The MARS - A Multi-Agent Recommendation System for Games on Mobile Phones

Transcription

1 The MARS - A Multi-Agent Recommendation System for Games on Mobile Phones Pavle Skocir, Luka Marusic, Marinko Marusic, Ana Petric University of Zagreb Faculty of Electrical Engineering and Computing Unska 3, HR-10000, Zagreb, Croatia {pavle.skocir, luka.marusic, marinko.marusic, ana.petric}@fer.hr Abstract. In order to achieve flow (i.e. complete focus on playing followed by a high level of enjoyment) and increase player retention (i.e. keep a user playing a game longer and more often) it is important that difficulty of the game that a user is playing matches her/his skills. Due to a large amount of different games which are available to users, it is not easy for them to find games which best suit their skills and abilities. In this paper we propose a recommendation algorithm based on the information gathered from users interaction with a game. We use that information to model users success and progress in the game as well as motivation for playing. Besides, the proposed algorithm also takes into account user preferences, mobile phone characteristics and game related information which is gathered from users once the game is available on the market. Before enough information is gathered from users, the algorithm uses the information gathered during the game development phase and acquired from game developers and testers. In the implemented multi-agent system, after a user finishes playing a game, she/he receives a notification with a list of games which best suit her/his skills and preferences. Keywords: computer and video games, user experience, recommendation system, multi-agent system 1 Introduction The amount of content which various service providers offer to their users is rapidly rising every day. Games, as a very popular content type, are no exception to that rule. According to Entertainment Software Association s (ESA 1 ) annual report [1] in almost three quarters of American households people play computer or video games and 55% of them play those games on their mobile phones and/or handheld devices. Since people (purchase and) play those games to have fun, entertaining the player and achieving player enjoyment is considered to be the most important goal for computer and video games. Different methods for evaluating gameplay experience are used in all phases of the game development process with the purpose of improving design of games 1

2 2 P. Skocir, L. Marusic, M. Marusic, A. Petric and various aspects of gameplay experience [2, 10]. Some of those methods include self-reporting (e.g., interviews, questionnaires), measuring psychological state of the player (e.g., hart rate, galvanic skin response) [15] and behaviour analysis from game metrics [21]. Two related concepts which are often used to describe and model gameplay experience are flow [6, 20] and immersion [3, 8]. Flow can be defined as the feeling of complete focus in the game followed by a high level of enjoyment [5] while immersion is used to describe the degree of engagement or involvement in the game [3]. In their own way, both concepts are closely related to the challenge a game represents to the user. The balance between a challenge of successfully completing a certain aspect of the game (i.e. game task or level) and user s skills and abilities to do so is considered to be an important precursor to flow [5, 20]. Another measure for successfulness of the game is player retention, i.e. the ability of a game to keep a user playing the game for longer periods of time and/or returning to play a game later [17]. Research has shown that it is necessary that the game quickly adapts to the user (i.e. finds a balance between her/his skills and game difficulty) in order to improve player retention [22]. In our research, we use those findings to model a multi-agent recommendation system where agents monitor users in-game behaviour and analyse their progress in a certain game. The results of the analysis are used to recommend games which best suit users skills and abilities in order to longer maintain users flow and increase the time users spend playing a certain game. The paper is organized as follows. Section 2 gives an overview of related work on recommendation systems for other content types, while Section 3 describes the model of the proposed recommendation system and the architecture of the implemented multi-agent system. Section 4 presents the recommendation algorithms which take into account modelled user experience parameters as well as user preferences. Section 5 concludes the paper and gives outline for future work. 2 Related Work Liang et al. [13] showed that the problem of content overload can be reduced using personalized services thus increasing users satisfaction and offering content which best suits user preferences. Therefore, if service providers recommend content that meets users interest, users save time and effort because they do not have to look for potentially interesting content on their own. There exist various content recommendation systems for different content types (e.g., video, music, books) and here we will briefly present the main concepts from a few familiar ones. One of the well known recommendation systems is the one used by YouTube 2 [7]. When forming recommendations it considers data collected from recent user actions (e.g., watched, liked video) and content data (e.g., raw video streams as well as video metadata) including periodically calculated relatedness score for 2

3 The MARS - A Multi-Agent Recommendation System for Games 3 each pair of videos depending on how often were those videos co-watched within sessions. Recommended videos are scored and ranked in dependence of video quality, user specificity and diversification criteria. Another known recommendation system is the one used by Amazon 3. It uses item-to-item collaborative filtering method in its online store [14]. Unlike most recommendation systems which organize similar users into groups and offer them products that other users from their group purchased/viewed/positively rated, Amazon s algorithm finds items that are similar to the ones that the user already purchased or rated. The found items are aggregated and the most popular or correlated ones are recommended to the user. The table of similar items is created offline and it is used by the algorithm s online component to create recommendations for users. There also exist systems that use several different means for recommending products to users, like the system proposed in [4] that uses fuzzy logic and data mining techniques to create recommendations. Based on its domain expert knowledge and information obtained from a user, the system uses multi-attribute decision making method to calculate the optimality of each product and recommends the best ranked ones. In recommendation systems, software agents can be used for collecting information about users content consumption by monitoring their actions and modelling their preferences [12] or for learning users preferences from their recent actions and giving personalized suggestions based on those preferences [16]. 3 A Multi-Agent System for Game Provisioning This section describes the Multi-Agent Recommendation System (MARS) designed to recommend games for mobile phones to users based on their skills and abilities demonstrated in previous games they played. The recommendation model and the architecture of the implemented multi-agent system are presented. 3.1 The Model of the Recommendation System Fig. 1 shows the recommendation model for games which is based upon four main processes: playing games, analysing game consumption data, recommending and provisioning games. While a user plays a game, an agent monitors her/his activities on her/his mobile phone which concern the game downloaded from the service provider. Collected information is stored in a database and analysed afterwards. The results of the analysis are used to recommend new games to users as well as to enhance the decision making process for purchasing content distribution rights for new games from various content providers [18]. Once the content provider delivers games to the service provider, new recommendation lists are created and games are recommended to users in correspondence with users skills and preferences. If a user likes the recommended game she/he downloads it and plays it on her/his mobile phone. 3

4 4 P. Skocir, L. Marusic, M. Marusic, A. Petric Fig. 1. The games recommendation model 3.2 The MARS Architecture Agents in the MARS system enable trading on two different electronic markets (e-markets): business-to-customer (B2C) and business-to-business (B2B) content e-market. As shown in Fig. 2, there are four types of agents in the MARS system: Service Provider Agent (SPA), Database Agent (DA), Content Provider Agent (CPA) and User Agent (UA). The SPA represents a network operator which acts as a service provider on the content e-market while the DA is in charge of SPA s database. The CPAs represent game publishing companies which publish (and produce) games and act as content providers on the B2B content e-market. On the B2C content e-market, the SPA provides game-based services to users which are represented by their UAs. The MARS system enables automated trading on the B2B and B2C content e-markets. Based on users skills and preferences, the SPA recommends suitable games to its users while, based on the gameplay statistics, it purchases content distribution rights from CPAs for new similar games. Digital games which a user purchased from the service provider as well as the UA are located on her/his mobile phone. The UA monitors and analyses user s interaction with the game and determines the game category and difficulty level which best correspond with her/his skills and preferences. Afterwards, the UA sends a request to the SPA asking for a new game recommendation and waits for a response. Upon receiving a request from the UA, the SPA carries out recommendations based on the game category and difficulty level specified in the UA s request but it also takes into account information specified in the user profile and the profile of the user s mobile phone. The SPA sends a new game recommendation to the UA which displays game information to the user. Additionally, the SPA forwards received information about game consumption (e.g. how often and how long did a user play a game) to the DA for the database update. Periodically, the SPA contacts the CPA and negotiates the purchase of content distribution rights for new games [19].

5 The MARS - A Multi-Agent Recommendation System for Games 5 Fig. 2. The architecture of the MARS system 4 Gameplay Data and User Preferences Analysis for Game Recommendation In this section, parameters used to model user s skills and gameplay experience are described and afterwards the algorithms for processing data collected during user-game interaction are explained in detail. 4.1 The parameters for analysing user s skills According to Komulainen et al. [11], gameplay experience consists of: cognition, motivation, emotion and focused attention. Since cognition is difficult to analyse using existing research techniques and emotions are analysed using questionnaires and measuring psychological state of the users, we observed the remaining two elements. In this paper, we propose models of motivation as well as of focused attention that is represented with user s success and progress in the game, as the parameters used for recommending trivia games. For the estimation of user s success and progress, following parameters are needed: t ij - average time that the user i needed to successfully complete game level j; t j - average time needed to successfully complete game level j for all players i [1, n]; and precision in mastering the game (i.e. was the given solution true or false or was the given task successfully performed). The parameter tij is defined as: p X tij = tijs + tijf, (1) f =1 where tijs represents the playtime of the level j when the user i successfully completed it, while the tijf represents the playtime of the level j when the user i failed to successfully complete the level assuming that p is the total number

6 6 P. Skocir, L. Marusic, M. Marusic, A. Petric of failed attempts. If the user i played the game more than once (e.g., k times) than the average time she/he needed to complete the level j is calculated as: t ij = 1 k k t l ij = 1 k l=1 k p (t l ijs + t l ijf ), (2) l=1 f=1 where t l ij, tl ijs and tl ijf denote playtimes measured during user s lth running of the game. Average time t j that a total of n users which played the game need to complete the level j is calculated as: t j = 1 n n t ij = 1 1 n k i=1 n i=1 l=1 k (t l ijs + p t l ijf ). (3) User s success is compared with the information how successfully other users play that game level. The gameplay metrics of all users are stored in the service provider s database and their average playtimes are calculated using Equations (1), (2) and (3). The UA receives information with those average results from the SPA. If the user i needs significantly less time to complete the game level j than other users do (i.e. t ij t j ), then she/he is considered to be more successful than other users. The progress is calculated from the information about user s earlier success results in that game. If user i s recent results are better than the past ones (i.e. the user i needs less time to complete the game level j than earlier; t l ij < t ij ), then it is considered that the user i progresses. We also model user s motivation which can be defined as a psychological state which reflects user s desire to play a game [9]. The parameters needed for the estimation of user s motivation are: t i - the time that the user i plays a certain game; and t - average time all users spent playing that game. The parameter t i is defined as: m p t i = (t ijs + t ijf ), (4) j=1 where m is the highest level that the user i reached while playing the game. Average time t that a total of n users spent playing a game is calculated as: t = 1 n n t i = 1 1 n i=1 n k i t l i = 1 1 n k i i=1 l=1 f=1 f=1 n k i m l (t l ijs + k i i=1 l=1 j=1 p t l ijf ), (5) where k i denotes how many times has the user i played the game while m l is the highest level that the user i reached during its l th running of the game. If the user i spends more time playing a certain game than the average user does (i.e. t i > t), than we assume she/he is highly motivated and would like to play other games of the same category. 4.2 The recommendation algorithms As shown in Algorithm 1, the UA compares user s current gameplay results, her/his historic results and average results of all users for that game which are f=1

7 The MARS - A Multi-Agent Recommendation System for Games 7 obtained from the SPA. While the user i plays a game, the UA collects the following information about the game: the level user played (i.e. j), the time she/he spent playing the level j (i.e. t ij ), the total time she/he spent playing the game (i.e. t i ) and the result she/he achieved (i.e. variable result in Algorithm 1). Since games are tested and evaluated by experts and regular players several times in different phases of their development [2], content providers (i.e. game publishing companies) provide information about the estimated average time needed to complete game tasks. Before it collects sufficient gameplay data from a certain game that is used to calculate the average time needed to complete a game task (i.e. t j ), the SPA uses the estimates received from the CPA. The UA measures the time (i.e. t ij ) the user i needs to successfully complete a task (i.e. level j) and compares it with the average time of other players (i.e. t j ) in order to determine user s success. Together with the number of times (i.e. p) the user i unsuccessfully attempted to solve a task those data are used to keep track of user s progress in the game. Content providers also estimate the average total time a user will spend playing a game (i.e. t). The UA keeps track how many times the user i played a game (i.e. k) and how long did each of those play sessions last (i.e. t ij ). From that information the UA calculates the time the user i played a game (i.e. t i ) Algorithm 1: Determining the category and difficulty of a new game playedcategory = category of the currently played game; playeddifficulty = difficulty of the currently played game; result = denotes was the user s solution of a task successful; switch result, t ij, t j, t i, t do case result == false && t i > t newcategory = playedcategory; newdifficulty = playeddifficulty 1; case result == false && t i < t newcategory = findanothercategory(); newdifficulty = playeddifficulty; case result == true && t ij < t j && t i > t newcategory = playedcategory; newdifficulty = playeddifficulty; case result == true && t ij < t j && t i < t newcategory = playedcategory; newdifficulty = playeddifficulty + 1; case result == true && t ij > t j && t i > t newcategory = playedcategory; newdifficulty = playeddifficulty 1; case result == true && t ij > t j && t i < t newcategory = findanothercategory(); newdifficulty = playeddifficulty; endsw

8 8 P. Skocir, L. Marusic, M. Marusic, A. Petric and uses it to determine user s motivation by comparing it with the information about average gameplay duration (i.e. t). If the user i plays a game longer than the average time calculated by the content provider (i.e. t), it is assumed that the user is highly motivated and a game from the same category is recommended, otherwise she/he is poorly motivated and a game from another category is recommended. If the user i is more successful than an average player, a more difficult game of the same category is recommended in order to maintain user s flow [20] and improve player retention [22]. To a highly motivated user which demonstrates satisfactory progress, a similar game from the same category and with similar difficulty is recommended. However, if user s progress is satisfactory but the motivation is poor a more difficult game is recommended to avoid boredom since it is assumed that this game is not challenging enough for this user [6]. We assume that the game is to difficult for the user if she/he is motivated but does not show progress in the current game so an easier game is recommended [5]. In case the motivation is also lacking, a game from another category is recommended since it is considered that the user is not interested in similar games. Based on this analysis of user s success, progress and motivation, the UA determines the category and difficulty of a new game that is believed to be suitable for its user. Those parameters are sent to the SPA which, as shown in Algorithm 2, finds the most appropriate games for the given user by taking into account received parameters and user profile (e.g., age group - children, teenagers, students, adults, seniors) and her/his mobile device profile (e.g., operating system, screen size, available memory). If games of desired category and difficulty cannot be found, a random game for user s mobile phone is recommended. Algorithm 2: Selecting a game to be recommended phoneid = findphone (userid); games = findgames (newcategory, newdifficulty, phoneid); if games.size > 0 then games2 = findnotdownloadedgames (games, userid); if games2.size > 0 then return randomgame (games2); else return randomgame (games); end else games3 = findgames (othercategory, newdifficulty, phoneid); games4 = findnotdownloadedgames (games3, userid); return randomgame (games4); end

9 The MARS - A Multi-Agent Recommendation System for Games 9 5 Conclusion and Future Work We presented the Multi-Agent Recommendation System (MARS) for games on mobile phones which enables game recommendation based on user s consumption information, game related information and user preferences (i.e. user profiles and mobile phone profiles). Software agents were chosen as representatives of stakeholders in the system because they can autonomously collect data for their owners, analyse that data and recommend business decisions which will maximize service provider s profit and enhance user experience for existing as well as for new services. We proposed a recommendation algorithm in which decisions are made based on user s motivation, success and in-game progress which were modelled by using the temporal dimension of games. The recommended games match user s skills and preferences and thus maintain user flow experience and improve player retention. The implemented B2C content e-market enables game downloading, collecting and processing of gameplay data and recommending games. For future work we plan to fully integrate the MARS system with the B2B content e-market for trading with content distribution rights [19] and introduce the results of the gameplay data analysis as one of the parameters taken into account in the decision making process used for purchasing content distribution rights [18]. Acknowledgments. The authors acknowledge the support of research project Content Delivery and Mobility of Users and Services in New Generation Networks ( ), funded by the Ministry of Science, Education and Sports of the Republic of Croatia. References 1. Essential Facts about the Computer and Videogame Industry, 2011 Sales, Demographic and Usage Data. Entertainment Software Association, Washington, DC, USA (2011), 2. Bernhaupt, R., Linard, N.: User Experience Evaluation for Multimodal Interaction in Games. In: Raymaekers, C., Coninx, K., Gonzlez-Calleros, J.M. (eds.) Design and Engineering of Game-like Virtual and Multimodal Environments. pp (2010) 3. Brown, E., Cairns, P.: A grounded investigation of game immersion. In: CHI 04 extended abstracts on Human factors in computing systems. pp ACM, New York, NY, USA (2004) 4. Cao, Y., Li, Y.: An intelligent fuzzy-based recommendation system for consumer electronic products. Expert Systems with Applications 33(1), (2007) 5. Chen, J.: Flow in games (and everything else). Communications of the ACM 50, (2007) 6. Cowley, B., Charles, D., Black, M., Hickey, R.: Toward an understanding of flow in video games. Computers in Entertainment 6, 20:1 20:27 (2008) 7. Davidson, J., Liebald, B., Liu, J., Nandy, P., Van Vleet, T., Gargi, U., Gupta, S., He, Y., Lambert, M., Livingston, B., Sampath, D.: The YouTube video recommendation system. In: Proceedings of the fourth ACM conference on Recommender systems. pp ACM, New York, NY, USA (2010)

10 10 P. Skocir, L. Marusic, M. Marusic, A. Petric 8. Ermi, L., Myr, F.: Fundamental components of the gameplay experience: Analysing immersion. In: de Castell, S., Jenson, J. (eds.) Changing Views: Worlds in Play, pp University of Vancouver (2005) 9. Fernandez, A.: Fun experience with digital games :a model proposition. In: Olli, L., Hanna, W., Amyris, F. (eds.) Extending Experiences. Structure, analysis and design of computer game player experience, pp Lapland University Press (2008) 10. Kim, J.H., Gunn, D.V., Schuh, E., Phillips, B., Pagulayan, R.J., Wixon, D.: Tracking real-time user experience (TRUE): a comprehensive instrumentation solution for complex systems. In: Proceedings of the twenty-sixth annual SIGCHI conference on Human factors in computing systems. pp CHI 08, ACM, New York, NY, USA (2008) 11. Komulainen, J., Takatalo, J., Lehtonen, M., Nyman, G.: Psychologically structured approach to user experience in games. In: Proceedings of the 5th Nordic conference on Human-computer interaction: building bridges. pp ACM, New York, NY, USA (2008) 12. Lee, W.P., Liu, C.H., Lu, C.C.: Intelligent agent-based systems for personalized recommendations in internet commerce. Expert Systems with Applications 22(4), (2002) 13. Liang, T.P., Lai, H.J., Ku, Y.C.: Personalized Content Recommendation and User Satisfaction: Theoretical Synthesis and Empirical Findings. Journal of Management Information Systems 23, (2007) 14. Linden, G., Smith, B., York, J.: Amazon.com recommendations: Item-to-item collaborative filtering. IEEE Internet Computing 7, (2003) 15. Mandryk, R.L., Inkpen, K.M., Calvert, T.W.: Using psychophysiological techniques to measure user experience with entertainment technologies. Behaviour & Information Technology 25(2), (2006) 16. Miao, C., Yang, Q., Fang, H., Goh, A.: A cognitive approach for agent-based personalized recommendation. Knowledge-Based Systems 20(4), (2007) 17. Nacke, L., Drachen, A.: Towards a framework of player experience research. In: Proceedings of the Second International Workshop on Evaluating Player Experience in Games at FDG Bordeaux, France (2011) 18. Petric, A., Jezic, G.: Multi-attribute Auction Model for Agent-Based Content Trading in Telecom Markets. In: Setchi, R., Jordanov, I., Howlett, R., Jain, L. (eds.) Knowledge-Based and Intelligent Information and Engineering Systems, Lecture Notes in Computer Science, vol. 6276, pp Springer Berlin / Heidelberg (2010) 19. Petric, A., Jezic, G.: A Multi-Agent System for Game Trading on the B2B Electronic Market. In: OShea, J., Nguyen, N., Crockett, K., Howlett, R., Jain, L. (eds.) Agent and Multi-Agent Systems: Technologies and Applications, Lecture Notes in Computer Science, vol. 6682, pp Springer Berlin / Heidelberg (2011) 20. Sweetser, P., Wyeth, P.: Gameflow: a model for evaluating player enjoyment in games. Computers in Entertainment 3(3), 3 3 (2005) 21. Thurau, C., Drachen, A.: Introducing archetypal analysis for player classification in games. In: 2nd International Workshop on Evaluating Player Experience in Games (epex11) (2011) 22. Weber, B.G., Mateas, M., Jhala, A.: Using data mining to model player experience. In: FDG Workshop on Evaluating Player Experience in Games. ACM, ACM, Bordeaux, France (2011)