A deep generative approach to personalized super mario level design.

Area of Science:

• Computer Science
• Artificial Intelligence
• Game Development

Background:

• Procedural content generation (PCG) faces challenges in matching game level difficulty to player skill, impacting engagement.
• Deep generative models, particularly Generative Adversarial Networks (GANs), offer potential for automated, scalable level synthesis.
• The effectiveness of different GAN architectures for skill-conditioned and personalized level generation requires further investigation.

Purpose of the Study:

• To investigate and compare the effectiveness of five GAN architectures for skill-conditioned procedural content generation.
• To evaluate GANs' ability to generate personalized game levels tailored to individual player skill levels.
• To identify optimal GAN architectures for balancing level diversity, playability, and training stability in adaptive PCG.

Main Methods:

• Utilized five GAN architectures: U-Net GAN, StyleGAN, Deep Convolutional GAN (DCGAN), ResNet-GAN, and Spectral Normalization GAN (SN-GAN).
• Employed Spectral Clustering to group player behavior into distinct skill levels, using these labels as conditioning signals for GANs.
• Evaluated models using quantitative metrics: tile distribution entropy, diversity score, discriminator accuracy, generation speed, and pairwise Hamming distance.

Main Results:

• Generated levels adapted to increasing player skill, showing fewer deaths and shorter completion times.
• Higher player skill levels encouraged more complex in-game interactions.
• ResNet-based and U-Net-style GANs demonstrated the best performance regarding diversity, playability, and training stability.
• Spectral clustering effectively enabled skill-conditioning for personalized level generation.

Conclusions:

• GANs, particularly ResNet and U-Net variants, are effective for skill-conditioned procedural content generation.
• The proposed spectral clustering approach successfully generates personalized game levels aligned with player abilities.
• This research supports the advancement of adaptive procedural content generation systems for enhanced player engagement.