Player Modelling for Cursor-Driven Games

Cursor-driven games compose the vast majority of digital games on the market. For humans and for agents that control a virtual cursor, precise cursor control is frequently the primary determining factor of skill. Past efforts to control computer difficulty by adding rule-based distortions to the virtual cursor are inflexible and often based on coarse difficulty settings. We propose a framework for modelling player-controlled cursor accuracy by abstracting the twodimensional cursor plane into a one-dimensional distance measure from a target’s center. We then infer a probability distribution from player data using a novel online learning technique. The proposed learner is shown to be more computationally efficient and robust to a range of player types than comparable methods. The player modelling techniques we describe can potentially be applied to developing interesting cursor-driven opponents, or providing players with feedback as to when and where their cursor control is weak.