Side-Scrolling Platform Game Levels Reachability Repair Method and Its Applications to Super Mario Bros

Jin Zhang, Tianhan Gao, Qingwei Mi

Abstract: Applying machine learning to game level design has become popular increasingly since it requires little human expertise. However, levels generated by machine learning may be unreachable. The traditional way to repair level unreachability is developing specific rule-based repairers to patch bugs. Due to the complexity of programming these constraints, we propose a new method to fix the reach ability of the level for the side-scrolling platform game in this paper. First, we fixed the logic errors in the level. Then we used the level break agent to locate the tiles that made the level unreachable. Finally, we used the deep learning model to replace the tiles to make the level reachable. The proposed method is to prove effective in the case study of fixing Super Mario Bros levels.