Procedural Level Generation for <i>Sokoban</i> via Deep Learning: An Experimental Study

Deep learning for procedural level generation has been explored in many recent works, however, experimental comparisons with previous works are rare and usually limited to the work they extend upon. The goal of this article is conduct an study on four deep generators Sokoban explore their strengths weaknesses. methods will be bootstrapping conditional generative models, controllable uncontrollable content via reinforcement (PCGRL), playing networks. We propose some modifications either adapt task or improve performance. For method, we using diversity sampling solution diversity, training auxiliary targets enhance models’ quality conditions from Gaussian mixture models (GMMs) sample quality. results show that generated solutions more diverse by at least 16% when used during training. It also shows GMMs can increase playability percentage. In our experiments, PCGRL superior while bootstrapped long-short term memory exhibit control confusion.