Artffkfd Isadlfgesw: Using Computers toTlslnk about Thirskfng. Part 2.

Some Pmctfcsd Applicatfom of AI

Number 52 December 26, 1983

This is the second of a two-part essayon artificial intelligence (AI). Part 1 dis-cussed “knowledge representations”-models of cognition that AI investigatorshave used in their attempts to buildthinking machines. 1 One of the mostambitious goals of AI research is to de-velop programs which enable machinesto perform commonsense reasoning. Itwill & many years, however, before th~

goal is achieved.In the meantime, AI research has

spawned numerous spin-offs that arejust beginning to enter the commercialmarket. These include programs that en-able robots to “see” and “feel” and ma-chines which can follow instructionswritten in natural language (NL). But themost ambitious and successful AI appli-cations thus far are “expert systems. ”These computer programs are designedto duplicate the problem-solving pro-cesses of experts in various fields. Thisessay will cover some of these expert sys-tems and review a liited number ofother applications of AI.

During the early years of AI research,investigators were intent upon dMcover-ing the general principles underlying in-telligence. In developing the knowledgerepresentations described in Part 1, AIresearchers tried to use these principlesto create an “inference engine.” Ideally,such an engine would solve any type ofproblem, from winning at chess to d~ag-nosing disease. But attempts at develop-ing computer-based problem-solvingstrategies for any situation met with

liited success. Techniques developedfor solving problems in one domain wereusually inadequate for other domains.However, programs equipped with agreat deal of information about a singledomain performed as well as, and some-times better than, experts in that field.zThe superior performance of theseknowledge-based, or expert, systemsconvinced many AI researchers thatproblem solving demands huge banks ofknowledge as well as reasoning proce-dures.3

Today, the goal of expert systems re-search is to transfer a specialist’s knowl-edge into a program so the informationcan be efficiently accessed and used bythe computer to solve problems. This in-cludes “textbook leaming’’—the factsobtained from training and reading.4 Italso includes heuristic knowledge, orrules of thumb developed through yearsof experience and judgment. Heuristicsare essentially educated guesses aboutwhich solutions to a problem are mostlikely to be successful. They don’t guar-antee correct answers. But they do savetime by limiting the search for solutionsto those most likely to be correct. Com-puter and other scientists called “knowl-edge engineers” sometimes spend yearspicking experts’ brains for these factsand heuristics, and then structure theminto computer programs.

Expert systems also include “infer-ence procedures” that determine whichheuristics and facts should be brought tobear on a problem. One such inference

430

procedure is backward chaiiing, inwhich you suggest a possible solution toa problem and work backward to see ifit’s correct. MYCIPJ, an expert systemthat assists in medical diagnoses, reasonsin this manner.z It advances a diseasehypothesis based on a few known symp-tom$. Then it looks for other symptomswhich support the hypothesis, request-ing additional tests and information asneeded. In most expert systems, the in-ference procedure for deciding whichfacts and heuristics to use is separatefrom the knowledge base of facts andrules. This makes it easier to add ormodify facts and rules as new informa-tion becomes available.

The fret, and probably best-known,AI programs that focused on a Ihniteddomain were chess programs. DonaldMichie, University of Edinburgh, Scot-land, explains that chess is ideal for mod-eliig specialist knowledge because it is avery well-defined domains A largeamount of formal information is avail-able in the form of instructional worksand commentaries. And numericalscales of performance are avaifable inthe national and international ratingsystems. Equally important, chess is agame that calls on a wide range of cogni-tive functions, from logical calculationto imaginative thinking. The numerouschess-playing programs developed in the1950s and 1960s tested the proficiencywith which various knowledge represen-tations used facts and heuristics to solveproblems.s

By the mid- 1960s, AI researchersbegan to expand beyond chess and puz-zle playing, or what Edward A. Feigen-baum, Stanford University, California,calls “toy problems, “G(p. 62) into practi-cal problems. The fiist such project re-sulted in DENDRAL, an expert systemthat identifies the chemicaf structures ofunknown compounds.7,8 DENDRALwas launched at Stanford University in1966 by Feigenbaum and Joshua Leder-berg, now president of Rockefeller Uni-versity, New York. Both were interested

in modeliig and assisting scientificthinkiig, and Lederberg, a geneti-cist/molecular biologist with expertisein chemistry, had developed a computerlanguage for describing the structure ofcomplex molecules.g As the projectgrew, Carl Djerassi, also at StanfordUnivemity, contributed his expertise. 10

The original DENDRAL program per-formed three basic operations on thechemical formulas and mass spectraldata with which it was provided. In thefmt phase, cafled “plan,” it translatedgeneral and specific prior knowledgeand heuristics into a spec~lc repertoireof constraints. In the next phase, called“generation,” it generated plausiblestructures based on such constraints asthe number of rings, double bonds, andatoms of various types in each molecule.In the final phase, caUed “test,” eachplausible structure was tested. The com-puter fiit generated sets of instrumentdata that would be expected to describeeach structure. Then it compared eachset of data to actual data about the com-pound. The closest fits were then rankedfor the user.

In the past few years, scientists work-ing on the DENDRAL project have fo-cused most of their attention on theplanning and generation portion of theprogram, and on making this portionavailable to users. Called CONGEN, forconstrained structure generation, thisgenerator has been expanded to inferplausible structures using a wide varietyof instrumental data. In 1982. CONGENwas made commercially avaifablethrough the computer network Compu-Serve. I I

Another spin-off of DENDRAL isMets-DENDRAL, a program that gen-erates its own rules from mass spectraldata on chemical compounds. After re-ceiving mass spectral data on a family ofcompounds, Mets-DENDRAL gener-ates planning and test rules that describehow these compounds fragment whenstudied with mass spectrometry. Someof the roles generated by Meta-

431

DENDRAL duplicated those formulatedby expert chemists, while others wereentirely original. 1Z

DENDRAL demonstrated that AItechniques could be used to solve realproblems within a Iiiited area of knowl-edge.b Paradoxically, it also demon-strated that it is easier to model the rea-soning processes of specialists than toprogram the steps a chfld goes throughin understanding language, or makingcommonsense inferences.z This isbecause the facts and judgments an ex-pert uses in making a decision are easierto identify and categorize than are thereasoning processes used for generalproblem solving.

So far, the most successful expert sys-tems have been programs that weigh andbalance evidence about data to deter-mine how they should be categorized.Differential diagnosis, for example, is “aclassical medical example of such aproblem,”z according to Richard O.Duda, Syntelligence, Menlo Park, Cali-fornia, and Edward H. Shortliife, Stan-ford University School of Medicine. Aphysician arrives at a diagnosis byevaluating a variety of symptoms andtest results. Although this is a fairlycomplicated procedure, it is based onidentifiable facts and heuristics and,therefore, lends itself to computermodeling. For this reason, and becausecomputers can consider many diseasesthat physicians might not encounter ineveryday practice, numerous expertsystems have been designed to assistdoctors in diagnosing and treatingdisease.z

The knowledge representation usedmost widely in these expert or “consulta-tion” systems is the “production rule”approach, according to Wi~lam B,Gevarter, National Aeronautics andSpace Administration, Washington,DC. 13 As mentioned earlier, each pro-

duction rule is a heuristic, also called an“if-then” rule or “condition-action” pair.Each rule or set of rules includes facts

about a domain which can be used to

solve problems in that domain. Figure 1shows one of the 500 such rules used bythe infectious diseases expert systemMYCIN. With MYCIN, a physician en-ters information about a patient into thecomputer. The computer then searchesfor the rules that can be applied to thisinformation. If more information isneeded, the computer will ask the physi-cian to supply it. Then the rules for theseadditional data are applied. This processcontinues until a diagnosis, and treat-ment, can be recommended.

Since the acceptability of an expertsystem depends on the confidence withwhich physicians can accept its sugges-tions, MYCIN and several other consul-tation systems can provide explanationsof their reasoning processes. At the phy-sician’s request, MYCIN will list the pro-duction rules it used in its diagnosis ortreatment recommendation, and citereferences to the literature that supportthese rules. MYCIN will also explainwhy it has requested additional tests orother information.z An example of thkexplanation feature is shown in Figure 2,

Ffgure 1: Sample MYCIN production rule.

Mycfri[f: 1) The infection which requires

therapy is meningitis,

And 2) The patient has evidence of aserious skin or soft tissue infection,

And 3) Organisms were not seen on thestain of the culture,

Amf 4) The type of infection is bacterial,

risen:

There is evidence that the organism(other than those seen on cultures orsmears) which might be causing the in-fection is staphyloccwcus-coagpos ( .75)or streptococcus (.5).

Source: Gevmtter W R. Expwt systems: limited but power-ful. IEEE Spemum 20(8):39-45, 1983.

@ IEEE

432

Figura 2: Explanation features of MYCIN, Thii excerpt from the MYCIN run illustrates someof the explanation facilities provided by the program. At entry 38, MYCINaskswhether the patient hasbeen seriously burned. Instead of answering, the user asks why this particular question is being asked, Inrespnn~, 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; inresponse to the HOW command, the program fists the five rules that it appfied, four of which supportedthis conclusion. This abiMy to inspect the program’s reasoning methcds provides the user with a firmerbasis for undemtandmg the final conclusinrrs.

34) Is the meningitis a hospital-acquired infection?““ NO35) Is Pt219’scfinical history consistent with EPIGLOTTITfS?““ NO36) Is Pt219’sclinical history consistent with OTITIS-fvfEDfA?““ NO37) Has the patient ever been to an area that is endemic for cnccidlomycoaes?““ UNKNOWNW 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) whichmight 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

thenthere is suggestive evidence that Paeudomonaa-aeruginoaa is one of the organisms (other than thoseseen on cultures or smears) which might be causing the infectionAuthor: 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-1970s at Stanford University by BruceBuchanan and Shortliife.2 In the late1970s, W .J. van Melle, Stanford Univer-sity, transformed MYCIN’S inferenceengine into an all-purpose program,

called EMYCIN.14 Since then,EMYCIN has been filled with produc-tion rules from a variety of fields, includ-ing structural engineering and computerrepair. Another MYCIN spin-off, calledGUIDON,lS was developed by W .J.

433

Clancey and colleagues, Stanford Uni-versity, to help med~cal students learnthe methods and swategies used by ex-perts in their field.

DENDRAL and MYCIN are two ofthe most influential expert systems. Butthey represent just a fraction of the workcurrently under way in this field. Dozensof systems have been, or are now being,developed at Edinburgh, Purdue, Rut-gers, Yale, and Carnegie-Mellon Uni-versities, Massachusetts Institute ofTechnology (MIT), and the Universityof Pittsburgh, to name only a few of theacademic institutions involved in basicand applied AI research. The leadingcenter for expert systems research in theUS, however, is Stanford University,which houses the Stanford UniversityMedical Experimental Computer forArtificial Intelligence in Medicine(SUMEX-AIM) network, a nationallyshared computer network devoted to AIsystems in biomedicine. Systems thathave come out of this project includePUFF, lb developed by respiratoryspecialists J. Osborn, R.J. Falfat, and B.Votteri, Pacific Medical Center, SanFrancisco, and Stanford Universitycomputer scientists L. Fagan, P. Nii,J.C. Kunz, J.S. Aikins, and D. McClung.PUFF diagnoses and recommends thera-pies for pulmonary dysfunction.PARRY, a program that simulates para-noid thought processes, was developedon SUMEX-AIM by K.M. Colby, Uni-versity of Caliornia, Los Angeles, IT andCASNET/Glaucoma was designed forophthalmology by C.A. Kulikowski andS.M. Weiss, Rutgers University, NewBrunswick, New Jersey, and A. Safir,Mount Sinai School of Medicine, NewYork. IS INTERNIST, a pioneering pro-gram developed by H.E. Pople and J.D.Myers, University of Pittsburgh, Penn-sylvania, diagnoses dkeases in internalmedicine. 19.20 It includes some 4,000symptoms cross-referenced to about Sot)diseases .21

Not afl expert systems are biomedical.

TEIRESL4S,22 developed by R. Da-

vis, now of MIT, while he was at Stan-ford University, is designed to assistknowledge engineers in creating andupdating the knowledge bases usedin expert systems. The success ofTEIRESIAS, and other programs that“build” expert systems, is crucial to thecommercial future of these systems.Knowledge engineers presently spendhundreds of hours collecting and struc-turing specialist knowledge before anexpert system can be built. Other non-biomedical expert systems includePROSPECTOR ,23 a geology consultantdeveloped by Duda and colleagues,while at SRI International, Menlo Park,California, and MACSYMA,ZJ designedby Joel Moses, MIT, for solving algebraand calculus problems.

A number of private firms have alsoentered the AI arena. Digital EquipmentCorp. and International Business Ma-chines Corp. (IBM) are working on sepa-rate expert systems that diagnose “sick”computers.j Xerox Corp. and Texas In-struments are developing systems toassist in the design of computer chips.And Schlumberger Ltd., which employsmany Al researchers in its three AIlaboratories, is developing a system thatanalyzes data on geologic formations.3

Several companies have also beenlaunched by established AI scientists.Feigenbaum and several of his col-leagues at Stanford University startedTeknowledge Inc. and IntelligeneticsInc., both in Palo Alto, California. Tek-nowledge markets expert systems, andoffers consulting and training services tocompanies that want to build their own.Intelligenetics markets expert systemsthat assist in gene splicing experiments.Computers designed for AI research aresold by two companies, Lisp Machkes,Inc., and Symbolics Inc., which werefounded by MIT researchers.

Expert systems aren’t the only AI ap-plications entering the marketplace.Machine Intelligence Corp., OpticalRecognition Systems Inc., and GeneralElectric Co. are among a growing num-

434

ber of companies marketing software for

recognizing visual images.25 ArWlcial

Intelligence Corp. and Cognitive Sys-

tems Inc. are marketing NL softwarethat enables users to communicate withcomputers in natural languages such aswritten English.3 The state of the art forthese applications is not as advanced indealing with technical domains as are ex-pert systems. 13 But these programs canfunction impressively within the limiteddomains for which they’ve been de-signed.

Part 1 of th~ essay reviewed some ofthe problems AI researchers confrontedwhile developing NL programs. At thattime, I noted that language is a highlycomplex cognitive function, involvingmuch more than the syntactical manipu-lation of words. Creating a program thatcan communicate effectively requiresthe inclusion of the assumptions, orlarge body of shared knowledge, speak-ers generally bring to a conversation.Colby points out that information aboutthe context in which language is usedand the speaker’s intentions are also re-quired.zb Consequently, as with expertsystems, the most successful NL pro-grams have been those equipped with agreat deal of information about arestricted topic. One of the first NL pro-grams to operate in such a domain wasSHRDLU, developed by Terry Wino-grad, Stanford University. It answeredquestions about an imaginary world ofblocks, pyramids, and an arm thatmoved them.27 A variety of systems havealso been built for studying NL and forquerying expert systems. Commercialinterest in NL, however, is focused ongeneric systems which can be fiiled withknowledge about the data bases towhich they are linked.

The knowledge in most NL systemsconsists of a pragmatic or semanticgrammar.m These are the rules for pars-ing, or breaking apart, and interpretingsentences Iiiely to be used to request in-formation from a particular data base.

LIFER,29 an NL program developed byG. G. Hendrix while at SRI Internation-al, for example, interprets the sentence,“What is the age of Mary’s sister, ” bymatching it against a series of grammati-cal questions, such as, “What is the attri-bute of the object?” Most NL programsare also equipped with contextual infor-mation about the data base. This in-cludes definitions of words liiely to beused in queries, and information on rela-tionships between objects in the database. Such information helps the pro-gram deal with ambiguities such as in-complete sentences (ellipses) and pro-nouns. An example of a conversation inwhich LADDER ,m also developed byHendrix and colleagues while at SRI, re-sponds to incomplete sentences is shownin Figure 3. Finally, several NL programswill paraphrase the user’s request so itcan be confirmed. Once confiied, thesystem translates the request into a for-mal computer language, which is thenused to retrieve information from thedata base.

F~ure 3: An actual conversation with the LADDERsystem. Boldface words represent user input.

l—Gfve me the feugth of the Kermdy.spelling > KENNEDY

PARSED!Give the length of the sMp KENNEDY JF

LEN = 10722—Wfdtb and &aftPARSED!Trying Elfipsis: GIVE ME WIDTH AND DRAFTOF THE KENEDYGive the beam and draft of the sldpKENNEDY JF(BEAM 130 DFT 36)3-Who b her commender?Give the rsnk and name of the commander of theship KENNEDY JF(RANK CAPT NAME MOFFETT P)4-Wlutclout the Fox?PARSED!Trying E@is MI: THE FOX

WHO IS THE FOX COMMANDERGive the rank and name of the commander of theship FOX(RANK C,WT NAME EVERE’IT J)

Source: He+ G G & ,%cardodE D. Naturd-tanguagepmessin8: the field in perspective. BYTE6:304-52, 1981.

435

Whereas NL systems match writtenstatements against grammatical rulesand contextual knowledge, the visionprograms developed by AI investigatorsmatch images picked up by a camera,Iaser, or light emitting diode (LED)against stored representations of thoseimages. These systems examine scenes,and operate by reducing their colors,shapes, and textures into “primalsketches,” which are essentially simpleline drawings. Then the program “blocksout” a rough approximation of thescene. This consists of simple shapes,such as lines, cylinders, and cones, thatthe computer has been programmed toidentify. Finally, the computer attemptsto match these “mental images” of ob-jects against three-dimensional sketchesof objects it has been programmed toidentify. 31

Although these vision programs are inuse for industrial quality control and in-spection, most are limited to recognizingobjects from only one perspective. Onlya few can identify moving objects, ordktinguish objects when backgroundlighting changes. However, in situationswhere these factors are controlled, com-puter vision systems can inspect objectswith remarkable speed and accuracy.Machtne Intelligence Corp. offers onesystem that can inspect stationary partsfor small dimensional defects at 900parts per minute. One of the most ad-vanced systems, the Optomation 11fromGeneral Electric Co., can inspect ran-domly oriented parts at the same rate,zs

ACRONYM,32 an experimental pro-gram developed by R.A. Brooks, Stan-ford University, can identify objects indifferent configurations and from per-spectives it has never seen before. It canalso “guess” the identity of an objectbased on a partial view of it. It matchesthe visible portion of the object with itsstored model of the corresponding por-tion of that object. Then it makes as-sumptions about what it sees.

Vision programs will initially havetheir greatest application in robotics. AI

researchers have also contributed to ro-botics by creating programs that recog-nize objects by their physical touch aswell as programs that understand alimited number of spoken words .33Several of these speech recognitionsystems—including HARPY, which wasdesigned for document retrieval-werediscussed in an earlier essay.% Anumber of researchers, including pio-neer John McCarthy, Stanford Universi-ty, believe that one day AI may evenmake possible a “universal manufactur-ing machine. “35 Essentially, this wouldbe a robot capable of tailoring productsto each person’s design.

So many AI applications are now en-tering the marketplace that it’s impossib-le to name alf of them in this brief es-say. These include educational pro-grams for teaching young children mathand more sophisticated programs forteaching medical students to reason likespecialists. AI is also used in developingautomatic programming systems. Likehigh-level computer languages, they willrelieve programmers from the drudgeryof specifying in detailed machine lan-guage what the computer should do.%And researchers at University of Califor-nia, Los Angeles, have developed aspeech prosthesis that helps patientswho have difficulty recalling words. 37I’ve already discussed how AI has madeit possible to query data bases in NL. ButAI techniques are also used to makedata base queries more explicit, and toimprove the efficiency with which infor-mation is retrieved by these systems.

ISI@ has been using AI techniques toimprove its data bases for some time.Our programs for verifying bibliographicinformation perform tasks that ordinan-Iy require intelligence. These programs,so to speak, “decide” how to edit cita-tions containing incorrect information,such as the volume, year, or the spellingof the author’s name. When a reader ex-amines a group of citations to the samepaper, he or she can quickly recognizetheir similarities and identify the errors.

436

Similarly, our “expert” programs canrecognize which version of a multicitedwork is accurate and can correct er-rors.sg

While a great deal of judgment is builtinto our computer programs for verify-ing references in Science Citation In-dexm (SCP), they only scratch thesur-face of the problem of artificially intelli-gent indexing. At a 1964 symposium ofthe National Bureau of Standards, I pre-sented a paper entitled, “Can citation in-dexing be automated?”, that is, “the ca-pability of the computer automaticallyto simulate human critical processes re-flected in the act of citation .“39

The paper pointed out that “a consid-erable standard~tion of document pre-sentations tiff be necessary, and prob-ably not achievable for many years if weare to achieve automatic referencing . . . .On the other hand, many citations, nowfortuitously or otherwise omitted, mightbe supplied by computer analysis oftext.”sg

When this paper was reprinted in Cur-rent Contents” in 1970, I explained thatthe original title was badly chosen—thata more appropriate title was, “Can crit-icismand documentation of research pa-pers be automated?”~ Even though theterm “artificial intelligence” was in useby 1970, it was new enough so that it wasnot obvious to use it in the title of theessay, although the original paper doesmention an “artificially intelligentmachine.” It is significant that a weekearlier, in a tribute to Lederberg,dl Ireferred to hls work, “Applications of ar-tificial intelligence for chemical infer-ence. ’942

ISI’s program for classifying docu-ments in ISI/BIOMED&, ISZ/Compu-Mathm, ISI/GeoSciTech ‘“ , and Index toScientific f?evie ws ‘“ also simulates hu-man judgment .43 Traditional indexingrequires the use of human indexers toclassify documents by subject. Our sys-tem establishes the classification systemalgorithmically by co-citation analysis,and then assigns each paper to one or

more categories. These programs alsoassess the relevance of each new paper.A current paper that cites many of thecore works is given a higher relevanceweight than a paper that cites only oneor two,

Researchers in Japan are also workingon AI-based programs for classifyingdocuments.’$’t And the work of R.S.Marcus, MIT, is interesting in connec-tion with expert systems.ds

Part 1 of this essay included a list ofthe AI research fronts identified throughour ZSI/CompuMath clustering pro-grams. The core documents for 1S1/Compukfath research front #80-0191,“Retrieval processes, computational lin-guistics, and language processing,” arelisted in Table 1. To show you how thecore papers in Table 1 are related, we’vealso included a multidimensionallyscaled map in Figure 4. The paper byA.M. Collins and M. Ross QuiMan, thenof Bolt Beranek & Newman, Inc.,Cambridge, Massachusetts, plays a cen-tral role in the research front. It dis-cusses the way humans and computersstore and retrieve information fromlong-term memory. If you decide tosearch this particular research frontthrough the 1S1 Search Network, you’llretrieve about 80 papers published be-tween 1976 and 1983.

W. W. Bledsoe, University of Texas,Austin, who wrote the three core papersin research front #8@0724 ,~-* hasplayed an important role in nonresolu-tion theorem proving, also called naturaldeduction. Thk type of theorem-prov-ing program uses heuristics to speed upparts of the proof.

Research front #80-0726, “Computer-aided diagnosis and clinical judgment ,“includes the work of Howard L. Bleich,Beth Israel Hospital, Boston.@ In anearlier essay,so I dscussed the Paper-chase system,sl which he developedwith Gary L. Horowitz. The other corepaper is by J.E. Overall, Kansas StateUniversity, Manhattan, and C.M. Wil-liams, University of Florida, Gainesville,

437

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

() Tuh@ E. (972

\

Yale Univ.

-------,

Smith E E. 1974

<

Stanford Univ.

8

sCoUins A M. 1969Bolt Bcranek & Newman, Inc.Cambridge, MA Kintsch W. 1974

Suniv. ColoradO

Colfim A M. 1975Bolt Beranek & Newman, Inc. Norman D A. l!?~5

Univ. California,San Diego

1 Anderson J R, 1973Yale LJniv,

T~bfe 1! Core papers of IS1/CompuMath c research front NWOI 91 “Retrieval processes, computational linguistics, andlanguage 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<r. 10:447-74, I%7.

JJomlmg J J. A study of gmmmaricd inference. PbDdws-ertation. Stanford, CA: Stanford University, Department ofComp.{er Science, August 1%9. No. CS- 139; AI memc-98. NTIWPC AOS MF AOI.

SoJasmxtofl SSJ. A formal theory of inductive inference. Part 1. Inform. Contr 7: I-22, 1964.SOJttzwmll R J. A formal theory of inductive inference. Part 11. Jnform. Conw 7224-54, 1964.

438

on a program for diagnosing thyroidfunction.sz About 18 current papers canbe retrieved by searching in this researchfront.

Another research front, #80-1963,“Knowledge-engineering and computer-aided rn?Aical decision-making, ” zerosin on the work of G.A. Gerry, MIT, andG.O. Barnett, Massachusetts GeneralHospital, Boston,5g and H .R. Warnerand colleagues, University of Utah, SaltLake City.~ These papers focus on con-genital heart disease and other diseasediagnoses using the computer.

In Table 2, we’ve listed the six core pa-pers that define research front #80-0739,“Nonrecursive grammars, natural lan-guages, and inductive inference of for-mal languages.” There were about 69papers published on these themes. Re-search front MM-1155 identifies the field,“Cognition, psychological epistemol-ogy, and experiments in artificial intelli-gence.” The two core works here in-clude Noam Chomsky’s Language andMind55 and a paper by R.C. Schank,Yale University, New Haven, Connecti-

cut, on conceptual dependency,% bothpublished in 1972.

Table 3 presents a selected lit ofhighly cited books and articles on AI.This list also shows how often each ofthese publications was cited in SCZ andSocial Sciences Citation Index@ from1961 to 1983. Three of the publicationson this list, Minsky’s paper in ThePsychology of Computer Vision, Quil&an’s “Semantic memory, ” and Schankand Abelson’s Scripts, Plans, Goals andUnderstanding, focus on knowledgerepresentations discussed in Part 1 ofthis essay. The paper by Waltz and thebook by Winograd are excellent reviewsof computer vision and NL, respectively,while Simon’s book is a short, nontech-nical discussion of AI and related topics.The most-cited publication is the 1%5book by N.J. Nilsson, SRI International,which reports early work on techniquesthat enable machines to learn by classi-fying and evaluating information. Hismore recent book, Problem -SolvingMethods in Artificial Intelligence, is aninfluential 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-tailed discussion of heuristic searchmethods.

Although university researchers havebeen involved in AI for some 30 years, anumber of governments are just now be-ginning to recognize the enormous im-pact AI may have on industrial progress.The UK, for example, recently commit-ted itself to a five-year, $300 million in-vestment in university and industrial re-search on advanced information tech-nology, including AI. S7 A multinationalcollaboration, cafled the European Stra-tegic Program for Research in Informa-tion Technology (ESPRIT), has alsobeen proposed by the European Com-mon Market.6 And though little litera-ture on AI has been coming out of theSoviet Union, where AI used to fallunder the rubric of cybernetics, com-puter scientists there have been pro-gramming with List Processor (LISP),the primary programming language forAI, since the 1960s.Ss

At present, the US is the leader in AIresearch and development.6 This domi-nant position is at least partially due totwo decades of support from the US De-fense Department’s Advanced ResearchProjects Agency (Darpa). The NationalInstitutes of Health Biotechnology Re-source Program of the Dhision of Re-search Resources also funds AI re-search, most notably by supporting theSUMEX-AIM computer system. In 1982alone, governmental agencies and pr-ivate firms channeled some $50 millioninto AI research.h And Darpa has justlaunched a massive new AI initiative,called “Strategic Computing and Sur-vivability.” According to Feigenbaum,21Darpa wilf spend about $50 million onth~ project in the 1983-1984 fiscal year

and increase funding to a level ap-proaching S200 mi~lon during the tenth,and final, year of the program.

This represents a substantial invest-ment in AI. But in their recently pub-lished book, The F#th Genemtion, Fei-genbaum and Pamela McCorduck warnthat the US may soon lose its two- tothree-year lead to the Japanese.b In1982, that country embarked on an am-bitious national ten-year plan to lead theworld in computer technology, includ-ing AI. Japanese government and in-dustry officials plan to spend about $800million over the next ten years on thisjoint government-industry effort. Fei-genbaum and McCorduck believe thatunless the US launches a comparable na-tionwide, collaborative effort, theJapanese may ultimately become thedominant factor in the computer world.

The AI applications discussed in thisessay demonstrate how basic researcheventualy has practical results. sgEquip-ping computers with commonsense rea-soning and general problem-solvingtechniques still remains a goal of basicAI research. Undoubtedly, there will benew spin-offs as we teach computers to“think.” Possibly of greater importance,we may learn more about how humansuse their brains to think and therebycreate what we call knowledge or natu-ral intelligence. It remains to be seenprecisely where the boundary betweenartificial and natural intelligence beginsor ends.

● ☛☛☛☛

My thanks to Joan Lipinsky Cochmnand Amy Stone for their help in theprepamtion of this essay.

G19S3 1%1

REFERENCES

1. GarlleldE. Artificialintelligence:usingcompulersto thhkaboutthinking.Part1. Representingknowledge.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-busedcon$ukation prognrm$. 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 formineral 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 tocomplex 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 computerprogram with a dynamic lexical-semantic memory for patients with anemia using an inteffigentspeech 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-freebibliographic 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 aninforwralion scienti$t. PhiJadelphk 1S1 Press, 1977. VOI. 1. p. S3.(Reprinted from: Current Contents (9):4, 4 March 1970.)

41. --------------- Joshua Lederberg-m ultidiaciplinarian extmotiinaim. Essays of on informationscientiw, 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 Thenumber 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 physicalsciences 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 informationretrieval 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 searchintermediaries, 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