Player modeling is, primarily, the study and use of artificial and computational intelligence techniques for the construction of computational models of player behavior, cognition and emotion, as well as other aspects beyond their interaction with a game (such as their personality and cultural background). Player modeling places an AI umbrella to the multidisciplinary intersection of the fields of user (player) modeling, affective computing, experimental psychology and human-computer interaction.
A. Model-based (Top-down) Approaches
According to a model-based or top-down 1 approach, a player model is built on a theoretical framework. As such, researchers follow the modus operandi of the humanities and social sciences, which hypothesize models to explain phenomena, usually followed by an empirical phase in which they experimentally determine to what extent the hypothesized models fit observations. Top-down approaches to player modeling may refer to emotional models derived from emotion theories (e.g., cognitive appraisal theory 2). Three examples are: (1) the emotional dimensions of arousal and valence 3, (2) Frome’s comprehensive model of emotional response to a single-player game 4, and (3) Russell’s circumplex model of affect 5, in which emotional manifestations are mapped directly to specific emotional states (e.g., an increased heart rate of a player may correspond to high arousal and therefore to excitement). Model-based approaches can also be inspired by a general theoretical framework of behavioral analysis and/or cognitive modeling such as usability theory 6, the belief-desire-intention (BDI) model, the cognitive theory by Ortony, Clore, & Collins 7, Skinner’s model 8, or Scherer’s theory 9. Moreover, theories about user affect exist that are driven by game design, such as Malone’s design components for ‘fun’ games 10 and Koster’s theory of ‘fun’ 11, as well as game-specific interpretations of Csikszentmihalyi’s concept of Flow 12.
B. Model-free (Bottom-up) Approaches
Model-free approaches refer to the construction of an unknown mapping (model) between (player) input and a player state representation. As such, model-free approaches follow the modus operandi of the exact sciences, in which observations are collected and analyzed to generate models without a strong initial assumption on what the model looks like or even what it captures. Player data and annotated player states are collected and used to derive the model. Classification, regression and preference learning techniques adopted from machine learning or statistical approaches are commonly used for the construction of a computational model. Data clustering is applied in cases where player states are not available.
In model-free approaches, we meet attempts to model and predict player actions and intentions 13, 14 as well as game data mining efforts to identify different behavioral playing patterns within a game 15, 16, 17. Model-free approaches are common for facial expression and head pose recognition since subjects are asked to annotate facial (or head pose) images of users with particular affective states in a crowd-sourcing fashion.
C. Hybrid Approaches
The space between a completely model-based and a completely model-free approach can be viewed as a continuum along which any player modeling approach might be placed. While a completely model-based approach relies solely on a theoretical framework that maps a player’s responses to game stimuli, a completely model-free approach assumes there is an unknown function between modalities of user input and player states that a machine learner (or a statistical model) may discover, but does not assume anything about the structure of this function. Relative to these extremes, the vast majority of the existing works on player modeling may be viewed as hybrids between the two ends of the spectrum, containing elements of both approaches.