ebook img

Encyclopedia of Physical Science and Technology PDF

136 Pages·2001·8.224 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Encyclopedia of Physical Science and Technology

P1:GAERevisedPages Qu:00,00,00,00 EncyclopediaofPhysicalScienceandTechnology EN001H-27 May26,2001 14:42 Artificial Intelligence Franz J. Kurfess CaliforniaPolytechnicStateUniversity I. TheHistoryofArtificialIntelligence II. FundamentalAIConcepts III. AIAreas IV. ResearchIssues V. Outlook GLOSSARY methodsrelyonpredicatelogicastheirunderlyingfor- malframework. Artificial intelligence Discipline of computer science Naturallanguageprocessing Enablesacomputertoin- that aims at building systems showing some intelli- terpret input provided in a natural language such as genceintheiroperationorbehavior;alsoinvestigates English; usually requires text as input, but it may be the way humans perform tasks that require intelli- combinedwithspeechrecognitionforspokeninput. gence. Neural network Collection of interconnected neurons Computervision Opticalsensorsareusedtoprovidethe that perform simple computations; usually used with equivalentofeyestoacomputersystem;oftenusedfor a learning method to extract regularities from sets of objectrecognitionandnavigationpurposes. exampledata. Expertsystem Computerprogramthatemulatesthebe- Robotics Aimsatthedesignofintelligentrobotscapable havior of a human expert in a specific domain, or for ofacquiringinputfromtheenvironmentthroughsen- aspecifictask;usuallycontainsaknowledgebaseand sorsandmovingaroundandmanipulatingobjectswith aninferencemechanism. thehelpofeffectors. Inferencemechanism Programthatgenerateshypothe- Speech recognition Computer-based identification of sesanddrawsconclusionsaccordingtosomereason- patternsinspokeninputfortheconversionofanaudio ing methods determined by inference rules, based on signalintoastringofwords. knowledgestoredinaknowledgebase. Knowledgebase Collectionofstatementsthatdescribes theknowledgeavailablewithinanexpertsystem. ARTIFICIAL INTELLIGENCE (AI) is a discipline Knowledgerepresentation Methodsusedforstoringthe within the field of computer science, with strong influ- statements capturing the knowledge about a specific encesfromcognitivescience/psychology,philosophy,lin- domain or task. Frequently used methods are rules, guistics,andmathematics.Itsbroadgoalistocreatesys- frames, scripts, or semantic networks. Many of these tems that incorporate or exhibit some intelligence. This 609 P1:GAERevisedPages EncyclopediaofPhysicalScienceandTechnology EN001H-27 May7,2001 15:31 610 ArtificialIntelligence canbedoneeitherbyemulatingthewayhumansperform aspects are the basis for many approaches and methods tasks that require intelligence (e.g., the use of analogies thatareusedinpractice. tofindasolutionforaproblem),orbyusingtechniques Some of the work performed by early computer pio- more suitable for computer processing (e.g., chess pro- neersalsostrikesachordwithlaterissuesthatcameupin gramsthatrelyheavilyonsystematicsearchmethodsand thefieldofAI.Inhis“analyticalengine,”CharlesBabbage databasesofendgames).Itincludessuchaspectsasrea- outlinesamechanicaldevicewithastunningsimilarityto soning,learning,searchingforsolutionstoproblems,and concepts found in modern computers, such as the sep- explainingthestepstakentosolveaproblem.Atthecoreof aration of memory and processor, digital operation, and mostAIsystemsliestheirabilitytodealwithknowledge: programmability.Healsodiscusses“patterns”ofintellec- acquiringknowledgefromexpertsorfromdatadelivered tual activity to be implementedby his analytical engine. throughsensorsorothercomputersystems,processingthe Othercomputerpioneersknownmostlyfortheiraccom- knowledge,ideallyaccordingtothesoundprinciplesofa plishmentsinthetheoryofcomputingandthedesignof formal reasoning method, selecting an action to solve a computers, such as Claude Shannon, Alan Turing, and problem, or part of it, and performing that action in the JohnvonNeumann,investigatedtheuseofcomputersfor internalrepresentationofthesystem,orintherealworld. chess programs and other games in the late 1940s and Tasksthatinvolveinteractionwiththerealworldcanbe early1950s. especiallychallenging,requiringthesystemtodealwith Aroundthesametime,athirdinfluentialstreamofde- possiblyhugequantitiesofrawdatafromvarioussenses velopmentswasinspiredbyresearchersworkingonmod- (e.g.,vision,hearing,touch),inconsistenciesbetweenex- elsofneurons.McCullochandPitts,Hebb,Minsky,and pertknowledgeandknowledgeextractedfromrawdata, others developed mathematical models of neurons and planning of sequences of action to achieve a goal, rec- learningprocessesinneurons,andalsobuiltneuralcom- onciliation of the results of actions performed with their puters. expectedeffects,andcommunicationwithothersystems orhumans. B. Birth Dartmouth,NewHampshire,inthesummerof1956was I. THE HISTORY OF ARTIFICIAL thelocationofaworkshopthatbroughttogetheragroup INTELLIGENCE of about 10 researchers with a common interest in the useofcomputerstosolveproblemsthatseemtorequire The“birth”ofthefieldnowknownasartificialintelligence intelligenceinhumans.Thisworkshopnotonlyledtoan isgenerallyattributedtoaworkshopthattookplaceinthe intensive exchange of ideas, it also gave rise to the term summer of 1956 at Dartmouth. Carrying the analogy to “artificialintelligence,”coinedbyJohnMcCarthy. human development one step further, we present impor- tantdevelopmentsaccordingtothephasesahumangoes C. BabySteps throughinhisorherlife. Therestofthe1950sbroughtforwardanumberofcom- puterprogramsthatsolvesimpleproblemsrequiringsome A. GestationPeriod intelligence,suchasthe“LogicTheorist”byAlanNewell ManyofthequestionsatthecenterofAIresearchofcourse and Herbert Simon (who later would be honored with a haveamuchlongerhistory,goingbacktoGreekphiloso- Nobel prize for his contributions to economics), and a phers such as Aristotle and Socrates. Probably the most seriesofcheckersprogramsbyA.L.Samuel.Thesepro- relevantinfluencesfromthatperiodcameoutoftheirat- gramsandothereffortsestablishedasubstantialnumber tempts to develop a framework for discourses in such a offormalandpracticalmethodsforknowledgerepresen- waythatitispossibletodeterminethewinnerofadebate, tationandreasoning,nottheleastamongthembeingthe according to well-established and generally accepted languageLISP,makingit(togetherwithFORTRAN)one rules.Thisestablishedthefoundationsforreasoningmeth- oftheoldestcomputerprogramminglanguagesstillinuse. ods, to be refined and formalized in the 19th and 20th centuriesbyGeorgeBoole,GottfriedWilhelmvonLeib- D. Kindergarten niz,GottliebFrege,BertrandRussell,KurtGo¨del,Alfred Tarski,andothers,intowhatwenowknowasmathemat- Whereas the activities mentioned so far were mainly ical logic. Although most forms of mathematical logic known to a relatively small community of researchers are computationally quite expensive, and often imprac- with common interest, the next phase in the early 1960s tical for real-world problems, their well-founded formal producedsome“childprodigies”thatastoundedalarger P1:GAERevisedPages EncyclopediaofPhysicalScienceandTechnology EN001H-27 May7,2001 15:31 ArtificialIntelligence 611 publicwiththeirskills.NewellandSimonpresentedtheir areas,andworkwascarriedonmainlybyacoreofded- “GeneralProblemSolver”endowedwithbasicproblem- icated researchers, and a few practitioners applying AI solving and reasoning skills, and Shakey the robot ex- methodstopracticalproblemsintherealworld. ploreditsenvironmentattheStanfordResearchInstitute. BasedontheearlyworkofHebbandothersonneuralnet- G. AIGetsaJob works,BernhardWidrowandFrankRosenblattsuggested improvedlearningalgorithmsforneuralnetworks.Fora Oneoftheareaswithearly,demonstrablesuccesseswas class of neural networks called perceptrons, Rosenblatt expertsystems,wherecarefulcapturingofhumanexper- wasabletoshowthattheycouldlearnamappingfroman tise into rules combined with heuristically guided rea- input vector to an output, provided that such a mapping soningcomponentsledtoanumberofpracticalapplica- exists. tions.DENDRAL,developedin1969byEdFeigenbaum, BruceBuchanan,andJoshuaLederberg,infersmolecular structurefromthechemicalformulaofamoleculeanda E. GradeSchool mass spectrogram. MYCIN, developed by Feigenbaum, In the late 1960s, Evans and Bobrow wrote programs Buchanan,andEdShortliffe,isamedicaldiagnosissys- thatknewhowtosolvegeometricanalogiesandalgebraic temforbloodinfections.LikeDENDRAL,itcapturesex- problems,andagrouparoundMarvinMinskystartedto pertiseintheformofrulesandgeneratesconclusionsfrom uselimiteddomains,ormicroworlds,asatestingground theserulesthroughareasoningcomponent.Sincemedical foralargevarietyofAIsystems,rangingfromreasoners knowledgeisencumberedbysomeuncertainty,MYCIN overimageprocessingandpatternrecognitiontorobots. usescertaintyfactorstoindicatehowconfidenttheexperts Manyofthesesystemsdemonstratedcapabilitiesthatwere are about the statement in a particular rule. During the generallybelievedimpossibleforcomputerstoachieve. 1970sandintothe1980s,expertsystemsweredeveloped for a number of applications in fields such as medicine, geology,andtheconfigurationofcomputersystems.Inthe F. TeenageYears lastdomain,thesystemR1isoneofthefirstexpertsys- Some of the early successes of AI, together with the temstobecommerciallysuccessful.Itwasdevelopedin general interest in the topic and a number of specula- theearly1980stohelpwiththecustomer-specificconfig- tive predictions by AI proponents, raised somewhat un- urationofDigitalEquipmentCorporation’slineofVAX realisticexpectationsforpracticalapplicationsandantic- computers, and it reportedly saved the company around ipated capabilities of “intelligent” systems. Many of the $40 million per year. Although the actual numbers are limited and isolated solutions to microworld problems, sometimes difficult to calculate, many U.S. corporations however, turned out to be problematic in realistic envi- derivedsubstantialbenefitsfromtheuseofexpertsystems ronments. Real-world problems frequently impose more duringthattime. parameters, more complex constraints, inconsistent and Together with the success of expert systems came the incompleteinformation,andtimelimitationsforsystems realizationthattheunderlyingknowledgerepresentation toworkwith.Forsuccessfulsolutions,generalknowledge schemesandreasoningcomponentsrequiresoundformal isfrequentlyrequiredinadditiontotheexpertknowledge, foundationstogetherwithatleastacceptableperformance and calculating a solution becomes quickly intractable forpracticalapplications.Thisfueledtheincorporationof since the number of candidate solutions to be considers aspectsofprobabilitytheory,logic,andlinguisticsforbet- grows extremely fast. In addition, some of the methods terfoundations;thedevelopmentanduseofexpertsystem employedbyAIresearchershadsubstantialfundamental shells;specializedlanguagessuchasLISPandPROLOG; limitations,discoveredsomewhatlatebecauseitwasof- andeventheconstructionofdedicatedcomputerworksta- tenmoreinterestingtotryoutthesemethods,ratherthan tionssuchasLISPmachines.In1981,Japanlaunchedtheir carefully investigate their basic properties. An example “FifthGeneration”project,alarge-scaleeffortinvolving hereisneuralnetworks.MinskyandPapert,intheir1969 companies and universities to build intelligent computer book Perceptrons, showed that although perceptrons in- systems, with a variant of PROLOG as system and ap- deed could learn anything within their capabilities, their plicationprogramminglanguage,anddedicatedhardware capabilitieswereseverelylimited,namelytolinearlysep- optimizedtoperformmillionsofinferencespersecond. arable functions. Thus, for example, perceptrons are not capableoflearningtheexclusiveor(XOR)function. H. AcquiringValuableWorkSkills These difficulties and unmet expectations led to some disillusionmentinresearchand,possiblymoreconsequen- Encouraged by the success of expert systems for practi- tial,fundingcircles.Duringthis“AIwinter,”fundingwas cal applications, AI methods started to make their way reduced,peopleandorganizationsconcentratedonother into various types of computer systems, ranging from P1:GAERevisedPages EncyclopediaofPhysicalScienceandTechnology EN001H-27 May7,2001 15:31 612 ArtificialIntelligence consumeritemssuchasvideocameras,oversoftwareap- and commercial products for lane following or keeping plications,tolargeindustrialcontrolsystems:Fuzzylogic asafedistanceshouldbeavailablesoon.Thesteptoau- helpsstabilizetheimageinhand-heldvideocameras,im- tonomous vehicles that shuttle passengers from home to provestheefficiencyofhouseholditemssuchaswashers workwithoutinputorattentionfromthedriver,however,is and dryers, optimizes various aspects in the engine and stillbeyondthereachofcurrentmethodsandtechnology. braking systems of automobiles, and controls the opera- Thistrendofbuildingsystemsthatassisthumans,instead tion of a metro system in Sendai, Japan. Expert systems oftakingfullcontrolofatask,hasbeenathreadthrough areroutinelyusedinalargevarietyofdomains,albeitnot manyareas;andforpsychological,legal,andsocietalrea- somuchasautonomoussystems,butmoreinanadvisory sons,itprobablywillcontinuefortheforeseeablefuture. functiontohumans.Neuralnetworksandmachinelearn- ing methods are regularly applied for classification and categorizationtasks,andinsituationswhereresponsesto II. FUNDAMENTAL AI CONCEPTS aninputcanbelearnedfromasetofexamplesandgener- alizedtounknowninputs. Intheprevioussectionsonthehistoricaldevelopmentof the field, some important concepts, methods, and appli- cations of artificial intelligence already made a brief ap- I. AIBlendsintotheWorkPlace pearance, albeit without much explanation. This section Startinginthe1990s,AImethodsbegantobeintegrated will describe the conceptual foundations of artificial in- into conventional computer-based systems and applica- telligence,tobefollowedbyadiscussionofpracticaland tions, often without the knowledge of the actual user. research-orientedareasinthefield. Wizards and agents use Bayesian networks to offer as- sistancetousersforcomplextasks,planningsystemsper- A. KnowledgeRepresentation formschedulingforfactoriesandspacemissions,neural networksandfuzzylogiccontrollersareimplementedin Thecapabilitytodealwithknowledgeinvariousformsis hardware and integrated into consumer products as well anessentialcharacteristicofalmostallAIsystems.Todis- asindustrialsystems,anddataminingtechniquesextract cussthisfurther,itmaybehelpfultodistinguishbetween valuable information from huge quantities of raw data. data,suchasthecontentsofadatabase,andtheknowl- Robots, already a staple in industrial environments such edgerequiredtoperformacertaintaskbyacomputer.Two as car manufacturing, are becoming more autonomous, importantaspectsherearetherepresentation,operations andinthelate1990s,commercialrobotsbegantoappear performedon,andtheinterpretationofdataorknowledge. fortaskssuchaswarehousing,lawnmowing,andclean- Inthecaseofdata,therepresentationisusuallyverysys- ing.Forothertaskssuchasspeechorhandwritingrecog- tematic,suchasthegridofcellsinaspreadsheet,thecol- nition,advancedcomputationalmethodssuchasHidden lectionofrecordsinadatabase,orthecharacters,lines, Markov Models together with improved microprocessor andpagesinaword-processingdocument.Typicaloper- performancesolutionsareinpracticaluseforspecialized ations performed on data are numerical operations such systems and applications (such as flight reservations via asaddition,subtraction,orcomparisonfornumbers,and phone,orhandwritteninputforhand-heldcomputers),and comparison,concatenation,changeofattributes,etc.,for more generic programs are available at affordable costs othertypesofdata.Theinterpretation,i.e.,theassignment for general-purpose tasks. To some degree, AI methods of meaning for data, is mostly left to the user, who will are also used for finding and locating information on a knowfromthecontextofthetask,throughadditionalde- computer disk, in a local area network, or on the World scriptiveinformationstoredwiththedata,orthroughthe WideWeb.Thebasicmethodherereliesonfindingmatch- values of the data themselves (in the case of text docu- ingstringsinhugeindicesextractedfromdocuments,and ments)whataparticulardataitem“means.”Forthecom- AIapproacheshelpidentifysimilardocuments,orcluster puter,alloftheessentialoperationscanbeperformedby founditemsintogroups. followingthealgorithmdescribedinaprogram,without Rather than spectacular events such as the defeat of having to know what the intended meaning of a specific chessGrandMasterGaryKasparovbytheIBMcomputer data item is. For knowledge, these issues become much DeepBlue(whichdidnotreallyusemuch“intelligence,” morecomplicated.Ingeneral,knowledgecannotbecap- butrather“bruteforce”)in1998,thesmoothintegration turedeasilyinsuchasystematicwayastablesorrecords. of AI methods into existing systems will continue over For knowledge, both the values of specific items, and thenextfewyears.Computer-assisteddriving,forexam- theirrelationshipswithotherrelevantitemsmustbecap- ple,hasbeendemonstratedforafewyearsinprototypes, tured.Thisleadstoamorecomplexrepresentation,often P1:GAERevisedPages EncyclopediaofPhysicalScienceandTechnology EN001H-27 May7,2001 15:31 ArtificialIntelligence 613 displayedasgraphs,orwrittendownascollectionsoffacts Obviously, a much larger collection of rules must be andrules.Inadditiontothecalculationofspecificvalues consideredforasatisfactoryidentificationofevensimple forsomeaspectofanitem,operationsonknowledgemust computer problems, but this simple example illustrates include the generation of new items from the already how knowledge is captured in rules. The next step, how knownones,accordingtosomegenerallyacceptedmeth- to use these rules for problem solving, requires reason- odscapturedinareasoningcomponent.Theinterpretation ingmethodscoveredlater;theoveralloperationofexpert ofknowledgeultimatelyisalsodonebyhumans,although systemsisdescribedinmoredetailinaseparatesection. itimposesmuchmoresevereconstraintsonthereasoning systemimplementedinacomputer:Themanipulationof 2. SemanticNetworksandConceptualGraphs theknowledgeitemsbythecomputermustbeperformed insuchawaythatitiscompatiblewiththeinterpretation An important aspect of knowledge representation is the of these items by humans. This can be compared to the explicit statement of relationships between concepts or commontaskofsolvingapuzzle:Theindividualpiecesof objects. In the human mind, these relationships are of- thepuzzlecorrespondtoknowledgeitems,andtheymust ten manifested through associations between concepts. becomposedaccordingtocertainconstraintssuchasthe Thefirstcomputerimplementationsofsemanticnetworks shape of the pieces. As long as all the pieces are distin- were developed in the early 1960s, based on research in guishablethroughtheirshapeonly,acomputershouldbe psychology,cognitivescience,andphilosophyonhowto able to put together the puzzle. As soon as some pieces representstructuresofconceptsandassociations.Inase- havethesameshape,however,themeaningofthepieces mantic network, concepts are represented as nodes, and withrespecttotheoverallpicturerepresentedbythepuz- linksbetweenthenodesindicatetherelationships.Since zlebecomesimportant.Ahumanpuzzlerwilltakethisinto thereareahugevarietyofpossiblerelationshipsbetween accounttofindtherightplaceforapiece;foracomputer, concepts,eitherthelinkshavetobelabeledtoindicatethe thisbecomesmuchmoredifficult.Withoutaninterpreta- specific relationships, or visual markers such as dashed tionofthepicture,andallthecommon-senseknowledge arrows,doublearrows,orcolorhavetobeusedforcom- thatbecomesavailablethroughthis(“Thisisapictureof mon relationships. One especially powerful relationship Neuschwanstein Castle, and the piece is sky-blue, with insuchnetworksisinheritance,whichindicatesthatmore cloud-likeshapes,soitmustbeintheupperportionofthe abstractconcepts(e.g.,“bird”)providemoreconcretecon- picture”), the computer must compare the pattern repre- cepts(e.g.,“canary”)orspecificindividualentities(e.g., sentedinthepieceagainstthereferenceimageandtryto “Tweety”) with certain properties (e.g., “capable of fly- identifythecorrectlocation.Ifnoreferenceimageisavail- ing”).Incasesuchapropertydoesnothold(e.g.,forpen- able,theshapeofthepiecetogetherwithcolorinformation guins,whicharebirds,butcannotfly),thisisindicatedfor mustbecomparedagainstallpossiblematchingpieces. thatnode. Efforts have been made to standardize the notations used for semantic networks, although it turned out to be 1. Rules difficult to identify a basic set of general relationships, Therepresentationofknowledgeinfactsandrulesisthe or conceptual dependencies, that is easy to comprehend mostwidelyusedmethod,especiallyforexpertsystems. and powerful enough to be used for different domains. Knowledgeisexpressedthroughcondition–actionpairsin Especiallyforlargernetworks,itbecomesquitedifficult theformofIF...THENrules.Thecondition(theIFpart) forhumanstocomprehendtheknowledgestoredinsucha specifiesthepremisesthatmustholdbeforetherulecanbe network.Computersprocesssuchnetworksbyexamining appliedor“fired”;ifallpartsoftheconditionaresatisfied, thelinksbetweennodesaccordingtoinstructionsforthe the conclusion (the THEN part) can be drawn, and the differenttypesofrelationships. respectiveactioninitiated.Asasimpleexample,consider Intheearly1980s,conceptualgraphsweredeveloped a few rules describing some problems with computers. byJohnSowaasanetwork-basedrepresentationlanguage. Incontrasttosemanticnetworks,relationsarerepresented Rule1: if the screen is dark by a different type of nodes, rather than through labeled then the problem is screen or links.Individualstatementsarerepresentedthroughsepa- cable or video board rateconceptualgraphs,andcomputersprocessconceptual Rule2: if the screen is bright and graphsbycomparingandmanipulatinggraphsfordiffer- the cursor does not move ent statements. As an example, the representation of the then the problem is mouse or sentence “John gives Mary the book” is given in Fig. 1 as cable or software asemanticnetworkandasaconceptualgraph. P1:GAERevisedPages EncyclopediaofPhysicalScienceandTechnology EN001H-27 May7,2001 15:31 614 ArtificialIntelligence FIGURE1 Anexampleofasemanticnetwork(left)andconceptualgraphrepresentation(right). Technically,suchagraph-basedrepresentationscheme erties from “parent” to “child” frames. To some degree, canbetranslatedwithmoderateeffortintoanequivalent frameshavesomesimilaritytoobjectsinobject-oriented rule-basedorpredicatelogicpresentation.Theexplicitvi- programming,althoughthecomputationalandformalun- sualdisplayoftherelationshipsbetweenconcepts,how- derpinningsaredifferent. ever,isofteneasierforhumanstounderstand,atleastfor relativelysmallgraphs.Forthisreason,someAIsystems B. Searching providedifferentviewsoftheknowledgeintheirknowl- Therepresentationofknowledgeinaformatthatissuit- edgebase. ableforacomputer,butstillcanbedisplayedsothathu- manscanunderstanditeasily,isoneessentialrequirement 3. FramesandScripts forthedesignofintelligentsystems.Inordertosolvesome Capturing knowledge through rules quickly can become task,thisknowledgemustbeaccessed,andingeneralpro- rathertediousandmayresultinlargecollectionsofrules cessedinsomeway.Intheconceptually(notnecessarily that are difficult to put together and maintain. Although computationally)simplestcase,aproblemcanbesolved not as rigid as in the case of data bases, in many situa- by identifying one or more solutions among a possibly tionstherearestructuresthatcanbeidentifiedinaknowl- largenumberofcandidates.Thestudyofsearchmethods edgebase,andthatapplytomanyitemsintheknowledge has been one of the cornerstones of AI problem-solving base.Thesestructurescanberepresentedasframes(some- methods,anditisfrequentlypossibletotranslateaprob- timesalsoreferredtoasschemas),suggestedbyMinsky lemthatdoesnotseemtobesolvablebysearchingintoan in 1975 as a data structure to represent abstract situa- equivalentsearchproblem,ortoallowtheextractionofa tions.Arelatedconceptcalledscripts,proposedbyRoger particular aspect that can be solved by searching. Many Schank and collaborators in 1977, captures situations as mind-benders,games,orpuzzlescanbeviewedassearch stereotypicalsequencesofevents.Thefollowingdiagram problems,suchasthe8-puzzle(wheremovabletileswith showsanexampleofaframedescriptionforacomputer numbersorlettershavetobearrangedinaparticularfash- system. ion on a 3×3 board) or cryptarithmetic puzzles (where Frames and scripts make it easier for knowledge en- letters have to be translated into numbers such that the gineers and users to capture and understand the knowl- resultingequationisarithmeticallycorrect).Gameswith edge needed for a particular task. They use a graphical opponents,suchascheckers,chess,Go,orbackgammon, representation that allows the visual display of relations also rely on search algorithms, although these cases are between items (Fig. 2), and some of the work for establish- more complex because of the presence of an opponent, ingaknowledgebasecanbereducedtogeneratingcopies andasinbackgammon,anelementofchance.Examples of frames and filling slots in frames. Thus, complex ob- ofreal-worldproblemsthatcanbesolvedthroughsearch jectscanberepresentedinasingleframe,ratherthanas methods are route-finding (e.g., for trucking companies, a set of rules, or a complex network or graph. Frames airlines,computernetworks),touringandtravelingsales- alsoenabletheassociationofprocedurestoperformspe- personproblems(multiplecitieshavetobevisitedexactly cific operations, and use inheritance to propagate prop- once),VLSIlayout,orrobotnavigation. P1:GAERevisedPages EncyclopediaofPhysicalScienceandTechnology EN001H-27 May7,2001 15:31 ArtificialIntelligence 615 FIGURE2 Anexampleofaframerepresentation. Inthefollowing,webrieflydiscussuninformedorblind formulateheuristicsthatcansteerthesearchtowardmore searchmethods,whichsystematicallygenerateallpossi- promisingcandidates. ble solution candidates until a solution is found, as well asinformedsearchmethods,whereavailableinformation 1. BlindSearch about the problem is used to eliminate unsuitable candi- dates, or to favor the exploration of candidates that are In cases where no additional information about a search morelikelytobeactualsolutions.Computer-basedsearch problem is available, uninformed or blind search meth- methodsgenerallytraversethesearchspace,whichcon- ods have to be employed. These methods generate solu- tains all possible solution candidates, or parts of it in a tioncandidatesinasystematicway,andthustraversethe systematic manner, generating a search tree that lists all searchspaceinaparticularmanner.Twoofthemostbasic thecandidatesalreadyvisitedandhowtheywerereached. ones are breadth-first and depth-first search. In breadth- Anobviouscriterionfortheperformanceofasearchalgo- firstsearch,allthenodesreachablefromthestartingpoint, rithmisthetimeittakestofindasolution(timecomplex- or root, are explored first, then all nodes reachable from ity);theamountofmemoryrequired(spacecomplexity), thefirstsetofnodes,andsoon.Indepth-firstsearch,the however,canalsoposeasevereconstraintonthesuitability first node reachable from the root is explored, then the ofanalgorithm.Inaddition,completeness(isthemethod firstnodereachablefromthatnode,andsoon.Thesetwo guaranteed to find a solution, provided one exists?) and methods are illustrated in Fig. 3; the nodes are numbered optimality(willthemethodidentifythebestamongsev- inthesequenceinwhichtheyarevisited. eral solutions?) may also have to be taken into account. Using such an uninformed or blind search method is Commonmeasuresforsearchperformancearethenum- appropriateiftheonlywaytofindasolutionistosystemat- ber of steps that are needed, and the path cost from the icallycheckallpossibilities.Thesesearchmethods,how- startingpointtoaparticularnode.Ifadditionalinforma- ever, suffer from potentially severe drawbacks. Check- tion about the problem is available, this can be used to ing all possible branches in the tree until one or several P1:GAERevisedPages EncyclopediaofPhysicalScienceandTechnology EN001H-27 May7,2001 15:31 616 ArtificialIntelligence FIGURE3 Breadth-first(left)anddepth-first(right)searchmethods. solutions are found is very expensive in terms of com- blocks away. With the depth-first approach, you would putationtimeandalsowithrespecttothememoryspace followonepathinthecorrespondingsearchtreebysys- requiredforstoringinformationaboutthenodesandpaths tematicallyandconsistentlyselectingonedirectionatan visitedsofar.Forbothmethods,thetimerequiredtofind intersection(e.g.,left,straightahead,orright).Obviously, a solution is exponential with respect to the number of thisstrategyisratherhopeless,sinceyoumaybecircling nodes.Itdependsonthelevelatwhichthesolutionislo- thesameblockoverandover,oryoumayendupwalking cated(thedepthofthesolutiond),themaximumdepthof totheendofthecontinent.Thebreadth-firstapproachis thetreem,andthenumberofoutgoingbranchesforthe guaranteedtohelpyoufindtheconferencecenter,butwith nodes (the branching factor b). For breadth-first search, considerableoverhead.Inthiscase,youwouldexploreall thetimeandthespacecomplexityarebd;fordepth-first, withinadistanceofoneblock,thenthosewithinadistance thetimecomplexityisbmandthespacecomplexityisbm. of three blocks, then four blocks, and so on. In addition Depth-firsthasalowerspacecomplexitythanbreadth-first todoingalotofwalking,youprobablywouldalsoneeda sinceitisnecessarytokeepinformationonlyforthecur- methodtokeeptrackofthelocationsalreadyvisited.With rentlyinvestigatedpath.Ontheotherhand,depth-firsthas aninformedsearchmethod,youcouldusesomeinforma- a problem with paths of infinite length, which may be tion about the relative location of the conference center causedbycyclicalsituationsinthesearchspace.Depth- withrespecttothehotel,andthenexploreonlyaselected firstisnotguaranteedtofindasolution(itisincomplete), numberofblocks.Forhumans,thisapproachisrathernat- andevenifitfindsone,thismaynotbethebestsolution uralandintuitive;forcomputers,however,theadditional (it is not optimal). Breadth-first will find a solution, and informationmustbemadeavailableinanappropriateway it will find the best (shallowest) solution: it is complete sothatitcanbeusedbythesearchmethod.Itshouldbe and optimal. In our considerations so far, we have only notedherethatfindingyourwaywiththehelpofamap, usedthedepthleveltodistinguishbetweenthequalityof whichisprobablyasmartthingtodo,providesyouwith solutions: A solution at a shallower depth is assumed to all the necessary information in advance, and thus does bebetterthanoneatadeeperlevel.Inmanycases,thisis notreallyconstituteasearchproblemanymore. notrealistic,anditmaybeadvantageousornecessaryto Ifitispossibletoobtainanestimateofthecostofreach- associateacostwitheachedgeleadingfromonenodein ing the goal from the current location (for geographical thetreetothenext.Thisinformationisusedbyamethod searchproblems,thiscanbethedistancetothegoal;for known as lowest-cost-first or uniform-cost search. Some othertypesofproblems,othercostmeasurescanbeused), disadvantagesofthesesearchmethodsmaybeovercome asearchmethodknownasgreedyorbest-firstsearchcan byusingvariationssuchasdepth-limited,iterativedeep- be applied. At each branching point, a heuristic is used ening,orbidirectionalsearch,butthegeneralproblemof toestimatethedistancetothegoalforallnodesthatare systematicallyexploringthefullsearchspaceremains. reachablewithinonestep,andtheonewiththelowestesti- mateisselected.Inourexample,anapplicationofgreedy searchwouldbetoselectatanintersectionthestreetlead- 2. InformedSearch inginthedirectionoftheconferencecenter;unlessthere Inmanysituations,additionalinformationabouttheprob- are obstacles such as dead ends, railroad tracks, rivers, lemtobesolvedisavailable,whichcanbeusedtoimprove etc.,greedysearchworksreasonablywell.Ingeneralitis theefficiencyofthesearch.Thegeneralprincipleistofo- averysimpleandeffectivestrategy,althoughitisneither custhesearchonareasofthesearchspacethataremore completenoroptimal.Combininggreedyandlowest-cost- promising than others, or to completely exclude certain firstsearchintoanalgorithmknownasA∗ searchgivesa areas that are guaranteed not to contain a solution. As complete and optimal search method, provided that the an example, let us assume that the task is to walk in an heuristicusedsatisfiessomeformalcriteria.A∗ attempts unfamiliar town from a hotel to conference center a few tofindthebestpathtothegoalbycombiningthepartial P1:GAERevisedPages EncyclopediaofPhysicalScienceandTechnology EN001H-27 May7,2001 15:31 ArtificialIntelligence 617 pathfromthestartingpointtothecurrentlocationwiththe icatelogic,attheexpenseofsomeefficiency.Logicpro- lowestcostsofar,withtheremainingpartthatisestimated gramminglanguages,themostpopularbeingPROLOG, tohavethelowestcosttoreachthegoalfromthecurrent achievebetterefficiencybyrestrictingtheexpressiveness location.AlthoughA∗greatlyreducesthesearchspaceto ofthelanguage,frequentlytoasubsetoffirst-orderpred- betraversed,thenumberofnodestobeexploredcanstill icatelogicknownasHornclauses. beexponential.Inpractice,A∗ tendstorunoutofmem- The simplest form of mathematical logic is proposi- orybecauseitstoresallgeneratednodesinmemory.This tional logic, relying on symbols that represent proposi- problem can be lessened by using variations of A∗ such tions,orfacts.Thesepropositionscanbeconnectedinto as IDA∗ (Iterative Deepening A∗) or SMA∗ (Simplified more complex ones using Boolean connectives such as Memory-bounded A∗); for many practical applications, AND (∧), OR (∨), NOT (¬), or IMPLIES (⇒). These SMA∗offersthebestperformanceamonggeneral-purpose propositionsmustbeformulatedaccordingtothesyntax searchalgorithms. of the language of propositional logic; the semantics of thelanguagespecifieshowthesentencesrelatetothecor- respondingentitiesintherealworld,andthusisusedfor C. Reasoning aninterpretationofthesentences.Inadditiontothesyn- The ability to generate new knowledge by drawing con- tactic and semantic aspects of the language, a calculus clusions from existing knowledge is one of the corner- requires a proof theory, which prescribes how sentences stones of human intelligence and has been an important can be formally deduced from a given set of sentences. challengeinthedomainofartificialintelligence.Reason- The implementation of such a formal system for propo- ingisinextricablylinkedwithknowledgerepresentation, sitionallogicbyacomputerisrelativelystraightforward. relyingonsomepropertiesoftheunderlyingrepresenta- Onemethodistoincrementallyconstructatruthtablefora tion techniques to generate new knowledge. In general, sentencefromitscomponentsbyenumeratingallpossible computers perform reasoning steps based on syntactical combinationsoftruthvaluesforthesymbolsoccurringin aspects of represented knowledge items (such as the ar- thesentence.Unfortunately,thisiscomputationallyvery rangement of words or symbols into logical formulas or expensive, requiring a table with 2n rows for a sentence rules), whereas humans rely more on semantical aspects with n symbols. Other proof methods for propositional (theintendedmeaningofasentence).Thisleadstoafun- logicarealsoexponentialintheworstcase,butformost damental gap in the reasoning capabilities of computers practicalapplications,proofscanbefoundwithreasonable and humans: Computers can perform a large number of effort. Since the expressiveness of propositional logic is individual reasoning steps in a short period of time, but ratherlimited,however,applicationseitherarerelatively theyaresimplyrearrangingandmanipulatingsymbolsac- simple to start with, or require an inordinate amount of cordingtosomereasoningmethod.Humans,ontheother effortforthelogicalspecificationoftheproblem. hand, rely heavily on the meaning and interpretation of Predicate logic offers much greater expressiveness statements and are usually best at performing quick, but throughasubstantiallyricherlanguagethanpropositional possiblyimpreciseandevenincorrect,reasoningsteps.In logic. It uses constant symbols that refer to objects or manycasesinreallife,comingtoafastbutpossiblyinac- concepts in the real world, predicate symbols to specify curateconclusionseemstoworkbetterthandeliberating relationsbetweenobjectsorconcepts,andtermsthatal- overalengthy,formallycorrectone. lowthecompactspecificationofreferencestoobjects.Just asinpropositionallogic,morecomplexsentencescanbe constructed from simple ones with the help of Boolean 1. Logic connectives. In addition to these connectives, predicate Theformulationofgeneralmethodsforreasoningsothat logicalsoutilizesquantifiers,usuallytheuniversalquan- statementscanbeprovencorrectorincorrectinaformal tifierFORALL(∀)andtheexistentialquantifierEXISTS way has long been a goal of mathematical logic. People (∃).Thequantifierscanbeusedtomakegeneralstatements interestedinartificialintelligenceverysoonrealizedthat aboutsetsorsubsetsofobjects;inpropositionallogic,each computersaresuitabletoolfortheimplementationoflog- objectwouldhavetobedescribedbyaseparatestatement. icalreasoningmethods,orcalculi,andsomeoftheearly Predicatelogiccanbeenhancedbyspecialnotationssuch AI systems were indeed reasoning systems for specific as equality, sets, or arithmetic, or for specific purposes applications such as algebra or geometry. In the mean- suchasthedescriptionofsituations,events,orplans.With time, automated theorem provers have been used to for- its good expressiveness, strong formal foundations, and mallyprovemathematicaltheoremsandverifycomputer suitabilityforcomputer-basedevaluation,predicatelogic programs, as well as for other scientific reasoning tasks. is a good general-purpose method for knowledge repre- Thesetheoremproverstypicallyusefullfirst-orderpred- sentation and reasoning. For many specific applications, P1:GAERevisedPages EncyclopediaofPhysicalScienceandTechnology EN001H-27 May7,2001 15:31 618 ArtificialIntelligence andwithrespecttocomputationalefficiency,othermeth- then extend them until a complete sequence of actions ods,inparticularrule-basedexpertsystems,arestillmore leadingtothegoalisfound.Sinceinmanyproblemsdif- widelyusedinpractice. ferentpartsoftheplanusuallydonotinterferewitheach other, this divide-and-conquer strategy generally works quitewell.Asfaraspossible,planningsystemusealeast- 2. Rule-BasedSystems commitment approach, delaying each decision until it is Rule-based systems provide the computational mecha- necessaryorthereisagoodreasonformakingone.These nismsfoundinmostexpertsystems.Knowledgeisspec- delayed decisions, however, may also obscure unresolv- ifiedviafactsandIF–THENrules,andmodusponensis ableconflicts,andtechniqueshavebeendevelopedtodis- used as the underlying inference method to derive new coversuchconflictsasearlyaspossible. conclusions from existing knowledge. These production rules in many cases allow a straightforward encoding of 4. FuzzyLogic expertise about a particular domain, often as situation– actionpairswheretheIFpartoftherulespecifiesaspects Conventionallogicassignsthetruthvalueofeithertrueor ofasituationleadingtooneormoreactionsasdescribed falsetoalogicalsentence,essentiallypaintinga“black- in the THEN part. In principle, the rules and facts in a and-white” world. In the real world, however, there are rule-basedsystemcanbetranslatedintoequivalentlogi- “shades of gray” in between, and a fact can be true to calsentences.Acombinationofrestrictionsandadditional a certain degree, or somebody can believe in a fact to constructs in their language, together with the tight in- a certain degree. Lotfi Zadeh’s fuzzy logic, or fuzzy set tegration between language and evaluation mechanism, theory,usesamembershipfunctionforsetsthatcantake offers substantial practical advantages for rule-based onrealvaluesbetween0and1,andlinguisticvariablesthat systems. describesuchsets.Incontrasttoconventionallogic,where anelementbelongseithertoasetortoitscomplement,in fuzzylogicamembercanbelongtoseveralsetsindifferent 3. Planning degrees (Fig. 4). The problem-solving methods described so far rely on Fuzzy logic also includes rules for constructing com- the generation and identification of states for finding a plexsentencesfromsimpleonesviatheBooleanconnec- solution,orthegenerationofnewknowledgeinorderto tives such as AND, OR, or NOT (¬); a simple version answersomequestionsinaparticulararea.Formanyprob- ofthesecombinationrulesappliestheminimumoftheir lems,asolutionalsorequiresthegenerationofaplanof membershipvaluestosentencesconnectedbyAND,and action that must be executed in order to achieve a goal. the maximum for sentences connected by OR. Starting Again,inprincipleagenerallogic-basedapproach(e.g., in the late 1980s, fuzzy logic has been very successful, situation calculus with a theorem prover) is expressive especiallyinindustrialcontrolapplications. enough for planning problems, but cannot compete in practicewithspecial-purposesystemsthatusearestricted 5. ReasoningwithUncertainty languageandtailoredevaluationmechanisms.Ratherthan generatingallpossiblesituationswiththehopeoffinding Reasoning systems must be capable of coming to con- onethatsatisfiesthegoalcriteria,planningsystemsgen- clusions even if the information available is incomplete, erate partial plans that achieve important subgoals, and inconsistent,unreliable,orotherwiseflawed.Itisclearthat FIGURE4 Thefuzzysetsofsmall,medium,andtallmales.

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.