Lectures TDDD10AIProgramming CooperationAndCoordination2 CyrilleBerger 1AIProgramming:Introduction 2Introductionto 3AgentsandAgents 4Multi-Agentand 5Multi-AgentDecision 6CooperationAndCoordination 7CooperationAndCoordination2 8MachineLearning 9AutomatedPlanning 10PuttingItAll 2/73 Lecturecontent TaskAllocation AssignmentProblems LinearSumAssignmentProblemFormulation andHungarianMethod SerialAssignmentProblem ContraintSatisfactionsProblems CentralizedContraintSatisfactionsProblems DistributedConstraintOptimizationProblems Task-SpecificationTrees 3/73 TaskAllocation AssignmentProblems Assigningnagentstomtasks Forexample,passengerstodrivers(carsharingdomain) Cleaningrobotstorooms Weddingmatchingwithindividualsympathymeasure Ambulancestocivilians,Firebrigadtofire … AssignmentProblems IndividualAssignmentCostscᵢⱼoccurfor eachassignmentofagentitotaskj Theobjectiveistofindaone-to-one matchingfromagentstotasksattheleast possibletotalcost 6 AssignmentProblems-Example AssignmentProblems-ExampleDynamicRideSharing GivenannxmcostmatrixC=(cᵢⱼ)match eachrowtoadifferentcolumninsucha waythatthesumofthecorresponding entriesisminimized Example:Fourpersonshavetobeassignedto 4cleaningjobs,eachonehasdifferentcosts: Thesingletripsof 4drivers BathroomFloorsWindowsKitchen Gabi 8€ 10€ 17€ 9€ Malte 3€ 8€ 5€ 6€ Tom 10€ 12€ 11€ 9€ Patrick 6€ 13€ 9€ 7€ Theirsharedrides computedbyanAPsolver Costscᵢⱼareassumednon7 8 DifferentRepresentationsofAssignmentProblems PermutationMatrix BijectiveMappingbetween twosets! Representationas Note:Therearen!validsolutions. Forexamplen=10,10!=3.63e6 Matrix Representation Graph Representation 9 10 AssignmentProblemFormulation AssignmentMatrix(xij): Formulation: LinearSumAssignmentProblemFormulationandHungarianMethod 12 HeterogenousAgents HungarianMethod Whathappenifnotallagentscanaccomplish thetasks TheHungarianMethodforsolvingLinear SumAssignmentProblems Theorem1: Whenadding(orsubtracting)aconstanttoeveryelementof anyrow(orcolumn)ofthecostmatrix(cᵢⱼ)thenanassignment whichminimizesthetotalcostforthenewmatrixwillalso minimizetheoriginalcostmatrix. Theorem2: Ifallcᵢⱼ≥0andthereexistsasolutionsuchthat∑cᵢⱼxᵢⱼ=0then thesolutionisoptimal. Wherexᵢⱼ=1ifagentiisassignedfortaskj, qᵢⱼ=1ifagentiisqualifiedfortaskj,andcᵢⱼthe TheHungarianMethod(originalversion) solvesLSAPsinO(n⁴)(KuhnandMunkers) 13 HungarianMethod(1/7) Step1:RowReduction 14 HungarianMethod(2/7) Step2:ColumnReduction Subtracttheminimumentryofeachrowfrom allentriesinthisrow Subtracttheminimumentryofeachcolumn fromallentriesinthiscolumn BathroomFloorsWindowsKitchen Gabi 8 10 17 9 Malte 3 8 5 6 Tom 10 12 11 9 Patrick6 13 9 7 BathroomFloorsWindowsKitchen Gabi 0 2 9 1 Malte 0 5 2 3 Tom 1 3 2 0 Patrick0 7 3 1 BathroomFloorsWindowsKitchen Gabi 0 2 9 1 Malte 0 5 2 3 Tom 1 3 2 0 Patrick0 7 3 1 BathroomFloorsWindowsKitchen Gabi 0 0 7 1 Malte 0 3 0 3 Tom 1 1 0 0 Patrick0 5 1 1 15 16 HungarianMethod(3/7) Step3:ZeroAssignment HungarianMethod(4/7) Finally,allzerosareeithercrossedoutor assignedandthenumberofassigned zerosisnandthusthesolutionisoptimal! Rows:Examineeachrow(startingwiththefirstone)untilfindingarowthat containsexactlyonezero.Markthiszeroastemporaryassignmentandcross allentriesinthecolumnwheretheassignmenthasbeenmade. BathroomFloorsWindowsKitchen Gabi 0 0 7 1 Malte 0 3 0 3 Tom 1 1 0 0 Patrick 0 5 1 1 BathroomFloorsWindowsKitchen Gabi 0 0 7 1 Malte 0 3 0 3 Tom 1 1 0 0 Patrick 0 5 1 1 Bathroom FloorsWindowsKitchen Gabi 0 0 7 1 Malte 0 3 0 3 Tom 1 1 0 0 Patrick0 5 1 1 Columns:Examineeachcolumn(startingwiththefirstone)untilfindinga columnthatcontainsexactlyonezero.Markthiszeroastemporaryassignment andcrossallentriesintherowwheretheassignmenthasbeenmade. Result:Gabidoesthefloors,Maltedoes thewindows,Tomdoesthekitchen,and Patrickdoesthebathroom. 17 HungarianMethod-Anotherexample(1/2) 18 HungarianMethod-Anotherexample(2/2) 5driverscanpick-uppassengersfrom5areasatdifferentcosts Step3a:ZeroRowAssignment Wiehre HerdernMerzhausenZähringenStühlinger Gabi 6 12 3 11 15 Malte 4 2 7 1 10 Tom 8 11 10 7 11 Patrick 16 19 12 23 21 Robert 9 5 7 6 10 WiehreHerdernMerzhausenZähringenStühlinger Gabi 2 9 0 8 8 Malte 2 1 6 0 5 Tom 0 4 3 0 0 Patrick3 7 0 11 5 Robert 3 0 2 1 1 Step1:Rowreduction Wiehre HerdernMerzhausenZähringenStühlinger Gabi 3 9 0 8 12 Malte 3 1 6 0 9 Tom 1 4 3 0 4 Patrick 4 7 0 11 9 Robert 4 0 2 1 5 Step3b:ZeroColumnAssignment WiehreHerdernMerzhausenZähringenStühlinger Gabi 2 9 0 8 8 Malte 2 1 6 0 5 Tom 0 4 3 0 0 Patrick3 7 0 11 5 Robert 3 0 2 1 1 Step2:Columnreduction Wiehre HerdernMerzhausenZähringenStühlinger Gabi 2 9 0 8 8 Malte 2 1 6 0 5 Tom 0 4 3 0 0 Patrick 3 7 0 11 5 Robert 3 0 2 1 1 19 20 HungarianMethod(5/7) Step4:Drawtheminimumnumberoflinestocoverallzero’s (a)Markallrowsinwhichtheassignmenthasnotbeendone HungarianMethod(6/7) Step5:Selectthesmallestelementfromthe uncoveredelements (a)Subtractthiselementfromalluncoveredones WiehreHerdernMerzhausenZähringenStühlinger Gabi 2 9 0 8 8 Malte 2 1 6 0 5 Tom 0 4 3 0 0 Patrick3 7 0 11 5 Robert3 0 2 1 1 WiehreHerdernMerzhausen Zähringen Stühlinger Gabi 2 9 0 8 8 Malte 2 1 6 0 5 Tom 0 4 3 0 0 Patrick 3 7 0 11 5 Robert 3 0 2 1 1 (b)Seethepositionofzerosinthemarkedrowsandmarkthecorrespondingcolumns WiehreHerdernMerzhausenZähringenStühlinger Gabi 2 9 0 8 8 Malte 2 1 6 0 5 Tom 0 4 3 0 0 Patrick3 7 0 11 5 Robert3 0 2 1 1 (b)Addthiselementtoallelementswhichareattheintersection oftwolines WiehreHerdernMerzhausen Zähringen Stühlinger Gabi 0 7 0 6 6 Malte 2 1 6 0 5 Tom 0 4 3 0 0 Patrick 1 5 0 9 3 Robert 3 0 2 1 1 (c)Markallotherrowswithazerointhosecolumns WiehreHerdernMerzhausenZähringenStühlinger Gabi 2 9 0 8 8 Malte 2 1 6 0 5 Tom 0 4 3 0 0 Patrick3 7 0 11 5 Robert3 0 2 1 1 21 HungarianMethod(7/7) 22 PracticalconsiderationsabouttheHungarianMethod Negativecostscanbeconvertedby addingtoeachelementofCthevalue v=-min{cij} Theobjectivefunctionistobe Step6:Nowwehave increasedthenumberofzeros RepeatStep3. WiehreHerdernMerzhausenZähringen Stühlinger Gabi 0 7 0 6 6 Malte 2 1 6 0 5 Tom 0 4 3 0 0 Patrick1 5 0 9 3 Robert 3 0 2 1 1 Multiplyallcijby-1or replaceeachcijbycmax-cijwherecmax=max{cij} Therearemoreagents(n>m)or moretasks(m>n) addasufficientnumberof‘‘dummy’’tasksoragents (whicheverisinshortersupply)withcostsof0(or inf) 23 24 Multi-RobotExploration Arewereallyinterestedin minimizingthesumofindividual costs? SerialAssignmentProblem 25 TheLinearBottleneckAssignmentProblem(1/2) TheLinearBottleneckAssignmentProblem(2/2) TheLinearSumAssignmentProblem minimizesthesummedtotalcostasif tasksareexecutedinseries Whenwehaveparallelmachines(MAS) wewanttominimizethetotalcompletion timeandthusthemaximaldurationofall parallelexecutedjobs! Wewanttominimizetheso-called bottleneckobjectivefunction: Canbesolvedin polynomialtime,using thethresholding algorithm. Whatabout dependency betweentasks? max{1≤i≤n}(cᵢφ(i)) 27 28 MultilevelGeneralizedAssignmentProblems(1/2) MultilevelGeneralizedAssignmentProblems(2/2) Appearsinthecontextoflarge-scaletask allocation,e.g.,RoboCupRescue Simulation! Theproblemofassigningnagentsto mtasksatllevels(i.e.timeslots) Theresourcerequirement(consumption) aᵢⱼₖdependentsontheagent,task,and level Asingleagentimayhavemorethan onetaskassignedbutthesumof neededresourcesaᵢⱼₖmaynotexceedbᵢ Therearealsocostscᵢⱼₖofthe ThisisNP-hard,anddeterminingwhetherafeasiblesolutionexistsisNPcomplete. AllMultilevelversionsofLinearSumandLinearBottleneckAssignment Problemsareintractable! 29 30 CaseStudy:Task-SequenceAllocationwithHumanintheLoop(1/3) CaseStudy:Task-SequenceAllocationwithHumanintheLoop(2/3) Coordinationofresponseteamsby resourceallocation Humanassistedtasksequenceassignment MixedIntegerLinearProgramming (MILP)solverbiasedbyhuman expertise TwoAdvantages:Fasterassignment computationwithoptimalsequencing 31 32 CaseStudy:Task-SequenceAllocationwithHumanintheLoop(3/3) ContraintSatisfactionsProblems 33 ConstraintOptimizationProblems Applications Insearchproblems,thestatedoesnot haveastructure(everythingisinthedata structure).InCSPs,statesareexplicitly representedasvariableassignments. ACSPconsistsof Timetabling(classes, rooms,times) Configuration(hardware,cars,…) Spreadsheets Scheduling Floorplanning Frequencyassignments … asetofvariables{x₁,x₂,…,xₙ}to values{d₁,d₂,…,dₖ}canbeassigned suchthatasetofconstraintsoverthevariablesis respected ACSPissolvedbyavariableassignment thatsatisfiesallgivenconstraints. 35 36 ExampleMapColoring:AustralianStatesandTerritories CentralizedContraintSatisfactionsProblems Coloreachoftheterritories/ statesred,greenorbluewithno neighbouringregionhavingthe samecolor ConstraintGraph:nodesare variablesarcsareconstraints 38 ExampleMapColoring OnePossibleSolution Variables:WA,NT,SA,Q,NSW,V,T Values:{red,green,blue} Constraints:adjacentregionsmusthavedifferentcolors,e.g.,NSW≠V Solutionassignment: {WA=red,NT=green,Q=red,NSW=green,V=red,SA =blue,T=green} 39 40 BacktrackingSearchoverAssignments Algorithm Assignvaluestovariablesstep bystep(orderdoesnotmatter) Consideronlyonevariableper searchnode! DepthFirstSearch(DFS)with single-variableassignments iscalledbacktrackingsearch Cansolven-queensforn≈25 41 Backtracking-ExampleMapColoring(1/4) 43 42 Backtracking-ExampleMapColoring(2/4) 44 Backtracking-ExampleMapColoring(3/4) Backtracking-ExampleMapColoring(4/4) 45 ImprovingEfficiency:CSPHeuristics&PruningTechniques 46 ConstraintOptimizationProblem ConstraintSatisfactionProblem(CSP) Variableordering:Whichoneto assignfirst? Valueordering:Whichvaluetotry first? Trytodetectfailuresearly Trytoexploitproblemstructure Note:allthisisnotproblemspecific! Objective:findanassignmentforallthe variablesinthenetworkthatsatisfiesall constraints ConstraintOptimizationProblem (COP) Objective:findanassignmentforallthe variablesinthenetworkthatsatisfiesall constraintsandoptimizesaglobalfunction Globalfunction=aggregation(typicallysum) oflocalfunctions.F(x)=∑iFi(xi) 47 48 LinearAssignmentisaConstraintOptimizationProblem Variations Binary,ternary,oreven higherarity example:x<y,x+y<z,x+y+z<w... Finitedomains(dvalues)=> dnpossiblevariable assignments Infinitedomains(reals, linearconstraints nonlinearconstraints example:log(x)<sin(y) 49 50 DistributedConstraintOptimizationProblem(DCOP) DistributedConstraintOptimizationProblems Ageneralmodelfordistributedproblemsolving DCOPsarecomposedofagents,eachholding oneormorevariables Constraintsamongvariables(possiblyheldby differentagents)assigncoststocombinationsof valueassignments Agentsassignvaluestotheirvariablesandtry togenerateasolutionthatisgloballyoptimal withrespecttothesumofthecostsofthe constraints 52 DistributedConstraintOptimizationProblem(DCOP) WhyDistributedsolvers? Costofformalization: Examples Incentralizedsolving,allimaginableoptionshaveto beformulatedbeforehand Indistributedconstraintsatisfaction,onlyaminimalnumber ofconstraintsareconsideredbyeachagent TargetAllocation(AgentstoTargets) Meetingscheduling Difficulties Privacyissues: Noglobalcontrol/knowledge Localizedcommunication Limitedtime Incentralizedscenariosthesolverseesallmeetings andconstraints Distributedsolutionscanbeconstructedinsuchaway thatagentsonlyrevealinformationpiecemealwhen evaluatingconstraints Robustness: Acentralizedsolvercreatesacentralpointof Distributedsolvingallowsloadbalancingand 53 54 Example:MeetingScheduling ConstraintNetworks Aconstraintnetworkisformallydefinedbya tuple<X,D,C>where: X={X₁,X₂,...,Xₙ}isasetofdiscretevariables D={D₁,D₂,...,Dₙ}isasetofvariabledomains, whichenumerateallpossiblevaluesofthe correspondingvariables C={C₁,C₂,...,Cₙ}isasetofconstraints,where eachconstraintCiisdefinedonsubsetSi⊂ Xandassociatedwithaarityr=|Sᵢ|(hardor soft) 55 56 DCOPformalizationfortheMeetingSchedulingProblem Example:MeetingScheduling Asetofagentsrepresenting Asetofvariablesrepresentingmeeting startingtimesaccordingtoa participant. Hard Startingmeetingtimesacrossdifferentagentsareequal Meetingsforthesameagentarenon-overlapping. Soft Representagentpreferencesonmeetingstartingtimes. Objective:findavalidschedulefor themeetingwhilemaximizingthe sumofindividuals’preferences 57 DCOPExample:MeetingScheduling 58 Example:TargetTracking(1/3) Asetofsensorstrackingasetoftargetsinordertoprovideanaccurate estimateoftheirpositions. 59 60 Example:TargetTracking(2/3) Example:TargetTracking(3/3) Sensorscanhavedifferentsensingmodalitiesthatimpactontheaccuracyof theestimationofthetargets’positions. Collaborationamongsensorsiscrucialtoimprovesystemperformance. 61 DCOPFormalizationfortheTargetTrackingProblem 62 DCOPAlgorithms Complete Agentsrepresentsensors Variablesencodethedifferentsensing modalitiesofeachsensor(positions...) Constraints Alwaysfindtheoptimal Exhibitanexponentiallyincreasingcoordination Verylimitedscalabilityongeneralproblems Search-based&Synchronous:SyncBB,AND/ORsearch Search-based&Asynchronous:ADOPT(Modietal.,2005),NCBB,AFB DynamicprogrammingbasedDPOP(PetcuandFaltings, relatetoaspecifictarget representhowsensormodalitiesimpactsonthe trackingperformance LocalGreedyApproximative Sacrifyoptimalityinfavorofcomputationaland communicationefficiency Activationprobabilitydomaindependentandneedstobe Well-suitedforlargescaledistributedapplications,e.g.sensor networksandrobotics DSA-A(Zhangetal.,2005),MGM-1(Maheswaranetal.,2004) Objective: Maximizecoverageoftheenvironment Provideaccurateestimationsofpotentiallydangerous targets 63 64 Centralizedvsdecentralized Usedecentralized Privacyisaconcern Thecostofsharinginformationishigherthan exchangingconstraintmessages Task-SpecificationTrees Otherwiseusecentralized,but makesureyourcentralnode canbereplaced 65 TaskSpecificationTrees Example ATaskSpecificationTree(TST) isadistributeddatastructure withadeclarative representationthatdescribes acomplexmulti-agenttask. AnodeinaTSTcorresponds toatask.Ithasanode interfacewithparametersand asetofnodeconstraintsthat restricttheparameters. 67 68 TaskSpecificationTreesStructure Example Therearecurrentlysix typesofnodes:Sequence, concurrent,loop,select, goal,andelementary action. ATSTisassociatedwitha setoftreeconstraints expressingconstraints betweentasksinthetree. 69 DelegationwithTaskSpecificationTrees 70 TSTSemanticsExample ConstraintSatisfactionProblem TSTSemantics Scan1 ToallowtasksspecifiedasTSTstobedelegated,the TSTsemanticsisdefinedintermsoftheCan predicate. Can(B,τ,[ts,te,…],cons)assertsthatanagentBhas thecapabilitiesandresourcesforachievingataskτ intheinterval[ts,te]withtheconstraintscons. Ataskτ=(α,φ)consistsofanactionα,whichcanbe compositeorelementary,andagoalφ,whichmaybe empty. Scan2 71 72 Summary LinearAssignmentProblem:usedfor problemwheretasksaresimpleandcan beexecutedinanyorderandagentsare homogenous Otherwise,useaConstraint SatisfactionandOptimizationSolver Godecentralized,ifsharinginformationis aproblem TaskSpecificationTrees:arepresentation ofacomplexsetoftasksanddelegation 73/73