A Hybrid Approach for Repairing Logic Errors in GAN-Generated Game Levels

With the development of deep learning, an increasing number of deep learning techniques are being employed in game level generation, which reduces the need for domain knowledge. Among them, level generation methods based on Generative Adversarial Networks (GANs) have garnered significant attention. The levels generated by GANs, however, exhibit logic errors that may be attributed to the inherent structure of the GAN model. To solve this problem, we propose a hybrid method based on the greedy algorithm and deep learning to repair logic errors in levels. The method comprises three fundamental components: candidate logic error detection, tile selection, and tile repair. The entire repair process is formed by iteratively executing these three components. The components are designed based on the analysis of the characteristics of the game level and the level constraint model, which is trained by deep learning on the original levels. Additionally, we establish five distinct repair strategies to control the repair process. We test our method on two game levels, Super Mario Bros. (SMB) and Lode Runner (LR), where the results show that our method is effective in repairing the levels’ logic errors.