Problem Solving in Artificial Intelligence 人工智能中的問題解決

19UCI816-ARTIFICIALINTELLIGENCEANDROBOTICS

UNIT1

ArtificialIntelligenceisamethodofmakingacomputer,acomputer-controlledrobot,orasoftwarethinkintelligentlylikethehumanmind.AIisaccomplishedbystudyingthepatternsofthehumanbrainandbyanalyzingthecognitiveprocess.Theoutcomeofthesestudiesdevelopsintelligentsoftwareandsystems.

Artificialintelligenceallowsmachinestounderstandandachievespecificgoals.AIincludesmachinelearningviadeeplearning.Theformerreferstomachinesautomaticallylearningfromexistingdatawithoutbeingassistedbyhumanbeings.Deeplearningallowsthemachinetoabsorbhugeamountsofunstructureddatasuchastext,images,andaudio.

HistoryofArtificialIntelligence

ArtificialIntelligenceisnotanewwordandnotanewtechnologyforresearchers.Thistechnologyismucholderthanyouwouldimagine.EventherearethemythsofMechanicalmeninAncientGreekandEgyptianMyths.FollowingaresomemilestonesinthehistoryofAIwhichdefinesthejourneyfromtheAIgenerationtotilldatedevelopment.

ThebirthofArtificialIntelligence(1952-1956)

Year1955:

AnAllenNewellandHerbertA.Simoncreatedthe"firstartificialintelligenceprogram"Whichwasnamedas

"LogicTheorist".Thisprogramhadproved38of52Mathematicstheorems,andfindnewandmoreelegantproofsforsometheorems.

Year1956:

Theword"ArtificialIntelligence"firstadoptedbyAmericanComputerscientistJohnMcCarthyattheDartmouthConference.Forthefirsttime,AIcoinedasanacademicfield.

Atthattimehigh-levelcomputerlanguagessuchasFORTRAN,LISP,orCOBOLwereinvented.AndtheenthusiasmforAIwasveryhighatthattime.

Thegoldenyears-Earlyenthusiasm(1956-1974)

Year1966:

Theresearchersemphasizeddevelopingalgorithmswhichcansolvemathematicalproblems.JosephWeizenbaumcreatedthefirstchatbotin1966,whichwasnamedasELIZA.

Year1972:

ThefirstintelligenthumanoidrobotwasbuiltinJapanwhichwasnamedasWABOT-1.

ThefirstAIwinter(1974-1980)

Thedurationbetweenyears1974to1980wasthefirstAIwinterduration.AIwinterreferstothetimeperiodwherecomputerscientistdealtwithasevereshortageoffundingfromgovernmentforAIresearches.

DuringAIwinters,aninterestofpublicityonartificialintelligencewasdecreased.

AboomofAI(1980-1987)

Year1980:

AfterAIwinterduration,AIcamebackwith"ExpertSystem".Expertsystemswereprogrammedthatemulatethedecision-makingabilityofahumanexpert.

IntheYear1980,thefirstnationalconferenceoftheAmericanAssociationofArtificialIntelligence

washeldatStanfordUniversity.

ThesecondAIwinter(1987-1993)

Thedurationbetweentheyears1987to1993wasthesecondAIWinterduration.

AgainInvestorsandgovernmentstoppedinfundingforAIresearchasduetohighcostbutnotefficientresult.TheexpertsystemsuchasXCONwasverycosteffective.

Theemergenceofintelligentagents(1993-2011)

Year1997:

Intheyear1997,IBMDeepBluebeatsworldchesschampion,GaryKasparov,andbecamethefirstcomputertobeataworldchesschampion.

Year2002:

forthefirsttime,AIenteredthehomeintheformofRoomba,avacuumcleaner.

Year2006:

AIcameintheBusinessworldtilltheyear2006.CompanieslikeFacebook,Twitter,andNetflixalsostartedusingAI.

Deeplearning,bigdataandartificialgeneralintelligence(2011-present)

Year2011:

Intheyear2011,IBM'sWatsonwonjeopardy,aquizshow,whereithadtosolvethecomplexquestionsaswellasriddles.Watsonhadprovedthatitcouldunderstandnaturallanguageandcansolvetrickyquestionsquickly.

Year2012:

GooglehaslaunchedanAndroidappfeature"Googlenow",whichwasabletoprovideinformationtotheuserasaprediction.

Year2014:

Intheyear2014,Chatbot"EugeneGoostman"wonacompetitionintheinfamous"Turingtest."

Year2018:

The"ProjectDebater"fromIBMdebatedoncomplextopicswithtwomasterdebatersandalsoperformedextremelywell.

GooglehasdemonstratedanAIprogram"Duplex"whichwasavirtualassistantandwhichhadtakenhairdresserappointmentoncall,andladyonothersidedidn'tnoticethatshewastalkingwiththemachine.

NowAIhasdevelopedtoaremarkablelevel.TheconceptofDeeplearning,bigdata,anddatasciencearenowtrendinglikeaboom.NowadayscompanieslikeGoogle,Facebook,IBM,andAmazonareworkingwithAIandcreatingamazingdevices.ThefutureofArtificialIntelligenceisinspiringandwillcomewithhighintelligence.

Actinghumanly

ThefirstproposalforsuccessinbuildingaprogramandactshumanlywastheTuringTest.Tobeconsideredintelligentaprogrammustbeabletoactsufficientlylikeahumantofoolaninterrogator.Ahumaninterrogatestheprogramandanotherhumanviaaterminalsimultaneously.Ifafterareasonableperiod,theinterrogatorcannottellwhichiswhich,theprogrampasses.

Topassthistestrequires:

naturallanguageprocessing

knowledgerepresentation

automatedreasoning

machinelearning

Thistestavoidsphysicalcontactandconcentrateson"higherlevel"mentalfaculties.A

total

Turingtestwouldrequiretheprogramtoalsodo:

computervision

robotics

ThinkingHumanly

Thisrequires"gettinginside"ofthehumanmindtoseehowitworksandthencomparingourcomputerprogramstothis.Thisiswhat

cognitive

science

attemptstodo.Anotherwaytodothisistoobserveahumanproblemsolvingandarguethatone'sprogramsgoaboutproblemsolvinginasimilarway.

Example:

GPS(GeneralProblemSolver)wasanearlycomputerprogramthatattemptedtomodelhumanthinking.ThedeveloperswerenotsomuchinterestedinwhetherornotGPSsolvedproblemscorrectly.Theyweremoreinterestedinshowingthatitsolvedproblemslikepeople,goingthroughthesamestepsandtakingaroundthesameamountoftimetoperformthosesteps.

ThinkingRationally

Aristotlewasoneofthefirsttoattempttocodify"thinking".His

syllogisms

providedpatternsofargumentstructurethatalwaysgavecorrectconclusions,givingcorrectpremises.

Example:Allcomputersuseenergy.Usingenergyalwaysgeneratesheat.Therefore,allcomputersgenerateheat.

Thisinitiatethefieldof

logic.Formallogicwasdevelopedinthelatenineteenthcentury.Thiswasthefirststeptowardenablingcomputerprogramstoreasonlogically.

By1965,programsexistedthatcould,givenenoughtimeandmemory,takeadescriptionoftheprobleminlogicalnotationandfindthesolution,ifoneexisted.The

logicist

traditioninAIhopestobuildonsuchprogramstocreateintelligence.

Therearetwomainobstaclestothisapproach:First,itisdifficulttomakeinformalknowledgepreciseenoughtousethelogicistapproachparticularlywhenthereisuncertaintyintheknowledge.Second,thereisabigdifferencebetweenbeingabletosolveaprobleminprincipleanddoingsoinpractice.

ActingRationally:Therationalagentapproach

Actingrationallymeansactingsoastoachieveone'sgoals,givenone'sbeliefs.An

agent

isjustsomethingthatperceivesandacts.

InthelogicalapproachtoAI,theemphasisisoncorrectinferences.Thisisoftenpartofbeingarationalagentbecauseonewaytoactrationallyistoreasonlogicallyandthenactononesconclusions.Butthisisnotallofrationalitybecauseagentsoftenfindthemselvesinsituationswherethereisnoprovablycorrectthingtodo,yettheymustdosomething.

Therearealsowaystoactrationallythatdonotseemtoinvolveinference,e.g.,reflexactions.

ThestudyofAIasrationalagentdesignhastwoadvantages:

Itismoregeneralthanthelogicalapproachbecausecorrectinferenceisonlyausefulmechanismforachievingrationality,notanecessaryone.

Itismoreamenabletoscientificdevelopmentthanapproachesbasedonhumanbehaviourorhumanthoughtbecauseastandardofrationalitycanbedefinedindependentofhumans.

Achievingperfectrationalityincomplexenvironmentsisnotpossiblebecausethecomputationaldemandsaretoohigh.However,wewillstudyperfectrationalityasastartingplace.

cognitivemodeling

Cognitivemodellingisanareaofcomputersciencethatdealswithsimulatinghumanproblem-solvingandmentalprocessinginacomputerizedmodel.Suchamodelcanbeusedtosimulateorpredicthumanbehaviourorperformanceontaskssimilartotheonesmodelledandimprovehuman-computerinteraction

Cognitivemodellingisusedinnumerousartificialintelligence(

AI

)applications,suchas

expertsystems

,

naturallanguageprocessing

,

neuralnetworks

,andinroboticsandvirtualrealityapplications.Cognitivemodelsarealsousedtoimproveproductsinmanufacturingsegments,suchas

humanfactors

,engineering,andcomputergameanduserinterfacedesign.

Anadvancedapplicationofcognitivemodellingisthecreationofcognitivemachines,whichareAIprogramsthatapproximatesomeareasofhumancognition.OneofthegoalsofSandia'sprojectistomakehuman-computerinteractionmorelikeaninteractionbetweentwohumans.

Typesofcognitivemodels

Somehighlysophisticatedprogramsmodelspecificintellectualprocesses.Techniquessuchasdiscrepancydetectionareusedtoimprovethesecomplexmodels.

Discrepancydetectionsystemssignalwhenthereisadifferencebetweenanindividual'sactualstateorbehaviorandtheexpectedstateorbehaviorasperthecognitivemodel.Thatinformationisthenusedtoincreasethecomplexityofthemodel.

Anothertypeofcognitivemodelistheneuralnetwork.Thismodelwasfirsthypothesizedinthe1940s,butithasonlyrecentlybecomepracticalthankstoadvancementsindataprocessingandtheaccumulationoflargeamountsofdatatotrain

algorithms

.

Neuralnetworksworksimilarlytothehumanbrainbyrunningtrainingdatathroughalargenumberofcomputationalnodes,calledartificialneurons,whichpassinformationbackandforthbetweeneachother.Byaccumulatinginformationinthisdistributedway,applicationscanmakepredictionsaboutfutureinputs.

R

einforcementlearning

isanincreasinglyprominentareaofcognitivemodeling.Thisapproachhasalgorithmsrunthroughmanyiterationsofataskthattakesmultiplesteps,incentivizingactionsthateventuallyproducepositiveoutcomes,whilepenalizingactionsthatleadtonegativeones.ThisisaprimarypartoftheAIalgorithmthatGoogle's

DeepMind

usedforitsAlphaGoapplication,whichbestedthetophumanGoplayersin2016

Thesemodels,whichcanalsobeusedinnaturallanguageprocessingandsmartassistantapplications,haveimprovedhuman-computerinteraction,makingitpossibleformachinestohaverudimentaryconversationswithhumans.

AgentsinArtificialIntelligence

AnAIsystemcanbedefinedasthestudyoftherationalagentanditsenvironment.Theagentssensetheenvironmentthroughsensorsandactontheirenvironmentthroughactuators.AnAIagentcanhavementalpropertiessuchasknowledge,belief,intention,etc.

WhatisanAgent?

Anagentcanbeanythingthatperceiveitsenvironmentthroughsensorsandactuponthatenvironmentthroughactuators.AnAgentrunsinthecycleof

perceiving,

thinking,and

acting.Anagentcanbe:

Human-Agent:

Ahumanagenthaseyes,ears,andotherorganswhichworkforsensorsandhand,legs,vocaltractworkforactuators.

RoboticAgent:

Aroboticagentcanhavecameras,infraredrangefinder,NLPforsensorsandvariousmotorsforactuators.

SoftwareAgent:

Softwareagentcanhavekeystrokes,filecontentsassensoryinputandactonthoseinputsanddisplayoutputonthescreen.

Sensor:

Sensorisadevicewhichdetectsthechangeintheenvironmentandsendstheinformationtootherelectronicdevices.Anagentobservesitsenvironmentthroughsensors.

Actuators:

Actuatorsarethecomponentofmachinesthatconvertsenergyintomotion.Theactuatorsareonlyresponsibleformovingandcontrollingasystem.Anactuatorcanbeanelectricmotor,gears,rails,etc.

Effectors:

Effectorsarethedeviceswhichaffecttheenvironment.Effectorscanbelegs,wheels,arms,fingers,wings,fins,anddisplayscreen.

IntelligentAgents:

Anintelligentagentisanautonomousentitywhichactsuponanenvironmentusingsensorsandactuatorsforachievinggoals.Anintelligentagentmaylearnfromtheenvironmenttoachievetheirgoals.Athermostatisanexampleofanintelligentagent.

FollowingarethemainfourrulesforanAIagent:

Rule1:

AnAIagentmusthavetheabilitytoperceivetheenvironment.

Rule2:

Theobservationmustbeusedtomakedecisions.

Rule3:

Decisionshouldresultinanaction.

Rule4:

TheactiontakenbyanAIagentmustbearationalaction.

RationalAgent:

Arationalagentisanagentwhichhasclearpreference,modelsuncertainty,andactsinawaytomaximizeitsperformancemeasurewithallpossibleactions.

Arationalagentissaidtoperformtherightthings.AIisaboutcreatingrationalagentstouseforgametheoryanddecisiontheoryforvariousreal-worldscenarios.

ForanAIagent,therationalactionismostimportantbecauseinAIreinforcementlearningalgorithm,foreachbestpossibleaction,agentgetsthepositiverewardandforeachwrongaction,anagentgetsanegativereward.

StructureofanAIAgent

ThetaskofAIistodesignanagentprogramwhichimplementstheagentfunction.Thestructureofanintelligentagentisacombinationofarchitectureandagentprogram.Itcanbeviewedas:

Agent

=

Architecture

+

Agent

program

FollowingarethemainthreetermsinvolvedinthestructureofanAIagent:

Architecture:

ArchitectureismachinerythatanAIagentexecuteson.

AgentFunction:

Agentfunctionisusedtomapapercepttoanaction.

ExampleofAgentswiththeirPEASrepresentation

Agent

Performancemeasure

Environment

Actuators

Sensors

1.MedicalDiagnose

Healthypatient

Minimizedcost

Patient

Hospital

Staff

Tests

Treatments

Keyboard

(Entryofsymptoms)

2.VacuumCleaner

Cleanness

Efficiency

Batterylife

Security

Room

Table

Woodfloor

Carpet

Variousobstacles

Wheels

Brushes

VacuumExtractor

Camera

Dirtdetectionsensor

Cliffsensor

BumpSensor

InfraredWallSensor

3.Part-pickingRobot

Percentageofpartsincorrectbins.

Conveyorbeltwithparts,

Bins

JointedArms

Hand

Camera

Jointanglesensors.

ProblemSolvinginArtificialIntelligence

ThereflexagentofAIdirectlymapsstatesintoaction.Whenevertheseagentsfailtooperateinanenvironmentwherethestateofmappingistoolargeandnoteasilyperformedbytheagent,thenthestatedproblemdissolvesandsenttoaproblem-solvingdomainwhichbreaksthelargestoredproblemintothesmallerstorageareaandresolvesonebyone.Thefinalintegratedactionwillbethedesiredoutcomes.

Onthebasisoftheproblemandtheirworkingdomain,differenttypesofproblem-solvingagentdefinedanduseatanatomiclevelwithoutanyinternalstatevisiblewithaproblem-solvingalgorithm.Theproblem-solvingagentperformspreciselybydefiningproblemsandseveralsolutions.Sowecansaythatproblemsolvingisapartofartificialintelligencethatencompassesanumberoftechniquessuchasatree,B-tree,heuristicalgorithmstosolveaproblem.

Wecanalsosaythataproblem-solvingagentisaresult-drivenagentandalwaysfocusesonsatisfyingthegoals.

Stepsproblem-solvinginAI:

TheproblemofAIisdirectlyassociatedwiththenatureofhumansandtheiractivities.Soweneedanumberoffinitestepstosolveaproblemwhichmakeshumaneasyworks.

Thesearethefollowingstepswhichrequiresolvingaproblem:

GoalFormulation:

Thisoneisthefirstandsimplestepinproblem-solving.Itorganizesfinitestepstoformulatetarget/goalswhichrequiresomeactiontoachievethegoal.TodaytheformulationofthegoalisbasedonAIagents.

Problemformulation:

Itisoneofthecorestepsofproblem-solvingwhichdecideswhatactionshouldbetakentoachievetheformulatedgoal.InAIthiscorepartisdependentuponsoftwareagentwhichconsistedofthefollowingcomponentstoformulatetheassociatedproblem.

Componentstoformulatetheassociatedproblem:

InitialState:

ThisstaterequiresaninitialstatefortheproblemwhichstartstheAIagenttowardsaspecifiedgoal.Inthisstatenewmethodsalsoinitializeproblemdomainsolvingbyaspecificclass.

Action:

Thisstageofproblemformulationworkswithfunctionwithaspecificclasstakenfromtheinitialstateandallpossibleactionsdoneinthisstage.

Transition:

Thisstageofproblemformulationintegratestheactualactiondonebythepreviousactionstageandcollectsthefinalstagetoforwardittotheirnextstage.

Goaltest:

Thisstagedeterminesthatthespecifiedgoalachievedbytheintegratedtransitionmodelornot,wheneverthegoalachievesstoptheactionandforwardintothenextstagetodeterminesthecosttoachievethegoal.

Pathcosting:

Thiscomponentofproblem-solvingnumericalassignedwhatwillbethecosttoachievethegoal.Itrequiresallhardwaresoftwareandhumanworkingcost.

Typesofsearchalgorithms:

Therearefortoomanypowerfulsearchalgorithmsouttheretofitinasinglearticle.Instead,thisarticlewilldiscuss

six

ofthefundamentalsearchalgorithms,dividedinto

two

categories,asshownbelow.

UninformedSearchAlgorithms:

Thesearchalgorithmsinthissectionhavenoadditionalinformationonthegoalnodeotherthantheoneprovidedintheproblemdefinition.Theplanstoreachthegoalstatefromthestartstatedifferonlybytheorderand/orlengthofactions.Uninformedsearchisalsocalled

Blindsearch.

Thesealgorithmscanonlygeneratethesuccessorsanddifferentiatebetweenthegoalstateandnongoalstate.

Thefollowinguninformedsearchalgorithmsarediscussedinthissection.

DepthFirstSearch

BreadthFirstSearch

UniformCostSearch

Eachofthesealgorithmswillhave:

Aproblem

graph,

containingthestartnodeSandthegoalnodeG.

A

strategy,

describingthemannerinwhichthegraphwillbetraversedtogettoG.

A

fringe,

whichisadatastructureusedtostoreallthepossiblestates(nodes)thatyoucangofromthecurrentstates.

A

tree,

thatresultswhiletraversingtothegoalnode.

Asolution

plan,

whichthesequenceofnodesfromStoG.

DepthFirstSearch

:

Depth-firstsearch(DFS)isanalgorithmfortraversingorsearchingtreeorgraphdatastructures.Thealgorithmstartsattherootnode(selectingsomearbitrarynodeastherootnodeinthecaseofagraph)andexploresasfaraspossiblealongeachbranchbeforebacktracking.

Ituseslastin-first-outstrategyandhenceitisimplementedusingastack.

Example:

Question.

WhichsolutionwouldDFSfindtomovefromnodeStonodeGifrunonthegraphbelow?

Solution.

Theequivalentsearchtreefortheabovegraphisasfollows.AsDFStraversesthetree“deepestnodefirst”,itwouldalwayspickthedeeperbranchuntilitreachesthesolution(oritrunsoutofnodes,andgoestothenextbranch).Thetraversalisshowninbluearrows.

Path:

?S->A->B->C->G

Breadth-firstsearch(BFS)isanalgorithmfortraversingorsearchingtreeorgraphdatastructures.Itstartsatthetreeroot(orsomearbitrarynodeofagraph,sometimesreferredtoasa‘searchkey’),andexploresalloftheneighbornodesatthepresentdepthpriortomovingontothenodesatthenextdepthlevel.

Itisimplementedusingaqueue.

Example:

Question.

WhichsolutionwouldBFSfindtomovefromnodeStonodeGifrunonthegraphbelow?

Solution.

Theequivalentsearchtreefortheabovegraphisasfollows.AsBFStraversesthetree“shallowestnodefirst”,itwouldalwayspicktheshallowerbranchuntilitreachesthesolution(oritrunsoutofnodes,andgoestothenextbranch).Thetraversalisshowninbluearrows.

Path:

S->D->G

InformedSearchingAlgorithms

Informedsearchalgorithmscontaininformationaboutthegoalstate.Thiswillhelpinmoreefficientsearching.Itcontainsanarrayofknowledgeabouthowcloseisthegoalstatetothepresentstate,pathcost,howtoreachthegoal,etc.Informedsearchalgorithmsareusefulinlargedatabaseswhereuninformedsearchalgorithmscan’tmakeanaccurateresult.

Informedsearchalgorithmsarealsocalledheuristicsearchsinceitusestheideaofheuristics.

Theheuristicfunctionisafunctionusedtomeasuretheclosenessofthecurrentstatetothegoalstateandheuristicpropertiesareusedtofindoutthebestpossiblepathtoreachthegoalstateconcerningthepathcost.

ConsideranexampleofsearchingaplaceyouwanttovisitonGooglemaps.Thecurrentlocationandthedestinationplacearegiventothesearchalgorithmforcalculatingtheaccuratedistance,timetaken,andreal-timetrafficupdatesonthatparticularroute.Thisisexecutedusinginformedsearchalgorithms.

InformedSearchAlgorithms:

Here,thealgorithmshaveinformationonthegoalstate,whichhelpsinmoreefficientsearching.Thisinformationisobtainedbysomethingcalleda

heuristic.

Inthissection,wewilldiscussthefollowingsearchalgorithms.

GreedySearch

A*TreeSearch

A*GraphSearch

SearchHeuristics:

Inaninformedsearch,aheuristicisa

function

thatestimateshowcloseastateistothegoalstate.Forexample–Manhattandistance,Euclideandistance,etc.(Lesserthedistance,closerthegoal.)Differentheuristicsareusedindifferentinformedalgorithmsdiscussedbelow.

GreedySearch:

Ingreedysearch,weexpandthenodeclosesttothegoalnode.The“closeness”isestimatedbyaheuristich(x).

Heuristic:

Aheuristichisdefinedas-

h(x)=Estimateofdistanceofnodexfromthegoalnode.

Lowerthevalueofh(x),closeristhenodefromthegoal.

Strategy:

Expandthenodeclosesttothegoalstate,

i.e.

expandthenodewithalowerhvalue.

Example:

Question.

FindthepathfromStoGusinggreedysearch.Theheuristicvalueshofeachnodebelowthenameofthenode.

Solution.

StartingfromS,wecantraversetoA(h=9)orD(h=5).WechooseD,asithasthelowerheuristiccost.NowfromD,wecanmovetoB(h=4)orE(h=3).WechooseEwithalowerheuristiccost.Finally,fromE,wegotoG(h=0).Thisentiretraversalisshowninthesearchtreebelow,inblue.

Path:

?S->D->E->G

Advantage:

Workswellwithinformedsearchproblems,withfewerstepstoreachagoal.

Disadvantage:

CanturnintounguidedDFSintheworstcase.

A*TreeSearch:

A*TreeSearch,orsimplyknownasA*Search,combinesthestrengthsofuniform-costsearchandgreedysearch.Inthissearch,theheuristicisthesummationofthecostinUCS,denotedbyg(x),andthecostinthegreedysearch,denotedbyh(x).Thesummedcostisdenotedbyf(x).

Heuristic:

ThefollowingpointsshouldbenotedwrtheuristicsinA*search.

Here,h(x)iscalledthe

forwardcost

andisanestimateofthedistanceofthecurrentnodefromthegoalnode.

And,g(x)iscalledthe

backwardcost

andisthecumulativecostofanodefromtherootnode.

A*searchisoptimalonlywhenforallnodes,theforwardcostforanodeh(x)underestimatestheactualcosth*(x)toreachthegoal.Thispropertyof

A*

heuristiciscalled

admissibility.

Admissibility:?

Strategy:

Choosethenodewiththelowestf(x)value.

Example:

Question.

FindthepathtoreachfromStoGusingA*search.

Solution.

StartingfromS,thealgorithmcomputesg(x)+h(x)forallnodesinthefringeateachstep,choosingthenod