A Comparative Evaluation of Procedural Level Generators in the Mario AI Framework

Evaluation is an open problem in procedural content generation research. The eld is now in a state where there is a glut of content generators, each serving dierent purposes and using a variety of techniques. It is dicult to understand, quantitatively or qualitatively, what makes one generator dierent from another in terms of its output. To remedy this, we have conducted a large-scale comparative evaluation of level generators for the Mario AI Benchmark, a research-friendly clone of the classic platform game Super Mario Bros. In all, we compare the output of seven dierent level generators from the literature, based on dierent algorithmic methods, plus the levels from the original Super Mario Bros game. To compare them, we have dened six expressivity metrics, of which two are novel contributions in this paper. These metrics are shown to provide interestingly dierent characterizations of the level generators. The results presented in this paper, and the accompanying source code, is meant to become a benchmark against which to test new level generators and expressivity metrics.