automatic theorem proving, graphics, and intelligent automata. Various problems relating to human-machine communication--tablets, graphic systems, hand-eye coordination--were all pursued with IPTO support" (Norberg and O'Neill, 1996).
These categories were sufficiently broad that researchers like McCarthy, Minsky, and Newell could view their institutions' research, during the first 10 to 15 years of DARPA's AI funding, as essentially unfettered by immediate applications. Moreover, as work in one problem domain spilled over into others easily and naturally, researchers could attack problems from multiple perspectives. Thus, AI was ideally suited to graduate education, and enrollments at each of the AI centers grew rapidly during the first decade of DARPA funding.
DARPA's early support launched a golden age of AI research and rapidly advanced the emergence of a formal discipline. Much of DARPA's funding for AI was contained in larger program initiatives. Licklider considered AI a part of his general charter of Computers, Command, and Control. Project MAC (see Box 4.2), a project on time-shared computing at MIT, allocated roughly one-third of its $2.3 million annual budget to AI research, with few specific objectives.
SUCCESS IN SPEECH RECOGNITION
The history of speech recognition systems illustrates several themes common to AI research more generally: the long time periods between the initial research and development of successful products, and the interactions between AI researchers and the broader community of researchers in machine intelligence. Many capabilities of today's speech-recognition systems derive from the early work of statisticians, electrical engineers, information theorists, and pattern-recognition researchers. Another key theme is the complementary nature of government and industry funding. Industry supported work in speech recognition at least as far back as the 1950s, when researchers at Bell Laboratories worked on systems for recognizing individual spoken digits "zero" through "nine." Research in the area was boosted tremendously by DARPA in the 1970s.
DARPA established the Speech Understanding Research (SUR) program to develop a computer system that could understand continuous speech. Lawrence Roberts initiated this project in 1971 while he was director of IPTO, against the advice of a National Academy of Sciences committee.19 Roberts wanted a system that could handle a vocabulary of 10,000 English words spoken by anyone. His advisory board, which included Allen Newell and J.C.R. Licklider, issued a report calling for an objective of 1,000 words spoken in a quiet room by a limited number of people, using a restricted subject vocabulary (Newell et al., 1971).
Roberts committed $3 million per year for 5 years, with the intention of pursuing a 5-year follow-on project. Major SUR project groups were established at CMU, SRI, MIT's Lincoln Laboratory, Systems Development Corporation (SDC), and Bolt, Beranek, and Newman (BBN). Smaller contracts were awarded to a few other institutions. Five years later, SUR products were demonstrated. CMU researchers demonstrated two systems, HARPY and HEARSAY-I, and BBN developed Hear What I Mean (HWIM). The system developed cooperatively by SRI and SDC was never tested (Green, 1988). The system that came the closest to satisfying the original project goals--and may have exceeded the benchmarks--was HARPY, but controversy arose within DARPA and the AI community about the way the tests were handled. Full details regarding the testing of system performance had not been worked out at the outset of the SUR program.20 As a result, some researchers--including DARPA research managers--believed that the SUR program had failed to meet its objectives. DARPA terminated the program without funding the follow-on.21 Nevertheless, industry groups, including those at IBM, continued to invest in this research area and made important contributions to the development of continuous speech recognition methods.22
DARPA began funding speech recognition research on a large scale again in 1984 as part of the Strategic Computing Program (discussed later in this chapter) and continued funding research in this area well into the late 1990s. Many of the same institutions that had been part of the SUR program, including CMU, BBN, SRI, and MIT, participated in the new initiatives. Firms such as IBM and Dragon Systems also participated. As a result of the controversy over SUR testing, evaluation methods and criteria for these programs were carefully prescribed though mutual agreements between DARPA managers and the funded researchers. Some researchers have hailed this development and praised DARPA's role in benchmarking speech-recognition technology, not only for research purposes but also for the commercial market.
By holding annual system evaluations on carefully designed tasks and test materials, DARPA and the National Bureau of Standards (later the National Institute of Standards and Technology) led the standards-definition process, drawing the participation of not only government contractors but also industry and university groups from around the world, such as AT&T, Cambridge University (of the United Kingdom), and LIMSI (of France). The overall effect was the rapid adoption of the most successful techniques by every participant and quick migration of those techniques into products and services. Although it resulted in quick diffusion of successful techniques, this approach may also have narrowed the scope of approaches taken. Critics have seen this as symptomatic of a profound change in DARPA's philosophy that has reduced the emphasis on basic research.
DARPA's funding of research on understanding speech has been extremely important. First, it pushed the research frontiers of speech recognition and AI more generally. HEARSAY-II is particularly notable for the way it parsed information into independent knowledge sources, which in turn interacted with each other through a common database that CMU researchers labeled a "blackboard" (Englemore et al., 1988). This blackboard method of information processing proved to be a significant advance in AI. Moreover, although early speech-recognition researchers appeared overly ambitious in incorporating syntax and semantics into their systems, others have recently begun to adopt this approach to improve statistically based speech-recognition technology.
Perhaps more important, the results of this research have been incorporated into the products of established companies, such as IBM and BBN, as well as start-ups such as Nuance Communications (an SRI spinoff) and Dragon Systems. Microsoft Corporation, too, is incorporating speech recognition technology into its operating system (DARPA, 1997; McClain, 1998). The leading commercial speech-recognition program on the market today, the Dragon Systems software, traces its roots directly back to the work done at CMU between 1971 and 1975 as part of SUR (see Box 9.2). The DRAGON program developed in CMU's SUR project (the predecessor of the HARPY program) pioneered the use of techniques borrowed from mathematics and statistics (hidden Markov models) to recognize continuous speech (Baker, 1975). According to some scholars, the adoption of hidden Markov models by CMU's research team owes much to activities outside the AI field, such as research by engineers and statisticians with an interest in machine intelligence.23
Other examples of commercial success abound. Charles Schwab and Company adopted DARPA technology to develop its VoiceBroker system, which provides stock quotes over the telephone. The system can recognize the names of 13,000 different securities as well as major regional U.S. accents. On the military side, DARPA provided translingual communication devices for use in Bosnia. These devices translated spoken English phrases into corresponding Serbo-Croatian or Russian phrases. The total market for these new personal-use voice recognition technologies is expected to reach about $4 billion in 2001 (Garfinkel, 1998).
SHIFT TO APPLIED RESEARCH INCREASES INVESTMENT
Although most founders of the AI field continued to pursue basic questions of human and machine intelligence, some of their students and other second-generation researchers began to seek ways to use AI methods and approaches to tackle real-world problems. Their initiatives were important, not only in their own right, but also because they were indicative of a gradual but significant change in the funding environment toward more applied realms of research. The development of expert systems, such as DENDRAL at SAIL, provides but one example of this trend (see Box 9.3).
The promotion in 1969 of Lawrence Roberts to director of IPTO also contributed to a perceived tightening of AI research. Under Roberts, IPTO developed a formal AI program, which in turn was divided into formal subprograms (Norberg and O'Neill, 1996). The line-item budget of AI research inevitably led to greater scrutiny owing to reporting mechanisms and the need to justify programs to the DOD, the Administration, and the U.S. Congress. Consequently, researchers began to believe that they were being boxed in by IPTO and DARPA, and to a certain extent they were. The flow of DARPA's AI research money to CMU, MIT, and Stanford University did not cease or even diminish much, but the demand grew for interim reports and more tangible results.
External developments reinforced this shift. The most important was the passage of the Mansfield Amendment in 1969.24 Passed during the Vietnam War amid growing public concern about the "military-industrial complex" and the domination of U.S. academic science by the military, the Mansfield Amendment restricted the DOD to supporting basic research that was of "direct and apparent" utility to specific military functions and operations. It brought about a swift decline in some of the military's support for basic research, often driving it toward the applied realm.25 Roberts and his successors now had to justify AI research programs on the basis of immediate utility to the military mission. The move toward relevance spawned dissatisfaction among both the established pioneers of the AI field and its outside skeptics.26
Another external development provided further impetus for change. In 1973, at the request of the British Scientific Research Council, Sir James Lighthill, the Lucasian Professor of Applied Mathematics at Cambridge University and a Fellow of the Royal Society of London, produced a survey that expressed considerable skepticism about AI in general and research domains in particular. Despite having no expertise in AI himself, Lighthill suggested that any particular successes in AI had stemmed from modeling efforts in more traditional disciplines, not from AI per se. He singled out robotics research for especially sharp criticism. The Lighthill report raised questions about AI research funding in the United States and led DOD to establish a panel to assess DARPA's AI program.
Known as the American Study Group, the panel (which included some of AI's major research figures) raised some of the same questions as did Lighthill's report and served to inform George Heilmeier, a former research manager from RCA Corporation who was then assistant director of Defense R&D and later became director of DARPA. The Lighthill report and its U.S. equivalent led to a shifting of DARPA funds out of robotics research (hurting institutions such as SRI that had committed heavily to the area) and toward "mission-oriented direct research, rather than basic undirected research" (Fleck, 1982).27
As a result of these forces, DARPA's emphasis on relevance in AI research grew during the late 1970s and 1980s. Despite the disgruntlement among some scientists, the changes led to increased funding--although not directly to widespread commercial success--for AI research. A magnet for these monies was the Strategic Computing Program (SCP), announced in 1983 (DARPA, 1983). DARPA committed $1 billion over the planned 10-year course of the program. The four main goals of the SCP were as follows:
To achieve these goals, DARPA established three specific applications as R&D objectives: a pilot's associate for the Air Force, an autonomous land vehicle for the Army, and an aircraft battle management system for the Navy. The applications were intended to spark the military services' interest in developing AI technology based on fundamental research. The SCP differed from some other large-scale national efforts in that its goals were extremely ambitious, requiring fundamental advances in the underlying technology. (By contrast, efforts such as the Apollo space program were principally engineering projects drawing from an established scientific base [Office of Technology Assessment, 1985]). The SCP also differed from earlier large AI programs in that some 60 percent of its funds were committed to industry. However, of the 30 prime contractors for the SCP involved in software or AI research, more than 20 were established defense contractors (Goldstein, 1992).
The SCP significantly boosted overall federal funding for AI research but also altered its character. Between 1984 and 1988, total federal funding for AI research, excluding the SCP, tripled from $57 million to $159 million (see Table 9.1). With support for the SCP included, federal funding increased from $106 million to $274 million. Because the SCP was budgeted as an applied program, it tipped the balance of federal funding toward applied research. Although DARPA's funding for basic AI research doubled from roughly $20 million to $40 million during this same period, the DOD's overall role in basic AI research declined (see Table 9.2). Meanwhile, it continued to play the dominant role in supporting applied research in AI (see Table 9.3). Although budget categorizations for programs such as the SCP are somewhat arbitrary and subject to political influence, researchers noted a change in DARPA's funding style.
The SCP also attracted a tremendous amount of industry investment and venture capital to AI research and development. Firms developing and selling expert systems entered the market, often basing their systems on the LISP machines developed by the AI community. Several new firms entered the market to design, make, and sell the very expensive LISP machines. Yet the rapid development of engineering workstations, especially those of Sun Microsystems, Inc., soon undermined the LISP machine industry. This segment of the market, which was clearly tied to the SCP, collapsed. Even with the development of expert-system shells to run on less-costly machines, doubts began to arise about the capabilities and flexibility of expert systems; this doubt hampered the commercialization of AI. In addition, commercial contractors had difficulty meeting the high-profile milestones of the major SCP projects because of difficulties with either the AI technologies themselves or their incorporation into larger systems. Such problems undermined the emergence of a clearly identifiable AI industry and contributed to a shift in emphasis in high-performance computing, away from AI and toward other grand challenges, such as weather modeling and prediction and scientific visualization.
ARTIFICIAL INTELLIGENCE IN THE 1990s
Despite the commercial difficulties associated with the Strategic Computing Program, the AI-driven advances in rule-based reasoning systems (i.e., expert systems) and their successors--many of which were initiated with DARPA funding in the 1960s and 1970s--proved to be extremely valuable for the emerging national information infrastructure and electronic commerce. These advances, including probabilistic reasoning systems and Bayesian networks, natural language processing, and knowledge representation, brought AI out of the laboratory and into the marketplace. Paradoxically, the major commercial successes of AI research applications are mostly hidden from view today because they are embedded in larger software systems. None of these systems has demonstrated general human intelligence, but many have contributed to commercial and military objectives.
An example is the Lumiere project initiated at Microsoft Research in 1993. Lumiere monitors a computer user's actions to determine when assistance may be needed. It continuously follows the user's goals and tasks as software programs run, using Bayesian networks to generate a probability distribution over topic areas that might pose difficulties and calculating the probability that the user will not mind being bothered with assistance. Lumiere forms the basis of the "office assistant" that monitors the behavior of users of Microsoft's Office 97 and assists them with applications. Lumiere is based on earlier work on probabilistic models of user goals to support the display of customized information to pilots of commercial aircraft, as well as user modeling for display control for flight engineers at NASA's Mission Control Center. These earlier projects, sponsored by the NASA-Ames Research Center and NASA's Johnson Space Center, were undertaken while some of the Lumiere researchers were students at Stanford University.28
Patent trends suggest that AI technology is being incorporated into growing numbers of commercial products. The number of patents in AI, expert systems, and neural networks jumped from fewer than 20 in 1988 to more than 120 in 1996, and the number of patents citing patents in these areas grew from about 140 to almost 800.29 The number of AI-related patents (including patents in AI, expert systems, neural networks, intelligent systems, adaptive agents, and adaptive systems) issued annually in the United States increased exponentially from approximately 100 in 1985 to more than 900 in 1996 (see Figure 9.1). Changes in the U.S. Patent and Trademark Office's rules on the patentability of algorithms have no doubt contributed to this growth, as has the increased commercial value of AI technology. The vast majority of these patents are held by private firms, including large manufacturers of electronics and computers, as well as major users of information technology (see Table 9.4). These data indicate that AI technology is likely to be embedded in larger systems, from computers to cars to manufacturing lines, rather than used as stand-alone products.
A central problem confronting the wider commercialization of AI today revolves around integration. Both the software and the hardware developed by the AI research community were so advanced that their integration into older, more conservative computer and organizational systems proved to be an enormous challenge. As one observer has noted, "Because AI was a leading-edge technology, it arrived in this world too early. As a consequence, the AI application community had to ride many waves of technological quick fixes and fads. . . . Many of these integration problems are now being addressed head on by a broad community of information technologists using Internet-based frameworks such as CORBA [common object request broker architecture] and the World Wide Web" (Shrobe, 1996).
The rapid development of computer hardware and software, the networking of information systems, and the need to make these systems function smoothly and intelligently are leading to wide diffusion of AI knowledge and technology across the infrastructure of the information age. Federal funding reflects these changes (see Box 9.4). Meanwhile, much of the knowledge acquired through AI research over the years is now being brought to bear on real-world problems and applications while also being deepened and broadened. The economic and social benefits are enormous. Technologies such as expert systems, natural-language processing, and computer vision are now used in a range of applications, such as decision aids, planning tools, speech-recognition systems, pattern recognition, knowledge representation, and computer-controlled robots.30 AI technologies help industry diagnose machine failures, design new products, and plan, simulate, and schedule production. They help scientists search large databases and decode DNA sequences, and they help doctors make more-informed decisions about diagnosis and treatment of particular ailments. AI technologies also make the larger systems into which they are incorporated easier to use and more productive. These benefits are relatively easy to identify, but measuring them is difficult.
Federal investments in AI have produced a number of notable results, some envisioned by the founders of the field and others probably not even imagined. Without question, DARPA's generous, enduring funding of various aspects of AI research created a scientific research discipline that meets the standard criteria of discipline formation laid out by sociologists of science.31 At least three major academic centers of excellence and several other significant centers were established, and they produced a large number of graduates with Ph.D.s who diffused AI research to other research universities, cross-pollinated the major research centers, and moved AI methods into commercial markets. (Figure 9.2 shows the production of Ph.D. degrees in AI and related fields at U.S. universities. Figure 9.3 compares Ph.D. production in AI and related disciplines to degree production in computer science more broadly.) In sum, the returns on the public investment are clearly enormous, both in matters of national security (which are beyond the scope of this study)32 and in contributions to the U.S. economy.
LESSONS FROM HISTORY
As this case study demonstrates, federal funding is critical in establishing new disciplines because it can sustain long-term, high-risk research areas and nurture a critical mass of technical and human resources. DARPA helped legitimize the AI field and served as the major source of research funds beginning in the 1960s. It created centers of excellence that evolved into today's major computer science research centers. This support was particularly critical given that some objectives took much longer to realize than was originally anticipated.
A diversity of approaches to research problems can be critical to the development of practical tools. A prime example is the field of speech recognition, in which the most effective products to date have used techniques borrowed from the mathematics and statistics communities rather than more traditional AI techniques. This outcome could not have been predicted and demonstrates the importance of supporting competing approaches, even those outside the mainstream.
Federal funding has promoted innovation in commercial products such as expert systems, the establishment of new companies, the growth of billion-dollar markets for technologies such as speech recognition, and the development of valuable military applications. AI technologies often enhance the performance of the larger systems into which they are increasingly incorporated.
There is a creative tension between fundamental research and attempts to create functional devices. Original attempts to design intelligent, thinking machines motivated fundamental work that created a base of knowledge. Initial advances achieved through research were not sufficient to produce, by themselves, commercial products, but they could be integrated with other components and exploited in different applications. Efforts to apply AI technology often failed initally because they uncovered technical problems that had not yet been adequately addressed. Applications were fed back into the research process, thus motivating inquiries into new areas.
1 Several histories of AI research have appeared over the last 25 years, some written by members of the AI community, some published by those outside the field, and still others produced by science journalists. These histories have been based largely on the published scientific literature, journalistic accounts of work in the field, published accounts of participants in the field, and interviews with participants. With some notable exceptions, few of these histories have relied on original source materials, such as manuscript records of participants or their funding agencies or editors.
2 The 1997 victory of IBM Corporation's Deep Blue Computer over world chess champion Gary Kasparov demonstrates the public's interest in AI. In the days leading up to the match and throughout the match itself, almost every major U.S. newspaper, news magazine, and television news or magazine program carried news and feature articles about the match and AI research in general.
3 Papers presented at the conference were published by Shannon and McCarthy (1956).
4 Minsky had an impressive background: a bachelor's degree in mathematics from Harvard University (1950); a doctorate in mathematics from Princeton University (1954); and the title of junior fellow, Harvard Society of Fellows (1954-1957). His 1954 dissertation was entitled "Neural Nets and the Brain Model Problem." The paper he presented at the 1952 conference was entitled "Some Universal Elements for Finite Automata." His early work at Lincoln Laboratory (Minsky, 1956) dealt with AI.
5 Rochester was the chief designer of IBM's 701 computer.
6 Dated August 31, 1955, "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence" was actually submitted, along with a cover letter to Morison, on September 2, 1955, according to the Rockefeller Foundation Archives grant files. An edited and lightly annotated version of this proposal can be found online at <http://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html>.
7 This quote comes from an article entitled "LISP Prehistory--Summer 1956 through Summer 1958," which can be found online at <http://www-formal.stanford.edu/jmc/history/lisp/node2.html#SECTION00020000000000000000>. See also McCarthy (1981).
8 Indeed, McCarthy's first memorandum on LISP (dated September 16, 1958) survives only because Rochester saved and annotated it. Rochester was the author of at least one of the foundational LISP memoranda; his Memo 5, November 18, 1958, provides clear evidence of Rochester's intellectual contributions to LISP. For an interesting treatment of LISP's history, see "Early LISP History (1956-1959)" by Herbert Stoyan, available online at <http://www8.informatik.uni-erlangen.de/html/lisp/histlit1.html>.
9 Rochester also published papers on AI (e.g., Rochester and Gelernter, 1958).
10 The founding of GSIA and its "new look" are described by Gleeson and Schlossman (1994).
11 The early history of RAND is described by Smith (1966), Jardini (1996), Hounshell (1997), and Collins (1998).
12 This report describes the Logic Theorist and the IPL2 list processing language (developed with J.C. Shaw) for the Johnniac. See also Newell, Simon, and Shaw (1957).
13 The history of the Johnniac is recounted by Gruenberger (1979).
14 Another of Simon's doctoral students, Edward Feigenbaum, as part of his 1960 doctoral dissertation developed a theory of human perception, memory, and learning and then modeled these processes successfully in his EPAM program. This program is still regarded as a major contribution both to theories of human intelligence and to AI research.
15 This quote comes from an article entitled "The Implementation of LISP" available online at <http://www-formal.stanford.edu/jmc/history/lisp/node3.html#SECTION00030000000000000000>.
16 The history and design of both the Logic Theorist and GPS are described by Newell and Simon (1972).
17 For example, bold predictions about the future of AI were made by Simon and Newell (1958); Simon, Newell, and Shaw (1958); Simon (1965); Minsky (1956, 1979).
18 DARPA definitely created a two-tier system in AI research, with CMU, MIT, Stanford University, and SRI occupying the top tier. The second tier included the University of Massachusetts, University of Maryland, Brown University, University of Pennsylvania, New York University, Columbia University, Rutgers University, University of Texas, and University of Illinois.
19 This was the so-called Pierce Committee, named after its chairperson J.R. Pierce, who, according to Roberts, said it would be impossible to make a computer understand human speech (Roberts, "Expanding AI Research"). A researcher at Bell Laboratories (where important speech recognition research had been done for decades), Pierce published an infamous letter to the editor of the Journal of the Acoustical Society of America in 1969 in which he railed against the "mad scientist and untrustworthy engineers" who believed that the development of a flexible, continuous speech recognition system was possible (Pierce, 1969).
20 One of the characteristics of AI that sets it apart from many other research domains of computer science is the degree to which the objectives can be described in general, nontechnical language that makes measuring performance difficult (e.g., an objective might be to make a machine that thinks or learns). Although specific metrics often can be developed, this may require a conscious effort by program managers and researchers.
21 Considerable controversy surrounded this decision and its motivations. J.C.R. Licklider claimed the project was turned off because "it didn't really prove itself by coming up with a conspicuously good demonstrable device. There was a lot of feeling that speech understanding was a bit ahead of its time, and DARPA did a good thing by stimulating the field but it wasn't really time yet to drive for a workable system. In this case, the speech project met its main objectives, but that wasn't enough to save it" (Licklider, 1988a). Some observers suggest that the demise of the SUR program illustrates "the dangers of prematurely imposing short-term goals and milestones on an underdeveloped field of research" such as speech recognition (Stefik, 1985). Marvin Denicoff, then at ONR, believed so strongly in the speech program and what it could do for future researchers that he convinced Robert Kahn, director of IPTO from 1979 to 1985, to fund a project to study SUR--to "take a year or two out, visit all the sites and create a document of everything that had been accomplished and what the issues were, what the failures were, and what the positives were" (Denicoff, 1988). The results were the ONR reports by W. Lea and J. Shoup (1979) and W. Lea (1980).
22 Here the allusion is to the work of Frederick Jelinek and the speech research group at IBM, which contributed enormously to the technology through statistical language modeling (e.g., N-gram models). Other firms that not only pursued speech recognition research but also entered commercial markets with related products included Verbex Voice Systems and Texas Instruments.
23 For example, L.E. Baum and J.A. Eagon of the Institute for Defense Analyses have been credited with introducing HMM theory (Makhoul and Schwartz, 1994). See also Baum and Eagon (1967). Moreover, even within CMU, the decision to use HMMs in speech recognition was, according to a recent analysis, "[w]ay out of step with the mainstream [of AI thought at CMU]. . . . The system had no knowledge of English grammar, no knowledge base, no rule-based expert system, no intelligence. Nothing but numbers" (Garfinkel, 1998).
24 The Mansfield Amendment was passed as part of the Defense Authorization Act of 1970 (Public Law 91-121) on November 19, 1969.
26 The Mansfield Amendment and the spirit of the times in which it was passed also established the conditions under which some members of the Congress would raise questions about AI research, particularly with regard to its focus on the playing of games such as chess and checkers but extending to the research on speech understanding, which is discussed in Box 9.2.
27 Fleck (1982) also said the Lighthill report "led to a considerable sharpening of the focus of AI research." Not all leaders in AI research agree with Fleck's assessment of the impact of the report. Amarel (1988), for example, maintained that "it didn't affect this country."
28 For more information, see Microsoft's home page at <www.research.microsoft.com/research/dtg/horovitz/lum.htm>.
29 These data were obtained from the U.S. Patent and Trademark Office database, available online at <http://patents.uspto.gov>, and IBM's patent database, also available online at <http://patent.womplex.ibm.com>.
30 These technologies are the results of the efforts of AI researchers as well as researchers in other, related fields.
31 See, for example, Fleck (1982).
32 A report by the American Association for Artificial Intelligence (1994) paraphrased a former director of ARPA in saying that DART (the intelligent system used for troop and materiel deployment for Operation Desert Shield and Operation Desert Storm in 1990 and 1991) "justified ARPA's entire investment in artificial-intelligence technology." (The report is also available on the association's Web site at <http://www.aaai.org/Policy/Papers/arpa-report.html>).