IntelligentWeb-Mining

人工網(wǎng)絡(luò)代理ArtificialIntelligenceandIntelligentArtificialintelligence(AI)isthescienceofsystematAgentsarecentraltoAI(andviceIntelligentagent=computationalagentthatactsNeedtotechnicallydefinethenotionof“actingintelligently”AI=ScienceofIntelligentSystemsarecalledcomputationalagentsinAI,

智能體是人工智能的核心（反之亦然）–=智能行不討論智能體就談?wù)撊斯ぶ悄軙e(cuò)過重點(diǎn)（反之亦然Whatisan

Anythingthatcanbeviewedasperceivingits

任何可以被看作感知其環(huán)境通過傳感?和

境通過

eyes,ears,and

eyes,ears,and

organsforsensors;hands,legs,mouth,andotherbodypartsfor organsforsensors;hands,legs,mouth,andotherbodypartsforRobotic

camerasandinfraredrangefindersforsensors;variousmotorsfor

距離測距儀for傳感?各 電機(jī)forSoftwareinterfaces,dataintegration,interpretation,data

軟 代Abstractions:Agentsand

Theagentfunctionmapsfrompercepthistoriesto Theagentfunctionmapsfrompercepthistoriesto[f:P*aphysicalarchitecturetoproduce

[f:P*The代理程序運(yùn)行在一個(gè)物理架 =+

ReactivevsGoal-based vsReactivevsGoal-based vsBalancingReactiveandGoal-Orientedchangingconditionsinanappropriatefashion(e.g.,WewantouragentstosystematicallyworktowardslotermgoalsAchievemaximumfreedomofactionifthereisnospecificshotermgoal(e.g.,keepbatteriescharged)

SocialTherealworldisamulti-agentenvironment:wecannintoaccountSocialabilityinagentsistheabilitytointeractwithotcommunicationlanguage…...withthegoaltoletotheragentstomakecommitme(ofothers)orreinforcements(aboutitsownbehavior)

?通信語言)RationalRationalAgent:Foreachpossibleperceptsequenshouldselectanthatisexpectedtomaximizeitslocalperformancewhateverbuilt-inknowledgetheagentRational=IntelligentThereismoretointelligencethanmeets

=智能RationalityisdistinctfromomniscienceRationalityisperceptssoastoobtainusefulinformation(informatioWhatmattersforthe"experience"isperceptsequence(whichtheagentscandetermine),staterepresentation,and

計(jì)算最佳行動通常（具備學(xué)習(xí)和適應(yīng)的能力Agentsactonbehalfofhumans,whospecifytheprovetheirbehaviorisbeneficialtoArtificialintelligence,agents,andAgentsmustactinanethical(andthose,whodevelopagents,possibly…thatthey(theagents)actinanethical

智能體代表人類（智能體）Human-awareAgentsinteractwithSelectedactionsmustmatchhumanSelectedactionsthatareassumedtonotmatchh

Human-aware智能體與人類符合人類期望–也許預(yù)期的動作可) 人類感知型 人類感知型AI系統(tǒng)的挑戰(zhàn)：Subbarao Everextentedperceptsequence(incl.moreorlessexp…withtheaimtobetter(faster)achieveWesay:Agentslearn(andwemean:whileacting,orOptimizeaperformanceSettingupagents’onlinelearningSettingupanagent’sinitialmodelbyexploitingAlsobasicallyoptimizingaperformance

學(xué)習(xí)代理（在線Ever擴(kuò)展的感 exp強(qiáng)化反饋或稀疏(nobigutriseto模型我們說：智能體學(xué)習(xí)(優(yōu)化性能指標(biāo)設(shè)置代理的在線學(xué)習(xí)引擎–專門的知識關(guān)于的知識關(guān)于機(jī)?學(xué)習(xí)–MachineLearning(ML):Statisticsvs.DataSciencevs.Machine“Whenyou’refundraising,it’sAI.Whenyou’rehiring,it’sML.Whenyou’rimplementing,it’slogisticAIismachine

ML

“Whenyou’refundraising,it’sAI.Whenyou’rehiring,it’sML.Whenyou’rimplementing,it’slogistic“contains”machine “contains”machine Machinelearningscales,butcanonlydosoAllfieldshaveevolvedandstilldo

Machinelearningscales,butcanonlydosoAllfieldshaveevolvedandstilldoReactivevsGoal-based vs

fX

f:PClaimingthatfis“anAI”isanindicationoflackofunderstanding……evenifthelastnperceptsare–f:P×…×POneislostw/oanunderstandingofintelligentf:P*

fPàfn個(gè)感知fP?àfP*Frame

Assumethatmachinelearningtechniquesare

Nointeractionw/environment,noAgentisfake(simplyaframearoundstandardasinitialpercepts(butnoactionstowardsgoalsarecomMaybeevenenlighteningforpracticalapplications,butagentidea…

fP?P代理是假的（Also持有when設(shè)置訓(xùn) 作為初始感知（Learning-basedSoftwareThereisnoneedtoconflatemachinelearningwithagentsandAINoneedtoinventframeagentsCanbuildextremelycoolSW/HW(e.g.,forindustrialimageprocessingProbabilisticDifferentialProgrammingTherearecaveats,

à概 微 編 à深度學(xué) Lab(CS5071-BacktotheFuture:Human-guidedArtificialagentsandhumanagents

A

代理在線學(xué)習(xí)–Agentsallowformoreorlesslearning(incl.noNext:Properagentwithno

代理允許或多或少的學(xué)習(xí)（包括不學(xué)習(xí)ProperAgent:An ProperAgent:AnProperAgent:AnCurrentstateoftheSetofactionFindsequenceofactions(afortransformingcurrentstateintogoalSelectfirstaction,andhopethatplancanbe

à（一個(gè)計(jì)劃）àSTRIPSCurrentstateaswellasgoalExample:BlocksOn_Table(A),On_Table(B),On_Block(C,B),On_Block(B,

STRIPS狀態(tài)被建模為地面原子集合（數(shù)據(jù)庫CurrentstateaswellasgoalExample:BlocksOn_Block(C,B),On_Block(B,STRIPS?STRIPS?斯坦福研究所問題解決?（1971 STRIPS=StanfordResearchInstituteProblemSolver STRIPSareCompleteare

Least

LeastRepresentationofPartial-OrderPlanconsistsPlanstepwithpartialorderwhereSi<SjiffSiistobeexecutedbefore

計(jì)劃步驟=STRIPSPlanconsistsSetofvariableassignmentsx=twherexisavariableandtisaconstantorSetofcausalSicSjmeansSicreatesthepreconditioncofSj(impliesSi

iSi為jSjcSTRIPS?斯坦福研究學(xué)院問題解決?（STRIPS?斯坦福研究學(xué)院問題解決?（1971

Scàc

STRIPS=StanfordResearchInstituteProblemSolver CompleteEverypreconditionofastepisforeverylinearizationitholds–?SkwithSi<Sk<Sj,?c?ConsistentIfx=A,thenx≠BfordifferentAandBforvariable

完整?Sjwithc∈SiSkSj，?c?一致 xAxBABx（唯一名稱假設(shè)）completeandconsistentPlanRefinementBackward

粗箭頭=+aftervariablePlanRefinementaftervariable

PlanRefinement...buyattheright

... PlanRefinement...butyoumustget

... PlanRefinementUptonownosearch,butsimple?backwardConflict!Aftergo(HWS)isexecuted,At(Home)nol

ProtectionofCausal ProtectionofCausalS3“threatens“causalrelationbetweenS1andConflictPromotion:PutthreatbeforecausalDemotion:Putthreataftercausal

降級：將威脅放在因果關(guān)系之后AnotherPlanRefinementAssumption:CannotresolveconflictbySelectx=HWS(withcausalrelation)whileinstantiaAt(x)ingo(SM)

AnotherPlanRefinement替代方案go(SMAt(xxHWS（具有因果關(guān)系A(chǔ)notherPlanRefinement TheCompleteTheCompleteSolutionPOPAlgorithm…andcorrecttreatmentPOP算法計(jì)算LastCenturyPlanningSystems(LastUCPOP(Weld,(\h/ai/ucpop.htmlSensoryGraphplan(Weld,Blum,andFurst:(\h/ai/sgp.htmlIPP(K?hlerandNebel:Univ.(https://idw-Prodigy:PlanningandLearning(Veloso:(\h/afs//project/prodigy/Web/prodigy-home.htmlAllsystemshavefoundinteresting

上個(gè)世紀(jì)規(guī)劃系統(tǒng)（LastUCPOP(Weld,(\h/ai/ucpop.htmlSensoryGraphplan(Weld,Blum,andFurst:(\h/ai/sgp.htmlIPP(K?hlerandNebel:Univ.(https://idw-Prodigy:PlanningandLearning(Veloso:(\h/afs//project/prodigy/Web/prodigy-home.htmlPlanningisanActiveFieldofCandealwithAboutstateAbouteffectsofVerypowerfulproblemsolverscanbesetupw/lesseffort/knowledgethanwithmathematical(https://\hwww.ifis.uni-luebeck.de/index.php?id=dski-aktuell-ss20&L=2

系 能dealwithw/less努力/知識 自動