-
FD 190 109
TITLEINSTITUTION
SPONS AGENCY
REPORT N6. PUB DATE
CONTRACTNOTE
ED15._.ERIC1DESCRIPTORS
DOCUMENT RESUME
IR 009 594
The Seed of Artificial Intelligence. SUMEX-AIM.Research
ResOurces Information Center, Rockville.,MdNational Institutes of
Flealth (DHEW) , Betht?sda, Md-Div. of Research
Resources..NIH-00-2071Mar 00N01-RR-9-2114.74p-; Photographs and
some other illpstrations maynot reproduce. . 5
MF01/PC03 Plus Postage.*Artificial'Intelligence;
*Biomedicine;-ComputerPrograms; Computers; *Computer Science;
InforMationNetworks; *Information Systems; Medical
Research;*Online'Systems; State of the Art Reviews
It'
4
ABSTRACTWritten tu provide an understanding of the ,broad
base
of information on,which the artificial intelligence
(AI)branch.ofcomputer science rests, this publication presents a
general view ofAI, the concepts frdm w.hich it, evolved, its
current abilities,' andits promise for research. ThefocuS is on a
community of prolectsthat kise the SUMEX-AIM (Staivford
Univers,3rty 'Medical Experimerital.
:Computer for Artificial Tntelligence'in Medicine) network, a
.nationally-shared computing resource4,eoted entirely to designing
Aapplications for the_biomedical scien6es:-Chapters explore
the:weaning of AI,.the histpry of computer sFience, the ptocesses
ofcomputit,g, the usesof UMEX 1n'biochemistry-0, clinical
medicine',psychology, and AI-tool 'building, and the futute
applicationsof AI.Appendices provide rinformation,on
the.orqabization of SUMEX, 71t8available.facilitifs, and its
management, as well as a AireCtory ofSUMEY-AIM project
investigatorS and-project-funding soUrceS.. .(FM)
V
A
C.
4
ReproductionSsupplied by EDRS,are th!Ce best thatcan be made *-,
:%,,from *he'original dOCument.' 1 , *
*SO'
e
teee;+'-
tW,
:s 4
t
-
TheSeeds ofArtificialIntelligence
e
A publication ofDivision of Research Resourcr)Nabonal Institutes
of I foalthBethesda, Marylahd 20205
1)1 PAR TM% NT OF HEAL IliF OUt AT ION IL WEL F JOSENA 1 IONAL
INSTITUTE OF
011CA 1 ION
l .... 1,0( t WI NI HAS III I N III 114(1tfot t If I lt Al I I N
AS III I I IV{ 6 I NOM1111 PI iv,on, WI C>PC,IINITAT ION
CINIfilN-A t IN, I 1 I,(ii tv 1 N OF VII W (lIt OPINION%% I A II il
DO Hill NI ( I SSA.I.tIl S 1.1t- PHIS I NJ I l l I I ( I Al N A T
IONAT U4'1111(171 ()I DM A 1 NON 1'0.'01)0N (Hi 1,00 I(
Prepared byResearch Resources Information
Center1776 East Jefferson StreetRockville, Maryland 20852under
contract N01-RR-9-2114
Mardi 1980
46,
1./.$. DEPARTMENT OF HEALTH-, ,EDUCATION, and OELFARE
'Public Health Service' .National _Institutes of Health
A
-
Acknowledgements
t
Tho Nrts of many peoplemado this publication posstble.Special
thanks go to Mr. ThomasRindfleisch; Drs. Joshua Leder-berg, Herbert
Simon, EdwardFeigenbaum, Bruce Buchanan,
_.113.?1! ........... William Baker, JackMyers, and ifarry
Pople: dIKTMr. Edward Post.
^41.
-
Foreword
4
4
1
In the past centui y, SCiOnce hilSnot only changed ,QUI
conceptitms
j,abolit io world, It has changed itself .1 'von by an explosion
of infor motion, specialties in-sciencehave sprung up, inevitably
givingrise to subsppcialties: But stayingabreast of new knowledge,
even innarrowly speCialized areas, is be-coming increasingly
difficult. Oneway to manage the continuingflood of now infoftation
may be to 'create entities of intelligence. .
The proposed tool is the intelli:gent machine, a device that
mimicsthe expert's reasoning power andcan retain in retrievable
form muchof the knowledge currently avail-able to experts in a
given specialty.-Most systems of this type are )stiimmature. But
some are akee'dymoving into the real world andothers will makt the
transitionwithin the next few years. As these
4 activities become apre formalized.,a new branch of apPied
sciencewill arise. Most likely it will becalled knowledge
engineering.
What systems will be available?Who will they help? How will
theywork?
Many answers are contained inexisting books and arlicles.
Buttechnical publications'Riffer from
4
the defect Qf their vir toes I hey aretoo detailed, too
exhaustive arid,inost minor tont, too focused onsin* areas of
rapidly expandingdisciplines To understand this newbranch of
computer science, calledartificial intelligence (Al), it is
nec-essary to understand tho founda-tion, the broader base, on
which itrests.
This publication will plesent ageneral view of Al, the
conceptsfrom which it evolved, its currentabilities, and its
promise for re-search. The focus is on a commu-nity of projects
that use theSUMEX-AIM (Stanford UniversityMedical Experimental
Computerfor Artificial Intelligence in Medi-cine) network. .
SUMEX-AIM is a nationallyshared computing resource de-voted
entirely to designing AIapplications for the,biomedical sci-ences.
It is funded by the NIHsion of Rqsparch Resources,BiotechnolOgy
Resources Pro-gram: Although SUMEX-AIM doesnot include all Al
projects directedtoward medicie and related re-search in
this/country, many of theprograms now using Al techniquesfor
medical decision-making weredeveloped using this facility.
-16
0
-
TableofContents
!.
foreword 4
IntroductionAt tit icial intelligenr,eWhat's in aName? -6
History of ComputingAbacus to ENIAC----and (3eyond 9
Processes of ,ComputingThe Heuristic Mind 20
SUMEX arid the ScienceCommun)tyThe Seeds of Artificial lntellig
Ice 24
Biochemistry' 25Clinical Medicine 33Psychology 54Al Tool
Building 62
Future of AlPrOspectus 64
Appendix AOrganization and FacilitiesAvailable 6-9
Appendix BManagement 71
Appendix CSUMEX-AIM Pirectory and ProjectFunding 73
%
-
Introduction
ArtificialIntelligenceWhat'sin aName?
\.(
For contiMos, phaupflots andlinguists have grappled with
the'question of defining intelligence.
, Most have approached the issue, by cribing the function of
lntelli-vn , or the way it appears in be-
yi An exact defi Rion for thistefort elusive.
its might be expected, machineintelligbnce islrqually, if not
more,difficult to define. According to Dr.Margaret Boden in her
book Mill-clal Intelligence and Natural Man,
oonipoldr-S aro only I ozioar 01 took,machines programmed le
dothings that would require intelli-gence if done by people. Dr.
Mar-vin L. Minsky, artificial intelligence(Al) research r at
the.Massachu-setts Institute f Technology andadvisor for S EX-AIM,
agrees.He says artificial intelligence is thescience of making
machines doWings that people need intellizgence to do.
Others take a somewhat differ-
4%.
' ..1.0ilt4VAVR,.` 5 ..t:*, :
": 1,
'4'; IL. 41,V.40 I"A'
ra ill
-
ent view. Dr. Edward.Feigenbaum,principal investigator of
SUMEX-AIM, says the field iS not primarilyoriented toward
technology, buttoward investigatingthe nature Ofintelligence as
information process-ing, whether the intelligence is ex-pressed by
man or machine.
One point of emphasis in current \Al research is to design
computerprograms that capture the knowl-edge and reasoning
processes ofhighly intelligent specialists. Thepracpcal goal of
such work is tomake specialized expertise moregenerally accessible.
To do so, .reSearchers are attempting tounderstand how experts go
aboutacquiring and using knowledgd.Principles of how knowledge
ac-crues and how it is retrieved in log-ical sequence are
extracted. Theyare then programmed into thecomputer.
Within the SUMEX-AIM system,the reasoning processes of
physi-cians, chemists, and other biomed-ical scientists are being
analyzed.At present, the ability of most pro-grams is limited and
much lessflexible than the correspondinghuman intellect. In
specialized
Dr. Herbert A. Simon; SUMEX-A1Madvisor: sorting out the recipe
ofintelligence,
4
areas of medical diagnosis andcheMical str mane analysis,
someprograms can already rival humancaPabilities. Still, maily
people areskeptical of the computer's potenT..tial.
Nobel Prize winner Dr. Herber tA. Simon, psychologist
computetscientist at Carnegie-Mellon Uni-versity and SUMEX-AIM
advisor, isconvinced that this potential isgenerally underrated. He
sayshuman behavior is based on acomplex but definite Set of laws.
Ifthese laws are discovered and re-duced to computer software,
Dr.SiMon believes machine intelli-gence comparable to man's
will_become a certainty in specificareas of expertise.
To capture these higher level Vfunctions, Al researchers are
de-veloping a new approach. It iscalled symbolic computation, a
setof methodeby which abstractionscan be expressed and managed
inthe computer to solve non-mathematical problems. They em-phasize
manipulations of symbolicrather than numeric-information,and they
yse largely informal or,heuristic decision-making rtSles
gained from real-wetH experienceWhen used in Al,
heoll'itic:,
.tocus the program's attention qnthese parts of the problem that
aremost critical and those parts of theknowledge base that are most
ml-event. The result is that these pro-gramspul S 10 a lirm of
reasoning,rather than a sequence of arith-metie steps.
Use of complex symbolic struc-tures is necessary when
construct-ing computer applications fordomains that cannot be
well-formulated in mathematical terms--either because they are not
fullyunderstood, as in medical diag-nosis, or because the
underlyingconcepts are intrinsically non-numerical. "Seldom are
thereequations, in the mathematicalsense, that felate
megisuremnntsof body parameterS to the diag-nosis of disease,"
sayis Mr.Thomas Rindfleisch, director of theSUMEX computing
facility. "Rather,the precsiss of diagnosis it charac-,terized by a
set of strategies hav-ing to do with rules of experienceand
judgmental knowledge..Theserules govern the interpretation
ofobservations and guide decisions1:
2t4,
rt '
IF--v,r1
. :' ',2i. '1 '. J, 4'',' " , 11111 ' ,.., 1.,,,i, q ,
..,,,,t,
i ",f ,. sti "2 V.1.11 ' ' % ''eq , l'ii Lqi.;; % . .) 1,1_:.
$
0' 1
%,
1.1i
VPk,,.)-r
T-.
-T1111111PFIL v141,...
1402-rav I
\ 3,
-`
-
Itiout what ()Mei intorrnItion isneeded tc) deter mine the
diseaseprOCIItis involved
or example, IN I ERNIS I. adiagnostic computer program inthe
SUMOK-AIM network, is fødised on the broadest of
medicalkpecialtiesinternal medicine Itanalyzes patient cases by
mimick-ing the expel Fs reasoning process"The method used by
physicians toarrive at diagnoses requires cornplex information
processing whichbears little resemblanceto thestatistical
manipulations of mostcomputer-based systems," saysDr. Jack D.
Myers, coprincipal in-v'estigator of the project at Me Uni-versity
of Pittsburgh. "As a reSult,the focus of reS4arch in this field
ofmedical applications has shiftedduring the past few years
frommodels of statistical inference tothose using the heuristics of
artkfi-cial intelligence."
"In final tom INTERNIST willamplify intelligence," Dr.
Feigen-baum says. It will supply ewertadvice to the general
Practitionekand physician's assistant, ac-celerating and improving
theirwork. "NI equally important out-
( owe ()I 1w5wow!) such rs this it!.;111v11,\X 1\llV1 is
elmcitnq. organ',mg, and polishing a body of knowleskje that rarely
sees the light otdray." he says "It is the knowledgethat underlies
the exper tise ofpiactice, the knowledge that isnor malty
transmitted by a kind ofOSMOSIS process from master toapprentice.
'that knowledge willnokbe codified, taught, used, andcritiqued." In
essence then, a keygoal of artificial intelligence research in the
SUMEX-AIM com-munity is to capture in computerprograrrts the
knowledge andproblem-solving abililies of ex-perts. After studying
this process inmany specialized areps of exper-tise, Mr.
Rindfleisch says, it may ul-timately be possible to capture in
\ computer programs something ofprocess of creativity and
dis-
covery itself. Programs.then wouldpossess the ability to detect
pat-terns that establish order fromchaos, to draw connections
be-tween seemingly unrelated ideas,and to establish the principles
forsolutions tO new classes ofproblems.
r
e
11 I
4.1010.00000.-
Dr. Edward FoigenbaUin, principalinvestigator of SUMEX-A1M:
"The,laws 15f.expertise will t taught,used, and critiqued."
4
-
HistoryofComputing
AbacustoENIACand Beyond
Boethius (left) and Pythagoras: afanciful battle between
arithmeticcalculation and the abacus.
9
A1144§04.1tUilli;:e. :4;.;. ".t.i4/ '(
In the millennium before Christ,'amid the great cities and
con-quests of Greece and Home,dreamers and theorists were layingthe
groundwork"for today's thinkingmachines. Like seed crystals in
asupersaturated solution, thesevi,ionaries drew from nature,
as-sembling conclusions from obser-vations about the
universe..Theirefforts brought important advancesin mathematics,
astronomy, andmedicine.
f)tuch of the early work in for-mulating the laws of
mathematicsmay appear to have little connec-tion with the computer
of today. Mteach step forward in this elaboratesciesce was
indispensable to theultimate arrival of the computer.
'Pythagoras, a 5th century 'B.0philosopher known as the
founderof Greek mathematics, was theharbinger. He, first described
the"mystical significance of numbers"and established the
relationshipbetween musical harmony andmathematics.
Perceiving in the skies a regu-larity similar to khat of
music,Pythagoras studied movements ofthe heavenly bodieS, or as
he
7 kepi.._1 '
1-- 11:2: 6 if
callmI it, "the music of thespheres I in became? the first
torealize the impoonlice of yoometric _shape, which governs all
naturefrom crystalline rock to the humanbody. In so doing,
Pythagoras setthe direction of mathematicalthought for centt ii es
to come.
Mathematics was soon iegardedas exact. It became the
corner-stone of all science. For centuriesscholars,believed that
its logic wasinfallible. BUt in the 19th centurythe first inklings
of doubt surfaced.Two mathematicians, one in Hun-gary and the other
in Russia, es-tablished irrefutably that it was im-possible to
prove Euclid's postulateof parallels, which states that nomote than
one line parallel to agiven straight line can passthrough a given
point. Alternatetheories sprang up, threatening toscatter the focus
of science. Math-ematics, the mainstay of scientificcertainty, was
suddenly unceitain.If there are two or more geomet-ries, which is
right?
After much thought and delibera-tion, Jules Henri Poincaré, a
19thceptury philosopher and mathe-matician, found the solutign.
He
%Mk
)4...i......,.....,,..
... al410,,.. .0. , ...J 1 ' 1 % .1,t .,`
, ` .^:".. ; ,t, j', cr. 107
-
10
answeied simply that the questionmeaningless Poiricare, de
scribed as one of the most eminentscientific thinkers of his
generation,asked, "Is the meter more truethan the foot? Are
Cartesian coordilates false arid polar coordinatescorrect?" One
geometry cannot be
'more true than another, just moreconvenient, he concluded.
Through his philosophy, Poln-care provided the flexibility
neces-sary for science to advance froman age of scientific
complacency.Few realized the significance ofPoincare's study of
mathematicaltruth. Even fewer guessed that, in2 decades, absolutes
of classicalscience such as space, time, andsubstance would.becokie
approx-imates, and the most respected as-tronorner would agree
that, if mancould look deep enough intospace, he would see the back
Ofhis head.
But the human mind is capableof much more than just
abstraction.Driven by the social pressures ofwar, business
competition, and
o , labor-saving machines weredeveloped. The first mechanicalaid
to calculation was the abacus.
I he Phoenician wont AllAR, thename of a flat slab covered in
sandIII which figures could he drawn,provide(I the root tor the I
nglr.diword ()tiring Greek and I tornantimes, the piunitiabocus was
aflat wooden boaid with counter s ItdevOoped into the now fainiliai
uirangement of beads threade(I Onwires or laid in grooves
With teithadvent of arithmeticsigns in the 15th century, the
popu-larity of the abacus began todecline in Europe. John
N9pierfurther reduced the labor of longmultiplication and division
with theinvention of logarithms Multiplica-tion and division.were
then facili-tated by adding or subtracting the"logs" of
numbers.
Before this technique could bewidely used, accurate tables
oflogs and antilogS had to be corn- 4.piled and.printed. Despite
valiantefforts by mathematicians to Makethese tables accurate,
thedrudgery of figuring, printing, andcopying the numbers led to
errors.Often, mistakes were handeddown from generation to
genera-tion as mathematicians built, all toofaithfully, on the
wobbly shoulders
...4.
-I. .11r -j1,..il .
,11/11 ;i'lipli I; 0:0 0
Ili.,. .4 ( 11';,,- i )0-,1 -1, -,-, .-.. i 1P -_, ,); v.1 i.
.'-) ''o`i lt .--; \'--I ) ti 0 ?,%+
.1 ' "c .,,' -- ..: Ar)) l' ( l ' ( ) ,i' ' C,O.
N1.- ,....1 (,.. ,'A,/ ) ...7-.. ir A.,.. ,.., (...,7...
..\;,-.c.' .1 yl"As i'' )'-) ',."' ! )r 5.,--:r--r.(:).
1 I. -c) ..'' --r "..- t 1'
'.
of those MR) had clomp beforeAn alternative to the use ot
mathematical tables was !soondeveloped MI ';kiin log deviceknown
as the IIrI i tile, whichconsists of two 0(1 scaiosmetliuted
side-by side in a mannerto permit slutilig them easily backail('
for th Whereas the modernoliqital computer counts, tbe analogdevice
measures quontities. 114nslide rule scale is arranged so
thatnumbers fall at distances corre-sponding to their logarithms.
Ls-sentially, multiplication is accom-plished by adding iwo
lengthstogether. Division is done by sub-tracting two lengths.
As \ni,ith all analog devices, theaccuracy of slide rules is
limited bythe accuracy of measurement.Their use did not solve the
proplemcaused by incorrect tables, butrather introduced a lack of
preci-sion. The solution, of course, wasto produce reliable
tables.
1[11812 this thought occurredto Charles Babbage and
JohnHerschel, tVvo young mathemati-cians, while they wew
checkinglogarithm tables for órrorl..As re-counted by
,. x
I 'rot:.
111111111111041R,I.
rr -
A I - Or,
1 )),
' r - .'(-}.
-
\
. ). /1.- 6.,\r. r -1
' .(14.; 7 (P ) , - . ,
S. '
r
'2 "5-cy)
A NI 1
t.
-
Schwatic of the Analytical Engine4esigned by Charles.Bpbbage In
the19th century,: a grand exercise infutility
,
.!
. 447,
(, V.) (' )
1111';111,111111
-
ings, ho exclaimed- "I wish to (lodthat those calculations had
b( enexecuted by steam." I lerschel re-plied: it is quite
possible." And sooccurred the idea that was to dom-inate Babbage's
life---elimination oferr or tIll ough mechanized calcula-tion.
Because his ideas were so ad-vanced and his standards so
high,Babbage experienced one disap-pointment after Mother. In
manyways the 19th century inventor'swork belongs more to oOr
timethan to his own.
Babbage's first project, the Dif-ference Engine, was to be a
largo,complex adding machine designedfor compding mathematical
tables.Unfortunately, the machine wasdoomed to fail. The mechanical
,tolerances required for the)na-chine to work exceeded
capabili-ties of the time. The accuracywithwhich gears could be out
was in-adequate. Clocks, the nearest me-chanical cousins to the
DifferenceEngine, were still laboriously fittedtogether by
hand.
Undaunted by this challenge,Babbage designed new machinetools.
He hired and trained a tech-
Scientific American illustrates use ofthe Hollerith Tabulator in
the 1890census: the' era of data-processingbegins.
assistant nut these piepara.tains est money arid the initialsum
pfuvided by the British .freaslily soon dwindled away Fiveyeais
after beginning the project,Ilabbago was asking the gover n-mont
for morn money Ills reque!Owas graitiod. Rut again, ihoamotint was
not --and could nothave beenenough.
After almbst a decade of workand some £35,000 of governmentarid
personal monies, the project
was rOlandoned. If completed, theengine would have !wen
uIllaikable piece of work 2 tons ofbrass, stool, and pewter, cut
totolerances never before attempted0.'mbittered by failure,
Babbage,
a man of considerable wealth, hay-ing In/willed 1,100,000 fun»
hisfather, devoted Much time andmoney to insulting and
slanderingfigureheads of the scientific andpolitical establishment
whom, heblamed for the engine's failure. But
. -r-:zrl-rx-rT-'rtr-r-rr-.--r-rv-I -1-r1 rr . .
-7,-, s_
A w y mull\ It or NI 0, \ roRNIATION, ART, SCIENCE, MECHANICS.
ilIEMISTIii% -PO) M 1IT10EHIM`i-. 111 .4 sJW 0111.,
I
Itry s .
I
-.11 r. 4 IP 44
,
.
/ .
7`:;:,,,c,, 4
LI" k) ':-1:. It t
d4;I
iLl'--ge'lig."
\`-`-,..;
I
\VINNI, ar_11111.._
t
.$
f
k,
j.t'h' ,."
41."
ri'5
,
AtrAFT *soft'
-
k.
ho did not Fibandon his gOal After111:33 Ilabbage elaborated on
a"gigantic-idea" which he had firstconceived while Working On the
Difforence Engine. If built, this mas-sive device, dubbed the
ArialyticalEngine, worild have been the firstgeneral-purpose
computing machino.
Babbage's scheme contained,for the first time. most of the
mis(r)tial features of tho modern com-puter An arithmetic unit
called "themi)l" was designed to carry out ad-dieron, Subtraction,
multiplication,and divisiorvA memory unit was tohave room tor 1,000
numbers,each 50,digits longa capacitybeyond technology until the
firstelectronic computer appeareda hundred years later.
lnstructipns and data were to befed into the machine on
punch,cards, which had been invented in1800 by Jos6ph Jacquard for
usewith his automatic loom. After cal-culations were completed,
resultingnumbers would be printed up to 29
The Analytical Engine was asfarsighted and intricate in designas
it wasimpossible to build. Once
The Groat Brass Brain: prodictingtidos accoratoly and
officiontly in1914.
Ilabbage's ambition hadtranscondod his time I von withtoday's
tochnotogy the 011(11111,would be difficult to constr (RA hoclitiso
of the mochanical tole!ances required ;ilill Ilabbage's offorts
were not altogother.in vain
.
tits enthusiasm spread to otheis,notably Her man I tollerith,
who dosigned tho first machine devoted todata piocessing
Hollerith's machine, which usedpunch cards, was the easy
winnerin a contest staged by the U.S.ConSus Office to pick an
efficientsystem for tabulating the 1 90census. His device completed
thetest in half the time needed by htscompetitors, whose entries
usedmanual methods.
Data in the form of "yes" or "no"answers were translated
ontopunch cards, which were compiledin a machine that
electromechani-cally sensed positions of holes.Cards passed under a
set -ofbrushes that transferred a pulse ofelectricity through each
hole to ametal cylinder.
After forming the Tabulating Ma-chine Company in 1896, which
was'one of several businesses that
41...
'sk
V,
1
later for builtMachine!; for sof Ong such cards,comparing 0110
to the other, andprinting &ILI I includn mere In-ter [nation
for business use, Hol-lerith increased his punch cards tothe siie
of the dollar bill of his time,which later became an
industrystandard
lis inventions opened the doorto an era of computing
machines,ushered in by the first efficientkey-driven calculating
machine.Called a oomptometer, it was, built .by Dorr E. Felt from A
macaronibox.
The riing popularity of calculat-ing aids and machines in
businesscharacterized Ft shift in attitude to7ward the kind of work
people coujdor should do. Calculating machinessoon entered into
head-on compe-tition with people hired as "rapidcalculators" by
businesses trying tokeep pace with expanding mar-kets. BeSides the
tedium associ-ated with mental calculation, healthwas also a
coriideration. Mentalcalculators, as experts in the tradewerR
called, often complained that,their -6Veninto wore haunted
byunending processions of figures .
,,,tte."'(* 4.N
50,e
`13
1,V:3.,:d Vt.:4Yr r4x:
, a. ,
\
"*1
,v,,,;\k 4-,
-
tof
shaped like numbers.William,S Out roughs, e hank
,clerk, was forced to changecareers because the "monotonousgrind
of clerical work" had de-stroyed his health. At the turn 'ofthe
century, Burroughs entered thecoMpteMeter field and from hisearly
venture grew one of today's.Major manufacturers of
digitalcomputers, Burroughs Corporation.
As the calculating marhinesgained acceptance, more andmore
applications were found. Onecalled Milli9naire, developed in1893
and widely useclaibe business,found immediate and key uses
inscience. Percival Lowell beganusing Millionaire in 1905 to
searchfor a "Planet X" located some-where beyond Neptune.
Calcula-tions were completed in 1914, butthe planet, later named
Pluto, wasnot sighted until 1930, 14 yearsafter Lowell's death.
At ttle same time that manufac-turers were converting to
massproduction techniques, Spanish in-ventor Leonardo Torres y
Quevedowas demonstrating a theory thatheralded the oncoming age of
theprogrammed machine in industry.
Vannevar Bush and the 1930's differ-ential analyzer: "I was
trying,to solvesuch problems of electric circuitry asthe one
connected with failures andblackouts in power networks. I hadbeen
thoroughly stuck because Icould not solve thetough equationsthe
investigation led to."
ir-ct"s
4,7
"roues combined electromechani-cal calculating techniques
withprinciples of automata, dernonstr at-ing that such machines can
per-form any desired sequence ofarithmetic operations.
Torres' electromechanicalArithoMeter, exhibited in 1920,realized
theories of automata thathe had pioneered 7 years
earlier.Arithmetic problems were typed inby theboperatct, and the
Arithome-ter printed the answers on a type-writer. Torres became
the first per7son to use a system of time-sharing when he linked
severaltypewriterb to one Arithometer.
One of his other inventions wasa remote-controlled guidance
sys-tem that successfully steered aboat through Spain's Bilblo
harbor,dramatizing the fact that machinescould perform tasks
formerly re-served for human intelligence. Tor-res later built the
first decision-making device-z--a chess-playingmachine that matched
a rook and'king against a human opponent'sking.
In 1914 Scientific American an-nounced the arrival of ."a
greatbrass brain" which computed
ocean tides On the basis of 37 fac-tors, displaying the i
et;ults on Owl':During the first World War, shipsused information
from the machineto maneuver into shallow water-and elude German
U-boats.
After World War I, VannevarBush Z7if the Massachu,setts
Insti-tute of Technology:invented the dif-ferential analyzer, an
analog deviceassembled from gears, cams, anddifferentials that
methanicallycompleted the various functionsnecessary to solve a
differentialeqUetion. Busfts machine wasapplied to many different
:tasks, re-pladng devices such as ':netweykanalyzers" built by
utility com-panies in the 1920s to analyzeload reOrements. These
ma-chines produced scale models ofpower networks, but they could
notpredict large power surges thatmight cause blackouts. The
differ-ential analyzemas the first ma-chine with such a capability.
Itssuccess seemed to indicate thatbig, general-purpose analog
com-puters would dominate scjentificcalctflatIon in the future.
In the 1930's servomecha-nismsautomatic devices that
-
controlled other machines bynlonitoring their output -came
intouse. Oil refineries and syr up-production plants were among
thofirst to use these "machines thatboss other machines." As
controlproblems were rethiced, more'andmore applications were found
forthe servomechanisms. Steam tur-bines, airplanes, and
chemicalprocesses were soon included inthe domain of the new
device.
As machines surprised societywith newfound abilities,
theircreators took to flights of fancy,building'robots in
exaggeratedhuman forms. Inventors built tin-can contraptions that
walked,talked, and responded to me-chanical commands. The
robots'lifelike actions were an elaborateillusion, as they were
controlled bysimple automatic devices or, re-motely, by human
operators. Assuch, they were no More thannovelties, commonly used
in prod-uct and company promotions .orfairs.
Willie Voce lite, built by Westing-t house in 1931, was one of
these.Willie had a stovepiPe head, ex-pressionless face, and
cauliflower
A.tlektro and Sparko en route to the1939 New York-World's Fair:
tin-cancontrapOons that walked, talked,and responded to
mechanicalcommands.
B.EN1AC, the world's first electroniccomputer, begins operation
in 1946:an unwieldy collection of vacuumtubes ,and relays that
could only beprogrammed by manually changingplug-and-socket
connections and bysetting switches.
oars At the inauguration of passenger air ser Vico between
NewYolk and San Francisco, Williemade i speech, wished Over yenebon
voyage, helped star t the MI-9nes, and after his official
dutieswere completed, relaxeciwith acigarette in the company of
alovely model hired for the occarsion.
Eight years later, Willie's metalcousin, Eloktr 0, a stocky,
tough-looking robot, appear0 at the NewYork-World's Fair with fUs
faithfulcompanion Sparko, the first robotdog. Elektro walked,
talked,counted on his fingers, puffed ciga-
rettes, arld collki distinguish hetwoen red and green with the
aid ofa photoelectricx:ell. Sparko barked,wagged his tail, sat up,
andbegged.
In the late nth tift, engineersturned thr*ir cellectiN're genius
toproblems raised by the secondcoming of world war. The its.Army
sot out to improve differential
Nralyzers'used at Maryland'sAberdeen Proving Grounds to
cal-,Wale firing tables for artillery bat-
- teries. Modifications increased\speed and accuracy by a factor
of80, allowing the machinelo pro-cluce one trajectory every 15
alio-
-
utos as compared to the 20 hoursneeded by a skilled
mathematician.ft it the machine was limited by itsdesign to
processing dif fer entialequations: it could only
calculatethe,functions of vectors.
"ThorQ exist problems beyondour abil4 to solve, not because
oftheoreti al difficulties, but becauseb,,of instil ticient means
of mechanicalcomputation," Howard H. Aikenisaid of the analyzer in
1937. He,then proposed a new kind of cal-culating machine.
In 1938 IBM began building aforerunner of the device for
Har-vard University. It was called theAutomatic Sequence
ControlledCalculator (ASCC). After its com-pletion in 1944, the
ASCC,nicknamed Mark 1, became the.first automatic,
general-purposedigital calculator.
Mechanical switches called re-lays routed electrical signals in
theASCC. During its 15 years of use,ASCC pfoved to be a reliable
and
, effective machine, but its morethan
three-quarters-of-a-millionparts and 500 miles of wiring
mademaintenance expensive and diffi-cult.
flw calculate); was mainly usedby the lf S Navy.for ballistics
andship design .;cienu4, and industrylater used the machine to
gondatoastionomical tables and specifi-cations for lens design It
was alsoused in military studies at WrightPatterson Air Force Base
and inr esearch for the Atomic I nergyCommission.
A year before ASCC das fin-ish9d, John Mauchly and J. Pies-per
Eckert, Jr., of the University ofPennsylvania, proposed the
nextlogical step in mechanized calcula-tion. First described as an
elec- .tronic difference analyzer, thescientists predicted their
new cal-culator Would execute all functionsin computing firing
tables, produc-ing each complete table in only 2days, The device
promised to getaround a major failing of the differ-ential analyzer
by allowing input or.such data as atmospheric resist-ance defined
by numbers ratherthan by mathematical formulae.
Built in secrecy at the Universityof Pennsylvtinia, the new
device,which ultimately became known asENIAC (Electronic
Numerical-lnte-Orator and Calculator), was moved
r14 4116
r.
to the Ilatlistit!; I towarch I obortories
People were necw;sary to genmato) firing tables on differ
arlalyzer, and the tiuriii r oleslowed production Cpmpletion
ofone table, on the average, toyk 2or 3 months.
.VVith the now machine, lengthyohd repetitive calculations for
oach60-second trajectory could becompleted in just 30 seconds.
ButENIAC was not completed until1946, and the huge device,
com-posed of some 18,000 vacuumtubes and 1,500 relays, was
neverused for ballistic computations. Itdid find wide-ranging
applicationsin scientific calculation, however.Until the early
1950s ENIACTIab-bled in weather predictionatomicenergy research,
cosmic ray,)studies, and thermal ignition!
Germany may have entered thefield of electronic computers
aheadof 'America, although little is knownabout the true dimensions
or oper-ation of these machines.. The most
fccessful version, Z4, was de-stroyed in an Allied bombifig
raid.Designed by Konrad Zuse and builtat the Gergran Aircraft
Research
pf,i,r11=-4---2------ t -1- ....ea
, ,'. ';;Jii ,! *c fik, iiiit-IP
.:tilicZt,, \,
k tiZ Ott,
..,.._. li.'t 1i ts N'iv-},,,s,'T'V''',.t's?
!4'ili:''tZQ 4 ,1'.
' ,
,, Zrtrtf"'":-'1,1'31"I'Al,<t
e
1 5,ki1.4.4.
,
,t
-
vt
Nnstitute, Z4 was used in develop-ing the tiS 293, a flying
bomblaunched from Nazi aircraft.
At the war's end, Zuse could riotcoavince Allied intorrogator9
thathe had any scientific expertise tootter, arid his research came
to asudden stop. Not until the midfittiesdid he resume his work,
this timeas owner oT a computer man-ufacturing'\compamk which
waslater absorbedby a large Germanelectronics firm.
As technology flourished duringthe 1940s, a major breakthrough
inthe burgeoning field of computerscience occurred. Althou9h
the/exact source of the concept isuncertainJohn von Neumann,Mauch
ly, Eckert, or British-mathe-matician Alan M. Turingit wassuggested
that instructions couldbe stored as numbers in the ma-chine itself.
The idea raised themechanized kingdom severalrungs on its
evoliilionary ladder. ,For the first time, logical choices
ofprogram sequences would bemade inside a machine.
Earlier, programming ENIAC andZ4 had been extremely tedious:
inENIAC. by changing plug-and-
socket connections arid by settingswitches; in Z4 by instr
fictionspunched into discarded 3!)rilliimovie film. The concept
ofsoftware programming provide(tthe basis for the next generation
ofcomputers.
The first machine with a com-pletely "logical design," which
vonNeumann described at the tir9e ofENIAC's construction, was to
becalled EDVAC (Electronic DiscreteVariable Automatic
Computer).While EDVAC Was still under con-struction 011948, ENIAC,
afterspecial wiring modifications, be-came the first computer to
embodythe stored-program concept. UsingENIAC's new
capabilities,-vonNeumann and severalmeteorologists completed the
firstcomputer-based weather forecast.Computations for the
hydrogenbomb were begun on ENIAC andcomplerd on its successor
MANI-AC (Mathematical Analyzer,Numerical Integrator and Com-puter).
MANIAC was one of manystored-program computers that fol-lowed in
the wake of the new pro-gramming concept, although eachdiffered
considerably in design.
I.DVAC, I mAC, .10IINNIAC(which was named for v.onNeumann),
;I-AC, SWAC, andNOliC were the first few to appiiar. .
As computers became increas-ingly power foil, these
machinesmoved into nokv areas. In his bookCybernetics, Norbert
Wiener ex-plored the potential uses of au-tomata. In 1948 W.
GreyWalter en-tered the field.of cybernetics (theComparative study
of automaticcontrol systems) with an elec-tromechanical "tortoise"
built tostudy simple reflex motion:"Thesemachines are perhaps the
simplestthat can be said to resemble ani-mals," Walter wrote,.
"Crude thoughthey are, they give an eerie im-pression of
purposefulness,Inde-pendence, and spontanerty."
Von Neumann, decidedly a ,"software scientist,"'hoped to
useautomatic machines such as themodern computer to draw
conclu-sions,about complex natural or-gantsms. He built on the idea
ofthe Universal Turing Machine, ad-vanced by Turing in 1936.
Turingdescribed, in theory, a machine -that could don y calculation
withinthe\icealm oMman intellect. The
A mode4integrated circuit: putting40-times the memory of ENIAC
on a \chip the size of an aspirin.
-
Universal luring Machine, whichcontains ideas later built into
al1general cowling) machines, pro-vides a standaid for measuring
thecomplexity of a computer.
Around 1950 the computeremerged as general tool. It h4dbecome
applicable not only to mill-/lary use, but also to functions
ingovernment, industiy,science, education, and social sci-ence. The
coThputor's spectaculargrowth in capability, applications,arid
numbers surprised mostpeople. In 1954, using cathode ray
se°
4a kr P!
*Pi
el.
tubes and magnetic di ums for information storage imd
vacuumtubes for I )gic and arithmetic filmic
ntiot;, expe s in the field had estimated that only some )()
COMpanies would eventually find usefor computers.
Eluis with the developrpent in thelate fifities, of computer
kinguages,which simplified pi ogi anmung, andwith the introduction
of integratedcircuits in the midsixties, compres-sing the
equivalent of 1,400 tran-sistors, resistors; and diodes ontosilicon
chips an eighttvot an inch
_
1
4
1, ,7,4'
11 Ftri
OS,
e
4x 4.. ,da'-driik"
Iv Ili. .'.'4.,..
-..
; t;"). , "''' i 4,.. 'c 44-4,,
k, qt. .gt,
', )0,1
rItN l
p %, ''tVN; -Pglit
N .1
*top
\S",,k
L
r
' s'4' , t. jki--
-..,,,,),1- 2',
, 11' -.eI'll. 'P,, 1" ,W.! v ,.. ..,A4i11 It/.
;
Sql1,110, !RI iez,tiaining widespread application of
cominitersHew off Private tuisinoss begantizong computers to
process Orders,inventories, and payrolls. Com-puters set copy for
newspapersand processed checks tpr banks.Airlines used computers to
makeand record seat reer vations.
Even in medicine, an alea bestcharagerized as inexact and
highlysubjeCtive, computers-fitted intovery specific and important
niches.Applications have included suchareas as electrocardiogram
analy-
,A, (.
171
,fait
t 17
-
515 systems, aids for managin,clinical routines, and
instrumentdata collection. Statistical cluster-ing techniques were
applied todiagnostic programs and were oneof the first diagnostic
approachesto prove useful in medicine.
Just as those new methods ofnumerical calculation'were
beingtested, the first inklings of a newapproach to computing were
intro-duced.,The first application of thisapproach----symbolic
comPutingwas the establishment of databanks stocked with patient
informa-tion. It was an application of sym-.bols to algorithnis,
matchingnames to numbers, and it carriedalong the guarantee that
ifpattern-matching was properlyapplied, the answers would
befound.
But medical diagnosis, on thesame level as practiced by
thephysician, required much more inthe way of programming
tech-niques. To grapple successfullywith the problem, prototypes
ofhigh level analysis and symbolicrepresentation were
developed.Many of these resulted from earlywork in applying
artificial intelli-
, ...... . .
genCe (Al) to biomedwal problemsAl Stanford liruversity. under
the
direction of Nobel PIII0 winner Dr.,fostrua I gletfierg.ilfid Dr
I dw.udiolyenbaum, a team of sc,ientistsbegan the development in
1966 ofDENDRAL., a chemistry programwhose offspring now rivals
exper tsin figuring out the structures of cer-tain organic
molecules. DENDHALwas the unlikely outgrowth of asystem c;illed the
Advanced Com-puter for Medical Research(ACME), which had been
sup-ported by the Biotechnology Re-sources Program(BRP) of the
MNDivision of Research Resources(DRR) since 165. OriginallyACME was
dedicated to real-timeanalysis of data gained during --biological
and clinical research.The computer was oriented entirelytoward
numerical calculation, orbatch processing of biomedical re-
,search data.
By the midsixties, principal'in-vestigator Dr. Lederberg
believedthat ACME had proved its worth asa data analyzer. At the
urging of Dr.Feigenbaum, he proposed t9at thesystem host a new type
of com-puter science application, artificial.
intelligom1 he first attempts to use ACMI
for Al were under taken al Stanford,and DLNDItAl was the first
majoreflor I In the begirming tfie machine's size and speed wet e
suffi-dent. Rut as DE-NDRAL grew, andother research was added,
moreand more computing time andmemory space were required. Bythp
early 1970s, Stanford's Alneeds completely outstripped themachine's
capacity and DENDRALappeared to be doomed.
During this period Dr. WilliamBaker of the BRP arrived for a
sitevisit. In light of ACtIviE's inadequatecapacity and the
interest in Al atStanford and elsewhere around thenation, Dr. Baker
suggested thatACME be dropped in favor of asystem devoted to the
develop-ment of this new type of comput-ing. Stanford applied for a
grant,another site visit was made; andthe NatiOnal Advisory
ResearchResources Council of the N11:I rec-ommended funding. A
machinedesigned to handle symbolic com-putation was installed and,
throughthe use of a nationwide communi-cations network called
ARPANET,
"All these deraye7a thousandth ofa second here, a millionth of
a
. second therewe'll have to get thedarn thing fixed."
-
a natiorlal computer communitywas established. An advisor
ygroup, Drs. Ledorberg, Feigen-baum, and Baker, recommendedthat 40
percent of the computer'scapacity be made available toStanford
researchers, another 40percent spread among the nationalcommunity,
and the remainder as-signed to development of the newsystem.
In 1973 SUMEX-AIM was formedas a community resource for
thedevelopment of AI techniques.From .ils inception, the
resourcehas Veen supported by the NIHDivi4ion of Research
Resource&Bi technology esources Pro--
am. Drs. Le rberg and Feigen-aum directed t e network that
/was to become a major medium forthe development of projects
likeDENDRAL among a national groupof biomedical researchers. In
mid-1978 Dr. Lederberg left to becomepresident of The Rockefeller
Uni-versity. He remains an advisor ofSUMEX-AIM. Dr. Feigenbaum
isnow principal investigator of theresource, which currently
includessome 20 autonomous projects,each targeted for
application
1 ,
:Ve't;vi.4.:::+44;;4414-.6..!;:.
in medicine, biochemistry, orpl>ychology
loday's Al expel ts believe thatcomputers will find many
non-professional and small businessapplications as the potentials
of thefield become better under stood.
computeis may somedayregulate heating and cooling sys-tems,
notify fire or police depart-ments in emergencids, and cto
taxstatements, among other chores.At work, letters will be typed,
or-ders placed, bills paid, and filessearcired by computer. At
neigh-borhood shopping centers, payrollrecords will be maintained
andstock automatically reordered ac-cording to need and
profitability.
Some scientists characterize this, revolution as a race to put
more
and moro function, processingppwer, and storage capacity
ontoeach semiconductor chip. Sincethe midsixties, the amount of
func-tion on a chip.has risen by a factorof 10,000. The cost of a
chip,meanwhile, has remained approx.-imately constant at $5 to $50.
,
BV the end of the century, com-puter buffs predict, this trend
ofbroadening applications Will be in
full stride, as w I tt -; trOnd towardMil riaturizafion, ich
made cornouting available to the general ptib-lic. A single silicon
chip, measuringonly a few millimeters sguare,,willbe able to follow
20 million instruc-tions per second, using 10 millioncells of
internal mummy storage.
And just as imagination andhardware have gone hand in handsince
the early 1900's, scientistspredict that programming tech-
cniques and the science softwarewill keep pace with deve
t(oping
technology.
-
AIn the course of 30 years theProcesses computer has graduated
(win
vacuum tubes and rnecharucallays to silicon chips, each one
nolarger than a pencil maser But the
mputing . transition from [NIAC, one of tilefirst electronic
number-crunchersof the late for ties, to the "thinking"machinevf
today required morethan advances in hardware. It re-quired advances
irlovzgramrning
/concepts. -Over the last few decades, there
.z/Ifas been an increasing emphasison the design of
knowledge-basedsystems. At the lowest level, thoseprograms differ
from traditionalprograms in two key ways."Theyemphasize
manipulations of sym-bolic rather than numeric informa-tion, and
they use largely informalor heuristic decision-Inaking rulesgained
from real-world experiencerather than mathematically
provedalgorithms. At a higher level, thesetools of symbolic
processing areused to construct understandablelines of reasoning in
solving prob-lems and to interact wp.humanusers.
Symbolic computation is neces-sary in certain domains, such
as
TheHeuristicMind
medical diagnosis, because cornprelmnsive mathematical
formulations do not exist. 1 or example, therelatioriship of a
symptom suchas "burning pain in the upperztbdor»en" =to disease
diagnosisrequires the manipulation of sym-bolic information.
Projects currently in SUME X-AIM include areas of
medicine,biochemistry, and psychology. Thekey goal of an Al program
is to ex-plain conclusions and allOw theprofessional to interadt in
the deci-sion process.
As a result, Al programs dependlargely on decision-making
strate-gies composed of heuristics, orrules based on Judgmerit and
ex-perience, which are expressedsymbolically. These
strategiesstarkly contrast with numeric corn--putation, which is
largelygorithmic, following a m e-metically fixed s rocedureswhen
evaluating functions ortabulating results, However, thetwo classes
of computation are nottotally dissimilar. .
Clinical flowcharts are al-gorithms used by diagnosticianswhen
deciding how best to man-
-
age a patient. Offen these fixedprocedures are designed by
expertptlysicions for use by pat aniedicscharged witp per for mins
mrtainroutine tasks. As suet!, data, arerepresented symbolically.
Becauseclinical algorithms are relativelysimple, computers are
seldomnecessary.
But automated record-keepingand data banks, more
intricateexamples of the clinical algorithm,require the computer.
In these sys-tems, patient !lames and historiesand other relevant
information aremanipulated as symbols, and areconnecIto numeric
data thatgive spoc1fivalues to theinformationfor example,
patientage: 21. Pattern-matching al-gorithms can be used to
locaterecords of similar individuals or
groups of patients to prochicestatistical summaries.
Although the oaf bust systemsseldom did mow than
maintainrecords, there have been recent at-tempts to create
programs that cancomplement this function by ana-lyzing the stored
infer illation.ARAMIS (American Rheumato4yAssociation Medical
InformationSystem) is one of the most suc-cessful projects in this
category. Inaddition to search and statisticalfunctions, the data
bank .offersanalysis of prognosis as it relatestd,a, specific type
of patient. Pro-gram6 systematically search thedata base to locate
case reportsand summarize the outcomes ofvarious alternative
treatments,matching recorded case historieswith descriptionsof
current pa-.
THE NATIONAL ARTHWTIS DATA RESOURCI
SC.I1 NUENHANCI Nil NTS
c'eA:
ST ANI 01111
DAR IMOtry H . i AICHI I A
I EU:NO
CINCINNAT I ARAMISIADoi I I
(WOUP?CI IWO
St.AM Nivol.;11H
t.81 I I IDA
tionts. In systews such as this, theanalysis of alternatives and
the de-cisitv about the best theiapy aresolely up to the
physician..
More complex decision-makingprograms attempt to assist
thephysician in evaluating the bestti eatment strategy. The
dedsioncriteria used in such firograms takevarious forms. Some
decision'rulesmay have rigorous statistical jus-tification, while
others may be onlyapproximate {Liles based onhuman experience and
judgment.These latter strategiesiare calledheuristics. Each type
can be effec-tive in providing solutions to prob-lemg.
In statistical approaches to diag-noses;the decision criteria
have
The ARAMIS data bank" meetingneeds in thetstudy and practice
ofrheumatology. (Abbreviations: Al -ar-tifiCial intelligence,
ARA-AmericanRheumatism Association,CCC=cooperating clinical
trialscommittee of the ARA, SLE-sys-temic lupus
erythematosus,SCCS=scleroderma cooperativecriteria study, JRA
=Juvenilerheumatoid arthritis, Canadian RA -Canadian Rheumatism
Association,UDB-uniform data base Mr rheu-matic disease, FDA=Food
and DrugAdministration, VA-Veterans Admin-istration)
( AEACOMPD 11
COMM1
1k:.1.CM11:,111.1A1
(*.HOW
SOODOCUMI N I ' CANADIAN
I OD
(LI INIU.AI;;i:DiD Al6PiDIGRAP"/
Appt
21L!
-a
4
21
-
been codified to a certain degree.Baye's theory of probability
is oneexample. Flsentially, Bayesiananalysis relates specific
patientdata to different disease signs ex-hibited by selected
groups of pa-tients In establishing these rela-tionships, it is
sometimes possibleto compute the most likely cadsefor symptoms
obserwki in a pa-tient.
One of the earliest such pro-grams,-developed in the 1960's,was
used to diagn9se congenitalheart disease. In some casestudies, the
program reacheddiagnoses with accuracy compara-ble to those
rendered by two expe-rienced physicians. As researchershoned and
polished the program,applications for other diseaseareas were
discovered. Todaymany types of diagnostickrogramsusing Bayesian
analysis are in op-eration. But Bayes' theory is justone of several
techniques used inmedical decision analysis.
Another displays sequences ofsteps representing various
possi-ble actions and events. Sequencesof this type resemble
tree-shapednetworks. Nbdes or junctions in the
_,. ,,iAtree are of two nf MS. t (Aversionnodes, the clinician
chooses fi(mi aset of possible Rctims. One actionmight be deciding
to perfor m a cottain test. At chance nodes, thepossible responses
ofpe patientto some action that ha6,been takenare represented. When
performinga diagnostic test, the patient's tresponse----whether he
developscomplications, for example----is amatter of statistical
likelihood. Byusing the dedsion Vein, a cliniciancan come to a more
infbrmed con-clusion about the range of alterna-tive
strategies.
Modifying the tree by attachingpatient-oriented values to
decisionnodes makes the simulation more-realistic. For exampld, a
definitivediagnosis might not be pursued ifthe required tests were
expensiveand painful, if the health of the pa-tient were not
threatened by thisinaction, and if rendering a defini-tive
diagnosis would not sirifi-canny improve tis health.
.
iThe effort to develo these ap-plications into program using
arti-ficial intelligence began n the early1970's. The intent was to
focusprimarily on the use of symbolic
reasoning tee;hniques. 1 he Objec-tives have been to capture
thejudgmental or heuristic knowledgeof expel ts for
decision-making, andto constf uct I oasoned and ()Wain-able
solutions
ifor diagnostic prob-
mles. Generla y the logic built intothese prograr s is composed
of sixmajor elemer*,_
Plan-Gene te-and-Test. ln thisframework;'- ie program uses
.heuristics t elect the generalarea in whit the answer is likelyto
be found; t generates plausi-ble Solution within these boun-daries,
and; sts conclusionsagainst ob rved data, appro-priately rev IN
conclusions untilone that be. fits the data is un-covered.
tDomain-SAcific-Knowledge.Much of th power thatdecision-m' ing
programs holdiis derived f m specific rules andknowledge:- bout the
target areaof applicati n. Such knowledgebases enc e factual
informa- .lion about e domain and theheuristic ri, s used by
expertsto rapidly fird solutions to prob-
,! ..terns. .
}artrc f;. 1
t44.1.
trg ..1'
"Why don't you chea yip the local data bank?"
-
Flexible Knowledge Base. Ifchosen pr operly, the knowledgebase
is small enough to b"e han-dled adequately by the com-puter, but
large enough to bemeaningful to the prospectiveuser. Once the basic
program iso Irating, knowledge can be
d, removed, Or changed byusing an explicit and flexible
en-coding of the knowledge.Line-of-Reasoning. Specialistsin the
target area of an applica-tion must be able to flollow thelogic
used by the program whenit generates conclusions. Al-though not
strictly necessery,specialists should also agreewith the route
chosen. To ac-complish these goals, computerscientists in SUMEX-AIM
teamup with experts in target fields tolearn the mechanics of
reason-ing. Human logic is then trans-lated into computer language
inthe form of symbolic rules.Multiple Sources of Know!-edge. Often
several practitionerslend their expertise to the designof Al
programs. Textbook knowl-
dgeis usually incorporated aswtI. Having access to knowl-
t_oti"s%'%,fro
edge representing varied points(-4 view ran speed HO
plerosslocating a solution and ieducethe chance of overlooking
alternative 9Q1itions.Explanation. The program mustbe able to
explain the lino ofronsoiling thalled to its conclu-sions. It not,
the user cannotunderstand the basis for theprogram's conclusions.
Also,thrqugh the explanatory function,flaws'in the program's logic
canbe located and fixed without ex-tensive study.
Over the last decade, computerscientists have used these
ele-ments to build many types of pro- 7grams. Some include the
ability tolearn. Others emulate crketivity.Those in the
SUMEX-A.IPtietworkare devoted to expert problem-solving in
medicine, biochemistry,or psychology.
A
tt,
-
SUMEXand theSOenceCommunity
TheSeeds ofArtificialIntelligence
,
A typical strategy in some Al re!Watch IS to choose a problem
thatis tightly foctismi and easy to concoptualize, such as a game
Ibisapproach offers cm tain advantages, most notably that ideas
rtribe tested with minimal experiso oftime and money
These games are called toyproblems because they serve
Flopractict.al use An example is themissionaries'_dilemma, a puzzle
inwhich three missionaries want tocross a river, but their efforts
arestymied by an equal number ofcannibals. A boat thtit holds
asmany as two people is available,but the missionaries must never
beoutnumbered, or they will becomethe main course of that
evening'scookout.
Projects in SUMEX-AIM gen-erally shun Joy problems. ''Theiruse
leads to sterility in that youquickly figure out the solution,
butare not faced With the additionalchallenges that a messy world
pro-vides," Dr. Feigenbaum says. "Weseek our inspiration from
programsdirected at diagnosing disease orassisting biologists in
planningDNA manipulation experiments,
t/tICArl!;F, plOWIll!, atopPll (lilted and rich
ho key to designing a socces'.ful Al pioject, he says, is to
pick aproblem limited enough to be congamed, Nit not so simple Phil
Oreprogram designed to solve It call.not be expanded into a
practicaltool. Most of the time, projects inSUMLX-AIM are
restricted to asubsection of an inteVed area ofapplication. When
that segment isadequately covered, boundariesare carefully
extended.
An equally important criterioncalls for an association
betweenthe project and at least one expertfrom the target field of
application.The collaboration must be a dedi-cated one, according
to Dr.Feigenbaum. "You cannot have thekind of inspirational meeting
ofminds needed f9r a project to suc-ceed if the specialist and
pro-grammer meet every once in awhile," he says. "It takes
aquarter-time to half-time effort bythe expert that stretches over
anumber Of years."
The seed from which SUMEX-AIM grew embodies this type
ofcollaboration. Known as DEN-
4:ZN
A.The missionarie dilemma: use ofsuch toy prob ms In Al research
Isnot produc ,
B.
A technician in the Stanford massspectrometry laboratory:
generatingdata for Ilse VENDRAL.
-
lAt it booan Ii too; when Iiieigenbaum tol(1 1)1 Joshuaoderberg,
then chairman of the
Stanford genetics depar tmont,about his interest in modeling
sci-entific ffintight with Al techniquesChemistr y was (:trusel) as
thotaeget field for tWO major reasons.Hist, fnuch knowledge in the
fieldalready existed in machine-readable Mon. Second, chemistrywas
tho field in which Dr. Leder-berg was expert. When the projectgrew
in scope, Dr. Cart Djerassi,Stanford professor Of organicchemistr
y, was recr uited.
Applying Al to science was inevi-table, according to Dr.
Feigen-baum. "As the computer grew inpower and the cost of its use
de-creased, more and more spe-cialties looked to the computer
forassistance in information process-ing," he says. "But very few
spe-cialties in medicine and other fieldsof science could be
modeled byformulas and calculations, whichare the-traditional means
of exploit-ing the computer.-
Programs of this type must em-ploy processes similar to
thosepresent int human reasoning and,
lheiefore. must he expressedsymbolically, I I ergenbilumJo
achieve this, it is necessary todeyekw techniques by which symbols
can be reptesented andmanipulated Rut when 1)1 NI)f-1,41vva!;
COIICPIVed 5011111 1!) year!; ago.
Intelligence was truly afledgling discipline.
BiochemistryDENDRAL
-The project was initially begunas a prototype to demonstrate
thatcomputerized symbolic reasoningcould be successfully, applied
tomolecular strycture pralltnis inchemistr y. The program
illustrargswell the evolution of Al work.
In solving problems, DLNDRALuses instrument data from a
massspectrometer (MS) and a nuclearmagnetic resonance (NMR)
pec-trometor, together with othiar con-straints on structural
features in themolecule. These constralbts de-scribe configurations
of atoms andprovide limits within which the an,-sWers, structural
candidates for anunknown compound, must fit. Suchconstraints
eliminate the produc-
lion of itilde',HrdWh101, 1),1!.,0t1 (n) /WWI( ,1) or
ohm gefic ghlurld, are ImplauslhleOrs I oder ber g and
roiguilbauni
quicklinalized the power provIdedby supplying severi-d soutcw.;
ofknowledge when ;II mlyringmolecular str witures litan earlycase
run on.D1..NDRAt., con-straints based on organic chomis
y principles alone would haveadmitted 1 25 million
plausiblecandidate structures fOr a singlecompound under study. The
scien-tists r esponded by adding informa-tion from proton NMR
analyses,from which the program could infera few additional
constraints. "Theset of 'plausible candidates wasthen reduced to
onothe nightstructure!" Dr. Feigenbaum recalls."This WIS not an
isolated result butshowed up dozens of times in sub-sequent
analyses."
The original DENDRAL programwas restricted to a small number
ofmolecular families for which theprogram had been giveh a
special-ist's knowledge, "namely thefamilies of interest to our
chemist-collaborators,-Dr. Feigenbaumsays. "Within these areas,
DEN-
to
01116,!.. ,
f\'
4 447 j
;Z:Pha 1;44,t,/,44t,i:Cer
;4 eiol:/ 1
.
4., ,,
>.,-.V.!`-:
-
DI1Al 's pe! foynance was usuallynut only rillIch 1ii!;tel tntt
1111,o MOW14.c-wilt() than ewer I human perf(i Malice
1)1 Bruce Iluchanim, a memberof the DLNDHAL team, explainsthe
general approach of DI_ N-DHAL "There are three phasesplan,
generates, and test," he says"In approaching a problem, DEN-DHAl,
makes some rough guessesas to what the solution should look
hat is the planning phase.The generation phase works withinthe
established constraints of theplan to develop plausible
solutions.
. Filially, each plausible solution istested."
Testing is accomjVhed in twosteps, which folloW. a
"model-drivenstrategy." First, the computer gen-erates sets of
instrument data thatwould be expected to describeeach candidate
structure. Thesesets are then compared to actualdata about the
compound. Theclosest fits are retained and rankedaccordingly.
Having enough knowl-edge about the characteristics of acertain type
of compound to domodel-driven analysis drasticallyreduces the
amount of data that
mut,t exammod, since the dataare use(i mainly to verify
pos,;ihlearroA/011;
1)11\JIMI1\1'; 11trmary limitationwas its restriction to only a
smalluubset of organic molecules, thesaturated, aliphatic,
monoluncholm! compounds. Work carriedout alter DENDIIM's early
suc-coss has focused on thestructure-generation aspects of theplan,
generate, and fest paradigmFrom this paradigm, the
structuregenerator, called CONGLN forCONstrained structure
GENera-tion, has been extracted. CONGENis the segment of the main
pro-
sgram that is not closely tied to spe-cific instrumental data
and is,therefore, of greatest use.
"Chemists have many sources ofdata for both planning and
testing,so the use of DENDRAL as awhole, which would restrict them
toNMR and mass spectral data,would be a hindrance," Dr. Bu-chanan
says. "That is why, in thelast 3 years, almost all the effort onOA
project has gone into develop-ing CONGEN, since it has thewkiest
possible applicability."
Now under the direption of Stan-
2 6
101(1 t;tlertil;;Is s Carl Djerassiand Denni:i niitli, the 1)1
NDliAlproject has evolved into one of thebest known mut nwst
successfulapplIcations of artificial intelli-gence the CONGI:N
program andrelated subprograms aid efiernistsin determining the
molecular struc-ture of unknown organic com-pounds. Because tho
molecularstructure-tf a compound must beknown before its other
propertiescan be studiedproperties relatedto pharmacology or
toxicology, forexampleDENDRAL promises animportant contribution to
biomedi-cine. Some investigators have al-ready capitalized on this
offer.
Durinthe past 5 years theCONGEN program has been
usedsuccessfully by chemists workingon biomedical problems at
Stan-ford and other institutions. Abouttwo dozen scientists use the
pro-gram each year when solvingquestions about the structures
ofcompounds. Invesfigator affiliationsare split about 50-50 between
uni--versities and private industry. Theprogram has been exported
toseveral laboratories in the UnitedStates. The British government
is
J.
-
r
now supporting work at the Uni-versity of Edinburgh aimed at
link-ing industrial researchers in theUnited kingdom with CONGEN.
Acopy of the program now runs onthe,Edinburgh computer. A col-
, league at the Australian NationalResearch Organization is
alsospeaiheading an effort to makeCONGEN available in that
country.
More recent research effortshave been directed to
extendingCONGEN's representation of struc-ture even further. The
program willsoon include principles of molecu-lar stereochemistry,
or three-dimensional representation of
'structures. Stereochemistry is ab-solutely essential in
understandingstructures and interactions ofmolecules in chemical
and bio-chemical *terns, Dr. Smith ex--plains. This new work is
pointedtoward a system of computer-based planning and testi%
whichincoiporates chemical ancispec-troscopic data from several
differ-ent techniques.
As the forerunner of Ar§ shift toknowledge-basecLanalysis,
DEN-DRAL holds a special place in'computer history. It
demonstrated
CONGEN printout currently one ofthe most successful applications
ofartificial intelligence, this progrdmhelps chemists determine
themolecular structure ()Organic com-pounds.
2 V
-
the superiority of domain-specificknowledge as a means to
achieveexpert performance and in sodoing raised impor tant
issuesconcerning knowledge representa-tion, acquisition, and
use.
But, more important than its ob-vious contributions, the
programdemonstrated that Al concepts andprogramming techniques were
ad-vanced enough to produce usefultools, although each could
dealwith only one limited specialty. Thisexample of competence,
accordingto Dr. Feigenbaum, vastly im-proved the credibility of Al
andpaved the way rcir other such sys-tems. "For us, the DENDRAL
sys-tem has been a fountain of ideas,many of which have found
theirway into our other projectS," Dr.Feigenbaum says.
Meta-DENDRAL
The project in SUMEX-AIM mostclosely associated with DENDRAL,as
might be expected by its dame,is meta-DENDRAL. Developed byDr.
Buchanan, professor of corn- .puter science at Stanford, t e
pro-gram learns rules,about a s cific
0')
type of compound by examiningdata from a set of examplesThese r
tiles can then be used tointerpret data concerning unknownorganic
compounds. Both DE_ N-DRAL and meta-DE:NORM_ usethe same rule-based
logic. Criteriasot up by expert chemists guidemeta-DENDRAL's
generation andselection of rules.
Dr. Feigenbaum says the pro-gram was evolved from DENDRALfor two
reasons. First, it was de-cided that DENDRAL has laid afoundation
firm enough to pursuethe deeper study of scientifictheory
formation. Second, it wasrecognized that acquiring expertknowledge
of a specific domain.was the bottleneck in building pro-grams
targeted for real-world use.
Meta-DENDRAL was originallyintended to complement the
parentprogram. Its job was to formulate .rules for interpreting
data frommass spectrometer analyses. Insuch analyses, molecular
frag-ments are sepatated according tomass and electrical
charge.Meta-DENDRAL's output is sets ofrules that deScribe how
moleculesfragment when studied with mass
28
5;p( wtt mliotr y (MS) MetaN1DIt At also includes evidence
suppor ting each fragmentation ruleand a serrnmar y of contr
adicter yevidence. Constraints, fed in bychemists, guide generation
of pilesalong desired lines.
The program, like DENDRAL,uses the plan-generate-testframework.
The process includesthree steps: interpret the data andsummarize
evidence; generate aset of plausible candidates; testand refine the
set of plausiblerules.
In the first step, meta-DENDRALcites each piece of MS data as
ahighly specific point of fragmenta-tion, then sums up the
evidencesupporting such fragmentation andthe configurations that
wouldcause these atoms to separate.The next step is a heuristic
searchfor general rules thar govern thefragmentations. The search
beginswith the single most general ruleand proceeds toward more
de-
jiailed specifications. This processcontinues until the program
de-cides that the rules being gener-ated are becoming too
specific.Meta-DENDRAL also includes a
. Dr. W. Todd Wipke, principal inves-tigator of the SECS
project: design-ing syntheses faster and without thebias of past
experience. .
-
criterion for deciding whether anemerging [UR? IS too goner
al
In the final stage, the programtests candidate rules,
comparing111;isitive and contradictory evi-dbnce. Those with a
negptive barancie are disregarded. Rules withiedundant features or
suppor tedby the same evidence are merged.
The end resultis a rule-set ofcomparable quality to those
thatcould be OonerEite'd by human ex-perts, according to Dr.
Buchanan."In some tests, meta-DENDRAL
kricreated rule-sets that we had '-ffeviously acquired from our
ex-perts during the DENDRAL proj-ect," he says. "In a more
stringenttest, involving a family of com-pounds for which the
mass-spectral theory had not been corn-pletely worked out by
chemists, theprogram discovered rufe-sets foreach subfamily."
These rules were judged by ex-perts to be "excellent." A
paperdescribing them was published inthe American Chemical
SocietyJournal in 1976.
Emphasis during the past year ,has been to make
meta-DENDRALmore-efficient. A major overhaul
f
CieP flt74-411
11:., vf f .11,1414011,Will
was accomplished, largely rem-(Jiff Illing UM mettkdc; by
whIchtho program Works:',With West)changes, the ability
ttl9eneraterules concerning a different type ofdata, carbon 13
nuclear magrieticresonance, was incluct,, . Severalpapers were
published r,.. 1979 Onthe rules generated irI Mrs area.
SEORe
The SECS (Simulation.andEvaluation of Chemical Synthesis)project
is aimed at describing thelogical principles used,,when con
Dstructing molecules. evelopedprimarily by Dr. W. Todd Wipke,
achemist and computer scientist atthe University of California,
SantaCruz, SECS is Intended to promotethe development of new
andmodified drugs, as well as syn-thetic compounds modeled
afterthose that occur 'naturally. In par-ticular, the project is
concentratedon assisting the chemist to designand select syntheses
of biologicallyimportant molecules. Dr. Wipkesays the computer
offers severaladvantages over conventionalMethods.
2 9
1.10ng chemists shouldbe able to (10!;icIfiand without the bias
of pa,;t experience," ho explains. :Many morepossible syntheses
will be consid-ered because of the system's ex-tensive library of
chemical reac-tions, which is larger than anyperson can remember.
And thecomputer can better process andrecord the many structures
that willresult."
Through on-site terminals ortelephone links, investigators
fromuniversity, industrial, and privatelaboratories are now using
SECS.Versions of SECS are available byacdessing SUMEX-AIM, or at
theUniversity of Pennsylvania MedicalSchool, the International
ADPNetwork Computers, or Merck &Company, Incorporated,
amongothers. Dr. Kenneth Williamson ofMount Holyoke College
usedSECS to build three-dimensionalmodels of some 50
compoundsparticularly important in nuclearmagnetic resonance
spectroscopy.Other scientists have successfullyused the program to
design chemi-cal syntheses. One chemist usedSECS to develop
procedufes for
ailMeAdin1111
-
making synthetic morphine.1 hese users have given us a lot
01 suggestions for improving theprogram," Dr. Wipke says.
"Somehave contribrad new reactionsand quite a few people from
indus-try have actually contributed laborto thbe projectquite
sophisticatedlabibr. In one case, an organic ,chemisf from
Hoffinan-LaRoche'who had worked in the fieki of . rheterocyclic
chemisio/ for 10 years3pent a year endowing the pro-gram with his
knowledge ofchemistry."
The scientist's lack of experiencein computer science was nqt
aproblem because S6CS uses aspecial language called ALCHEM,which
was develop d by Dr.Wipke's group, He says it is com-posed of
declarati es that describehow the environment of
moleculesinfluences chemical reactions.
"A chemist can understand thelanguage and rotild a reaction
withonly about 5 minutes of explana--tion;-" he says. "To actually
use thelanguage well takes only a coupleof days."
When working on a problem, theprogram studies data about the
30
natur al tar got molecule and conttucts a three dimow;ional
model
for diSplay on 0 graphics tem mina!!lased on the analysis, SI
CSdraws from its kriowledge base toselect reactions that could be
usedin the last step of the synthesis andthen backtracks through
the re(wired peocursorst. het system stimulates thechemist's own
creativity," Dr. Wipkesays. "It presents many differentand unbiased
approaches to thesynthesis." The chemist guides thecomputer through
the process bypointing out the most interestingtechniques.'
"This is a unique feature of ourproject in terms of Al
research.Usually programs are designed tofind one good way to
accomplish atask. We are interested in findingall the good
syntheses, and thatinvolves dealing with plans, plansthat have many
branches and "many contingencies," he says.
There is another feature thatsets the project apart from
othersin the field, according to Dr. Wipke."Our program is
interactive. We aretackling the problem of synthesisfrom the
viewpoint of how best to
-s
tl!e the heinist and the compoteily; a team and to have each
teammember doin() the tasks tot whichthat memhet is hest !railed. A
10 ofAl has heel] diiected at how tomalio the computei do the
wholething with very little emphasis onpresenting intermediate
results tothe user in a form that allowst.search process to he
guided, inter-rupted, stopped, or redirected."
Dr. Wipke and colleagues,mostly synthetic:organic chemists, .are
currently expandipg the pro-gram to include mbre complexstrategies
for designing syntheses."Essentially, the prograffwill havea more
preciely directed search,and it will be more selective inwhat it
generates," he says.
But before these strategies canbe piq into the computer, they
mustbe exPlicitly defined, which is often,difficult to accomplish.
For in-stEince, strategies based on princi-ples of symmetry are
learned from.experience ratheethan fromtextbooks, Dr. Wipke
explains. Forthe computer to recognize a sym-metrical design, these
principlesmust be'dissected and reassem-bled In the form of
software.
-
Unlike the cur rent version ofSF.C.,S, which uses Al CIII M
toexpress r ules eonceinipg chemicalreactions, strategies., will be
wr ittenusing mathematical equations. Ex-pressing knowledge in this
formwill allow the Wipke team to buikiae explanatory function into
theprogram. If questioned by thechemist as to why a certain
reac-tion was chosen fakthe synthesis,the computer will brPable to
reply,citing strategies of chemistry.
In final form the strategy portionof the program will complement
thepart that deals only with reactions."The current program deckles
whatto do by consulting a list of goals,"Dr_ Wipke says, "and that
goal listwill be created by this higher levelreasoning process
which picks outthe strategies applicable to thesituation and
explains why. Theprogram will then select ways toimplement the
strategies and, fi-natty, decide hew to modify themolecule's
structure. This multi-step procedure allows a view of theproblem
free from human bias."
Dr. Wipke hopes to demonstratethat computer-based
synthesistechniques can also be applied te
SECS printout: helping the chemistdesign and select syntheses
ofbiologically important compounds.
4
the study of metabolism. nursed ontechnology from the Sf:CSgram,
a new compute! programcalled XENO has been developedto predict
metabolic pathways forxenobiotic compoundS--chemicalsnot normally
found in the body. Theobjective is to predict plausiblemetabolites
of a given xenobiotic."What you put in is t6e chemicalstr ucture of
the foreign compound,"he says. "What you get out is thechemical
stRicture of the metabo-lites."
Predictions of plausible metabo-lites result from knowledge of
howcompounds are activated by en-zymes. Many of the
mechanismsinvolved in these processes areknown and more are being
discov-ered. j
In addition, the metabolite'sstereochemistry is predicted.
Acompound may exist in two forms,each the mirror irpage of the
other.s.One may be active while the otheris not, or they may
botfvbe activebut produce different effects.
"Stereochemistry in metabolismis a new frontier," Dr. Wipke
says."In tApast, instruments were notsensitive enough to explore
this
angle using the amount of motabolito that was obtained
In recent months problefli;been submitted to the progrwn totest
its ability-to predict metabo-lites. Dr. Wipke says XLNO hasbeen
fairly successful at identifyingmetabolites found in
laboratorystudies, and also predicts metabo-lites that have not
been found.When discrepanbies occur, Dr.Wipke says, they sometimes
canbe traced to errors in the knowl-edge base. "The computer
maypredict more metabolism than isactually going on in living
sys-tems," he says. 'That's really nOttoo bad, because the
metabolitesthat can be isolated will always beincluded in the set
of metabolitespredicted by the computer. Theprogram defines a set
of candi-dates to look for."
Dr. Wipke and colleagues havenow foCused on expanding
theknowledge base, particularly to in-clude models of more
species.Only the rat and the mouse arecurrently described in
detail.
An index of biological activitiesassociated with
metabolitesforexample, nrcinogensis slated
*/
31-
-
#
for inclusion Iho functrtni willapply pattern r ecognition to
Awnpounds not listed in the index as ameans of classifying
moiatwlites
The XENO project is not the onlyspin-off from SECS. In 1978
theSECS program led to developmentof a daughter project that
extendscomputer-assisted synthesis intophosphorus chemistry. Under
thedirec3ion of Dr. Wipke, Drs. GerardKaufman and Francois Ctioplin
atthe University of Strasbourg inFrance created a knowledge
basecomposed of reactions pertainingto phosphorus. In analyzing
sev-eral compounds and searching theappropriate literature, the new
sys-tem found most of the existingsyntheses and, more
Importantly,suggested new techniques thatappear to be equally geod
or bet-er, according to Dr. Wipke.
OLGEN
Experiment-planning in themanipulation of DNA is the goal
ofMOLGEN, a Stanford project beingconducted in collaboration
withScientists at the University of NewMexico (UNM). Program
develop-
4k,
.,.,,,
_ _.,l's;-,!.
, ,y, ..,.:*c ,wit-,".' IY.'3 ,:t '' ,:..1..,,,t4 it '''' '0'
ril4 '''''.'',',:,,,V,:.1,,,r'Y,, ,,c. , ,, ,c,, 10,4,...2L.1,
0,;disL-,. .rstty-k!,44,4,t,.viv'P,',
'cl1, c 4', ' ,''''. "1-..cs rill t'`I 0
.} ,' . \ ,a-1,(...-thzo.ts,'',1[0 ,-,.
e,l'.r.'.1,',))t%.4),t,'; s::'-,
,.,!.... mj. 4
g4.,,,.... :5 Ye'4 ,i,
..........--n,.,,,,. ,,,, pp;.:,,, ',,,i,A.:?.. .k- .
.r 0,- si! 4.1
merit is primarily ution of I)rs 1 auron
dward I eigenbatscientists sPeter I riodland, anDoug utlag
MO1.to advise geneticists about the dosign of labor ator y
experiments.These include methods used toanalyze and modify
nucleic:acids.
MOLGEN is mainly foc(rsed onorganizing experimental tech-niques
and determining the orderin which they should beapplied toachieve
specified goals, Dr. Brutlagsays. "The enormous volume ofdetailed
knowledge makes it likelythat good experiments am beingmissed," Dr.
Feigenbaum says."We believe that an intelligentplanning assistant
can offer help inanticipating the results of combin-ing
experimental methods in manyways."
Dr. Peter Friedland saysMOLGEN makes near,expert deci-sions when
selecting physicalmethods, such as electron micro-scopy or
enzymatic modification, toanalyze molecular structure. Even-tually
the program will be ex-panded to include the design of
del the direc:0 Ii. Kedes andn, compute!Stefik andbiochemist
IV
lEN's task is
synthesis experiments in whichmethod; for buildiN molecules
willbe described I midi() c ial analyses
,uhieti as well, ,ilk)wing ttnprOgr am to klentify the products
of
:nucleic acidsThe success of MOl. GEN as an
experiment designer depends onthe quality of its knowledge
base.Much effort has boon expendedlosupply the base with explicit
infor-mation about DNA str uctures,restriction enzymes, a hierarchy
oflaboratory techniques, and a grow-ing collection of
genetics-orientedstrategies for discovering informa-tion about
various aspects of DNAmolecules. Most common ana-lytical and
manipulative methodshave already been pul fn the base.
Results of the research includesome special-purpose programs
inthe area of molecular genetics.The most useful are highly
refinedversions of previou.sly existingstrategies. Many of these
concerndetermination of the sequences ofnucleic acids in DNA.
Modificationsare focused on technical aspectsof the prodrams; those
leading to'impftived efficiency, for &le,and those
addressing human en-
04 4 t
II
41;
A t0,1,`,46.'.
DNA's spiral ladder of heredity: help-ing chemists manipulate
themolecule through well-planned ex-periments is the tOCUS of
MOLGEN.
sit
-
gineering concerns to make iteasier for icientists not
familiarwith c4Anputors to use the progrfims.
In addition to its applied,or ienta-MOLGEN includes an Al
re-
search dimension: use of theknowledge domain of
moleculargenetics to create a generally ap-plicable problem-solving
program.lhe system is designed to allowgeneralization into domains
be-yond genetics in future,researchand application.
"Integrating the many diversesources of knowledge is a
centralproblem in constructing MOLGENbecause the expert-planning
pro- ecess requires a blend of biological, .genetic, chemical,
topological, antiinstrument knowledge," Dr,Feigenbaum says. "The
expert'sknowledge Of experimental strate-gies must also be
representI antiput to use."
PROTEIN STRUCTUREPROJECT
fluilding computer models of,pro-tein structures from cryStallo-
-graphic data, particularly electrO
Wpwamere-vo
u.o,.......--..w.n-eyeNwim.. no...- a
density maps, is 1,he goal Of thePRO II IN SHIM:flit IF 1)10)0(1
at:Aanforo I lectr On dt.msrly nhipsarc data f ()presenting the sti
tic-tures in three dirnensivs. Unfor to-nately, these maps well
liallycrude and ambiguous As .1 result,the program depends largely
enbackground ir\for [nation, such asthe amino acid s'equence in a
pro-tein, for guidance and support informing hypotheses about
thecompound's three-dimensionalstructure.
f3ecause the shape of amolecule exerts a major effect onits
performance, accurate analysesand representations of
molecularstructure are seen by medical re-searchers as essential-to
under-standing the biological function of.these complex
molecules.
Interprating electron densitymaps is the prt of a
proteinchemist, which the system's logicscheme attempts to capture
.through the use of heuristicrules. Due to(the size of
proteinMolecules, which often containmany thousands of atoms,
theplan-generate-and-test strategyusedby,DENDRAL cannot be em-
.
.0 MOMow.w-.111, Nu. 411..
...ft OIMONON~Oon....enft onononrfto nano.
Re AIM, IMMO* alsepwrog...Amp mow wommishilywe
oft o dnnoo or o o .01 go ann. von... =wow .....onr-
iNry w OROO. NNW /WNW.. nnftno.
Ow *OW=...j'"1".'"
ployed !lather, the :;\;!Aern precestogether hypothows try (
rcei rIm ntint.) :,11(Te!;!>ively on !.pecrfW atex;of the
protein 1 he project is underthe direction of Drs. I ergenbaurnand
Rober t kngelmor 0 of StanfordUniversity with assistance born
Mr.Allan -Fer y, it the t fniversity of Calanima (1 Ir vine, and
the strongcollabor ition of Dr. Stephen itemat the C San Diego.
Clinical Medicine
INTERNIST
Heuristic search tethnigues areused in all SUMEX-AIM
projects,although each differs according tothe purpose.of the
project. Drs.Jack D. Myers and Harry E. Pople,mentors of INTERNIST
at the Uni-versity of Pittsburgh, reasoned thatthe best way to
design a computerprogram for solving difficult prob-lems is to
simulate the mentalprocesses used-by people. Theyare primarily
interested in buildinga Program thal will aid skilled vpe-cialists
in solving cOmplicated prob-lems concerning internal
medicine.Spin-offs from the program might
rod.
,
41 '
Afilk 44"."-'111
-
,
,4F
be used tIy physicians assistantsor ill rur al health care
clinics, milltar y outposts, and spacecraft. tobe effective, the
program must beable to diagnose several diseasesif they are present
in a tingle pa-tient, and it must render diagnosesquickly to
reflect the current statusof the patient.
Drs. Myers 6nd Po le analyzedthe diagnostic routlnq followtd
bythe expert clinician ahd establisheda set of criteria:
pbservations foci into the corn-PUtcar must evoke Ihe
appropri-ate.hypotheses of disease.Hypotheses must generate a
listof manifestations that would bepresent in the patient if the
diag-npsis is correct,The computer must be able torank models of
disease accord-ing to their probability of beingcorrect and must be
able to delcide when the weight of evi-dence is sufficient to
permit rea-,sonably Confident judgments-1-he program must be able
todroup hypotheses into Mutuallyexclusive,subsets correspondingto
different diagnoses, in order tohandle cases in which more than
si
At
Ono disease may bo presentSince beginning their work in
1970, pls. Myers and Pople havedeveloped an operative system,and
in so doing have par tiallyachieved these
objectives.INTEIINISTUccepts descriptionsof disease manifestations
in anyorder and asks for more infor mo-tion, stich as historical
items,symptoms, signsl and laboratorydata. These facts are not
enteredin a specific order, but rather asthey are gained through
tests andobservations. As factsjaccumulate,nodes of recognition are
triggeredand a pattern begins to develop.
Now, with a specific direction,the computer fits the data
togetherlike pieces of a jigsaw puzzle. Aninterlocking web of
programmeddata isiset up, beginning with cate-gories uch as liver
disease lead-ing into specifics like hepatitis A.After sufficient
data have been fedinto the computer, disease modelsare developed.
The models arethen compared and ranked.
"The computer holds the profilefor each disease in its memory
andif the model fits that standard pro-file very closely, it could
make a
3 4
di gnosis,- Dr Myers says "If thatis 1 possible, it will set out
askingquestions to obtain fur ther )nfor malion, so that one or
more of themodels can tnconfir med."
A second generation programdubbed INTERNIST II, which mayspeed
up the diagnostic procedure,is now being designed.
Althoughexperimental, the new programhas shown promising results,
rais-ing hopes that it will lead to-a moreefficient workup of
clinical prob-lems when the program is applied.
Dr. Myers predicts that within 5years INTERNIST might be
diag-nosing disease on a practical,rather than experimental,
basis.When.completed, the 'system owillbe able to assist physicians
work-ing on difficult cases and .paramedics serving in reMote
or.medically underserved areas.
"The computer's assessment ofa patient's condition will be
re,garded as evidence to help thepractitiondr form a diagnosis,"
hesays. "The program is intended to .serve as a conwltant, not as a
re-placement for the physician."
Currently, more than three-fourths of the knowlelge appli-
Drs. Jack Myers (poin4) and Harry-Pople VERAYST: "No one
canpossibly emorize all the data inmedicine."
ttl
-
cable to intern' medicine, 000 ofthe broa(1est )ecialties, has
beentranslated ir o symbolic data str tic-tures and stored in the
computermemory. Over the past 2 years IN-TERNIST's ability to
translate thisvast store of knowledge into accu-rate diagnoses has
been proved,using a variety of difficult casestudies that were
published inmedical jour nals or occurred inPittsburgh teaching
hospitals. "Inthe great majority of cases theprogram has been
effective ia sort-ing out the pieces of the puzzleand coming to a
correct diagnosis,"Dr. Myers says. '`The knowledgebase is too
incomplete for a com-prehensive test in a clinical situa-tion,
although it is used on anad hoc basis at Presbyterian-University
Hospital, Pittsburgh, forclinical guidance."
Within 1 year, the knowledgebase is expected to reach a
"criti-cal stage of completeness," ac-cording to Dr. Myers. Soon
after,field trials of INTERNIST arescheduled to begin at
Presbyterian-University Hospital. If successfulthere, a half-dozen
other healthcare centers will take part in the
testing At each institution, Dr.Myers estimates, 20 caw
analyseswin be run each day. During thetrials, physicians'
reactions to thesystm and their patter n of use willbe retorded. On
the basis of thisinformation, INTERITST will be re-vised, if
necessary, to improve ser-vice to future users.
In the past year, Drs. Myers andPople have devised a
programcalled ZOG, which makes it possi-ble for a physician only
casuallyacquainted with computer scienceto master the use of
INTERNIST,reportedly within 5 minutes. Testsshow that ZOG,
developed atCarnegie-Mellon.University, is veryversatile and easy
to use. Dr.Myers says ZOG is important be-cause the computer must
be easyto operate if it fk. to bridge theever-widening gly between
what isknown in medicine and whatphysicians are able to
remember.
"No one can possibly memorizeall the data in medicine," he
says."There's just too much kngwledgeand that pool of information
is con-stantly increasing. The computerhas a perfect memory and is
ad-mirably suited for a large knowl-
4
;
ilk