Artificial intelligence (AI) is the intelligence of machines and the branch of computer

science that aims to create it. AI textbooks define the field as "the study and design of intelligent
agents"
[1]
here an intelligent agent is a system that percei!es its en!ironment and takes actions
that maximi"e its chances of success.
[#]
$ohn %c&arthy' ho coined the term in 1()*'
[+]
defines it
as "the science and engineering of making intelligent machines."
[,]
-he field as founded on the claim that a central property of humans' intelligence.the sapience
of Homo sapiens.can be so precisely described that it can be simulated by a machine.
[)]
-his
raises philosophical issues about the nature of the mind and limits of scientific hubris' issues
hich ha!e been addressed by myth' fiction and philosophy since anti/uity.
[*]
Artificial
intelligence has been the sub0ect of optimism'
[1]
but has also suffered setbacks
[2]
and' today' has
become an essential part of the technology industry' pro!iding the hea!y lifting for many of the
most difficult problems in computer science.
[(]
AI research is highly technical and speciali"ed' deeply di!ided into subfields that often fail to
communicate ith each other.
[13]
4ubfields ha!e gron up around particular institutions' the
ork of indi!idual researchers' the solution of specific problems' longstanding differences of
opinion about ho AI should be done and the application of idely differing tools. -he central
problems of AI include such traits as reasoning' knoledge' planning' learning' communication'
perception and the ability to mo!e and manipulate ob0ects.
[11]
5eneral intelligence (or "strong
AI") is still among the field6s long term goals.
[1#]
History
-hinking machines and artificial beings appear in 5reek myths' such as -alos of &rete' the
golden robots of 7ephaestus and 8ygmalion6s 5alatea.
[1+]
7uman likenesses belie!ed to ha!e
intelligence ere built in e!ery ma0or ci!ili"ation9 animated statues ere seen in :gypt and
5reece
[1,]
and humanoid automatons ere built by ;an 4hi'
[1)]
7ero of Alexandria'
[1*]
Al<$a"ari
[11]
and =olfgang !on >empelen.
[12]
It as also idely belie!ed that artificial beings had been
created by $?bir ibn 7ayy?n'
[1(]
$udah @oe
[#3]
and 8aracelsus.
[#1]
Ay the 1(th and #3th centuries'
artificial beings had become a common feature in fiction' as in %ary 4helley6s Frankenstein or
>arel Bapek6s R.U.R. (Rossum's Universal Robots).
[##]
8amela %c&orduck argues that all of
these are examples of an ancient urge' as she describes it' "to forge the gods".
[*]
4tories of these
creatures and their fates discuss many of the same hopes' fears and ethical concerns that are
presented by artificial intelligence.
%echanical or "formal" reasoning has been de!eloped by philosophers and mathematicians since
anti/uity. -he study of logic led directly to the in!ention of the programmable digital electronic
computer' based on the ork of mathematician Alan -uring and others. -uring6s theory of
computation suggested that a machine' by shuffling symbols as simple as "3" and "1"' could
simulate any concei!able act of mathematical deduction.
[#+]
-his' along ith recent disco!eries in
neurology' information theory and cybernetics' inspired a small group of researchers to begin to
seriously consider the possibility of building an electronic brain.
[#,]
-he field of AI research as founded at a conference on the campus of Cartmouth &ollege in the
summer of 1()*.
[#)]
-he attendees' including $ohn %c&arthy' %ar!in %insky' Allen Deell and
7erbert 4imon' became the leaders of AI research for many decades.
[#*]
-hey and their students
rote programs that ere' to most people' simply astonishing9
[#1]
computers ere sol!ing ord
problems in algebra' pro!ing logical theorems and speaking :nglish.
[#2]
Ay the middle of the
1(*3s' research in the E.4. as hea!ily funded by the Cepartment of Cefense
[#(]
and laboratories
had been established around the orld.
[+3]
AI6s founders ere profoundly optimistic about the
future of the ne field9 7erbert 4imon predicted that "machines ill be capable' ithin tenty
years' of doing any ork a man can do"
[+1]
and %ar!in %insky agreed' riting that "ithin a
generation ... the problem of creating 6artificial intelligence6 ill substantially be sol!ed".
[+#]
-hey had failed to recogni"e the difficulty of some of the problems they faced.
[++]
In 1(1,' in
response to the criticism of :ngland6s 4ir $ames @ighthill and ongoing pressure from &ongress to
fund more producti!e pro0ects' the E.4. and Aritish go!ernments cut off all undirected'
exploratory research in AI. -he next fe years' hen funding for pro0ects as hard to find'
ould later be called an "AI inter".
[+,]
In the early 1(23s' AI research as re!i!ed by the commercial success of expert systems'
[+)]
a
form of AI program that simulated the knoledge and analytical skills of one or more human
experts. Ay 1(2) the market for AI had reached o!er a billion dollars. At the same time' $apan6s
fifth generation computer pro0ect inspired the E.4 and Aritish go!ernments to restore funding for
academic research in the field.
[+*]
7oe!er' beginning ith the collapse of the @isp %achine
market in 1(21' AI once again fell into disrepute' and a second' longer lasting AI inter began.
[+1]
In the 1((3s and early #1st century' AI achie!ed its greatest successes' albeit somehat behind
the scenes. Artificial intelligence is used for logistics' data mining' medical diagnosis and many
other areas throughout the technology industry.
[(]
-he success as due to se!eral factors9 the
incredible poer of computers today (see %oore6s la)' a greater emphasis on sol!ing specific
subproblems' the creation of ne ties beteen AI and other fields orking on similar problems'
and abo!e all a ne commitment by researchers to solid mathematical methods and rigorous
scientific standards.
[+2]
Problems
-he general problem of simulating (or creating) intelligence has been broken don into a
number of specific sub<problems. -hese consist of particular traits or capabilities that researchers
ould like an intelligent system to display. -he traits described belo ha!e recei!ed the most
attention.
[11]
Deduction, reasoning, problem solving
:arly AI researchers de!eloped algorithms that imitated the step<by<step reasoning that humans
ere often assumed to use hen they sol!e pu""les' play board games or make logical
deductions.
[+(]
Ay the late 1(23s and 6(3s' AI research had also de!eloped highly successful
methods for dealing ith uncertain or incomplete information' employing concepts from
probability and economics.
[,3]
For difficult problems' most of these algorithms can re/uire enormous computational resources
. most experience a "combinatorial explosion"9 the amount of memory or computer time
re/uired becomes astronomical hen the problem goes beyond a certain si"e. -he search for
more efficient problem sol!ing algorithms is a high priority for AI research.
[,1]
7uman beings sol!e most of their problems using fast' intuiti!e 0udgments rather than the
conscious' step<by<step deduction that early AI research as able to model.
[,#]
AI has made some
progress at imitating this kind of "sub<symbolic" problem sol!ing9 embodied agent approaches
emphasi"e the importance of sensorimotor skills to higher reasoningG neural net research
attempts to simulate the structures inside human and animal brains that gi!e rise to this skill.
Knowledge representation
>noledge representation
[,+]
and knoledge engineering
[,,]
are central to AI research. %any of
the problems machines are expected to sol!e ill re/uire extensi!e knoledge about the orld.
Among the things that AI needs to represent are9 ob0ects' properties' categories and relations
beteen ob0ectsG
[,)]
situations' e!ents' states and timeG
[,*]
causes and effectsG
[,1]
knoledge about
knoledge (hat e kno about hat other people kno)G
[,2]
and many other' less ell
researched domains. A complete representation of "hat exists" is an ontology
[,(]
(borroing a
ord from traditional philosophy)' of hich the most general are called upper ontologies.
Among the most difficult problems in knoledge representation are9
Default reasoning and the qualification problem
%any of the things people kno take the form of "orking assumptions." For example' if
a bird comes up in con!ersation' people typically picture an animal that is fist si"ed'
sings' and flies. Done of these things are true about all birds. $ohn %c&arthy identified
this problem in 1(*(
[)3]
as the /ualification problem9 for any commonsense rule that AI
researchers care to represent' there tend to be a huge number of exceptions. Almost
nothing is simply true or false in the ay that abstract logic re/uires. AI research has
explored a number of solutions to this problem.
[)1]
The breadth of commonsense nowledge
-he number of atomic facts that the a!erage person knos is astronomical. Hesearch
pro0ects that attempt to build a complete knoledge base of commonsense knoledge
(e.g.' &yc) re/uire enormous amounts of laborious ontological engineering . they must
be built' by hand' one complicated concept at a time.
[)#]
A ma0or goal is to ha!e the
computer understand enough concepts to be able to learn by reading from sources like the
internet' and thus be able to add to its on ontology.
The subsymbolic form of some commonsense nowledge
%uch of hat people kno is not represented as "facts" or "statements" that they could
actually say out loud. For example' a chess master ill a!oid a particular chess position
because it "feels too exposed"
[)+]
or an art critic can take one look at a statue and instantly
reali"e that it is a fake.
[),]
-hese are intuitions or tendencies that are represented in the
brain non<consciously and sub<symbolically.
[))]
>noledge like this informs' supports
and pro!ides a context for symbolic' conscious knoledge. As ith the related problem
of sub<symbolic reasoning' it is hoped that situated AI or computational intelligence ill
pro!ide ays to represent this kind of knoledge.
[))]
Planning
Intelligent agents must be able to set goals and achie!e them.
[)*]
-hey need a ay to !isuali"e the
future (they must ha!e a representation of the state of the orld and be able to make predictions
about ho their actions ill change it) and be able to make choices that maximi"e the utility (or
"!alue") of the a!ailable choices.
[)1]
In classical planning problems' the agent can assume that it is the only thing acting on the orld
and it can be certain hat the conse/uences of its actions may be.
[)2]
7oe!er' if this is not true'
it must periodically check if the orld matches its predictions and it must change its plan as this
becomes necessary' re/uiring the agent to reason under uncertainty.
[)(]
%ulti<agent planning uses the cooperation and competition of many agents to achie!e a gi!en
goal. :mergent beha!ior such as this is used by e!olutionary algorithms and sarm intelligence.
[*3]
!earning
%achine learning
[*1]
has been central to AI research from the beginning.
[*#]
Ensuper!ised learning
is the ability to find patterns in a stream of input. 4uper!ised learning includes both classification
and numerical regression. &lassification is used to determine hat category something belongs
in' after seeing a number of examples of things from se!eral categories. Hegression takes a set of
numerical inputIoutput examples and attempts to disco!er a continuous function that ould
generate the outputs from the inputs. In reinforcement learning
[*+]
the agent is rearded for good
responses and punished for bad ones. -hese can be analy"ed in terms of decision theory' using
concepts like utility. -he mathematical analysis of machine learning algorithms and their
performance is a branch of theoretical computer science knon as computational learning
theory.
"atural language processing
Datural language processing
[*,]
gi!es machines the ability to read and understand the languages
that humans speak. %any researchers hope that a sufficiently poerful natural language
processing system ould be able to ac/uire knoledge on its on' by reading the existing text
a!ailable o!er the internet. 4ome straightforard applications of natural language processing
include information retrie!al (or text mining) and machine translation.
[*)]
#otion and manipulation
-he field of robotics
[**]
is closely related to AI. Intelligence is re/uired for robots to be able to
handle such tasks as ob0ect manipulation
[*1]
and na!igation' ith sub<problems of locali"ation
(knoing here you are)' mapping (learning hat is around you) and motion planning (figuring
out ho to get there).
[*2]
Perception
%achine perception
[*(]
is the ability to use input from sensors (such as cameras' microphones'
sonar and others more exotic) to deduce aspects of the orld. &omputer !ision
[13]
is the ability to
analy"e !isual input. A fe selected subproblems are speech recognition'
[11]
facial recognition
and ob0ect recognition.
$ocial intelligence
:motion and social skills
[1+]
play to roles for an intelligent agent. First' it must be able to
predict the actions of others' by understanding their moti!es and emotional states. (-his in!ol!es
elements of game theory' decision theory' as ell as the ability to model human emotions and the
perceptual skills to detect emotions.) Also' for good human<computer interaction' an intelligent
machine also needs to display emotions. At the !ery least it must appear polite and sensiti!e to
the humans it interacts ith. At best' it should ha!e normal emotions itself.
%reativity
A sub<field of AI addresses creati!ity both theoretically (from a philosophical and psychological
perspecti!e) and practically (!ia specific implementations of systems that generate outputs that
can be considered creati!e). A related area of computational research is Artificial Intuition and
Artificial Imagination.
eneral intelligence
%ost researchers hope that their ork ill e!entually be incorporated into a machine ith
general intelligence (knon as strong AI)' combining all the skills abo!e and exceeding human
abilities at most or all of them.
[1#]
A fe belie!e that anthropomorphic features like artificial
consciousness or an artificial brain may be re/uired for such a pro0ect.
[1,]
%any of the problems abo!e are considered AI<complete9 to sol!e one problem' you must sol!e
them all. For example' e!en a straightforard' specific task like machine translation re/uires that
the machine follo the author6s argument (reason)' kno hat is being talked about
(knoledge)' and faithfully reproduce the author6s intention (social intelligence). %achine
translation' therefore' is belie!ed to be AI<complete9 it may re/uire strong AI to be done as ell
as humans can do it.
[1)]
Approaches
-here is no established unifying theory or paradigm that guides AI research. Hesearchers
disagree about many issues.
[1*]
A fe of the most long standing /uestions that ha!e remained
unansered are these9 should artificial intelligence simulate natural intelligence' by studying
psychology or neurologyJ Kr is human biology as irrele!ant to AI research as bird biology is to
aeronautical engineeringJ
[11]
&an intelligent beha!ior be described using simple' elegant
principles (such as logic or optimi"ation)J Kr does it necessarily re/uire sol!ing a large number
of completely unrelated problemsJ
[12]
&an intelligence be reproduced using high<le!el symbols'
similar to ords and ideasJ Kr does it re/uire "sub<symbolic" processingJ
[1(]
%ybernetics and brain simulation
In the 1(,3s and 1()3s' a number of researchers explored the connection beteen neurology'
information theory' and cybernetics. 4ome of them built machines that used electronic netorks
to exhibit rudimentary intelligence' such as =. 5rey =alter6s turtles and the $ohns 7opkins
Aeast. %any of these researchers gathered for meetings of the -eleological 4ociety at 8rinceton
Eni!ersity and the Hatio &lub in :ngland.
[#,]
Ay 1(*3' this approach as largely abandoned'
although elements of it ould be re!i!ed in the 1(23s.
$ymbolic
=hen access to digital computers became possible in the middle 1()3s' AI research began to
explore the possibility that human intelligence could be reduced to symbol manipulation. -he
research as centered in three institutions9 &%E' 4tanford and %I-' and each one de!eloped its
on style of research. $ohn 7augeland named these approaches to AI "good old fashioned AI" or
"5KFAI".
[23]
%ognitive simulation
:conomist 7erbert 4imon and Allen Deell studied human problem sol!ing skills and
attempted to formali"e them' and their ork laid the foundations of the field of artificial
intelligence' as ell as cogniti!e science' operations research and management science.
-heir research team used the results of psychological experiments to de!elop programs
that simulated the techni/ues that people used to sol!e problems. -his tradition' centered
at &arnegie %ellon Eni!ersity ould e!entually culminate in the de!elopment of the
4oar architecture in the middle 23s.
[21][2#]
!ogic based
Enlike Deell and 4imon' $ohn %c&arthy felt that machines did not need to simulate
human thought' but should instead try to find the essence of abstract reasoning and
problem sol!ing' regardless of hether people used the same algorithms.
[11]
7is
laboratory at 4tanford (4AI@) focused on using formal logic to sol!e a ide !ariety of
problems' including knoledge representation' planning and learning.
[2+]
@ogic as also
focus of the ork at the Eni!ersity of :dinburgh and elsehere in :urope hich led to
the de!elopment of the programming language 8rolog and the science of logic
programming.
[2,]
&Anti'logic& or &scruffy&
Hesearchers at %I- (such as %ar!in %insky and 4eymour 8apert)
[2)]
found that sol!ing
difficult problems in !ision and natural language processing re/uired ad<hoc solutions L
they argued that there as no simple and general principle (like logic) that ould capture
all the aspects of intelligent beha!ior. Hoger 4chank described their "anti<logic"
approaches as "scruffy" (as opposed to the "neat" paradigms at &%E and 4tanford).
[12]
&ommonsense knoledge bases (such as Coug @enat6s &yc) are an example of "scruffy"
AI' since they must be built by hand' one complicated concept at a time.
[2*]
Knowledge based
=hen computers ith large memories became a!ailable around 1(13' researchers from
all three traditions began to build knoledge into AI applications.
[21]
-his "knoledge
re!olution" led to the de!elopment and deployment of expert systems (introduced by
:dard Feigenbaum)' the first truly successful form of AI softare.
[+)]
-he knoledge
re!olution as also dri!en by the reali"ation that enormous amounts of knoledge ould
be re/uired by many simple AI applications.
$ub'symbolic
Curing the 1(*3s' symbolic approaches had achie!ed great success at simulating high<le!el
thinking in small demonstration programs. Approaches based on cybernetics or neural netorks
ere abandoned or pushed into the background.
[22]
Ay the 1(23s' hoe!er' progress in symbolic
AI seemed to stall and many belie!ed that symbolic systems ould ne!er be able to imitate all
the processes of human cognition' especially perception' robotics' learning and pattern
recognition. A number of researchers began to look into "sub<symbolic" approaches to specific
AI problems.
[1(]
(ottom'up, embodied, situated, behavior'based or nouvelle AI
Hesearchers from the related field of robotics' such as Hodney Arooks' re0ected symbolic
AI and focused on the basic engineering problems that ould allo robots to mo!e and
sur!i!e.
[2(]
-heir ork re!i!ed the non<symbolic !iepoint of the early cybernetics
researchers of the )3s and reintroduced the use of control theory in AI. -his coincided
ith the de!elopment of the embodied mind thesis in the related field of cogniti!e
science9 the idea that aspects of the body (such as mo!ement' perception and
!isuali"ation) are re/uired for higher intelligence.
%omputational Intelligence
Interest in neural netorks and "connectionism" as re!i!ed by Ca!id Humelhart and
others in the middle 1(23s.
[(3]
-hese and other sub<symbolic approaches' such as fu""y
systems and e!olutionary computation' are no studied collecti!ely by the emerging
discipline of computational intelligence.
[(1]
$tatistical
In the 1((3s' AI researchers de!eloped sophisticated mathematical tools to sol!e specific
subproblems. -hese tools are truly scientific' in the sense that their results are both measurable
and !erifiable' and they ha!e been responsible for many of AI6s recent successes. -he shared
mathematical language has also permitted a high le!el of collaboration ith more established
fields (like mathematics' economics or operations research). 4tuart Hussell and 8eter Dor!ig
describe this mo!ement as nothing less than a "re!olution" and "the !ictory of the neats."
[+2]
Integrating the approaches
Intelligent agent paradigm
An intelligent agent is a system that percei!es its en!ironment and takes actions hich
maximi"es its chances of success. -he simplest intelligent agents are programs that sol!e
specific problems. -he most complicated intelligent agents are rational' thinking humans.
[(#]
-he paradigm gi!es researchers license to study isolated problems and find solutions
that are both !erifiable and useful' ithout agreeing on one single approach. An agent
that sol!es a specific problem can use any approach that orks . some agents are
symbolic and logical' some are sub<symbolic neural netorks and others may use ne
approaches. -he paradigm also gi!es researchers a common language to communicate
ith other fields.such as decision theory and economics.that also use concepts of
abstract agents. -he intelligent agent paradigm became idely accepted during the 1((3s.
[(+]
Agent architectures and cognitive architectures
Hesearchers ha!e designed systems to build intelligent systems out of interacting
intelligent agents in a multi<agent system.
[(,]
A system ith both symbolic and sub<
symbolic components is a hybrid intelligent system' and the study of such systems is
artificial intelligence systems integration. A hierarchical control system pro!ides a bridge
beteen sub<symbolic AI at its loest' reacti!e le!els and traditional symbolic AI at its
highest le!els' here relaxed time constraints permit planning and orld modelling.
[()]
Hodney Arooks6 subsumption architecture as an early proposal for such a hierarchical
system.
Tools
In the course of )3 years of research' AI has de!eloped a large number of tools to sol!e the most
difficult problems in computer science. A fe of the most general of these methods are discussed
belo.
$earch and optimi)ation
%any problems in AI can be sol!ed in theory by intelligently searching through many possible
solutions9
[(*]
Heasoning can be reduced to performing a search. For example' logical proof can be
!ieed as searching for a path that leads from premises to conclusions' here each step is the
application of an inference rule.
[(1]
8lanning algorithms search through trees of goals and
subgoals' attempting to find a path to a target goal' a process called means<ends analysis.
[(2]
Hobotics algorithms for mo!ing limbs and grasping ob0ects use local searches in configuration
space.
[*1]
%any learning algorithms use search algorithms based on optimi"ation.
4imple exhausti!e searches
[((]
are rarely sufficient for most real orld problems9 the search space
(the number of places to search) /uickly gros to astronomical numbers. -he result is a search
that is too slo or ne!er completes. -he solution' for many problems' is to use "heuristics" or
"rules of thumb" that eliminate choices that are unlikely to lead to the goal (called "pruning the
search tree"). 7euristics supply the program ith a "best guess" for hat path the solution lies
on.
[133]
A !ery different kind of search came to prominence in the 1((3s' based on the mathematical
theory of optimi"ation. For many problems' it is possible to begin the search ith some form of a
guess and then refine the guess incrementally until no more refinements can be made. -hese
algorithms can be !isuali"ed as blind hill climbing9 e begin the search at a random point on the
landscape' and then' by 0umps or steps' e keep mo!ing our guess uphill' until e reach the top.
Kther optimi"ation algorithms are simulated annealing' beam search and random optimi"ation.
[131]
:!olutionary computation uses a form of optimi"ation search. For example' they may begin ith
a population of organisms (the guesses) and then allo them to mutate and recombine' selecting
only the fittest to sur!i!e each generation (refining the guesses). Forms of e!olutionary
computation include sarm intelligence algorithms (such as ant colony or particle sarm
optimi"ation)
[13#]
and e!olutionary algorithms (such as genetic algorithms
[13+]
and genetic
programming
[13,][13)]
).
!ogic
@ogic
[13*]
is used for knoledge representation and problem sol!ing' but it can be applied to other
problems as ell. For example' the satplan algorithm uses logic for planning
[131]
and inducti!e
logic programming is a method for learning.
[132]
4e!eral different forms of logic are used in AI research. 8ropositional or sentential logic
[13(]
is the
logic of statements hich can be true or false. First<order logic
[113]
also allos the use of
/uantifiers and predicates' and can express facts about ob0ects' their properties' and their
relations ith each other. Fu""y logic'
[111]
is a !ersion of first<order logic hich allos the truth
of a statement to be represented as a !alue beteen 3 and 1' rather than simply -rue (1) or False
(3). Fu""y systems can be used for uncertain reasoning and ha!e been idely used in modern
industrial and consumer product control systems. 4ub0ecti!e logic models uncertainty in a
different and more explicit manner than fu""y<logic9 a gi!en binomial opinion satisfies belief M
disbelief M uncertainty N 1 ithin a Aeta distribution. Ay this method' ignorance can be
distinguished from probabilistic statements that an agent makes ith high confidence. Cefault
logics' non<monotonic logics and circumscription
[)1]
are forms of logic designed to help ith
default reasoning and the /ualification problem. 4e!eral extensions of logic ha!e been designed
to handle specific domains of knoledge' such as9 description logicsG
[,)]
situation calculus' e!ent
calculus and fluent calculus (for representing e!ents and time)G
[,*]
causal calculusG
[,1]
belief
calculusG and modal logics.
[,2]
Probabilistic methods for uncertain reasoning
%any problems in AI (in reasoning' planning' learning' perception and robotics) re/uire the
agent to operate ith incomplete or uncertain information. AI researchers ha!e de!ised a number
of poerful tools to sol!e these problems using methods from probability theory and economics.
[11#]
Aayesian netorks
[11+]
are a !ery general tool that can be used for a large number of problems9
reasoning (using the Aayesian inference algorithm)'
[11,]
learning (using the expectation<
maximi"ation algorithm)'
[11)]
planning (using decision netorks)
[11*]
and perception (using
dynamic Aayesian netorks).
[111]
8robabilistic algorithms can also be used for filtering'
prediction' smoothing and finding explanations for streams of data' helping perception systems
to analy"e processes that occur o!er time
[112]
(e.g.' hidden %arko! models
[11(]
or >alman
filters
[1#3]
).
A key concept from the science of economics is "utility"9 a measure of ho !aluable something
is to an intelligent agent. 8recise mathematical tools ha!e been de!eloped that analy"e ho an
agent can make choices and plan' using decision theory' decision analysis'
[1#1]
information !alue
theory.
[)1]
-hese tools include models such as %arko! decision processes'
[1##]
dynamic decision
netorks'
[1##]
game theory and mechanism design.
[1#+]
%lassifiers and statistical learning methods
-he simplest AI applications can be di!ided into to types9 classifiers ("if shiny then diamond")
and controllers ("if shiny then pick up"). &ontrollers do hoe!er also classify conditions before
inferring actions' and therefore classification forms a central part of many AI systems. &lassifiers
are functions that use pattern matching to determine a closest match. -hey can be tuned
according to examples' making them !ery attracti!e for use in AI. -hese examples are knon as
obser!ations or patterns. In super!ised learning' each pattern belongs to a certain predefined
class. A class can be seen as a decision that has to be made. All the obser!ations combined ith
their class labels are knon as a data set. =hen a ne obser!ation is recei!ed' that obser!ation is
classified based on pre!ious experience.
[1#,]
A classifier can be trained in !arious aysG there are many statistical and machine learning
approaches. -he most idely used classifiers are the neural netork'
[1#)]
kernel methods such as
the support !ector machine'
[1#*]
k<nearest neighbor algorithm'
[1#1]
5aussian mixture model'
[1#2]
nai!e Aayes classifier'
[1#(]
and decision tree.
[1+3]
-he performance of these classifiers ha!e been
compared o!er a ide range of tasks. &lassifier performance depends greatly on the
characteristics of the data to be classified. -here is no single classifier that orks best on all
gi!en problemsG this is also referred to as the "no free lunch" theorem. Cetermining a suitable
classifier for a gi!en problem is still more an art than science.
[1+1]
"eural networs
-he study of artificial neural netorks
[1#)]
began in the decade before the field AI research as
founded' in the ork of =alter 8itts and =arren %c&ullough. Kther important early researchers
ere Frank Hosenblatt' ho in!ented the perceptron and 8aul =erbos ho de!eloped the
backpropagation algorithm.
[1+#]
-he main categories of netorks are acyclic or feedforard neural netorks (here the signal
passes in only one direction) and recurrent neural netorks (hich allo feedback). Among the
most popular feedforard netorks are perceptrons' multi<layer perceptrons and radial basis
netorks.
[1++]
Among recurrent netorks' the most famous is the 7opfield net' a form of attractor
netork' hich as first described by $ohn 7opfield in 1(2#.
[1+,]
Deural netorks can be applied
to the problem of intelligent control (for robotics) or learning' using such techni/ues as 7ebbian
learning and competiti!e learning.
[1+)]
$eff 7akins argues that research in neural netorks has stalled because it has failed to model
the essential properties of the neocortex' and has suggested a model (7ierarchical -emporal
%emory) that is loosely based on neurological research.
[1+*]
%ontrol theory
&ontrol theory' the grandchild of cybernetics' has many important applications' especially in
robotics.
[1+1]
!anguages
AI researchers ha!e de!eloped se!eral speciali"ed languages for AI research' including @isp
[1+2]
and 8rolog.
[1+(]
*valuating progress
In 1()3' Alan -uring proposed a general procedure to test the intelligence of an agent no
knon as the -uring test. -his procedure allos almost all the ma0or problems of artificial
intelligence to be tested. 7oe!er' it is a !ery difficult challenge and at present all agents fail.
Artificial intelligence can also be e!aluated on specific problems such as small problems in
chemistry' hand<riting recognition and game<playing. 4uch tests ha!e been termed sub0ect
matter expert -uring tests. 4maller problems pro!ide more achie!able goals and there are an
e!er<increasing number of positi!e results.
-he broad classes of outcome for an AI test are9
Kptimal9 it is not possible to perform better
4trong super<human9 performs better than all humans
4uper<human9 performs better than most humans
4ub<human9 performs orse than most humans
For example' performance at draughts is optimal'
[1,3]
performance at chess is super<human and
nearing strong super<human'
[1,1]
and performance at many e!eryday tasks performed by humans
is sub<human.
A /uite different approach measures machine intelligence through tests hich are de!eloped
from mathematical definitions of intelligence. :xamples of these kinds of tests start in the late
nineties de!ising intelligence tests using notions from >olmogoro! &omplexity and data
compression.
[1,#]

[1,+]
-o ma0or ad!antages of mathematical definitions are their applicability to
nonhuman intelligences and their absence of a re/uirement for human testers.
Applications
Artificial intelligence techni/ues are per!asi!e and are too numerous to list. Fre/uently' hen a
techni/ue reaches mainstream use' it is no longer considered artificial intelligenceG this
phenomenon is described as the AI effect.
[1,,]
%ompetitions and pri)es
-here are a number of competitions and pri"es to promote research in artificial intelligence. -he
main areas promoted are9 general machine intelligence' con!ersational beha!ior' data<mining'
dri!erless cars' robot soccer and games.
Platforms
A platform (or "computing platform")is defined as "some sort of hardare architecture or
softare frameork (including application frameorks)' that allos softare to run." As Hodney
Arooks
[1,)]
pointed out many years ago' it is not 0ust the artificial intelligence softare that
defines the AI features of the platform' but rather the actual platform itself that affects the AI that
results' i.e.' e need to be orking out AI problems on real orld platforms rather than in
isolation.
A ide !ariety of platforms has alloed different aspects of AI to de!elop' ranging from expert
systems' albeit 8&<based but still an entire real<orld system to !arious robot platforms such as
the idely a!ailable Hoomba ith open interface.
[1,*]
Philosophy
Artificial intelligence' by claiming to be able to recreate the capabilities of the human mind' is
both a challenge and an inspiration for philosophy. Are there limits to ho intelligent machines
can beJ Is there an essential difference beteen human intelligence and artificial intelligenceJ
&an a machine ha!e a mind and consciousnessJ A fe of the most influential ansers to these
/uestions are gi!en belo.
[1,1]
Turing+s &polite convention&
If a machine acts as intelligently as a human being then it is as intelligent as a human
being. Alan -uring theori"ed that' ultimately' e can only 0udge the intelligence of a
machine based on its beha!ior. -his theory forms the basis of the -uring test.
[1,2]
The Dartmouth proposal
!"very aspect of learning or any other feature of intelligence can be so precisely
described that a machine can be made to simulate it.! -his assertion as printed in the
proposal for the Cartmouth &onference of 1()*' and represents the position of most
orking AI researchers.
[1,(]
"ewell and $imon+s physical symbol system hypothesis
!# physical symbol system has the necessary and sufficient means of general intelligent
action.! Deell and 4imon argue that intelligences consists of formal operations on
symbols.
[1)3]
7ubert Creyfus argued that' on the contrary' human expertise depends on
unconscious instinct rather than conscious symbol manipulation and on ha!ing a "feel"
for the situation rather than explicit symbolic knoledge. (4ee Creyfus6 criti/ue of AI.)
[1)1][1)#]
,-del+s incompleteness theorem
# formal system (such as a computer program) can not prove all true statements. Hoger
8enrose is among those ho claim that 5Odel6s theorem limits hat machines can do.
(4ee $he "mperor's %e& 'ind.)
[1)+][1),]
$earle+s strong AI hypothesis
!$he appropriately programmed computer &ith the right inputs and outputs &ould
thereby have a mind in e(actly the same sense human beings have minds.!
[1))]
4earle
counters this assertion ith his &hinese room argument' hich asks us to look inside the
computer and try to find here the "mind" might be.
[1)*]
The artificial brain argument
$he brain can be simulated. 7ans %ora!ec' Hay >ur"eil and others ha!e argued that it
is technologically feasible to copy the brain directly into hardare and softare' and that
such a simulation ill be essentially identical to the original.
Prediction
%ain articles9 Artificial intelligence in fiction' :thics of artificial intelligence' -ranshumanism'
and -echnological singularity
AI is a common topic in both science fiction and pro0ections about the future of technology and
society. -he existence of an artificial intelligence that ri!als human intelligence raises difficult
ethical issues' and the potential poer of the technology inspires both hopes and fears.
In fiction' AI has appeared fulfilling many roles' including a ser!ant (H#C# in )tar *ars)' a la
enforcer (>.I.-.-. ">night Hider")' a comrade (@t. &ommander Cata in )tar $rek+ $he %e(t
,eneration)' a con/uerorIo!erlord ($he 'atri()' a dictator (*ith Folded Hands)' an assassin
($erminator)' a sentient race (-attlestar ,alacticaI-ransformers)' an extension to human abilities
(,host in the )hell) and the sa!ior of the human race (H. Caneel Kli!a in the Foundation
)eries).
%ary 4helley6s Frankenstein
[1)2]
considers a key issue in the ethics of artificial intelligence9 if a
machine can be created that has intelligence' could it also feelJ If it can feel' does it ha!e the
same rights as a humanJ -he idea also appears in modern science fiction' including the films I
Robot' -lade Runner and #.I.+ #rtificial Intelligence' in hich humanoid machines ha!e the
ability to feel human emotions. -his issue' no knon as "robot rights"' is currently being
considered by' for example' &alifornia6s Institute for the Future'
[1)(]
although many critics belie!e
that the discussion is premature.
[1*3]
-he impact of AI on society is a serious area of study for futurists. Academic sources ha!e
considered such conse/uences as a decreased demand for human labor'
[1*1]
the enhancement of
human ability or experience'
[1*#]
and a need for redefinition of human identity and basic !alues.
[1*+]
Andre >ennedy' in his musing on the e!olution of the human personality'
[1*,]
considered
that artificial intelligences or 6ne minds6 are likely to ha!e se!ere personality disorders' and
identifies four particular types that are likely to arise9 the autistic' the collector' the ecstatic' and
the !ictim. 7e suggests that they ill need humans because of our superior understanding of
personality and the role of the unconscious.
4e!eral futurists argue that artificial intelligence ill transcend the limits of progress. Hay
>ur"eil has used %oore6s la (hich describes the relentless exponential impro!ement in
digital technology) to calculate that desktop computers ill ha!e the same processing poer as
human brains by the year #3#(. 7e also predicts that by #3,) artificial intelligence ill reach a
point here it is able to impro!e itself at a rate that far exceeds anything concei!able in the past'
a scenario that science fiction riter Pernor Pinge named the "technological singularity".
[1*#]
Hobot designer 7ans %ora!ec' cyberneticist >e!in =arick and in!entor Hay >ur"eil ha!e
predicted that humans and machines ill merge in the future into cyborgs that are more capable
and poerful than either.
[1*#]
-his idea' called transhumanism' hich has roots in Aldous 7uxley
and Hobert :ttinger' has been illustrated in fiction as ell' for example in the manga ,host in
the )hell and the science<fiction series .une.
:dard Fredkin argues that "artificial intelligence is the next stage in e!olution'"
[1*)]
an idea first
proposed by 4amuel Autler6s "Carin among the %achines" (12*+)' and expanded upon by
5eorge Cyson in his book of the same name in 1((2.
8amela %c&orduck rites that all these scenarios are expressions of the ancient human desire to'
as she calls it' "forge the gods".
[
.eferences
1. / 8oole' %ackorth Q 5oebel 1((2' p. 1 (ho use the term "computational intelligence"
as a synonym for artificial intelligence). Kther textbooks that define AI this ay include
Dilsson (1((2)' and Hussell Q Dor!ig (#33+) (ho prefer the term "rational agent") and
rite "-he hole<agent !ie is no idely accepted in the field" (Hussell Q Dor!ig
#33+' p. ))) -hese textbooks are the most idely used in academic AI. 4ee -extbooks at
AI topics
#. / -his definition' in terms of goals' actions' perception and en!ironment' is due to
Hussell Q Dor!ig (#33+). Kther definitions also include knoledge and learning as
additional criteria.
+. / Although there is some contro!ersy on this point (see &re!ier 1((+' p. )3)' %c&arthy
states une/ui!ocally "I came up ith the term" in a cRnet inter!ie. (4ee 5etting
%achines to -hink @ike Es.)
,. / 4ee $ohn %c&arthy' =hat is Artificial IntelligenceJ
). / 4ee the Cartmouth proposal' under 8hilosophy' belo.
*. S
a

b

c
-his is a central idea of 8amela %c&orduck6s 'achines $hat $hink. 4he rites9 "I
like to think of artificial intelligence as the scientific apotheosis of a !enerable cultural
tradition." (%c&orduck #33,' p. +,) "Artificial intelligence in one form or another is an
idea that has per!aded =estern intellectual history' a dream in urgent need of being
reali"ed." (%c&orduck #33,' p. x!iii) "Kur history is full of attempts.nutty' eerie'
comical' earnest' legendary and real.to make artificial intelligences' to reproduce hat
is the essential us.bypassing the ordinary means. Aack and forth beteen myth and
reality' our imaginations supplying hat our orkshops couldn6t' e ha!e engaged for a
long time in this odd form of self<reproduction." (%c&orduck #33,' p. +) 4he traces the
desire back to its 7ellenistic roots and calls it the urge to "forge the 5ods." (%c&orduck
#33,' p. +,3<,33)
1. / -he optimism referred to includes the predictions of early AI researchers (see optimism
in the history of AI) as ell as the ideas of modern transhumanists such as Hay >ur"eil.
2. / -he "setbacks" referred to include the A@8A& report of 1(**' the abandonment of
perceptrons in 1(13' the the @ighthill Heport of 1(1+ and the collapse of the lisp machine
market in 1(21.