Experience Assessment and Design in the Analysis of Gameplay

Psychophysiology and Emotions

Psychophysiological methods study psychological phenomena, emotions in the present case, by measuring human autonomous physiological activity, which is real time and relatively objective. Psychophysiological studies often use a model of emotions positing that emotional experiences can be organized along two dimensions: valence and arousal (Lang, 1995). The valence dimension corresponds to the extent to which an emotional experience is unpleasant or pleasant and the arousal dimension indicates the level of (bodily) activation associated with the emotional experience. However, when studying responses to complex, interactive, and multimodal activity such as digital game playing, it is important to notice that positive and negative emotions are not mutually exclusive. Indeed theory and experimental evidence both suggest that they are, at least to some extent, separable and independent (see, for example, Cacioppo, Gardner, & Berntson, 1999; Larsen & McGraw, 2011; Park, 2008). Furthermore, the valence and arousal dimensions have been suggested to be the subjective components of the two primary brain motivational systems: the behavioral activation system (BAS) that regulates approach/appetitive behavior and the behavioral inhibition system (BIS) that regulates withdrawal/aversive behavior, respectively (Watson, Wiese, Vaidya, & Tellegen, 1999).

The current iteration of the PPAX framework focuses on those signals that combine practicality of use with relevance to the valence-arousal model. The signals in this core set are relatively simple to record, functionally cohesive and have strong empirical and theoretical foundations in the game research literature. These are electrodermal activity (EDA), associated with arousal; facial electromyography (EMG), which indexes conscious and non-conscious emotional expression; and electrocardiography (ECG), which measures the activity of the heart and can help to index many processes such as mental workload.

Processing pleasant or unpleasant emotions, or stress, is associated with increasing activity over facial muscle areas, which can be measured using EMG (Bradley, 2000; Ekman, Davidson, & Friesen, 1990; Lang, 1995; Tassinary & Cacioppo, 2000). Several encouraging reports on using EMG specifically to index responses to digital game events have already been published (see Kivikangas et al., 2011; Kivikangas & Ravaja, 2013).

Arousal is most often measured with EDA (or skin conductance level; sometimes inaccurately called galvanic skin response; Bradley, 2000; Lang, 1995). EDA is a reliable tool when studying gaming experiences (e.g., Mandryk & Atkins, 2007; Schneider, Lang, Shin, & Bradley, 2004; Staude-Müller, Bliesener, & Luthman, 2008), as it is less susceptible to misinterpretations than either ECG or facial EMG.

In addition, features of cardiac activity such as heart rate (HR) and certain frequency bands of HR variability are among the most widely used physiological signals, despite the fact that interpretation in the game context can be challenging, given that the heart and circulatory system are regulated by many different bodily processes. This difficulty is demonstrated by the fact that, in different studies, cardiac activity has been interpreted as an index of both valence and arousal, but also of attention, cognitive effort (see, for example, Cowley, Ravaja, & Heikura, 2013), stress, and the orientation reflex (see, for example, Ravaja, 2004). Still, cardiac indices have been used in many game studies especially for indexing arousal, and often with claims of success (e.g., Ravaja et al., 2006).

Formal Game Features and Game Design

The PPAX framework seeks to integrate experiential data with a structural understanding of game events. A major challenge when discussing the evaluation of design is how to classify individual game features in meaningful ways that allow us to link particular elements or game events to the playing experience, but also adequately contextualizes elements within the greater plan of the game. The development of formal systems for investigating games has largely been design-driven, addressing practitioners’ need for talking about design decisions in terms of gameplay outcomes. However, the conceptual roots are sociological, lying for example in the 1960s work of Caillois (2001), who distinguished four types of games as having different structural and experiential qualities. We can extrapolate from Caillois’s work that these different types of games would arise from different types of play preferences; in other words, different forms of play arose from different ways of having fun. Even a complex game is usually built from simpler elements that are based on either competition, chance, role-playing, or other mechanics. The typical way of structuring discussion of these archetypal forms of play is the genre model. Although it is true that genres do not facilitate well-defined taxonomies (see, for example, the review in Arsenault, 2009), Zammitto (2010) has shown it is possible to leverage them for modeling purposes.

A number of authors have proposed formal solutions to the problem of describing games (Church, 1999; Costikyan, 2002; Zagal, Mateas, Fernández-Vara, Hochhalter, & Lichti, 2005). Björk and Holopainen (2005) proposed expressing formalized gameplay features as game design patterns, also related to the concept of game grammars (Cook, 2006; Koster, 2005). Other authors have proposed the use of game metrics to explore the play space (Cousins, 2005; Drachen, Canossa, & Yannakakis, 2009; Yannakakis & Maragoudakis, 2005). The main idea behind all of these is to generalize and describe recurring design problems and their solutions. These approaches allow us to use patterns to trace measurable outcomes (such as jumping time in a platform game), which describe and quantify game designs. Metrics are an important part of the PPAX framework’s approach to deriving insight into player preferences and abilities, providing a data structure to contextualize the lower level data from physiology.

Personality and Player Profiling

The third important part of the PPAX framework is to accommodate analysis without assuming that all players have the same preferences. By allowing the integration of player profiling, the framework is free to abandon a one-size-fits-all approach and to take individual differences into account as parameters. In one sense, the idea of a player profile springs from observations of game design patterns, as it appears that players may have preferences for certain patterns, whether due to aptitude or taste.

Hartmann and Klimmt (2006) reviewed research on personality factors and game preferences, arguing for “the importance of distinguishing between game types: the landscape of computer games appears to be too diverse to hypothesize about the role of personality factors in gaming in general” (p. 123). Bartle (1996) made the well-known early attempt to categorize player types based on such observations, grouping players by their tendency to behave in a certain way during play. The typology originated in a very specific context (early Multi-User Dungeons or MMOs), but has also been applied to more general contexts. Yee (2006) continued this study line with his research on motivations to play (modern) MMOs and used a factor-analytic approach to produce a more trait-based categorization instead of a straight typology. Bateman, Lowenhaupt, and Nacke (2011) analyzed the studies that created associations between psychological personality testing and gaming. They found the typology approach to be inherently weak, and instead called for specialized trait models of play to profile players. They reported on building on their trait model (BrainHex) and theorized on how neurobiological factors are related.

In addition to player typologies, player profiling has been built on factor-analytic psychological theories or models. In this approach, researchers have sought to find associations between the well-established personality theory, Big Five or Five-Factor Model, and gaming (e.g., Zammitto, 2010; see Goldberg, 1993), albeit with contradictory and limited results. A particularly interesting study converted personality questionnaire items into in-game decisions, in order to automatically assess the personality trait from player actions (van Lankveld, Schreurs, & Spronck, 2009). Also relevant has been the self-determination theory of motivation (Deci & Ryan, 1985), which has resulted in comprehensive empirical investigations (Ryan, Rigby, & Przybylski, 2006; Tamborini, Bowman, Eden, Grizzard, & Organ, 2010) and a game motivation questionnaire (Lafrenière, Verner-Filion, & Vallerand, 2012). Tekofsky, Spronck, Plaat, van den Herik, and Broersen (2013) also demonstrated the possibility of assessing player personality directly from gameplay data by analyzing game behavior.

Although trait models can be useful approximations, some results also show that another factor should be considered: people play differently in different situations and at different times (Kallio, Mäyrä, & Kaipainen, 2010). It seems likely that on a general level players’ preferences do not arise only from stable traits, but also from player states that change from day to day and moment to moment. A comprehensive framework should be able to take into account both the more stable traits and the more transitory state factors, related to psychological concept of moods that correspond to the positive-negative valence dimension (see, for example, Gendolla, 2000).

Without using experimental conditions to illustrate the differences that depend on them, trait and state factors are more like the parameters to a model: They change over time at rates depending on their stability, and their representation in the model depends on the level of approximation used. In PPAX, player profiling instruments are used as instantiating parameters to a machine learning model of player psychophysiology and game events.

Machine Learning and Psychophysiology

In supervised learning, the algorithm learns classification rules based on examples of an existing compression scheme (i.e., categorically labeled data) given in a “training set.” By contrast, unsupervised learning seeks a novel compression of the data by exploiting whatever structural similarities of redundancy exist (assuming the algorithm has been properly configured to fit the data). In short, supervised learning requires a priori knowledge of what the classification should be and only lets the algorithm label the data with these predetermined classes; unsupervised learning generates new information by creating classes according to the data itself. Thus, given that the typologies are often basically categories built from qualitative data (such as self-report), they do not lend themselves well to unsupervised learning. This further suggests the superiority of trait-continuums over typologies.

In terms of the machine learning background of player profiling, some articles have proposed using personality as a factor in their models (e.g., Charles et al., 2005), although real implementations and results are more scarce. Previous work has reported some success in classifying players by a typology of playing styles (Cowley, Charles, Black, & Hickey, 2013), indicating that, to some degree, players do play in a characteristic way by which they can be distinguished. However, player type classification tends to require supervised learning methods.

Learning from psychophysiology is affected by some of the same issues. Physiological data can be labeled by indices drawn from the literature, such as the link between EDA and arousal. However, such data are always relative (the absolute electrical potentials differ from person to person and even from day to day), therefore it does not lend well to precise labeling and is not ideal for supervised methods, because approximate labels lead to inaccurate generalizations. The same problem arises in unsupervised learning when learned concepts shift, because this discovered compression must be interpreted somehow, and only approximate tools exist to aid interpretation. Nevertheless, Pagulayan, Keeker, Wixon, Romero, and Fuller (2003) suggested that games evaluation should be based on player emotions instead of performance metrics. Indeed, a number of papers describe recognizing emotional or mental states by learning from psychophysiological data using computational methods (e.g., Fantato, Cowley, & Ravaja, 2013; Mandryk & Atkins, 2007; Yannakakis & Hallam, 2008).

Liu, Agrawal, Sarkar, and Chen (2009) used EMG, EDA, and cardiac activity signals to automatically recognize anxiety, boredom, engagement, and frustration. Based on anxiety detection, they designed a dynamic difficulty adjustment (DDA) system, which resulted in a slight improvement in player performance and self-reported experience compared with a performance-based DDA. Some efforts have been made to combine a larger number of signals. Chanel among others (Chanel, Rebetez, Bétrancourt, & Pun, 2011; Tijs, Brokken, & Ijsselsteijn, 2008) have described DDA systems using facial EMG, EDA, cardiac indices, respiration, and other signals to discriminate between boredom, anxiety, and engagement. The approaches vary on how they attempt to capture and model feelings, partly depending on what components are deemed relevant for assessing gameplay. For example, Mandryk and Atkins (2007) described a fuzzy logic method to process EMG, EDA, and HR signals into valence and arousal levels, which in turn were modeled into fun, challenge, boredom, frustration, and excitement.

Some work has linked events and their predicted emotional reactions in a proactive manner. McQuiggan, Lee, and Lester (2006) used machine learning for predicting EDA and cardiac signals from changes in the virtual environment (n = 20). Finally, Yannakakis and others (Yannakakis & Hallam, 2008; Yannakakis, Martínez, & Jhala, 2010) employed affective modeling for predicting children’s play preferences based on EDA and cardiac responses during earlier instances of play, and later used a computational model of affective states (based on the EDA and cardiac responses) to investigate effects of camera control.

Toward a New Approach

Emotion recognition and/or DDA based on physiological signals have also been recently addressed by major game companies. Valve presented their experiments and positive results using biofeedback at the Game Developers’ Conference 2011 (Ambinder, 2011). However, the work on automatically classifying emotion lacks usefulness for design evaluations for two reasons. First, the scope of emotion models for assessing responses to varying difficulty is not enough to capture the experiences of playing. Existing approaches provide fragmentary experiential categories, selectively covering only a few experience aspects in a supervised approach. Second, more critically, the findings of different studies are not easily consolidated into a greater framework. In a review on emotion recognition, Arroyo and Romano (2010) similarly contended that “most of the emotion recognition systems follow an ad hoc strategy ([i.e. such systems are] . . . specific . . . non-generalizable . . . not easily adaptable).” (p. 17)

The major lack in the field is an integrative approach to unsupervised classification of physiological signals, which would not be limited by the state-based labels for emotion required in supervised learning. Although the theoretical and empirical basis of the valence-arousal model (or positive/negative activation) is solid, the fact remains that outside very specific events that elicit clearly recognizable responses, the physiological activity in general is not so easily classified in terms of emotional states; especially given the many factors influencing physiological processes with complex multimodal stimuli such as digital games. An unsupervised approach permits learning of continuous representations, and thus the creation of novel information.

The PPAX framework provides a theoretical basis for the integration of psychophysiology, profiling, and machine learning in a structured manner. The empirical part of this article applies the PPAX framework in practice, using machine learning to cluster game events, and psychophysiological data. The aim is to create an open framework that can take in data from various sources. Our approach differs from previous work primarily by using unsupervised learning to examine the user experience in a non-specific fashion (in contrast to predetermined application contexts).

Context and Parameters

The outermost level of PPAX, labeled “Player” and “Game” in Figure 1, is represented by contextualizing categorical data, which are gathered using existing self-report or player profiling instruments and existing taxonomies of games. In an application of PPAX, these layers will be instantiated as parameters to the model.

Personality

In order to accommodate individual response patterns, we need a way to incorporate player-specific information into the framework. We intentionally define player profiling as widely as possible: It accepts any validated source of information providing actionable descriptions of the player. Player descriptions include both long-term (more or less fixed) traits, as well as state-based descriptors, including the game situation, the socio-spatial context (De Kort & Ijsselsteijn, 2008), and the player’s point of view. The profiling further distinguishes between domain-specific personality and general personality traits. Domain-specific personality descriptors allow using play-specific information that is highly contextualized. General personality descriptors offer a connection point to personality-salient theories within psychological research. Thus, we have four profiling components to which we assign specific instruments.

Trait-domain

The Game Engagement Questionnaire (GEnQ; Brockmyer et al., 2009) is domain-specific and captures traits: It scores the player on their likelihood to be engaged in gameplay in general. The domain-specific trait model BrainHex (Bateman et al., 2011) describes play preferences and scores the player on seven categories of play preference: Achiever, Conqueror, Daredevil, Mastermind, Seeker, Socializer, and Survivor. Finally, the Gaming Motivation Scale (GAMS; Lafrenière et al., 2012) measures trait intrinsic motivation in the domain of games, as well as lack of motivation and relevant forms of self-regulation.

State-domain

The Game Experience Questionnaire (iGEQ; Ijsselsteijn, de Kort, & Poels, 2008) was designed to assess a broad range of player experience constructs based on self-report, which allow an estimate of the player’s response to the game, including positive and negative valence, flow (Csikszentmihalyi, 1975), immersion, competence, tension, and challenge. In addition, the Perceived Competence for Learning Scale (PCS) is derived from self-determination theory and provides a short-form measure of subjective ability to learn, which can be adapted to the domain of games.

Trait-general

Self-Determination Theory (SDT) provides instruments to assess trait intrinsic motivation, applied to games by Tamborini et al. (2010). The BIS/BAS Questionnaire measures the player’s general tendency to activation or inhibition of behavior. The Big 5 factor model (Goldberg, 1993) provides needed background on the player’s personality, that is, the developed expression of their temperament.

State-general

Within the player-state assessment, we must distinguish between baseline factors that give rise to player experience and outcomes that arise from play. For baseline state assessment, we use the Karolinska Sleepiness Scale (KSS; Kaida et al., 2006), which has a good agreement with ability to concentrate and mood; also we measure the spontaneous blink rate, which indexes dopamine and thus capacity of low-level attention (Slagter et al., 2012). Outcome general state assessments include the Self-Assessment Manikin (SAM; Bradley & Lang, 1994) that captures subjective feelings of excitement/arousal, positive/negative valence, and dominance; and the NASA Task Load Index (TLX; Hart & Staveland, 1988) for players to report their feelings of demand, effort, and associated frustration.

Figure 2 shows a simple schematic of this profiling approach based on existing instruments. Currently, the majority of instruments are self-report; however, the model accommodates features drawn from other sources as well (recorded play behavior, etc.)

OPEN IN VIEWER

Figure 2.

A simple schematic for player profiling and the hierarchy of various sources of information.

Design patterns

As with player profiling, the game can also be approximated by a loose structuring, illustrated in Figure 3. This schema is a simple way to visualize the elements of a game and how they lead from the player to the design and play patterns. It does not attempt to describe “what is a game,” since that is a deep and out-of-scope topic.

OPEN IN VIEWER

Figure 3.

Game structure.

On the highest level, the model differentiates between genre models or game types. Between genres, games can differ not only by game mechanics and style, but genre generally also influences (via marketing, player segment targeting, or technical adaptation/systems preference) the player base involved with the game and the expectations they project on the game. Genre is a loose construct without hard distinctions between classes; nevertheless, it is a valid starting point as some predictive power is still possible. Zammitto (2010) showed that up to 7.5% of players’ preferences for genre could be predicted by a model based on Big 5 personality traits. Thus, to tie in with our use of Big 5 in player profiling, the Zammitto model of genres is the first data point in the PPAX game structuring.

When interpreting games as formal structures that create the patterns of play, we are inspired by the aforementioned Björk and Holopainen’s (2005) work on game design patterns as well as the design pattern approach (that Björk and Holopainen used as an inspiration) first introduced by Alexander, Ishikawa, and Silverstein (1977). Within a game, we can distinguish between patterns of play (or gaming) and patterns of design. This differentiation is particularly pertinent when we want to examine the realization, effects, and quality of particular game design solutions by way of observing actual play. If players play as the designer intended, actual play (event sequences) will match design structures. In these cases, the play data can be used to verify whether the responses (experience) match design intentions. However, it is also possible that players find ways of playing the game that were unanticipated by the designer or restrict their actions to a very small subset of the designed possibilities; such subversive play can have impacts on the final experience that cannot be anticipated by looking only at the designed structures. The empirical example study in “Example on Application of the Framework” section utilizes the design/play division to structure data.

The computational analysis of recorded game events contributes to a general model of the game; clustered by unsupervised machine learning the events feed into the patterns of play. These then can be related to patterns of design, extracted using Björk and Holopainen’s (2005) game design patterns.

Patterns of play are instantiated as gameplay metrics. These capture moment-to-moment features of player activity from the game log and allow further inferences to be built, which can give insight into the player’s abilities and preferences. It is a quite game-specific process to derive metrics and features of gameplay; however, existing methods can guide the process, two of which are relied upon here. Drachen et al. (2009) described a powerful algorithmic approach to learning from simple metrics, whereas Cowley, Charles, et al. (2013) used simpler machine learning, but described a more in-depth approach to deriving features of gameplay using expert knowledge. Whereas the latter approach requires specialized knowledge of the game, the former approach puts the knowledge requirement in interpretation of algorithm output.

Play patterns elicit reaction patterns in players, which in PPAX are represented by psychophysiological metrics, as described next.

Play/Reaction Patterns

The play patterns represent the predictable, interpretable qualities of play as expressed by players. In the practical application of PPAX, this takes the form of a model built on top of the fundamental sources of observable real-time data, psychophysiology, and the game log. The model is parameterized by profiling and game structure data.

In the current iteration, player physiology is a keystone of the PPAX approach, since we assume that it lends the richest real-time data sourcing and analysis opportunities. As stated, the signals measured are EDA, ECG, and EMG.

Psychophysiological models of play have been proposed before (e.g., Bateman & Nacke, 2010); the PPAX approach is novel in that it draws on a rich body of prior studies by the authors themselves, but applies new forms of analysis to advance the method. At heart, the approach uses EDA, ECG, and EMG signals mapped as indices of experience constructs, based on robust and reliable findings. For instance, it is known that tonic electrodermal activity can act as an index of arousal or the activity of facial muscles such as Zygomaticus major can be taken as an index of valence (Lang, 1995). The combination of these electrical potential measurements, via some signal processing functionality as shown in Figure 4 (Arroyo & Romano, 2010; Mandryk & Atkins, 2007), is used to locate the player within the emotional circumplex described by the valence-arousal model.

OPEN IN VIEWER

Figure 4.

Measurements of facial muscles and skin conductance can index the arousal and valence of a person (anatomical drawings adapted from Gray, 2000, where gray circles indicate approximate electrode placement locations).

Higher order inferences can be made: For instance, the combination of medium high arousal with low heart rate variability can be taken as a sign of learning in the right context (Cowley, Ravaja, et al., 2013). The role of context in making inferences is important to our approach, thus we aim to incorporate context into the computational methods. The framework proposes how to do this by drawing on the contextual knowledge in play patterns, game design patterns, and player profiling to parameterize machine learning algorithms.

Framework

With the parts of the PPAX framework described, it is worth pausing to consider how the whole arises from the sum of these parts.

The PPAX framework aims to allow predictive analytics by computational methods. The end result of PPAX is thus an interlocking framework of theoretical approaches with methods of application. It suggests an approach to describe experiences from a particular game within a holistic structure of empirical data. The ultimate aim is a theoretically based scaffold for structuring unsupervised machine learning of hierarchical dependencies among data sources and levels of detail. Discovery of these dependencies relies on appropriate application of machine learning. The PPAX approach faces two disadvantages, but at the same time derives two main benefits from the characteristics of the learning-problem domain.

The subjectivity of the experience of play means that measurements cannot be easily validated or verified. The mutability of the play experience means that because players learn from games at a rapid rate, the experience they have in one session may alter significantly within another session. Thus, player learning needs to be accounted for by some refinement of machine learning.

On the other hand, the learning problem confers two advantages. First, homomorphic mapping means that with rich data sets on both play and game sides, it may not be necessary to compress the data space to over-simplified categories. Second, multiple solutions exist: Because the patterns of play and reaction are represented by game log activity and psychophysiology, it is valid to assume that multiple domains will be used to learn the same concepts (Bedek et al., 2011). This not only makes the whole approach more robust, but also facilitates a form of cross-validation.

We envision state and trait data as both being fed to the model as parameters, especially when the model acts on much shorter timescales than the profiling instruments. If the model is to remain accurate even over the timescale of state constructs (e.g., to model long-term play behavior), the model should be able to handle potential changes to the mappings that it learns between play patterns and game patterns, known as concept drift (Black & Hickey, 1999). For example, players may receive a thrill from winning in a given scenario, an emotional reaction that diminishes over time. It is anticipated that future iterations of the framework will incorporate methods to manage concept drift, in which case state changes can be modeled as continuous instead of parametric. This would allow the model to accurately represent the play experience even over mid-scale time frames as the players learn the game and their state responses to it change.

For this reason, and because the supporting research on player profiling and game structure is evolving, PPAX and its components may also evolve in future iterations. To test the current iteration, we have conducted an evaluation study that investigates the relationship between event clusters and game design patterns. The following section describes how the framework performs using unsupervised classification to identify experience patterns.

Experimental Procedure

The EDA, EMG, and ECG psychophysiological experience data and gameplay video were collected in a laboratory setting. The collection consisted of 3 hours of play from 36 (25 males and 11 females) Finnish undergraduates, 5 minutes per participant. The following sections describe the procedure and specific approach for gathering the data. For a more detailed description of the experimental setup, refer to Ravaja, Timo, Jani, Kari, and Mikko (2005).

The event scoring was defined by a group of experts (including experts on game studies and on psychophysiology, plus the assistant who knew the particular game) based on the videos and gameplay. Two main criteria were used in selecting events which were considered relevant enough to score: (a) Everything that would plausibly cause a notable psychophysiological reaction and (b) everything that either promotes or hinders success/progress in the game.

The following game events were coded:

The exact onset times of the predefined game events were determined by examining the played games, frame by frame; onset times were then imported into Psylab7 analysis software.

Method of Analysis

The psychophysiological measurements that we dealt with in this study are related to several pattern groups. However, some patterns can be realized only over longer periods of play or are hard to map precisely to individual actions. Although many of these pattern groups are interesting for design, it seems necessary to first establish the connection between play experience and design on a pattern group with the most straightforward mapping to identifiable game events. Therefore, for our current purposes, the pattern group “Actions and Event Patterns” from Björk and Holopainen (2005, p. 145) has been chosen as the basis of analysis.

The patterns in “Actions and Event Patterns” focus on the actions available to the player, how they relate to changes in the game state, and how they relate to the goals of the players. We will apply this set of patterns to examine how categories retrieved by computational analysis correspond to design features.

Results

Machine learning

When analyzing event data, the clustering found three very distinct play patterns (Figure 6):

15. Banana—5. Seeing Stars—6. Sparkles—4. Small Fall.

19. Out of Control—18. Bump—20. In Control.

7. Home Stretch—9. Finish—10. Time Bonus—12. Replay.

OPEN IN VIEWER

Figure 6.

The three meta-events clusters showing how each event cluster relates to the rule set generated.

The last cluster, occurring in the order {7. Home Stretch—9. Finish—10. Time Bonus—12. Replay}, happens automatically when you reach the goal of the level without bumping into goal posts. The two others are more interesting. The cluster {15. Banana—5. Seeing Stars—6. Sparkles—4. Small Fall} describes risk-taking behavior: More bananas are available if you take high speed routes with small platforms, and therefore pursuing bananas increases the risk of falling. When you go really fast, sparkles represent your high speed; when you jump, the character will see stars. When you steer correctly, you are able to reach bananas that are placed in difficult-to-reach locations. ³

Running the k-means algorithm on a data set consisting of psychophysiological data and game events showed distinct upward and down trends in EDA, from events around the start of a race (Start—Increased arousal) and events where the player died (Big Fall—Decreased arousal). These clusters matched the events with an accuracy of between 70% and 80%. ⁴ In other words, in four out of five cases, the algorithm can correctly distinguish between these two events by the psychophysiological data alone.

Discussion

Two interesting results were found by applying the machine learning approach described. First, we were able to automatically detect relevant clusters of game events, allowing us to determine interesting patterns in the gameplay and widen the analytical scope beyond single events. Second, we were able to cluster the psychophysiological data so that the algorithm was able to separate two different events into two clusters with accuracy of 80%, based only on the physiological responses of the players during these events. The results of unsupervised learning in the evaluation study are in line with the earlier results found for SUPER MONKEY BALL 2 reported in Ravaja et al. (2005).

These results indicate the potential for connecting psychophysiology to formal approaches in game studies. Concretely, on the basis of this, we have evidence to establish links between two of the levels in Figure 1: game event to player reaction and play pattern to reaction pattern. Although far from conclusive, this is however encouraging, since the quality of the machine learning results is quite high. The results of machine learning could be used to revise the game design patterns to match better empirical data.

However, our machine learning found only a few event patterns out of all possible patterns. Some less frequent, but relevant event patterns were not found, such as {19. Out of Control—13. Big Fall—25. Fall Replay}. Moreover, machine learning is less useful to distinguish emotional reactions when the reactions are indistinct, unless very large data sets are used to improve signal-to-noise ratio.

We did not include personality data in our example study, in order to focus on the first level of the PPAX within this introductory article. However, personality dimensions could be included as factors to the machine learning algorithm. With such future work, we expect to be able to see whether or not their personality influences how players experience event patterns. Based especially on Cowley, Charles, et al. (2013), we have grounds to expect that PPAX’s modeling can be improved by personality profiling; the test of this remains for future work.

Lankoski and Björk (2011) pointed out that psychological implications are underdeveloped in the pattern collection by Björk and Holopainen (2005). Our method extends Lankoski and Björk’s argument in describing psychological implications of gameplay patterns, by including psychology and psychophysiology in a data-driven way. Even though connecting gameplay event patterns to design patterns cannot be performed automatically, the event patterns can be used to refine the design patterns: For example, what kinds of gameplay does a design pattern imply or what kinds of playing experiences does the event sequence of a design pattern relate to.

In terms of practical game design, it is important to understand the implications of the design choices that are made. For example, Adams and Dormans (2012) argued that patterns can be used in the brainstorming stage of the design. PPAX is intended to connect game experience research to practical design work. As argued above, the current understanding of how formal game features relate to the playing experience is still hypothetical. However, our evaluation study indicates that we can connect the player’s psychophysiology, as an empirical indicator of their play experience, to patterns of play and even design patterns. When this kind of data can be used to pre-screen the design patterns used in brainstorming, it will be possible to focus on the patterns that are likely to contribute toward the desired kind of play experience (Kreimer, 2002). It may require some efforts to practically implement PPAX, so improving ease-of-use would be important for future work.