Artffkfd Isadlfgesw: Using Computers to Tlslnk about Thirskfng. Part 2. Some Pmctfcsd Applicatfom of AI

Number 52 December 26, 1983

This is the second of a two-part essay liited success. Techniques developed on (AI). Part 1 dis- for solving problems in one domain were cussed “knowledge representations”- usually inadequate for other domains. models of cognition that AI investigators However, programs equipped with a have used in their attempts to build great deal of information about a single thinking machines. 1 One of the most domain performed as well as, and some- ambitious goals of AI research is to de- times better than, experts in that field.z velop programs which enable machines The superior performance of these to perform commonsense reasoning. It knowledge-based, or expert, systems will & many years, however, before th~ convinced many AI researchers that goal is achieved. problem solving demands huge banks of In the meantime, AI research has knowledge as well as reasoning proce- spawned numerous spin-offs that are dures.3 just beginning to enter the commercial Today, the goal of expert systems re- market. These include programs that en- search is to transfer a specialist’s knowl- able robots to “see” and “feel” and ma- edge into a program so the information chines which can follow instructions can be efficiently accessed and used by written in natural language (NL). But the the computer to solve problems. This in- most ambitious and successful AI appli- cludes “textbook leaming’’—the facts cations thus far are “expert systems. ” obtained from training and reading.4 It These computer programs are designed also includes heuristic knowledge, or to duplicate the problem-solving pro- rules of thumb developed through years cesses of experts in various fields. This of experience and judgment. Heuristics essay will cover some of these expert sys- are essentially educated guesses about tems and review a liited number of which solutions to a problem are most other applications of AI. likely to be successful. They don’t guar- During the early years of AI research, antee correct answers. But they do save investigators were intent upon dMcover- time by limiting the search for solutions ing the general principles underlying in- to those most likely to be correct. Com- telligence. In developing the knowledge puter and other scientists called “knowl- representations described in Part 1, AI edge engineers” sometimes spend years researchers tried to use these principles picking experts’ brains for these facts to create an “inference engine.” Ideally, and heuristics, and then structure them such an engine would solve any type of into computer programs. problem, from winning at chess to d~ag- Expert systems also include “infer- nosing disease. But attempts at develop- ence procedures” that determine which ing computer-based problem-solving heuristics and facts should be brought to strategies for any situation met with bear on a problem. One such inference

430 procedure is backward chaiiing, in in modeliig and assisting scientific which you suggest a possible solution to thinkiig, and Lederberg, a geneti- a problem and work backward to see if cist/molecular biologist with expertise it’s correct. MYCIPJ, an in chemistry, had developed a computer that assists in medical diagnoses, reasons language for describing the structure of in this manner.z It advances a disease complex molecules.g As the project hypothesis based on a few known symp- grew, , also at Stanford tom$. Then it looks for other symptoms Univemity, contributed his expertise. 10 which support the hypothesis, request- The original DENDRAL program per- ing additional tests and information as formed three basic operations on the needed. In most expert systems, the in- chemical formulas and mass spectral ference procedure for deciding which data with which it was provided. In the facts and heuristics to use is separate fmt phase, cafled “plan,” it translated from the knowledge base of facts and general and specific prior knowledge rules. This makes it easier to add or and heuristics into a spec~lc repertoire modify facts and rules as new informa- of constraints. In the next phase, called tion becomes available. “generation,” it generated plausible The fret, and probably best-known, structures based on such constraints as AI programs that focused on a Ihnited the number of rings, double bonds, and domain were chess programs. Donald atoms of various types in each molecule. Michie, University of Edinburgh, Scot- In the final phase, caUed “test,” each land, explains that chess is ideal for mod- plausible structure was tested. The com- eliig specialist knowledge because it is a puter fiit generated sets of instrument very well-defined domains A large data that would be expected to describe amount of formal information is avail- each structure. Then it compared each able in the form of instructional works set of data to actual data about the com- and commentaries. And numerical pound. The closest fits were then ranked scales of performance are avaifable in for the user. the national and international rating In the past few years, scientists work- systems. Equally important, chess is a ing on the DENDRAL project have fo- game that calls on a wide range of cogni- cused most of their attention on the tive functions, from logical calculation planning and generation portion of the to imaginative thinking. The numerous program, and on making this portion chess-playing programs developed in the available to users. Called CONGEN, for 1950s and 1960s tested the proficiency constrained structure generation, this with which various knowledge represen- generator has been expanded to infer tations used facts and heuristics to solve plausible structures using a wide variety problems.s of instrumental data. In 1982. CONGEN By the mid- 1960s, AI researchers was made commercially avaifable began to expand beyond chess and puz- through the computer network Compu- zle playing, or what Edward A. Feigen- Serve. II baum, Stanford University, California, Another spin-off of DENDRAL is calls “toy problems, “G(p. 62) into practi- Mets-DENDRAL, a program that gen- cal problems. The fiist such project re- erates its own rules from mass spectral sulted in DENDRAL, an expert system data on chemical compounds. After re- that identifies the chemicaf structures of ceiving mass spectral data on a family of unknown compounds.7,8 DENDRAL compounds, Mets-DENDRAL gener- was launched at Stanford University in ates planning and test rules that describe 1966 by Feigenbaum and Joshua Leder- how these compounds fragment when berg, now president of Rockefeller Uni- studied with mass spectrometry. Some versity, New York. Both were interested of the roles generated by Meta-

431 DENDRAL duplicated those formulated solve problems in that domain. Figure 1 by expert chemists, while others were shows one of the 500 such rules used by entirely original. 1Z the infectious diseases expert system DENDRAL demonstrated that AI MYCIN. With MYCIN, a physician en- techniques could be used to solve real ters information about a patient into the problems within a Iiiited area of knowl- computer. The computer then searches edge.b Paradoxically, it also demon- for the rules that can be applied to this strated that it is easier to model the rea- information. If more information is soning processes of specialists than to needed, the computer will ask the physi- program the steps a chfld goes through cian to supply it. Then the rules for these in understanding language, or making additional data are applied. This process commonsense inferences.z This is continues until a diagnosis, and treat- because the facts and judgments an ex- ment, can be recommended. pert uses in making a decision are easier Since the acceptability of an expert to identify and categorize than are the system depends on the confidence with reasoning processes used for general which physicians can accept its sugges- problem solving. tions, MYCIN and several other consul- So far, the most successful expert sys- tation systems can provide explanations tems have been programs that weigh and of their reasoning processes. At the phy- balance evidence about data to deter- sician’s request, MYCIN will list the pro- mine how they should be categorized. duction rules it used in its diagnosis or Differential diagnosis, for example, is “a treatment recommendation, and cite classical medical example of such a references to the literature that support problem,”z according to Richard O. these rules. MYCIN will also explain Duda, Syntelligence, Menlo Park, Cali- why it has requested additional tests or fornia, and Edward H. Shortliife, Stan- other information.z An example of thk ford University School of Medicine. A explanation feature is shown in Figure 2, physician arrives at a diagnosis by evaluating a variety of symptoms and Ffgure 1: Sample MYCIN production rule. test results. Although this is a fairly complicated procedure, it is based on Mycfri identifiable facts and heuristics and, [f: 1) The infection which requires therefore, lends itself to computer therapy is meningitis, modeling. For this reason, and because And 2) The patient has evidence of a computers can consider many diseases serious skin or soft tissue infection, that physicians might not encounter in everyday practice, numerous expert And 3) Organisms were not seen on the systems have been designed to assist stain of the culture, doctors in diagnosing and treating disease.z Amf 4) The type of infection is bacterial, The knowledge representation used risen: most widely in these expert or “consulta- tion” systems is the “production rule” There is evidence that the organism approach, according to Wi~lam B, (other than those seen on cultures or Gevarter, National Aeronautics and smears) which might be causing the in- Space Administration, Washington, fection is staphyloccwcus-coagpos (.75) DC. 13 As mentioned earlier, each pro- or streptococcus (.5). duction rule is a heuristic, also called an Source: Gevmtter W R. Expwt systems: limited but power- “if-then” rule or “condition-action” pair. ful. IEEE Spemum 20(8):39-45, 1983. Each rule or set of rules includes facts about a domain which can be used to @ IEEE

432 Figura 2: Explanation features of MYCIN, Thii excerpt from the MYCIN run illustrates some of the explanation facilities provided by the program. At entry 38, MYCINaskswhetherthe patient has been seriously burned. Instead of answering, the user asks why this particular question is being asked, In respnn~, the program states bnth the goal of the question and the rule it was attempting to apply. This ex- planation leads the user to wonder how MYC2N established that the infection might be bacterial; in response to the HOW command, the program fists the five rules that it appfied, four of which supported this conclusion. This abiMy to inspect the program’s reasoning methcds provides the user with a firmer basis for undemtandmg the final conclusinrrs.

34) Is the meningitis a hospital-acquired infection? ““ NO 35) Is Pt219’scfinical history consistent with EPIGLOTTITfS? ““ NO 36) Is Pt219’sclinical history consistent with OTITIS-fvfEDfA? ““ NO 37) Has the patient ever been to an area that is endemic for cnccidlomycoaes? ““ UNKNOWN W Is Pt219 a bum patient? “● WHY

[i.e. WHY is it important to determine whether PATIENT-219 has been seriously burned?]

[4.0] This wiff aid in determining the organisms (other than those seen on cultures or smeara) which might be causing the infection. It has afmady been established that [4. 1] the infection which requires therapy is meningitis, and [4.2] organisms were not seen on the stain of the pending csf culture (CULTURE-1), and [4.3] the type of the infection maybe bacterial Therefore, if [4.41 Pt219 has been aerioualy burned then there is suggestive evidence that Paeudomonaa-aeruginoaa is one of the organisms (other than those seen on cultures or smears) which might be causing the infection Author: Yu [RULE578]

[back to question 38... ] ‘* HOW 4.3

[i.e. HOW was it established that the type of the infection maybe bacterial?]

The following rules concluded about: the likelihood that the type of the infection may be bacterial

Rule Yes No

[4.5] RULE500 (.4) [4.61 RULE501 (.17) [4.7] RULE5412 (.1) [4.8] RULE524 (,42) [4.9] RULE504 (.2.s)

Thus, it haa been established that the type of the infection may be bacterial.

Source: Dub R O & ShortftSle E H. Expert systems research. Science 220(4594):261-8, 1983. @ 1953 by the AAAS

MYCIN was developed in the mid- called EMYCIN.14 Since then, 1970s at Stanford University by Bruce EMYCIN has been filled with produc- Buchanan and Shortliife.2 In the late tion rules from a variety of fields, includ- 1970s, W .J. van Melle, Stanford Univer- ing structural engineering and computer sity, transformed MYCIN’S inference repair. Another MYCIN spin-off, called engine into an all-purpose program, GUIDON,lS was developed by W .J.

433 Clancey and colleagues, Stanford Uni- vis, now of MIT, while he was at Stan- versity, to help med~cal students learn ford University, is designed to assist the methods and swategies used by ex- knowledge engineers in creating and perts in their field. updating the knowledge bases used DENDRAL and MYCIN are two of in expert systems. The success of the most influential expert systems. But TEIRESIAS, and other programs that they represent just a fraction of the work “build” expert systems, is crucial to the currently under way in this field. Dozens commercial future of these systems. of systems have been, or are now being, Knowledge engineers presently spend developed at Edinburgh, Purdue, Rut- hundreds of hours collecting and struc- gers, Yale, and Carnegie-Mellon Uni- turing specialist knowledge before an versities, Massachusetts Institute of expert system can be built. Other non- Technology (MIT), and the University biomedical expert systems include of Pittsburgh, to name only a few of the PROSPECTOR ,23 a geology consultant academic institutions involved in basic developed by Duda and colleagues, and applied AI research. The leading while at SRI International, Menlo Park, center for expert systems research in the California, and MACSYMA,ZJ designed US, however, is Stanford University, by Joel Moses, MIT, for solving algebra which houses the Stanford University and calculus problems. Medical Experimental Computer for A number of private firms have also Artificial Intelligence in Medicine entered the AI arena. Digital Equipment (SUMEX-AIM) network, a nationally Corp. and International Business Ma- shared computer network devoted to AI chines Corp. (IBM) are working on sepa- systems in biomedicine. Systems that rate expert systems that diagnose “sick” have come out of this project include computers.j Xerox Corp. and Texas In- PUFF, lb developed by respiratory struments are developing systems to specialists J. Osborn, R.J. Falfat, and B. assist in the design of computer chips. Votteri, Pacific Medical Center, San And Schlumberger Ltd., which employs Francisco, and Stanford University many Al researchers in its three AI computer scientists L. Fagan, P. Nii, laboratories, is developing a system that J.C. Kunz, J.S. Aikins, and D. McClung. analyzes data on geologic formations.3 PUFF diagnoses and recommends thera- Several companies have also been pies for pulmonary dysfunction. launched by established AI scientists. PARRY, a program that simulates para- Feigenbaum and several of his col- noid thought processes, was developed leagues at Stanford University started on SUMEX-AIM by K.M. Colby, Uni- Teknowledge Inc. and Intelligenetics versity of Caliornia, Los Angeles, IT and Inc., both in Palo Alto, California. Tek- CASNET/Glaucoma was designed for nowledge markets expert systems, and ophthalmology by C.A. Kulikowski and offers consulting and training services to S.M. Weiss, Rutgers University, New companies that want to build their own. Brunswick, New Jersey, and A. Safir, Intelligenetics markets expert systems Mount Sinai School of Medicine, New that assist in gene splicing experiments. York. IS INTERNIST, a pioneering pro- Computers designed for AI research are gram developed by H.E. Pople and J.D. sold by two companies, Lisp Machkes, Myers, University of Pittsburgh, Penn- Inc., and Symbolics Inc., which were sylvania, diagnoses dkeases in internal founded by MIT researchers. medicine. 19.20 It includes some 4,000 Expert systems aren’t the only AI ap- symptoms cross-referenced to about Sot) plications entering the marketplace. diseases .21 Machine Intelligence Corp., Optical Not afl expert systems are biomedical. Recognition Systems Inc., and General TEIRESL4S,22 developed by R. Da- Electric Co. are among a growing num-

434 ber of companies marketing software for LIFER,29 an NL program developed by recognizing visual images.25 ArWlcial G. G. Hendrix while at SRI Internation- Intelligence Corp. and Cognitive Sys- al, for example, interprets the sentence, tems Inc. are marketing NL software “What is the age of Mary’s sister, ” by that enables users to communicate with matching it against a series of grammati- computers in natural languages such as cal questions, such as, “What is the attri- written English.3 The state of the art for bute of the object?” Most NL programs these applications is not as advanced in are also equipped with contextual infor- dealing with technical domains as are ex- mation about the data base. This in- pert systems. 13 But these programs can cludes definitions of words liiely to be function impressively within the limited used in queries, and information on rela- domains for which they’ve been de- tionships between objects in the data signed. base. Such information helps the pro- Part 1 of th~ essay reviewed some of gram deal with ambiguities such as in- the problems AI researchers confronted complete sentences (ellipses) and pro- while developing NL programs. At that nouns. An example of a conversation in time, I noted that language is a highly which LADDER ,m also developed by complex cognitive function, involving Hendrix and colleagues while at SRI, re- much more than the syntactical manipu- sponds to incomplete sentences is shown lation of words. Creating a program that in Figure 3. Finally, several NL programs can communicate effectively requires will paraphrase the user’s request so it the inclusion of the assumptions, or can be confirmed. Once confiied, the large body of shared knowledge, speak- system translates the request into a for- ers generally bring to a conversation. mal computer language, which is then Colby points out that information about used to retrieve information from the the context in which language is used data base. and the speaker’s intentions are also re- quired.zb Consequently, as with expert F~ure 3: An actual conversation with the LADDER system. Boldface words represent user input. systems, the most successful NL pro- grams have been those equipped with a l—Gfve me the feugth of the Kermdy. great deal of information about a spelling > KENNEDY PARSED! restricted topic. One of the first NL pro- Give the length of the sMp KENNEDY JF grams to operate in such a domain was SHRDLU, developed by Terry Wino- LEN = 1072 2—Wfdtb and &aft grad, Stanford University. It answered PARSED! questions about an imaginary world of Trying Elfipsis: GIVE ME WIDTH AND DRAFT blocks, pyramids, and an arm that OF THE KENEDY Give the beam and draft of the sldp moved them.27 A variety of systems have KENNEDY JF also been built for studying NL and for (BEAM 130 DFT 36) querying expert systems. Commercial 3-Who b her commender? interest in NL, however, is focused on Give the rsnk and name of the commander of the ship KENNEDY JF generic systems which can be fiiled with (RANK CAPT NAME MOFFETT P) knowledge about the data bases to 4-Wlutclout the Fox? which they are linked. PARSED! Trying E@is MI: THE FOX The knowledge in most NL systems WHO IS THE FOX COMMANDER consists of a pragmatic or semantic Give the rank and name of the commander of the grammar.m These are the rules for pars- ship FOX (RANK C,WT NAME EVERE’IT J) ing, or breaking apart, and interpreting sentences Iiiely to be used to request in- Source: He+ G G & ,%cardodE D. Naturd-tanguage pmessin8: the field in perspective. BYTE formation from a particular data base. 6:304-52, 1981.

435 Whereas NL systems match written researchers have also contributed to ro- statements against grammatical rules botics by creating programs that recog- and contextual knowledge, the vision nize objects by their physical touch as programs developed by AI investigators well as programs that understand a match images picked up by a camera, limited number of spoken words .33 Iaser, or light emitting diode (LED) Several of these speech recognition against stored representations of those systems—including HARPY, which was images. These systems examine scenes, designed for document retrieval-were and operate by reducing their colors, discussed in an earlier essay.% A shapes, and textures into “primal number of researchers, including pio- sketches,” which are essentially simple neer John McCarthy, Stanford Universi- line drawings. Then the program “blocks ty, believe that one day AI may even out” a rough approximation of the make possible a “universal manufactur- scene. This consists of simple shapes, ing machine. “35 Essentially, this would such as lines, cylinders, and cones, that be a robot capable of tailoring products the computer has been programmed to to each person’s design. identify. Finally, the computer attempts So many AI applications are now en- to match these “mental images” of ob- tering the marketplace that it’s impossib- jects against three-dimensional sketches le to name alf of them in this brief es- of objects it has been programmed to say. These include educational pro- identify. 31 grams for teaching young children math Although these vision programs are in and more sophisticated programs for use for industrial quality control and in- teaching medical students to reason like spection, most are limited to recognizing specialists. AI is also used in developing objects from only one perspective. Only automatic programming systems. Like a few can identify moving objects, or high-level computer languages, they will dktinguish objects when background relieve programmers from the drudgery lighting changes. However, in situations of specifying in detailed machine lan- where these factors are controlled, com- guage what the computer should do.% puter vision systems can inspect objects And researchers at University of Califor- with remarkable speed and accuracy. nia, Los Angeles, have developed a Machtne Intelligence Corp. offers one speech prosthesis that helps patients system that can inspect stationary parts who have difficulty recalling words. 37 for small dimensional defects at 900 I’ve already discussed how AI has made parts per minute. One of the most ad- it possible to query data bases in NL. But vanced systems, the Optomation 11from AI techniques are also used to make General Electric Co., can inspect ran- data base queries more explicit, and to domly oriented parts at the same rate,zs improve the efficiency with which infor- ACRONYM,32 an experimental pro- mation is retrieved by these systems. gram developed by R.A. Brooks, Stan- ISI@ has been using AI techniques to ford University, can identify objects in improve its data bases for some time. different configurations and from per- Our programs for verifying bibliographic spectives it has never seen before. It can information perform tasks that ordinan- also “guess” the identity of an object Iy require intelligence. These programs, based on a partial view of it. It matches so to speak, “decide” how to edit cita- the visible portion of the object with its tions containing incorrect information, stored model of the corresponding por- such as the volume, year, or the spelling tion of that object. Then it makes as- of the author’s name. When a reader ex- sumptions about what it sees. amines a group of citations to the same Vision programs will initially have paper, he or she can quickly recognize their greatest application in robotics. AI their similarities and identify the errors.

436 Similarly, our “expert” programs can more categories. These programs also recognize which version of a multicited assess the relevance of each new paper. work is accurate and can correct er- A current paper that cites many of the rors.sg core works is given a higher relevance While a great deal of judgment is built weight than a paper that cites only one into our computer programs for verify- or two, ing references in Science Citation In- Researchers in Japan are also working dexm (SCP), they only scratch thesur- on AI-based programs for classifying face of the problem of artificially intelli- documents.’$’t And the work of R.S. gent indexing. At a 1964 symposium of Marcus, MIT, is interesting in connec- the National Bureau of Standards, I pre- tion with expert systems.ds sented a paper entitled, “Can citation in- Part 1 of this essay included a list of dexing be automated?”, that is, “the ca- the AI research fronts identified through pability of the computer automatically our ZSI/CompuMath clustering pro- to simulate human critical processes re- grams. The core documents for 1S1/ flected in the act of citation .“39 Compukfath research front #80-0191, The paper pointed out that “a consid- “Retrieval processes, computational lin- erable standard~tion of document pre- guistics, and language processing,” are sentations tiff be necessary, and prob- listed in Table 1. To show you how the ably not achievable for many years if we core papers in Table 1 are related, we’ve are to achieve automatic referencing . .. . also included a multidimensionally On the other hand, many citations, now scaled map in Figure 4. The paper by fortuitously or otherwise omitted, might A.M. Collins and M. Ross QuiMan, then be supplied by computer analysis of of Bolt Beranek & Newman, Inc., text.”sg Cambridge, Massachusetts, plays a cen- When this paper was reprinted in Cur- tral role in the research front. It dis- rent Contents” in 1970, I explained that cusses the way humans and computers the original title was badly chosen—that store and retrieve information from a more appropriate title was, “Can crit- long-term memory. If you decide to icismand documentation of research pa- search this particular research front pers be automated?”~ Even though the through the 1S1 Search Network, you’ll term “artificial intelligence” was in use retrieve about 80 papers published be- by 1970, it was new enough so that it was tween 1976 and 1983. not obvious to use it in the title of the W. W. Bledsoe, University of Texas, essay, although the original paper does Austin, who wrote the three core papers mention an “artificially intelligent in research front #8@0724 ,~-* has machine.” It is significant that a week played an important role in nonresolu- earlier, in a tribute to Lederberg,dl I tion theorem proving, also called natural referred to hls work, “Applications of ar- deduction. Thk type of theorem-prov- tificial intelligence for chemical infer- ing program uses heuristics to speed up ence. ’942 parts of the proof. ISI’s program for classifying docu- Research front #80-0726, “Computer- ments in ISI/BIOMED&, ISZ/Compu- aided diagnosis and clinical judgment ,“ Mathm, ISI/GeoSciTech ‘“ , and Index to includes the work of Howard L. Bleich, Scientific f?evie ws ‘“ also simulates hu- Beth Israel Hospital, Boston.@ In an man judgment .43 Traditional indexing earlier essay,so I dscussed the Paper- requires the use of human indexers to chase system,sl which he developed classify documents by subject. Our sys- with Gary L. Horowitz. The other core tem establishes the classification system paper is by J.E. Overall, Kansas State algorithmically by co-citation analysis, University, Manhattan, and C.M. Wil- and then assigns each paper to one or liams, University of Florida, Gainesville,

437 FIsum 41 A cluster map of ISI/CompuMath a research front #8QO19 I “Retrieval processes, computational linguistics, and language processing.”

() Tuh@ E. (972 Yale Univ.

\

sCoUins A M. 1969 Bolt Bcranek & Newman, Inc. Smith E E. 1974 Cambridge, MA Kintsch W. 1974 Stanford Univ. Suniv. ColoradO 8

Colfim A M. 1975 ------, Bolt Beranek & Newman, Inc. Norman D A. l!?~5 Univ. California, < San Diego

1 Anderson J R, 1973 Yale LJniv,

T~bfe 1! Core papers of IS1/CompuMath c research front NWOI 91 “Retrieval processes, computational linguistics, and language pmes.sing, ‘“

Amiemom I R & Bower G Il. Human associative memory. New York: Wiley, 1973.524 p. Coffht$ A M & f.dtas E F. A spreading-activation theory of semantic processing. Psycho/. Rev. 82:407-28, 1975. Coflfw A M & QuJffJ%nM R. Retrieval time from semantic memory. J. Verb. Learn Verb. JJeha v 8:240-7, 1%9. CamAegfsma J P & Shepard R N. Monotone mapping of similarities into a general metric space. J. hfd Psychol. 11:335-63, 1974. Kkutach W. The mpresenmtion of meaning in memory. New York: Wiley, 1974.279 p. Norman D A, RummShart D E h LNR ResemcJI Group. Explomtions i. cognitton. San Francisco, CA: W .H, Freeman, 1975.430 p. RIP L 1, Shoben E J & Smith E E. .%mantic distance and the verification of semantic relations. I. Verb. LQ.rn, Verb. Behav. 12: I-20, 1973. Smftk E E, Shoben E J & R@ L 1. Stmcture and process in semantic memory: a fentural model for semantic decisions. P@m(. Rev. 81,214-41, 1974, Tufvtag E & DocMJdwn W, eb. Organization 01 memory. New York: Academic Press, 1972. 423p.

Tnbk? 28 Core papem of ISI/CompuMatha research front #80-0739 “Nomecwsive grammars, natmd Lmg”ages, and induc- tive inference of formal languages. ”

mm L & iffam M. Toward ● mathematical theory of inductive inference. tnform. Contr 28125-55, 1975. Fehh!sm J. time decidnbility resuJts on grammatical inference and complexity. Inform. Cont.. 20:244-62, 1972. Gold E M. Language identification in the limit Inform, Con

438 on a program for diagnosing thyroid cut, on conceptual dependency,% both function.sz About 18 current papers can published in 1972. be retrieved by searching in this research Table 3 presents a selected lit of front. highly cited books and articles on AI. Another research front, #80-1963, This list also shows how often each of “Knowledge-engineering and computer- these publications was cited in SCZ and aided rn?Aical decision-making, ” zeros Social Sciences Citation Index@ from in on the work of G.A. Gerry, MIT, and 1961 to 1983. Three of the publications G.O. Barnett, Massachusetts General on this list, Minsky’s paper in The Hospital, Boston,5g and H .R. Warner Psychology of Computer Vision, Quil& and colleagues, University of Utah, Salt an’s “Semantic memory, ” and Schank Lake City.~ These papers focus on con- and Abelson’s Scripts, Plans, Goals and genital heart disease and other disease Understanding, focus on knowledge diagnoses using the computer. representations discussed in Part 1 of In Table 2, we’ve listed the six core pa- this essay. The paper by Waltz and the pers that define research front #80-0739, book by Winograd are excellent reviews “Nonrecursive grammars, natural lan- of computer vision and NL, respectively, guages, and inductive inference of for- while Simon’s book is a short, nontech- mal languages.” There were about 69 nical discussion of AI and related topics. papers published on these themes. Re- The most-cited publication is the 1%5 search front MM-1155 identifies the field, book by N.J. Nilsson, SRI International, “Cognition, psychological epistemol- which reports early work on techniques ogy, and experiments in artificial intelli- that enable machines to learn by classi- gence.” The two core works here in- fying and evaluating information. His clude Noam Chomsky’s Language and more recent book, Problem -Solving Mind55 and a paper by R.C. Schank, Methods in Artificial Intelligence, is an Yale University, New Haven, Connecti- influential text on theorem proving and

Tabk 3: A selected fist of fdghly cited publications in artificial inteUigence. A =number of citations from SCP, 1961-1983. and SSCP, 1%6-1983. B = bibliographic data.

A B

70 Bobrow D G & Wbmgmd T. An overview of KRL. a knowledge representation language. Cognitive Sci. 1:3-46. 1977. 49 Boden M A. Artificial intelligence 4 natuml man. New York: Basic. 1977.537 p. 44 Brown J S & Burton R R. Diagnostic modefs for prcxedural bugs in basic mathematical skiffs. Cogniiive Sci. 2:155-92, 19’78. 56 Ckwe# M B. On seeing things, Artif. hwel{. 2:79-116, 1971, 40 Kossfyn S M & Sbwmsx S P. A simulation of visual imagery. Cognitive Sci. I :26 S-95. 1977. 45 McCarthy J & Hayes P J. Some phIlosoph&d problems from tbe standpoint of artificial intefhgence. (Meltzsr B & Micbie D, eds. ) Machine intelligence. 4. New York: American Elsevier. 1%9. p. 463-502. 119 Mb.wky hi, ed. Semantic information processing. Cambridge, MA: MIT Press., 1969.440 p. 154 Mkusky M. A framework for representing knowledge. (Winston P H, ed. ) The psychology of computer vision. New York: McGraw-Hill, 1975. p. 21 [-77. 45 NcweSf A, Show I C & Shnon H A. Report on a general problem-solving program. Information processing. Pmceedinga of the Imemarional Conference on information Pmcessi.g. 15-20 June 19S9, Paris, France. Paris: UNESCO. 1960. p. 256-64. 462 NfksoII N J. Learning machines. New York: McGmw-Hilf, 1965. 137 p. 263 Nffsson N J. Problem-solving methods in artificial intelligence. New York: McGmw-Hill, 1971.255 p. 143 Quflfkn M R. Semantic memory. (Minsky M, cd. ) Semamic information processing. Cambridge, MA: MIT Press, 1969. p. 227-70, 3SJ Schamk R C & Abekon R P. Scrip/r, plarw gods and .ndemlmding. Hilkdale. NJ: Lawrence Erlbaum, 1977.248 p. 119 ShortJfffe E H. Compuler.based medical con.ndratiotu, MYCIIV. New York: Elsevier, 1976, 26-4 p. 413 Sfmon H A. The sciences of the artificial. Cambridge, MA: MJT Press. 1969. 123 p. 75 Wsht D. Understanding line drawings of scenes with shadows. (Winston P H, cd. ) 7’he p~ychology of ecwnptder vi$ion. New York: McGraw-Hill. 1975. p. 19-91. 186 Wfmgrad T. Understanding mxum( language. New York: Academic Press. 1976. !95 p. 107 Wkmton P H, cd. The psychology of compuier vision. New York: McGraw-Hill. 1975.282 p. 65 Winston P H, cd. Artificial intelligence. Reading, MA: Addison-Wesley, 1977, 444p.

439 AI search techniques. It includes a de- and increase funding to a level ap- tailed discussion of heuristic search proaching S200 mi~lon during the tenth, methods. and final, year of the program. Although university researchers have This represents a substantial invest- been involved in AI for some 30 years, a ment in AI. But in their recently pub- number of governments are just now be- lished book, The F#th Genemtion, Fei- ginning to recognize the enormous im- genbaum and Pamela McCorduck warn pact AI may have on industrial progress. that the US may soon lose its two- to The UK, for example, recently commit- three-year lead to the Japanese.b In ted itself to a five-year, $300 million in- 1982, that country embarked on an am- vestment in university and industrial re- bitious national ten-year plan to lead the search on advanced information tech- world in computer technology, includ- nology, including AI. S7A multinational ing AI. Japanese government and in- collaboration, cafled the European Stra- dustry officials plan to spend about $800 tegic Program for Research in Informa- million over the next ten years on this tion Technology (ESPRIT), has also joint government-industry effort. Fei- been proposed by the European Com- genbaum and McCorduck believe that mon Market.6 And though little litera- unless the US launches a comparable na- ture on AI has been coming out of the tionwide, collaborative effort, the Soviet Union, where AI used to fall Japanese may ultimately become the under the rubric of cybernetics, com- dominant factor in the computer world. puter scientists there have been pro- The AI applications discussed in this gramming with List Processor (LISP), essay demonstrate how basic research the primary programming language for eventualy has practical results. sgEquip- AI, since the 1960s.Ss ping computers with commonsense rea- At present, the US is the leader in AI soning and general problem-solving research and development.6 This domi- techniques still remains a goal of basic nant position is at least partially due to AI research. Undoubtedly, there will be two decades of support from the US De- new spin-offs as we teach computers to fense Department’s Advanced Research “think.” Possibly of greater importance, Projects Agency (Darpa). The National we may learn more about how humans Institutes of Health Biotechnology Re- use their brains to think and thereby source Program of the Dhision of Re- create what we call knowledge or natu- search Resources also funds AI re- ral intelligence. It remains to be seen search, most notably by supporting the precisely where the boundary between SUMEX-AIM computer system. In 1982 artificial and natural intelligence begins alone, governmental agencies and pr- or ends. ivate firms channeled some $50 million into AI research.h And Darpa has just launched a massive new AI initiative, ● ☛☛☛☛ called “Strategic Computing and Sur- vivability.” According to Feigenbaum,21 My thanks to Joan Lipinsky Cochmn Darpa wilf spend about $50 million on and Amy Stone for their help in the th~ project in the 1983-1984 fiscal year prepamtion of this essay. G19S3 1%1

REFERENCES

1. GarlleldE. Artificialintelligence:usingcompulersto thhkaboutthinking.Part1. Representing knowledge.CwrenfContenf.(49):5-13,5 December1983. 2. DudaR O& SliortlirfeE H. Expertsystemsresearch.Science 220(4594):261-8, 1983. 3. Artificial inteU1gence. The second computer age begins. Bus. Week (2729):66-75, 1982. 4. Roberu S K. ArWlcial intelhscnce. BYTE 6(9): 164-78, 1981.

440 5. Mkbfa D. Computer chess and the humanization of technology. Nature 299:391-4, 1982, 6. FekgenbmrmE A & McCorduck P. The fifth genenrtion. Reading, MA: Addison- Wesley, 1983.275 p. 7. faie~erg J & Fefgenbaum E A. Mechanization of inductive inference in organic chemistry, (Kleinmuntz B, cd. ) Formal representation of human judgment. New York: Wiley, 1968. p. 187-218. 8. Lhrdxay R K, Buchanan B G, Fefgenbaum E A & f.ederberg 1. Applications of artificial intelligence for chemical inference: the DENDRAL project. New York: McGraw-HiU, 1980, 256 p. 9. Laferberg 1. Topological mapping of organic molecules. Proc. Nat. A cad. Sci. US 53:134-9, 1965. 10. Garfield E. A tribute to Carl Djeraaai: reflections on a remarkable scientific entrepreneur. Essays of an information scientist. Philadelphia: 1S1 Press, 1983. Vol. 5. p. 721-30. (Reprinted frum: Current Corrfeno (42):5-14, 18 October 1982. ) 11. AJaxander T. Practical uses for a “useless” science. Fortune 105:138-45, 1982, 12. Fefgenbaum E A. The art of artificial intefhgencc--themes and case studies of knowledge engineering. AFIPS Conf. Proc. 47:227-40, 1978. 13. Gevarter W B. Expert systems: limited but powerful. IEEE Spectrum 20(8):39-45, 1983. 14. van MeUe W J. A domain-independent system that aids in constructing knowledge-bused con$ukation prognrm$. PhD dissertation, Stanford University, 1980. 15. CJancey W J. Tmnsfer of rule-based expertise through a tutonkl dirrlogue. PhD d~scrtation, Stanford University, 1979. 16. Alkfna J S, Kunz 1 C, Sbortffffe E H & FafJat R J. PUFF: an expert system for interpretation of pulmonary function data. Cornput. Biomed. R... 16(3):199-208, 1983. 17. Cofby K M. Mudefing a paranoid mind. Behav. Bruin Sci, 4:51 S-60, 1981. 18. Weka S M, KuUkowafd C A & Safir A. A model-baaed consultation system for the long-term management of gfaucoma. IJCA [- 77. Proceedings of the 5th International Joint Conference on Artificial Intelligence, 22-25 August 1977, Cambridge, MA. Pittsburgh, PA: Carnegie-Mellon University, 1977. Vol. 2. p. 826-32. 19. Pople H. The formation of composite hypotheses in diagnostic prublem solving: an exercise in synthetic reasoning. IJCA I- 77. Proceedings of the 5th Intemationa[ Joint Conference on Artificial Intelligence, 22-25 August 1977, Cambridge, MA. Pittsburgh, PA: Carnegie-Mellon University, 1977. Vol. 2, p. 103@7. 20. MfUer R A, Pople H E & Myers J D. Intemis[-1, an experimental computer-based diagnostic consultant for genersl internal medicine. New Eng/. J. Med. 307:468-76, 1982. 21. Feigenbaum E A. Telephone communication. 28 November 1983, 22. Dad R. Use of meta [evel knowledge in the construction and maintenance of large kno w!edge bases. PbD d=rtation, Stanford University, 1976. (Reprinted in: Davis R & Lenat D, eds. Knowledge-based systems in orf~ici.1 intelligence. New York: McGraw-Hill, 1982, p. 229-490, ) 23. Duds R, Gaacfmlg J & Hart P. Model design in the PROSPECTOR consultant system for mineral exploration, (Michie D, ed. ) Expert systems in the micro-electronic age. Proceedings of the 1979 AISB Summer Schoo/, July 1979, Edinburgh, Scotland. Edinburgh: Edinburgh University Press, 1979. p. 15367. 24. NASA. Proceedings of the 1977 MA CSYMA Users’ Conference, 27-29 July 1977, Berkeley, CA, Washington DC: National Aeronautics and Space Administration, 1977. S01 p, 25. AlJan R. Robotics comes into its own with improved vision, sense perception, teaching languages, Electron. Des. w9):87-lCNl, 1982. 26. Cofhy K M. Personal communication. 10 November 1983. 27. Wfnugrad T. Understanding natuml language. New York: Academic Press, 1972, 195 p. 28. Hendrfx G G & !$acerdotf E D. Natural-language processing: the field in perspective. BYTE 6:304-52, 1981. 29. Hendrfx G G. Human engineering for apptied natursl language processing. IJCA1. 77. Proceedings of the 5th Irrtemational Joint Conference on Artificial Intelligence, 22-25 August 1977, Cambridge, MA. Pittsburgh, PA: Camegie-MeUon University, 1977. Vol. 1. p. 183-91. 30. HendrLx G G, Sacersfod E G, .%galowkz D & SIocum J. Developing a natural language interface to complex data. ACM Trans. Database Syst. 3(2): 105-47, 1978. 31. AJexasrder T. Teaching computers the art of reason, Fortune 105(10):82-92, 1982. 32. Bmnka R A. Mdel-based three dnensional interpretations of two dimensional images. (Drinan A, cd.) IJCAI-81. Proceedings of the Seventh International Joint Conference on Artificial Intelligence, 24-28 August 1981, Vancouver, BC, Canada. Vancouver, BC: Univemity of British Columbia, 1981. Vol. II. p, 619-24. 33. Strobfe WT. The year of the robot. Techno/. Rev, 85(2):87, 1982, 34. GarJJefd E. From Vucuder to Vocaluck-speech recognition machines stilf have a long way to go. Essays of an information scientisr. Phdadelphki: 1S1 Press, 1981. Vol. 4. p. 579-85. (Reprinted frum: Current Contents (34):5-11, 25 August 1980.) 35. H9ta P J. Interview: John McCarthy, OMNI 5(7): 101-6; 122-6, 1983.

441 36. Automatic programming. (Barr A & Feigenbaum E A, eda. ) The handbook of artificial intelligence. Los Altos, CA: William Kaufrrmnn, 1982. Vol. JJ. p. 295-379. 37. Colby K M, Chrlsthraz D, Pukkon R C, Gmbam S & K8rpf C. A word-finding computer program with a dynamic lexical-semantic memory for patients with anemia using an inteffigent speech prosthesis. Brain Lung. 14:272-81, 1981. 38. GarfleM E. Quality control at 1S1: a piece of your mind can help us in our quest for error-free bibliographic information. Current Contents (19):5- 12, 9 May 1983. 39, ------Can citation indexing be automated? (Stevens M E, Giuliano V E & HeiJprin LB, eds.) Statistical association methods for mechanized documentation. Symposium proceedings, Washington, 1964. Washington, DC: NBS, 15 December 1%5. Misc. Publ. No. 269. p. 189-92. 40. ------Can criticism and dcrcumentation of rcsearcb papers be automated? Emays of an inforwralion scienti$t. PhiJadelphk 1S1 Press, 1977. VOI. 1. p. S3. (Reprinted from: Current Contents (9):4, 4 March 1970.) 41. -------m ultidiaciplinarian extmotiinaim. Essays of on information scientiw, Philadelphia: 1S1 press, 1977. Vol. 1. p. 81-2. (Reprinted from: Cumem Contents (8):4-5, 25 February 1970. ) 42. Led.arharg 1, Sutherland G L, BuchmraB B G, Feigenbaum E A, Robertam A V, Duffkld A M & Djemaal C. Applications of artificial intelhgence for chemical inference, L The number of possible organic compounds. Acycfk structures containkig C, H, O, and N. J. Amer. Chem. Sot. 91:2973-6, 1%9. 43, Garffdd E. ABCS of cluster mapping. Parts 1 & 2. Most active fields in the life and physical sciences in 1978. Essays of an information scientist. Philadelphia: 1S1 Press, 1981. Vol. 4. p. 634-49. (Reprinted from: Current Contents (40Y5-12, 6 CJctobcr 1980 and (41):5-12, 13 October 1980.) 44. Mfyariroto S, Mfyake T & Nekaymaa K. Generation of a pseudothessurus for information retrieval baaed on cuwcumences and furzy set operations. IEEE Trans. Syst. Man Cybem. SMC-13(1 ):62-70, 1983. 45. Marcus R S. An experimental comparison of the effectiveness of computers and humans as search intermediaries, J. Amer. Sot. Inform. Sci. 34(6):381-404, 1983. 46. Bladane W W. Splittingandreductionheuristicsinautomatictheoremproving. Artf. Irrtell. 2;55-77, 1971. 47. Bledaea W W, Boyer R S & Henneman W H. Computer pruofs of Jimit theorems. Artrf Intell, 3:27-60, 1972. 48. Bkdaea W W & Brrdl P. A man-machine theorem-proving system. Artrf Irrte[l, 5:51-72, 1974. 49. fflaklr H L. Computer evaluation of acid-base disorders. J. C/in. Invest. 48:1689-%, 1%9. 50. Garfkkl E. The impact of hospital libraries on the quafi~y and coat of health care delive~. Cur-rent Contents (8):5-10, 21 February 1983. 51. Horowftz G L & Blekh H L. Paperchaae: a computer program to search the medical literature. N. Engl. J. Med. 305:924-30, 1981, 52. flvemfl J E & Wffffams C M. Conditional prnbabifity program for dkgnoxis of thyroid function. J. Amer. Med. Ann. 183:X)7-13, 1963, 53. Gerry G A & fkmett G O. Experience with a model of sequential diagnosis. Comput, Biomed. Res. 1:490-507, 1968, 54. Warner H R. Tommto A F & Veasy L G. Experience with Baye’s theorem for computer dmgtmsis of congenital heart dwaae. Ann, NY A cad. Sci, 115:558-67, 1964. 55. ChrMky N. .Lmrguage and mind. New York: Harcourt Brace Jovanovich, 1972.194 p. 56. SCbaak R C. Conceptual dependency: a theory of natural language understanding, Cognitive Psychol. 3:552-631, 1972. 57. Dkbon D. Britain rises to Japan’s computer chaJJenge, Science 220(45$9):799-803, 1983. 58. Glrmhkov V M. USSR, computing in. (Belser J, Holzman A G & Kent A, da.) Encyclopedia of computer science and technology. New York: Dekker, 1979. Vol. 13. p. 498-507. 59. Garffekf E. How can we prove the value of basic research? Essays of an information scientist. Phifadelphla: 1S1 Press, 1981. Vol. 4, p. 285-9. (Reprinted from: Current Corrtuno (40):5-9, 1 October 1979.)

442