Data is becoming an important central point for making design decisions for most software. Game development is not an exception. As data-driven methods and systems start to populate these environments, a good question is: can we make models developed from this data transparent to users? In this paper, we synthesize existing work from the Intelligent User Interface and Learning Science research communities, where they started to investigate the potential of making such data and models available to users. We then present a new area exploring this question, which we call Open Player Modeling, as an emerging research area. We define the design space of Open Player Models and present exciting open problems that the games research community can explore. We conclude the paper with a case study and discuss the potential value of this approach.
As data science and Machine Learning (ML) become
increasingly adopted in a wide range of everyday digital
products, there is a rapidly growing demand to make the
underlying algorithmic processes and models more transparent. As a
response, new research areas such as eXplainable AI (XAI)
and Interactive Machine Learning (IML) have emerged to
investigate how to make data science and ML models less
opaque and more interpretable
With a few exceptions
Copyright © 2021 for this paper by its authors. Use permitted
under Creative Commons License Attribution 4.0 International
(CC BY 4.0).
by different human stakeholders. In the game development
process, analysts use player models to gain further insights
into different player types and improve the gameplay
This paper proposes a new research area — Open Player Modeling(OPM) — to make computational models of players accessible and more transparent to players themselves. This involves answering questions such as WHAT data or models would be useful for players, HOW would such data be communicated, and WHEN to show these models or the resulting processes, such as recommendations or predictions, to players in the gameplay experience. Our core position is that open player modeling is a promising research area that can potentially empower players to learn from their own data, increase their trust in AI, and enhance the effectiveness of some games as interventions. In e-Sports, OPM can be used to enhance the experience of spectators and to help players learn from other players’ gameplay strategies. In this paper, we illustrate the usefulness of OPMs through a case study showing how an OPM can be applied to an educational game on parallel programming. We argue that OPM can increase players’ reflection on their learning strategy and self-regulate learning more effectively.
The contribution of this paper is the following. 1) We identify Open Player Modeling as a promising new research area by synthesizing related research in separate communities: games, intelligent user interfaces, and intelligent tutoring systems. 2) We provide a theoretical framework to map out the design/problem space of open player modeling and identify key open problems. 3) We offer a case study of how OPM can be used to enhance players’ experience and capabilities.
Constructing computational models of how users interact
with digital systems is the subject of user modeling
research
Most existing research efforts on showing users
computational models of their behavior concentrate in the Intelligent
User Interfaces (IUI) community. Most notably, Open User
Modeling (OUM)
Open Player Models (OPM) extend the main idea behind OUM/OLM to computer games. The majority of existing work on OUM/OLM is based on adaptive web applications where hyperlinks are the primary source of user interaction. As argued below, player interactions within games tend to be more complex than with web applications. Thus, OPMs face unique challenges in terms of how to open the models and how to make the resulting information useful to the players.
Traditional player models, a component of the AI systems
in games, are used directly by the game system to, for
example, adapt the gameplay experience. They are typically
not designed to be open to human users. As game user
research develops, post-hoc player models start to be
developed and are visualized for user researchers, game analysts,
game business teams, and game developers. Such models are
then used in the game production pipeline for various
purposes, including enhancing the game design or tuning the
technical aspects of the game
The most developed applications of OPMs for
nontechnical stakeholders are dashboards of player
performance. Dashboards are broadly used in modern computer
games to display player performance stats, such as win rate,
quality of solutions in puzzle games, and average speed in
racing games. More recently, with the increased
popularity of e-Sports, more sophisticated dashboards have been
developed. Such dashboards have also been extended for
spectators to show what is happening within the match
On the research side, several works investigated
dashboards for various games
We propose the new term Open Player Modeling (OPM) to unify the range of practices of showing players computational representations of their gameplay data. These models can vary in their computational complexity — ranging from players’ in-game performance through descriptive statistics (e.g., average speed) to cognitive, affective, or behavioral models of player characteristics (e.g., learning preferences in educational games). Figure 1 shows the key components of OPM. First, player’s gameplay log data is processed to produce a player model. In some cases, human experts are involved to provide labeling and domain knowledge to the modeling process. Next, these models are explained and communicated to the players and other stakeholders (e.g., game analysts) through text, visualization, or other media. In cases of complex player models with low interpretability, such as deep neural network-based player models, extra effort is needed in producing explanations to the underlying models before they can be communicated to the players.
As shown above, there are pockets of knowledge about open models in different research communities. So is OPM simply an extension of the prior work to a new domain, or is it fundamentally different from them? Below, we argue that there are four reasons why OPM is the latter.
First, player data is typically more complex than the type of user data applied to OLM and OUM research. As summarized above, most existing OLM and OUM research are conducted based on how users interact with hypermedia applications. In comparison, players interact with games with higher frequency and longer duration. A 2020 international report1 shows that although gamers typically play for an average of 1.37 hours at a time, many of them play for longer sessions. In the U.S. players’ average longest consecutive play sessions are 5.10 hours. All the above statistics indicate the size and complexity of player data.
Second, player modeling techniques are different from
those used in other open models domains. Because play
experience often involves time and virtual space, most player
models involve a way to deal with the time and space
continuum. Some models aggregate over these dimensions
Third, most player models will integrate some way of
measuring or integrating preference, emotions, and
engagement (some examples of these models have been discussed
in Yanakakis et al.’s paper and book (2013; 2018)). These
factors are important for personalizing the play experience
and also for developing game content. Due to the influence
of game characters on team performance within a game like
DoTA or League of Legends, such aspects will also need to
be modeled (e.g.,
online-gaming-2019/n#spend, last accessed on January 31, 2021. recommendation systems, including in Senpai.gg and Mobalytics.gg, which recommend items, characters, or other aspects of gameplay that are based on player state modeling.
Fourth, how players interact with OPM may be different from other open models, imposing new research questions for open models in the context of play. Players’ desire to improve their in-game performance and win the game may provide them with the extra motivation to use OPM. Compared to users of other open models, gamers are more familiar with viewing information about their activities. Many games already use player stats to communicate how well each player is doing. In some complex games, players have to make sense of a wide range of stats and use them to inform gameplay decisions. Having this background means players may be more prepared to make use of more sophisticated open models in ways they may not be outside games. It thus makes OPM an exceptionally suitable domain to investigate eXplainable AI problems of communicating complex AI decisions to end-users.
This section maps the problem space of OPM along two main dimensions: the nature of the player models and their level of openness.
I. Types of Player Models There are different ways that different models in the literature have been categorized. For example, Yanakakis et al. (2013) categorized computational player models in terms of their computational structure defining them as model-based techniques (techniques that embed a theoretical or expert model that is used to derive the player model), model-free techniques (those that construct a model based on data without a specific knowledge or theory), and hybrids that mix model-based and model-free techniques. Another approach to categorize models is the approach borrowed from statistics or engineering, where models are categorized into: (a) descriptive, where data is used to show or describe what has happened, this can be done at different levels of abstraction, (b) predictive, where data is used to predict an outcome or a state, (c) diagnostic, which extends a descriptive model to specifically answer specific questions, and (d) prescriptive extends predictive models to prescribe what to do next.
While these frameworks are useful conceptual tools from
the algorithmic perspective, they contain information that
players may find unnecessarily complicated. Since OPM is
used by players, we propose an adapted framework to
classify player models from the players’ perspective.
1. Descriptive Models tend to describe a system or a
phenomenon, constraints, complexity, and interaction. An
example is a descriptive model that contains a player’s
problem-solving process using the gameplay sequence of
how she solves a puzzle in a level (e.g.,
An alternative way to categorize player models is based
on the type of service they offer. For example, for reminders,
these models tend to be context-aware and use expert
knowledge to show specific information to players based on their
play and the data analyzed. A good example of this model
can be seen in the Senpai.gg tool, where players can use the
reminders to improve their gameplay. Another service type
is recommendations. These models offer recommendations
for items, characters, or specific elements to player pre- or
during play
It communicates to players the result of the player
models, without detailed information of how the models were
calculated from the players’ gameplay behavior. For
example, an OPM of an adaptive game similar to Left4Dead
may communicate to the player that its model of her is
that she is a highly-skilled player, and therefore the game
is sending her more difficult enemies. Notice this type of
OPM does not contain information about how the user
model is computed. It focuses on what the model is. The
main benefit of this type of OPMs is its simplicity.
Especially when players are engaging in time-critical
gameplay, such as combat, extra information of the model can
be overwhelming.
2. Open Model Process reveals the computational process
through which the outcome of player models are
calculated from the input of player behavior. For instance, this
type of OPM can use visualization to display the
clusters of players based on their gameplay and show a player
where her gameplay is situated. It can also be used to
explain certain predictions that the game AI has made (e.g.,
probability of winning a game). The main benefit of this
type of OPMs is that they can help players understand
WHY and HOW they are modeled in a particular way.
3. Editable Open Models affords the highest level of
openness. In addition to making the player model open to the
players, they allow players to modify their own models.
When designed well, editable models can improve the
accuracy of the models by allowing players to make
corrections. For instance, editable OPMs can be used in adaptive
games for physical rehabilitation patients. The game can
model a player’s joint limitations based on her movements
in the gameplay and other information
Opening player models to players requires the underlying player models to be explainable, especially for the models with higher levels of openness. The game AI community has developed a large body of work to model players, but the vast majority of them are black boxes. A key open problem for OPM research is how to improve the transparency of player modeling. For systems that use AI modeling techniques that have low interpretability (e.g., deep learning), how can we still offer insights to players about why the player model classified them in a certain way. For other models, how can we choose the appropriate level of abstraction to reduce the noise of low-level actions and highlight the main trends that are useful to differentiate among the different players?
explainable player model, research is needed to effectively
communicate the model to players. Currently, information
visualization has been a primary tool to convey player stats.
As OPMs may contain more complex information, more
research is needed to expand the current work on
visualization and multi-modal explanations (e.g., visualization, text,
and voice). What is the information that the players want to
know about their own models? How can we visualize
highdimensional player data and gameplay traces? For OPMs
of open model outcome, how can we help players
establish the connection between their gameplay actions and the
model outcome? For OPMs of open model process, how can
we provide relevant information without overwhelming the
player? And for editable open models, can we provide
sufficient scaffolding so that the players can provide more
informed edits to the model?
Design Player-AI/Model Interaction in the Eco-System
of Games Compared with other eXplainable AI domains,
a unique aspect of OPM is that it needs to be embedded in
the rich game-player interaction context. This allows us to
investigate new OPM and explainable AI research questions
in a wide variety of contexts. For example, we can explore
when and how to provide OPM information to players in
the entire player experience. In open learner modeling
literature, the models are often displayed at the end of a
pedagogical module for reflection.However, in educational games, it
is possible that OPM can be used to provide guidance and
encourage reflection during the gameplay. How can we
address different player needs at various stages of the game
and incorporate OPM accordingly? Can we even align the
gameplay incentives with OPM so that players are
motivated to provide accurate information to the editable player
models? One notable design challenge is the potential
conflict between OPM and immersion. OPM is not for every
game. Directing players’ attention into the player models,
like any extra-diegetic features, may interrupt the narrative
or emotional immersion. Game designers need to be
mindful of these interruptions. Alternatively, OPM can be directly
incorporated into the core game mechanics (e.g., as a
metagaming feature
By increasing the transparency of player models, OPMs will
highlight the ethical issues that have underlined user
modeling in games. Similar to the research of ethics of AI
Another example of the ethical considerations of OPM is privacy. A good example is alluded to earlier on emotions or visual identifiable information embedded in the models. How do we share players’ traces in a socially responsible way? By showing the different gameplay strategies, the OPM can reveal struggling players. Would it expose these players to potential toxic behaviors in games? Or, can we steer what OPM reveals towards self-improvement and assistance from the game and its community? Finally, player modeling is rarely completely accurate. Opening these models up to players further raises questions around accountability. What happens when the model incorrectly characterizes a player? Editable open models allow players to identify and repair these instances, but they also impose additional challenges of explainability and preventing players from potentially exploiting the system.
Below we discuss a case study of OPM in an educational game. Note this work is still ongoing and the validation of this model has yet to be measured.
Parallel is a single-player 2D puzzle game, designed to teach
parallel and concurrent programming core concepts,
especially non-determinism, synchronization, and efficiency, to
Computer Science undergraduate students
From an in-class user study
Open Player Modeling for Parallel Player Model Type We believe that the use of OPM here can potentially enable players to reflect on their own problem-solving patterns and explore new ways when comparing them with other players’. Here we decided to use a descriptive model, where low-level gameplay actions, such as mouse clicks and button presses, are abstracted into higher-level actions such as level test, placing a semaphore, linking a semaphore to a signal, etc., which focuses on decision-making and problem-solving strategies. A key challenge here is that, directly showing players’ every gameplay action will be overwhelming in terms of the number of actions as well as the complexity of the action sequences. Yet the established technique of showing aggregated player stats hides necessary information about problem-solving strategies. This challenge creates a rich domain for OPM to explore how to model player action sequences and how to make them open and transparent to players for reflection (Villareale et al. 2020).
Level of Openness Since our goal is to improve players’ reflection and learning, we intend our player model to be open in a way that players can use to compare their gameplay/problem-solving process with others’. We decided to explore Open Model Outcome as the foundation type of OPM. In our case, the model is the individual players’ problem-solving processes, expressed as sequences of abstracted player actions during a level.
We use visualization, through the Glyph visualization
system, to communicate the OPM to the players. Glyph
To facilitate comparison of individual action sequences, we augmented the state graph with a synchronized sequence graph showing the popularity and similarity of sequence patterns exhibited by users (see Figure 2 Right). Each node in the sequence graph represents a full play trace, the size is an indication of popularity, i.e. how many players made these actions. Further, the distance between each node provides a visual representation of similarity/dissimilarity. For this particular level, the Glyph-based OPM revealed three distinctive ways that players in our dataset have solved it. The sequence graph (Figure 2 Right) shows three distinct clusters of the gameplay sequences. A player can locate her gameplay sequence and see how it relates to others.
To inspect this even further, the visualization system allows users to interact with the graphs, highlighting different sequences to inspect the decision-making steps in detail. Figure 2 shows such rendering of four different paths towards solutions for the level. This system then demonstrates an example Open Player Model - visualizing a descriptive process-oriented Model. One can thus imagine a player who may be struggling with a level to inspect the data from another player who was successful and follow through the sequence: “added a wait statement for red arrow”, “tested and failed”, “adjusted the wait statement for the red arrow”, “tested and passed”. Visualizing such patterns can give the player a way to inspect their own actions, reflect on them and also compare them to others. The specificity of the design and the visualization will depend on the model used.
We are currently tackling several open problems. First, we are investigating whether Glyph-based visualizations are interpretable by regular players. It is possible that players may find the state graph too difficult to understand. Through user studies, we will test and refine appropriate techniques that are informative and understandable to our target player group. There needs to be a balance between the expressivity of the visualization and its understandability. Second, we are also exploring the type of interactions and features needed to make our OPM useful for reflection. For example, we anticipate using annotations to allow users to leave notes or reflections on detailed problem-solving processes. Further, we also plan to use AI techniques to guide attention to similar sequences or add filtering options. These interactions will need to be tested through user studies. Third, we also need to investigate the type of player model to construct for an effective OPM. This is due to the fact that OPMs are tightly coupled with how explainable the resulting visualization is to the players. As mentioned above, there is a tension between models that capture sufficient information about problem-solving patterns (tend to be complex player models) and models that can be easily understood by players (tend to be simple player models).
While the case study above specifically showcased one example in the space of possible OPMs, there are many other examples that one can imagine for Parallel. These include recommendation models where the system recommends actions or strategies to players based on their models and level. We can also imagine models that give players reminders about specific aspects of the level to pay attention to, again given players’ models. These types of models are different from the descriptive models described above and can be effective for learning. However, they also need to be equipped with more explanations for the information presented. These could include different ways to establish the openness of the model. We can imagine, for example, a system that explains the process by which the model is developed, i.e. Open Model Process. We can also imagine that sometimes such models need to be edited or adjusted due to low accuracy or issues with the model, in such cases, designers can investigate the use of editable models. Researchers can develop different types of OPMs given the framework we outlined through adjusting the model’s type and level of openness. We hope these dimensions can help establish and further expand the work on player modeling to deliver better game mechanics as well as ways to explore engendering trust and fairness in the process of player modeling.
This work is supported by the National Science Foundation (NSF) under Grant #1917855. The authors would like to thank all past and current members of the project.