Procedural maze level generation with evolutionary cellular automata

Maze running games represent a popular genre of video games and the design of playable mazes provides an interesting research challenge in procedural content generation for computational intelligence research in games. In this paper, we attack the problem of creating playable mazes by using genetic algorithms to evolve cellular automata rules that lead to playable mazes. More specifically, a fixed number of evolved-rule applications generates maze like patterns on a cellular automata grid and a region merging algorithm then generates the final, playable maze. Since maze path lengths correlate with maze playability, the genetic algorithm searches for cellular automata rules that lead to longer path lengths. Results from two types of cellular automata and three different fitness functions of path length show that our approach results in a variety of interesting, playable mazes with longer path lengths and complex paths.