1

Social Characters for Computer Games

Bill Tomlinson

ACE (Arts Computation Engineering) program, University of California, Irvine430A Computer Science Building, Irvine, CA 92612 USA

Email: wmt@uci.edu

Many current video games feature virtual worlds inhabited by autonomous 3D animated characters.These characters often fall short in their ability to participate in social interactions with each other or withpeople. Increasing the social capabilities of game characters could increase the potential of games as aplatform for social learning. This article presents advances in the area of social autonomous characterdesign. Specifically, a computational model of social relationship formation is described. This modelformed the basis for a game entitled “AlphaWolf” that allows people to play the role of newborn pups in apack of virtual wolves, helping the pups to find their place in the social order of the pack. This articleoffers the results from a 32-subject user study that assessed the social relationship model, showing that iteffectively represents the core elements of social relationships in a way that is perceivable by people.Additionally, this article proposes a game that will allow parents, teachers and children to experiment withcomputational social behavior through social virtual characters. This research contributes to thedevelopment of games for social learning by offering a set of viable algorithms for computationalcharacters to form social relationships, and describing a project that could utilize this model to enablechildren to learn social skills by interacting with game characters.

Keywords: Computer Games, Video Games, Autonomous Characters, Social Dominance

IntroductionMost modern video games feature interactive 3Danimated characters. At the time of writing, nineof the ten top-selling PC games in the US, andeight of the ten top-selling console games in theUS, feature 3D animated virtual worldsinhabited by real-time autonomous characters(Gamasutra.com, 2004b, Gamasutra.com,2004a). (The other three games in these toptwenty also feature characters, but in 2Dworlds.) The characters in these games havemany strengths. They are delicately modeled,carefully animated and elaborately rendered, allat real-time frame rates. Despite thesecapabilities, game characters on the whole lackthe complex social competence to participate in

the relationships that form the heart of humaninteractions.

Video games show great potential asplatforms for social learning (Jenkins, 2002). Ifgames are to serve as an effective platform forsocial learning, the characters in those gamescould benefit from the ability to engage in socialrelationships with each other and with people.This article presents a robust computationalmodel that enables game characters to formsocial relationships. It offers an originalevaluation of this computational model thatsupports its validity. In addition, it proposes agame, based on the computational model,through which children, parents and teachers

International Journal of Interactive TechnologySpecial Issue on Social Learning through Gaming.

Vol. 2, No. 2, p. 101-115. 2005.

2

may explore social interactions by interactingwith social autonomous characters.

This research was developed as part of aproject called AlphaWolf, an interactive gamedeveloped at the MIT Media Lab. This gamefeatured a pack of six virtual wolves – threefully autonomous adults and three semi-autonomous pups whom people could direct byhowling, growling, whining or barking intomicrophones. The focus of the game was thesocial relationships that the wolves formed witheach other, enabled by a novel computationalmodel of emotional memories.

Figure 1 shows an example of the animatedimages that a player saw in AlphaWolf. Thegraphical buttons at the bottom of the screenshow the player which other wolves their puphas met. The posture of the wolf on each buttonshows the mental picture that the player’s wolfhas of that other wolf. For example, a wolfwhom the player’s wolf knows to be submissiveto it would be shown rolled over on its back in atypical wolf submission posture.

The evaluation of this model was conductedby means of a 32-subject human user study.Results from this evaluation support thehypothesis that the relationship model enablescomputational systems to form socialrelationships in a way that is perceptible bypeople.

A direction for future work that was inspiredby AlphaWolf is a social learning game thatallows children, parents and teachers to exploresocial relationships together by interacting with

a community of autonomous characters. Thisgame is described in more detail below.

Socially capable autonomous characters arean important building block for character-basedgames in general (Tomlinson and Blumberg,2002), and are particularly relevant to gamesthat focus on social learning. This articlepresents recent advances in computationalmodeling of social relationships, including aquantitative evaluation of a working model. Inaddition, it proposes a system for incorporatingthis model into a game for learning about socialskills. This work contributes to the basicresearch that will facilitate further explorationsof socially capable autonomous characters andthe social learning games in which they will besituated.

Related WorkThe development of computational systems withsocial competence is an active area of researchin a number of fields, including autonomousagents and multi-agents systems (Hales andEdmonds, 2003), human computer interaction(Bickmore and Picard, 2004), computer graphics(Tomlinson and Blumberg, 2002), robotics(Vaughan et al., 2000), artificial intelligence(Dautenhahn, 2000), video games (Mateas andStern, 2003) and artificial life (Breazeal et al.,2004). This section describes the work in theseand other areas that are most relevant to theresearch presented here.Autonomous CharactersThe computer/video game industry hasexpended great effort to make compelling real-time characters. Many of these games showcasecutting edge research in graphics, charactermotion, AI and other areas. Games such as EA’sThe Sims have focused on game characters,building characters with simple personalities andinteraction dynamics. Nevertheless, most gamecharacters are weak in areas such asexpressiveness and believability.

Academic researchers have looked moreclosely at the topics of expressiveness andbelievability, focusing most frequently oninteractive virtual humans (e.g., (Gratch et al.,2002)). Perlin and his colleagues have donepioneering work in creating synthetic actors(e.g., (Perlin and Goldberg, 1996)). By applying

Figure 1: Two virtual wolf pups from theAlphaWolf game.

3

procedural modifiers to characters' motion andoffering an engine for scripting the characters’behaviors, they have created virtual characterswho move and interact very naturally. Blumbergand his colleagues have contributed in severalareas including action-selection (Isla et al.,2001), learning (Blumberg et al., 2002) andinteractive cinematography (Tomlinson et al.,2000). Cassell and her colleagues built virtualhumanoids with the ability to expressthemselves like real people, in particular in thearea of embodied conversational agents (e.g.,(Bickmore, 2003, Cassell et al., 1999)). Batesand his colleagues have created computationalcharacters who appear lifelike and are able tointeract with people in real time (Bates et al.,1992, Reilly, 1996). More recently, Schaub andhis colleagues have presented research onempathic characters (Schaub et al., 2003).Prendinger and his colleagues offered anevaluation mechanism for empathic charactersbased on skin conductivity and otherphysiological signals (Prendinger et al., 2004).Eladhari and Lindley offer a design for a virtualmind for player characters (Eladhari andLindley, 2003). Lee and colleagues have lookedat ways of controlling the motions of avatars(Lee et al., 2002). Robots have also beenendowed with characteristics that allow them toengage people in expressive ways (Velasquez,1998, Breazeal, 2000). Many other researchershave also contributed to the creation ofcompelling synthetic characters, e.g., (Cavazzaet al., 2002, Funge et al., 1999, Paiva et al.,2001, Thalmann et al., 1997, Hodgins andPollard, 1997).

Despite the abundance of research in virtualcharacters, the field is still in its infancy. As onepair of researchers put it, "[t]he ethos of thetireless, constant, friendly human companion canquickly become an ethos of a broken, verynonhuman machine." (Randall and Pedersen,1998, p. 68).Social BehaviorThroughout this research, a social relationshipwill be defined as "a learned and rememberedconstruct by which an entity keeps track of itsinteraction history with another entity, andallows that history to affect its current and futureinteractions with that entity" (Tomlinson, 2002,

p. 22-23). Social relationships serve asmechanisms for context preservation (Cohen etal . , 1999), allowing an entity to behavedifferently toward different social partners,rather than interacting with all other entities inthe same way. Social relationships are closelylinked with emotion, both in their formation(Damasio, 1994) and in their expression(Darwin, 1965 (originally published 1872)).Damasio offers the Somatic Marker Hypothesis,which proposes that people (and animals) attachemotional significance to stimuli that theyencounter in their environment, and then re-experience those emotions when they encounterthose stimuli on future occasions.

Various models of emotion have beenproposed. Ekman (1992), for example, offered acategorical model with six cardinal emotions inhumans – happiness, sadness, anger, fear,surprise and disgust. Another model of emotionmaps a range of emotional phenomena onto anexplicitly dimensioned emotional space;Mehrabian and Russell (Mehrabian and Russell,1974) offer an emotion space defined bypleasure, arousal and dominance. Pankseppdraws a distinction between basic emotions,such as anger and fear, and social emotions,such as love and grief (Panksepp, 1998). Nardioffered three dimensions of connection thatcontribute to social relationships - affinity,commitment, and attention (Nardi, 2005); thesedimensions will be included in the evaluations ofhuman relationships with autonomouscharacters. Social relationships also form thebuilding blocks from which many socialnetworks may be formed (Watts, 2003, Fisherand Dourish, 2004, boyd, 2004).Computational ModelThe social relationship mechanism describedand evaluated in this paper combines Damasio’sSomatic Marker Hypothesis (Damasio, 1994)with Mehrabian and Russell’s dimensionalmodel of emotions (Mehrabian and Russell,1974) to create social relationships in a pack ofvirtual wolves. These relationships aremechanisms by which context is preservedamong different dyadic pairs of wolves (Cohenet al., 1999). The behavior of the virtual wolvesis informed by the study of real wolves (e.g.,(Mech et al., 1998, Fox, 1971, Klinghammer and

4

Goodmann, 1985)), and by previous work insynthetic social systems (e.g., (Hemelrijk,1996)). Wolves, humans and other primatespecies all exhibit social hierarchies of variousforms (Mech, 1999, Wrangham et al., 1994,US_Army, 2004), so it is possible that some ofthe relationship structures developed for wolvesmay have some application to humans as well.Immerman and Mackey (2003) point to possibleconvergent evolution of social behavior betweenhumans and canids.

The social relationship mechanism issituated in the Synthetic Characters Group’sToolkit, described in (Burke et al., 2001), and(Isla et al., 2001). In the evaluation section, themechanism is compared to other mechanisms bywhich entities may interact, in particular thosedescribed in (Vaughan et al., 2000). In addition,the AlphaWolf installation is evaluated withsimilar criteria to those used by other believablesocial and emotional agent systems – e.g.,(Reeves and Nass, 1996, Reilly, 1996, Perlin andGoldberg, 1996, Cassell et al., 1999, Breazeal,2000). Inspired by Craig Reynolds’ Boids(Reynolds, 1987), we have tried to keep themechanism as simple as possible. None of thesystems that we have encountered uses anemotional memory mechanism to enable long-term social relationships among computationalentities.

The AlphaWolf GameThe computational model of social relationshipformation described in this article was originallydesigned as part of an interactive game entitledAlphaWolf. In the game, three participants help

direct the actions of virtual wolf pups in asimulated litter. By howling, growling, whiningor barking into a microphone, each participantcan tell her pup to howl, dominate, submit orplay (see Fig. 2). The actions that the pup takesaffect emotional relationships that the pup formswith its littermates and with the adults of thepack. The pups autonomously maintain theserelationships and display them by means of theemotional style in which they take the actionssuggested by the participants. The relationshipsare also displayed to the participant throughdynamic buttons at the top and bottom of eachscreen, which show each of the social partners ofthat wolf in a dominant or submissive pose thatreflects how the participant’s pup views thatpartner.

There are six wolves in all, three pups andthree adults. The wolves take various colors –black, white or gray. Each puppy grows up frompup to adult size over approximately fiveminutes. By the end of the five-minuteinteraction, the pups, guided by the users, haveworked out their places in the social order of thepack.

AlphaWolf premiered in the EmergingTechnologies section of SIGGRAPH 2001(Tomlinson et al., 2001), and has been exhibitedin seven venues in five countries since that time.To see a short video describing the project,please view a short segment from 10/22/02edition of the Public Broadcasting Service (PBS)television show “Scientific American Frontiers”with host Alan Alda, available at:

http://pbs-saf.virage.com/cgi-bin/visearch?user=pbs-saf&template=template.html&query=Alpha+Wolf&category=0&viKeyword=Alpha+Wolf

Computational Model of SocialRelationshipsThe mechanism by which the virtual wolvesform social relationships with each otherinvolves models of emotion, perception andlearning. The relationships in turn are expressedthrough the behavior of the wolves. This sectionelaborates on each of these elements.

Figure 2: A participant interacts with thewolves using a microphone interface.

5

EmotionThe model of emotion in AlphaWolf is verysimple – a single floating-point value calleddominance , which varies from 0.0 to 1.0.(Dominance is a subset of a more complexdimensional representation of emotion presentedby Mehrabian and Russell (Mehrabian andRussell, 1974) including pleasure and arousal astwo other orthogonal axes). Each wolf’sdominance value is affected by its interactions.For example, being bitten causes a wolf’sdominance to drop, and being the target ofanother wolf’s submission causes a wolf’sdominance to increase.

The wolf’s emotional state affects the stylein which the wolf takes its actions. For example,a wolf with a current dominance value near 0.0might walk with its tail between its legs and itsears back, while a wolf with a higher dominancevalue will hold its tail and ears erect. Thesystem blends between example animations togive an expressive range to the behavior of thewolves (Downie, 2001).

In addition, when a wolf is behavingautonomously, its emotional state affects whatactions it chooses to take towards its socialpartners. For example, a submissive wolf mightchoose to roll over on its back, while a dominantwolf might growl.PerceptionThe second major section of the relationshipsystem is the way in which the wolves perceiveeach other. The wolves are able to tell theirpack mates apart in order to form differentrelationships with each of them. They alsoknow when they are in an interaction with eachother. (For example, being bitten is a muchdifferent experience than seeing someone elseget bitten.)

The wolves disambiguate among their socialpartners by means of a unique identification tag.Being able to identify social partners is one ofthe three principles that make up the “bedrock ofcooperation” (the other two being thatindividuals must be likely to encounter eachother more than once, and that individuals musthave information about how the individual hasbehaved in the past – see below) (Kollock, 1998,Axelrod, 1984). While it may seem simplisticfor entities to “know” each other by means of a

unique ID, real wolves appear to be able todistinguish each other by scent (Beaver, 1999).

From the point of view of an individual wolfA, an interaction with a partner B begins whenA has B as its Object of Attention, and perceivesthat B is reciprocally attending to him. For A,the interaction ends when it changes its Objectof Attention so that it is no longer attending toB, regardless of whether B is still attending to A.In fact, A would have no way of telling whetherB was still attending to it, since that informationshould be unavailable to it unless B is its Objectof Attention. Giving an individual only oneObject of Attention does preclude the socialrelationship mechanism from making correctassociations in certain circumstances (e.g.,attending to one wolf in front, while being bittenby another wolf from behind). However, thissimplified form should be able to be extended toinvolve more complex perceptual mechanisms.In addition, if a wolf bit me from behind everytime I attended to a certain social partner, Iwould not be surprised if that affected myrelationship with that partner, rather than justwith the wolf doing the biting.

To summarize, a character is able to identifyindividuals, to have an Object of Attention, andto assess when it is the Object of Attention ofanother individual.LearningThe first time individual A ends an interactionwith individual B , it forms an “emotionalmemory” of B . The model of an emotionalmemory contains three pieces of information –the unique ID of B, an emotional value, and aconfidence value.1 When the emotional memoryis first formed, the emotional value stores theemotion that A was feeling at the time itsinteraction with B ended. Since the interactionsthat A has had with B may have altered itsemotional state over the course of theirinteraction, forming the emotional memory atthe end of the interaction rather than at thebeginning will reflect the emotional content ofthe relationship more accurately.

The next time A switches to have B as itsObject of Attention, its emotional memory of Bwill influence its current emotional state inproportion to its confidence in that model. The

6

emotional memory is applied to the currentemotional state by:

E′ = (C × Em) + ((1 – C) × E)

Equation 1

where E′ is the wolf’s new emotional value, C isthe confidence value of the emotional memorybeing applied, E is the wolf’s emotional valueprior to the application of the emotionalmemory, and Em is the emotional value that isstored in the emotional memory.

At the end of each successive interaction, Arevises its emotional memory of B . Uponrevision, two of the three elements of anemotional memory are changed – confidenceand emotional value. Confidence is revisedbefore the emotional value so that the change inthe emotional value will reflect the change inconfidence, thereby preserving the effect ofdeviations from the expected emotionalinteraction.

The confidence value is revised by:C′ = C + ((Min(C, 1-C)) × (Tc – |Em – E|) × L)

Equation 2

where C′ is the new confidence value for theemotional memory, C is the previousconfidence, Tc is some confidence thresholdbetween 0 and 1, Em is the emotional value thatis stored in the emotional memory, E is thewolf’s current emotional state, and L is alearning rate. (Multiplying the learnedcomponent by the Min of C and 1-C effectivelyclamps the confidence value to between 0 and 1,and helps to polarize relationships.)

The emotional value stored in the emotionalmemory is then revised by:

Em′ = (C × Em) + ((1 – C) × E)

Equation 3

where Em′ is the new emotional value stored inthe wolf’s emotional memory, C is confidence,

Em is the emotional value in the memory prior tothe revision, and E is the wolf’s currentemotional state.

These equations represent a simple buteffective implementation of the emotionalmemory mechanism. For more elaborate socialbehavior, any or all of these equations might bemade more complex.

The emotional memories described herefunction as remembered constructs by which anindividual keeps track of its interaction historywith another individual. They allow that historyto affect its current and future interactions withthat individual. These two elements satisfy thedefinition of a social relationship offered earlier.ExpressionEach wolf’s emotional state affects what it doesand how it does it. The most sublimely complexrepresentation of a social relationship isn’t ofmuch use without an equally expressive range ofbehavior. The AlphaWolf system uses a motorblending system described in (Downie, 2001) toenable the characters have dynamic expressiveranges (see Figure 3). Downie’s expressivemotor system allows a wolf’s current emotionalstate to affect the style in which the wolfbehaves (e.g., run dominantly or submissively).Since emotional memories affect the wolf’scurrent emotional state, they also affect the styleof the wolf’s behavior.

EvaluationTo evaluate the effectiveness of the socialrelationship mechanism in AlphaWolf, a set ofuser studies was conducted. Several hypotheseswere tested through these experiments, of whichthe one presented here was most central. For adiscussion of the other hypotheses and theirresults, please see (Tomlinson, 2002). Thissection presents the core hypothesis,experimental method, results and discussion in

Figure 3: An emotional range from submissive to dominant. The extreme poses were hand-crafted byan animator. The center three poses are automatically generated.

7

this study.HypothesisThe central hypothesis of this study was asfollows: the mechanism in AlphaWolf encodessocial relationships created by a first person in away that can be perceived by a second, naïveperson. This hypothesis has two sub-hypotheses: a) the mechanism successfullyencodes social relationships, and b) naïve userscan effectively perceive the relationships amongthe characters.Experimental MethodThis hypothesis was evaluated by means of anexperiment involving 32 human subjects. Thesubjects ranged in age from 17 to 55 (mean =26.2, standard deviation = 7.8). Half (16) werefemale (min age = 18, max age = 55, averageage = 26.1, standard deviation = 10.1), and halfwere male (min age = 17, max age = 37, averageage = 26.4, standard deviation = 4.7).

In the user tests, each of the 32 subjectsinteracted with the AlphaWolf game. Eachsubject participated in three 4-minuteinteractions with different versions of the game.Two of each subject’s runs were interactive, anda third run was non-interactive.

In the interactive runs, all subjects directedthe gray pup, and were instructed to formspecific social relationships with its siblings, thewhite and black pups. Various runs featured theAlphaWolf mechanism (described earlier in thispaper) or one of three alternate relationshipmechanisms (Random, Emotional or Fixed),which lacked the persistent, expressive qualitiesof the AlphaWolf mechanism. At the end of

each interactive run, the relationships of thepups were written to a file.

The non-interactive runs for each subjectfeatured pups with relationships that wereloaded in from the file created by a previoussubject’s interaction. Subjects simply sat andwatched the wolves interact, and observed therelationships that they exhibited.

The two interactive and one non-interactiveruns occurred in random order.

The evaluation of the hypotheses abovedraws upon Shannon’s Theory ofCommunication (Shannon, 1948) (see Figure 4).Shannon’s theory subdivides the process ofcommunication into five parts – an informationsource, a transmitter, a channel, a receiver and adestination. Each of these parts has acorresponding element in the AlphaWolf usertests. Each act of communication involved twosubjects – Subject A and Subject B. Subject A’ssession with the wolves preceded that of SubjectB; one of Subject A’s runs created therelationships that were loaded into Subject B’sNon-interactive run.

The simple model presented by Shannonoffered a clear way to represent an act of socialcommunication. Subject A is assigned the taskof forming two social relationships. He or sheencodes these relationships into a computationalrepresentation through an interactive run of theAlphaWolf system. These relationships areloaded in at the beginning of Subject B’s Non-interactive run. If Subject B can perceive therelationships, they have been successfullydecoded. If both encoding and decoding work,transmission occurs. This model allowed the

AssignedRelationships

–Subject A

RecordedRelationships

–Digital file

PerceivedRelationships

–Subject B

Encoding Decoding

Transmission

Figure 4: Shannon’s model of communication, adapted to represent the transmission of socialinformation between two subjects in the AlphaWolf study.

8

experiment to determine which elements of thesystem were capturing or transmitting the socialinformation effectively.Information SourceThe first stage in the communication of socialrelationships is specifying the relationship to betransmitted. These relationships were specifiedin advance – the experimenter informed SubjectA of the relationships that he or she shouldattempt to have the gray pup form with the whitepup and the black pup. All subjects wereinstructed to be dominant to one pup andsubmissive to the other. Therefore, theexperimenter is the information source, and theassigned relationships are the message thatmight or might not be transmitted to thedestination.TransmitterSubject A’s interaction with the virtual wolfpack is the transmitter, taking the assignedrelationships and converting them into theA l p h a W o l f d o m i n a n c e / c o n f i d e n c erepresentation. This conversion encodes therelationships in a way that can be transmittedover the channel.ChannelWhereas Shannon’s theory deals primarily witha noisy channel (a telephone cable, for example),the channel in the experiment was a digital filethat was loaded in without any loss of precision(barring experimenter error). This file stored therelationships formed by Subject A’s interactionuntil they were loaded into Subject B’s Non-interactive run.ReceiverSubject B’s interaction serves as the receiver,decoding the social relationship representationinto a human-readable form. (While the currentrepresentation is simple enough that it, too, issomewhat human-readable, a more elaborateversion with three emotional axes, for example,would be completely opaque.) By viewing thesocial interactions among the virtual wolves in aNon-interactive run, Subject B attempts torecognize the relationships among the pups.DestinationThe final destination of the message is thequestionnaire that Subject B fills out. If SubjectB reliably perceives the same relationships that

were assigned to Subject A, and the only linkbetween the two subjects is the AlphaWolfrepresentation, then it is very likely that therepresentation successfully transmitted therelationships.

The dissection of the process ofcommunication in the above sections and inFigure 4 demonstrates that there might bedifficulty at a variety of points in a single act ofcommunication. In the experiment, Subject Amight not understand the instructions. He or shemight fail to encode the relationships because ofa failure of the interface. There could beproblems writing the file at the end of SubjectA’s run, or loading it at the beginning of SubjectB’s. Subject B could fail to recognize therelationships. Subject B could get confused andfill out the form incorrectly. The experimentalmethod was constructed in order to minimize thechance of each of the peripheral problems and tofocus on the two central issues at hand – whetherSubject A can encode the social relationships,and whether Subject B can decode them. Onlyif both of these conditions are true can theAlphaWolf system be said to transmitinformation that people would attribute to asocial relationship.

Strong graphics and animation are keycomponents of this system. Muddy graphics orpoor animation could certainly render the systemnon-functional, preventing people fromrecognizing the relationships of the wolves.However, the results do not rely solely on thegraphics. Without the underlying relationshiprepresentation, Subject B might accuratelyperceive certain relationships, but theserelationships could have no correlation withthose assigned to Subject A. Graphics andanimation are crucial to the transparency of thesystem, but do not in themselves account for therelationship mechanism’s functioning.ResultsThis section describes the results of theencoding, decoding and transmission elementsof the above study.EncodingEach of the 32 subjects viewed one Non-interactive run during their session. This Non-interactive run was randomly assigned to be thefirst second or third run that the subject

9

encountered. Thirty-one of the 32 subjectsviewed Non-interactive runs based on othersubjects’ interactive runs. (The first subject hada Non-interactive run based on an interactive runfrom the pilot study, which had a slightlydifferent setup. That subject will therefore notbe included in certain portions of this analysis.)

Half of these runs were encoded by previoussubjects interacting with runs that featured theAlphaWolf mechanism, while the other halfwere encoded by the Random algorithm or theFixed algorithm. Of those encoded with theAlphaWolf mechanism, half again were donethrough the Emotional algorithm in whichemotional memories are still formed, eventhough they are not allowed to influence awolf’s current emotional state. Each runincluded two assigned relationships, onebetween the gray pup and the white pup, and onebetween the gray pup and the black pup.

To determine whether or not encoding hadoccurred, the assigned relationship wascompared to the recorded representation of therelationship. The relationship between twowolves (A and B) is determined by an average ofthe dominance values, weighted by theconfidence values:

((DAB x CAB) + ((1-DBA) x CBA))(CAB + CBA)

Equation 4

where DAB is A’s dominance value with respectto B , CAB is A’s confidence in that dominance,DBA is B’s dominance value with respect to A,and CBA is A’s confidence in that dominance.

This formula yielded a single value from 0.0to 1.0 that captured the essence of therelationship between the two wolves. If thisvalue was less than 0.5, the relationship was“submissive” from the point of view of the graypup. If it was greater than 0.5, the relationshipwas “dominant.” Because of the relativecomplexity of the system, it was exceedinglyunlikely that a relationship would ever equalexactly 0.5. However, if two pups never met (assometimes occurred between the black and thewhite pups, who might accidentally pass the fourminute run without ever meeting), the valuewould end up at exactly 0.5 (a weighted averageof the two starting values, 0.5 and 0.5). Thiscondition did not happen in any of the assigned

relationships involving the gray pup, though,and is therefore not relevant to these results.

Of the 32 total relationships encoded withthe AlphaWolf mechanism (16 runs x 2relationships per run), the internal representationmatched the assigned relationship in 30 cases,and did not match in 2 cases (93.8% success, p<0.00012, see Figure 5). These figuresdemonstrate that the subjects and the systemwere quite successful at encoding socialrelationships.

Even in subjects' first run of virtual wolves,15 out of 16 were encoded successfully (93.8%success, p = 0.0021). These figures demonstratethat there was not a significant learning curve inthe experiment; people were just as successful atthe beginning of the session as they were at theend.

In the runs without the AlphaWolfmechanism, in which relationships were eitherpredetermined or random, 20 of the 30 runs wereencoded correctly (66.7% success, p = 0.1261).While pure randomness would give 50%encoding, the p value demonstrates that it isreasonable that these data arose by chance. Alarger sample size would be necessary todetermine for certain if the elevated encodingresulted from some code error or problem in theexperimental method, or simply from five wrongrolls of the dice.

0102030405060708090

100

Non-AlphaWolf AlphaWolf

Per

cent

age

Suc

cess

Figure 5: Encoding success with Non-AlphaWolfand AlphaWolf social relationship mechanisms.The AlphaWolf mechanism offers significantly

improved encoding over random chance, or overalternate mechanisms.

One 29-year-old male subject wrote acomment that suggests a possible reason the

10

encoding success rate wasn’t even higher. Hereported a “[s]trong urge to test relationships –growl at black, for instance.” The fact thatpeople like to “poke at” a system is an importantconsideration in the design of interactivesystems. Because the AlphaWolf system allowsthe computational entities to switch theirrelationships in response to adequate user input(e.g., it took about three clear interactions, inwhich both pups took on opposite roles, toreverse a relationship) the urge for people tovary their behavior could have had a significantimpact on the results. Nevertheless, it appearsthat encoding was largely successful despite thepossibility of intentional disobedience.DecodingThere were 31 subjects whose Non-interactiveruns featured relationships created duringanother subject’s run. Each of these Non-interactive runs included two relationshipsinvolving the gray pup (gray-black and gray-white). In the questionnaire at the end of thatrun, subjects were asked to specify which ofeach pair of pups was more dominant.

0102030405060708090

100

Non-AlphaWolf AlphaWolf

Per

cent

age

Suc

cess

Figure 6: Decoding success with Non-AlphaWolfand AlphaWolf social relationship mechanisms.Decoding is successful regardless of relationship

mechanism.

The way in which a match in decoding wasdetermined was to simplify both the recordedrepresentation and the subjects’ answers to thequestionnaire. Using Equation 4 above, therepresentation was reduced from four numbersto one, which in turn was converted to a simple“dominant” or “submissive” by the methoddescribed above. So, for example, if Wolf A had

a dominance of 0.7 and a confidence of 0.5 withrespect to Wolf B, and Wolf B had a dominanceof 0.1 and a confidence of 0.8 with respect towolf A, the equation would reduce to:

((0.7 x 0.5) + ((1-0.1) x 0.8))/(0.5 + 0.8) =(0.35 + 0.72)/1.3 =0.82

Since 0.82 is greater than 0.5, Wolf A would beconsidered to be dominant with respect to WolfB.

Subjects’ responses to the questionnaireswere also simplified, with an answer of 1, 2 or 3counting as “submissive”, an answer of 5, 6 or 7counting as “dominant”, and an answer of 4counting as “couldn’t tell.”

Of the 62 relationships (31 runs x 2relationships per run), subjects were able todetermine which pup was dominant in 53 cases,couldn’t tell in 2 cases3, and chose incorrectly in7 cases (88.3% success, p < 0.0001, see Figure6). These figures, well above chance levels,show that people were able to watch a run ofvirtual wolves and comprehend the relationshipsthat they had seen there. Subjects weresuccessful at decoding social relationships,according to the definition offered earlier: “alearned and remembered construct by which anentity keeps track of its interaction history withanother entity, and allows that history to affectits current and future interactions with thatentity" (Tomlinson, 2002, p. 22-23)).

Even in their first run when they could nothave learned about which animations correlatewith dominance and submission from previousinteractive runs, subjects were successful at“reading” the relationships among the pups; of22 first run relationships, subjects decoded 18correctly, couldn’t tell for 1, and choseincorrectly in 3 (85.7% success, p = 0.0048).These figures demonstrate that fully naïvesubjects were able to determine relationshipsfrom only their pre-existing knowledge and theirviewing of real wolves; they did not just learnwhat this system “meant” by social relationshipsover the course of their interactive runs.

There was no significant difference betweendecoding success when the relationships hadbeen encoded by the AlphaWolf mechanismversus when they had been encoded by a randomor fixed mechanism. Of the 32 decoded fromAlphaWolf-created relationships, 26 were

11

correct, 2 couldn’t tell, and 4 were incorrect(86.7% success, p < 0.0001). Of the 30 decodedfrom non-AlphaWolf origins, 27 were correctand 3 were incorrect (90.0% success, p <0.0001). This result suggests that having theAlphaWolf encoding mechanism in the firststage of the transmission did not significantlydisrupt the decoding process in the second stage.The Relationships between Black and WhiteAlthough the black and white pups were formingrelationships with each other as well as with thegray pup, these pups were not the camerasystem’s “lead actor,” and therefore it was notvery common for the camera to capture a cleardominance interaction between the two. As one30-year-old male subject put it: “It was hard todetermine the relationship between the two otherpups (in this case between white & black) as Irarely saw them interacting with each other(only at the end).” Added to which, only thetwo relationships involving the gray wolf wereassigned to the encoding subjects. Therefore,the relationships between black and white werenot included in the above results for decoding.

Nevertheless, it appears that subjects did justas well decoding the relationships between thesetwo pups as they did in the relationshipsinvolving the gray pup. Of the 32 black/whiterelationships, subjects successfully decoded 26,subjects couldn’t tell in 2 cases, got it wrong in2, and in 2 cases the black and white pups nevermet4 (and therefore had no relationship) (92.9%success, p < 0.0001). Even in their first run,subjects were successful; of the 11 subjectswhose Non-interactive run was in the firstposition, there were 10 correct decodings, 1 runwhere the pups did not form relationships, andno mistakes. (100% success, p = 0.0079) Itappears that it doesn’t take much screen-time forpeople to recognize a relationship.TransmissionConsidering the entire process (both encodingand decoding) as a single act of communication,there were 16 subjects whose Non-interactiveruns were based on another subject’s interactiverun that featured the AlphaWolf algorithm.Each of these subjects’ Non-interactive runs hadtwo relationships, for a total of 32 relationships.Of these, 26 were successfully transmitted(assigned relationship matched recorded

representation and answer given onquestionnaire), 1 was a double error resulting insuccessful communication of the message(assigned relationship matched answer given onquestionnaire, but recorded relationship wasopposite5), 2 couldn’t tell, and 3 were incorrect(89.7% success, p < 0.0001, see Figure 7).These figures verify that the AlphaWolfalgorithm and the entire AlphaWolf systemsuccessfully captures, transmits and displayssocial relationships in a way that is readilyapparent to humans. The fact that subjects wereable to decode the wolves’ behaviorssuccessfully regardless of the encodingmechanism suggests that the animation systemwas working in a way that was apparent topeople. The most important feature of theAlphaWolf relationship algorithm is that itenabled people to encode the wolves’relationships in a format that could later beperceived by a separate subject.

0102030405060708090

100

Non-AlphaWolf AlphaWolf

Per

cent

age

Suc

cess

Figure 7: Transmission of relationships with Non-AlphaWolf and AlphaWolf social relationships.

Transmission is significantly better with theAlphaWolf mechanism than with the alternate

mechanisms or by random chance.

Transmission was much less successful incases where the relationships had not beenencoded by the AlphaWolf mechanism. Of the30 cases where encoding occurred by one of theother three mechanisms, transmission occurredin 19 cases, there were 2 double errors, and 9incorrect responses (67.9% success, p = 0.0574).These figures are a bit high (in line with theslightly elevated encoding described above), butwell below the level attained by the AlphaWolfmechanism. The two double errors are appear tobe partially responsible for the artificially high

12

number. The disparity between the transmissionin the AlphaWolf and Non-AlphaWolf casessuggests that the success of the AlphaWolf caseswas due to the AlphaWolf algorithm (rather thanany other cause), since all other factors wereheld constant.Summary and DiscussionTo summarize the above study: 32 humansubjects interacted with packs of virtual wolvesunder controlled conditions. Each person playedthe role of a gray wolf pup and attempting todirect that pup to form certain dominancerelationships with its two siblings. At the end ofthe run (approx. 4 minutes long), the systemrecorded all of the pups' internal representationof the relationships. A second experimentalsubject later viewed a Non-interactive pack ofautonomous wolves whose relationships werespecified by the recorded relationships from theprevious subject's interactive run. If the secondsubject's perception of the relationships matchedthe relationships that had been assigned to thefirst subject, then the AlphaWolf systemsucceeded in encoding, transmitting anddecoding those relationships. This method ofanalysis is derived from Shannon's Theory ofCommunication (Shannon, 1948).

0

10

20

30

40

50

60

70

80

90

100

Encoding Decoding Transmission

Per

cent

age

Suc

cess

Non-AlphaWolfAlphaWolf

Figure 8: The AlphaWolf mechanism offerssignificantly better encoding of relationships than

chance (50%) or than a system without theAlphaWolf mechanism. Since there were no

statistically significant differences in decoding, thegreater success at encoding translated directly to

significantly improved transmission.

The results from that study demonstrate thatthe AlphaWolf mechanism successfullyrepresents social relationships (see Figure 8).Using the AlphaWolf mechanism, the firstsubjects successfully encoded 93.8% of the

relationships that they had been assigned tocreate (p < 0.0001). The second subjectsdecoded 86.7% of the relationships that theyviewed (p < 0.0001). These two factors resultedin a successful transmission rate of 89.7% (p <0.0001). Each of these figures demonstratessignificantly better performance than the 50%success rate predicted by chance.

Using the non-AlphaWolf mechanism,66.7% of the encoded relationships matchedthose that had been assigned to the first subject(p < 0. 1261). The second subjects decoded90.0% of the relationships that they viewed (p <0.0001). These two factors resulted in asuccessful transmission rate of 67.9% (p < 0.0574). Both encoding and transmission are notstatistically different from chance (although, asdiscussed above, they are somewhat higher thanthe predicted value of 50%).

The significantly lower levels of encodingand transmission among non-AlphaWolf runsconfirm that these effects are a direct result ofthe AlphaWolf algorithm. The human userevaluation of the AlphaWolf system supports thehypothesis that the system and its relationshipalgorithm successfully encode, decode andtransmit a valid representation of socialrelationships.

Nevertheless, this system is still greatlysimplified from wolf social relationships, letalone from human relationships. It could beargued that the system doesn’t scale well –89.7% accuracy on a single-axis model ofemotion drops down to 72.2% accuracy on athree-axis model. However, having an exactprediction of emotion is not absolutely necessaryin a model such as this, and in fact might notreflect the realities of biological relationships.Confusion over emotions is standard in humanrelationships, and a system that is very good, butnot perfect, might more accurately model thenatural phenomenon. Also, the partial successesof the model (where, for example, it wouldpredict two of the three axes effectively) wouldstill provide reasonable approximations ofemotionally responsive behavior. Therefore,this model should scale reasonably well, and inbiologically plausible ways. Despite thesimplification of the system to represent onlyone emotional axis in this case, it should provide

13

a solid grounding for more complex relationshipsystems for modeling both animals and humans.

It could also be argued that creatures thatalways exhibit the emotions that they are feelingdo not capture the complexities of human socialand emotional interactions. This is a fair point,and this system is limited in that it does notallow entities to mask their emotional responsesas humans sometimes do. However, in order tomask one’s emotions, it is necessary for one tohave emotions first. This system does notcapture the full complexity of human (or evenwolf) relationships, but it does provide a centralbuilding block that could be used to developmore believable emotions and relationships.

Limitations of the StudyThere are a number of limitations to the designof this study. First, it is possible that the “socialrelationships” transmitted by the system are notreal social relationships, but instead just a wayof transmitting patterns of behaviors thathumans associate with simple social relationshipphenomena. This difference could be relevanton two levels. First, there is a practicaldifference, where perhaps a more complex oreffective computational mechanism could do abetter job of capturing social relationships thanthe AlphaWolf mechanism. If this is the case,then hopefully AlphaWolf can provide a startingpoint or inspiration for the design of thissuperior mechanism. A deeper issue is whethera computational simulation can have a socialrelationship, or whether social relationships areby definition the domain of human beings.Similar questions have been discussed in variousother aspects of computational systems (e.g.,intelligence, emotion). The author prefers afunctional definition of social relationships thatwould allow computational systems to formthem, but respects that other people woulddisagree with this position.

A second limitation is that the AlphaWolfmechanism captures only simple dyadicrelationships, and does not explicitly representrelationships among three or more individuals.Since people are able to comprehend triadicrelationships, subjects may have used this skillto intuit more information about the wolves’relationships than the experimenters expected.For example, if a decoding subject had a strong

view of the gray pup’s relationship with theblack pup and a strong view of gray’srelationship with white, and if the encodingsubjects were always assigned to play gray, todominate one pup, and to submit to the other,then, by the transitive property of socialrelationships, subjects might be able to predictthe relationship between black and white.Nevertheless, packs encoded with the randomalgorithm would arguably detract from thisphenomenon, and pups encoded during a Fixedrun would have strong relationships, but wouldnot be guaranteed to have gray in the middlerank. Despite these facts, it is relevant thatcertain “whole-world” effects may havecontributed to people achieving some indirectunderstanding of relationships to which theywere not well-exposed directly.

A third limitation of this study involves therelatively small sample size used in it. While 32subjects were adequate to find statisticalsignificance in several important areas, therewere additional findings that might have beenmore effectively understood with moreadditional data, such as the effectiveness ofvarious interaction paradigms (e.g., microphonevs. keyboard) at enabling subjects to encoderelationships.

Proposed Interactive Game forExamining Social BehaviorDuring the development of AlphaWolf, itbecame apparent that subtle changes to thesocial relationship algorithms could havesweeping repercussions on the personalities ofthe characters. Different relationship algorithmsmight lead to “Might makes right” characters,“Do unto others as you would have them do untoyou” characters, “An eye for an eye” characters,or pacifist characters. In addition, changing theperceptual mechanism so that the charactersrecognized each other by color rather than byunique ID led to characters that exhibited simplesocial stereotyping, or even racist, behavior.

The design of systems like AlphaWolf is asubtle and potentially problematic domain, sincethey deal with human-like behaviors andultimately will have an effect on the humanswho interact with them. It is very easy forimplicit biases to find their way into the design

14

of the system. By thinking through the potentialsocial ramifications of the system, the designersof AlphaWolf attempted to be aware of theideologies that were embedded in the system’scode. Raybourn addressed similar issues indescribing the design of intercultural agents(Raybourn, 2003).

The realization of the profound impacts ofsmall modifications to the relationship code hasled to a potentially interesting direction forfuture work – creating a game that allows peopleto explore the ramifications of these kinds ofsocial interactions. This game will be designedso that children, teachers and parents are able tolearn about social behavior together. This systemwill be based on the core research describedabove. The system will feature a virtualenvironment with a community of interactiveanimated characters in it. People will be able tocontrol several relevant algorithms in thecharacters' social relationship mechanisms. Thiscontrol will allow them to explore and discussthe ramifications of various kinds of socialbehavior. While the social phenomena will bevastly simplified compared to the complexitiesof human social behavior, they will neverthelessprovide scaffolding through which parents,teachers and children may talk about these issuesand hopefully come to understand them better.In addition, by introducing students to theinteractions between computer programmingand interactive animation, the system couldencourage students to consider pursuingprogramming and computer science in generalas careers.

Prior to the deployment of the system,several schools and families will be sought outto serve as a user group for evaluating theeffectiveness of the system. The system will beevaluated in two ways. First, the parents andteachers in the user group will be asked tocomplete questionnaires before and after usingthe system to assess whether they found iteffective. Second, the children involved will beasked to interact with a version of the systembefore and after interacting with it with theirparents/teachers, and will fill out questionnairesregarding why they made certain decisions insocial situations in the game. Their responseswill be analyzed to see if they gave moreconsidered responses in the post-case. After the

initial evaluation and revision, the system willbe released on the web so that families andschools are able to download and utilize it.

ConclusionEntertainment-oriented video games are largelybased in virtual worlds inhabited by avatars andautonomous characters. In order for learninggames to compete with these games, they toomay need to have compelling animatedcharacters. Giving these characters socialcompetence will make it much more likely thatthey could provide opportunities for sociallearning to the people who interact with them.This article has offered a simple computationalmodel for enabling autonomous characters tohave social capabilities. In addition, it providedquantitative support for the effectiveness of themodel, by means of a 32-subject user study.Finally, it described a potential game in whichthese social characters could be situated toprovide an opportunity for people to explore arange of social phenomena through interactionswith virtual characters.

This research clearly only scratches thesurface of social algorithms for autonomouscharacters. Nevertheless, it points toward afruitful area of research into the creation ofcomputational characters as building blocks foreducational games. Social characters couldserve as virtual teachers, synthetic students andlearning companions, and could help providerich virtual worlds in which people may learnand grow.

AcknowledgementsThe author would like to thank severalanonymous reviewers who contributedsignificantly to its improvement of this paper.This research would not have been possiblewithout the collaboration and support of theother members of Professor Bruce Blumberg’sSynthetic Characters Group at the MIT MediaLab, in particular Marc Downie, Matt Berlin,Jesse Gray, Adolph Wong, Derek Lyons, JennieCochran, Bryan Yong and Michael PatrickJohnson. The research also benefited greatlyfrom discussions with Professors BruceBlumberg, Cynthia Breazeal, Rosalind Picardand Richard Wrangham.

15

Notes1 Although the social relationship mechanism

that we describe treats i nd i v idua l s asemotional significant stimuli, a stimulus doesnot have to be an individual – only a causativeentity (Damasio, 1994). Forming emotionalmemories of other kinds of stimuli (e.g., thepresence of two wolves at the same time)could result in other kinds of relationships(e.g., alliance formation).

2 Throughout this analysis, the Mann-Whitney Utest was used to calculate the p values.

3 For the purposes of calculating averages and pvalues, runs in which subjects couldn’t tell therelationship were discarded.

4 For the purposes of calculating averages and pvalues, runs in which pups never formed arelationship were discarded.

5 Throughout this analysis, double errors werethrown out in determining percentages and pvalues. Nevertheless, it is an interestingpossibility that “whole-world” effectssomehow contribute to the prevalence ofdouble errors – that some aspect of the inter-relation among the three pups makes peopleable to divine the correct relationship despite aflawed representation. In this particulardouble error, the recorded relationshipssuggested that both pups had decided that theywere submissive to the other. This unusualcase most likely created behavior that wasdifficult to understand for the perceivingsubject. The subject gave a response of 5,suggesting a slight leaning towards gray beingdominant. In fact, gray was slightly moredominant, feeling 0.12 dominance towardswhite, while white felt 0.08 dominant towardsgray. However, gray had a higher confidencein its submissive state (0.87 vs. white’s 0.65),so the weighted average showed a slight leantowards gray being more submissive (0.46).

ReferencesAxelrod, R. (1984) The Evolution of

Cooperation, Basic Books, New York,USA.

Bates, J., Loyall, A. and Reilly, W. (1992) AnArchitecture for Action, Emotion, and

Social Behavior, Lecture Notes in ComputerScience Vol. 830: Selected Papers from theFourth European Workshop on ModelingAutonomous Agents in a Multi-Agent World,pp. 55-68.

Beaver, B. V. (1999) Canine Behavior: A Guidefor Veterinarians, W. B. SaundersCompany, Philadelphia, PA, USA.

Bickmore, T. (2003) Relational Agents:Effecting Change through Human-ComputerRelationships, Ph.D. dissertation, MediaArts & Sciences, Massachusetts Institute ofTechnology, Cambridge, MA, USA.

Bickmore, T. W. and Picard, R. W. (2004)Towards caring machines, In Extendedabstracts of the 2004 conference on Humanfactors and computing systems, ACM Press,pp. 1489-1492.

Blumberg, B., Downie, M., Ivanov, Y., Berlin,M., Johnson, M. P. and Tomlinson, B.(2002) Integrated learning for interactivesynthetic characters, Proceedings of the 29thannual conference on Computer graphicsand interactive techniques, San Antonio,Texas, USA, July 23-26, 2002, pp. 417-426.

boyd, d. m. (2004) Friendster and publiclyarticulated social networking, In Extendedabstracts of the 2004 conference on Humanfactors and computing systems, ACM Press,pp. 1279-1282.

Breazeal, C. (2000) Sociable Machines:Expressive Social Exchange Between Robotand Human, Sc.D. Dissertation, ArtificialIntelligence Laboratory, MIT, Cambridge,MA, USA.

Breazeal, C., Buchsbaum, D., Gray, J., Gatenby,D. and Blumberg, B. (2004) Learning fromand about Others: Towards Using Imitationto Bootstrap the Social Understanding ofOthers by Robots, In: L. Rocha and F.Almedia e Costa (eds.) Artificial Life (toappear).

Burke, R., Isla, D., Downie, M., Ivanov, Y. andBlumberg, B. (2001) CreatureSmarts: TheArt and Architecture of a Virtual Brain,Proceedings of the Game DevelopersConference, San Jose, CA, USA, March 20-24, pp. 147-166.

16

Cassell, J., Bickmore, T., Billinghurst, M.,Campbell, L., Chang, K., Vilhjálmsson, H.and Yan, H. (1999) Embodiment inConversational Interfaces: Rea, Proceedingsof the CHI'99 Conference on Human Factorsin Computing Systems, Pittsburgh, PA,USA, May 15-20, pp. 520-527.

Cavazza, M., Charles, F. and Mead, S. J. (2002)Interacting with Virtual Characters ininteractive storytelling, Proceedings of thefirst international joint conference onautonomous agents and multiagent systems,Bologna, Italy, July 15-19, pp. 318-325.

Cohen, M. D., Riolo, R. L. and Axelrod, R.(1999) The Emergence of SocialOrganization in the Prisoner's Dilemma:How Context-Preservation and OtherFactors Promote Cooperation, Santa FeInstitute Working Paper 99-01-002(technical report). Available at:http://www.santafe.edu/research/publications/wpabstract/199901002 (January, 2005).

Damasio, A. (1994) Descartes' Error: Emotion,Reason, and the Human Brain, G. P.Putnam's Sons, New York, NY, USA.

Darwin, C. (1965 (originally published 1872))The Expression of the Emotions in Man andAnimals, The University of Chicago Press,Chicago, IL, USA.

Dautenhahn, K. (2000) Socially IntelligentAgents and The Primate Social Brain -Towards a Science of Social Minds., 2000AAAI Fall Symposium: Socially IntelligentAgents: The Human in the Loop, NorthFalmouth, MA, USA, November 3-5, pp.35-51.

Downie, M. (2001) behavior, animation, music:the music and movement of syntheticcharacters, S.M. Thesis, Media Arts &Sciences, MIT, Cambridge, MA, USA.

Ekman, P. (1992) An Argument for BasicEmotions, In Basic Emotions, (Ed, Stein, N.,and Oatley, K.), Lawrence Erlbaum, Hove,UK, pp. 169-200.

Eladhari, M. and Lindley, C. (2003) PlayerCharacter Design Facilitating EmotionalDepth in MMORPGs, Digital GamesResearch Conference 2003, University of

Utrecht, The Netherlands, 4-6 November2003, pp.

Fisher, D. and Dourish, P. (2004) Social andtemporal structures in everydaycollaboration, Proceedings of the CHI2004conference on human factors in computingsystems, Vienna, Austria, April 24-29, pp.551-558.

Fox, M. W. (1971) Behaviour of Wolves, Dogsand Related Canids, Harper & Row, NewYork, NY, USA.

Funge, J., Tu, X. and Terzopoulos, D. (1999)Cognitive modeling: knowledge, reasoningand planning for intelligent characters,Proceedings of the 26th annual conferenceon Computer graphics and interactivetechniques (SIGGRAPH 99), Los Angeles,CA, USA, August 8-13, pp. 29-38.

Gamasutra.com (2004a) Latest US ConsoleGames Sales Charts, 7/26/04, Available at:http://www.gamasutra.com/php-bin/news_index.php?story=4070 (January,2005).

Gamasutra.com (2004b) Latest US PC GamesSales Charts, 7/28/04, Available at:http://www.gamasutra.com/php-bin/news_index.php?story=4081 (January,2005).

Gratch, J., Rickel, J., André, E., Badler, N.,Cassell, J. and Petajan, E. (2002) CreatingInteractive Virtual Humans: Some AssemblyRequired, IEEE Intelligent Systems,July/August, pp. 54-63.

Hales, D. and Edmonds, B. (2003) Evolvingsocial rationality for MAS using "tags",Proceedings of the second international jointconference on Autonomous agents andmultiagent systems, Melbourne, Australia,July 14-18, pp. 497-503.

Hemelrijk, C. K. (1996) Dominance interactions,spatial dynamics and emergent reciprocity ina virtual world., From Animals to Animats4: Proceedings of the Fourth InternationalConference on Simulation of AdaptiveBehavior, Cape Cod, MA, USA, September9-13, pp. 545-552.

Hodgins, J. K. and Pollard, N. S. (1997)Adapting simulated behaviors for newcharacters, Proceedings of the 24th annual

17

conference on Computer graphics andinteractive techniques (SIGGRAPH 97), LosAngeles, CA, USA, August 3-8, pp. 153-162.

Immerman, R. S. and Mackey, W. C. (2003)Perspectives on human attachment (pairbonding): Eve’s unique legacy of a canineanalogue, Evolutionary Psychology, 1 pp.138-154.

Isla, D., Burke, R., Downie, M. and Blumberg,B. (2001) A Layered Brain Architecture forSynthetic Creatures, Proceedings of theInternational Joint Conferences on ArtificialIntelligence (IJCAI), Seattle, WA, USA,August 4-10, pp. 1051-1058.

Jenkins, H. (2002) Game Theory: DigitalRenaissance, MIT Technology Review, pp.1-3.

Klinghammer, E. and Goodmann, P. (1985) TheManagement and Socialization of CaptiveWolves (Canis lupus) at Wolf Park, NorthAmerican Wildlife Park Foundation, Inc.,Battle Ground, IN, USA.

Kollock, P. (1998) Design Principles for OnlineCommunities, PC Update, 15, (5), pp. 58-60.

Lee, J., Chai, J., Reitsma, P. S. A., Hodgins, J.K. and Pollard, N. S. (2002) Interactivecontrol of avatars animated with humanmotion data, In SIGGRAPH '02:Proceedings of the 29th annual conferenceon Computer graphics and interactivetechniques, ACM Press, pp. 491-500.

Mateas, M. and Stern, A. (2003) TowardsBuilding a Fully Realized InteractiveDrama, Game Developers Conference, SanJose, CA, USA, March 6-8, pp. 24.

Mech, L. D. (1999) Alpha status, dominance,and division of labor in wolf packs.,Canadian Journal of Zoology, 77 pp. 1196-1203.

Mech, L. D., Adams, L. G., Meier, T. J., Burch,J. W. and Dale, B. W. (1998) The Wolves ofDenali, University of Minnesota Press,Minneapolis, MN, USA.

Mehrabian, A. and Russell, J. (1974) AnApproach to Environmental Psychology,MIT Press, Cambridge, MA, USA.

Nardi, B. (2005) Beyond Bandwidth:Dimensions of Connection in InterpersonalCommunication, Journal of Computer-Supported Cooperative Work, (in press).

Paiva, A., Machado, I. and Prada, R. (2001)Heroes, villians, magicians…: dramatispersonae in a virtual story creationenvironment, Proceedings of the 6thinternational conference on Intelligent userinterfaces, Santa Fe, NM, USA, January 14-17, pp. 129-136.

Panksepp, J. (1998) Affective Neuroscience: TheFoundations of Human and AnimalEmotions, Oxford University Press, NewYork, NY, USA.

Perlin, K. and Goldberg, A. (1996) Improv: ASystem for Scripting Interactive Actors inVirtual Worlds, Proceedings of ACMSIGGRAPH 96, New Orleans, LA, USA,August 4-9, pp. 205-216.

Prendinger, H., Mori, J., Saeyor, S., Mori, K.,Okazaki, N., Juli, Y., Mayer, S., Dohi, H.and Ishizuka, M. (2004) Scripting andEvaluating Affective Interactions withEmbodied Conversational Agents,Künstliche Intelligenz (KI) Zeitschrift, 1 pp.4-10.

Randall, N. and Pedersen, I. (1998) Who exactlyis trying to help us? The ethos of helpsystems in popular computer applications,Proceedings of the 16th annual internationalconference on Computer documentation,Quebec, Quebec, CA, September 24 - 26,pp. 63-69.

Raybourn, E. M. (2004). Designing InterculturalAgents for Multicultural Interactions. InSabine Payr & Robert Trappl (Eds.), AgentCulture: Human-Agent Interaction in aMulticultural World, Lawrence Erlbaum &Associates, Mahwah, NJ, USA, pp. 267-285.

Reeves, B. and Nass, C. (1996) The MediaEquation, Cambridge University Press,Cambridge, UK.

Reilly, W. S. N. (1996) Believable Social andEmotional Agents, Ph.D. dissertation,School of Computer Science, CarnegieMellon University, Pittsburgh, PA, USA.

Reynolds, C. (1987) Flocks, Herds and Schools:A Distributed Behavioral Model,

18

Proceedings of ACM SIGGRAPH 87,Anaheim, CA, USA, July 27-31, pp. 25-34.

Schaub, H., Zoll, C. and Aylett, R. (2003)Modelling Empathy: The EU-projectVICTEC (Virtual Information andCommunication Technology with EmpathicCharacters), Fifth International Conferenceon Cognitive Modeling, Bamberg, Germany,April 10-12, pp. 291-294.

Shannon, C. E. (1948) A Mathematical Theoryof Communication, The Bell SystemTechnical Journal, 27, (July, October,1948), pp. 379-423, 623-656.

Thalmann, D., Noser, H. and Huang, Z. (1997)Autonomous Virtual Actors Based onVirtual Sensors, Creating Personalities,Heidelberg, Germany, pp. 25-42.

Tomlinson, B. and Blumberg, B. (2002) Socialsynthetic characters (Visfiles column, editedby Bill Hibbard), SIGGRAPH ComputerGraphics, 36, (2), pp. 5-7.

Tomlinson, B., Blumberg, B. and Nain, D.(2000) Expressive AutonomousCinematography for Interactive VirtualEnvironments, Proceedings of 4th Int'l.Conf. on Autonomous Agents, Barcelona,Catalonia, Spain, pp. 317-324.

Tomlinson, B., Downie, M., Berlin, M., Gray, J.,Wong, A., Burke, R., Isla, D., Ivanov, Y.,Johnson, M. P., Lyons, D., Cochran, J.,Yong, B., Stiehl, D., Soetjipto, R.,Zaharopol, D. and Blumberg, B. (2001)AlphaWolf, Proceedings of SIGGRAPH2001: conference abstracts and applications,pp. 2.

Tomlinson, W. (2002) Synthetic SocialRelationships for Computational Entities,Ph.D. dissertation, Media Arts & Sciences,Massachusetts Institute of Technology,Cambridge, MA, USA.

US_Army (2004) The United States Army,Available at: http://www.army.mil/(December 2004).

Vaughan, R. T., Stoy, K., Sukhatme, G. S. andMataric, M. J. (2000) Go ahead, make myday: robot conflict resolution by aggressivecompetition, In From Animals to Animats 6:Proc. Int. Conf. Simulation of AdaptiveBehaviour, (Ed, Berthoz A, F. D., Meyer J,

Roitblat H, Wilson S), MIT Press,Cambridge, MA, USA, pp. 491-500.

Velasquez, J. (1998) When Robots Weep:Emotional Memories and Decision-Making,Proceedings of the Fifteenth NationalConference on Artificial Intelligence,Madison, WI, USA, pp. 70-75.

Watts, D. J. (2003) Six Degrees: The Science ofa Connected Age, W.W. Norton &Company, New York, NY, USA.

Wrangham, R. W., McGrew, W. C., de Waal, F.B. M. and Heltne, P. G. (Eds.) (1994)Chimpanzee Cultures, Harvard UniversityPress, Cambridge, MA, USA.