A Hybrid Approach Combining Generative Adversarial Networks and Graph Traversal Algorithms for Procedurally Generating Maps in Games
Soham Chavan, Sachi Nandan Mohanty · 2025
Procedural content generation is a core component of video games. Conventionally, procedural content generation for maps in games is done using algorithms that are finetuned using strict guidelines that make it really game-specific and are also bound to require a lot of effort in the process of finetuning, and making the rules based upon the design and goal of the game. While these classical methods have proven to be robust and provide structural integrity and overall connectivity in the game maps, they lack at providing overall diversity and adaptability without either extensive manual tuning or making a good amount of content by hand and tying it up with algorithms as is in the case of games like Dead Cells and Spelunky. On the other hand, there have been recent approaches which suggest that GANs could procedurally generate maps using image generation, but they suffer from high computational overhead, and at times fail to ensure proper connectivity in the maps, i.e. lack the robustness of algorithms. In this paper, we propose a hybrid system that integrates Generative Adversarial Networks(GANs) with path-finding algorithms used for dungeon generation. Unlike fully neural approaches that employ a neural network to generate an entire dungeon from scratch end-to-end, our method uses the neural networks only from producing a weighted-grid that encodes traversability and structural bias while delegating path finding to lightweight algorithms such as A*. This allows the system to retain the controllability, efficiency and overall map connectivity that algorithms bring in while also having the advantage of diversity and adaptability brought in by learned structures. The discriminator in this approach grades the generator either adaptively using evolutionary learning principles or uses a data set of weighted grids to make it so that the dungeon maps generated are well routed and diverse.