人工智能課件

人工智能

ObjectivesInthisclass,youwilllearnaboutWhatisartificialintelligenceKnowledgerepresentationRecognitiontasksReasoningtasksRoboticsIntroductionto

ArtificialIntelligenceWhatisintelligence?Thecapacitytoacquireandapplyknowledge.Thefacultyofthoughtandreason.Theabilitytolearnorunderstandortodealwithnewortryingsituations.MajorSubdivisionsofAIUnderstandingThinkingActingAI:UnderstandingComputerVision–understandingwhatyouseeAI:ThinkingCapturingStructureandReachingGoalsMachineLearningPlanningClusteringAI:ActingRoboticsConsiderAIuseinonecompanySearchSponseredLinksGoogleNewsGooglemapsIntroductionTuringtestAtestforintelligentbehaviorofmachinesAllowsahumanbeingtointerrogatetwoentities,bothhiddenfromtheinterrogatorAhumanbeingAmachine(acomputer)TheTuringTestIntroduction(continued)Turingtest(continued)Iftheinterrogatorisunabletodeterminewhichentityisthehumanbeingandwhichisthecomputer,thecomputerhaspassedthetestArtificialintelligencecanbethoughtofasconstructingcomputermodelsofhumanintelligenceADivisionofLaborCategoriesoftasksComputationaltasksRecognitiontasksReasoningtasksComputationaltasksTasksforwhichalgorithmicsolutionsexistComputersarebetter(fasterandmoreaccurate)thanhumanbeingsADivisionofLabor(continued)RecognitiontasksSensory/recognition/motor-skillstasksHumanbeingsarebetterthancomputersReasoningtasksRequirealargeamountofknowledgeHumanbeingsarefarbetterthancomputersFigure14.2HumanandComputerCapabilitiesKnowledgeRepresentationKnowledge:AbodyoffactsortruthsForacomputertomakeuseofknowledge,itmustbestoredwithinthecomputerinsomeformKnowledgeRepresentation(continued)KnowledgerepresentationschemesNaturallanguageFormallanguagePictorialGraphicalKnowledgeRepresentation(continued)RequiredcharacteristicsofaknowledgerepresentationschemeAdequacyEfficiencyExtendabilityAppropriatenessRecognitionTasksAneuronisacellinthebraincapableofReceivingstimulifromotherneuronsthroughitsdendritesSendingstimulitootherneuronsthroughitsaxonFigure14.4ANeuronRecognitionTasks(continued)Ifthesumofactivatingandinhibitingstimulireceivedbyaneuronequalsorexceedsitsthresholdvalue,theneuronsendsoutitsownsignalEachneuroncanbethoughtofasanextremelysimplecomputationaldevicewithasingleon/offoutputRecognitionTasks(continued)Humanbrain:AconnectionistarchitectureAlargenumberofsimple“processors”withmultipleinterconnectionsVonNeumannarchitectureAsmallnumber(maybeonlyone)ofverypowerfulprocessorswithalimitednumberofinterconnectionsbetweenthemRecognitionTasks(continued)Artificialneuralnetworks(neuralnetworks)SimulateindividualneuronsinhardwareConnecttheminamassivelyparallelnetworkofsimpledevicesthatactsomewhatlikebiologicalneuronsTheeffectofaneuralnetworkmaybesimulatedinsoftwareonasequential-processingcomputerRecognitionTasks(continued)NeuralnetworkEachneuronhasathresholdvalueIncominglinescarryweightsthatrepresentstimuliTheneuronfireswhenthesumoftheincomingweightsequalsorexceedsitsthresholdvalueAneuralnetworkcanbebuilttorepresenttheexclusiveOR,orXOR,operationFigure14.5OneNeuronwithThreeInputsFigure14.8TheTruthTableforXORRecognitionTasks(continued)NeuralnetworkBoththeknowledgerepresentationand“programming”arestoredasweightsoftheconnectionsandthresholdsoftheneuronsThenetworkcanlearnfromexperiencebymodifyingtheweightsonitsconnectionsReasoningTasksHumanreasoningrequirestheabilitytodrawonalargebodyoffactsandpastexperiencetocometoaconclusionArtificialintelligencespecialiststrytogetcomputerstoemulatethischaracteristicIntelligentSearchingState-spacegraphAfteranyonenodehasbeensearched,thereareahugenumberofnextchoicestotryThereisnoalgorithmtodictatethenextchoiceState-spacesearchFindsasolutionpaththroughastate-spacegraphFigure14.12AState-SpaceGraphwithExponentialGrowthIntelligentSearching(continued)EachnoderepresentsaproblemstateGoalstate:ThestatewearetryingtoreachIntelligentsearchingappliessomeheuristic(oraneducatedguess)toEvaluatethedifferencesbetweenthepresentstateandthegoalstateMovetoanewstatethatminimizesthosedifferencesSwarmIntelligenceSwarmintelligenceModelsthebehaviorofacolonyofantsSwarmintelligencemodelUsessimpleagentsthatOperateindependentlyCansensecertainaspectsoftheirenvironmentCanchangetheirenvironmentMay“evolve”andacquireadditionalcapabilitiesovertimeIntelligentAgentsAnintelligentagent:SoftwarethatinteractscollaborativelywithauserInitiallyanintelligentagentsimplyfollowsusercommandsIntelligentAgents(continued)OvertimeAgentinitiatescommunication,takesaction,andperformstasksonitsownusingitsknowledgeoftheuser’sneedsandpreferencesExpertSystemsRule-basedsystemsAlsocalledexpertsystemsorknowledge-basedsystemsAttempttomimicthehumanabilitytoengagepertinentfactsandcombinetheminalogicalwaytoreachsomeconclusionExpertSystems(continued)Arule-basedsystemmustcontainAknowledgebase:SetoffactsaboutsubjectmatterAninferenceengine:MechanismforselectingrelevantfactsandforreasoningfromtheminalogicalwayManyrule-basedsystemsalsocontainAnexplanationfacility:AllowsusertoseeassertionsandrulesusedinarrivingataconclusionExpertSystems(continued)Afactcanbe

AsimpleassertionArule:Astatementoftheformif...then...Modusponens(methodofassertion)ThereasoningprocessusedbytheinferenceengineExpertSystems(continued)InferenceenginescanproceedthroughForwardchainingBackwardchainingForwardchainingBeginswithassertionsandtriestomatchthoseassertionsto“if”clausesofrules,therebygeneratingnewassertionsExpertSystems(continued)BackwardchainingBeginswithaproposedconclusionTriestomatchitwiththe“then”clausesofrulesThenlooksatthecorresponding“if”clausesTriestomatchthosewithassertionsorwiththe“then”clausesofotherrulesExpertSystems(continued)Arule-basedsystemisbuiltthroughaprocesscalledknowledgeengineering

BuilderofsystemacquiresinformationforknowledgebasefromexpertsinthedomainRoboticsRobot:DevicethatcangathersensoryinformationautonomouslyManyusesforrobots(automanufacturing,bombdisposal,exploration,microsurgery)Deliberativestrategy:RobothasaninternalrepresentationofitsenvironmentReactivestrategy:

Usesheuristicalgorithmstoallowrobottoresponddirectly