/LMAR=0/XLINE=3/FONT#0=BAXL30/FONT#1=BAXM30/FONT#2=BAXB30/FONT#3=SUB/FONT#4=SUP/FONT#5=BASL35/FONT#6=NGR25/FONT#7=MATH30/FONT#8=FIX25/FONT#9=GRK30/FONT#10=ZERO30/FONT#11=BAXI30/FONT#12=MS25



␈↓ ↓H␈↓α␈↓ ∧rARTIFICIAL INTELLIGENCE

␈↓ ↓H␈↓␈↓ ¬Lby John McCarthy

␈↓ ↓H␈↓␈↓ β_[draft of article on artificial intelligence for Collier's Encyclopedia]

␈↓ ↓H␈↓␈↓ α_Artificial␈α∞intelligence␈α∞is␈α∞the␈α∞branch␈α∂of␈α∞computer␈α∞science␈α∞concerned␈α∞with␈α∂making␈α∞machines
␈↓ ↓H␈↓behave␈α∩intelligently.␈α∩ Its␈α⊃scientific␈α∩part␈α∩involves␈α⊃discovering␈α∩what␈α∩intellectual␈α∩mechanisms␈α⊃are
␈↓ ↓H␈↓needed␈α
to␈αsolve␈α
various␈αproblems,␈α
and␈αits␈α
engineering␈αpart␈α
involves␈αwriting␈α
computer␈αprograms␈α
or
␈↓ ↓H␈↓building␈α∂machines␈α∂to␈α∞solve␈α∂one␈α∂or␈α∞another␈α∂kind␈α∂of␈α∞problem.␈α∂ It␈α∂ultimate␈α∞goal␈α∂is␈α∂to␈α∞understand
␈↓ ↓H␈↓intelligence␈α
well␈α
enough␈α
to␈α
make␈α
computers␈α
more␈α
intelligent␈α
than␈α
people,␈α
but␈α
we␈α
are␈α
far␈α∞from␈α
it
␈↓ ↓H␈↓today, and some people don't believe it possible.

␈↓ ↓H␈↓␈↓ α_The␈α∞idea␈α∂of␈α∞intelligent␈α∂machines␈α∞is␈α∂old␈α∞in␈α∂legend␈α∞and␈α∂fiction,␈α∞but␈α∂legend␈α∞and␈α∂fiction␈α∞are
␈↓ ↓H␈↓more␈αconcerned␈αwith␈αhow␈αan␈αartificial␈αintelligent␈αbeing␈αmight␈αbehave␈αthan␈αwith␈αhow␈αit␈αmight␈αbe
␈↓ ↓H␈↓made.␈α Until␈αthe␈α1950s,␈αrobots␈αwere␈αusually␈αdepicted␈αas␈αenemies␈αof␈αmankind␈αusually␈α
arising␈αfrom
␈↓ ↓H␈↓human␈α
weakness␈α
or␈α
greed.␈α
 They␈α∞then␈α
played␈α
a␈α
brief␈α
role␈α∞as␈α
an␈α
oppressed␈α
minority,␈α
and␈α∞in␈α
the
␈↓ ↓H␈↓1960s they became neurotic and even guilt ridden just like the human characters in literature.

␈↓ ↓H␈↓␈↓ α_In␈α∂the␈α∂early␈α∂1900s,␈α∂the␈α∂Spanish␈α∂inventor␈α∂Torres␈α∂y␈α∂Quevedo␈α∂made␈α∂a␈α∂machine␈α⊂that␈α∂could
␈↓ ↓H␈↓checkmate␈α
its␈α
opponent␈α
with␈α
a␈α
rook␈α
and␈α
king␈α
against␈α
a␈α
king,␈α
and␈α
speculated␈α
intelligently␈αabout␈α
the
␈↓ ↓H␈↓possibilities␈α∩of␈α∩intelligent␈α∩machines.␈α∪ However,␈α∩the␈α∩first␈α∩serious␈α∪work␈α∩on␈α∩AI␈α∩began␈α∪with␈α∩the
␈↓ ↓H␈↓invention␈αof␈αdigital␈α
computers␈αafter␈αWorld␈α
War␈αII.␈α Perhaps␈αthe␈α
first␈αserious␈αscientific␈α
article␈αon
␈↓ ↓H␈↓the␈α⊂subject␈α⊃was␈α⊂written␈α⊂by␈α⊃the␈α⊂British␈α⊂mathematician␈α⊃Alan␈α⊂Turing␈α⊂in␈α⊃1950␈α⊂and␈α⊃was␈α⊂entitled
␈↓ ↓H␈↓␈↓↓Computing␈αMachinery␈αand␈αIntelligence␈↓.␈α It␈αwas␈αin␈α1954␈αthat␈αthe␈αfirst␈αfull␈αtime␈αresearch␈αgroup␈αwas
␈↓ ↓H␈↓founded at Carnegie-Mellon University in Pittsburgh by Allen Newell and Herbert Simon.

␈↓ ↓H␈↓␈↓ α_Like␈α∞any␈α∂other␈α∞complex␈α∞problem,␈α∂artificial␈α∞intelligence␈α∞is␈α∂studied␈α∞by␈α∂finding␈α∞subproblems
␈↓ ↓H␈↓that␈α⊃can␈α∩be␈α⊃worked␈α∩on␈α⊃separately␈α⊃and␈α∩then␈α⊃combining␈α∩the␈α⊃solutions.␈α⊃ Here␈α∩are␈α⊃some␈α∩of␈α⊃the
␈↓ ↓H␈↓research areas that have been studied.

␈↓ ↓H␈↓␈↓αSearch␈↓

␈↓ ↓H␈↓␈↓ α_When␈α∞the␈α∞computer␈α∞has␈α
to␈α∞decide␈α∞on␈α∞an␈α
action,␈α∞it␈α∞may␈α∞have␈α
a␈α∞number␈α∞of␈α∞choices.␈α
 Each
␈↓ ↓H␈↓possible␈α
action␈α
may␈α
a␈α
number␈α
of␈α
different␈α
consequences␈α
depending␈α
on␈α
unknown␈α
facts␈α∞about␈α
the
␈↓ ↓H␈↓world,␈α⊗and␈α⊗each␈α⊗consequence␈α⊗may␈α⊗give␈α⊗new␈α⊗possible␈α⊗actions,␈α⊗and␈α⊗so␈α⊗on.␈α⊗ Sometimes␈α⊗the
␈↓ ↓H␈↓possibilities␈α
come␈αto␈α
an␈α
end␈αin␈α
a␈α
state␈αof␈α
affairs␈α
that␈αcan␈α
be␈α
conclusively␈αevaluated,␈α
but␈αusually␈α
the
␈↓ ↓H␈↓programmer␈α∪has␈α∪to␈α∪make␈α∀the␈α∪analysis␈α∪stop␈α∪short␈α∪of␈α∀a␈α∪conclusive␈α∪answer,␈α∪and␈α∀devise␈α∪some
␈↓ ↓H␈↓approximate way of evaluating the resulting situation.

␈↓ ↓H␈↓␈↓ α_The␈α
first␈α
experimental␈α
studies␈α
of␈α
searching␈αsuch␈α
␈↓↓trees␈α
of␈α
possibilities␈↓␈α
were␈α
made␈αusing␈α
games
␈↓ ↓H␈↓like␈αchess,␈αand␈αthe␈αperformance␈αof␈αchess␈αprograms␈αis␈αone␈αgood␈αmeasure␈αof␈αprogress␈αin␈αthe␈αsearch
␈↓ ↓H␈↓part␈αof␈αartificial␈αintelligence.␈α
 The␈α1950s␈αprograms␈αperformed␈α
much␈αworse␈αthan␈αexpected,␈αand␈α
one
␈↓ ↓H␈↓reason␈αbecause␈αis␈αunexpectedly␈αdifficulty␈αfor␈αa␈αchess␈αmaster␈αwho␈αis␈αalso␈αa␈αprogrammer␈αto␈αmake␈αa
␈↓ ↓H␈↓program␈α
that␈α
analyzes␈α
positions␈α
the␈αway␈α
he␈α
analyzes␈α
them␈α
himself.␈α When␈α
we␈α
try␈α
we␈α
miss␈αmany
␈↓ ↓H␈↓aspects of our own ways of thinking that seem obvious when someone finally points them out.

␈↓ ↓H␈↓␈↓ α_Thus␈α
one␈αchess␈α
master's␈α
program␈αlooked␈α
at␈α
the␈αseven␈α
most␈α
plausible␈αmoves␈α
and␈α
the␈αseven
␈↓ ↓H␈↓␈↓ εH␈↓ 91


␈↓ ↓H␈↓plausible␈αreplies␈αto␈αeach␈α
of␈αthese␈αand␈αseven␈αreplies␈α
to␈αeach␈αof␈αthese␈αand␈α
seven␈αto␈αeach␈αof␈αthese␈α
for
␈↓ ↓H␈↓a␈α∂total␈α∂of␈α∂7␈↓πx␈↓7␈↓πx␈↓7␈↓πx␈↓7␈α∂=␈α∂2401␈α∂final␈α∂positions.␈α∞ He␈α∂didn't␈α∂notice␈α∂that␈α∂most␈α∂of␈α∂these␈α∂positions␈α∞were
␈↓ ↓H␈↓examined␈α∂unnecessarily,␈α∂because␈α∂when␈α∂a␈α∂move␈α∂has␈α∂been␈α∂shown␈α∂to␈α∂be␈α∂worse␈α∂than␈α∂a␈α∂previously
␈↓ ↓H␈↓examined␈α∂move␈α⊂of␈α∂the␈α∂same␈α⊂player␈α∂by␈α∂finding␈α⊂one␈α∂reply␈α∂that␈α⊂refutes␈α∂it,␈α∂it␈α⊂is␈α∂not␈α⊂necessary␈α∂to
␈↓ ↓H␈↓examine␈α⊂other␈α∂replies.␈α⊂ This␈α⊂observation␈α∂(illustrated␈α⊂in␈α∂the␈α⊂figure)␈α⊂has␈α∂been␈α⊂elaborated␈α⊂into␈α∂a
␈↓ ↓H␈↓method␈α⊂called␈α⊂the␈α⊂alpha-beta␈α⊂heuristic␈α⊂for␈α⊂reducing␈α⊂search␈α⊂used␈α⊂by␈α⊂all␈α⊂modern␈α⊃game␈α⊂playing
␈↓ ↓H␈↓programs.␈α The␈α
joke␈αis␈α
that␈αit␈α
is␈αalso␈αused␈α
by␈αall␈α
human␈αplayers,␈α
modern␈αor␈α
not,␈αand␈αwhether␈α
they
␈↓ ↓H␈↓realize␈α∂it␈α∂or␈α∂not.␈α∂ Most␈α∂likely,␈α∂our␈α∂programs␈α∂today␈α∂lack␈α∂features␈α∂of␈α∂human␈α∂play␈α∂that␈α⊂will␈α∂seem
␈↓ ↓H␈↓equally obvious once pointed out.

␈↓ ↓H␈↓␈↓ α_In␈α⊃1967␈α⊃an␈α⊃M.I.T.␈α⊂chess␈α⊃program␈α⊃won␈α⊃the␈α⊂class␈α⊃D␈α⊃trophy␈α⊃in␈α⊂a␈α⊃tournament.␈α⊃In␈α⊃1977␈α⊂a
␈↓ ↓H␈↓Northwestern␈α⊃University␈α⊃program␈α⊃by␈α⊃Atkins␈α∩and␈α⊃Slate␈α⊃won␈α⊃the␈α⊃Minnesota␈α∩open␈α⊃tournament,
␈↓ ↓H␈↓achieving␈α
for␈α
itself␈α
a␈α
rating␈α
of␈α
Master,␈α
but␈α
was␈α
defeated␈α
in␈α
the␈α
subsequent␈α
invitational␈α
tournament
␈↓ ↓H␈↓for␈α⊗the␈α↔state␈α⊗championship.␈α↔ This␈α⊗improvement␈α↔is␈α⊗not␈α⊗as␈α↔good␈α⊗as␈α↔it␈α⊗looks,␈α↔because␈α⊗the
␈↓ ↓H␈↓Northwestern␈α
program␈αruns␈α
on␈α
an␈αextremely␈α
fast␈α
computer␈αand␈α
looks␈α
at␈αhundreds␈α
of␈αthousands␈α
of
␈↓ ↓H␈↓positions␈αin␈α
deciding␈αeach␈α
move.␈α Champion␈αlevel␈α
play␈αwithout␈α
excessive␈αcomputation␈αwill␈α
require
␈↓ ↓H␈↓new ideas.

␈↓ ↓H␈↓␈↓αGoals and subgoals␈↓

␈↓ ↓H␈↓␈↓ α_We␈α∞often␈α∞want␈α∞a␈α∞computer␈α∂to␈α∞find␈α∞a␈α∞sequence␈α∞of␈α∞actions␈α∂that␈α∞will␈α∞achieve␈α∞a␈α∞goal␈α∂on␈α∞the
␈↓ ↓H␈↓basis␈α∃of␈α∀information␈α∃about␈α∃the␈α∀effects␈α∃of␈α∀the␈α∃actions␈α∃and␈α∀the␈α∃preconditions␈α∃for␈α∀successful
␈↓ ↓H␈↓performance␈αof␈αactions.␈α A␈αsimple␈αexample␈αis␈αto␈αmake␈αa␈αtower␈αof␈αblocks␈αwhere␈αa␈αprecondition␈αfor
␈↓ ↓H␈↓moving␈α
one␈α
block␈α
on␈α
top␈α
of␈α
another␈α
is␈α
that␈αboth␈α
the␈α
block␈α
to␈α
be␈α
moved␈α
and␈α
the␈α
place␈α
where␈αit
␈↓ ↓H␈↓goes␈αshould␈αbe␈αclear,␈αwhich␈αmay␈αinvolve␈αmoving␈αother␈αblocks␈αfirst.␈α Much␈αrecent␈αAI␈αresearch␈αhas
␈↓ ↓H␈↓concerned how to express and use information about actions, their effects and preconditions.

␈↓ ↓H␈↓␈↓αReasoning and the representation of information␈↓

␈↓ ↓H␈↓␈↓ α_Many␈α∂of␈α∞the␈α∂difficulties␈α∞in␈α∂making␈α∂machines␈α∞perform␈α∂tasks␈α∞turn␈α∂out␈α∞to␈α∂be␈α∂difficulties␈α∞in
␈↓ ↓H␈↓deciding␈α
what␈α
information␈α
the␈αprogram␈α
should␈α
have,␈α
what␈αfurther␈α
conclusions␈α
can␈α
be␈αdrawn␈α
from
␈↓ ↓H␈↓initial␈αinformation␈αand␈αhow␈αthe␈αinformation␈αshould␈αbe␈αstored␈αin␈αthe␈αcomputer.␈α These␈αdifficulties
␈↓ ↓H␈↓have␈αled␈αto␈αresearch␈αon␈αthe␈αsubject␈αof␈αwhat␈αis␈αknowledge␈αasking␈αmany␈αof␈αthe␈αsame␈αquestions␈αthat
␈↓ ↓H␈↓philosophers␈α_have␈α_studied␈α_under␈α_the␈α→name␈α_of␈α_␈↓↓epistemology␈↓␈α_-␈α_the␈α_theory␈α→of␈α_knowledge.
␈↓ ↓H␈↓Mathematical␈αlogic␈αhas␈αprovided␈α
powerful␈αmeans␈αof␈αrepresenting␈α
facts␈αin␈αcomputers␈αand␈α
powerful
␈↓ ↓H␈↓modes␈α∞of␈α∞reasoning.␈α
 However,␈α∞it␈α∞has␈α∞turned␈α
out␈α∞that␈α∞not␈α∞all␈α
modes␈α∞of␈α∞reasoning␈α∞performed␈α
by
␈↓ ↓H␈↓humans␈αand␈αneeded␈αfor␈αproblem␈αsolving␈αare␈αrepresented␈αin␈αpresent␈αsystems␈αof␈αmathematical␈αlogic.
␈↓ ↓H␈↓Logic␈α
is␈αexcellent␈α
for␈αthe␈α
safe␈αmethods␈α
of␈α
reasoning␈αthat␈α
never␈αlead␈α
to␈αerror␈α
from␈α
true␈αpremises,
␈↓ ↓H␈↓but intelligence also requires methods of reasoning that only generate conjectures.

␈↓ ↓H␈↓␈↓ α_McCarthy␈α∩at␈α∩Stanford␈α∩University␈α∩has␈α∩studied␈α∩how␈α∩to␈α∩represent␈α∩facts␈α∩about␈α∪sources␈α∩of
␈↓ ↓H␈↓knowledge,␈αsuch␈αas␈αthe␈αfact␈αthat␈αtravel␈αagents␈αknow␈αairline␈αschedules␈αwhich␈αwould␈αbe␈αrequired␈α
by
␈↓ ↓H␈↓a␈α
program␈α
for␈α
planning␈α
trips.␈α
 It␈αhas␈α
proved␈α
especially␈α
difficult␈α
to␈α
represent␈α
facts␈αabout␈α
concurrent
␈↓ ↓H␈↓actions - one action starts while others contine.

␈↓ ↓H␈↓␈↓αObserving and acting in the physical world␈↓

␈↓ ↓H␈↓␈↓ α_Besides␈α
purely␈α
symbolic␈α
problems␈α
like␈α
chess␈α
and␈α
mathematics,␈α
an␈α
intelligent␈α
machine␈α
must
␈↓ ↓H␈↓␈↓ εH␈↓ 92


␈↓ ↓H␈↓be␈α
able␈α
to␈α
see␈αand␈α
hear␈α
and␈α
it␈α
must␈αbe␈α
able␈α
to␈α
control␈α
motion␈αand␈α
to␈α
make␈α
things.␈α Programs␈α
have
␈↓ ↓H␈↓been␈α∞written␈α∞that␈α∞find␈α∞objects␈α∞by␈α∞tracing␈α
outlines␈α∞of␈α∞color␈α∞or␈α∞brightness␈α∞changes␈α∞or␈α∞by␈α
growing
␈↓ ↓H␈↓regions␈α∞of␈α∞a␈α∞single␈α∞color␈α∞or␈α∞texture.␈α∞ Programs␈α∞that␈α∞recognize␈α∞and␈α∞produce␈α∞human␈α∞speech␈α
have
␈↓ ↓H␈↓been␈αwritten.␈α Some␈αof␈αthem␈αrecognize␈αhundreds␈αof␈αisolated␈αwords,␈αand␈αsome␈αof␈αthem␈αcan␈αhandle
␈↓ ↓H␈↓connected␈αspeech␈αif␈αit␈αis␈αnot␈αtoo␈αdifficult.␈α There␈αare␈αalso␈αprograms␈αthat␈αdrive␈αa␈αvehicle␈αavoiding
␈↓ ↓H␈↓obstacles and programs that assemble objects like automobile water pumps out of parts.

␈↓ ↓H␈↓␈↓ α_Industrial␈α∞robots␈α∞that␈α∂do␈α∞strenuous␈α∞or␈α∂dangerous␈α∞tasks␈α∞have␈α∞been␈α∂in␈α∞use␈α∞for␈α∂many␈α∞years.
␈↓ ↓H␈↓Almost␈αall␈αof␈αthem␈αare␈αprogrammed␈αto␈αrepeat␈αthe␈αsame␈αexact␈αsequence␈αof␈αactions␈αsuch␈αas␈αputting
␈↓ ↓H␈↓an␈α
object␈α
found␈α
in␈α
a␈α
fixed␈α
location␈α
in␈α
a␈α
punch␈α
press␈α
or␈α
a␈α
furnace␈α
and␈α
removing␈α
it␈α
later.␈α The␈α
only
␈↓ ↓H␈↓variability␈αin␈αtheir␈αprograms␈αis␈αthat␈αthey␈αcan␈αstop␈αif␈αan␈αexternal␈αdevice␈αdetects␈αsomething␈αwrong.
␈↓ ↓H␈↓Nevertheless,␈α
using␈α
such␈α
a␈α
robot␈αis␈α
more␈α
flexible␈α
than␈α
building␈αa␈α
special␈α
machine␈α
for␈α
each␈αsuch
␈↓ ↓H␈↓job␈α∀that␈α∀has␈α∀to␈α∀be␈α∀scrapped␈α∀when␈α∪the␈α∀production␈α∀task␈α∀changes.␈α∀ Industrial␈α∀robots␈α∀can␈α∪be
␈↓ ↓H␈↓reprogrammed for new tasks.

␈↓ ↓H␈↓␈↓ α_Industrial␈α
robots␈α∞that␈α
use␈α∞television␈α
cameras␈α
to␈α∞find␈α
objects␈α∞and␈α
guide␈α∞their␈α
manipulation
␈↓ ↓H␈↓have␈α∂come␈α∂recently␈α∂into␈α∂use␈α∂at␈α∂Hitachi␈α∞in␈α∂Japan␈α∂and␈α∂at␈α∂General␈α∂Motors.␈α∂ Thomas␈α∂Binford␈α∞at
␈↓ ↓H␈↓Stanford␈α∞University␈α∞has␈α∞developed␈α∞a␈α∞language␈α∞called␈α∞AL␈α∞for␈α∞writing␈α∞the␈α∞programs␈α∞that␈α
control
␈↓ ↓H␈↓industrial robots.

␈↓ ↓H␈↓␈↓αPattern matching␈↓

␈↓ ↓H␈↓␈↓ α_All␈α∂these␈α∂tasks␈α∂require␈α∂computers␈α∂to␈α∂store␈α∂patterns␈α∂and␈α∂to␈α∂recognize␈α∂objects␈α∂in␈α∂scenes␈α∂by
␈↓ ↓H␈↓matching␈αthe␈αpattern.␈α For␈αexample,␈αa␈αcomputer␈αcan␈αstore␈αits␈αidea␈αof␈αa␈αdog␈αas␈αa␈αcollection␈αof␈αlegs,
␈↓ ↓H␈↓tail␈αetc.,␈αeach␈αof␈αa␈αgiven␈αshape␈αand␈αconnected␈αin␈αthe␈αright␈αway.␈α Then␈αit␈αcan␈αtry␈αto␈αfind␈αdogs␈αin␈αa
␈↓ ↓H␈↓picture␈αby␈α
matching␈αthe␈α
dog␈αpattern␈α
with␈αparts␈αof␈α
the␈αpicture.␈α
 The␈αprogram␈α
must␈αcompute␈αhow␈α
a
␈↓ ↓H␈↓dog␈αwill␈αlook␈αwith␈αits␈αlegs␈α
in␈αsome␈αposition␈αlooked␈αat␈αfrom␈α
some␈αangle␈αand␈αpartly␈αhidden␈αby␈α
other
␈↓ ↓H␈↓objects.␈α
 Actually,␈α
in␈α
order␈α
to␈αrecognize␈α
dogs␈α
it␈α
must␈α
do␈αthis␈α
backwards,␈α
deciding␈α
what␈α
kind␈αof␈α
dog
␈↓ ↓H␈↓would lead to the appearance it sees.

␈↓ ↓H␈↓␈↓ α_General␈α⊃principles␈α⊃of␈α⊃pattern␈α⊃description␈α⊂and␈α⊃techniques␈α⊃for␈α⊃pattern␈α⊃matching␈α⊃apply␈α⊂to
␈↓ ↓H␈↓recognizing dogs and recognizing chess positions that permit attacks on the opponent's king.

␈↓ ↓H␈↓␈↓ α_Minsky␈αand␈αhis␈αstudents␈αat␈αMassachusetts␈αInstitute␈αof␈αTechnology␈αhave␈αstudied␈αpatterns␈αof
␈↓ ↓H␈↓actions␈αthat␈αthey␈αcall␈αframes.␈α A␈αtypical␈αframe␈αis␈αthe␈αevent␈αof␈αvisiting␈αa␈αrestaurant␈αwhich␈αcontains
␈↓ ↓H␈↓subframes␈αof␈αbeing␈αseated,␈αordering,␈αwaiting,␈αeating,␈αpaying␈αthe␈αbill␈αand␈αleaving.␈α Schank␈αat␈αYale
␈↓ ↓H␈↓has␈αused␈αsuch␈αframes␈αin␈αprograms␈αthat␈αanswer␈αquestions␈αabout␈αstories␈αand␈αcan␈αfill␈αin␈αinformation
␈↓ ↓H␈↓omitted from the story, because it is implicit in the frame.

␈↓ ↓H␈↓␈↓αUnderstanding English␈↓

␈↓ ↓H␈↓␈↓ α_One␈α∞idea␈α∞for␈α∞making␈α∞an␈α∞intelligent␈α∂machine␈α∞is␈α∞to␈α∞first␈α∞make␈α∞it␈α∞capable␈α∂of␈α∞understanding
␈↓ ↓H␈↓English␈αand␈αthen␈αletting␈αit␈αread␈α
textbooks,␈αencyclopedias␈αand␈αscientific␈αarticles.␈α At␈α
present,␈αthere
␈↓ ↓H␈↓are␈αcomputer␈αprograms␈αthat␈αcan␈αread␈αstories␈αfrom␈αfirst␈αgrade␈αreaders␈αand␈αanswer␈α
some␈αquestions
␈↓ ↓H␈↓about␈αthem.␈α Other␈αprograms␈αcan␈αconverse␈αwith␈αphysicians␈αabout␈αbacterial␈αdiseases␈αof␈α
the␈αblood,
␈↓ ↓H␈↓but␈α∩all␈α∩present␈α⊃programs␈α∩are␈α∩quite␈α∩limited␈α⊃in␈α∩the␈α∩subject␈α⊃matter␈α∩they␈α∩can␈α∩understand.␈α⊃ The
␈↓ ↓H␈↓difficulty␈αis␈αnot␈αone␈αof␈αvocabulary,␈αbecause␈αit␈αis␈αeasy␈αto␈αread␈αa␈αwhole␈αdictionary␈αinto␈αa␈αcomputer.
␈↓ ↓H␈↓The␈αproblem␈αis␈αthat␈αunderstanding␈αrequires␈αsome␈αinitial␈αknowledge,␈αand␈αwe␈αhaven't␈αbeen␈αable␈αto
␈↓ ↓H␈↓give computers that initial knowledge in a sufficiently general way.
␈↓ ↓H␈↓␈↓ εH␈↓ 93


␈↓ ↓H␈↓␈↓ α_In␈α∩the␈α⊃1950s␈α∩programs␈α∩were␈α⊃written␈α∩to␈α⊃translate␈α∩from␈α∩one␈α⊃language␈α∩to␈α∩another.␈α⊃ They
␈↓ ↓H␈↓weren't␈α⊂very␈α⊂good,␈α⊂and␈α⊂it␈α⊂was␈α⊂some␈α⊂years␈α⊂before␈α⊂it␈α⊂was␈α⊂understood␈α⊂that␈α⊂successful␈α∂translation
␈↓ ↓H␈↓requires␈α
that␈α
the␈α
program␈α
understand␈α
the␈α∞material␈α
being␈α
translated.␈α
 The␈α
effort␈α
that␈α
used␈α∞to␈α
go
␈↓ ↓H␈↓into␈α
automatic␈α
translation␈α
is␈αnow␈α
going␈α
into␈α
making␈αcomputers␈α
understand␈α
better␈α
and␈αbetter␈α
larger
␈↓ ↓H␈↓and␈αlarger␈αfragments␈α
of␈αnatural␈αlanguage.␈α
 This␈α"understanding"␈αis␈α
tested␈αby␈αthe␈α
performance␈αof
␈↓ ↓H␈↓␈↓↓question␈α
answering␈α
programs␈↓␈αthat␈α
answer␈α
questions␈α
about␈αa␈α
text␈α
on␈α
the␈αbasis␈α
of␈α
information␈αin␈α
the
␈↓ ↓H␈↓text and the common sense knowledge posessed by the program.

␈↓ ↓H␈↓␈↓ α_Terry␈α⊃Winograd's␈α⊂program␈α⊃SHRDLU␈α⊂combined␈α⊃English␈α⊂dialog␈α⊃and␈α⊂problem␈α⊃solving␈α⊂to
␈↓ ↓H␈↓answer␈α
questions␈α
and␈α
perform␈α
requested␈α
actions␈α
in␈α
a␈α
simulated␈α
"blocks␈α
world".␈α
 SHRDLU␈αcould
␈↓ ↓H␈↓be␈αrequested,␈α"Pick␈α
up␈αthe␈αred␈α
pyramid␈αand␈αput␈α
it␈αon␈αthe␈α
green␈αblock".␈α SHRDLU␈α
would␈αfigure
␈↓ ↓H␈↓out␈αthat␈α"it"␈αreferred␈αto␈αthe␈αred␈αpyramid,␈αand␈αit␈αwould␈αclear␈αoff␈αthe␈αgreen␈αblock␈αif␈αnecessary,␈αand
␈↓ ↓H␈↓would ask "Which green block" if there was more than one.

␈↓ ↓H␈↓␈↓αLearning from experience␈↓

␈↓ ↓H␈↓␈↓ α_Humans␈αand␈αanimals␈αlearn␈α
from␈αtheir␈αexperience␈αhow␈αto␈α
do␈αbetter␈αnext␈αtime,␈αand␈α
machines
␈↓ ↓H␈↓should␈α
do␈α
the␈α
same.␈α∞ An␈α
early␈α
example␈α
was␈α
a␈α∞program␈α
by␈α
Arthur␈α
Samuel␈α
for␈α∞playing␈α
checkers.
␈↓ ↓H␈↓The␈α∩behavior␈α∩of␈α∪the␈α∩program␈α∩was␈α∪determined␈α∩by␈α∩a␈α∩some␈α∪numbers␈α∩that␈α∩determined␈α∪how␈α∩it
␈↓ ↓H␈↓evaluated␈α∞positions,␈α
for␈α∞example␈α∞the␈α
relative␈α∞values␈α
of␈α∞a␈α∞king␈α
and␈α∞a␈α
single␈α∞man.␈α∞ The␈α
program
␈↓ ↓H␈↓would␈αread␈αbooks␈αof␈αgames␈αplayed␈αby␈αmaster␈αplayers␈αand␈αadjust␈αthe␈αnumbers␈αuntil␈αthey␈αpredicted
␈↓ ↓H␈↓as␈α∞many␈α
as␈α∞possible␈α∞of␈α
the␈α∞moves␈α
regarded␈α∞as␈α∞good␈α
by␈α∞the␈α
masters.␈α∞ Combined␈α∞with␈α
lookahead
␈↓ ↓H␈↓this␈α∩made␈α∩an␈α∩excellent␈α⊃program.␈α∩ A␈α∩version␈α∩of␈α⊃Samuel's␈α∩program,␈α∩without␈α∩the␈α∩learning,␈α⊃has
␈↓ ↓H␈↓recently been marketed for playing through a TV set.

␈↓ ↓H␈↓␈↓ α_The␈α∪ability␈α∀of␈α∪a␈α∪program␈α∀to␈α∪learn␈α∪from␈α∀experience␈α∪depends␈α∪on␈α∀how␈α∪its␈α∀behavior␈α∪is
␈↓ ↓H␈↓represented␈α⊃within␈α⊃the␈α∩machine.␈α⊃ If␈α⊃it␈α⊃is␈α∩to␈α⊃learn␈α⊃efficiently,␈α⊃then␈α∩what␈α⊃we␈α⊃regard␈α∩as␈α⊃simple
␈↓ ↓H␈↓changes␈α⊗in␈α⊗behavior␈α⊗must␈α⊗be␈α⊗represented␈α⊗by␈α⊗small␈α⊗changes␈α⊗in␈α⊗the␈α⊗way␈α⊗the␈α↔behavior␈α⊗is
␈↓ ↓H␈↓represeented.␈α∩ With␈α⊃present␈α∩computer␈α∩programs,␈α⊃this␈α∩isn't␈α⊃usually␈α∩true;␈α∩in␈α⊃order␈α∩to␈α∩change␈α⊃a
␈↓ ↓H␈↓program's␈α
behavior,␈α
one␈α
must␈α
understand␈α
the␈α
program␈α
thoroughly␈α
and␈α
make␈α
changes␈α
in␈α
a␈α
number
␈↓ ↓H␈↓of places.

␈↓ ↓H␈↓␈↓ α_This␈αmust␈αbe␈αcontrasted␈αwith␈αteaching␈αa␈αhuman.␈α If␈αyou␈αwant␈αto␈αteach␈αsomeone␈αto␈αplay␈αtic-
␈↓ ↓H␈↓tac-toe␈α∂or␈α∂solve␈α∞an␈α∂algebra␈α∂problem,␈α∞you␈α∂can␈α∂tell␈α∂him␈α∞the␈α∂rules␈α∂and␈α∞rely␈α∂on␈α∂his␈α∂intelligence␈α∞to
␈↓ ↓H␈↓apply them.  With computers it as though all education were by brain surgery.

␈↓ ↓H␈↓␈↓ α_One␈α∩of␈α∩the␈α∩research␈α∩goals␈α∩of␈α∩artificial␈α∩intelligence␈α∩is␈α∩to␈α∩to␈α∩change␈α∩this␈α∪by␈α∩developing
␈↓ ↓H␈↓programs with ␈↓↓common sense␈↓ that can take instructions and combine it with their own knowledge.

␈↓ ↓H␈↓␈↓ α_Samuel's␈α
program␈αis␈α
an␈αexample␈α
of␈αlearning␈α
to␈αdo␈α
something.␈α Learning␈α
to␈αrecognize␈α
objects
␈↓ ↓H␈↓has␈αbeen␈αeasier.␈α Early␈αpattern␈αrecognition␈αprograms␈αlearned␈αto␈αclassify␈αobjects␈αsuch␈αas␈αimages␈αof
␈↓ ↓H␈↓letter␈α
of␈α
the␈α∞alphabet␈α
into␈α
categories,␈α
and␈α∞more␈α
recent␈α
programs,␈α
such␈α∞as␈α
one␈α
written␈α∞by␈α
Patrick
␈↓ ↓H␈↓Winston␈α∞learn␈α∞to␈α
classify␈α∞objects␈α∞according␈α∞the␈α
the␈α∞presence␈α∞of␈α
subobjects␈α∞related␈α∞in␈α∞a␈α
specified
␈↓ ↓H␈↓way.␈α∂ Winston's␈α∂example␈α∂is␈α∂recognizing␈α∂an␈α∂arch␈α∂as␈α∂consisting␈α∂of␈α∂three␈α∂objects␈α∂one␈α∂of␈α⊂which␈α∂is
␈↓ ↓H␈↓supported␈αby␈αthe␈αother␈αtwo␈αwhich␈αare␈αnot␈αtouching␈αeach␈αother.␈α An␈αarcade␈αis␈αlearned␈αas␈αa␈αline␈αof
␈↓ ↓H␈↓arches arranged so that there is a path under all of them.

␈↓ ↓H␈↓␈↓αExpert programs␈↓
␈↓ ↓H␈↓␈↓ εH␈↓ 94


␈↓ ↓H␈↓␈↓ α_Feigenbaum␈α∀and␈α∀Lederberg␈α∪at␈α∀Stanford␈α∀pioneered␈α∪the␈α∀development␈α∀of␈α∀programs␈α∪that
␈↓ ↓H␈↓embody␈αthe␈αknowledge␈αof␈αan␈αexpert␈αin␈αsome␈αfield.␈α Such␈αprograms␈αare␈αdeveloped␈αby␈αinterviewing
␈↓ ↓H␈↓experts␈αand␈αgetting␈αthem␈αto␈αhelp␈αimprove␈αfurther␈αversions␈αof␈αthe␈αprogram.␈α DENDRAL␈αis␈αexpert
␈↓ ↓H␈↓in␈α
determining␈α
the␈α
structure␈α
of␈α
an␈α
organic␈α
compound␈α
from␈α
mass␈α
spectrograph␈α
observations,␈αand
␈↓ ↓H␈↓MYCIN␈α∪helps␈α∪a␈α∀doctor␈α∪diagnose␈α∪bacterial␈α∪infections␈α∀of␈α∪the␈α∪blood,␈α∪recommending␈α∀tests␈α∪and
␈↓ ↓H␈↓treatment␈α∀and␈α∀recommending␈α∀further␈α∪tests␈α∀and␈α∀treatment␈α∀based␈α∪on␈α∀the␈α∀results␈α∀of␈α∀the␈α∪first.
␈↓ ↓H␈↓MYCIN␈α
is␈α
intended␈α
only␈αto␈α
make␈α
suggestions,␈α
and␈α
the␈αdoctor␈α
still␈α
must␈α
understand␈α
the␈αreasons␈α
for
␈↓ ↓H␈↓everything he does.

␈↓ ↓H␈↓␈↓ α_The domain of each of these expert programs is very narrow.

␈↓ ↓H␈↓␈↓αInformation processing psychology␈↓

␈↓ ↓H␈↓␈↓ α_The␈α∞development␈α∞of␈α∞artificial␈α∞intelligence␈α∞and␈α∞the␈α∞study␈α∞of␈α∞human␈α∞intelligence␈α∞are␈α
closely
␈↓ ↓H␈↓related,␈αand␈αthe␈αinformation␈αprocessing␈αapproach␈αto␈αpsychology␈αhas␈αmainly␈αreplaced␈αbehaviorism
␈↓ ↓H␈↓which␈α⊂was␈α⊂concerned␈α⊂with␈α∂finding␈α⊂direct␈α⊂relations␈α⊂between␈α∂stimuli␈α⊂received␈α⊂by␈α⊂organisms␈α∂and
␈↓ ↓H␈↓their␈α⊂responses.␈α∂ The␈α⊂information␈α⊂processing␈α∂approach,␈α⊂initiated␈α⊂by␈α∂Allen␈α⊂Newell␈α⊂and␈α∂Herbert
␈↓ ↓H␈↓Simon␈α∂in␈α⊂the␈α∂1950s,␈α⊂writes␈α∂computer␈α∂programs␈α⊂that␈α∂match␈α⊂complex␈α∂problem␈α⊂solving␈α∂behavior.
␈↓ ↓H␈↓Unlike␈αstimulus-response␈α
theories,␈αinformation␈α
processing␈αtheories␈α
do␈αnot␈α
postulate␈αdirect␈α
relations
␈↓ ↓H␈↓between␈α
inputs␈α
and␈α
outputs,␈α
because␈α
the␈α
internal␈α
processes␈α
can␈α
be␈α
very␈α
complex.␈α
 Both␈α
artificial
␈↓ ↓H␈↓intelligence␈α
and␈α∞information␈α
processing␈α∞psychology␈α
must␈α∞determine␈α
what␈α∞intellectual␈α
mechanisms
␈↓ ↓H␈↓are required to solve different kinds of problems.

␈↓ ↓H␈↓␈↓αStudying artificial intelligence␈↓

␈↓ ↓H␈↓␈↓ α_Artificial␈α⊃intelligence␈α⊃is␈α⊃a␈α⊃young␈α⊃and␈α⊂difficult␈α⊃branch␈α⊃of␈α⊃computer␈α⊃science.␈α⊃ Studying␈α⊂it
␈↓ ↓H␈↓requires␈αthe␈αability␈αto␈αprogram␈αa␈αcomputer␈αespecially␈αusing␈αprogramming␈αlanguages␈αsuch␈αas␈αLISP
␈↓ ↓H␈↓popular␈αin␈αAI␈αresearch.␈α Other␈αbranches␈αof␈α
computer␈αscience␈αsuch␈αthe␈αtheory␈αof␈αthe␈αcorrectness␈α
of
␈↓ ↓H␈↓computer␈αprograms␈αand␈αcomplexity␈αtheory␈αthat␈αstudies␈αhow␈αmuch␈αcomputing␈αis␈αrequired␈αto␈αsolve
␈↓ ↓H␈↓various kinds of problems are also important.

␈↓ ↓H␈↓␈↓ α_There␈α
is␈αless␈α
to␈αlearn␈α
than␈αin␈α
physics␈αor␈α
mathematics␈αin␈α
order␈αto␈α
reach␈αthe␈α
frontier␈α
of␈αthe
␈↓ ↓H␈↓subject.␈α Indeed␈αmuch␈α
of␈αwhat␈αthe␈αstudent␈α
does␈αlearn␈αis␈αcontroversial␈α
and␈αsome␈αof␈αit␈α
is␈αprobably
␈↓ ↓H␈↓wrong.

␈↓ ↓H␈↓␈↓ α_Besides␈αconnections␈αwith␈αpsychology,␈α
artificial␈αintelligence␈αneeds␈αfacts␈αfrom␈α
and␈αcontributes
␈↓ ↓H␈↓to␈α⊂mathematical␈α∂logic,␈α⊂linguistics␈α∂and␈α⊂the␈α∂physiology␈α⊂of␈α∂the␈α⊂nervous␈α∂system.␈α⊂ Finally,␈α⊂it␈α∂studies
␈↓ ↓H␈↓many questions that are also studied by philosophers from a different point of view.

␈↓ ↓H␈↓␈↓ α_All␈α
present␈α
artificial␈α
intelligence␈αprograms␈α
lack␈α
any␈α
general␈α
view␈αof␈α
the␈α
world␈α
and␈αof␈α
actions
␈↓ ↓H␈↓and␈αevents.␈α Until␈αa␈αbreakthrough␈αin␈αthis␈αarea␈αoccurs,␈αhuman-level␈αintelligence␈αexcept␈α
in␈αnarrow
␈↓ ↓H␈↓domains is far off.

␈↓ ↓H␈↓␈↓αReferences␈↓

␈↓ ↓H␈↓␈↓↓Artificial␈α⊂Intelligence␈α⊂and␈α⊂Natural␈α⊂Man␈↓␈α⊂by␈α∂Margaret␈α⊂Boden␈α⊂is␈α⊂critical␈α⊂and␈α⊂detailed␈α⊂but␈α∂non-
␈↓ ↓H␈↓technical.

␈↓ ↓H␈↓␈↓↓Artificial Intelligence␈↓ by Patrick Winston is the most popular textbook as of the middle of 1978.
␈↓ ↓H␈↓␈↓ εH␈↓ 95


␈↓ ↓H␈↓␈↓↓Problem Solving Methods in Artificial Intelligence␈↓ by Nils Nilsson is an older textbook.

␈↓ ↓H␈↓␈↓↓Human␈α
Problem␈α
Solving␈↓␈α∞by␈α
Allen␈α
Newell␈α
and␈α∞Herbert␈α
Simon␈α
connects␈α
artificial␈α∞intelligence␈α
and
␈↓ ↓H␈↓psychology.
␈↓ ↓H␈↓␈↓ εH␈↓ 96


␈↓ ↓H␈↓search,␈α
pattern␈α
matching,␈α
learning,␈α
representation␈α
What␈αman␈α
can␈α
understand␈α
about␈α
how␈α
man␈αor
␈↓ ↓H␈↓machine␈α∃can␈α⊗solve␈α∃problems.␈α∃ Torres␈α⊗y␈α∃Quevedo␈α∃Turing␈α⊗Chess␈α∃Generality␈α⊗Heuristics␈α∃and
␈↓ ↓H␈↓epistemology Plenty for the ambitious lad to do.  What to read and what to learn.

␈↓ ↓H␈↓␈↓ α_computer␈α
programming␈α
including␈α
LISP,␈αmathematical␈α
logic,␈α
mathematics␈α
in␈αgeneral,␈α
analytic
␈↓ ↓H␈↓philosophy,␈αStanford,␈α
M.I.T.␈αand␈αCMU.,␈α
Winston,␈αNilsson␈α
AI␈αand␈αthe␈α
future␈αof␈αman.␈α
 comparison
␈↓ ↓H␈↓with science fiction

␈↓ ↓H␈↓1. Introduction

␈↓ ↓H␈↓␈↓ α_a. definition

␈↓ ↓H␈↓␈↓ α_b. importance and difficulty

␈↓ ↓H␈↓␈↓ α_c. computers

␈↓ ↓H␈↓␈↓ α_d. chess 4.6 and the famous match

␈↓ ↓H␈↓␈↓ α_2. History and present state

␈↓ ↓H␈↓␈↓ α_3. Problems and methods

␈↓ ↓H␈↓␈↓ α_a. epistemology and heuristics

␈↓ ↓H␈↓␈↓ α_b. patterns and pattern matching

␈↓ ↓H␈↓␈↓ α_c. heuristic search

␈↓ ↓H␈↓␈↓ α_d. problem solving, e.g. finding a sequence of actions

␈↓ ↓H␈↓␈↓ α_e. the problem of generality

␈↓ ↓H␈↓␈↓ α_f. AI languages

␈↓ ↓H␈↓␈↓ α_g. natural language

␈↓ ↓H␈↓␈↓ α_h. learning

␈↓ ↓H␈↓␈↓ α_i. representation

␈↓ ↓H␈↓␈↓ α_j. physiology

␈↓ ↓H␈↓␈↓ α_4. AI and the future

␈↓ ↓H␈↓␈↓ α_a. AI in science fiction

␈↓ ↓H␈↓␈↓ α_b. the prospects

␈↓ ↓H␈↓␈↓ α_c. the problem for humanity
␈↓ ↓H␈↓␈↓ εH␈↓ 97


␈↓ ↓H␈↓␈↓ α_5. Literature

␈↓ ↓H␈↓␈↓ α_Pictures and diagrams

␈↓ ↓H␈↓␈↓ α_a. a scene and representations of information about it