This project combines unsupervised learning and generative modeling to create structured, playable dungeon mazes. Using K-Means clustering and a Variational Autoencoder (VAE), it learns spatial patterns from thousands of randomly generated grids and then samples new layouts from a compressed latent space.
Tags: unsupervised learning · generative modeling · VAE · K-Means
Core Pipeline -generate_dungeons() creates mazes with guaranteed entry → exit paths.
-KMeans clusters dungeons based on structural similarity.
-VAE learns a low-dimensional latent representation to generate new layouts.
-simulate_player() uses BFS (Breadth-First Search) to check if a generated dungeon is solvable.
-visualize_player_path() overlays the player’s exact solution path (if one exists)
-Applies key ML concepts: clustering, latent variables, probabilistic sampling
-Demonstrates generative design for game maps / intelligent systems
-BFS validation ensures outputs are not just realistic—but playable