Generating automatic Mario levels with a Genetic Algorithm

Procedural Content Generation (PCG) is one of the key features of modern entertainment, showcased in a vast set of applications such as, animation, film, and of course many diverse genres of games. Using the fundamental elements of levels such as the environment, enemies, and player, procedural content generation has the capability to construct levels, maps, and even entire games. This thesis concentrates on the application of a Genetic Algorithm (GA) to autonomously create levels for 2-D platformer games, exemplified by games like Super Mario. These generated auto-levels are compared through randomization, parameter tweaks and level space restrictions. We successfully uncovered insights in PCG indicating that the generation of complex levels is intrinsically linked to the appropriate investment of generation time and a balanced utilization of essential level tiles. Notably, our findings highlight the importance of these factors in achieving optimal level design. To enhance the complexity of the generated levels, we have introduced a “least-block” fitness function. This novel approach not only sheds light on positive aspects but also identifies areas for improvement, distinguishing between cluttered and sparse generated levels.