Session: Paper Presentation CHI PLAY 2018, October 28–31, 2018, Melbourne, VIC, Australia Towards Deep Player Behavior Models in MMORPGs Johannes Pfau Jan David Smeddinck Rainer Malaka Digital Media Lab, TZI Open Lab, School of Comp. Digital Media Lab, TZI University of Bremen Newcastle University University of Bremen Bremen, Germany Newcastle upon Tyne, UK Bremen, Germany [email protected] [email protected] [email protected] ABSTRACT issues with large-scale systems, and cheating or other unethical Due to a steady increase in popularity, player demands for behavior. We approach the closing of multiple unsolved gaps video game content are growing to an extent at which consis- in these areas of concern for game research and development tency and novelty in challenges are hard to attain. Problems in based on an uncommon building block: deep player behav- balancing and error-coping accumulate. To tackle these chal- ior modeling (DPBM). We discuss the potential of DPBM lenges, we introduce deep player behavior models, applying with regard to the challenges indicated above. To establish machine learning techniques to individual, atomic decision- apt representation techniques we also explore the potential of making strategies. We discuss their potential application in different machine learning techniques for player modeling in personalized challenges, autonomous game testing, human massively multiplayer online role-playing games (MMORPGs) agent substitution, and online crime detection. Results from and implement a pilot study which provides a first data set a pilot study that was carried out with the massively multi- and enables the comparison between selected models. We player online role-playing game Lineage II depict a bench- hypothesize that different advantages can be attained from mark between hidden markov models, decision trees, and deep Hidden Markov Models (HMMs), decision trees (DTs) and learning. Data analysis and individual reports indicate that deep learning (DL), in terms of analytic capabilities and pre- deep learning can be employed to provide adequate models of diction power. After outlining the concept of employing user individual player behavior with high accuracy for predicting modeling with machine learning for approaching challenges skill-use and a high correlation in recreating strategies from in game research and development, we present the selected previously recorded data. techniques and illustrate the results of an exploratory study that was carried out with the established MMORPG Lineage II. CCS Concepts Deep learning appears most adequate in terms of prediction ac- •Information systems ! Massively multiplayer online curacy and behavior representation similarity, whereas HMMs games; •Human-centered computing ! User models; and DTs offer useful visualization and analysis features. These •Computing methodologies ! Machine learning ap- models also constitute the basis for ongoing subsequent future proaches; work focusing on the evaluation of user experience in different game modes. Author Keywords Through this paper, we contribute a general discussion of po- Neural networks; deep learning; HMM; decision trees; games; tential application fields of DPBM and machine learning in player modeling; personalization; game testing; adaptive the context of digital games, provide an overview of the state agents; dynamic difficulty adjustment of the art in research and industry, and point out distinct advan- tages of player behavior models. Additionally, our exploratory INTRODUCTION study exemplifies an early working utilization, highlighting Video game production and maintenance, especially for flag- advantages and disadvantages of the different models. ship productions, is reaching the limits even of what large companies can sustain. Following the demand of players, BACKGROUND games grow more complex in terms of content and mechanics, In related work, player behavior modeling has been ap- where the action spaces become nearly endless, greatly in- proached mostly with the goal of facilitating dynamic difficulty creasing the number of things that could potentially go wrong. adjustments (DDA) [1, 8, 20, 43, 50, 58]. Further application This includes players facing unbalanced challenges, software areas that have been discussed are the modeling of behavior execution or gameplay bugs that go undetected, connectivity impressions for informing game development [6,9,26] and the reproduction of atomic tasks [12, 47]). After briefly highlight- © thePermission authors, to make 2020. digital This or hard is copiesthe author’s of all or part version of this work of forthe personal work. or ing MMORPGs as an especially fitting class of games for the It classroomis posted use ishere granted for without your feepersonal provided thatuse. copies Not arefor not maderedistribution. or distributed application of DPBM with machine learning in the next sec- Thefor definitive profit or commercial version advantage was published and that copies as: bear this notice and the full citation Pfau,on the J., first Smeddinck, page. Copyrights J. forD., components & Malaka, of thisR. work(2018). owned Towards by others Deep than thePlayer tion, the subsequent sections introduce general categories that Behaviorauthor(s) Models must be honored.in MMORPGs. Abstracting CHI with PLAY credit ’18 is permitted.Proceedings To copy of otherwise, the 2018 or encompass the application areas of these isolated reports and Annualrepublish, Symposium to post on serverson Computer-Human or to redistribute toInteraction lists, requires priorin specificPlay, permission381–392. https://doi.org/10.1145/3242671.3242706and/or a fee. Request permissions from [email protected]. discuss the potential of individual behavior models for tack- CHI PLAY ’18, October 28âA¸S31,2018,˘ Melbourne, VIC, Australia ling common challenges in games research and development Âl’ 2018 Copyright is held by the owner/author(s). Publication rights licensed to ACM. ACM 978-1-4503-5624-4/18/10...$15.00 A-D. https://doi.org/10.1145/3242671.3242706 1 381 Session: Paper Presentation CHI PLAY 2018, October 28–31, 2018, Melbourne, VIC, Australia The Case for Player Modeling in MMORPGs human player. Given such abilities, challenging “oneself” can MMORPGs typically encompass several aspects that highlight present continuous and powerful DDA, since players have to the potential benefits of the application of player modeling. adapt and overcome their own behavior in order to be less The vast amount of data that each individual player is gen- predictable. In this light, genuinely balanced challenges can erating constantly along with the immediate opportunity to be provided to players on a generative basis. Moreover, if the compare it to the global behavior data of a big community skill level of a player can be quantitatively assessed by means provides a rich basis for powerful but data-demanding ma- of player modeling, novel PvE modes are conceivable that chine learning algorithms. Moreover, the player behavior in confront the player with ever-changing enemies originating MMORPGs can most often be broken down into movement from a potentially large set of human players with adequate behavior and action selection (skill usage). Each skill in this skill levels. context is unique and discrete, which allows for less noisy be- B. Autonomous Game Testing havioral state categorization compared to other popular video Automatic simulations of video game play have proven to game genres, such as first-person shooters (FPS) or real-time be usable in situations where human testing is too tedious or strategy (RTS) games, where behavioral data quickly gets noisy [53]). Also, the continuously required internet connec- not exhaustive enough for the purpose of finding bugs and tion simplifies centralizing and outsourcing the computational glitches [4, 15, 38, 45], parameter tuning [60], and assuring effort (e.g. through cloud computing). In return, player behav- solvability [42]. Based on the insights and the potential of our ior models can enrich MMORPGs in various ways, such as previous work on a tool for completing and debugging adven- increasing novelty and the prevalence of interesting challenges ture games [35]), we want to further extend the possibilities of (see A), human tester relief and predicting the effects of game autonomous game testing. changes (see B), preventing game or match disruptions (see For developers, one of the most difficult and time-consuming C), and preventing unfair or unethical player behavior which phases of the game design process [22] is the balancing of dif- can not only harm the player community but can also cause ferent character classes. Following the definition of Sirlin [44], financial losses for players and companies (see D). These as- a multiplayer game is “balanced if a reasonably large number pects are further discussed in the following sections to clarify of options available to the player are viable” (where viabil- how these potential benefits of DPBM can come to play in ity sets the requirement of having many meaningful choices (massively) online multiplayer games. throughout a game), while “players of equal skill should have an equal chance at winning” A. Personalized Challenges . Together with frequently desired asymmetrical character configuration possibilities this inher- Due to the complexity and scale of the game environments ently leads to combinatorial explosions, which can become and interactions it can be difficult to consistently present mo- tivating, well-perceived and evenly balanced