Quatrosem

Analyzing Player Behavior in Pacman using Feature-driven Decision Theoretic Predictive Modeling
Ben Cowley, Darryl Charles, Michaela Black and Ray Hickey

Abstract!We describe the results of a modeling methodology that combines the formal choice-system representation of decision theory with a human player-focused description of the behavioral features of game play in Pacman. This predictive player modeler addresses issues raised in previous work [1] and [2], to produce reliable accuracy. This paper focuses on using player-centric knowledge to reason about player behavior, utilizing a set of features which describe game-play to obtain quantitative data corresponding to qualitative behavioral concepts.

I. INTRODUCTION With this paper we present the results from a novel realtime player modeling algorithm that attempts to predict each move that a human player makes in Pacman. These predictions are compared with actual player activity. We suggest that the results of this comparison form a model, and after enough instances of play, analysis of this model should form a valuable picture of player behavior. Whether they follow the predicted path or deviate from it, some insight should still be possible. Predictive player modeling works by considering the playerEs in-game goals as equivalent to some target function of the game state and calculating this function using observed player data [3], [4], [1]. Consideration of the playerEs goals or utilities is central to this; broadly speaking the predictive modeler will use real-time observation of play habits to determine player preference for competing potential game states. In [5] we discussed how games can be described using formal or semi-formal structured specification systems. By extension of this approach, we can use a decision theoretic formulation to model the short-term elements of the game that are important to a player P the playerEs utilities P and predict the playerEs next actions. That is, we can calculate a