Model-Cheking IteratedGames
Chung-HaoHuang 1
SvenShewe 2
FarnWang 1;3
1:GraduateInstituteofEletroniEngineering,NationalTaiwanUniversity
2:DepartmentofComputerSienes,UniversityofLiverpool
3:DepartmentofEletrialEngineering,NationalTaiwanUniversity
Abstrat. Wepropose alogiforthedenition oftheollaborativepowerof
groupsofagentstoenforedifferenttemporalobjetives.Theresulting
tempo-ralooperationlogi(TCL)extendsATLbyallowingforsuessivedenition
ofstrategiesforagentsand agenies.Differenttopreviouslogiswithsimilar
aims,ourextensionutsanelinebetweenextendingthepowerand
maintain-ingalowomplexity:model-hekingTCLsentenesisEXPTIMEompletein
thelogi,andxedparametertratableforspeiationsofboundedsize.This
advanementovernon-elementarylogisisboughtbydisallowingatoolose
en-tanglementbetweenooperationandompetition.Weshowhowallowingsuh
anentanglementimmediately leadsto anon-elementary omplexity. Wehave
implementedamodel-hekerforthelogiandshownthefeasibilityof
model-hekingonafewbenhmarks.
1 Introdution
WhiletheveriationoftraditionallinearandbranhingtimelogislikeLTL,CTL,
andCTL*[17,8℄hasbeenreduedto(repeated)reahability[11,13℄,thesatisability
hekingandsynthesisproblemhasbeentightlylinked withgametheoryeversine
theseminalworksofBühiandLandweber[5,4℄.Withtheintrodutionof
alternat-ing timelogi (ATL)by Alur,Henzinger,andKupferman[1℄andinautomatabased
-alulusmodel-heking(e.g.,[22℄),gameshaveenteredintotheveriationofthe
orretnessofreativesystems.Withgametheoretihallengesmovingintothefous
ofresearherswhostudythespeiationanddesignofreativesystems, traditional
problemsofmulti-playergamesarereplaingtheformerdistintionbetweenan
adver-sarial environmentandasupportivesystem. Instead,we havegroupsofplayersthat
ooperateonsomeobjetiveswhileompetingonothers.
Forpartiularproperties,theintuitionthatsomeplayersrepresentthesystemwhile
otherplayersrepresentthe environmentis,however,still useful.Followingthis
intu-ition,thesystemwinsthegameinanexeution(oraplayinthejargonofgametheory)
ifthe systemspeiationisfullledalongit,anditwins thegameifit anforea
winningplay.Systemdesignasawholeforspeiationsingamelogisanratherbe
omparedtodesigningagameboardandtoshowthattherespetivegroupofplayers
(or:ageny)hastheoalitionpowerrequiredbythesystemspeiation.
?
TheresearhwassupportedbytheNationalSieneCounilgrant97-2221-E-002-129-MY3
ATL
,thealternating-alulus(AMC),andgamelogi(GL)[1℄,strategylogis[7,
9,15,14℄,oordinationlogi[10℄,stohastigamelogi[3℄,andbasistrategy
inter-ationlogi(BSIL)[21℄ forthe speiationofthe interplayinopen systems.Eah
languagealsoomeswithaveriationalgorithmthatdetermineswhetherawinning
strategyforthesystemexists.However,thereisagapbetweentheavailabletehniques
andthesalabilityrequiredforindustrialappliations.Franklyspeaking,noneofthe
languagesaboverepresents,inourview,aproperombinationofexpressivenessfor
loseinterationamong agentstrategiesandefienyforthe veriationor
refuta-tionofomplianewithaspeiation.Ononehand,logislikeATL,ATL
,AMC,
andGL[1℄allowustospeifytheollaborativepowerofgroupsofplayerstoenfore
aommonobjetive.Thisfallsshortfromspeifyingeventhesimplepropertiesina
typialgame.Forexample,itwasshownin[21℄thatATL,ATL
,AMC,andGL[1℄
annotexpressthatthesamestrategyofabankingsystemmustallowthelientsboth,
towithdrawandtodepositmoney:astrategyquantierintheselogisalwaysrefersto
thestrategiesofallagents,whereasthispropertyrequirestobindrstthestrategyof
thebank,andthenrefertodifferentstrategiesofthelients.Thisisarguablyasevere
restritionwhenreasoningaboutreal-worldproblems.
Tosolvethe expressiveness problem in the aboveexample,strategy logis (SL)
wereproposedin[3,7,15,14℄.Theyallowfortheexiblequantiationoverstrategies
inlogiformulas.However,theirveriationomplexityisprohibitivelyhighandhas
inhibitedpratialappliation.
Apreviousattempttotametheomplexityofstrategyinteration[21℄,ontheother
hand,resultsinafulltemporalisation.Thisleadstosevererestritionsinthe
entangle-mentbetweentemporaloperatorsandstrategybindingandthusprevents,forexample,
reasoningaboutNashequilibria.
Wethusproposetoadaptthelogiintroduedin[21℄toanewtemporallogialled
temporalooperationlogi(TCL)forthispurpose.LetusintrodueTCLinformallyon
agameamongthreeprisoners.
Example:IteratedPrisoners'Dilemma. Inspiredbythefamousprisoners'dilemma,
weonsideramodelwherethreesuspets,whoareinitiallyinustody,areinterrogated.
Inoursimpliedversion,theyplayinturns(ratherthanonurrently),andhavethe
hoiestoeitheradmitordenythehargesmadeagainstthem.Ifalldeny,theywillbe
releasedbasedonlakofevidene.
However,asuspetmaydeidetoollaboratewiththepolieandbetrayherpeers.A
soleollaboratorwillbeaquittedasarownwitness,whileherpeerswillbesentened.
Butiftwoormoresuspetsollaboratewiththepolie,allwillbesentened.
Inaniteratedprisoners'dilemma,theinterplayanontinueuptoanunbounded
numberoftimes.Suhagameisveryusefulinmodellingollaborationandompetition
innetworks.Forexample,astrategyinprisoners'dilemmaisnieifitdoesnotsuggest
betrayalinitiallyandonlysuggestsbetrayalif,inthepreviousround,anotherprisoner
betrayed[2℄.ThefollowingTCLsentenestatesthatPrisoner1hasaniestrategy.
h1i((h+i:betray
1 )_
W
a6=1
betray
a
toahievehertemporalgoal.h+iisastrategyinterationquantier(SIQ)thatinherits
thestrategyfromitsparentformula.Propositionbetray
i
isanatomipropositionfor
thebetrayalofprisoneriatthepresentstate.Similarly,weanreetmoreinvolved
strategies,suhas`Prisoner2willalwaysbetraywhenshedoesnothavethepowerto
forePlayer1toalwaysplaynie.'
h2i((h+ibetray
2
)_h+i((h+i:betray
1 )_
W
a6=1
betray
a
)) (B)
Similarpropertiesanbeusedtospeifyforgiving 1
orotherrelatedstrategies[2℄.
AforgivingstrategyofPrisoner1isreetedbythefollowingTCLproperty.
h1i((h+i:betray
1 )^
W
a6=1
betray
a
) (C)
Weanalsoreasonabout theexisteneofPrisoner2'sstrategythatavoidbetrayalif
Prisoner1anbeunforgivingunderthisstrategy.
h2i((h+i:betray
2
)_h+1i((h+i:betray
1 )^
W
a6=1
betray
a
)) (D)
Asanbeseen,propertieslike(B)and(D)arerelevantinnetworkenvironmentswhere
playsanbeextendedroundbyroundwithouttermination.Everyagentmaytrakeah
others'reordstodeidewhetherornottoooperate.Suhapropertyannotbe
ex-pressedinATL
,GL,AMC,orBSIL.WhileitanbeexpressedwithSL,the
veria-tionomplexityofSLisprohibitive.
In[21℄,SIQsanneitheroverridenorrevokestrategiesassignedbytheSQorSIQs
inwhosesopetheyare.Consequently,BSILannotexpressdeterministiNash
equi-libria.Tooveromethisrestrition,weintrodueastrategyresetoperatorthatrevokes
previousstrategyassignments.
Letjail
a
beaproposition,whihstatesthatPrisoneraisinjail.InTCL,
h1;2;3i V
a2[1;3℄
(h+;i:jail
a
)_h aijail
a
(E)
requiresthatthetreeagentsanooperatesuhthateveryagenteithereventuallyleaves
prism,orstaysforeverinprismregardlessofherownstrategyundertheurrent
strate-giesoftheremainingprisoners.TheSIQh ai revokesthebindingofagentatoher
strategy.
Inthiswork,weestablishthatTCLisinomparablewithATL
,GL,andAMCin
expressiveness.Althoughthestrategylogisproposedin[3,7,9,15℄subsumeTCLwith
theirexiblequantiationofstrategiesandbindingtostrategyvariables,their
model-hekingomplexities arealldoublyexponentialtimehard.Inontrast, TCLenjoys
an EXPTIME-ompletemodel-heking omplexity andxed parametertratability
whenusingthelengthoftheformulaasparameter,aswellas2EXPTIMEompleteness
ofthe TCL satisability problem forturn-basedgamegraphs. TCL thusprovidesa
betterbalanebetweenexpressivenessandomplexity/efienyonsiderationsthan
ATL
,GL [1℄,andSL[7,15,14℄.Giventhe expressivepowerasexempliedby the
speiationsfromabove,TCLanbeviewedasanexpressiveyetinexpensivesublass
ofSL[15,14℄.
1
w 0
fqg u
0
fpg ;
w
; ;
u
belongstoAgent1andbelongstoAgent2.
Fig.1.Aturn-basedgamegraph
OrganisationofthePaper. Setion2explainsturn-basedgamegraphsforthe
desrip-tionofmulti-agentsystemsandpresentsthesyntaxandsemantisofTCL.Setion3
disussestheexpressivenessofTCL,establishingthatCTL,ATL,LTL,andCTL*an
beviewedassyntatifragmentsofTCL.WeshowthatTCLismoreexpressivethan
anyoftheselogiswhileinomparablewithATL
,AMC,andGL[1℄in
expressive-ness,anddisusstheeffetofamildextensionofTCL.Inthefollowingsetions,we
develop an automatabased model-hekingalgorithmand establish the
EXPTIME-ompletenessand2EXPTIME-ompletenessoftheTCLmodel-hekingand
satisa-bilityproblem,respetively.Finally,wehaveimplementamodel-hekerandvalidated
thefeasibilityofusingTCLonasetofbenhmarks.
2 SystemModelsandTCL
2.1 Turn-basedgamegraphs
Aturn-basedgameisplayedbyanitenumbermofagents,indexed1throughm.A
gameisatupleG=hm;Q;r;!;P;;Ei,where
Parametermisthenumberofagentsinthegame.
Qisthesetofstatesandr2Qistheinitialstate(orroot)ofG.
!:Q7![1;m℄isafuntionthatspeiestheownerofeahstate.Onlytheowner
ofastatemakeshoiesatthestate.
Pisanitesetofatomipropositions.
:Q7!2 P
isapropositionlabellingfuntion.
EQQisthesetoftransitions.
Foreaseofnotation,wedenotewithQ
a
=fq2Qj!(q)=agthestatesownedbyan
agenta.
InFigure1,wehavethegraphialrepresentationofaturn-basedgamegraph.The
ovalsandsquaresrepresentstateswhilethearsrepresentstatetransitions.Wealsoput
downthevaluesinsidetheorrespondingstates.
Foronveniene,intheremainingpartofthe manusript,weassumethatwe are
alwaysintheontextofagivengamegraphG=hm;Q;r;!;P;;Ei.Thus,whenwe
writeQ;r;!;P;,andE,werespetivelyrefertotheomponentsQ,r,!,P,,and
EofthisG.
Aplayisaninnitepathq
0 q
1
:::inGsuhthat,foreveryk2N,(q
k ;q
k +1 )2E.
isinitialifq
0
=r.Foreveryk0,welet(k)denoteq
k
.Also,givenhk,welet
aplayprex = q
0 q
1 :::q
n
,jj = n+1denotesthelengthoftheprex.Givena
k2 [0;jj 1℄,welet(k)=q
k
.Foronveniene,weuselast()todenotethelast
statein,i.e.,(jj 1).
Foranagenta2[1;m℄,astrategy foraisafuntionfromQ
Q
a
toQsuhthat
forevery2Q
Q
a
,()2Qwith last();()
2E.
AnagenyAof[1;m℄isasubsetof[1;m℄.Inashorthandnotation,weoftendrop
the urlybraketsinthesetnotation,inpartiularforsingletonandemptysets.For
example,1;3;4isashorthandforf1;3;4g.
Aplayisompatiblewithastrategy
a
ofanagenta2[1;m℄iff,foreveryk2N,
!((k))=aimplies(k+1)=([0::k℄).
2.2 TCLSyntax
ATCLformulaisonstrutedwiththefollowingthreesyntaxrules.
::=pj:
1 j 1 _ 2 jhAi
::=jj
1 _ 2 j 1 ^ 2 jh+Ai 1 jh+Ai 1 jh+Ai 1 U 1 jh+Ai 1 R 1
j h Ai
1
jh Ai
1
jh Ai
1 U
1
jh Ai
1 R
1
::=j
1 _ 2 j 1 ^ 2 jh+i 1 jh+i 1 U 2 jh+i 1 R 2
j h Ai
1
jh Ai
1 U
2
jh Ai
1 R
2
Here,pisanatomipropositioninP andA f1;:::;mgisanageny.Property
hAi
1
isan(existential)strategyquantiation(SQ)speifyingthatthereexist
strate-giesoftheagentsinAthatmakeallplaysonsistentwiththesestrategiessatisfy
1 .
Propertyh+Ai
1
isan(existential)strategyinterationquantiation(SIQ)andan
onlyourboundbyanSQ.Intuitively,h+Ai
1
meansthatthereexiststrategiesofthe
agentsinAthatworkwiththestrategiesintroduedbytheanestorformulas.Likewise,
h AiindiatesarevoationofthestrategybindingfortheagentsinA.h+iisan
ab-breviationforh+;or,equivalentlyh ;i.Thus,itneitherbindsnorrevokesthebinding
ofthestrategyofanyagent.Yet,itprovidesatemporalisationinthatitprovidesatree
formulathatanbeinterpretedatapartiularpoint.
`U'istheuntiloperator.Theproperty
1 U
2
speiesaplayalongwhih
1 istrue
until
2
beomestrue.Moreover,alongtheplay,
2
musteventuallybefullled.`R'is
thereleaseoperator.Property
1 R
2
speiesaplayalongwhiheither
2 isalways trueor 2 U( 1 ^ 2
)issatised.(Releaseisdualtountil::(
1 U 2 ) , : 2 R: 1 .)
Inthefollowingwemayuseh?Ai toonvenientlydenoteanSQorSIQformula
with`?'isempty,`+',or`-'.AnSIQhAi isallednon-trivialifAisnotempty,and
trivialotherwise.
FormulasarealledTCLformulas,sentenes,orstateformulas.Formulas and
arealledtreeformulas.Notethatwestritlyrequirethatnon-trivialstrategy
intera-tionannotrosspathmodaloperators.Thisrestritionisimportantbeauseitoffersa
sufientlevelofloalitytoefientlymodel-hekasystemagainstaTCLproperty.
Toillustratethisandtoprovideasimpleextensionthatoffersmoreexpressivepowerto
theostofamuhhigheromplexity,weinformallydisussasmallextension,extended
usedtoenodeATL ,andtherealisabilityproblemofprenexQPTLanbereduedto
ETCLmodel-heking.
Foronveniene,wealsohavethefollowingshorthandnotations.
truep_(:p) false:true
1 ^
2
:((:
1 )_(:
2
))
1
)
2
(:
1 )_
2
1
trueU
1
1
falseR
1
:
1
:
1
hAi
1
hAih+i
1
hAi
1 U
2
hAih+i
1 U
2
hAi
1 R
2
hAih+i
1 R
2
Ingeneral,itwouldalsobenietohavetheuniversalSQsandSIQsasdualsof
exis-tentialSQsandSIQs,respetively.Couldn'tweadd,orenodebypushingnegationsto
stateformulas,apropertyoftheform[+A℄
1
,meaningthat,forallstrategiesofageny
A,
1
willbefullled?Inpriniple,thisisindeednoproblem,andextendingthe
se-mantiswouldbesimple.Thislogiwouldbeequivalenttoallowingfornegationsin
theprodutionruleof .Theproblemwiththislogiisthatitistoosuint.Wewill
brieydisussinthefollowingsetionthatmodel-hekingbeomesnon-elementary
ifweallowforsuhnegations.
Fromnowon,weassumethatwearealwaysintheontextofagivenTCLsentene.
2.3 TCLSemantis
Inordertopreparethe denitionofasemantisforTCLformulas,westartwiththe
denitionofasemantisforsentenesoftheformhAi ,where doesnotontainany
SQs.WealltheseformulasprimitiveTCLformulas.
DuetothedesignofTCL,strategybindingsanonlyeffetivelyhappenat
non-trivialSQshAiandwhenanon-trivialSIQh+Biisinterpreted.Toeasereferringto
thesestrategies,werstdenetheboundagenyofasubformulasofaTCLsentene
,denotedbnd(),asfollows.
Forstateformulas,bnd()=;.
ForstateformulashAi ,bnd( )=A(unless isastateformula).
Fortreeformulas
1
=h+Ai
2 ,bnd(
2
)=bnd(
1 )[A.
Fortreeformulas
1
=h Ai
2 ,bnd(
2
)=bnd(
1 )rA.
Forallothertreeformulas
1 or
2
with =
1 OP
2
,withOP2f^;_;U;Rg,
wehavebnd(
1
)=bnd( )orbnd(
2
)=bnd( ),respetively.
bndshows,whihagentshavestrategiesassignedtothembyanSIQorSQ.Notethat
thisleavesthebndundenedforallstateformulasnotinthesopeofanSQformulas.
Forompleteness,weoulddenebndasemptyintheseases,butadenitionwillnot
berequiredinthedenitionofthesemantis.
Astheintrodutionofadditionalstrategiesthroughnon-trivialSIQ h+Biis
gov-ernedbyapositiveBooleanombination,allstrategyseletionsanbeperformed
on-urrently.Suhadesignleadsustotheoneptofstrategyshemes.
Astrategyshemeisthesetofstrategiesintroduedbyanynon-trivialSQhAior
SIQh+Ai.Byabuseofnotation,weuse[;a℄toidentifysuhastrategy.Readinthis
way,anbeviewedasapartialfuntionfromsubformulasandtheirboundagenies
h2iq),thestrategyusedinbyAgent1toenforethewholeformulaanbereferred
toby
[h1i((h+2ip)^h2iq);1℄;
butalsoby[h+2ip;1℄,while[h2iq;1℄isundened.
WeuseasimpletreesemantisforTCLformulas.A(omputation)treeT
r is
ob-tainedbyunravellingGfromrandexpandtheownershipandlabellingfuntionsfrom
GtoT
r
inthenaturalway.Tehnially,wehavethefollowingdenition.
Denition:ComputationTree. AomputationtreeforaturnbasedgameGfroma
stateq,denotedT
q
,isthesmallestsetofplayprexesthatontainsqand,forall2T
and(last();q 0
)2E,q 0
2T.
Thestrategy-prunedtreeforatreenode,astrategysheme,andasubformula
1
offromastateq,insymbolsT
q h;;
1
i,isthesmallestsubsetofT
q suhthat: 2T q h;; 1 i; forall 0 2 T q h;; 1
i with! (last( 0 ) = 2 bnd( 1
)and(last( 0
);q 0
) 2 E,
0 q 0 2T q h;; 1 i; forall 0 2 T q h;; 1
i, a = ! (last ( 0
)
,and q 0
= [
1 ;a℄(
0
)with a 2
bnd( 1 ), 0 q 0 2T q h;; 1 i.
Given aomputationtreeorastrategy-prunedtreeT anda node 2 T,forevery
q2T,wesaythatqisasuessorofinT.AplayisalimitofT (oraninnite
pathinT),insymbols 1
2T,ifthereareinnitelymanyprexesofinT.
WenowdenethesemantisofsubformulasofprimitiveTCLformulasindutively
as follows.Given theomputationtreeT
q
ofG,atree node 2 T
q
,andastrategy
sheme,wewriteT
q ;;j=
1
todenotethatT
q
satises
1
atnodewithstrategy
sheme.
Whilethenotation mightseemheavyonrstglane,notethatthe truthforstate
formulasmerelydependsonthestatelast()inwhihtheyareinterpreted,andthetree
formulasaresimplyinterpretedonastrategyprunedtreerootedinanddenedbythe
strategysheme.
ForstateformulasotherthanSQformulas,weusethestateformulasemantis:
T
q
;;j=iffG;last()j=,withtheusualdenition.
G;qj=pif,andonlyif,p2(q),
G;qj=:if,andonlyif,G;q6j=,
G;qj=
1 _
2
if,andonlyif,G;qj=
1
orG;qj=
2 ,and
G;qj=
1 ^
2
if,andonlyif,G;qj=
1
andG;qj=
2 .
(Notethatthisallowsforusingnegationforstateformulas.)
T
q
;;j=
1 _
2 iffT
q
;;j=
1 orT
q
;;j=
2 .(The
i
arenostateformulas.)
T
q
;;j=
1 ^
2 iffT
q
;;j=
1 andT
q
;;j=
2 hold.
T
q
;; j= hAi iff, forall suessorsq 0
of inT
q
h;;hAi
1 i, T q ;q 0
;j= holds.
T
q
;; j= hAi
1 U
2
iff, forall limits 1
2 T
q
h;;hAi
1 U
2
i,there is
ak jj 1suhthat T
q
;[0;k℄; j=
2
and,forall h 2 [jj 1;k 1℄,
T
q
;[0;h℄;j=
T
q
;; j= hAi
1 R
2
iff,foralllimits 2 T
q
h;;hAi
1 R
2
i,oneofthe
followingtworestritionsaresatised.
Forallkjj 1,T
q
;[0;k℄;j=
2 .
Thereisak jj 1suhthatT
q
;[0;k℄; j=
1 ^
2
,and,forallh 2
[jj 1;k℄,T
q
;[0;h℄;j=
2 .
T
q
;;j=hAi
1 iffT
q ;;j=
1 .
G;qj=hAi
1
iffthereisastrategyshemesuhthatT
q ;q;j=
1 .
If
1
isaTCLsentenethenwewriteGj=
1
forG;rj=
1 .
Notethat,whileaskingfortheexisteneofastrategyshemereferstoallstrategies
introduedbysomeSQorSIQintheTCLsentene,onlythestrategiesintroduedby
therespetiveSQandtheSIQsinitssopearerelevant.
Thesimpliityofthesemantisisowedtothefatthatitsufestointroduenew
strategies atthepointswhereeventualitiesbeometrue forthe rsttime. Thus,they
donotreallydependonthepositioninwhihtheyareinvokedandweanguessthem
up-front.(Or,similarly,togetherwiththepointsontheunravellingwheretheyare
in-voked.)Thisispossible,simplybeausethevalidityofstate formulas(andheneof
TCLsentenes)annotdependonthevalidityofthelefthandsideofanuntil(orthe
righthandsideofarelease)afterthersttimeithasbeensatised.
3 ExpressivenessofTCL
NotethatTCLisnotasuperlassofBSILsineBSILallowsfornegationinfrontof
SIQswhileTCLdoesnot.However,byexaminingtheproofsin[21℄forthe
inexpress-ibilityofBSIL properties byATL
, GL,andAMC,we nd thatthe BSILsentene
usedintheproofsisalsoaTCLsentene.Thisleadstotheonlusionthatthereare
propertiesexpressibleinTCLbutannotbeexpressedinATL
,GL,andAMC.
Lemma1. ThereareTCLsentenesthatannotbeexpressedinanyofATL
,GL,or
AMC.
TCLis,infat,notonlyapowerfullogi,butalsoontainsimportantlogiseither
assyntatialfragmentsoranembedtheminastraightforwardway.ATLandCTL
anbeviewedassyntatifragmentsofTCL.
ButitisalsosimpletoembedLTLandevenCTL
.Westartwith9LTL,theless
usedvariantwhereoneisontentifonepathsatisestheformula.Wethentranslatean
LTLformula,whihweassumew.l.o.g.tobeinnegativenormalform(negationsonly
infrontofatomipropositions).Thenthereisapaththatsatisesisequivalentto
h1;:::;mi b
,where b
isderivedfrombyreplaingeveryourreneof,U,andR
byh+i,h+iU ,andh+iR,respetively.Thesimpletranslationispossiblebeausethe
formula b
isde-fatointerpretedoverapath,thepathformedbythejointstrategyof
theageny[1;m℄.Theh+ioperatorswehaveaddedhavenoeffetonthesemantisin
suhaase,justasaCTLformulaanbeinterpretedastheLTLformulaobtainedby
deletingallpathquantierswheninterpretedoveraword.
Consequently,we havetheexpetedsemantisfor8LTL:allpathssatisfyis
equivalentto:hAi
:,where:isassumedtobere-writteninnegativenormalform.
Theenodingof9LTLand8LTLaneasilybeextendedtotheenodingofCTL
2 3 4 5
n
Fig.2.Theturn-basedgamegraphfromthenon-elementaryhardnessproofofextendedTCL.
Lemma2. TCLismoreexpressivethanCTL
andLTL.
ThisenodingdoesnotextendtoATL
.h1i((p)_q)isanATL
propertythat
annotbeexpressedwithTCL.
Thisis different from the ATL property (h1ip)_h1iq or the TCL property
h1i((h+ip)_h+iq).Infat,theproofsandexamplesin[21℄analsobeapplied
inthisworktoshowthattherearepropertiesofATL
(orGL,orAMC)thatannotbe
expressedwithTCL.Thisleadstothefollowinglemma.
Lemma3. TCLisinomparableinexpressivenesswithATL
,GL,andAMC.
Note,however,thatallowingforanegationinthedenitionof wouldhangethe
situation.ThenanATL
formulahAi (assumingforthesakeofsimpliitythat isan
LTLformula),wouldbeomehAi:h+[1;m℄rAi
: intheextendedversionofTCL.
ThetranslationextendstofullATL
,butthisexamplealsodemonstrateswhynegation
is banned:even withoutnesting, we an, byenoding ATL
, enodea 2EXPTIME
ompletemodel-hekingproblem,losingtheappealingtratabilityofourlogi.
Infat,itiseasytoreduetherealisabilityproblemofprenexQPTL,andhenea
non-elementaryproblem,tothemodel-hekingproblemofextendedTCL.Usingthe
gamestruturefromFigure2,weanenodetherealisabilityofaprenexQPTLformula
withn 1variables,forsimpliityoftheform8p
2 9p
3 8p
4 :::9p
n
,wherep
2 ;:::;p
n
areallpropositionsourringin.Wereduethistomodel-hekingtheformula
0
=h1i:h+2i:h+3i:h+4i:::::h+ni(
^h+ip
1 );
where
anbeobtainedfrom b
byreplaing
everyliteralp
i
byh 1ih+1i(p
i
^h+ip
i ),and
everyliteral:p
i
byh 1ih+1i(p
i
^h+i:p
i ).
TheseformulasaretehniallynotextendedTCLformulasash+ii
1
isnotpartof
theprodutionruleof ,buth+ii
1
anbeusedasanabbreviationforh+iifalseU
1 .
Chekingsatisabilityofisisequivalenttomodel-heking 0
onthegameshown
inFigure2.Thegamehasn+1nodes,agents,andatomipropositions.Thenodesin
Figure2arelabeledwiththeagentthatownedthenodes,andtheatomipropositionp
i
istrueexatlyinnodei.Fromhisstate,Agent1anmovetoanyotherstate,whileall
otheragentsaneitherstayintheirstateorreturntothestateownedbyAgent1.
ThegamestartsinthenodeownedbyAgent1,andinordertoomplywiththe
speiation,theoutermoststrategyprolehosenbyAgent1mustbetostayinthe
initialstateforever.
ishosentoalignthe truthofp
i
atpositionj 2 N withthe
deisionthat Agentimakes onthehistory1 j
i: trueorrespondstostayinginiand
falsewithreturningto1.
r
f
1 t
1
f
2 t
2
f
3 t
3
f
h+k t
h+k
1 .
.
.
Fig.3.Theturn-basedgamegraphfromtheEXPTIMEhardnessproof.
4 ComplexityofTCL
Inthissetion, we show that model-hekingTCL formulas isEXPTIME-omplete
inthe formulaandP-ompleteinthe model(andforxed formulas),while the
sat-isabilityproblemis2EXPTIME-omplete.Astheproofofinlusionofthe
satisa-bilityproblemin2EXPTIMEbuildsontheproofoftheinlusionofmodel-heking
inEXPTIME, we startwithan outlineofthe EXPTIME hardnessargumentfor the
TCLmodel-hekingproblemandthenontinuewithdesribingEXPTIMEand
2EX-PTIME deisionproeduresforthe TCLmodelandsatisability hekingproblem,
respetively.2EXPTIMEhardnessforTCLsatisabilityisimpliedbytheinlusionof
CTL*asade-fatosub-language[20℄.
WeshowEXPTIMEhardnessbyaredutionfromthePEEK-G
6
[19℄game.An
in-staneofPEEK-G
6
onsistsoftwodisjointsetsofbooleanvariables,P
1 =fp
1 ;:::;p
h g
(ownedbyasafetyagent)andP
2 =fp
h+1 ;:::;p
h+k
g(ownedbyareahabilityagent),
asubsetI P
1 [P
2
ofthemthatareinitiallytrue,andabooleanformulainCNF
overP
1 [P
2
thatthereahabilityagentwantstobeometrueeventually.Thegameis
playedinturnsbetweenthesafetyandthereahabilityagent(say,withthesafetyagent
movingrst),andeahplayeranhangethetruthvalueofoneofhisorhervariables
inhis/herturn.
Lemma4. TCLmodel-hekingisEXPTIMEhardforprimitiveTCLformulas.
Proof. ToreduedeterminingthewinnerofaninstaneofaPEEK-G
6
gametoTCL
model-heking,we introduea2-agentgameG = h2;Q;r;!;P;;Eiasshownin
Figure3,whereAgent1(he,foronveniene)representsthesafetyagentwhileAgent
2(she,foronveniene)representsthereahabilityagent.t
h+k andf
h+k
aretheonly
statesownedbyAgent2.
Thegameisplayedinrounds,andaroundstartseahtimethegameisatstater.
Ifthegamegoesthrought
i
thisisidentiedwiththevariablep
i
tobetrue.Likewise,
goingthroughf
i
isidentiedwiththevariablebeingfalse.
ItissimpletowriteaTCLspeiationthatforesthesafetyplayertotogglethe
valueofexatlyoneofhisvariablesineahround,andtotogglethevalueofthevariable
p
h+i
ofthereahabilityplayerdenedbythestateishehaspreviouslymovedto,while
maintainingallothervariablevalues.Requiringadditionallythatthesafetyagentan
gamewithonlytwoagentssufesfortheproof.Twoagentsarealsosufienttoshow
Phardnessforxedformulas,assolvingareahabilityproblemforAND-ORgraphs
[12℄naturallyreduestoshowingh1ip.
Lemma5. TCLmodel-hekingforxedformulasisPhardforprimitiveTCL
formu-las.
InordertoestablishinlusioninEXPTIMEandP,respetively,weuseanautomata
basedargument.
Theorem1. Themodel-hekingproblemofTCL formulasagainstturn-basedgame
graphsisEXPTIME-omplete,andP-ompleteforxedformulas.
Proof. WerstshowthelaimforprimitiveTCLformulas=hAi .
Tokeep the proofsimple, we rst onsider a treeautomatonU thatheks the
aeptaneof foragivenstrategysheme .Thatis,U heksifT
q +
;q; j=
undertheassumptionthatbothandthetruthvaluesforthesubformulasstartingwith
ahBiareenodedinthenodesofT
q +
.
Suhanautomatonwouldmerelyhavetorunsimpleonsistenyheks,anditis
simpletoonstrutasuitableuniversalweaktreeautomatonU,whihispolynomialin
thesizeof.FromthereitissimpletoinferadeterministiBühitreeautomatonD,
whihisexponentialintheweakuniversaltreeautomaton[16℄.
Itisthenatrivialstep(projetion)toguess andthetruthannotationofthe
sub-formulasonthey,turningthedeterministiBühitreeautomatonDthatrequiresa
orretannotationinto anondeterministiBühi automatonN ofthe samesizethat
heksG;qj=.Aeptaneanbehekedintimequadratiinthesizeofthe
prod-utofN andG[6℄.
TotakethesteptofullTCL,weanmodel-hekthetruthofprimitiveTCL
formu-lasandthenusetheresultofthismodel-hekinginsteadoftherespetivesubformula.
HardnessisinheritedfromLemmata4and5.
Thisargumentshowsmore:theomplexityofTCLmodel-hekingforxed
for-mulasdoesnotdependontheformula.ItsufestosolveanumberofBühigames,
whereboththesizeofthegameandthenumberofgamestobeplayedislinearinG.
Corollary1. ViewingthesizeofaTCLsenteneasaparameter,TCLmodel-heking
isxedparametertratable.
TheautomataonstrutionfromtheproofofTheorem1extendstoaonstrution
forsatisabilityheking.
Theorem2. TheTCLsatisabilityproblemis2EXPTIME-omplete.
Proof. Asusual,itisonvenienttoonstrutanenrihedmodelthatontainsthetruth
ofallsubformulasforaTCLsentenethatstartwithanSQ.
Inarst step, we onstrutan alternating treeautomatonA thatreognises the
thetruthassignmentofeahSQisonsistent.Butthisissimple,asweanusethetree
automatonN
0
fromtheproofforTheorem1tovalidatethelaimthatasubformula 0
ofthatstartswithanSQistrue,anditsdualtovalidatethatitisfalse.Hene,suhan
automatonhasonlytwostatesmorethanthesumofthestatesoftheindividualN
0.In
partiular,itisexponentialin.
Fortheresultingalternatingautomaton,weanagaininvokethesimulationtheorem
[16℄toonstrutanequivalentnondeterministiparityautomaton,whihhasdoubly
exponentiallymanystatesin(andwhosetransitiontableisdoublyexponentialin)
andwhoseoloursareexponentialin .Solvingtheemptinessgameofthisautomaton
reduestosolvingaparitygame,whihanbedoneintimedoublyexponentialin ,
e.g.,using[18℄.
HardnessisinheritedfromCTL
satisabilityheking[20℄.
5 ImplementationandExperiment
Asaproofofonept,wehaveimplementedamodel-heker,tl,inC++.tl
a-eptsmodelsomposedofextendedautomatathatommuniatewithsynhronisersand
sharedvariables,withanexpliitsharedvariableturnthatspeiestheturnofagents
atastate.Aturn-basedgamegraphisthenonstrutedastheprodutoftheextended
automata.Suhaninputformatfailitatesmodulardesriptionoftheinterationamong
theagents.
TheimplementationbuildsonaprototypeforaPSPACElogi[21℄.Theextension
ispossiblebeauseweanreduethe omplexityofTCL toPSPACEbysimply
re-stritingthenumberofoperatorsintheprodutionrulesinthesopeofanySQtobe
logarithmiinthesizeoftheTCLsentene.WeshowthisforprimitiveTCLsentenes.
Lemma6. Model-hekinganbedoneinspaebilinearinthesizeoftheturnbased
gamestrutureandthestateandtreeformulasthatareproduedusingthe prodution
rulesandexponentiallyonlyinthenumberofproduedtreeformulas.
Proof. Wehaveseenthat,foraprimitiveTCLsentene,weanuseasinglestrategy
shemeandonlyhavetorefertotherstpositionthattherighthandsideofanuntil
orthelefthandsideofareleaseoperatoristrue.Moreover,itsufestoguessjusta
minimalsetofpositionswheretreeformulasaretrue.Inpartiular,thelefthandside
ofarelease,the righthandsideofanuntil,andanextformulaarethenmarkedtrue
exatlyone,andtherespetivereleaseanduntilformulasneverneedtobemarkedas
trueaftersuhanevent.
Weanthereforeuseanalternatingalgorithmthatguessessuhminimaltruthlaims.
Thealgorithmalternatesbetweenaverierwhoguessesatruthassignmentandthe
ur-rent deisionsofthe strategysheme,andafalsier,whoguessesthe diretion into
whihtoexpandthepath.
Itis noweasytosee thattheywill produean innite pathinthis way,andon
thispatheahobligationthatreferstoatreesubformulafroma produtionrulean
appear onlyon a ontinuousinterval.The points wherethese obligationshangeis
by produtionrules, thisalsoinludes amarkerthatdistinguishesaleading until,
whihishangedinaroundrobinfashionwhentheleadinguntilityisfullled.)
Thenumberofpossibleassignmentsisthenexponentialinthenumberoftree
sub-formulasfromprodutionrules.Notethatformulasanbeexemptfromthisrule:
theyaremonotonousandheneinurasmallimpatsimilartotheformulasintrodued
usingthe produtionrule.
Hene,ifjGjdenotesthesizeoftheturnbasedgameandkthenumberoftemporal
operators(differentto)introduedby produtionrules,we endup inayle if
thereisnohangeinthetruthassignmenttemporaloperatorsthatareintroduedby
produtionrulesoroperatorswereahaylewithinjGjk2 k
steps.Hene,we
reahayleinanumberofstepsthatislinearinjGjandthesizeof,andexponential
onlyinthesizeof-produedtemporaloperators(differentto).
Uponreahingayle,issufestohekiftheyleisaepting.(Nostanding
obligationbyanuntil.)
Themodel-hekerusesastaktoexpliitlyenumerateallpathsofalltreetopswith
depthpresribedbyLemma6.ThetoolanbedownloadedfromSoureforgeatprojet
REDLIBat:http://soureforge.net/projets/redlib/.
Weusetheparametrisedmodelsoftheiteratedprisoners'dilemmaas our
benh-markstohektheperformaneofourimplementation.Abriefexplanationofthe
mod-elsanbefoundintheintrodution.Theuniqueparametertothemodelsarethenumber
ofprisonersm.Thereisalsoapoliemaninthemodels.Webuildaturn-basedgame
graphforeahvalueofmintheexperiments.Theparametrisationhelpsustoobserve
howouralgorithmandimplementationsaletomodelandformulasizes.Tosimplify
theonstrutionofthestate-spaerepresentation,weassumethat,ineahiteration,the
prisonersmaketheirdeisionsinaxedorder.Afterallprisonershavemadetheir
de-isions,thepoliemanmakeshisdeision.Subsequently,thewholegamemovestothe
nextiteration.Weusesevenbenhmarkformulasonthesemodelsinourexperiments.
Therstvebenhmarksaretakenfromtheexamples(A)through(E)fromthe
intro-dution.Benhmarks(F)and(G)arethefollowingtwoproperties,takenfrom[21℄.
Property(F)speiesthatallprisonersexeptPrisoner1anollaboratetorelease
Prisoner1andletPrisoner1deidetheirfate.
h2;:::;mi (h+i:jail
1 )^
V
i2f2;:::mg
(h+1i:jail
i
)^(h+1ijail
i
(F)
Property(G)speiesthatPrisoner1hasastrategytoputallotherprisonersinjail
whileleavingherfatetothem.
h1i ( V
i2f2;:::mg
h+ijail
i
)^(h2;:::;mi:jail
1
)^h2;:::;mijail
1
(G)
Forthesebenhmarks,wehaveolletedtheperformanedataforvariousparameter
valuesinTable1. Forsmall models,thememory usageisdominatedbythe normal
overhead,suhastherepresentationofvariabletables,state-transitiontables,formula
strutures,et.Thedatashowsthatourprototypeanhandlethevariousbenhmarks,
andsaleswellonveofthesevenbenhmarks.Ignoringtheoverhead,italsoshowsthe
exponentialgrowth.Themodels,however,aregrowingexponentially,too.Weassume
P
P
P
P
P
P
P properties
m
2 3 4 5 6 7 8 9 10
(A) 0.71s 0.94s 5.41s 66.3s 945s >1000s
163M165M185M350M1307M
(B) 0.50s 0.52s 0.61s 0.71s 1.11s 1.62s 5.77s 20.9s 68.1s
163M163M164M165M 168M 176M214M270M376M
(C) 0.51s 0.51s 0.6s 0.82s 1.01s 1.81s 5.54s 18.2s 48.3s
163M163M164M165M 168M 176M200M241M318M
(D) 0.5s 0.51s 0.57s 0.74s 1.01s 1.79s 7.41s 33.8s 141s
163M163M164M165M 168M 175M232M312M430M
(E) 0.51s 0.66s 19.1s >1000s
163M164M194M
(F) 0.51s 0.53s 0.61s 0.71s 1.01s 1.70s 5.38s 15.2s 53.7s
163M163M163M165M 168M 175M202M243M295M
(G) 0.52s 0.52s 0.65s 0.72s 1.03s 1.85s 4.86s 16.1s 93.5s
163M163M164M165M 169M 177M189M208M235M
s:seonds;M:megabytes.
Themodelsarewith1poliemanandmprisoners.TheexperimentwasarriedoutonanIntel
i52.4Gnotebookwith2oresand4Gmemory,runningubuntuLinuxversion11.10.
6 Conlusion
TCLisapromisinglogiforthespeiationofgroupsofagentswhobalanetheir
strategiesinordertoooperatewithdifferentpartnerstoahievedifferentobjetives.It
isaninexpensivelogiinmanyways.Firstandforemost,itisxedparametertratable.
Followingfolklore,speiationsaretinywhilemodelsarehuge.Inthissituation,xed
parametertratabilityisaveryimportantproperty,inpartiularas itisahievedbya
naturalandsimpledeisionproedure,whihismerelyexponentialintheformula.
Thisappealingpropertyisnotboughtwithinexpressiveness.Inpartiular,the
pop-ulartemporallogisLTL,CTL,ATL,andCTL
areontainedas de-fatosublogis.
Consequently,itanbeexellentlyusedtoextendexistingspeiationsinthese
lan-guages,withouttheneedtodevelopompetitivemodels.
Theappliabilityisunderlinedbyompellingdatafromourbenhmarks.Thisis
inspiteofthefatthatourimplementationisratherbasedonanadhoextensionof
anexistingalgorithmforadifferentlogi,andneitherfullyexploitthelowomplexity,
norisa fullysymboliimplementation.Itwill beinterestingtosee bywhihextent
symbolirepresentationlikeBDDswillenhanetheperformaneandhowanautomata
basedtoolwouldfare.
Referenes
1. R.Alur,T.A.Henzinger,andO.Kupferman. Alternating-timetemporallogi. Journalof
24(1):325,1980.
3. C. Baier,T. Brázdil,M.Gröser, andA.Kuera. Stohastigamelogi. InQEST,pages
227236.IEEEComputerSoiety,2007.
4. J.BühiandL.Landweber. Denabilityinthemonadiseond-ordertheoryofsuessor.
JournalofSymboliLogi,34(2):166170,1969.
5. J.BühiandL.Landweber. Solvingsequentialonditionsbynite-statestrategies. Trans.
AMS,138(4):295311,1969.
6. K.ChatterjeeandM.Henzinger. AnO(n 2
)timealgorithmforalternatingBühigames.
InProeedingsoftheTwenty-ThirdAnnualACM-SIAMSymposiumonDisreteAlgorithms
(SODA2012),Kyoto,Japan,January17-19,2012,pages13861399.SIAM,2012.
7. K.Chatterjee,T.A.Henzinger,andN.Piterman.Strategylogi.Informationand
Computa-tion,208:677693,2010.
8. E.M.ClarkeandE.A.Emerson. Designandsynthesisofsynhronizationskeletonsusing
branhing-time temporallogi. InWorkshop onLogi of Programs,volume LNCS131.
Springer-Verlag,1981.
9. A. D.Costa,F.Laroussinie, andN. Markey. Atl withstrategy ontexts:Expressiveness
andmodelheking. InIARCSAnnualConfereneonFoundationsofSoftwareTehnology
andTheoretialComputerSiene(FSTTCS2010),volume8ofLeibnizInternational
Pro-eedingsinInformatis(LIPIs),pages120132.ShlossDagstuhlLeibniz-Zentrumfuer
Informatik,2010.
10. B.FinkbeinerandS.Shewe.Coordinationlogi.InCSL,pages305319,2010.
11. G.J.Holzmann.Themodelhekerspin.IEEETrans.SoftwareEng.,23(5),1997.
12. N.Immerman. Numberofquantiersisbetterthannumberoftapeells. Journalof
Com-puterandSystemSienes,22(3):6572,1981.
13. O.Kupferman,M.Y.Vardi,andP.Wolper. Anautomata-theoretiapproahto
branhing-timemodelheking.JournalofACM,47(2):312360,2000.
14. F.Mogavero,A.Murano,G.Perelli,andM.Y.Vardi. Whatmakesatl
deidable?a
de-idablefragmentofstrategylogi. InConurrenytheory(CONCUR2012),volumeLNCS
7454,pages193208.Springer-Verlag,2012.
15. F. Mogavero,A.Murano, andM. Y. Vardi. Reasoning aboutstrategies. In IARCS
An-nualConfereneonFoundationsofSoftwareTehnologyandTheoretialComputerSiene
(FSTTCS2010),LIPIs8,pages133144,2010.
16. D.E.MullerandP.E.Shupp. Simulatingalternatingtreeautomataby nondeterministi
automata:newresultsandnewproofsofthe theoremsofRabin,MNaughtonandSafra.
TheoretialComputerSiene,141(1-2):69107,1995.
17. A.Pnueli. Thetemporallogiofprograms. In18thannualIEEE-CSSymposiumon
Foun-dationsofComputerSiene,pages4557,1977.
18. S.Shewe. Solvingparitygamesinbigsteps. InProeedingsofthe27thConfereneon
Foundations of Software Tehnologyand TheoretialComputerSiene (FSTTCS2007),
1214 Deember, NewDelhi,India,volume4805ofLetureNotesinComputerSiene,
pages449460.Springer-Verlag,2007.
19. L.J.StokmeyerandA.K.Chandra.Provablydifultombinatorialgames.SIAMJournal
onComputing(SICOMP),8(2):151174,1979.
20. M.VardiandL.Stokmeyer. Improvedupperandlowerboundsformodallogisof
pro-grams:Preliminaryreport. InProeedingsofthe17thAnnualACMSymposiumonTheory
ofComputing(STOC'85),May6-8,Providene,RhodeIsland,USA,pages240251,1985.
21. F.Wang,C.-H.Huang,andF.Yu.Atemporallogifortheinterationofstrategies.In22nd
ConurrenyTheory(CONCUR),volumeLNCS6901.Springer-Verlag,Sept.2011.
22. T.Wilke. Alternatingtreeautomata,paritygames,andmodal-alulus. Bulletinof the