AI, Virtual Worlds, and Massively Multiplayer Online Games Hsinchun CHEN

View metadata, citation and similar papers at core.ac.uk brought to you by CORE

provided by Institutional Knowledge at Singapore Management University

Singapore Management University Institutional Knowledge at Singapore Management University

Research Collection School Of Information Systems School of Information Systems

1-2011 Trends and controversies: AI, virtual worlds, and massively multiplayer online games Hsinchun CHEN

Yulei ZHANG

W. S. Bainbridge

Kyong Jin SHIM Singapore Management University, [email protected]

N. Pathak

See next page for additional authors

DOI: https://doi.org/10.1109/MIS.2011.21

Follow this and additional works at: https://ink.library.smu.edu.sg/sis_research Part of the Communication Technology and New Media Commons, and the Graphics and Human Computer Interfaces Commons

Citation CHEN, Hsinchun; ZHANG, Yulei; Bainbridge, W. S.; SHIM, Kyong Jin; Pathak, N.; Ahmad, M. A.; DeLong, C.; Borbora, Z.; Mahapatra, A.; and Srivastava, J.. Trends and controversies: AI, virtual worlds, and massively multiplayer online games. (2011). IEEE Intelligent Systems. 26, (1), 80-89. Research Collection School Of Information Systems. Available at: https://ink.library.smu.edu.sg/sis_research/1515

This Journal Article is brought to you for free and open access by the School of Information Systems at Institutional Knowledge at Singapore Management University. It has been accepted for inclusion in Research Collection School Of Information Systems by an authorized administrator of Institutional Knowledge at Singapore Management University. For more information, please email [email protected]. Author Hsinchun CHEN, Yulei ZHANG, W. S. Bainbridge, Kyong Jin SHIM, N. Pathak, M. A. Ahmad, C. DeLong, Z. Borbora, A. Mahapatra, and J. Srivastava

This journal article is available at Institutional Knowledge at Singapore Management University: https://ink.library.smu.edu.sg/ sis_research/1515 Published in IEEE Intelligent Systems, 2011 February, Volume 26, Issue 1, Pages 80-89 https://doi.org/10.1109/MIS.2011.21 TRENDS & CONTROVERSIES Editor: Hsinchun Chen, University of Arizona, [email protected]

AI, Virtual Worlds, and Massively Multiplayer Online Games

Hsinchun Chen, University of Arizona Yulei Zhang, University of Arkansas

undreds of thousands of people from dif- of Second Life residents has grown dramatically, ferent physical locations can join virtual with 2.2 million residents in December 2006 and H 8.3 million by August 2007. worlds whenever they want to interact with each Collecting social media data is diffi cult, how- other and create objects in a computer-based ever, because most virtual worlds do not provide simulated environment, typically in an interac- easy data collection functions for regular users. tive 3D format. The rich social media data gener- Most of the data collected by virtual worlds and ated in virtual worlds has important implications MMOGs are proprietary content and are not for business, education, social science, and soci- available for academic research. In an earlier ety at large. Similarly, massively multiplayer on- study, we explored a technical framework to col- line games (MMOGs) have become increasingly lect avatar-related data by using a combination popular and have online communities compris- of bot- and spider-based approaches.4 To demon- ing tens of millions of players. They serve as un- strate our proposed framework, we conducted a precedented tools for theorizing about and empiri- case study on Second Life exploring behavioral cally modeling the social and behavioral dynamics differences between male and female avatars. of individuals, groups, and networks within large Figure 1 shows the proposed framework, where communities.1–3 we use the bot-based approach to collect avatar Some technologists consider virtual worlds and behavioral data and the spider-based approach to MMOGs to be likely candidates to become the collect avatar profi le information. Web 3.0. In such a “brave new world” of diverse The bot-based approach uses LibOpenMetaverse, online virtual civilizations, we believe AI can play a software library for communicating with the a signifi cant role, from multiagent avatar research Second Life server. LibOpenMetaverse lets each and immersive virtual interface design to virtual bot conduct data collection on a much larger scale world and MMOG Web mining and computa- than the basic Linden Script Language (LSL). By tional social science modeling. logging onto the Second Life server using Lib- OpenMetaverse as clients, we used avatars as bots Avatar Data Collection Example to automatically collect other avatars’ behavioral Virtual worlds will become another type of im- data in a given region. We sent a bot to a given portant world in people’s lives in addition to the region where it collected both the behavioral data real world where they physically live. The Gartner and the avatar identifi ers (such as names) at the Group estimated that 80 percent of active Internet same time. The avatar names help the spider-based users will have a “second life” in the virtual world approach collect the avatars’ profi le data. The by the end of 2011 (www.gartner.com/it/page. spider-based approach consists of three compo- jsp?id=503861). Second Life (http://secondlife. nents. Avatar profi le spidering collects the profi le com), developed by Linden Lab and publicly pages of all the avatars whose behavioral informa- launched in 2003, is currently the most popular tion had been gathered using the prior bot-based 3D virtual world. Since its inception, the number approach. Profi le data parsing generates avatar

IS-26-01-TandC.indd 80 25/01/11 4:23 PM Improved bot-based approach

Avatar data fields including avatar name, behavioral data Avatar behavioral birthday, and picture as well as a data collecting short self-introduction. Lastly, avatar personal data are used to infer ava- Bots Avatar identifiers tars’ other personal characteristics, such as gender and virtual age. Following the proposed integrated Spider-based approach framework, we conducted a case study to examine the differences in Profile data Avatar profile virtual world actions (such as flying, Avatar profile parsing spidering running, and walking) between the two virtual genders—that is, the self- reported gender in virtual worlds. Avatar personal characteristics Avatar personal Avatar names, self-introductions, and extraction/coding characteristics profile pictures were parsed out to three avatar gender coders. Only the avatars whose gender information, either male or female, was able to be Figure 1. Proposed framework for avatar data collection. The bot-based approach collects avatar behavioral data, and the spider-based approach collects avatar clearly decided without any disagree- profile information. ment or confusion among all three coders were kept. Independent group Table 1. Results of the virtual world gender analysis. t-tests were conducted to compare the Male N = 537 Female N = 797 action differences between the two (mean/standard (mean/standard T-test genders as well as two age groups. Avatar action deviation) deviation) (t-stats/p-value) Table 1 shows the analysis results. Flying 0.020 / 0.044 0.015 / 0.037 2.16 / 0.015* For flying (p-value = 0.015), run- Hovering 0.042 / 0.062 0.036 / 0.069 1.51 / 0.065 ning (p-value = 0.023), turning left Running 0.020 / 0.045 0.016 / 0.040 2.00 / 0.023* (p-value = 0.010), turning right (p- value = 0.046), and walking (p-value = Striding 0.018 / 0.028 0.016 / 0.025 1.23 / 0.109 0.005), male avatars were signifi- Turning left 0.099 / 0.071 0.090 / 0.071 2.33 / 0.010** cantly more likely to perform these Turning right 0.095 / 0.071 0.089 / 0.073 1.68 / 0.046* actions than female avatars. Female Walking 0.143 / 0.105 0.129 / 0.099 2.62 / 0.005** avatars were significantly more likely to stand as compared to male ava- Standing 0.775 / 0.161 0.795 / 0.160 −2.26 / 0.012* tars (p-value = 0.012). Male ava- *Significant at 0.05 level. **Significant at 0.01 level. tars seemed more likely to hover and stride, but the results were not statis- patterns. All these data can be read- “Secondary Avatars and Semiautono- tically significant. Overall, male ava- ily obtained (some with further pro- mous Agents,” William S. Bainbridge tars tended to perform more “active” cessing) based on our proposed bot- presents semiautonomous agent as- actions than female avatars. spider framework for virtual world sistants that have been incorporated These findings are consistent with avatar data collection. into recent MMOGs. Although the real people in the physical world, AI behind these agents is basic, their where males are in general more In This Issue behavior in groups and complex sit- physically active than their female This issue includes two articles with uations can provide insights into peers. Our future research will in- additional research examples from how future intelligent systems might volve examining other behavioral distinguished experts in social sci- function in many realms of activity. and avatar characteristics in virtual ence and computer science. Each arti- Bainbridge reports results from ex- worlds, including virtual world ex- cle presents a unique research frame- tensive explorations of two popular perience, age, education level, social work, computational methods, and game worlds, Dungeons and Drag- network formation, and interaction selected results. In the first article, ons Online (DDO) and Star Trek

January/february 2011 www.computer.org/intelligent 81

IS-26-01-TandC.indd 81 25/01/11 4:23 PM Second Life,” IEEE Intelligent Systems, vol. 25, no. 4, 2010, pp. 17–23.

Hsinchun Chen is the McClelland Profes- sor of Management Information Systems and the director of the Artificial Intelligence Lab at the University of Arizona. Contact him at [email protected].

Y ulei Zhang is a visiting assistant professor in the Information Systems Department at the University of Arkansas and a PhD can- didate in management information systems at the University of Arizona. Contact him at Figure 2. A primary avatar and six secondaries in Dungeons and Dragons Online. [email protected]. The game lets players rent secondary agents to help with game play when friends and other human players are unavailable. Secondary Avatars and Online (STO). Both MMOGs empha- The authors argue that the close Semiautonomous Agents size team play, and AI assistants serve collaboration between social scientists as substitutes for human beings when and computer scientists is creating an William Sims Bainbridge, National the player’s friends are not online or emerging area called ”computational Science Foundation are unable to participate. The author social science,” where computation is believes that for more advanced fu- used as an integral mechanism in so- Recent massively multiplayer online ture virtual world and MMOG re- cial science research. Given current (MMO) games are prototyping search, scientists will need to partici- trends in virtual-and-real-world inter- methods through which users can in- pate actively in the design of virtual actions, this area of research will interact with multiple semiautonomous worlds, incorporating their experi- crease in importance and impact. agent assistants.1–3 Although these ments while helping the game cre- agents use simple AI, their behav- ators innovate. ior in groups and complex situations Acknowledgments In the second article, Kyong Jin can provide insights into how future This material is based on work supported Shim, Nishith Pathak, Muhammad in part by grants DOD HDTRA-09-0058, mixed-initiative systems might func- Aurangzeb Ahmad, Colin DeLong, NSF CNS-070933, NSF CBET-0730908, tion in many realms of activity. Here Zoheb Borbora, Amogh Mahapatra, and NSF IIS-0428241. I report results from extensive explo- and Jaideep Srivastava report re- ration of two game worlds, Dungeons search results from their multiyear re- References and Dragons Online (DDO) and Star search project on MMOGs. The Uni- . 1 W.S. Bainbridge, ed., Online Worlds: Trek Online (STO). Both emphasize versity of Minnesota group adopts Convergence of the Real and the Vir- team play, and AI assistants serve as knowledge discovery as a core mech- tual, Springer, 2010. substitutes for human beings when anism to analyze MMOG logs and 2. W.S. Bainbridge, The Warcraft Civiliza- the player’s friends are not online or build models of social interactions tion: Social Science in a Virtual World, when other players are unavailable. in games. They constructed com- MIT Press, 2010. bat, mentoring, and trust networks 3. J. Srivastava, tech. reports, Dept. of Hirelings in DDO based on data from Sony’s popular Computer Science and Eng., Univ. AI assistants in DDO are mercenaries, MMOG Everquest II. They also pro- of Minnesota, 1998–2009; www. called hirelings, who must be rented, vide an excellent summary of their cs.umn.edu/research/technical_reports. either with virtual coins earned by ongoing research in performance and php?page=author&author_id=68. completing missions inside DDO or learning, player churn analysis, and 4. Y. Zhang et al., “An Integrated Frame- with points bought in an online store identifying undesirable behavior in work for Avatar Data Collection from with real dollars. Figure 2 shows MMOGs. the Virtual World: A Case Study in my avatar, Sagittarius, standing

82 www.computer.org/intelligent I Eee INTELLIGENT SYSTEMS

IS-26-01-TandC.indd 82 25/01/11 4:23 PM on a footbridge over the Menechta- One mission carried out in the ruins team attacked them: Coyle, his two run desert with three agents on each of Threnal illustrates how complex assistants, and three of the hirelings. side—fi ve hirelings who cost US$2 the action of a simulated team can be- The two cleric hirelings stood passively for a one-hour rental, plus his pet come. Buried deep beneath the Earth watching the battle, only reacting hyena. These are secondary avatars, for 500 years, the Library of Threnal when Sagittarius took damage or when in complex ways under the user’s has just been entered by archeolo- given specifi c instructions. Sagittarius control. gists. An ambitious man named Coyle set his attention on Coyle and, when- For each hireling, there is a hot is seeking an ancient manuscript that ever Coyle took damage, instructed bar holding 10 icons, and clicking could give him magical powers, but one of the clerics to heal him. In the one either gives a specifi c command he has kept secret this reason for en- pause after one attack and before the or sets the agent’s general behavior. tering the library. In DDO, he is rep- next, Sagittarius would rapidly use Four of these are actions that only resented as an autonomous nonplayer the clerics plus his own limited heal- the particular hireling can perform. character, as are two assistants who ing abilities to restore the health of any For example, a cleric might have one accompany him. They stand in the members of the team who needed it. icon to give signifi cant healing to a library’s central book repository, re- The nine-member team included selected wounded ally and another sponding as waves of monsters of dif- three agents who could not be con- to give minor healing to all mem- ferent kinds enter from random direc- trolled at all, three secondary ava- bers of the party. All hirelings also tions. As soon as Coyle sees a monster tars who were set to function autono- have six icons that perform these mously but could have been controlled, actions: and three characters (counting the primary avatar and two secondaries) un- • toggle between stand still or follow DDO and STO are der the player’s direct control. Given the primary avatar; that groups of attacking monsters • summon the hireling to the avatar, fantasy games, but they came from random directions, and the even over a long distance; behaviors of the AIs were infl uenced • set the hireling to behave autono- offer a glimpse into the by their exact locations in relation mously; to the others, even simple program- • engage defensive mode, responding real future that many ming produced a remarkably complex only if it or the primary avatar is result. attacked; people will inhabit. • engage passive mode, doing noth- Secondary Avatars in STO ing; and STO chiefl y takes place in outer • interact with a target. he attacks it, and his assistants follow space, as fl eets of spaceships bat- his lead. Unfortunately, this will re- tle each other, but some “away mis- The last of these might, for exam- sult in his quick death, and the goal sions” take crews of fi ve avatars to ple, have the secondary avatar pull of the mission is keeping him alive for a planet’s surface or inside a ship. a lever in one of the many dungeon 15 minutes. A mission begins when the play- mazes, when two widely separated Sagittarius rented fi ve hirelings— er’s ship enters a solar system, and levers must be pulled simultaneously two cleric healers and three warriors STO automatically assembles play- to open a gate. The hireling follows with different attack modes to fi ght the ers into teams if they have not al- the primary avatar to one lever, is different monsters. On ordinary mis- ready teamed up with friends and if told to stand still, and the user right sions, he often instructs all his hire- other players are starting the mission clicks the lever to select it as the hirelings to support him in defensive mode, at roughly the same time. When this ling’s target. The primary avatar then which means they heal him and at- does not happen, the team is com- walks to the second lever and acti- tack only enemies who have harmed pleted by secondary avatars concep- vates it, while the player clicks the him. This time, he gave the three war- tualized as officers from the bridge hireling’s icon to tell it to interact riors instructions to act autonomously, of the player’s ship. with the fi rst lever. These combined which means they immediately attack By the time he became an admi- actions open the gate, something that enemies they detect. Thus, when mon- ral, my avatar Rho Xi had 10 of one avatar could not do alone. sters approached, six members of the these bridge offi cers but only used the

January/february 2011 www.computer.org/intelligent 83

IS-26-01-TandC.indd 83 25/01/11 4:23 PM Table 2. Characteristics of five secondary avatars in Star Trek Online. Character Ability Description Thuvia, Andorian Suppressing fire Slows down and reduces the damage caused by an enemy. female tactical officer Leg sweep A short distance attack with a high chance of knocking back any nearby foes. Photon grenade Throws an explosive grenade at the target. Fire on my mark Reduces the enemy target’s resistance to damage. Marx, Ferengi male Overwatch Increases nearby allies’ resistance to damage, and impedes nearby enemies. tactical officer Lunge Leaps at the target, delivering a melee attack and possibly knocking it down. Focus fire Lowers the damage resistance of the target with each attack from allies. Ambush Increases damage from next attack, best used while in stealth mode or not under attack. Tonga, Trill female Shield generator Creates a fixed position generator that recharges personal shields for allies in the engineering officer fabrication immediate area. Reroute power to shields Strengthens the target character’s shield against damage. Shield recharge Restores lost personal shield energy to you or an ally. Fuse armor Snares, roots, and then holds the target, reducing its mobility. Flash, Human male Medical generator Creates a fixed position generator that heals injured allies in the immediate area. engineering officer fabrication Phaser turret fabrication Creates a fixed position weapon platform that fires at nearby enemies. Chroniton mine barrier Deploys five mines, which explode when hostiles get close. Cover shield Creates a broad shield that deflects attacks from a particular direction. Azura, Vulcan Neural neutralizer Placates nearby enemies, discouraging them from attacking. female science officer Vascular regenerator Rapid recovery from serious injury. Medical tricorder Quickly heals injury to your avatar or an ally. Nanite health monitor Scans nearby allies to determine if they are seriously injured, then heals wounded.

five listed in Table 2, typically leav- perform, putting the emphasis on future that many people will inhabit. ing one or the other engineer behind others. Their AI systems possess primitive when beaming down on an away mis- Unlike the situation in DDO, a machine learning, as the secondar- sion. Each secondary avatar has four STO player can tell each secondary ies concentrate on the enemies that special abilities, in addition to us- to go to a particular location within have done most damage to them. The ing whatever weapon it carries, and the field of view or send all four to systems also appear to be hierarchi- the descriptions are adapted from the same spot. When encountering a cal because secondaries often choose STO’s wiki (http://stowiki.org/wiki/ group of enemies, the user can leave at random among the actions avail- Bridge_officer_abilities). the secondaries free to respond to the able to them, but they respond cor- Operating STO secondary ava- situation autonomously, control each rectly when approached or attacked tars involves much end-user pro- one in detail, or have them all attack by an enemy or when they notice that gramming, although not generally the same enemy officer to remove him a team member needs healing. conceptualized as such because it is from action quickly and then respond In the future, our lives might done by clicking or dragging icons. autonomously as the combat unfolds. incorporate many semiautonomous The user decides which of several Visually, the result is exceedingly agents, and we may set the degree of possible abilities each bridge offi- complex, especially when engineers autonomy of any secondary avatar as cer will have and what weapons, set up equipment that then functions events require. DDO and STO place a armor, and special equipment they autonomously. great premium on a good user inter- will carry. In action, the hot bars for face because decisions must be made the four secondaries have icons rep- Implications of quickly. Future use of agents might resenting their abilities, and clicking Secondary Agents take place over longer periods of one during battle removes it from DDO and STO are fantasy games, time and thus provide more scope for the list of actions the secondary can but they offer a glimpse into the real advanced machine learning. On the

84 www.computer.org/intelligent I Eee INTELLIGENT SYSTEMS

IS-26-01-TandC.indd 84 25/01/11 4:23 PM team level, if not that of the individ- Analyzing Human unmatched by any real-life experi- ual agent, these games already repre- Behavior from mental setup. The data collected pro- sent highly intelligent systems. Thus, Multiplayer Online Game vides an excellent means of studying they have become both a laboratory Logs: A Knowledge human social behavior with respect for research on human-multiagent in- Discovery Approach to the complete context of the envi- teraction and a training ground for ronment for a large population. the future users and creators of such Kyong Jin Shim, Nishith Pathak, Because MMOGs usually con- systems. Muhammad Aurangzeb Ahmad, sist of a thin-client architecture, all Among the interesting research Colin DeLong, Zoheb Borbora, player actions are captured in the questions is when it will be advan- Amogh Mahapatra, and click-stream logged at the server. tageous to conceptualize an agent as Jaideep Srivastava, University Along with the players’ behaviors, a secondary avatar, possessing some of Minnesota the games often also store a brief re- degree of personhood. In terms of cord of their real-life profile. And it human-computer interaction, giving Observing and analyzing a phe- is common for popular virtual worlds an agent a person’s appearance might nomenon at unprecedented levels of to have hundreds of thousands of be helpful for inexperienced users, but granularity not only furthers our un- players generating copious amounts sophisticated users might do better derstanding of it but also transforms of data at any given time. A key chal- with agents represented through a the way it is studied. With the mass lenge is developing methods that can variety of metaphors that best ex- adoption of the Internet in our daily analyze relationships while scaling press their different functions. Re- lives and the ability to capture high- to terabytes of data. The data also searchers can initially study such resolution data on its use, we are have a temporal component that is questions by observing how users of at the threshold of a fundamental often an integral part of the analysis different levels of sophistication op- shift not only in our understand- and introduces further relationships. erate diverse resources inside existing ing of the social and behavioral sci- Thus, while providing an exciting game worlds. For more advanced re- ences, but also the ways in which we new tool for social science research, search, scientists will need to partici- study them. Massively multiplayer the virtual worlds also present a set pate actively in the design of virtual online games (MMOGs) have be- of difficult and novel computational worlds, incorporating their experi- come increasingly popular, with tens challenges. ments while helping the game cre- of millions of users.1 They serve as Our research uses knowledge dis- ators innovate. unprecedented tools to theorize and covery as a core mechanism to ana- empirically model the social and be- lyze MMOG logs and build models Acknowledgments havioral dynamics of individuals, of social interactions in games and The views expressed in this essay do not nec- groups, and networks within large understand how users’ relationships essarily represent the views of the National communities. are affected by the variety of envi- Science Foundation or the United States. MMOGs are online spaces provid- ronmental factors present in each ing users with comprehensive virtual world. MMOGs provide mechanisms References universes, each with its own unique to foster social activities among us- . 1 W.S. Bainbridge, Online Multiplayer context and mechanics. They range ers, letting them form groups, guilds, Games, Morgan and Claypool, 2009. from the fantastical world of elves, corporations, and so on and tackle 2. W.S. Bainbridge, ed., Online Worlds: dwarfs, and humans to space far- collaboration-oriented tasks, such as Convergence of the Real and the Vir- ing corporations and mirrors of our raiding a dungeon for gold. Data col- tual, Springer, 2010. world. Large numbers of users in- lected from these mechanisms pro- 3. K.E. Merrick and M.L. Maher, Moti- teract and role-play via the in-game vide an excellent means of studying vated Reinforcement Learning: Curi- mechanics. With their increasing human social behavior with respect ous Characters for Multiuser Games, popularity, researchers are realizing to the complete context of the en- Springer, 2009. that MMOGs are a means to fully vironment. Such research is infea- observe an entire isolated universe. sible for real-life activities because William Sims Bainbridge is a director of Each action is logged, and the level it is impossible to track and record the Human-Centered Computing Program of granularity and completeness with complete information on a large at the US National Science Foundation. which information is collected is population. Knowledge discovery

January/february 2011 www.computer.org/intelligent 85

IS-26-01-TandC.indd 85 25/01/11 4:23 PM Data Techniques Applications Marketing Game and *Customer churn analysis virtual worlds * Avatar characteristics * Game mechanics Military * Player versus environment P-star * Player versus player *Team formation * Combat training Classiﬁcation Feature extraction Game vendors *Gold farming detection * Behavior proﬁling * Social network Clustering *Game design analysis Education psychology Regression *Student assessment and performance management Social networks *Curriculum design (combat, housing, Others mentoring, friends, Economics enemies) Sociology ...

Figure 3. Data obtained from massively multiplayer online games (MMOGs) can be used for various scientific and commercial purposes. Scientific purposes include data-intensive empirical studies of various social science theories and models, including organizational behavior, team dynamics, competition and cooperation, mentoring and teamwork, and performance and achievement. Commercial applications include predicting player churn, identifying undesirable behavior, and social-influence- based recommendations.

is a key computational approach to when a user kills a monster, com- require multiple avatars with dif- realizing this promise, and in the pletes a quest, forms a group with ferent sets of skills. Team members process, it creates opportunities to other users, or engages in trade—a must rely on each other while play- push the frontiers of knowledge dis- record is created in the in-game be- ing their respective set of capabili- covery itself. havior logs. The behavior logs are ties and abilities during combat. time-stamped and can be parsed to Studying the formation of task- Game Data and uncover a wealth of information. oriented groups is an important first Social Networks For our research, the most impor- step to understanding the dynamics Our research focuses on data from tant information gleaned from the of team collaboration. Team for- Sony’s popular MMOG EverQuest logs is about user relationships. Play- mation is highly influenced by the II. Figure 3 provides an overview of ers can socially interact with each common interests of players dur- the overall approach to our project. other using various methods pro- ing challenging tasks. Combat team The data primarily consist of avatar vided by the game mechanics. They performance is positively correlated characteristics, player demograph- can form groups to tackle monsters, with group size (to a certain extent) ics, and in-game behavior. Avatar go on quests together, trade with each and group level. More than 12 mil- characteristics provide data on play- other, express different levels of trust, lion teams form monthly, making ers’ virtual avatars in the game. This create and join guilds, mentor other games such as EverQuest II an ex- includes the avatars’ gender, race, players, and so forth. Each of these cellent venue for studying human faction, health and damage statis- can be represented as a social net- and organizational behaviors. tics, and more. Player demograph- work among players, which we can • Mentoring network. Mentoring oc- ics include real-life age, gender, geo- then study to understand how they curs in MMOGs when an experi- graphic location, and so on. Survey are impacted by and affect both the enced player helps a novice player results from a small sample of players players and the environment. The fol- gain skills. Mentors move down are also available that include their lowing are examples of the types of to the level of their apprentices to responses to multiple facets such as relationship networks that we can teach them combat skills, transmit feelings toward the game and other construct from game logs. knowledge about different aspects game users, psychological fitness, and of the game, and perform one or real-life behavior outside the game. • Combat network. Team formation more tasks with the apprentices. The most unique data are the in- is a common occurrence in many Mentors gain achievement points game behavior logs, which contain MMOGs because the games are de- from successfully completing tasks each user’s game activities. Anytime a signed to encourage social interac- with their apprentices. Apprentices significant event occurs—for example, tions. For instance, certain quests also gain XP points (plus bonus

86 www.computer.org/intelligent I Eee INTELLIGENT SYSTEMS

IS-26-01-TandC.indd 86 25/01/11 4:23 PM XP for participating in mentoring) player churn, identifying undesirable study of individual player characteris- from, for instance, killing mon- behavior, and social-infl uence-based tics, team composition and character- sters. More importantly, they attain recommendations. istics, social interactions among team valuable lessons and experience members, and game environments by interacting with their mentors. Performance and Learning can reveal a great deal about the rec- Skilled and knowledgeable men- In recent years, researchers have ipes for success in achieving various tors will help their apprentice pick found games and virtual worlds game objectives.2,3 up the game quickly and effi ciently. to be a sound venue for studying A player can have multiple mentors human behavior. In particular, in Player Churn Prediction in a session and can interact with learning sciences and educational The estimated worth of the MMOG one mentor at a time. psychology communities, systematic market currently stands at US$6 • Trust network. Trust networks are studies of player performance and billion, with little expected future diffi cult to study because mapping social interactions captured in game letdown in growth.1 Activision- trust relationships in the real world logs have led to a better understand- Blizzard’s game World of Warcraft is diffi cult. Online interactions, ing of learning, collaboration, social has the largest market share of any however, provide a unique oppor- participation, literacy, and learning MMOG by far at roughly 60 per- tunity to study such relationships. trajectory at the individual and group cent, and the revenue potential has In some MMOGs such as Ever- levels. attracted several new game makers to Quest II, for example, players can this market segment. With increased buy virtual houses and grant other competition, game companies are es- players varying levels of access to pecially concerned with customer ac- them. Thus, trust in these games The estimated worth quisition and retention. is described with respect to a com- Using Web-mining techniques,4 we modity. Analyzing these trust net- of the MMOG market analyzed players’ in-game behaviors. works reveals that their properties When combined with subscription- are similar to trust networks in currently stands at related information, we gained sig- other domains. nifi cant insight into the mechanics US$6 billion, with little of player engagement and their likeli- Interestingly, networks generated hood to churn.4 Game developers can by in-game processes are closely anal- expected future letdown use this knowledge to improve mar- ogous to the real world. For example, keting and retention strategies as well money transactions are similar in the in growth. as to tweak game mechanics to en- game world and the real world. In hance player engagement. contrast, role-playing networks dif- As part of churn analysis, we try fer from any real-world networks In addition, the military has widely to understand different player moti- because this is a unique in-game used simulation-based training in vations. Two key indicators of player phenomenon. virtual environments. Traditionally, engagement in MMOGs are their in- human instructors are tasked to ab- dividual involvement and achieve- Applications stract from individual soldier behav- ments as well as their level of in-game As Figure 3 shows, data obtained iors to the team behaviors and goals. socialization with other players. from MMOGs can be used for various One challenge with this approach is Churn-prediction models built on scientifi c and commercial purposes. that the interpretation is subject to a combination of these factors have Scientific purposes include data- the instructor. Another challenge greater predictive power than those intensive empirical studies of various is that human instructors cannot cap- that are not. MMOGs have differ- social science theories and models, in- ture information about rapidly evolv- ent business approaches ranging from cluding organizational behavior, team ing events or low-level information. subscription-based to free-to-play dynamics, competition and coopera- Therefore, effi ciently and effectively models. The churn analysis and pre- tion, mentoring and teamwork, and determining what teams are doing as diction models also vary based on the performance and achievement. Com- well as how and why they are doing it nature of the underlying subscription mercial applications include predicting remain challenging tasks. A systematic model.

January/february 2011 www.computer.org/intelligent 87

IS-26-01-TandC.indd 87 25/01/11 4:23 PM Identifying Undesirable Behavior proven to be highly engaging, given here is an example of data-driven mod- Gold farming refers to the illicit prac- the increasing rate of adoption and eling, which is a hard sell at best to a tice of gathering and selling virtual the amount of time people spend community trained in and used to hy- goods in online games for real-world playing them. This has opened up pothesis-driven model building. currency. Gold farmers are especially new avenues for studying social sci- The close collaboration between frowned upon by game developers ences, as this article illustrates. social scientists and computer sci- because they disrupt the economic As the science’s success becomes entists is creating an emerging area balance of a game’s virtual economy, increasingly visible, commercial ef- called computational social sciences, go against the ethos of fair play, and forts will not lag far behind. There where computation is used as an inte- might make certain areas of the game is already talk in commercial fo- gral mechanism to do social science diffi cult to access for legitimate play- rums, such as the annual Game research. Given current trends of user ers. Consequently, game administra- Developer Conference (GDC, www. behavior, this area can only increase tors actively seek to ban gold farmers gdconf.com), of the need to use game in importance. Emerging experiences from their games. analytics commercially. Thus, one of show that our proposed approach can The problem of gold farmer detec- the major trends expected for this re- aid in the process of classical model tion can be posed as a binary classi- search is to transition into the indus- building. However, the computa- fi cation task using in-game features try. Next, the success of commercial tional and social science communities’ such as the amount of time played, analytics will lead to game developers mode of thinking has suffi cient gaps sequence of play, associations with that such collaborations are unlikely other players, and offl ine character- to arise easily. We sincerely hope to be istics such as language, location, and proven wrong on this prediction, to gender.5 We can further divide gold Online games have the benefi t of both communities. farmers into four subtypes: proven to be highly • gatherers accumulate gold (virtual Acknowledgments currency), engaging and opened The Virtual World Exploratorium proj- • bankers are low-activity accounts ect is a major collaborative effort involv- up new avenues for ing numerous educational institutions, that hold gold in reserve, industrial partners, and sponsoring agen- • mules (or dealers) are one-time cies. The academic institutions are North- characters that act as a link in the studying social sciences. western University, University of Southern chain to distance the customer California, University of Illinois at Urbana- from the operation, and Champaign, University of Chicago, and University of Toronto. The data for this re- • marketers (or spammers) advertise inserting more data collection hooks search has been provided by Sony Online the company’s services. into the system and building more Entertainment, Linden Lab, and Kingsoft. personalized games. Subsequently, we This research has been sponsored by the Different in-game and network fea- might also see personalized recom- US National Science Foundation, the Army tures must be used to detect each mendations and product placements Research Institute, the Air Force Research Laboratory, and the Army Research Labo- type. Not all gold farmers are labeled in the games on a regular basis—for ratory through a Network Science Collab- as such. Thus, administrators must example, see www.xconomy.com/ orative Technology Alliance (CTA) award. use labeled propagation techniques to san-francisco/2010/09/30/npario-shows- Although this article focuses on computa- identify other potential gold farmers. ea-how-to-track-and-target-consumers- tional aspects, our ideas have been shaped across-web-mobile-social-internet-tv- by cross disciplinary interaction with colleagues from all partnering institutions, Trends and Controversies and-game-consoles. and we owe them all a debt of gratitude. Games fulfi ll many needs for humans Any major change is not without The ideas and positions expressed herein, at multiple levels and thus have been controversy, and this area is no excep- however, are solely the authors’ and do not popular since the dawn of time. The tion. On the commercial side, there are represent the offi cial positions, either stated Internet medium, with its new ca- issues of privacy violation and ruined or implied, of the sponsoring agencies. More information about this project can pabilities, lets game developers and game experience. For the social sci- be obtained from http://vwobservatory.com/ players exercise their imagination in ences, the controversy is even more fun- drupal, and all publications are available at new ways. Online games have clearly damental. The approach we describe www.vwobservatory.com/fi leshare.php.

88 www.computer.org/intelligent Ieee InTeLLIGenT SySTeMS

IS-26-01-TandC.indd 88 25/01/11 4:23 PM References . 5 B. Keegan et al., “Dark Gold: Statistical Colin DeLong is a graduate student in . 1 A. Bagga, “The Emergence of Games Properties of Clandestine Networks in the Department of Computer Science and as a Service,” industry report, Think Massively Multiplayer Online Games,” Engineering at the University of Minnesota. Equity, 2009. Proc. IEEE 2nd Int’l Conf. Social Contact him at [email protected]. 2. K. Jin Shim, R. Sharan, and J. Srivas- Computing, IEEE CS Press, 2010, tava, “Player Performance Prediction in pp. 201–208. Zoheb Borbora is a graduate student in the Massively Multiplayer Online Role- Department of Computer Science and En- Playing Games (MMORPGs),” Proc. Kyong Jin Shim is a graduate student and gineering at the University of Minnesota. Pacific-Asia Conf. Knowledge Discov- research fellow in the Department of Com- Contact him at [email protected]. ery and Data Mining (PAKDD), vol. 2, puter Science and Engineering at the Univer- LNAI 6119, Springer, 2010, pp. 71–80. sity of Minnesota. Contact her at kjshim@ A mogh Mahapatra is a graduate student 3. Y. Huang et al., “The Formation of cs.umn.edu. in the Department of Computer Science and Task-Oriented Groups: Exploring Engineering at the University of Minnesota. Combat Activities in Online Games,” N ishith Pathak is a graduate student in Contact him at [email protected]. Proc. IEEE Int’l Conf. Computational the Department of Computer Science and Science and Eng., vol. 4, IEEE CS Press, Engineering at the University of Minnesota. Jaideep Srivastava is a professor in the 2009, pp. 122–127. Contact him at [email protected]. Department of Computer Science and En- 4. J. Kawale, A. Pal, and J. Srivastava, gineering at the University of Minnesota. “Churn Prediction in MMORPGs: Muhammad Aurangzeb Ahmad is a grad- Contact him at [email protected]. A Social Influence Based Approach,” uate student in the Department of Computer Proc. IEEE Int’l Conf. Computational Science and Engineering at the University Selected CS articles and columns Science and Eng., vol. 4, IEEE CS Press, of Minnesota. Contact him at mahmad@ are also available for free at 2009, pp. 423–428. cs.umn.edu. http://ComputingNow.computer.org.

Call for artiCles IT Professional seeks original submissions on technology solutions for the enterprise. Topics include • emerging technologies, • social software, • cloud computing, • data management and mining, • Web 2.0 and services, • systems integration, • cybersecurity, • communication networks, • mobile computing, • data center operations, • green IT, • IT asset management, and • RFID, • health information technology. We welcome articles accompanied by Web-based demos. For more information, see our author guidelines at www.computer.org/itpro/author.htm. www.Computer.org/itpro

January/february 2011 www.computer.org/intelligent 89

IS-26-01-TandC.indd 89 25/01/11 4:23 PM