=Paper=
{{Paper
|id=Vol-2480/GHItaly19_paper_05
|storemode=property
|title=Towards a Model to meet Players' Preferences in Games
|pdfUrl=https://ceur-ws.org/Vol-2480/GHItaly19_paper_05.pdf
|volume=Vol-2480
|authors=Mattia Bellini
|dblpUrl=https://dblp.org/rec/conf/chitaly/Bellini19
}}
==Towards a Model to meet Players' Preferences in Games
==
https://ceur-ws.org/Vol-2480/GHItaly19_paper_05.pdf
Towards a model to meet players’ preferences in games
Mattia Bellini
University of Milan
Milan, Italy
bellinitia@gmail.com
using a Recurrent Neural Network [6]) and music (e.g.
ABSTRACT
through «a probabilistic model based on distribution
Different have been the attempts to use Procedural Content
estimators conditioned on a recurrent neural network» [1]).
Generation via Machine Learning in game development.
Some of the techniques used in such studies have been
Among the others, some researchers have tried to adapt a
applied in other domains, by knowledge transfer. The
game, or some part of it, to the user playing it. This
target1 domains of these transfers included game
approach has been called “adaptive game design”.
development. More precisely, PCGML has been applied to
Contrarily to what it may seem, apparently the most
game contents rather than games themselves, or, better
interesting findings in this field have been made for drama
told, on level design rather than game design. Researchers
managers, i.e. for the artificial intelligences that
generally tried to automatically generate contents
procedurally generate story flow. The paper takes the move
constituting levels, like maps, while the attempts to
from what seems to be a missing in current literature and it
automatically generate or to adapt entire game
is aimed at proposing and discussing a possible procedural
environments have been less frequent. This is a direct
content generation via machine learning model that takes
reflection of the limits of present ML: the scarcity of
the latest approaches in machine learning applied to drama
available data for full game generation and the difficulty of
managers and combine them with findings from adaptive
creating a model able to generate an entire game from
game design. The objective of the proposed model is to give
scratch.
players the best possible gaming experience of a highly
branched game, depending on their attitudes towards the As Summerville et al. [18] outline in their work, researchers
gaming world. of PCGML have applied different machine learning
approaches in game studies, including artificial neural
networks, Markov models, clustering and matrix
KEYWORDS factorization. Most of the works focused on the autonomous
Video games; Storytelling for video games; Procedural generation of levels, particularly for platformer games like
Content Generation via Machine Learning; PCGML; Super Mario Bros. [11] (e.g. [18], [8]). There had been
Adaptive game design; Drama manager. attempts to generate also contents different from mere
game level, like Magic: The Gathering [24] cards [17] or
stories for interactive fictions [7]. Some of the most
INTRODUCTION interesting approaches in the field applied ML to drama
Procedural Content Generation via Machine Learning managers (DM), to procedurally generate stories following
(PCGML) is a new paradigm for the self-driven creation of players’ behaviour. Similar studies have been made in the
new content. The main difference with the mere procedural field of adaptive game design, e.g. to balance game difficulty
generation is the generally higher quality of the created to player’s abilities. Following the idea of applying ML for
content, achieved by integrating the procedural content a recognition of a player’s attitudes, the focus of this paper
generation (PCG) algorithm with a machine learning (ML) is to discuss whether is possible an application of PCGML
model trained on existing content. for the creation of a player-aware model capable of
PCGML has been applied to a variety of different content predicting user preferences and serving an adapted level
types and, by the time this paper is being written, it progression, to maximize appreciation.
performed well particularly for the creation of images (e.g.
GHItaly19: 3rd Workshop on Games-Human Interaction, September 23rd, 2019,
Padova (Italy)
Copyright © 2019 for this paper by its authors. Use permitted under Creative
Commons License Attribution 4.0 International (CC BY 4.0).
ADAPTIVE GAME DESIGN changes happening in run-time, just like the Mario Kart Wii
Using Procedural Content Generation, some researchers [12] AI example seen above.
tried to adapt games to the player. Traditionally, these
In the survey made by [10], however, an important lack in
works were interested in modifying difficulty settings,
research emerges: the procedural generation of quests has
using different techniques. Examples include the mechanics
been studied only from the point of view of placing goals
of different published games, like the aim assistant in Max
and “keys” to reach them. Even though that is a relatively
Payne [13] that is more precise the less the player’s abilities,
out-to-date survey, being dated back in 2011, this lack
or the opposing AI in Mario Kart Wii [12] that increases its
seems to be still present. Indeed, at the best of my
skills when the player is performing too well. These studies
knowledge, no new impactful studies have been conducted
are of little interest for the purposed of the model, since
in this sense. On the contrary, many have been the works
their aim is to adapt the game mechanics and difficulty but
on improving the storytelling mechanisms on which DM
in no way the game itself, that is the aim of this research.
are based, as we have already mentioned.
More interesting are semantic and declarative approaches,
like the one found in Tutenel et al. [21]. Their model is
based on a semantic definition of objects that includes all APPROACHES IN DRAMA MANAGERS
game-relevant information of a particular game object. There have been several attempts in the field of DM to
These include functional information, possible relations to identify and classify a player in order to give her the story
other objects and metadata of the game content. After progression that best fits her tastes. In the survey made by
having declared and assigned these data to the objects, it is Roberts and Isbell [14] we can find multiple examples, as
possible to use them to better drive the new content also outlined by the more recent study conducted in [22].
creation process. Giving objects a semantic layer «helps For analytical purposes, the same four features presented in
convey the meaning and the role of an object in the virtual this lastly mentioned paper will be used to describe models
world, and consists of generic descriptions of classes of found in literature. The four features are: replayability
features, including attributes, properties, roles, relations, (possibility to play again the game without receiving the
etc. This encourages the incorporation of further semantic same gaming experience), authorial control (control over the
information about player-dependent gameplay purposes, game design process left in the hands of the author), player
and how these can be used to control object generation» autonomy (freeness of the player in the gaming experience)
[10]. and adaptability (capability of modifying the game to meet
player’s tastes). Later in the paper, we will also address the
[10] points out also that a model that aims at providing a
problem presented by a fifth feature, namely the
better game experience to the players by adapting the game
coordination, i.e. the ability to orchestrate Non-Player
to their play-time behaviour, needs a firm knowledge over
Characters (NPCs) and other game elements to present
what a player expects to play, to feel and in general to find
specific experiences to the player.
in a product. This means, basically, three needings:
Researches on DM present the closest approaches towards
1. Have a solid player model and a way to capture
the model that this paper is aimed to propose. In particular,
player’s expectation;
the approach of the PaSSAGE system [20] provides a good
2. Quantify the expectation to a measurable level;
degree of adaptability through the identification of pre-
3. Process them and adapt the game consequently.
defined players’ styles. By assigning at each event of the
These three steps are of essential need for the creation of a game a weight for each style, the model chooses the most
model really capable of adapting contents to the player. attractive event at every stage, ensuring autonomy to the
Charles et al. [5] supports the idea of shaping a player player. To be noted is that this approach is deterministic, i.e.
model to capture her interests and playstyle, but also points there is no degree of randomness. Thus, replayability is
a fourth need: the necessity of monitoring the player to virtually zero, since the same actions will always result in
constantly check for the effectiveness of a generated the same reactions in the interactive fiction. This is also the
solution. reason for the high degree of authorial control of this
model.
A last useful distinction made by [10] is the one between
off-line customized generation and on-line adaptivity: the Very interesting is also the approach that emerges in
former is intended as a generation of contents while the Implementation and Analysis of a Non-Deterministic Drama
game is not running, typically during the loading of a Manager [22]. The aim of the authors here is to serve the
gaming session; the latter, on the contrary, describes the best possible match between emergent player attitude
towards the story and story progression, to provide a high way to maximize player’s appreciation of a game by
degree of adaptability. The job is entrusted to a DM built on adapting it at the levels of game design, story development
a genetic algorithm that ensure non-deterministic results, and «the logical flow of events and actions that follow» [10].
thus replayability, and the possibility to deal with a large However, Yu and Riedl’s [26] decision to not rely on a pre-
number of blocks (here quests). To match player’s defined PM presents us a huge knowledge gap between what
preferences, the developers rely on a player model (PM) can be designed and what really players want.
composed by a vector of three dimensions, representing the Notwithstanding the advantages of having self-built patterns
attitudes of the player towards the three regiments that do not rely on any abstract theory, this approach is not
theorized by Durand in The anthropological structures of the returning any clearly readable data on user preferences but
imaginary. Quests are human-authored and described using just, indeed, opaque patterns. For this reason, it might be
similar vectors. This way, it becomes possible a quite more profitable to rely on a well-established PM. For
computation of the concatenation, to evaluate the distance the purpose of this model, I decided to base the PM on
of the results of each possible sequence with player’s Stewart’s theory[16], grounded in turn on Bartle’s
preferences. After a probabilistic tournament selection, the psychographic taxonomies [1]. Other theories might have
tournament-winner concatenation is manipulated to been used, among the others Yee’s model [25], or Bartle’s
minimalize the possibility of two identical results given the three dimensions model [2]. I decided to discard Yee’s
same premises. The model grants player autonomy by because, with Bartle’s word [3], «if you want a theory for
setting a minimum number of choices available to the […] studying player psychology, then you may be better
player at each stage of the interactive fiction. served by a straight taxonomy […] such as Nick Yee’s
The model presented in [22] is an extremely useful starting motivations». It should be clear that my intent is not to
point for the present theory. However, in the field of DM, study the psychology of the players, but rather their
the most interesting approach is the one showed in preferred game style. On the other hand, Bartle’s three
Personalized Interactive Narratives via Sequential dimensions model [2] presents a too-broad categorization,
Recommendation of Plot Points [26]. In the paper, the authors that becomes nearly impossible to handle in the
present a collaborative filtering approach, similar to the development phase. Another possibility is to develop a
ones used in recommendation systems for services like custom categorization of players, but this is going further
Netflix and YouTube, applied to DM for the autonomous beyond the scope of the paper. I decided to base the model
building of a story. Thanks to the collaborative filtering on Stewart’s expansion of Bartle’s taxonomies mainly for
algorithm, the developers handed-off the complex problem two reasons: firstly, it provides a clear and simple
of defining PMs to determine current users, as the different categorization of players that is well-grounded in literature
categories of users were grouped by the algorithm itself. An and which validity has been tested different times.
interesting advancement of this model is that it Secondly, describing the PM with four values guarantees an
demonstrates how a progression-aware model has amount of information that is easily processable both by the
impactful benefits in the recommendation of subsequent algorithm and, more importantly, by the developers.
plot nodes. On the other hand, a limit of the approach is that However, the proposed model is not dependent on the way
it is largely based on manual and explicit expression of of describing the players and it is possible and easy to
positive and negative feedbacks via a review system, due to change the PM description if a better method is found.
the non-pre-defined PM. The authors also implicitly Stewart in [16] describes the four taxonomies (Socializer,
pointed out a good practice to retain authorial control over Killer, Explorer and Achiever) by binding them with
story generation, i.e. the use of a branched scheme as a specific actions performable in a generic game. On the basis
starting point from which to pick the blocks to be appended of this theory, a player of the hypothetical single-player
at each stage. game based on the proposed model will be described by soft
clusters, «in the sense that a particular player can have a
degree of membership in each player type» [26], i.e. she will
THE PROPOSED MODEL be characterized by a vector of four values ranging from 0
The approach presented in [26] points out an extremely to 1, that quantify her degree of membership to each of the
promising case of knowledge transfer: they demonstrated four taxonomies. The vector will then describe the gaming
that sequential recommendation, frequently used to suggest persona of the current player in a pretty accurate way.
complete fictional artefacts, is eligible to be applied also to
shape only parts of a product, to best fit the tastes of
audience in almost real-time. This promises to be a smart
The proposed model relies therefore on PM recognition to
procedurally generate content. The content is customized
on player’s profile and based on a PM built during an online
(in-game) opaque survey: while normally playing, the
algorithm registers player’s path, formed by each action she
chooses to perform, and updates in run-time her profile
accordingly. However, the proposed procedural generation
is offline (pre-game), meaning that the algorithm will select
a block to present to the player as next level during the
loading screen (i.e. in the time moving from one level to the
subsequent) and not during an active game session, mainly
to avoid slow-downs.
To decide which block to select, it will be used an approach
similar to the one that can be found in [22]. The model will
Figure 1 – Visual description of Stewart’s expansion of have a pool of developers-defined levels at each stage, in
Bartle’s taxonomies [16] which are encapsulated n number of actions, having each a
Using such a model would surely require some care during vector of traits corresponding to Bartle’s taxonomies. A
level design, particularly for early-in-the-game levels. What fitness function evaluates the levels in the pool by
would be needed are multi-solution problems to overcome. calculating the distance between available actions vectors
This might include both enemies and riddle to solve, and player’s profile vector and return for each level a fitness
obstacles, pathfinding moments, etc. The first step than is rate. The fitness rate is then downscaled to a percentage,
to assign a value for each of the taxonomies to the different that is in turn used to probabilistically pick the level to show
possible actions, i.e. to the different ways of progressing in to the player as game progression. It might be more efficient
the events. Later, we can easily obtain a well-defined to evaluate each block in advance, attaching a vector of
gaming persona of the player by adding those values to the properties to the levels and not to the single actions, in
player’s profile. Indeed, to evaluate her disposition towards order to accelerate the concatenation process. This way, the
a taxonomy it will suffice to register the actions she algorithm would only have to evaluate levels as a whole,
performs and their semantic description, made using instead of each action separately. However, unfortunately,
Stewart’s expanded taxonomies. For example: a player that this approach probably presents its drawbacks, too: even if
during the game tends to speak with all the NPCs and to it is true that the algorithm could work faster, presenting
solve problems in a “diplomatic way” might be labelled as a the next level a few milliseconds in advance, on the other
Socializer, while a player who tends to attack whatever is hand the blocks may need to be specifically designed to
in sight might be labelled as a Killer. Again, this is an please a particular part of the audience. This might lead, in
example: as said, the gaming persona are not defined as a set turn, to a regression to an almost deterministic model,
of mutually exclusive booleans, but rather as a set of values especially after the PM will have reached a good level of
floating between 0 and 1. Indeed, a much more realistic precision. In addition, such an approach could mean more
representation of the player would be formed using “floats restricted possibilities in designing actions due to the
approach”. specificity of the preferences of the audience for that block.
This could make level blocks even more specific. In turn, it
Furthermore, to better represent a player’s attitude, it will
could lead to an even more specifically designed level,
be needed to weight different actions in a reasonable way.
ending up with a system that, due to a “wrong” choice of
An accurate weighting is necessary to not overbalance an
the player in the early stage, could keep her in an unwanted
action regarding the others and thus to obtain a valid
path for the entire game. On the other hand, by evaluating
gaming persona of the player. Taking back the previous
each single action the designers can also include different
example: if actions are not differently weighted, a player
taxonomies in each level. This way it will be possible to not
that kills an evil slaver would end up with the same “Killer
present the player with actions belonging only to her
rate” as one who murders an innocent just for fun. In
preferred taxonomy, in order not to bore her with too
addition, a single action might have a (positive or negative)
similar tasks. To obtain the same result, if needed, it will
weight in two or more taxonomies, thus actions, too, need
also be possible to include a random factor during the pick
to be described as vectors of weights.
of the blocks.
Picking from a set of author-made blocks, the algorithm will scalability, as it is possible to add or remove levels
choose the one that would probabilistically provide the best at each point of the game without impacting on the
possible experience for the player. Thus, for our previously game progression, since the algorithm picks the
instantiated player booleanly labelled as Socializer, the best-fitting block in the provided pool. This being
algorithm will more likely choose to concatenate levels with said, a nota bene is that the model does not evaluate
the most “problems” solvable via socializing, while for the the coherence of the game progression, that has to
Killer are more likely to be chosen the levels with the be addressed during the game design phase. This
greatest number of possible enemies, and so on. approach is scalable also in the sense that it is
possible to modify the PM description to best fit
The advantages of the proposed approach are multiple and
the needing of each game built on the model;
can be summarized as follows:
▪ Single-person collaborative filtering: taking the
▪ Authorial control: using a defined pool of possible example of the collaborative filtering approach
levels at each stage ensures a high degree of found in [26], the proposed model will rely on a
authorial control over the result. With this “single-person collaborative filtering”. The model
approach, each level is entirely created by authors: is based on a prediction of likeliness built on a
the PCG applies only on the concatenation. It series of positive and “non-positive” feedbacks.
generates the game, but not single levels. Authors The feedbacks are given by player’s choices of the
can design the game and its story with a normal actions to perform: the chosen action is a positive
branching tree, just as [26] suggested; feedback, while all the other discarded possibilities
▪ Player autonomy: the player keeps the autonomy are “non-positive” ones;
she has in a general game, since there is no ▪ Data scarcity: the main issue of the approach found
autonomy retention in the model itself. in [26] is the heavy reliance on human-provided
Constraints might be decided in the phase of actual data quantity. For the model proposed in this
development of game and levels; paper, data scarcity is not an issue, mainly for two
▪ Adaptability: the whole model is intended to have reasons: 1) the concatenation can be delayed until
a good level of adaptability, given the constraints a certain amount of data over the player are
of human capabilities to create levels. The collected, and 2) is virtually possible to design
proposed approach is not a PCGML model aimed extremely dense levels that would give a relatively
at the autonomous generation of an uncountable huge amount of information.
number of games, or levels, of stories, but rather it
This model finds its place between level design and game
is aimed at the maximization of the appreciation of
design. Taking player’s choices in levels as inputs, and
a widely branched game. In addition, thanks to the
outputting game design options through a recommendation
evaluation of the single next step, players can
system, might be a play-changing approach in PCGML
change their attitude towards the game and its
applied to games. However, this approach shows us a
fictional world without being constraint in a
challenge: on the one hand we would end up with a game
narrow path, pre-determined by her early-in-the-
that has the highest possible appreciation rate, due to the
game choices;
very fact that the game itself is shaped on the individual
▪ Replayability: due to the probabilistic
player attitudes. On the other hand, however, designing the
concatenation of levels, the model keeps a medium
proceeding of such a game requires particular care, above
level of replayability, since the concatenation is
all for story progression. This is the main weak point of the
not deterministic but, indeed, probabilistic;
model: it needs an expertise in storytelling and game design
▪ Coordination: coordination in the model is
to keep the consistency of the story. To address this
addressed incidentally, since there is no direct
problem is probably preferable to keep levels relatively
control of the model on the behaviour of NPCs and
small-scaled: keeping in mind the Aristotelian unities of
other game elements. The coordination arises here
time, place and action when designing levels might be a
from the very fact that depending on the actions of
good practice when this model is applied. Of course,
the player, the concatenated levels will ideally be
constraints to the level concatenation can be applied in
built to present a reaction of the environment to
order to prevent a certain level being shown after another
player’s actions;
one that has nothing to do with the previous story. This
▪ Scalability: the evaluation of vectors of the
does not mean, obviously, that the player cannot occur in
individual actions found in a level ensures
major crossroads in the story.
FINAL CONSIDERATIONS AND FUTURE STUDIES [3] R. A. Bartle. 2009. Understanding the Limits of Theory in
Beyond Game Design: Nine Steps to Creating Better Videogames.
The current paper was aimed at proposing a PCGML model Chris Bateman. Delmar.
capable of adapting a game to players’ attitudes and [4] N. Boulanger-Lewandowski, Y. Bengio, and P. Vincent. 2012.
preferences. The starting point has been different models Modeling temporal dependencies in high-dimensional
sequences: Application to polyphonic music generation and
aimed at an adaptive game design. Notwithstanding some transcription, arXiv preprint arXiv:1206.6392.
extremely useful best practices pointed out by such [5] D. Charles, A. Kerr, M. McNeill,M.McAlister, M. Black, J.
Kücklich, A. Moore, and K. Stringer. 2005. Player-centred game
researches, no models actually similar to the one I was
design: Player modelling and adaptive digital games, in Proc.
aiming to obtain have been found in this field. Instead, DiGRA Conf. Changing Views—Worlds in Play, Vancouver, BC,
researches on DM seemed to be oriented more towards the Canada, pp. 285–298.
[6] K. Gregor, I. Danihelka, A. Graves, D. J. Rezende, and D.
direction my model was aimed at facing. Indeed, approaches Wierstra. 2015. Draw: A recurrent neural network for image
that relied on a non-deterministic blocks concatenation [22] generation arXiv preprint arXiv:1502.04623.
and on a collaborative filtering recommendation system [7] M. Guzdial, B. Harrison, B. Li, and M. O. Riedl. 2015.
Crowdsourcing open interactive narrative in the 10th
[26] have been extremely useful for the theorization of the International Conference on the Foundations of Digital Games,
proposed model. This is based on the definition of a PM as Pacific Grove, CA.
[8] R. Jain, A. Isaksen, C. Holmgård, and J. Togelius. 2016.
a vector of four dimensions: each time the player performs
Autoencoders for Level Generation, Repair, and Recognition in
an action, the PM is updated accordingly. At each new level, ICCC.
the algorithm probabilistically picks a subsequent block [9] B. Li. 2015. Learning knowledge to support domain-
independent narrative intelligence, Ph.D. Dissertation, Georgia
that is more likely to be interesting for the player, according Institute of Technology.
to her current PM. A particular care is needed during the [10] R. Lopes, and R. Bidarra. 2011. Adaptivity Challenges in Games
storytelling and game design phases, but every game that is and Simulations: A Survey in IEEE transactions on
computational intelligence and AI in games, vol. 3, no. 2.
not linear-paced needs expertise in game design and [11] Miyamoto, T. Tezuka. 1985. Super Mario Bros.
storytelling and the little additional care needed here is a [12] Nintendo EAD. 2008. Mario Kart Wii.
[13] Remedy Entertainment. 2001. Max Payne, Gathering of
little mite to be paid for what the model promises to do.
Developers.
[14] D. L. Roberts, and C. L. Isbell. 2008. A survey and qualitative
Further progress of the research will be, first of all, the analysis of recent advances in drama management. in
development of the proposed model and its implementation Transactions on Systems Science and Applications, no. 3.
in a game. However, there are also different other [15] M. Sharma, S. Ontañon, M. Mehta, and A. Ram. 2010. Drama
Management and Player Modeling for Interactive Fiction
advancements that might be needed to obtain a completely Games in Computational Intelligence Journal, Volume 26 Issue 2
valid result. Among the others, a better way of drawing a [16] B. Stewart. 2011. Personality and Play Styles: A Unified Model,
PM might be found. Indeed, as mentioned, the model is not Retrieved June 03, 2019 from:
https://www.gamasutra.com/view/feature/134842/personality_
dependent on the way of describing the players and it is and_play_styles_a_.php
always possible to modify the way the PM is calculated, [17] A. Summerville, and M. Mateas. 2016. Mystical tutor: A magic:
The gathering design assistant via denoising sequence-
described and stored. In addition, from the point of view of tosequence learning.
a storyteller, it might be very useful to conduct a proper [18] A. Summerville, and M. Mateas. 2016. Super Mario as a string:
research to analyse the new paradigm for addressing Platformer level generation via LSTMs, in Proceedings of 1st
International Joint Conference of DiGRA and FDG.
interactivity that emerges from the application of the [19] A. Summerville, S. Snodgrass, M. Guzdial, C. Holmgård, A.
proposed model. Also, it might be interesting to examine Hoover, A. Isaksen, A. Nealen, and J. Togelius. 2017. Procedural
whether my approach presents restrictions in the stories or Content Generation via Machine Learning (PCGML) in IEEE
Transactions on Games.
in the mechanics of a game based on it. Lastly, contrarily to [20] D. Thue, V. Bulitko, M. Spetch, and E. Wasylishen. 2007.
what might be found in many researches on the field, I Interactive storytelling: A player modeling approach in Proc. of
the 3rd AIIDE07.
strongly believe that a PCGML model able to improve [21] T. Tutenel, R. Bidarra, R. M. Smelik, and K. J. de Kraker. 2008.
human design - rather than substituting it - can help the The role of semantics in games and simulations, in Comput.
improvement of such approaches both in literature and in Entertain., vol. 6, no. 4, pp. 1–35.
[22] M. N. R. Utsch, G. L. Pappa, L. Chaimowicz. 2018.
the industry. As of little help as it might be, I hereby Implementation and Analysis of a Non-Deterministic Drama
encourage any studies aimed at this purpose. Manager in Proceedings of SBGames.
[23] J. Valls-Vargas, S. Ontañon, J. Zhu. 2013. Towards Story-Based
Content Generation: From Plot-Points to Maps, IEEE Conference
on Computational Inteligence in Games (CIG)
REFERENCES [24] W. of the Coast. 1993. Magic: The Gathering.
[1] R. A. Bartle. 1996. Hearts, clubs, diamonds, spades: Players who [25] N. Yee. 2006. Motivations for play in online games in
suit MUDs, in Journal of MUD research, Colchester. CyberPsychology & Behavior, 9, 6, pp. 772-775.
[2] R. A. Bartle. 2005. Virtual Worlds: Why people play in Massively [26] H. Yu, and M. O. Riedl. 2014. Personalized Interactive Narratives
Multiplayer Gamer Development, 2, 1. via Sequential Recommendation of Plot Points in IEEE
Transactions on Computational Intelligence and Artificial
Intelligence in Games.