Box 9.3

Pioneering Expert Systems

The DENDRAL Project was initiated in 1965 by Edward Feigenbaum (one of Herbert Simon's doctoral students in AI); Nobel Prize-winning geneticist and biochemist Joshua Lederberg; and Bruce Buchanan, a recent recipient of a doctorate in philosophy from Michigan State University.¹ DENDRAL began as an effort to explore the mechanization of scientific reasoning and the formalization of scientific knowledge by working within a specific domain of science, organic chemistry. Developed in part with an initial research grant from the National Aeronautics and Space Administration (in anticipation of landing unmanned spacecraft on other planets), but also picked up under DARPA funding, DENDRAL used a set of knowledge- or rule-based reasoning commands to deduce the likely molecular structure of organic chemical compounds from known chemical analyses and mass spectrometry data. The program took almost 10 years to develop, combining the talents of chemists, geneticists, and computer scientists. In addition to rivaling the skill of expert organic chemists in predicting the structures of molecules in certain classes of compounds, DENDRAL proved to be fundamentally important in demonstrating how rule-based reasoning could be developed into powerful knowledge engineering tools. Its use resulted in a number of papers published in the chemistry literature. Although it is no longer a topic of academic research, the most recent version of the interactive structure generator, GENOA, has been licensed by Stanford University for commercial use.

DENDRAL led to the development of other rule-based reasoning programs at the Stanford Artificial Intelligence Laboratory (SAIL), the most important of which was MYCIN, which helped physicians diagnose a range of infectious blood diseases based on sets of clinical symptoms.² Begun in 1972 and completed in 1980, the MYCIN project went further than DENDRAL in that it kept the rules (or embodied knowledge) separate from the inference engine that applied the rules. This latter part of the MYCIN project was essentially the first expert-system shell (Buchanan and Shortliffe, 1984).³

The development of these pioneering expert systems not only constituted major achievements in AI but also gave both researchers and research funders a glimpse of the ultimate power of computers as a tool for reasoning and decision making. Moreover, the apparent success of these projects helped to touch off the rapid development of expert systems. Promoted by SAIL's Edward Feigenbaum, expert systems became the rage in AI research in the late 1970s and early 1980s and a commercial tool in the 1980s, when corporations were seeking to embody the knowledge of their skilled employees who were facing either retirement or downsizing (Feigenbaum et al., 1988). Expert-system shells, based in large part on the "Empty MYCIN" (EMYCIN) shell, moved on to the commercial software market.

Starting in the mid-1980s, numerous start-up AI companies began to appear, many with products akin to expert systems. Many such companies came and went, but some flourished. For example, Gensym Corporation, founded in 1986 by an alumnus of the Massachusetts Institute of Technology's Artificial Intelligence Laboratory, built a substantial business based on its G2 product for development of intelligent systems. More recently, Trilogy Development Group, Inc., went public, selling both software and services that apply rule-based reasoning and other AI methods to marketing operations. One of Trilogy's founders (a Stanford University graduate) learned about the expert system that Carnegie Mellon University (CMU) had developed for Digital Equipment Corporation to configure its VAX computers (XCON).⁴ Basing their work in part on the systems that had emerged from DENDRAL and MYCIN and what they learned about XCON, Trilogy's founders also used constraint-based equations and object-oriented programming methods, derived in part from AI research.⁵ Another of Trilogy's founders applied the company's methods to the marketing of personal computers (PCs) over the Internet. This new firm, pcOrder.com.Inc., promises to simplify the configuration of PCs and drastically lower the cost of buying (or selling) one (McHugh, 1996).

Many corporations committed substantial capital and human resources to the development of expert systems, and many reported substantial returns on these investments. Others found that, as AI pioneer McCarthy (1990) had argued, these expert systems were extremely "brittle" in that a small development in knowledge or change in practice rendered such programs obsolete or too narrow to use. In one study of AI (Office of Technology Assessment, 1985), expert systems were singled out as evidence of "the first real commercial products of about 25 years of AI research" but were also criticized for "several serious weaknesses" that demanded "fundamental breakthroughs" to overcome. But expert systems represented a failure to meet expectations as much as a failure of technology. They provided valuable help for users who understand the limitations of a system that embodied narrow domains of knowledge. One of the biggest problems with expert systems was the term itself, which implied a certain level of capability; a number of users started calling them knowledge-based systems to refer to the technology instead of the goal.

Despite these criticisms, work on expert systems continues to be published; some corporations with strong knowledge-engineering capabilities continue to report substantial savings from expert systems and have demonstrated a continued commitment to expanding their use. Expert-system shell programs continue to be developed, improved, and sold. By 1992, some 11 shell programs were available for the MacIntosh platform, 29 for IBM-DOS platforms, 4 for Unix platforms, and 12 for dedicated mainframe applications.⁶ A recent review of expert systems reported that the North American market for expert systems is roughly $250 million (representing about 70 percent of the total commercial AI market). Estimates suggest that more than 12,000 stand-alone expert systems are in use (Liebowitz, 1997). Moreover, small expert systems are being incorporated into other types of computer software, most of it proprietary.

² Another important program, carried out in the early 1970s, was META-DENDRAL. This inductive program automatically formulates new rules for DENDRAL to use in explaining data about unknown chemical compounds. Using the plan-generate-test paradigm, META-DENDRAL has successfully formulated rules of mass spectrometry, both by rediscovering existing rules and by proposing entirely new rules. Although META-DENDRAL is no longer an active program, its contributions to ideas about learning and discovery are being applied to new domains. These ideas suggest, for example, that induction can be automated as a heuristic search; that, for efficiency, search can be broken into two steps--approximate and refined; that learning must be able to cope with noisy and incomplete data; and that learning multiple concepts at the same time is sometimes inescapable. More information is available online at <http://www-camis.stanford.edu/research/history.html#DENDRAL>.

³ The MYCIN team also developed an important program known as TEIRESIAS, which made the basis of MYCIN's reasoning transparent to its users and allowed MYCIN's knowledge base to be changed or upgraded more easily. The literature on these programs is extensive.

⁴ XCON (for eXpert Configurer) has been widely hailed as one of the first successful expert systems programs. It was the work of John McDermott and his team at CMU. See Crevier (1994).

⁵ Trilogy's history is discussed by McHugh (1996).

⁶ These data were compiled from R.R. Bowker Company (1992) and Table 3 in Pickett and Case (1991).