Article Preview
TopIntroduction
Player modeling has been attracting the interest of game design and development experts for several years, as a means to increase player satisfaction and immersion. According to the inclusive reviews in (Smith et al., 2011) and (Hooshyar et al., 2018), modeling techniques vary from empirical (data-driven) (Thue et al., 2007; Thurau & Bauckhage, 2010; Roberts et al., 2007; Geisler, 2002; Drachen et al., 2013), where the application of machine learning or statistical analysis to gaming data enables predictions of playing styles, to theoretical (i.e. analytical), mostly applicable to board-like games, where search and optimization techniques are used to determine the moves towards the best outcome (Bellman, 1965). The term ‘play style’ indicates the manner in which each player behaves while playing, i.e. the choices he makes, his reactions, his response time etc.
Regarding empirical approaches to player modeling, various learning techniques have been experimented with; supervised, like support vector machines for predicting difficulty adjustment (Missura & Gaertner, 2009), Bayesian networks for classification (He et al., 2008), statistical analysis of the distribution of player actions (Thawonmas & Ho, 2007), and unsupervised, like self-organizing maps (Drachen et al., 2009), reinforcement learning (Kang & Tan, 2010), transfer learning (Shahine & Banerjee, 2007) and preference learning (Yannakakis et al., 2009). Supervised techniques (stand-alone or in combination with unsupervised approaches) have been gaining in popularity (Bisson et al., 2015; Luo et al., 2016; Min et al., 2016; Tamassia et al., 2016; Falakmasir et al., 2016; Gao et al., 2016) during the last three-four years, compared to purely unsupervised approaches (Drachen et al., 2009; Anagnostou & Maragoudakis, 2009; Cowley et al., 2014), mostly due to their improved performance. Dimensionality reduction techniques, other than self-organizing maps, have been experimented with for unsupervised modeling: Linear Discriminant Analysis has been applied to arcade-style as well as combat-style games (Gow et al., 2012), where match data are annotated with the players’ identity to enable the supervised application of Linear Discriminant Analysis, and then k-means clustering groups together players of the same gaming style.
All previous approaches use a limited number of high-level game and player features to perform modeling, that are game-dependent (vary from game to game, and a game expert is required to define them) and whose impact on the player model is to some extent a-priori sensed. High level features indicate directly and almost explicitly the game status. High-level features in combat-style games may, for instance, include the number of weapons obtained, the number of shots performed, the number of spare lives accumulated. The high-level features pose significant demands on knowledge resources, while they minimize expectations to extract new knowledge and unforeseen relations and dependencies between game and player features. Low level features, on the other hand, are features that describe the morphology of the game terrain, at specific time intervals, and the game status needs to be indirectly deducted.