This project explores the use of Generative Adversarial Networks (GANs) to generate playable and stylistically consistent Super Mario Bros-style platformer levels. By training on real Mario level data, the model learns implicit design patterns and produces new level segments with a balance of diversity and structure.
The implementation leverages deep learning techniques with TensorFlow and Keras to create a GAN architecture specifically designed for level generation. The discriminator network learns to distinguish between real Mario levels and generated ones, while the generator network learns to create increasingly realistic level layouts that fool the discriminator.
The project includes comprehensive evaluation metrics for generated levels, including playability analysis using the Mario AI Framework, structural coherence assessment, and visual similarity comparisons with original Nintendo levels. The generated levels maintain the classic Mario aesthetic while introducing novel design patterns and challenges.
This work demonstrates the potential of AI-driven procedural content generation in game development, offering insights into how machine learning can be applied to create engaging, varied, and balanced level designs for platformer games.
Navigate through the presentation slides to learn about the methodology and results:
View the complete research paper with detailed technical information and findings: