Dynamic Player Modelling: A Framework for Player-centred

Digital Games.

Darryl Charles1 and Michaela Black2

School of Computing and Information Engineering,
University
1
of Ulster, Northern Ireland.
[email protected]
2
[email protected]

KEYWORDS needs of the player in-game. While the concept of an

Adaptive Games, Dynamic Player Modelling, Neural adaptive game is a controversial topic among some gamers
Networks. and developers, there are clear benefits to tailoring the game
experience to particular player types – especially for
ABSTRACT educational games (Beal et al 2002). Catering for the
individual more effectively could help attract a wider
In this paper we outline a framework for creating player- participation, if for no other reason that it will be easier for
centred digital games. At the core of our proposal is the players to get started, progress and complete a game. In a
requirement for games to be more responsive to different recent edition of the Edge (Edge magazine 2004) Poole
player types and their individual needs; for games to have provides an insightful discussion on the problem of
the capability to adapt so as to provide an appropriate level “beginnings”, teaching the player, and lack of game
of challenge for each player, to smooth the learning curve, completions by most players, while in the same issue Redeye
and enhance the gameplay experience for each player highlights the niche quality to current games and their lack
individually. This is not an easy objective to achieve and of accessibility to a wider group of people.
adaptive game technology, although a popular ideal in some Adaptivity within games may primarily be implemented
quarters, is still fraught with difficulties and some by auto dynamic difficulty technologies (Miller, 2004) but
controversy. We address some of the most well known there are a number of other ways in which adaptivity can be
issues and outline a proposal for dealing with two of the advantageous. For example, in helping players avoid getting
more recent issues: that of monitoring the effectiveness of stuck, adapting the gameplay more to the player’s
game adaptation on the basis of player intention and/or preference/taste, or perhaps detecting deviant player
frustration, and dealing with dynamic player profiles – behaviour and modifying the game in response. What we
because players learn in different ways and at a different mean by deviant player behaviour is, for example, when a
speed. player uses or abuses an oversight in the game design to their
advantage. Often this means that the player finds it easier to
INTRODUCTION succeed in the game but their enjoyment of the game is
lessened because the challenge that they face is reduced and
All game players are different; each has a different they are not encouraged to explore the full features of the
preference for the pace and style of gameplay within a game, game – i.e. players will often repeat a successful strategy
and the range of game playing capabilities between players over and over again because it leads to a predictable win,
can vary widely. Even players with a similar level of game even if it is boring and somewhat ruins the game. This
playing ability will often find separate aspects of a game to happens frequently in real-time strategy games such as
be more difficult to them individually and the techniques that “Warcraft” or “Command and Conquer”. Bungie, the
each player focuses on to complete separate challenges can creators of “Halo 2” – a game much praised for its AI –
also be very different. For these reasons and others it can be acknowledged the importance of this when they designed the
very difficult to design a game that caters for a wide range of AI deliberately to prevent the player using “boring” tactics
player capability and preference. Game developers have but positively reinforced the player when they used
traditionally dealt with the range of player abilities in a very imaginative or adventurous tactics (Griesemer & Butcher,
straightforward manner, for example, by allowing the player 2002).
to select a difficulty level at the beginning of the game, as In this paper we propose and discuss a novel framework
with the classic first person shooter “Doom”. Once a player incorporating advanced ideas about player-centred game
selects their level of difficulty for a game designed in this design. This comprises of four key aspects: player
way, then there is usually no attempt within the game to modelling, adaptive game environments in response to
monitor how a player is performing in order to adjust the player needs, monitoring the effectiveness or appropriateness
level of challenge or gameplay experience. Recent games are of any adaptation, and dynamic player remodelling or
better at allowing a player to set up preferences for their classification.
gameplay experience, e.g. as with “DeusEx” where a player
can tailor their own avatar’s characteristics, but this relates
more to setting up the gameplay experience before starting
the game than intelligently recognizing and adapting to the
PLAYER-CENTRED GAME DESIGN the player to press a button that “rewinds” a sequence back
in time by up to approximately 10 seconds. This is
Most game design is, of course, already centred on the player particularly useful for dealing with mistakes, accidents, or
but it tends to focus on large groups of players rather than misjudgements by a player, for example, let’s say a player
catering for individual players – in this paper we hope to makes his/her character jump across a gap and the jump is
persuade the reader that games that are adaptive in catering miss-judged so that the character falls to his death. Usually
for the individual will be one of the key innovations in future in a game – particularly with game consoles – this would
games. One of the novel aspects of the framework that we mean that the level would have to be restarted or the player
propose in this paper for player-centric games is the ability would have to go back to a previous save point, but with
of a game to dynamically model, remodel, or reclassify a “Prince of Persia” a player simply rewinds that mistake and
player as they play the game. Players differ not only in their tries it again (up to a limited number of times obviously).
characteristics and ability as they begin to play the game, but This mechanism proves to be excellent in reducing
every player will learn at a different rate and each player will frustration simply by adding a quality game design feature.
excel in (or just simply enjoy) different aspects of the game. A different type of help mechanism used in 3rd person
The most common game model for differentiating view games is to have a game character look at areas that are
between player – and even this is quite rare – is shown in interesting as with “Eternal Darkness” where the player
Fig. 1. A player may set their difficulty preference (and character will turn his or her head to look at pictures etc. that
perhaps make a few other choices relating to their ability or perhaps should be examined. “Ico” is even more impressive
preference) before beginning the game. Within the gameplay in this regard in that the non player character “Yorda” who
itself there may be a simple hinder/help mechanism, as in the accompanies the main player character will often wander
racing game “Mario Kart” where a player who is doing well around independently, looking and pointing at things that the
will not receive good power-ups or weapon bonuses while a player should examine after he/she has been stuck in an area
player who is struggling will gain a lot of help through a for a while.
discreet speed up or by receiving more powerful item drops. These approaches supplement clear, directional level
Most of the simpler methods used – and often most effective design but are not particularly dynamic. So the basis that we
– are straightforward help mechanisms, for example in the propose for a player-centred framework should build on
“Crash Bandicoot” series if a player repeatedly fails at the recent research within the AI community with methods such
same point in the game then a mask is provided to the player as intelligent interfaces (Rogers & Iba, 2002, and
character which acts as a shield. This essentially allows the Livingstone & Charles, 2004). Mainstream AI research is
player to make one mistake and still be able to progress, e.g. relatively unused within the game development world yet
the character may hit a land mine once without losing a life. progress in this area for games has the potential to
Off-line information about players

Player
preferences

Monitor Help or
player hinder the
performance player.

In-game system

Fig. 1 A typical current game system that changes in response to the player.

Often such systems are “life” based as with “Maximo” revolutionise gameplay (Charles, 2003) as much as 3D game
where a player is provided with a coin by the angel of death technology has in the past. AI can provide a perceptual and
character “Grim” in order to buy another go when they die – functional interface between the player and game (Charles &
when a character fails at a challenge they may go back to a Livingstone, 2004) to enhance the experience for an
save point within the game level if, and only if, they have a individual player.
life/death coin. In this way a weaker player still has an
opportunity to progress while a stronger player is encouraged PLAYER MODELLING AND ADAPTIVE
to play sensibly – because they have a limited number coins.
However, as well designed as this mechanism is, the game GAMES
can still be prohibitively difficult at times for the novice A few game developers and researchers are now considering
player. When a player runs out of coins then he/she has to player modelling (Houlette, 2004) and adaptive games
reload a save and restart the level, and this inevitably is one (Charles, 2003, and Charles & Livingstone, 2004), though
of the reasons that many people will never finish this game. work in this area is still relatively rare. Fig 2 illustrates our
“Prince of Persia: The Sands of Time” provides another view of how a basic adaptive game system could be set up.
mechanism which operates in a similar manner by allowing Two sources of information can be used to identify the
player-type for a game: firstly the information that a player as a starting point for the dynamic modelling process in-
provides when they begin the game by setting basic game or to help label player groups. For example, we know
preferences and inputting information about themselves. The that there are certain differences, in general, between some
second source of information should be taken from the of these groups in terms of reaction time and in game play
player’s gameplay habits and performance in-game. deliberation. Of course, caution must be taken when
Together this information can be used to match the player to adopting this approach, because this initial classification
pre-defined models and the game can then be adapted to process will be quite coarse, e.g. girls may generally like
cater specifically to their needs and abilities. games like “The Sims” and “Everquest” due to the pace of
Players may not necessary need to be modelled by one the game and other factors but many prefer action/adventure
single object but several object models may be used to model or sports games. Identification of which type of information
them that cover different aspects of the gameplay and their produces the most informative profiles is a very important
relationship with the player. A more refined object model is initial task. Key fields of data can be identified as attributes
obviously better because player modelling can be complex, of information, for example gender attribute with values:
for example one player may be excellent at combat but male, female, and once the necessary attributes have been
terrible at problem solving, while for another the opposite identified and the information collected, some pre-analysis
may be true. In this case it is clearly better to model both can done. If predetermined profiles are not obvious we can
aspects separately, rather than try to fit them into a coarser use unsupervised machine learning techniques such as
single model. clustering to partition groups of players. We demonstrate
how both may be achieved with neural networks in the next

Off-line information about players

Models of Player
player types preferences

Monitor Adapt the

player game to
performance individuals

On-line
Adaptive game system

Fig 2. A basic adaptive game system.

chapter.
Player Modelling Once the most appropriate attributes have been identified
we then may produce our separate player profiles where each
It could be said that there are two main reasons for player
cluster group represents a different profile of player. If we
modelling in digital games. Firstly, modelling a player in
wish to be able to interpret the properties of these individual
order to instil human-like qualities into a non-player
groups, they can be labelled and the individual examples of
character, as was demonstrated by an example in a recent
each group provided to a supervised machine learning
paper (McGlinchey, 2003) where it was shown that the
technique such as a tree induction classifier (Quinlan, 1986).
characteristics of an individual player could be captured
This common inference task consists of making discrete
while playing a game of “Pong” by a Self Organising Map
predictions about a concept, in this case each profile, and this
(SOM) neural network. The SOM could then be used as the
prediction problem is referred to as the classification
“AI” for an artificial computer opponent in subsequent
problem. The task of a classification algorithm is to accept a
games. The second reason for player modelling and the
set of training examples which will depict the current state of
approach that we are interested in within this paper, is
knowledge for that concept/profile. These training examples
modelling players – or perhaps classifying typical player
are a set of descriptive attributes with an associated class,
types or behaviour – so that we may recognise predefined
and this class represents a value for the concept. The
player types or behaviour within the game. The reason that
algorithm will induce a knowledge structure to distinguish
we want to recognise the type of player currently playing is
between the values of the concept. A tree induction
because we wish the game to adapt the needs of the player.
algorithm will produce a classifier in the form of a tree from
To enable the creation of initial user profiles, some
which rules can be interpreted as one for each path from the
monitoring of game players is required to attain information.
root of the tree to each leaf. These rules depict knowledge
Additionally, information about the player – provided by the
which represents the concept. As will be demonstrated in the
player themselves – such as whether they are a novice or
next chapter we can also use other supervised learning
advanced, male or female, young or old, and other basic
algorithms such as neural networks.
general factors may be used as part of the player modelling
or clustering process. This information can be used as part of
the initial modelling or classification process and it can serve
Adaptive games possibilities with regard to our ideas about adapting to
individual players and the dynamic re-modelling of players.
Adaptation can have two related meanings: one meaning that
relates simply to change, and another related to learning and
Supervised Approaches
transformation. In the first case the adaptation from one form
to another has been predetermined and the adaptive states are In-game data is very valuable in the process of tailoring a
known in advance, and in the second case the adaptation game to the individual player and building accurate player
occurs after some learning from experience and the models. For example, we can use reaction times, choices
transformed stated may be previously unknown. Both forms made, styles of play, accuracy of shots/hits, how often a
of adaptation are relevant for games, but adaptation from stage needs to be repeated before completing, average health,
learning is the most interesting and also the more number of deaths per level, kills per level per possible kills
controversial. The reason for this controversy is that as with “Max Payne” Auto-dynamic Difficulty technology
mechanisms within games that have online learning are (Miller, 2004). This data may be used directly to decide how
unpredictable and therefore are very difficult to test to change the parameters of the game environment, attributes
thoroughly. Scepticism (or even anger!) is also often of the player character, or non-player character behaviour
expressed by gamers and developers with regard to games dynamically through the training of a neural network such as
that change according to player performance. “Mario Kart” the Backpropagation network. With this approach player
provides one of the most well-known examples; in this game entered game data may also be used alongside the in-game
if a player is winning then he/she does not get any of the player data to moderate the response of the network. This
powerful power-ups and the computer controlled cars often aspect could be important because it may provide a clue to
speed up – and the opposite is the case if a player is losing. how rapidly or how much the game should be adapted to the
While this can be annoying if a player is dominating a race, player. For example, if an advanced game player is currently
it does even out player ability disparity in a multiplayer playing then they may be less frustrated by not completing a
competition and thus the race may be more evenly matched challenge after a few attempts than a novice and therefore
and thus exciting. However, if players are aware of the game adaptation may be by a smaller amount or not at
“cheating” AI they may alter their gameplay accordingly; i.e. all. There are problems with using user-entered profile data
a player may decide to remain in second or third place until (or perhaps any type of profiling), for example, profiling
near the end of the race so that they may receive a significant may become frustrating or even redundant if more than one
power-up or weapon to unleash on the leader on the last player plays the game at the same time (taking turns) and
corner of the race – thus a new (perhaps unpredicted) thus sharing the same profile, in this case it would be
gameplay mechanic is introduced. There is some evidence impossible for the profiling and adaptation to be accurate.
that adaptive game technology is more effective when the Also, every type of game would require a different approach
player is unaware that it is happening, for example, the and the technology may not be appropriate for many types
primary author of this paper played and completed “Max multi-player games because players would be playing
Payne” without realizing that it incorporated “auto-dynamic against each other on an uneven playing field. For example,
difficulty” technology (Miller, 2004). in “Soul Calibur II” it is possible for a weaker player to
There are two opposing desires in players that we need to increase their “life bar” relative to their opponent but it
take into account: the desire of a player to learn the rules so actually unusual for this to occur in practice because players
as to master the game, and the requirement to avoid like to feel that they are competing on a level playing field.
“sameness” or lack of variety of gameplay. Thus, while we We can also take another neural network approach to
believe that there is a clear need for player-centric adaptive player modelling by using a clustering algorithm. In this way
technology within games to cater for individual players we use the neural networks to cluster player types according
needs, to help them learn and play the game, to enhance their to out-of-game and in-game data, grouping player with a
playing experience, to recognise when the player is stuck or similar profile into the same group type. There is a wide
frustrated and help out. There is also a requirement that the range of ways in which this may be done, for example we
rules of the game do not change significantly, which would could use a radial basis network with fixed cluster centres to
frustrate many players, and ideally either the player should classify the players, with the centres fixed on different areas
not aware of the adaptive nature of the game or they should of the data space that we believe to provide a good “centre”
have the option to switch it off. for our player classification. By monitoring and adapting the
player profile throughout the game then the player may
achieve a new classification, and thus the game would
NEURAL NETWORKS FOR THE
respond differently. Radial basis networks may also have
MODELLING PROCESS moving “centres” and so the centres can be moved
The use of neural networks for the player modelling process automatically during training to fit the data more
is quite an obvious approach but the authors are not aware of appropriately. It is also possible to retrain the full network
them having been used much for this purpose in games yet during gameplay on the basis of new data, although this is
and so we provide an overview to a few possible supervised not necessarily an easy thing to do. For example, a single
and unsupervised approaches below. Neural networks are player, depending on the method, may only provide one new
good at detecting patterns and clustering data (depending on data point and so re-training may be futile. This is generally
the method) and so we can use a variety of neural network an issue with online learning in games; it is not only slow but
techniques in different ways to identify or understand often there is not enough new data to significantly impact the
different players. Additionally, as neural networks are training of the network, and needs to be taken into account
essentially learning machines they hold a number of
when choosing which method to use and how to implement relationships between sub-sets of the data variables that may
it. be used to identify more refined aspects of player behaviour,
e.g. output one could identify the overall capability of the
Unsupervised Approaches player and output two may identify whether the player is
cautious or just dashes in etc. Being able to identify more
There is very little digital game research going on that
subtle or complex aspects of player behaviour could be very
involves unsupervised learning, perhaps because of a lack of
valuable in tailoring the game experience to the player, and it
expertise in this area. However, we would like to
also potentially opens up new possibilities for dynamic
demonstrate here that there are a few very positive and
gameplay. For example, if we are able to discover patterns
promising uses for unsupervised neural networks for forming
that relate more to player emotion or motivation then this
a statistical understanding of player data. Unsupervised
may be used with other sensory devices to discern the needs
neural networks are generally used to explore or investigate
or desires of the player and the game can be adapted to
structure or patterns in data on the basis of statistics or
account for this.
information theory (or similar). It is not known, a priori
(though we may have an idea), what the relationship is
between the data variables and we would like to investigate AN ADVANCED FRAMEWORK FOR
this. This is similar to data mining the player data and we PLAYER-CENTRED GAMES
can use this approach to help us understand the difference
between player styles or capabilities then use this knowledge Two particular novel technology aspects that we discuss in
in our player modelling process. Once the neural network this paper are monitoring adaptation through sensory
has been trained to our satisfaction then it may be used equipment and dynamic player modelling and we explore
directly in-game to identify player types or behaviours. these in more detail within this section. Detection of the need
Many of these algorithms are also quick to train and so may for the game to adapt based, for example, on measuring
be more suitable than other approaches for on-line re- player frustration (Gilleade & Dix, 2004) is one approach for
training. The techniques that we focus on in the examples game adaptation but we propose a slightly different model,
below are known as projection methods. With projection one in which the game is adapted on the basis of detecting
methods we typically want to explore the relationship player type coupled with game performance. The
between the input variables but with clustering approaches effectiveness of adaptation can then be measured by a
we treat each data example as a data point (e.g. a player reduction in the level of frustration and other measures. If
description) and attempt to group data points together based adaptation does not improve player performance or their
on some similarity measure. frustration levels then perhaps this is because the player has
Let us say that we wish to explore the relationship been classified incorrectly, or more likely as they have
between the variables that we have chosen to uniquely progressed through the game the model that fitted the player
describe a player in a game, e.g. average health, times shot, initially is no longer applicable. Therefore in this scenario it
enemies shot, enemies killed, etc. Then using statistical may make sense to reclassify or dynamically remodel the
neural network approaches such as Principal Component player – Fig 3 illustrates how this advanced framework may
Analysis or Factor Analysis we may explore the data so as to be executed.
identify the correlational (or high order statistical)
relationship between the variables. Factor analysis is Measuring the Effectiveness of Adaptation
particularly interesting in this regard because it is frequently We need to know when to adapt the game to a player
used by statisticians in an exploratory mode. A well known (Gilleade & Dix, 2004) but also we should monitor if our
example of the use of this method is where the statistical adaptation has been effective or appropriate. If we make a
relationship for different forms crime in different cities are change based on the game data coupled with the player
explored, e.g. murder, theft, robbery etc. Factor analysis can profile and this frustrates, or hinders the player more (or vice
decipher which input variables have the strongest correlation versa) then we may make one of two conclusions: our
and the statistician can interpret what this means. It may be adaptation is inappropriate or our model of the player is
found that there is a strong link between robbery and murder inaccurate. In either case this is a good reason to have the
and so the output of the network that identifies this feedback loop in our model illustrated in Fig. 3.
relationship may be said to have identified a correlational Assuming that there are discrete changes to the adaptation
link which can be explained because these are violent of the game and that these have tested these thoroughly
crimes. Similarly, a non-violent crime correlational may be before game release, we then can focus on making sure that
discovered. Using this method to explore player data we may we classify the player correctly so that the state of the game
have an advantage in our interpretation of the data because is appropriate to them. This is especially important because
we can also collect information additional about the player players learn at different rates and so we need to take
that can help us interpret the statistical relationships, e.g. account of concept drift (Black & Hickey, 1999) in the
how old are they, sex, what type of games they like to play, classification process (see next section).
how often do they play etc. These values could also be used The manner in which we measure the requirement or
in the statistical analysis but we would suggest that they may appropriateness of adaptation may be most effectively
be better served in helping us interpret the correlations achieved using affective computing techniques by
discovered by the outputs of the network. Whereas a monitoring a player’s emotional state through input devices,
clustering method would group players together so that we coupled with in-game data. It seems clear by initial research
can label these groups as novice, normal or advanced, on the that attempts to detect a player’s emotion through input
basis of the complete data point. Factor Analysis can identify devices that it is not very straightforward. For example, the
 Off-line information about players

Models of Player
player types preferences

Monitor Adapt the Measure

player game to effectiveness
performance individuals of adaptation

Re-model
player types
On-line
Adaptive game system

Fig. 3 Adaptive game system diagram illustrating the three phases that that takes account of
errors in adaptation.
emotional affect detected in the player through a gamepad new area for digital game research. On a basic level a player
analog button (Sykes & Brown, 2003) will vary with each model may be thought of as a statistical representation of the
player, game type and even perhaps when it’s played. The player based on the frequency of repeated actions or average
information may be corrupted by interaction stress in, for values of the parameters of their player character etc. It
example, playing an action game by altering physiological should be obvious then that an individual player’s profile is
factors that would normally infer emotion, such as skin likely to change throughout the progress of a game. This can
conductance (when using an appropriate sensor). Facial be for all sorts of reasons, for example they are learning the
expressions or body movement may be used to infer the action aspects of game more quickly than adventure aspects
emotional state of the player – whether they are happy, or perhaps they have reached a new gameplay dynamic in
content or frustrated – and game cameras such as the PS2 the game that they can’t quite get to grips with – all players
EyeToy is becoming more popular and widespreadly will be different so these things are very difficult to predict.
incorporated into games. The difficulty with using a camera Because of the nature of game playing there will be new
for facial expression though is that, to some degree, there is examples available about the player’s profile as they play the
an expectation that the player will roughly maintain their game, hence the requirement for on-line learning. These on-
position relative to the camera (Gilleade & Dix, 2004) – this line learning systems will receive examples on a continual
is particularly an issue with game consoles. basis and are required to induce and maintain a basis for
With simple modification of existing input devices classification and thus may have to deal with concept drift
temperature or pulse (i.e. heart rate) sensors may be added (Black & Hickey, 1999). Game players will adapt their
like those on a typical exercise bicycle. These would not be strategy to survive or win as the game adapts to suit their
expensive to implement but could potentially revolutionize profile. This change in the player behaviour, as discussed
game design with respect to a games’ responsiveness to an previously, may be part of their learning process: i.e. they
individual players’ needs. Even in casual way this get better at the game over time, or a may be forced into a
information may introduce interesting new directions in strategic change of tactics. This adaptation, known as
gameplay – if you imagine a game from the horror genre concept drift, can therefore be an immediate change in tactic
such as “Silent Hill” or “Resident Evil”. In this example the or a slow progression to another. By concept drift we mean
game could wait until a player seems at their more relaxed that some, or all, of the basis for defining a profile is
before landing that shocking surprise on him or her. changing as a function of time.
Normally, games of this type must craft the levels and script Typically there are a number of sub-tasks involved in the
events very cleverly to achieve the same effect, and it is very handling of drift within incremental classification learning.
difficult to perfect. It will probably prove to be the case that In increasing order of difficulty these are:
one method alone will not be enough to accurately gauge a
player’s mood. That a mix information from standard 1. Identifying that drift is occurring;
sources such as the mouse or joy pad, along with more 2. Updating classification in the light of drift;
advanced sources of information provided by cameras or 3. Tracking and modeling/analysing the pattern of
other sensory devices along with the player’s profile, will be drift over a period of time.
necessary to make decisions that tailor the game to
individual players on the basis of their emotional state. Machine learning techniques have been used with this
Statistical methods such as neural networks will then be form of user profiling/modelling in other domains such as
necessary to decipher the structural relationships in the data. cellular fraud in telecommunications (Fawcett & Provost,
1999). The aim was to analyse calling behaviour and detect
Dynamic Player Modelling and Reclassification anomalies. It also highlighted that patterns of fraud are
dynamic; bandits constantly change their strategies to avoid
The idea that a player’s model needs to be adapted has been
detection. This links very well into game players having
recognised recently (Houlette, 2004) but this is still a very
profiles which change/evolve through the life of a game. Charles D, Livingstone D, "AI: the Missing Link in Digital Game
Game players may be thought of as behaving like the Interface Design?", 3rd International Conference on Entertainment
fraudsters; they adapt and change their strategies as a Computing, September 1st-3rd, 2004, Eindhoven, The Netherlands.
mechanism to win/survive and so profiles can be monitored Edge Magazine (UK) #139, Trigger Happy and Redeye articles,
August 2004.
and adapted using existing machine learning techniques. For
example, recent work (Black & Hickey, 1999) has Fawcett T. & Provost F, “Activity Monitoring: Noticing Interesting
demonstrated that profiles may be induced from Changes in Behavior”, 5th ACM SIGKDD Int. Conference on
telecommunication customers in relation to using a product, Knowledge Discovery and Data Mining, San Diego, pp 53 – 62,
and that changes may be detected in the customers who are 1999.
currently using the product.
As already indicated, player’s profiles may change in Gilleade K. & Dix A, “Using Frustration in the Design of Adaptive
many ways. We can break these down into two aspects of Videogames”, Proceedings of ACE 2004, Advances in Computer
change: a progressive move – referred to as evolutionary Entertainment Technology, ACM Press, 3-5 June, 2004.
adaptation, or immediate change – referred to as Griesemer J. & Butcher C, “The Illusion of Intelligence”, Game
revolutionary adaptation (Black & Hickey, 1999). This work Developer Conference, San Jose, 2002, slides available from (last
also introduces a methodology called TSAR (Time Stamp accessed 11/07/2004),
Attribute Relevance) which has been used successively to http://halo.bungie.org/misc/gdc.2002.haloai/talk.html
adapt to concept drift in telecommunication customer data
(Black & Hickey, 2002). This methodology can be applied to Houlette R, “Player Modelling for Adaptive Games”, pp 557-566,
neural network approaches, as discussed earlier, so as to deal AI Game Programming Wisdom II, Charles River Media, 2004.
with concept drift in online learning within digital games.
Livingstone D. & Charles D, “Intelligent Interfaces for Digital
Games”, AAAI-04 Workshop on Challenges in Game AI, 25-26th
CONCLUSIONS July 2004.
Modern digital games are extraordinarily good at many McGlinchey S, “Learning of AI Players from Game Observation
things but even the best examples of these games are still not Data”, GAME-ON 2003, 4th International Conference on Intelligent
very capable at monitoring players, distinguishing between Games and Simulation, pp. 106-110, Nov. 2003.
different player groups and altering the game state to meet
individual players’ needs. In this paper we described a Miller S, “Auto-dynamic Difficulty”, website forum debate, last
framework for dealing with this issue and providing more accessed 11/07/2004,
adaptable games, and in particular approaches for dealing http://dukenukem.typepad.com/game_matters/2004/01/autoadjustin
g_g.html
with two particularly current issues: that of monitoring the
effectiveness of adaptation through affective and statistical Quinlan, J.R., Induction of Decision Trees, Machine Learning. 1(1),
computing approaches, and the dynamic remodelling of 81-106, 1986.
players based on ideas from concept drift. We proposed
several neural network approaches as part of the realisation Rogers S. & Iba W, “Adaptive User Interfaces”, 2000 AAAI Spring
of this framework and in future work intend to test these Symposium, Technical Report SS-00-01.
ideas further. The improvements that may come from
positive developments in this area could be as Sykes J. & Brown S, “Affective Gaming: Measuring Emotion
straightforward as helping the player in learning how to play through the Gamepad”, Human Factors in Computing, pp 732-733
CHI 2003.
the game, through to encouraging gameplay innovation in
digital games.

References
Beal C, Beck J, Westbrook D, Atkin M & Cohen P, “Intelligent
Modeling of the User in Interactive Entertainment”, pp 8 – 12,
AAAI Spring Symposium on Artificial Intelligence and Interactive
Entertainment, Standford, 2002.

Black, MM, Hickey RJ, “Maintaining the Performance of a Learned

Classifier under Concept Drift”, Intelligent Data Analysis 3 (1999)
453-474.

Black MM & Hickey RJ, “Classification of Customer Call Data in

the Presence of Concept Drift and Noise” Soft-Ware 2002:
Computing in an Imperfect World, Belfast, 8-10 April 2002, in
"LNCS 2311", Springer-Verlag (Berlin), ISBN 3-540-43481-X,
Pages 74-87.

Charles D, "Enhancing Gameplay: Challenges for Artificial

Intelligence in Digital Games", 1st World Conference on Digital
Games 2003, University of Utrecht, The Netherlands, 4-6th
November 2003.