Cognitive Modeling of Video Game Players

Cognitive Modeling of Video Game Players
Cognitive Modeling of Video Game Players
Corey J. Bohil (Blank if Blind Review)
Affiliation (Blank if Blind Review)
Address (Blank if Blind Review)
e-mail address (Blank if Blind Review)
Optional phone number (Blank if Blind Review)
Frank A. Biocca (Blank if Blind Review)
Affiliation (Blank if Blind Review)
Address (Blank if Blind Review)
e-mail address (Blank if Blind Review)
Optional phone number (Blank if Blind Review)
then posttesting using the same standard laboratory task
again to see if performance has changed.
This paper argues for the use of cognitive modeling to gain
a detailed and dynamic look into user experience during
game play. Applying cognitive models to game play data
can help researchers understand a player’s attentional focus,
memory status, learning state, and decision strategies
(among other things) as these cognitive processes occurred
throughout game play. This is a stark contrast to the
common approach of trying to assess the long-term impact
of games on cognitive functioning after game play has
ended. We describe what cognitive models are, what they
can be used for and how game researchers could benefit by
adopting these methods. We also provide details of a single
model – based on decision field theory – that has been
successfully applied to data sets from memory, perception,
and decision making experiments, and has recently found
application in real world scenarios.
These approaches take a bird’s eye view of the cognitive
phenomena underlying game play. They treat the mind like
an impenetrable black box, observing or manipulating
inputs to the cognitive system, and observing the
concomitant outcomes. Although these research endeavors
are valuable, they take an indirect route to understanding
cognition during play. In both approaches outlined above,
game play effects are measured after the fact. In the case of
questionnaire methods, the data is subjective – participants
give some indication of the game’s effects through their
answers to various questions. And although the second
research approach aims at understanding something about
cognition as it pertains to games, it too focuses on effects
and measures indirectly by assessing changes after the fact.
What is needed is an approach that allows one to track
cognitive effects of games during the game play session.
As the player progresses through the game, seeking to
accomplish various goals, making decisions, all manner of
cognitive phenomena come to bear. Learning is required
(e.g., what strategies worked before?). Attention has to be
allocated. Memories of previous outcomes have to be
accessed. Decisions have to be made. The outcome of
these cognitive events translate into the player’s observable
performance in the game, their level of enjoyment or
accomplishment, their learning.
Understanding these
dynamic events as they unfold throughout the course of
play, rather than trying to infer something about them
subjectively or indirectly after the game is over, would be
of great value to the designer of games concerned with
changing behavior, communicating messages, or just
maximizing engagement and fun.
In attempting to
maximize game efficacy, the designer would likely benefit
from knowing what a player is looking at when making
decisions, what dimensions are most salient, which
dimensions are routinely ignored, and when options are
Author Keywords
Cognitive, model, game, user experience, sequential
ACM Classification Keywords
H5.m. Information interfaces and presentation (e.g., HCI):
A major goal of video game research is to understand and
influence what a player is thinking during game play, and
perhaps to effect long term changes in the game player.
One major theme in this research includes game impact on
personality traits and emotional states – most famously,
effects on player aggressiveness. This type of research is
typically carried out via survey methods. Players answer a
battery of questions before the game begins to assess their
personal traits and current emotional state. After game play
has completed, the player answers more questions and any
changes are attributed to the effects of the intervening game
play period. Another prominent theme is the effect of
games on some aspect of cognition (e.g., spatial skills [5]).
Although this second research theme is concerned with
cognitive effects of games, it is often carried out by
pretesting the participant using a standard laboratory task
(e.g., a speeded search task) to assess the trait of interest
(e.g., visual acuity), followed by a game play period, and
In recent years, there have been efforts to peer into the inner
workings of the mind during the game play events that
cause them. Brain imaging techniques (e.g., fMRI) have
been used to associate brain activity known to occur during
aggressive thought with violent game content [8]. Other
studies have tracked psychophysiological events (e.g.,
EEG) to infer mental states during play [1]. This approach
to studying player cognition during game play is a welcome
addition to the field. This research gives us valuable data
synched in time to game events, and we can learn a lot by
trying to interpret it. However, at this point in time, both
imaging and psychophysiolical data are difficult and
expensive to obtain, the measures are still relatively crude,
and findings are often difficult to interpret. The degree to
which one can relate the observed bodily states to aspects of
cognition such as attention or executive function is an issue
of lively (sometimes withering) debate [7].
Luckily, there is another approach to understanding
cognitive phenomena at our disposal – cognitive modeling.
For the past several decades, cognitive psychologists have
devised and tested scores of detailed mathematical models
that offer precise accounts of the cognitive underpinnings of
behavior, and demonstrated their links to theoretical
structures like memory and attention. Given the mature
state of this field, as well as its widespread representation
throughout academia, it is surprising to find that it has very
little representation in the game studies literature (although
some applications are noted below). The objective of this
paper is to provide a basic understanding of what cognitive
models can provide researchers, and to advocate their use in
studying video games.
What, exactly, is a “cognitive model”? A cognitive model
is a mathematical interpretation (i.e., specification) of the
set of principles embodied in a theory of cognition.
Cognitive models make specific assumptions about the
information represented in the cognitive system (e.g., words
and their meanings), along with the processes acting on this
information to produce observable cognitive behavior (e.g.,
classifying an object). More concretely, a model receives
inputs like a person in an experiment (e.g., size of objects
on a screen), performs mental operations (e.g., like
comparing perceived stimulus information to information
stored in memory), and outputs a response (e.g., emits a
classification of the object).
Models such as these are valuable for several reasons.
First, they require a researcher to move past the initial
stages of theorizing – often characterized by vague verbal
descriptions of mental entities and their interactions – to
taking a detailed, specific stance on these quantities and
relationships. Doing this affords the research community a
better opportunity to evaluate and criticize a theory’s
quality. Second, making detailed quantitative statements in
a cognitive model allows a researcher to make precise,
testable predictions. A third benefit is that simulating
model behavior on a computer can lead to unexpected
observations and insights that the researcher might not
otherwise have reached. It is widely agreed in the modeling
community that this is an important benefit of modeling.
There currently exists a wide array of cognitive models that
have been vetted over the years by many experiments and
data sets. These models elucidate a range of topics. Many
models are designed to capture steady-state performance in
cognitive tasks like recognition memory, discrimination
ability, attention allocation, to name a few. These models
are intended to account for specific, circumscribed aspects
of cognition such as recognition, categorization, attention,
etc. Another class of models – known as connectionist
models (also called neural net or parallel distributed
processing models) – mimic fundamental aspects of brain
anatomy (i.e., populations of single processing units or
artificial neurons communicating activation levels back and
forth) and capture learning over the course of many training
trials. A third class of models – known as cognitive
architectures (e.g, ACT-R, EPIC, Soar) – attempt to capture
several aspects of cognition in a single unified framework
(e.g., attentional processes, memory, visual search
tendencies), reflecting the fact that all these processes come
into play simultaneously in the human cognitive system.
Cognitive architectures have found wide application in
human-computer interaction research and have even made
their way into game research to some extent [4]. Existing
applications of cognitive modeling in game research tends
to take on a computer-science flavor. These models are
valuable tools for making the game respond to the player in
interesting ways or to create “smarter” non-player
characters [3]. Our aim in this paper is to encourage much
more widespread adoption of these techniques for gaining
general understanding of the cognitive capacities invoked
during video game play.
The value of models for game research lies in the fact that
models require inputs and produce outputs. In between
they offer precise statements about attention, learning,
decision strategies and biases, and so on. In doing so, a
model often tells the researcher why performance looks as it
does. Although a model can’t tell the designer exactly how
to craft a game environment that teaches or entertains,
discovering that current inputs place unrealistic demands on
attention might offer guidance by narrowing the range of
necessary modifications to gain desired results. An
important detail, of course, is how one goes about applying
these models.
Within a cognitive model lie parameters that capture the
modeled quantities (e.g., attention weights, learning rate,
response biases). These values are indicators of the mental
underpinnings of observable behavior. In order to make
inferences about cognition, these models are often “fit” to a
set of data. The computer takes the output of the model
(i.e., predicted responses to events), compares it to player
data (i.e., actual responses) and adjusts the internal
parameter values (i.e., changes assumptions about attention,
etc.) until the predicted responses are as close as possible to
the data. The resulting adjusted parameter values indicate
things like how confusable the stimuli were or which
stimulus dimensions garnered the most attention. These
parameter values can be used to make predictions for the
player in later game sessions or scenarios.
It is also important to verify that what the model tells us is
correct. In order to do this, researchers often attempt to fit a
model to data using fixed parameter values gleaned from
prior knowledge of the research participant. Achieving a
good model fit (i.e., a good prediction of player
performance) by setting parameter values a priori is a
powerful demonstration that one understands the player’s
cognitive processes during play.
Among the many aspects of cognition that can be modeled
and examined in games, perhaps the most natural starting
point is to look at decision making. One popular class of
models that illuminates decision making is known as
“sequential sampling” models. Sequential sampling models
simulate the accumulation of information (i.e. sampling)
over time in support of each choice alternative, leading to
the eventual selection of one option over others. Decisions
are triggered by internal choice thresholds – the first
accumulation process to reach threshold wins, and the
corresponding choice is made. Figure 1 depicts this
sampling process for three choice options.
One way to obtain fixed parameter values for a priori
prediction is to fit the model (by adjusting free parameters)
to one data set, and then use the best-fitting parameter
values to see if the model accounts for additional data sets
(without re-adjusting the parameter values). Another way
to demonstrate our understanding is to set the model
parameters based on something else we already know about
the player.
For example, one could take advantage of the kinds of data
acquired through the survey methods described above. One
recent (non-game) study used results from a survey
designed to assess whether a person has an “action”
orientation (tendency to accept risks to expedite achieving a
goal) or a “state” orientation (tendency to be more
deliberative in order to avoid risks). Scores on this
questionnaire were converted into parameter values in a
cognitive model and used to predict response probabilities
and response time distributions in a sports-related task [6].
Such an approach grounds model parameters in knowledge
about the participant even before experimental
manipulation begins, and can still enable the model to make
interesting predictions about behavior.
Figure 1. Information accumulation for three choice options
Another possibility would be for the researcher to set model
parameters to reflect instructions given to the player (either
before the game or inside the game). Instructing a player to
pay attention only to RED enemies, for example, should be
reflected in a model’s attention weight parameters
(assuming the model has them) and consequently in the
model’s predicted response probabilities (and hopefully
lead to a good model fit). An important long-term goal of
modeling is to find parameter values that can lead to valid
predictions across several experimental conditions without
the need to adjust parameters to account for each data set.
In this section, we describe a sequential sampling model
based on Decision Field Theory [2]. Variants of this model
have been successfully applied to a wide range of
phenomena, including decision making, perception, and
memory, among others. The model has mostly been
applied to data from standard laboratory tasks, but has
recently been used to explain decision making in a sports
judgment task [6].
On a given experimental trial (in the context of games, an
operationally defined recurring event), the model assumes
that each set of choice options can be characterized by
values along salient dimensions. For example, when trying
to choose the best weapon for a fight, the player might
consider three weapons along dimensions such as strength,
range, and ammunition supply. Each weapon has its own
set of values on these dimensions, and the player makes
some assessment of these values. Table 1 illustrates some
hypothetical values.
One challenge to applying cognitive models to data from
video games is that events of interest must be operationally
defined. For example, some agreement might need to be
reached about what constitutes “fighting or fleeing” in a
game scenario. Another example would be determining
what qualitifies as a response option. Depending on the
question under study, it may be wise to compare
performance only in situations with a constant number of
response options. Such apples-to-apples comparisions
might be necessary when trying to determine response
probabilities or response time distributions.
Strength Option 1 1.0 Option 2 0.5 Option 3 0.7 Range 50 150 100 Ammo 0.6 0.8 0.8 Table 1. Hypothetical dimension values for weapon choices
Also, each player is likely to display some difference in
preference for the choice dimensions. For example, the
player’s decision might be most strongly influenced by the
strength dimension 70% of the time. Range might be the
most influential dimension 20% of the time, and ammo only
10% of the time. These values are model parameters. The
model uses these values on each trial, along with other
parameters representing initial biases, memory from trial to
trial, and similarity between options, to produce a decision.
Figure 1 displays a characteristic example. The figure
shows the (simulated) stochastic accumulation over time of
evidence (to the cognitive system) in favor each of the three
options. The first option to reach an internal decision
threshold “wins” the race, and supplies the response. As
the figure shows, not only is a choice determined from this
process, but also the time to reach threshold. From trial to
trial, responses and termination times will vary, and over
trials the model will provide response probabilities and
response time distributions that can be compared to a
player’s data. By adjusting the internal model parameters
in order to fit the observed responses, the model tells a tale
about the player’s attention focus, memory, biases, and the
confusability of the response alternatives.
This decision-field theoretic model could potentially answer
many interesting questions. For example, how does action
orientation predict game play? How does decision strategy
change as a result of learning throughout the game? Which
dimensions receive the most attention, and which the least?
We’ve argued in this paper that cognitive modeling
provides a detailed and dynamic view into cognition – at
the individual player level – as it unfolds during video game
play. Currently, this powerful approach is seldom utilized
in game research. This is a shame, since cognitive
modeling is a mature field, and there are many useful
models available that have been affirmed by decades of
research in carefully controlled experiments.
Models can offer clues into the inputs required to produce
the outputs desired. If a game is to have educational value,
(or for communication or even just for fun), then variables
that influence model behavior should be manipulated to
moderate player behavior. Currently, most game design is
guided by heuristics, prior experience, and flashes of
Many of the cognitive models in existence today are ready
for extension to new areas. In fact, the field of cognitive
psychology is increasingly marked by attempts to extend
the reach of cognitive theories to real-world scenarios. The
application of these tools is especially timely considering
the recent explosion of research into serious games (games
designed to communicate and educate players). Designers
of such games would likely benefit from a tool that can help
foster a deeper understanding of what players focus on and
are affected by during game play.
Finally, cognitive modeling dovetails well with the imaging
and psychophysiological research mentioned above. The
relatively recent emergence of the field of cognitive
neuroscience attests to this. Cognitive models have become
so powerful that competition between theories is often
difficult to assess on the basis of behavioral data alone.
Neuroscience data is now routinely used to place biological
plausibility constraints on computational models. In turn,
cognitive modeling imparts a deep level of meaning to
neuroscience results.
Models help neuroscientists
understand the cognitive implications of their data.
In conclusion, cognitive modeling presents a powerful
method for understanding what a player is thinking about
while playing a video game. Research papers that describe
cognitive models often report their model derivations in
detail so that interested readers can adopt these methods.
Our hope is that we’ve been able to convince readers of the
allure of cognitive modeling for their own game research.
We thank Bradly Alicea, Nick Bowman, Andy Boyan,
Allison Eden, & Mayur Mudigonda for their comments on
an earlier draft of this article. This work was funded by an
AT&T endowment to the second author.
1. Baumgartner, T., Valko, L., Esslen, M., & Jancke, L.
Neural correlate of spatial presence in an arousing and
noninteractive virtual reality: An EEG and
psychophysiology study. Cyberpsychology & Behavior
9,1 (2006), 30-45.
2. Busemeyer, J. R., & Townsend, J. T. Decision field
theory: a dynamic-cognitive approach to decision
making in an uncertain environment. Psychological
Review, 100 (1993), 432-459
3. Funge, J. D. AI for Computer Games and Animation: A
cognitive Modeling Approach. AK Peters Ltd.,
Wellesley, MA, USA, 1999
4. Gray, W. D. (Ed.) Integrated Models of Cognitive
Systems. Oxford University Press, New York, 2007
5. Kaveri, S., & Greenfield, P. M. Effect of video game
practice on spatial skills in girls and boys. Journal of
Applied Developmental Psychology 15,1 (1994), 13-32.
6. Raab, M., & Johnson, J. G. Individual differences of
action-orientation for risk-taking in sports. Research
Quarterly for Exercise and Sport, 75,3 (2004), 326-336.
7. Uttal, W. R. The New Phrenology: The Limits of
Localizing Cognitive Processes in the Brain. The MIT
Press, MA, USA, 2003
8. Weber, R., Ritterfeld, U., & Mathiak, K. Does playing
violent video games induce aggression? Empirical
evidence of a functional magnetic resonance imaging
study. Media Psychology 8,1 (2006), 39-60.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF