1This paper gives a tical description of Creatures,a coercial home-ent - PDF document

1This paper gives a tical description of Creatures,a coercial home-entertainment software package.Creatures prides a sulated environment inwhich exist a number of synthec agents that a usercan interact with in real-e. The agents (known asreatures”) are intended as “virtual pets”. Theinternal architture of the creatures is inspired byanal biology. Each creature has a neural networkresponsible for sensory-motor crdinaon andbehavior selon, and an “arficial biochemistry”that models a sple energy metasm along wh a“hormonal” system that interacts wh the neuralnetwork to model diuse molaon of neuronalacvity and staged ontogenec development. AHeian learning mhanism allows the neuralAddonally, th the network archture anddetas of the biochestry for a creature are specifiedby a variable-length “genec” encoding, aowing foreluonary adaptaon throseal reprocon.Creatures, avaable on Windows95 and Macintoshplatforms from late , oers users an oortunyto engage wh Arficial Life tecologies. Inaddon to describintical details, this paperconcludes wh a discuion of the scienficAutonomous software agents have significant potential forappcation in the entertainment instry. In this paper, wediscuss an interacve entertainment proct based onties developed in Arficial Life and AdapveBehavior research (e.g. Brks and Maes , Cff et al1994). The proct, caed Creatures, aows human usersto interact in real-e wh synthec agents which inhaba closed environment. The agents, known as reatures”,have arficial neural networks for sensory-motor controland learning, arficial biochestries for energymetasm and hormonal relaon of behavior, and ththe network and the biochestry are “genecally”spified to aow for the poiby of eluonaryAhoh a coercial proct,1 we beevaspects ofCreatures will be of interest to the science andenginring counities. This paper discues the most1 Creatures was develed by Meium Interacve Ltd and is pubshedby Warner Interacve and Inscape. The core Arficial Life teciquesdeveled by Meium for use in Creatures are refeed to assignificant aspts of the proct relevant to autonomousagent researchers. The proct, avaable in NorthAmerica and Europe from late , runs in real-e onSon 2 discues related work. Son 3 presents adescripon of tical aspts of Creatures, and Son4 concludes wh some speculavoents on theHere we briefly suarize work in Arficial Life andAdapve Behavior research that is relevant to Creatures.For bacround discuioand descripon of a selonof other entertainment-oriented research projts, s(Senal work by Reynolds () estabshed thepoiby of using autonomous agents for behavioralanation, a tie which aows mie seensshowing behavior in synthec agents to be prod whthe human anator giving only broad horeographic”coands, rather than detaed frame-by-frame posespificaons. Subseent related projts, such as that byTerzopolous et al (), where fahful kinemacsulations of fish are modeled wh preive visualauracy and considerable biological plausibility in thebehavioral control, have shared wReynolds' originalwork a reance on skful manual design of the agent'sphysical morphology, behavioral control mhanism, orth. This can often reire significant investment ofFad wh the diicult task of designing felikesynthec agents for entertainment acations, severalresearchers have drawn inspiraofrom biology. Forexample, Blumberg (1994) has developed a behavioralcontrol mhanism inspired by findings in ethology (thescience of anal behavior) which is used to control asynthec dog that iabs a 3D software environment,interacng wh a human user and wh other virtualOther researchers have worked on developingties that rece the reanon sked lar byincorporang some type of automac adaptaon orlearning mhanism in the agent software. Reynolds(1994) has elored the use of genec prograng(Koza ), a tie related to genec algorhms(Goldberg 1986), to develop control programs forsynthec agents ming in 2D worlds wh splifiedkinemacs. Ss () has employed sar arficialeluon ties to develop th the physicalmorphology and the arficial neural network controers 2for synthec autonomous agents that iaba 3D worldwh reastic kinemacs. There is also a large dy ofwork on learning in arficial neural networks (see e.g.Pubcations in the scienfierature describingcoercial interactive entertainment software proctsare rare, so the cations in this son are to promoonalOne of the first pis of entertainment softwareexpcitly promoted as drawing on Arficial Life researchwas SLife by Maxis, released in 1993 (Maxis, )Ineen, SLife aowed a user to observe and interactwh a “sulated osystem” wh a variable teain andcate, and a variety of spies of plant life, herbirousanals, and carnirous anals. The ecosystem wassulated using ular automata ties, and soMore recent procts have had stronger nks toautonomous agent research. Another Maxis proct, El-Fish, was presented as an "eltronic aarium" whereusers could design and brd virtual tropical fish whichcould then be served swnin a virtual fishtank.The sarities betwn this proct and the work ofIt should also be noted that Maxis pionred the conceptof “software toyss oosed to omter games”. Themetaphor of “toy” rather than “game” is intended tohighght a dierent style of interacon: a game is usuayplayed in one (extended) seion, until an nd condon”is reached, and a score or high-score is awarded; incontrast, use of a toy does not imply a score or an aim toachieve some end-condon, and interacon wa toy is aSubsequeny, Magic Inc. (Magic ) hasreleased two procts, Dogz and Catz which give users on-scrn anations of virtual pets based on dogs and cats.Users can interact with their virtual pets and train them toperform simple tricks. There is some sarity betwnthese procts and Blumberg's work menoneae.Two procts aound but not released at the ofwrng are Fin-Fin by Fujsu Interacve Inc. (Fujsu1996) and Galapagos by Anark (Anark 1996). Of the two,Galapagos has stronger nks to Arficial Life, inlvinga 3D kinemacally reastic model of a six-leed agent ina 3D me-ke environment with an adapve neural-network-ke controer based on Anark's proprietary“NERM” (Non-staonary Entropic Recon Maing)tologyThe Fin-Fin product inlves 3D rendering ofa hybrid dolphin/bird creature which the user can engagein sple interacons wh via a spiazed input devicombining a proxy dettor wh a crophone whichAll of Dogz, Catz, Fin-Fin and Galapagos arepresented as inlving Arficial Life tecologies, butnone of them (yet) employ genecallencoded neuralnetwork archtures or arficial biochestries as usedin Creatures, which we describe in So3. Nor do theyaow for the development of uure” in counities ofarficial agents, a poiby wh Creatures which weThe creatures iabit a “two-and-a-half densional”world: evely a 2D platform environment with mu-plane depth cueing so that jts can aear, relave tothe user, to be in front of or behind each other. On atypical Windows95 system, the world measuresapproxately 15 scrns horizontay by 4 scrnsvercally, wh the window scrong smthly to foow aselted creature. Whin the world there are a number ofjts which the creature can interact with in a variety ofways. The system has bn wreusing jt-orientedprograng ties: virtual jts in the world suchas toys, fdetc. have scripts aached that determine howthey interact with other jts, incling the creatureagents and the staparts of the environment. Somejts are “automated”, such as elevators which rise/fawhen a buon is pressed. Other jts and environmentsmay be aed later. A scrnshot showing a view of part1When the user's mouse pointer is anywhere whin theenvironment window, the pointer changes to an age of ahuman hand. The user can me jts in theenvironment by picking them up and dropping them, andcan aract the aention of a creature by waving the handin front of , or by stroking it (which generates a posve,“reward” reinforment signal) or slaing it (to generateA typical creature is shown in Fire 2. All creaturesare bipedal, but minor morphological details such ascoloring and hair type are genecally spified. As theygrow older, the on-screen size of the creature increases, until “maturity”, aroxately one third of the way 3through their fe. The fe-span of each creature isgenecally influend: if a creature manages to survive toold age (measured in game-hours) then senesnce genesmay bome acve, kng the creature. The creature hassulated senses of sight, sound, and touch. All aremodeled using se-symc aroxation ties.For example, the sulation of vision does not inlve asulation of opcs or proing of renal images.Rather, if a rtain jt is whin the nof sight of acreature, a neuron represenng the presence of that jtin the visual field bomes acve. Such aroxations tothe end-result of sensory proing are fairly coon inneural network research. Sounds aenuate er distanand are muled by any jts betwn the creature andthe sound-source. An jt can only be seen if thecreature's eyes are poinng in s diron. There is also asple focus-of-aention mhanism, described further2Creatures can learn a sple verb-jt lanage, ehervia keyard input from the user, or by playing on ateaching-machine in the environment, or from interactionswh other creatures in the environment. On typical targetplatforms, up to ten creatures can be acvat one ebefore serious degradaon of response-e occurs. Thefoowing sons describe in more detail the neuralEach creature’s brain is a heterogeneous neural network,sub-divided into objts caed ‘les’, which define theeltrical, checal and morphological characteristics of agroup of s. Ces in each le form coons to oneor more of the ces in up to two other source les toperform the various funcons and sub-funcons of the net.The network archture was designed to be biologicayplausible, and computable from the ‘om-’, wh veryThe inal model contains aroxately 1,neurons, groed into 9 les, and intercotethroughroughly 5,000 synapses. However, all these parameters aregenecally controed and may vary ring laterLe 1Le 23The structure of the neural architture was designed to· it must be very eicient to comte (a world wtencreatures reirethe proceing of some ,000neurons and 100,000 synapc coons everysond, in aoto the load poseby the display· it must be capable of sorng the plaed brainmodel, i.e. the neural confiraon which controls the· it must be capable of ereing many other poible· it must not be too bre—mutation and rombinaonshould have fair chance of construcng new systemsNeurons. All the neurons whin a single le share thesame characteriscs, but these characteriscs can varyer a wide range of poible behaviors. Some aspects ofthe neurons’ dynacs are splparameters, whe othersare defined as ereions. All of these factors arecontroed genecally ring the construcon of a le.1StateExponential recovery rate from(A neuron’s internal state is comted via a genecallydefined funcon known as a State-Variable Rule, orSVRule. SVRules are composed of interpreted opcodesand operands, and are also used to control several asptsof synapc behavior. An SVRule ereion is designed tobe interpreted extremely rapidly, and also to be non-breand fa-safeenec mutaons can never cause syntax 4eors. SVRules can comte new state values in manyways (see Table 2). Many of these poiblfuncons gowell beyond the present nds of the ‘brain model’, but areprided in order that a powerful tl-ks avaable forfuture man-made or eluonary prements to the2 State is sum of type0 inputs or zero if not all intsState is raised by cuent inolated by chemo-After a neuron’s State is comted, a ‘relaxaon’funcon is aed to , which eonenally returns towards a definable ‘rest state’. One portant use of thisrelaxaon funcon isto act as a dampingmhanism, since thefurther the neuron’sstate gets fromequbrium, the fasterit relaxes, and so theharder it bomes todisturb it further. Thistension betwintand relaxaootonly kps the systemreasonably stable, butcan also provide an integraoof int signals, such thatthe state of the neuron reflts th the intensy and theDerites. Eaceuron is fed by signals from one ormore dendres. Each ll may caone or two dierentclaes of dendre, each wdierent characteristics andsource les, thus aowing for the integraon of differenttypes of data. The main parameters for a dendre/synapse3The signal aiving at the synapse is molated by theShort-term igho provide an outt value. A rise inSTW can be triered by a reinforment SVRule usuayin response to acvity at a chemo-rptor. Afterdisturban, th the STW and the LTW relaxexponenally towards other, wh the LTW being theslower. The STW therefore reacts strongly to indivialreinforment episodes, whe the LTW evelycomputes a ming average of many STW disturbances: ifa creature meets wh suation X and finds that its chosencourse of acon was undesirable, then it shouldediately be strongly disincned to repeat the acon,espiay as many of the incenves to do so may still bepresent. However, suation X may not always be as bad asfirst eerience sests, and so the creature’s long-termDeritic Migration. The inal wiring is defined atbirth aording to a small number of genec rules.Generay, neurons aempt to cot from one ltoanother in a dirt spatial maing, wh mupledendres faing out in a spified distribuon to eherside of the opum source cell (see Fir1). After birth,however, indivial dendrites may grate and form newcoons (always whin the same source le).Periodicaya Strength value change is comted for eachsynapse using SVRules, often in response to checalchanges. If the Strength fas to zero, the dendrediscots and foows the aropriatrule aut how tofind a new coon. These gration rules were chosenin order to fulfill the reirements for the inal brainmodel. It was hoped that a more general scheme could beinvented, but this was not poible in the avaable.An extra gration funcon, inlving a survival-of-the-fest competition betws for the right to represent aparcular int pattern, was plemented as part of themodel’s generaon system, but has caused prlemsThe aarchture is a generazed engine for neuron-ke comtaon, whose circury can be definedgenecally. This son describes the specific modelwhich has bn serposed onto the system torest levelReltionthresld4 5Conpt Le(640 neus)Percepon LeDecision LeVerbsMiscDrivesSmSourAentionNosAangement of lobes5: Brain ModelAention. Some of the neural circuits are deted torelavely nor tasks. For example, two lobes are used toplement an aention-dirng mhanism. Suliaiving from jts in the environment cause parcularll to fire in an int le (where each ll represents adierent class of jt). These signals are maed one-on-one into an outt le, which sums the intensy andfrequency of those suli er e. Sulated lateraliibon aows these ces to compete for control of thecreature’s aention. The creature’s ge (and thereforemuch of his sensory aaratus) is fixed on this jt, andit bomes the recipient for any acons the creaturechses to take. Such a mhanism limits creatures to“verb objt”, as oosed to “subjt verb objt” modes ofthought, but serves to rece sensory and neuralproing to aptable levels, since the net nd onlyDision Making. The bulk of the remaining neuronsand coons make up three les: a ‘perpon’ le,which combines several gros of sensory ints into onepla; a large region known as Conpt Spa, in whichevent memories are laid down and eked; and a small butmaively dendrc laed the Decision Layer, whererelaonship memories are stored and acon disions gettaken. The erall model is behaviorist and based onCes in Conpt Space are splpaern-matchers.Each has one to four dendres and computes outt bying the analog signals on s ints, which come viathe Percepon le from sensory systems. Each thereforefires when all of s ints are firing. These ces arerandoy wired at birth, but sk out new paerns as theyour. Once a ll has coed to a parcular paern, remains coted until its dendres’ strengths all fall tozero. A biochecal fdback land two SVRulesaempt to maintain a pl of uncoed neurons wheleaving ‘useful’ (i.e. repeatedly reinford) s cotedfor long periods. The Percepon le has around 128sensory ints, and so the total number of s that wouldbe reired to represent all poible sensory permutaonsof up to four ints is unfeasibly large. Thisreinforment, atrophy and gratiomhanism isdesigned to get round this prlem by rording only theporon of int space which turns out to be relevant.There are a number of prlems associated wthisThe Decision layer comprises only 16 s, eachrepresenng a single poiblacon, such as “acvate ”,“deacvate ”, “walk west”, and so on, where “” is thecueny aended-to objt. The Decision neurons arehighly dendrc and ffrom ConpSpa. Thedendres’ job is to form relationships betwn Conpts and acons, and to rorin their synapcweighngs how aropriate each acois in any givenAn SVRule on each dendre decides the cuentsynapc ‘suscepbility’, i.e. sensvity to molaon byreinforrs. This is raised whenever that dendrite isconducng a signal to a ll and that s firing (i.e. thecoon represents th a ‘true’ condon and also thecuent action). It then days eonenally er e.Synapses are therefore senszed when they representrelaonships betwn cuent sensory schemata and thelatest action dision, and remain sensve for a period inorder to respond to any share of a more-or-less defeedThere are not enoh dendres to cot every aconto every Conpt , and so these dendres are alsocapable of grating in search of new sours of signal.Again a biochecal fdback lcontrols atrophy, wheDision s sum their ints into their current state(in fact they sum their type 0 inputs (excory) and subtractthe sum of their type 1 (iibory) ints). The relaxaonrate of Dision s is moderate, and so each aumulates a number of nes er a short period, basedon the number of Conpt s which are firing, plus theirintensy. The strongest-firing Dision ll is taken to bethe best course of acon, and whenever the wierGenazation. Bause Concept Space seeks torepresent all the various permutaons of one to four intsthat exist within the total sensory suation taining at agiven moment, the system is capable of generazing frompreviously learned relaonships to novel situations. Twosensory suations can be deemed related if they share oneor more indivial sensory features, for example suationABCD, which may never before have been eeriend,may eke memories of related suations such as D, ABD,etc. (ahoh not BCDE). Each of these sub-suationsrepresents previously learned experience from one or morerelated suations and so each can oer useful advice onhow to react to the new suation. For example, “I findmyself lking at a big, grn thing wh staring eyes,which ve never sn before. I remember that going up tothings wh staring eyes and kiinthem is not a gdidea, and that hitting big things, parcularly big, grnthings, doesn’t work well either. So, all in a, I think try something else this e.” Of course, if the newsuation turns out to have dierent aes frompreviously eeriend sub-suations (an ‘expoto the 6rule, then tthe new, total ‘conpt’ and thepreviously learned sub-conpts will be reinforcedaordingly. As long as ser-conpts fire more stronglythan sub-conpts, and as long as reinforment issupped in proporoto ll outt, the creature cangraduay learn to discrnate between these aciredmemories and so form ever more useful generalions forLearning. Delayed-reinforment learning is pridedby changes to Dision Layer short-term weights inresponse to the existence of eher a Reward checal (forexcatory synapses) or a Punishment chemical (foriibory ones). These checals are not generateddiry by environmental stimuli but ring checalreacons inlved in Drive level changes. Each creaturemaintains a set of checals represenng ‘drives’, such as“the drive to avoid pain”, “the drive to rece hunger”,and so on. The higher the concentraon of each checal,the more preing that drive. Environmental stimuli causethe procon of one or more drive raisers or driveredurs: chemicals which react to increase or drease thelevels of drives. For example, if the creature takes ashower by acvating a shower jt, the shower ghtrespond by recing hotness and coldness (normazingtemperature), dreasing redness and increasingslpineDrive raisers and redurs proce Punishment®®Drive recon therefore increases the weights ofexcatory synapses whe drive increase reinforcesiibory ones. Of course, recing a non-present drivehas no et, and so the balance of nishment to rewardmay reverse. Thus, many acons on jts can return anet nishmenor a net reward, aording to thecreature’s internal state at the e. Creatures thereforeThe brain model is not an ambous one, and severelys the range of cognve funcons which can arise. Itis also prvely Behaviorist in s reinformentmhanism. However, it serves rpose by priding alearned logic for how a creature chses acons, anddoesn’t suer from too many non-fe-like side ets: itsin-built generalion mhanism res arbrarineinthe face of ney, and the dynacal structure, albedamped and close to ebrium, pros a sasfactorilycomplex and beevable seenof behaviors,surprisingly free from limit cycles or irretrievable coapseCentral to the funcon of the neural net is the use of asplified, sulated biochestry to control widespreadinformaon flow, such as internal fdback lps and theexternal drive-control system. This mhanism is alsoused to sulate other endocrine funcons outside thebrain, plus a basic metasm and a very sple unesystem. The biochestry is very straightforward and isChemicals. These are just arbitrary numbers in therange 0 to , each represenng a dierent chemical andeach aociated wh a number represenng s cuentconntraon. Checals have no ierent properties—the reacons which each can undergo are definedgenecally, wh no restricons based on any in-buchecal or physical characteristics of the moleculesEmis. These checals are prod by Chemo-eer jts, which are genecally defined and can beaached to arbrary byte values whin other systemjts, such as neurons in the brain or the outts ofsensory systems. The locus of aachment is defined by adescriptor at the start of an eer gene, represenng‘organ’, ‘ue’ and ‘se’, foowed by codes for thechecal to be eed and the gain and othercharacteriscs of the eer. Changes in the value of abyte to which an eer is aached will automaticallycause the eer to adjust its outt, whout the codewhich has caused the change needing to be aware of theReactions. Checals undergo transformaons asdefined by Reacon jts, which spify a reacon inthe form iA + [jB] [kC] + [lD], where i,j,k deterneraos. Most transformations are aowed, expt forReacons are not defined by utable checal laws butby genes, which spify the reactants and reaconproducts and their proporons, along wh a value for thereacon rate, which is conntraon-dependent andRptors. Checal conntraons are monored byChemo-rptor jts, which aacto anset arbitrarybytes defined by locus IDs, as for eers. Rptor genesspify the locus, the checal that the receptor respondsto, the gain, the threshold and the nonal outt. Manyparts of the brain and dy can have receptors aached,Biochemical structus. Aachinrptors andeers to various loci within brain les aowswidespread fdbacaths whin the brain, parcularly incombinaon wreacons. Paths have been plementedto control synaptic atrophy and gration, and also topride drive-recon and learning reinforment. Otherneurochecal interactions are poible, such as thecontrol of arousal. However, these have not bnplemented, and we wait to see whether Nature canAs well as controng val neural systems,biochestry is used to plement those systems which are 7not actually naror comlsory whin digalorganisms, yet which would be eted by the generalpubc. For example a sple metasystem is Û
22Sarly, a selon of biochecals and reaconsproduce the ets of toxins, which may be ingested fromplants or eed by the various synthec ‘bacteria’ whichiabit the environment. These bactericay various‘angens’, which inke ‘andy’ procon in thecreaturescausing a very splified une response. Thebacterial polaon is aowed to mutate and elve,potenally aing a e co-eluonary spice to theAs much as possible of the creature’s structure andfuncon are deterned by s genes. Prarilythisgenome is prided to aow for iered characteriscs—our users et their new-rn creatures to showcharacteriscs idenfiably drawn from each parent.However, we have also gone to considerable trouble toensure that genomes are capable of eluonarydevelopment, incling the introcon of nelThe genome is a string of bytes, divided into isolatedgenes by means of ‘nctuaon marks’. Genes ofparcular types are of characterisc lengths and containbytes which are interpreted in spific ways, ahoh anybyte in the genome (other than gene markers) may safelymutate into any 8-bit value, whout fear of crashing theThe genome forms a single, haploid chromosome.During reproducon, parental genes are croed andspd at gene undaries. Oasional croer errorscan introce gene oions and dupcations. A smanumber of random mutations to gene dies is alsoapped. To prevent an exive faurrate e toreproducon eors in crcal genes, each gene is prdedby a header which spifies which operaons (oion,dupcation and mutaon) may be performed on .Croing-er is performed in such a way that genenkage is proporonal to separaon distan, aowing fornked characteriscs such as ght be eted (forexample, temperament with facial type). Bause thegenome is haploid, we have to prevent useful sex-linkedcharacteriscs from being eradicated sply bause theywere iered by a creature of the oossex. Therefore,each gene caies the geneinstrucons for th sexes,but only the un-sexed and appropriately sexed genes getEach gene’s header also contains a value deternings swch-on e. The genome is re-scaed at intervals,and new genes can be ereed to cater for changes in acreature’s structure, aearance and behavior, for exampleSome of our genes sply code for outwardcharacteriscs, in the way we speak of the “gene for redhair” in humans. However, the vast majority code forstructure, not function. We could not emulate the fact thatreal genes code only for proteins, which prostructureswhich in turn proce characteriscs. However,we have tried to stay as true as we can to the principle thatgenotype and phenotype are separated by several orders ofabstracon. Genes in our creatures’ genomes thereforecode for structures such as chemo-rptors, reacons andbrain les, rather than outward phenomena such asIt is diicult to provide any “resus” in this paper, sinthe project was essenally an exercise in enginring,rather than scien. The erall jve was to createsynthec, biological agents, whose behavior wassuicieny fe-like to sasfy the etaons of thegeneral bc. Sales fires will be our resus, and at thee of wrng, the proct is still to be launched.However, in subjve terms, we have achieved most ofour as: the behavior of the creatures is dynacally“interesng” and varied and they do indd appear tolearn. Oasional examples of aareny emergent“social” behavior have been served, such as cperaonin playing wh a ba, or hase” snes resung from“unrequed le”. However, it is very diicult to estabshhow much of this is genuine and how much is confeedby an server’s tendency to anthropomorphism. Thedynacal behavior of the agents and erall environmenthas bn grafyingly stable, and configuring a usablegenotype has not bn a prlem, despe reiringapproxately 320 interacnenes, each wseveralparameters. From that point of view, our beef that such acomplex synthesis of sub-systems was an achievable aWe beevthat Creatures is prably the only coercialproduct available that allows home users to interact witharficial autonomous agents whose behavior is controedby genecally-spified neural networks interacng wh agenecally-spified biochecal system. As the creaturesare responsible for crdinang perpon and acon forextended periods of e, and for maintaining suicientinternal energy to survive and mature to the point wherethey are capable of seal reprocon, it could plausiblybe ared that they are instances of “strong” arficial life,i.e. that they exhibit the neceary and suicientcondons to be described as an instance of fe. Naturay,formulang such a st of condons raises a number ofphosophical diicues, and we do not claim here thatthe creatures are ave. Rather, we note that thephosophical debate concerning the poiby of, andrequirements for, strong arficial life, will be raised in thends of many of the users of Creatures. As such, the“general bc” will be engaging wh arficial lifetologies in a more complete maer when using 8Creatures than when using the other procts menonedin Son 2.2. Furthermore, if we aumthat each userruns 5 to 10 creatures at a e, then after a few months ofreasonable sales around the world, it is possible that therewill be ons, or even tens of ons, of creaturesexisng in the “cyberspa” prided by the machines ofthe glal Creatures user counity. In this sense, theuser counity will be helping to create a “digalbiodiversy reserve” sar to that adcated by T.Ray inhis ongoing work on NetTiea, a major glal ArtificialLife research eerent (Ray ,1996). If we chseto, we can monor the eluon of parcular features ingroups of creatures: on a local scale there may be evariaon, but national or glal comparisons may revealdivergent eluonary paths. Also, bause the creaturescan learn whitheir fetimes, th from humans andfrom other creatures, it should be poible to sty thespread of uure” or the emergence of “dialts” ascreatures, med from machine to machine via eltronicma, teach each other behaviors or lanage variants. Inthis sense, Creatures users could be considered to betaking part in an internaonal Artificial-Life scienceThanks to all at MilleiuInteractive Ltd. for their( ("( ((( ( (Reynolds ) C. Reynolds "Flocks, herds and schls:A distributed behavioral model". Computer Graphics(Reynolds ) C. Reynolds "Eluon of Coidor(((