OptimizingRandomizedPatrolsSteveAlpernyAlecMortonKaterinaPapadakiNovember52009AbstractAkeyoperationalproblemforthosechargedwiththesecurityofvulnerablefacilitiessuchasairportsorartgalleriesisthes ID: 135698
Download Pdf The PPT/PDF document "Working Paper LSEOR 09.116 ISSN 2041-46..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Working Paper LSEOR 09.116 ISSN 2041-4668 (Online)Optimizing Randomized Patrols Steve Alpern, Alec Morton, Katerina Papadakia. Department of Mathematics, London School of Economics and Political Science, London. b. Operational Research Group, Department of Management, London School of Economics and Political Science, London. OptimizingRandomizedPatrolsSteveAlperny,AlecMorton,KaterinaPapadakiNovember5,2009AbstractAkeyoperationalproblemforthosechargedwiththesecurityofvulnerablefacilities(suchasairportsorartgalleries)istheschedulinganddeploymentofpatrols.Motivatedbytheproblemofoptimizingrandomized,andthusunpredictable,patrols,wepresentaclassofpatrollinggamesongraphs.Thefacilitycanbethoughtofasagraphofinterconnectednodes(e.g.rooms,terminals)andtheAttackercanchoosetoattackanynodeofwithinagiventimeT:Herequiresconsecutiveperiodsthere,uninterruptedbythePatroller,tocommithisnefariousact(andwin).ThePatrollercanfollowanypathonthegraph.Thusthepatrollinggameisawin-losegame,wheretheValueistheprobabilitythatthePatrollersuccessfullyinterceptsanattack,givenbestplayonbothsides.Wedetermineanalyticallyoptimal(minimax)patrollingstrategiesforvariousclassesofgraphs,anddiscusshowourresultscouldsupportdecisionsabouthardeningfacilitiesorchangingthetopologyoftheterraintobepatrolled.Subjectclassi cations:Games,noncooperative;Military,search/surveillance;Deci-sionAnalysis,riskAreaofreview:MilitaryandHomelandSecurity DepartmentofMathematics,LondonSchoolofEconomicsandPoliticalScience,London.OperationalResearchGroup,DepartmentofManagement,LondonSchoolofEconomicsandPoliticalScience,London. 1IntroductionAkeyoperationalproblemforthosechargedwiththesecurityofvulnerablefacilitiesistheschedulinganddeploymentofpatrols.Thisproblemisencounteredby,forexample:securityguardspatrollingamuseumorartgalleryantiterroristo¢cerspatrollinganairportorshoppingmallpoliceforcespatrollingacitycontaininganumberofpotentialtargetsfortheftsuchasjewelrystoressoldierspatrollinganoccupiedcityorterritoryairmarshalspatrollinganairlinenetworkinspectorspatrollingacontaineryardorcargowarehouseSuchproblemshavebeenstudiedindiverseliteratures.Forexample,awell-knownproblemincomputationalgeometrydealswiththepositionofsecurityguardsinartgalleries(Urrutia,2000)andaclassicalOperationsResearchliteratureexistsontheschedulingofpolicepatrols(seee.g.Larson(1972)andreferencestherein).Theimportanceofrandomizedpatrolshasbeenrecognizedinlawenforcementforsometime,butnotthenatureoftherandomization(e.g.ShermanandEck(2002,p.297)).Muchoftheoptimizationliteratureonthissubject(e.g.Chelst,1978)concentratesontheimportantproblemofhowtodeployrandomizedpatrolstomaximizetheprobabilityofinterceptingacrimeinprogress,whenthecrimefrequencyofdi¤erentlocationsistakenasgiven(oftenarealisticassumption,atleastintheshortterm).Suchmodelshoweverarenotgametheoreticanddonotcapturetheideaofapatrollingscheduleasastrategyselectedinthefaceofanintelligentandmalignadversary,forexampleanartthieforterrorist,whichisadistinctivefeatureoftheclassofmodelswestudyinthispaper.Althoughtheredoexistdi¤erentialgameformulationsoftherelationshipbetweenpoliceandcriminal (Isaacs,1999,Feichtinger,1983)thesetendtofocusonadynamic(andoftenstrategic)processofmutualadjustmentratherthanconfrontingtheproblemconfrontedbytheschedulerwhositsdowntodeterminethepathwhichthepatrolwilltake.GametheoreticanalyseshaverecentlyfeaturedprominentlyinORstudiesofhomelandsecurityandcounterterrorism(e.g.Brownetal,2006,BierandAzaiez,2009,Lindelaufetal.,2009).AnattractiveanduniquefeatureofgametheoreticformulationsinthecontextofpatrollingisthattheyprovideinsightintohowaPatrollershouldrandomizeherpatrols.Thereisaclearcommon-senserationaleforrandomization:apredictablePatrollerisanine¤ectiveone.Yetanaive"maximumentropy"heuristic(Foxetal,2005)maybenotfaremuchbetter:facedwithtargetsitmaynotmakesensetospend=noftheavailablepatrollingtimewitheachofthem.Thisdilemmahasattractedconsiderableattentionrecentlyamongstpractitionersandtheresearchcommunityhasrespondedtothischallenge:inparticular,theworkofParuchuriandcolleagues(2007)providesanumberofheuristicmodelswhichillustratehowequilibriumrandomizedstrategiescanbeapproximatedwhentheproblemisformulatedasaStackelberg(leader-follower)game,andsuchmodelshavefounduseinrealsecuritysituations(Gordon,2007;Newsweek,2007).Ourworkonthisproblemisinspiredbythetheoryofsearchgames,onwhichanextensivemathematicalliteraturedevelopedoverthelastfewdecades.ThistheorycapturessituationsinwhichaSearcheraimsminimizethetimetakento ndastationaryormobileHiderwhodoesnotwanttobefound(AlpernandGal,2003).TherearealsorelatedliteraturesonInspectiongames(Avenhaus,vonStengelandZamir,2002),inwhichanInspectorwhoseekstocatchanInspecteered-handed,andIn ltrationgames(Auger,1991;Garnaev,GarnaevaGoutal,1996;Garnaev,2000andAlpern,1992)inwhichaGuardseekstopreventanIn ltratorfrompenetratingsomesensitivefacility.Similarattack/defencegameshavebeenstudiedinmilitaryoperationsresearch(Washburn,2003),datingasfarbackasMorseandKimball(1950).Manysuchgamesareofindependentmathematicalinterestandhavebeenstudiedin apurelymathematicalsettings(e.g.Baston,BostockandFerguson,1989).VariousresultsareavailableforhowtheSearcher/Inspector/Guard/Defendershouldproceed,dependingontheassumptionsaboutthestructureofthemathematicalspacewhichsheinhabits.Aparticularlyproductivelineofresearchinthesearchgameliteraturehasbeentoexplorethecasewherethesearchspacecanbethoughtofasagraph,aswedohere.InthispaperweformulateagamewhichwecallthePatrollingGame.UnliketheworkofParuchuriandcolleagues,ourproblemisazero-sumgame,andprovidesforadefenderwhoismobile,beingabletotravelbetweenlocationsinthecourseofhisshift(a"Patroller").Unlikesearchgames,our"Attacker"(theequivalentofthesearchgame"Hider")maycommencehisattackatanytimeandhastobedetectedwithinagiventime-windowinordertoforestalltheperformanceofsomemisdeed.Ourgameiswin-lose-agameoftyperatherthandegreeintheterminologyofIsaacs(1999).Ourproblemissu¢cientlyidealizedthatitispossibletoobtaininsightfulanalyticresults,butsu¢cientlyrealisticthatitisrecognizableasapracticalproblemfacedbypractitionersinvariousdomains.Inthispaperwepresentsomeanalyticresultsforthisgame,anddemonstratethatityieldspatrolling(andattacking)strategieswhicharenaturalandintuitive.Moreover,weshowhowthegamecanbeusedtoguidedecisionsaboutinvestmentinhardeningvulnerablesitesorinaddingadditionalpassagewaystoenablethePatrollertoshortenthetimerequiredtogobetweendi¤erentsitesthatmightbeattacked.Weareinthispaperunabletopresentgeneralanalyticresultsforallgamesofthistype,anditseemsunlikelythatsuchsolutionsexist.Indeed,evencomputingoptimalstrategiesmaybequitechallenging,becauseofthecombinatorialexplosioninthePatrollersstrategyspace;inacompanionpaper(Alpern,MortonandPapadakiinprepa-rationa),wepresentsomealgorithmsfore¢cientlycomputingthevalueofthisgameformorecomplex(andrealistic)examples.Thispaperisorganizedasfollows.WepresentinSection2arigorousformulationofpatrollinggames,togetherwithsomeelementaryobservationsonpropertiesoftheValue.As thenumberofpurestrategiesfortheplayerscanbeverylarge,wegiveinSection3threemethodsforreducingthenumberthatwehavetoconsider:symmetrization,dominanceanddecomposition.Section4discussescertainclassesofstrategiesthattheplayerscanuseonanygraph,andwhichareoptimaloncertainclassesofgraphs.Section5solvespatrollinggamesoncertainclassesofgraphs:Hamiltonian,bipartiteandlinegraphs.Section6considershowthegamecanbeusedtoguidedecisionsininvestmentabouthardeningnodesoraddingedges.Section7presentsextensionsofthemodelandconcludes.2ThePatrollingGameInthissectionwegiveaformaldescriptionofthePatrollingGameQ;T;mwhereisthegraphwhosenodesareunderattack,isthetotalnumberoftimeunitsthegameisplayedover,and)isthenumberof(consecutive)periodsrequiredtosuccessfullycarryoutanattackonanode.Roughlyspeaking,theAttackerpicksanodetoattackandchoosessometimeinterval;+1;:::;oflengthinwhichtoattackit.ThePatrollerfollowsawalkonQ;thatis,hechoosesnodes(0);:::;w1)withconsecutivenodesthesameoradjacentinQ:ThePatrollerwinsifhiswalkisatthenodeinsomeperiodinwhichitisbeingattacked,thatis,if)=forsomeI:OtherwisetheAttackerwins.Wealsodemonstratesomesimplemonotonicityresultsandsomeboundsonthevalue,whichwillbeusefullateron.2.1FormulationMoreformally,thePatrollingGameQ;T;misawin-lose(andhencezero-sum)gamebetweenamaximizingPatroller(female)andaminimizingAttacker(male).Itcomesintwoforms,theone-o¤gameQ;T;mandtheperiodicgameQ;T;mTheone-o¤gameisplayedoutoveragiventimeinterval;:::;Toflengthonagraphwithnodesandedges.Wewilltendtoassumethatisconnectedunlessstated otherwise.ApurestrategyfortheAttackerisapairpairi;Iwhere2NiscalledtheattacknodeandTisan-intervalcalledtheattackinterval.ApurestrategyforthePatrollerisawalkT!calledapatrol.Ifwesaythatthepatrolinterceptstheattack,inwhichcasethePatrollerwinsandthepayo¤is=1;otherwisewehave=0Thusthepayo¤isgivenbyw;;i;I])=(Patrollerwins)if(Patrollerinterceptsattack),(Attackerwins)ifi=(attackissuccessful)TheValueofthisgameisthustheprobabilitythattheattackissuccessfullyintercepted.Exceptintrivialcases,optimalstrategiesmustbemixed.Theperiodicgameissimilarexceptthatthepatrols(Patrollerpurestrategies)arenowwalksofperiod(satisfying)=forallAttackintervalsarenow-intervalsinthetimecirclemod(soforexampleifis24andistheattackcouldbecarriedoutovernight,duringtheinterval2223(10oclockto2inthemorning).Wecanalsoviewthepatrolsaswalks!N.Theperiodicgameissimplertoanalyzebecausetheattackcanbeassumedtotakeplaceequiprobablyinanytimeinterval,whichsimpli estheanalysis(seeSubsection3.1).Whenthevaluesofthegamesdi¤er,wewillusethesuperscriptsandtodistinguishbetweentheValues,usingwhentheresultappliestobothcases.)canbeconsideredasparameterizedby,andjustasis,butmostofthetimewritingQ;T;misdistractingandconfusingandwewilltendtosuppresssomeorallofthesearguments.Wedenotebyi;ithedistancefunctiononthenodeset,theminimumnumberofedgesbetweenand2NThisformulationmakesanumberofassumptionswhicharenotinfactasrestrictiveastheymightappear.The rstistheassumptionanattackwilltakeplace.Animmediateresponsetothisisthateventhoughattacksoccurveryrarely,oneshouldpatrolontheassumptionthatanattackwillhappen-otherwisewhatisthepointofpatrollingatall?Amoresophisticated Figure1.asaprojectionofLemma11.Q;T;misnondecreasinginm:2.Q;T;mcannotdecreaseifanadditionaledgeisaddedbetweentwononadjacentnodesofthegraph.Thatis,isnondecreasingin(withtheorderingonthelatterunderstoodinthesenseofsetinclusion).3.Q;T;mQ;T;m4.Ifisobtainedfrombynodeidenti cation,Proof.The rstpartfollowsfromtheobservationthatapatrolthatinterceptsanattackki;Ialsointerceptstsi;IifI:Thenexttwoarebasedonthefactthatinazerosumgameaplayercannotdoworseifhegetsadditionalstrategies.Thelastisbasedonthefollowingobservation:IfapatrolinterceptsanattackonanodeofthenthepatrolinterceptstheassociatedattackonthenodeofSothePatrollercanensurethattheexpectedpayo¤isatleastbychoosingpatrolsforaccordingtosomeoptimalmixedstrategy,andthenplayingtheprojectedpatrol ThenextresultgiveseasygeneralboundontheValue.Lemma2 nVm ,forequaltoorandanyparametersand.Moregenerally,!=n;whereisthemaximumnumberofnodesthatanypatrolcancover(dependsonwhethertheone-o¤orperiodicversionisbeingplayed).Proof.ThePatrollercanobtaintheleftinequalitybyrandomlypickinganodeandwaitingthere.TheAttackercanobtaintherightinequalitybyattackingarandomnodeduringsome xedtimeinterval.Ofthesepurestrategies,thePatrollercaninterceptatmostjofthem,givingtheboundm=n;ormoregenerallythebound!=n;sincejbyde nition. 8 therearenoattacksatpenultimatenodes(usingLemma3).Sincenodesandandnodesandareequivalentundersymmetry,wecanassumetheyareattackedwithequalprobability.SothemostgeneralAttackermixedstrategyisshown,where+6=1Thepatrol=(33)interceptsallattacksatnodeandtwoattacksatnode,sowinswithprobability+2similarly=(1anyinterceptsoneattackatnodeand veattacksatnodesand,andwinswithprobability+5y:Thesetwopatrolstogetherdominateallothers.SotheAttackerminimaxeswhen+2+5y:Thisoccurswhen=110and=115withminimaxvalue=1330=43333:::.(AneasycalculationthenshowsthatthePatrollershouldadoptandwithprobabilitiesand i\t1 2 3 4 5 y x x x y y y y y 1/81/81/81/81/81/81/81/8 Figure3.Optimalattackingandpatrollingstrategiesforwith=5and=3Inthe(unrestricted)game3)itishardertoderivetheequilibriumstrategypair,butitisfairlyeasytodemonstratethattheValueis=38=0375whichshowsthatnotimeinvariantmixedAttackerstrategycanbeoptimal.Toseethat,considertheAttackstrategyshowninthemiddledrawing,andobservethatnopatrolcaninterceptmorethanthreeoftheeightequiprobableattacks.AnoptimalPatrollermixedstrategyistoadoptthefourstrategies(24)(26)andtheirreections(53)(51)(drawninthinblueslantedlines)withprobabilityeach;andadoptthetwoequivalentstrategies(33)and(44)(drawninthickredlines)withprobabilityeach.Theassertionfollowsfromtheobservationthateveryoneoftheeighteenpossibleattacksisintercepted(hit,totherightof,totheleftof)byatleastthreeofthepatrols,countingthethickredonesas12 two.Itisalsointerestingtoobservethatalltenattackswhicharenotusedatallinthemiddledrawingareinterceptedbymorethanthreeofthesepatrols.Thus,theValueoftheone-o¤game3)isbutitrequirestheuseoftimedependentAttackerstrategies:themiddlenodeisonlytobeattackedinthemiddletimeintervalNextweanalyzetheperiodicversion,thegame3)Thisissimilartotherestrictedversionoftheone-o¤gamediscussedabove,exceptthatthemiddleoftheattackcanbeatanytime,socomparingwiththeleftdrawingofFigure3,thesandswouldextendthroughouttherows,andsowehave10+10=1.The(periodic)patrol=(33)interceptsalltenattacksatmiddlenodesandandwinswithprobability10y:The(periodic)patrol=(11)interceptsfouroftheattacksatnode(allexcepttheoneduringtimes2,3,4)andthreeoftheattacksatnodesowinswithprobability+3y:Thesetwo(togetherwiththeirsymmetrictranslations)dominateallotherpatrols.SotheAttackerminimaxeswhen10=4+3y:Thishassolution=7110and=4110withminimax=411ForthePatroller,(anditsequivalents)shouldbeusedwithprobability11;wwithprobability1011Tosummarize,forlinegraph,with=5and=3wehave36364V375V43333ThustheAttackerdoesbetterintheone-o¤game,andthustheboundstatedinLemma1Part3neednotbetight.Further,inthisinstance,theAttackerhastoadoptatimedependentstrategyinordertobene tfully.3.4DecompositionSometimeswecanthinkofagraphasbeingmadeupofsimplergraphsandWecallthisadecompositionof.Thenodesoftheoriginalgrapharetheunionofthenodes13 3.5Example:thekitegraphTodemonstratetheuseofallofourstrategyreductiontechniques,weanalyzetheperiodicgameforthekitegraphillustratedinFigure5with=3.ThedominanceargumentofLemma3showedthattheAttackerwouldneverattacknode4,asitisalwaysbetterforhimtoattacktheadjacentleafnode5.Moreover,intheperiodiccasefor=3,thereisnofeasiblePatrollerstrategywhichvisitsbothnode5andanyoneof1,2,or3.Therefore,wecanremovenode4andbecon dentthattheperiodicgameontheresultinggraphKThasthesameValueasthegameonKT: Q1Q23251 Figure5.DecompositionofKitegraphKTintoKTLemma4showsthatforandasinFigure5,wehaveKT =V)+1=V(1)Obviously)=1,anditcanbeeasilyshownthat)=1for=3.Henceby(1)wehaveKT)=KT 1+2 ThisisananotherexamplewherethePatrollerdoesstrictlybetterintheone-o¤game,inwhich=3Toseethis rstnotethatbyLemma2theAttackercanensurethatm=n=3byattackingequiprobablyatthe venodes.Thenobservethatbyusingthefourpatrols(23)(25)(35)(15)withrespectiveprobabilitiesthe15 leastoneofthepatrols.Thecoveringnumberistheminimumcardinalityofanycoveringset.De nition9If,foranytwonodesand,anypatrolwhichinterceptsanattackatnodeinattackinterval,cannotalsointerceptanattackatinattackinterval,thenandwillbesaidtobeindependent.Intheone-o¤gamethisisequivalenttorequiringanytwonodestosatisfyi;iintheperiodicgame,theymustsatisfyi;iori;i(becausethePatrollerhastoreturntohisstartingpointbytheendoftheperiod).Theindependencenumberisthecardinalityofamaximalindependentset.ObviouslyIJObservethatbothanddependontheparametersQ;T;mandontheversionofthegamethatisplayed,orForexample,when=3and=3thenodesubsetofisindependentfortheperiodicgamebutnotfortheone-o¤game.FortheAttacker,theindependentstrategyisto xanattackintervalandthenchoosetheattacknodeequiprobablyfromsomemaximalindependentset.ForthePatroller,thecoveringstrategyistochooseequiprobablyfromaminimalsetofcoveringpatrols.Notethatfor=2patrolscanbeidenti edwithedgesof,sothesede nitionsreducetotheusualnotionofanindependentsetnothavingadjacentnodesandacoveringsetconsistingofedges.Lemma10 JV1 (with=1when).Proof.TheAttackersindependentstrategygivestheupperboundandthePatrollerscoveringstrategygivesthelowerbound. Thecaseswheredealwithmanypatrollinggames.Forexample,wecanusethistechniquetogiveanothersolutiontothekitegraphKTofFigure2fortheperiodicgamewith=3Herethenodes2,3,and5formanindependentset(becausei;i)=43=andinterceptingpatrolsonthetopleft,toprightandbottomedges(periodpatrolsofthe18 Proof.Firstobservethatineithercasewehavem=nbyLemma2.Intheone-o¤game,supposethePatrolleradoptsarandomHamiltonianpatrol.ThenforanyattackintervalI;isarandom-arcoftheHamiltoniancycle,andassuchcontainstheattacknodewithprobabilitym=n;asclaimed.Ifisamultipleofn;thisstrategyisalsofeasibleintheperiodicgame.Toobtainthelimitingresult,notethatifmod=0theperiodicPatrollercanmodifytherandomHamiltonianpatrolbywaitingatarandomnodeduringarandominterval.Thiswillnothurthimunlesstheattackintervaloverlapsthewaitinginterval,whichhasprobability1)=T;so Tm nVpVo=m andso (2) Sincetheaboveresultappliestothecyclegraph,wecanuseittosolvethegameonsomegraphswhichcanbeobtainedfromthecyclegraphbyidenti cationofnodes.WenowsolvetheperiodicPatrollingGamefortheeightnodegraphshownbelowontheleftofFigure6inthecase=10and=4Firstnotethatsincethediameteris=5wehavefromTheorem7thatthediametricalAttackstrategyensuresthatm==410Byviewingthegraphasaprojectionof10(withValuem=10=410fromTheorem11)weconcludefromLemma1Part4that10so=410 Figure6.Agraphshownasprojectionof1020 ingeneralandwithasimilarlimitingresultfortheperiodiccase Inallthesecases,informallyspeaking,anoptimalstrategyfortheAttackeristo xanattackintervalandchoosetheattacknodeequiprobablyfromthelargerhalfset;anoptimalstrategyforthePatrolleristorandomizeoveracollectionofstrategieswhichvisitthelargerhalfseteverysecondtimeperiod.Inthecaseof=2,Patrollerchoosesanedgejoiningthehalfsets;theAttackersandPatrollersstrategiescanbeseenasarandomchoicefromanindependenceandcoveringsetrespectively;inthiscasetheTheoremcanbeunderstoodasaversionofKönigsTheorem(Harary,1971,Theorem10.2)inourcontext,sinceKönigsTheoremstatesthattheindependenceandcoveringnumbersofabipartitegraphareidentical.Toillustratetheproof,considerthespecialcaseofthestargraph;nconsistingofacentralnodeconnectedtoextremenodes.ThismodelsthesituationwherethePatrollerhasresponsibilityforthesafekeepingofabuildingwhichhasmultiplewings,accessiblethroughacommonlobbyarea.Wecanviewasobtainedfromtheevencyclegraph2(1)byidentifying(say)allevennumberednodes,asinFigure7. Figure7.obtainedfrombynodeidenti cation.ThismodeofreasoningleadsustodiscoveradditionalequilibriumpairsfortheHamiltoniancase.Considerthecyclegraphforevenandamultipleofn:WesawearlierthattheuniformstrategywasoptimalfortheAttacker.Butsincen=;n=isbipartiteTheorem12nowgivestheadditionaloptimalstrategyofattackingequiprobablyontheodd(oreven)nodes.InfactthereisonemoreoptimalAttackerstrategy:sincethediameterisn=thediametricalstrategyalsogivesm=m=nbyTheorem7.22 probability foranoverallprobabilityofinterceptionof 4 3 7+1 7=3 .Allotherpatrolsyieldalowerexpectedpayo¤andso3) Nextconsider2)Weclaimthatthevalueofthisgameis .Wehave=4andsofromTheorem10wehaveToensurewinningwiththisprobability,thePatrollermustuse"biasedoscillations"onedgesi;ioftheformi;i;i;i;i,whichwedenoteas ,witharandomtimerotation.Clearly interceptsanyattackonandinterceptsanyattackonwithprobability(thatisunlesstheattackcoincideswitharepeatedTheoptimalprobabilitiesofthebiasedoscillationsonconsecutivenodesareshownbelow.161616161616Attacksonanynodeareinterceptedwithprobabilityatleastwithequalityexceptforthecentralnode(whichshouldneverbeattacked).Forexampleanattackonnodeisinterceptedwithprobabilityifeither or isadoptedbythePatroller,thatiswithprobability(416+516)=516Soitisinterceptedwithprobability(45)(516)=16HardeningNodesorAddingEdgesUptonowwehavetakenthenetworkasgiven.ButtheagencythatcontrolsthePatroller(the"Defender")maybeabletopaytoeitherhardenasitesothatitisimmunetoattackortobuildapassagewaytohelpthePatrollermovemorequicklybetweensites.InourformulationthiscorrespondseithertoreducingtheAttackersstrategyspacebyremovinganattacknode,oraddinganedgetothegraph.Inthissectionwegivesomeverysimpleexamplesintheeasycasewhereallsites(nodes)areequallyexpensivetohardenandnewpassageways(edges)areequallyexpensivetobuild(morecomplexexampleswillbedealtwithinourforthcomingpaperonthecomputationalaspectsofthisproblem).6.1Hardeningnodes24 6.2AddingedgesWeobservedearlier(Lemma1)thataddingedgescannothurtthePatroller,andusedthistoanalyzecertainnetworks.Herewetakeadi¤erentpointofviewandaskwhetheritpaysforthePlayerthatcontrolsthePatrollertospendmoneytoaddanedgebetweentwonodesinordertoincreasethee¢cacyofthePatroller.WeconsideragainthenetworkandA;3)andaskhowtheinterceptionprobabilityincreases,ifatall,byaddingedgesbetweennonadjacentnodesofA:TheindependentsetsareindicatedinallcasesbydisksatnodesinFigure9.Thecoveringsetsareeitheridenti edwithedgesorcycles,whicharethickened.Weobservethataddingedgesasinorachievesnoincreaseatallintheinterceptionprobability(theValue,butasinorincreasesittofromTheintuitionisthattheseedgescreatecycles,whicharemoree¢cientforpatrolsthanoscillationsonedgeswitharepeatednode. A: I=J=3, V=1/3 B: I=J=3, V=1/3C: I=J=2, V=1/2 D: I=J=2, V=1/2 E: I=J=3, V=1/3 Figure9.Thegraphandfourwaysofaddinganedge7ConclusionInthispaperwehavedescribedasimple,intuitivemodelwhichcanserveasthebasisforobtainingoptimalrandomizedpatrols.Theassumptionsofthepresentpapercanberelaxedinthefollowingextendedmodels:26 Thenodescanhavedi¤erentvalues(e.g.paintingsofvaryingartisticmeritinanartgallery).ThesimplestwaytomodelthisistokeepthesamestrategiesbutviewthePatrolleraswantingtominimizethevalueofasuccessfullyattackednode.Sothisversioncomeswithacostvectorc;wheredenotesthecostofasuccessfulattackonnodei:Somenodesmaybeunequallyhard,orevenimpossible,toattack.WethusreplacetheparameterbyavectorM;wheredenotesthenumberofperiodsrequiredtosuccessfullyattacknodei:Inarelatedway,edgesmayhavelengthsattachedtothem.(Twonodeswithaninterveningnodewhichcannotbeattackedareinsomesensetwotimeunitsapart.)Theremaybemultiplepatrollersand/orattackers.Someoftheresultsofthecurrentpapertransferovereasilytothissituation-forexampleonecansimply"multiplyup"thenumeratorofLemma10tohandlethesituationwherethereareseveralpatrollers-butingeneralthisseemstoberathermorecomplex.PerhapsthePatrollermuststartataknownnode=0intheone-o¤game.Ofcourse,inthiscasethediameterofcannotbelargewithrespectto,otherwisetheAttackerwillalwayswin.AnaturalconjectureisthattheAttackerwouldattackearlierinordertotakeadvantageofhisgreaterknowledgeofwherethePatrollerwillbe,butisthisalwaysso?Itmaybenaturaltoconsideracontinuoustimeformulationofthisproblem.Anattacktakesplaceatanypointofthenetwork(notnecessarilyanode)onacontinuoustimeintervalof xedlength.ThePatrollerusesaunitspeedpathandwinsifheisattheattackedpointatsometimeduringtheattackinterval.Thiswouldmodel,forexample,thedefenseofapipelinesystem.ThePatrollermaybealerted(perhapsnoisilyandwithsomeerror)tothepresenceofanAttacker;andtheAttackermaybealertedbyaconfederatewhocanidentifywhena27 Patrollerleavesaparticularnode(forexample,ifthePatrollerisinamarkedpolicecar).Manyoftheseproblemsarenotanalyticallytractable,andsomeofthemwillbediscussedinourforthcomingpaperoncomputationalaspectsofthesegames(Alpern,MortonandPapadakiinpreparationa).Acknowledgement14WewouldliketothankDelofvonWinterfeldtforsuggestingthisprob-lemtous,andtoMilindTambeforinterestingdiscussions.AMandKParealsogratefultotheCentreforRiskandEconomicAnalysisofTerroristEvents(CREATE)attheUniversityofSouthernCaliforniafortheirsupportandhospitality.SAwassupportedbyNATOCollaborativeLinkageGrant983583onDefenseAgainstTerrorism.8ReferencesAlpern,S.(1992).In ltrationgamesonarbitrarygraphs.J.Math.Anal.Appl.163,no.1,286-288.Alpern,S.andM.Asic(1985)."TheSearchValueofaNetwork."Networks15(2):229-238.Alpern,S.andS.Gal(2003).TheTheoryofSearchGamesandRendezvous.Dordrecht,Kluwer.Alpern,Morton,andPapadaki(inpreparationa).Computingoptimizedpatrols.Alpern,Morton,andPapadaki(inpreparationb).Patrollingachannelforaslowin ltrator.Auger,J.(1991)Anin ltrationgameonkarcs.Nav.Res.Logist.38,511-530.Avenhaus,R.,B.VonStengel,etal.(2002).Inspectiongames.HandbookofGameTheorywithEconomicApplications.R.J.AumannandS.Hart.AmsterdamElsevier.Baston,V.,F.Bostock,etal.(1989)."TheNumberHidesGame."ProceedingsoftheAmericanMathematicalSociety107(2):437-447.Bier,V.M.andM.N.Azaiez(eds)(2009).GameTheoreticRiskAnalysisofSecurityThreats.NewYork,Springer.28 Brown,G.,Carlyle,M.,Salmeron,J.andWood,K.(2006)."DefendingCriticalInfrastruc-ture".Interfaces36,530-544.Chelst,K.(1978).Analgorithmfordeployingacrimedirected(tactical)patrolforce.Man-agementScience,Vol.24,No.12,pp.1314-1327Feichtinger,G.(1983).Adi¤erentialgamessolutiontoamodelofcompetitionbetweenathiefandthepolice.ManagementScience,Vol.29,No.6,pp.686-699.Fox,C.R.,D.Bardolet,etal.(2005).Partitiondependenceindecisionanalysis,resourceallocationandconsumerchoice.ExperimentalBusinessResearch.VolumeIII.R.ZwickandA.Rappoport.Doredrecht,Kluwer.Garnaev,A.(2000).Searchgamesandotherapplicationsofgametheory:LecturenotesinEconomicsandMathematicalSystems485.Berlin,Springer.Garnaev,A.,G.GarnaevaandP.Goutal(1997).Onthein ltrationgame.Int.J.GameTheory26:215-221.Gordon,L.(2007)."USCstudentscomputerprogramenlistedinsecuritye¤ortsatLAX."LosAngelesTimesOctober1(http://teamcore.usc.edu/ARMOR/LATIMES.pdf).Harary,F.(1971).Graphtheory.Reading,MA,Addison-Wesley.Isaacs,R.(1999).Di¤erentialGamesDover.Larson,R.C.(1972).UrbanPolicePatrolAnalysis.Boston,MITPress.Lindelauf,R.,Borm,P.,Hamers,H.(2009).Theinuenceofsecrecyonthecommunicationstructureofcovertnetworks.SocialNetworks,31,126-137.Morse,P.M.andG.E.Kimball(1950).MethodsofOperationsResearch.London,MITPressandWiley.Newsweek(2007).TheElementofSurprise.Newsweek.http://teamcore.usc.edu/ARMOR/newsweek.pdf.Paruchuri,P.,J.P.Pearce,etal.(2007).Ane¢cientheuristicapproachforsecurityagainstmultipleadversaries.AAMAS07,Honolulu,Hawaii,ACM.Rosenshine,M.(1970).Contributionstoatheoryofpatrolscheduling.OperationalResearch29 Quarterly,Vol.21,No.1,pp.99-106Ruckle,W.H.(1983).GeometricGamesandTheirApplication.Boston,Pitman.Sherman,L.andEck,J.E.,(2002).Policingforcrimeprevention.In:Evidence-basedcrimeprevention,L.Shermanet.al.,eds,Routledge,TaylorandFrancis,NewYork.Urrutia,J.(2000).Artgalleryandilluminationproblems.Handbookofcomputationalgeometry.J.-R.SackandJ.Urrutia.Amsterdam,Elsevier.Washburn,A.(2003).Two-personzero-sumgames,3rdedition.Linthicum,MD,INFORMS.Zoroa,N.andP.Zoroa(1993)."Somegamesofsearchonalattice."NavalResearchLogistics40:525-541.30