Learning How Design Choices Impact Gameplay Behavior

Designers structure a game to provide a desired range of player behaviors: the play space of a game. Any given game is one instance from a design space of alternatives. Navigating a design space to achieve a designer's goals requires knowledge of how design choices shape the play space in a game. We present algorithms to automatically measure play patterns using statistical models that predict how design choices alter player behavior. We present Monte Carlo tree search as a way to sample behaviors from a play space, action metrics to automate play space measurement, and predictive modeling techniques to model design spaces. We demonstrate these techniques in two simplified, perfect information, adversarial game domains based on Scrabble and Hearthstone showing their use for automated design space modeling.