The Fun Facets of Mario: Multifaceted Experience-Driven PCG via Reinforcement Learning

Ziqi Wang, Jialin Liu, Georgios N. Yannakakis

Abstract: The recently introduced EDRL framework approaches the experience-driven (ED) procedural generation of game content via a reinforcement learning (RL) perspective. EDRL has so far shown its effectiveness in generating novel platformer game levels endlessly in an online fashion. This paper extends the framework by integrating multiple facets of game creativity in the ED generation process. In particular, we employ EDRL on the creative facets of game level and gameplay design in Super Mario Bros. Inspired by Koster’s theory of fun, we formulate fun as moderate degrees of level or gameplay divergence and equip the algorithm with such reward functions. Moreover, we enable faster and more efficient game content generation through an episodic generative soft actor-critic algorithm. The resulting multifaceted EDRL is not only capable of generating fun levels efficiently, but it is also robust with respect to dissimilar playing styles and initial game level conditions.