Generative Network Automata: A Generalized Framework for Modeling Adaptive Network Dynamics Using Graph Rewritings HirokiSayamaandCraigLaramee 9 0 0 2 n a J Abstract A variety of modeling frameworks have been proposed and utilized in 2 complex systems studies, including dynamical systems models that describe state ] transitionsona systemoffixedtopology,andself-organizingnetworkmodelsthat O describe topologicaltransformationsof a network with little attention paid to dy- A namical state changes. Earlier network models typically assumed that topological . transformations are caused by exogenous factors, such as preferential attachment n ofnewnodesandstochasticortargetedremovalofexistingnodes.However,many i l real-worldcomplexsystemsexhibitbothofthesetwodynamicssimultaneously,and n [ theyevolvelargelyautonomouslybasedonthesystem’sownstatesandtopologies. Here we show that, by usingthe conceptof graphrewriting,both state transitions 1 and autonomoustopology transformationsof complex systems can be seamlessly v 6 integratedandrepresentedinaunifiedcomputationalframework.Wecallthisnovel 1 modelingframework“GenerativeNetwork Automata(GNA)”. In this chapter,we 2 introducebasicconceptsofGNA,itsworkingdefinition,itsgeneralitytorepresent 0 otherdynamicalsystemsmodels,andsomeofourlatestresultsofextensivecompu- . 1 tationalexperimentsthatexhaustivelysweptoverpossiblerewritingrulesofsimple 0 binary-stateGNA.TheresultsrevealedseveraldistincttypesoftheGNAdynamics. 9 0 : v i 1 Introduction X r a Avarietyofmodelingframeworkshavebeenproposedandutilizedforresearchon thedynamicsofcomplexsystems[1,2,3].Amajorclassofmodelingframeworks HirokiSayamaandCraigLaramee CollectiveDynamicsofComplexSystemsResearchGroup/DepartmentofBioengineering,Bing- hamtonUniversity,StateUniversityofNewYork,P.O.Box6000,Binghamton,NY13902-6000, USA,e-mail:[email protected],[email protected] HirokiSayamaisalsoanAffiliateoftheNewEnglandComplexSystemsInstitute,24Mt.Auburn St.,Cambridge,MA02138,USA,e-mail:[email protected] 1 2 HirokiSayamaandCraigLaramee isthatofdynamicalsystemsmodels,includingordinaryorpartialdifferentialequa- tionsanditerativemaps[4],artificialneuralnetworks[5,6],randomBooleannet- works[7,8,9],andcellularautomata[10,11].Whiletheyarecapableofproducing strikingly complex and even biological-like behaviors [12, 13, 14, 15, 16], these toolsgenerallyassumeanetworkmadeofafixednumberofcomponentsorganized in a fixed topology.Theirdynamicsare consideredastrajectoriesofsystem states inaconfinedphasespacewithtime-invariantdimensions. The recent surge of network theory in statistical physics has demonstrated yet anothergraph-theoreticapproachtocomplexsystemsmodeling[17,18,19].Itad- dressestheself-organizationofnetworkstructurevialocaltopologicaltransforma- tionssuchasrandomorpreferentialaddition,modificationandremovalofcompo- nentsandtheirinteractions(i.e.,nodesandlinks).Amongthemostactivelyinvesti- gatedissuesinthisfieldishowstatisticalpropertiesoftheentirenetworktopology will be affected by additions (growth or augmentation) and removals (failures or attacks) of nodes and links, and in particular, how networks can be more robust against the latter [20, 21, 22, 23, 24]. Those additions and removals are typically assumed as perturbationscomingfrom externalsources, not incorporatedinto the dynamicsofthenetworkitself.Theyarealsolimitedinthatnotmuchattentionhas beenpaidtodynamicalstatechangesonthenetwork.Researchersrecentlystarted investigatingdynamicalstatechangesoncomplexnetworks[25,26,27,28,29,30]. Theyarestilllargelyfocusingonfixednetworktopologiesortopologiesvariedby exogenousperturbations. When looking into real-world complex networks, however, one can find many instancesofnetworkswhosestatesandtopologies“coevolve”,i.e.,theykeepchang- ingoverthesametimescalesduetothesystem’sowndynamics(Table1).Inthese networks, state transitions of each component and topological transformations of networksaredeeplycoupledwitheachother.Understandinganddescribingtheco- evolutionofstatesandtopologiesofnetworksisnowrecognizedasoneofthemost importantproblemstoaddress[21,31].Severaltheoreticalmodelsofcoevolution- arynetworkshavebeenproposedandstudiedmostrecently[32,33,34,35,36],yet each of these studies used different model formulations for different phenomena, withlimitedimplicationsgivenforhowthesecoevolutionarynetworkmodelscould belinkedtootherexistingcomplexsystemsmodels. Hereweaimtoaddresstheabove-mentionedlackoflinkagesbetweencoevolu- tionarynetworkmodelsandotherexistingcomplexsystemsmodelsbydeveloping amorecomprehensiveformulation.Specifically,weshowthat,byusingtheconcept ofgraphrewriting,bothstatetransitionsandautonomoustopologytransformations ofcomplexsystemscanbeseamlesslyintegratedandrepresentedinaunifiedcom- putationalframework.Wecallthisnovelmodelingframework“GenerativeNetwork Automata(GNA)”[37].Thenameindicatestheintegrationofknowledgeaccumu- latedindynamicalsystemstheory,networktheory,andgraphgrammartheory. In the following sections, we will introduce basic concept of graph rewriting, a working definition of GNA, its generality to represent other dynamical systems models,andsomeofourlatestresultsofextensivecomputationalexperimentsthat GenerativeNetworkAutomata 3 Table1 Real-worldexamplesofcomplexnetworkswhosestatesandtopologieschangeoverthe sametimescalesduetothenetwork’sowndynamics. Network Nodes Links Exampleof Exampleof Exampleof nodestates nodeaddition topological orremoval changes Organism Cells Celladhesions, Gene/protein Celldivision, Cellmigration intercellular activities celldeath communica- tions Ecological Species Ecological Population, Speciation, Changesin community relationships intraspecific invasion, ecological (predation, diversities extinction relationships symbiosis, viaadaptation etc.) Epidemio- Individuals Physical Pathologic Death, Reductionof logical contacts states quarantine physical network contacts Socialnetwork Individuals Social Sociocultural Entrytoor Establishment relationships, states,political withdrawal or conversations, opinions, from renouncement collaborations wealth community ofrelationships exhaustivelyswept overpossible rewriting rules of simple binary-state GNA. The resultsrevealedseveraldistincttypesoftheGNAdynamics. 2 About GraphRewriting The key characteristic of GNA is that it should have mechanisms for transforma- tionsoflocalnetworktopologiesaswellastransitionsoflocalstates. Topological transformationsmaybemodeledasarewritingprocessoflocalnetworkconfigura- tions.Wewillthereforeadoptmethodsandtechniquesdevelopedingraphgrammar theory[38]toconstructgeneralformulationsofGNA. Graphgrammars,studiedsincelate 1960’sin theoreticalcomputerscience[39, 40, 41, 42], are an extensionof formalgenerativegrammarsin computationallin- guistics to discuss similar rule-based generativeprocesses of graphs,or networks. They recursively define a set of “valid” graph topologies that can be generated throughrepetitiveapplicationsofagivensetofnodeand/orlinkreplacementrules. Acomputationalimplementationofsuchprocessesiscalledagraphrewritingsys- tem, often used to simulate particular generative processes of network topology. Heretheword“generative”meansthatthereplacementsaretriggeredbylocaltopo- logicalfeaturesofthenetworkitself,andnotbyexternalsourcesofperturbationas typicallyassumedinmodernnetworktheory. Aclassic,andprobablymostwidelyknown,exampleofgraphrewritingsystems is the Lindenmayersystem, or L-system [43]. It is a simple rewriting system that 4 HirokiSayamaandCraigLaramee canproduceself-similarrecursivestructuresinasequentialstring(inthissense,the L-system remains within the range of classic formalgrammars).What makes this system outstandingis thatit comeswith an interpretationthat convertsa resultant stringintoatree-liketopologicalstructure,whichmayappearjustlikeanaturaltree ifparametersareappropriatelychosen.Thisexampleshowsthecapabilityofgraph rewritingsystemsto describetheemergenceofnaturalcomplexstructuresusinga setofsmalllocalrules. Although their relevance to biology was initially recognized [39, 44], appli- cations of graph grammars have so far remained within computer science, such as pattern recognition, compiler design, and data type and process specification [38,40,41,42],andtheirusehasbeennotsocommonevenwithincomputerscience duetounintuitive,complicatedformulationandlackofsoftwaretoolsformodeling [45].Moreover,mostapplicationswereprimarilyfocusedoncontext-freerewriting rules,andtheyrarelyconsidereddynamicalstatetransitionsonnetworks.Recently, context-dependentgraphgrammarshavebeenappliedtodescribereactionrulesin artificial life/artificial chemistry, includingmodels of self-replication[46, 47, 48], self-assembly[49], morphogenesis[50, 51] anddynamicstate changes[51] ofar- tifacts. However, none of them integrated graph grammars into complex systems modelinginaflexible,generalizablewaysoastobereadilyapplicabletonetworks studiedinotherdomains. To the best of our knowledge, our GNA framework is among the first to sys- tematicallyintegrategraphrewritingsintherepresentationandcomputationofthe dynamics of complex networks that involve both state transition and autonomous topologicaltransformation.Ourlong-termgoalistodevelopacomprehensivethe- oryofGNAandasetofanalytical/computationaltoolsthatcanbebroadlyapplied tothemodelingofvariouscomplexsystems. 3 Definition ofGNA A workingdefinitionofGNA isa networkmadeofdynamicalnodesanddirected linksbetweenthem.Undirectedlinkscanalsoberepresentedbyapairofdirected linkssymmetricallyplacedbetweennodes.Eachnodetakesoneofthe (finitelyor infinitely many) possible states defined by a node state set S. The links describe referential relationships between the nodes, specifying how the nodes affect each otherinstatetransitionandtopologicaltransformation.Eachlinkmayalsotakeone ofthepossiblestatesinalinkstatesetS′.AconfigurationofGNAataspecifictime t isacombinationofstatesandtopologiesofthenetwork,whichisformallygiven bythefollowing: • V: A finite set of nodes of the network at time t. While usually assumed as t time-invariant in conventional dynamical systems theory, this set can dynami- callychangeintheGNAframeworkduetoadditionsandremovalsofnodes. • C :V →S:Amapfromthenodesettothenodestate setS.Thisdescribesthe t t globalstate assignmentonthenetworkattimet.Iflocalstates arescalar num- GenerativeNetworkAutomata 5 bers,thiscanberepresentedasasimplevectorwithitssizepotentiallyvarying overtime. • L :V →{V ×S′}∗:Amapfromthenodesettoalistofdestinationsofoutgoing t t t linksandthestatesoftheselinks,whereS′isalinkstateset.Thisrepresentsthe globaltopologyofthenetworkattimet,whichisalso potentiallyvaryingover time. States and topologies of GNA are updated through repetitive graph rewriting events,eachofwhichconsistsofthefollowingthreesteps: 1. ExtractionofpartoftheGNA(subGNA)thatwillbesubjecttochange. 2. ProductionofanewsubGNAthatwillreplacethesubGNAselectedabove. 3. EmbeddingofthenewsubGNAintotherestofthewholeGNA. ThetemporaldynamicsofGNAcanthereforebeformallydefinedbythefollowing triplethE,R,Ii: • E:AnextractionmechanismthatdetermineswhichpartoftheGNAisselected fortheupdating.ItisdefinedasafunctionthattakesthewholeGNAconfigura- tionandreturnsaspecificsubGNAinittobereplaced.Itmaybedeterministic orstochastic. • R: A replacementmechanism that producesa new subGNA from the subGNA selected by E and also specifies the correspondence of nodes between the old andnewsubGNAs.ItisdefinedasafunctionthattakesasubGNAconfiguration andreturnsapairofanewsubGNAconfigurationandamappingbetweennodes in the old subGNA and nodes in the new subGNA. It may be deterministic or stochastic. • I:AninitialconfigurationofGNA. ThereareacoupleofothercommonlyusedproceduresneededtosimulateGNA dynamics,such as the removalof the selected subGNA fromthe whole GNA and the re-connectionof “bridge”links (i.e., links that were between the old subGNA and the rest of the GNA) when embedding the new subGNA. Because the work- ingsoftheseproceduresarefairlyobvious,weomitdetailedexplanationsforthem. TheaboveE,R,I aresufficienttouniquelydefinespecificGNAmodels.Theentire pictureofarewritingeventisillustratedinFig.1,whichvisuallyshowshowthese mechanismsworktogether. Thisrewritingprocess,ingeneral,maynotbeappliedsynchronouslytoallnodes or subGNAs in a network, because simultaneous modifications of local network topologiesatmorethanoneplacesmaycauseconflictingresultsthatareinconsis- tent with each other. This limitation will not apply though when there is no pos- sibility of topologicalconflicts, e.g., when the rewriting rules are all context-free, orwhenGNAisusedtosimulateconventionaldynamicalnetworksthatinvolveno topologicalchanges. We notethatitisauniquefeatureofGNAthatthemechanismofsubgraphex- traction is explicitly described in the formalism as an algorithm E, not implicitly assumedoutsidethegrammaticalruleslikewhatothergraphrewritingsystemstyp- ically adopt (e.g. [51]). Such algorithmic specification allows more flexibility in 6 HirokiSayamaandCraigLaramee Fig.1 GNArewritingprocess.(a)TheextractionmechanismEselectspartoftheGNA.(b)The replacementmechanismRproducesanewsubGNAasareplacementoftheoldsubGNAandalso specifiesthecorrespondenceofnodesbetweenoldandnewsubGNAs(dashedline).Thisprocess mayinvolvebothstatetransitionofnodesandtransformation oflocaltopologies. The“bridge” linksthatusedtoexistbetweentheoldsubGNAandtherestoftheGNAremainunconnectedand open.(c)ThenewsubGNAproducedbyRisembeddedintotherestoftheGNAaccordingtothe nodecorrespondencealsospecifiedbyR.Inthisparticularexample,thetopgraynodeintheold subGNAhasnocorrespondingnodeinthenewsubGNA,sothebridgelinksthatwereconnected tothatnodewillberemoved.(d)Theupdatedconfigurationafterthisrewritingevent. representing diverse network evolution and less computational complexity in im- plementingtheirsimulations,significantlybroadeningtheareasofapplication.For example, the preferential attachment mechanism widely used in modern network theory to construct scale-free networks is hard to describe with pure graph gram- marsbutcanbeeasilywritteninalgorithmicforminGNA,asdemonstratedinthe nextsection. Whilethedefinitiongivenaboveisoneofthesimplestpossibleformulationsof GNA, it already has considerable complexity compared to conventional dynami- cal systems models. The possibility of temporalchanges ofV and L particularly t t GenerativeNetworkAutomata 7 makesitdifficultto investigateits dynamicalpropertiesanalytically.However,the updating process of GNA is algorithmically described and hence their dynamics canbeexperimentedthroughcomputersimulationrelativelyeasily.Wehavedevel- opeda packagein Wolfram Research Mathematicaforsmall-scale simulationand visualizationofGNAwithnodestates1.Theresultspresentedinthischapterwere obtainedusingthispackage. 4 Generality ofGNA TheGNAframeworkishighlygeneralandflexiblesothatmanyexistingdynamical networkmodelscanberepresentedandsimulatedwithinthisframework. Forexample,ifRalwaysconserveslocalnetworktopologiesandmodifiesstates ofnodesonly,thentheresultingGNAisaconventionaldynamicalnetworkmodel, including cellular automata, artificial neural networks, and random Boolean net- works(Fig.2(a),(b)).AstraightforwardapplicationofGNAtypicallycomeswith asynchronousupdatingschemes,asintroducedintheprevioussection.Sinceasyn- chronousautomatanetworkscanemulateanysynchronousautomatanetworks[52], the GNA framework covers the whole class of dynamics that can be produced by conventionaldynamicalnetwork models. Moreover,as mentionedearlier, syn- chronousupdatingschemescouldalso be implementedin GNA forthis particular classofmodelsbecausetheyinvolvenotopologicaltransformation. Ontheotherhand,manynetworkgrowthmodelsdevelopedinmodernnetwork theorycanalsoberepresentedasGNAifappropriateassumptionsareimplemented inthesubGNAextractionmechanismEandifthereplacementmechanismRcauses nochangeinlocalstatesofnodes(Fig.2(c)). 5 Computational Exploration ofPossibleDynamicsofSimple Binary-State GNA In this section we reportour latest results of extensivecomputationalexperiments that exhaustivelyswept over possible rewriting rules of simple binary-state GNA. Theresultsshownherewereobtainedwithmuchlessrestrictedrulesetsthanthose assumedinourpreviouswork[37]. 1TheMathematicapackageisstillunderactivedevelopmentbutmaybeavailableuponrequest. 8 HirokiSayamaandCraigLaramee Fig. 2 Various dynamical network models simulated using GNA. These examples were repre- sentedinthesameformatofhE,R,Ii(seetext)andsimulatedusingthesamesimulatorpackage implementedinMathematica.(a)Simulationofasynchronous2-Dbinarycellularautomatawith vonNeumannneighborhoods andlocalmajorityrules.Spacesize:100×100.(b)Simulationof anasynchronousrandomBooleannetworkwithN=30andK=2.Timeflowsfromlefttoright. NodesofrandomBooleannetworksarenon-homogeneous,i.e.,theyobeydifferentstate-transition rules.Hereeachnode’sownstate-transitionruleisembeddedaspartofitsstate,andthereplace- mentmechanismRreferstothatinformationwhencalculatingthenextstateofanode.(c)Simu- lationofanetworkgrowthmodelwiththeBaraba´si-Albertpreferentialattachmentscheme.Time flowsfromlefttoright.Eachnewnodeisattachedtothenetworkwithonelink.Theextraction mechanismEisimplementedsothatitdeterminestheplaceofattachmentpreferentiallybasedon thenodedegrees,whichcausestheformationofascale-freenetworkinthelongrun. GenerativeNetworkAutomata 9 5.1 Assumptions There are infinitely many possible mechanisms for E and R because there are no theoreticalupperboundsin termsof the size of the old subGNA selected by E (it could be infinitely largeas the GNA grows) and the new subGNA producedby R (itcouldbearbitrarilylargebythedesignofR).Makingreasonableassumptionsto restrictthepossibilityofmechanismsforE andRiscriticaltofacilitatesystematic studyonthedynamicsofGNA.Herewemakethefollowingassumptions(Fig.3): 1. Nodestatesarebinary(0or1). 2. Nolinkstateisconsidered(i.e.,linkshomogeneouslytakeonlyonestateandit willneverchange). 3. Links are undirected(i.e., every connectionbetween nodes is represented by a pairofsymmetricallyplaceddirectedlinks). 4. TheextractionmechanismE alwaysselectsasubGNAby a. randomlypickingonenodeufromtheentireGNA(Fig.3(a)), b. takingallthedestinationnodesofitsoutgoinglinksL(u)(Fig.3(b)),and t c. producinga subGNA “induced”by these nodes{u}∪L(u), i.e., a subGNA t that includes all these nodes as well as all the links present between them (Fig.3(c)). 5. ThereplacementmechanismRonlyreferstothestateofthecentralnodeuand thelocalmajoritystate withintheinducedsubGNA.Ifthereareequalnumbers of 0’s and 1’s within the subGNA, one of the two states is randomly chosen. Thistwo-bitinformationwillbeusedtodeterminewhatwillhappentothelocal configuration (Fig. 3 (d)). The following ten possible rewriting outcomes are madeavailable(whichareextendedfrom[37]): 0) Thecentralnodeudisappears. 1) Everythingremainsinthesamecondition. 2) Thestateofthecentralnodeuisinverted. 3) Thecentralnodeudividesintotwowiththestatepreservedinbothnodes. 4) Thecentralnodeudividesintotwowiththestateinvertedinbothnodes. 5) Thecentralnodeudividesintotwowiththestateinvertedinonenode. 6) Thecentralnodeudividesintothreewiththestatepreservedinallthreenodes. 7) The central node u divides into three with the state inverted in all of three nodes. 8) The central node u divides into three with the state inverted in two of three nodes. 9) The central node u divides into three with the state inverted in one of three nodes. Incaseswherenodedivisionoccurs,thelinksthatwereconnectedtothecentral nodeuisdistributedasevenlyaspossibletoitsdaughternodes(Fig.3(e)). 6. TheinitialconditionI consistsofasinglenodewithstate0. 10 HirokiSayamaandCraigLaramee Fig.3 SimplifiedGNArewritingprocessusedfortheexhaustivesweepexperiments.Theextrac- tionmechanismE(a)randomlypicksonenodeu,(b)takesallthedestinationnodesofitsoutgo- inglinksLt(u),and(c)producesasubGNAinducedbythosenodes{u}∪Lt(u).Thereplacement mechanism R(d)refers tothestateofthecentral node uintheselected subGNAandthelocal majoritystatewithinittodeterminewhathappenstothelocalconfiguration,(e)producesanew subGNAaswellasthecorrespondenceofnodesbetweentheoldandnewsubGNAsbasedonthe choicemadein(d),andthen(f)embedsthenewsubGNAintotherestoftheGNA. Notethattheabovemodelassumptionswillalwaysgenerateplanargraphsinwhich thenodedegreesareboundeduptothreewheninitiatedwithasinglenode.There- forealltheresultsshowninthischapteraretopologicallyplanar. 5.2 Methods We carried out an exhaustivesweep of all the possible rewriting rules that satisfy the assumptions discussed above. Since the extraction mechanism E is uniquely defined,itisonlythereplacementmechanismRthatcanbevaried.HereRisdefined as a function that maps each of the four possible two-bit inputs to one of the ten possibleactions.Thereforethe numberofallthe possibleR’sis1022 =10000.To indicateaspecificR,wewilluseits“rulenumber”rn(R)thatisdefinedby rn(R)=a ×103+a ×102+a ×101+a ×100, (1) 11 10 01 00 where a is a numerical representation (numbers associated with each of the ten ij possibleactionsshownabove)ofthechoicethatRwillmakewhenthestateofthe centralnodeuisiandthelocalmajoritystateis j.