Generative Adversarial Networks for Procedural Content Generation with Aesthetic Constraints

Abstract

This paper explores the application of Generative Adversarial Networks (GANs) to procedural content generation (PCG) with explicit aesthetic constraints. Traditional PCG methods often rely on predefined rules and algorithms, which can lead to repetitive or predictable outcomes. We propose a novel framework utilizing a GAN architecture to address this limitation. The generator network produces content according to specified procedural rules, while the discriminator network learns to distinguish between aesthetically pleasing and displeasing content, effectively incorporating subjective aesthetic preferences into the generation process. We formulate this as a minimax game, where the generator attempts to fool the discriminator, and the discriminator strives to correctly identify aesthetically flawed content. Mathematical formulations detail the loss functions for both networks, emphasizing the adversarial training process. The core claim is that GANs can be trained to generate content that satisfies both procedural rules and subjective aesthetic criteria. This approach offers a pathway to creating more diverse and engaging content, particularly in domains such as game level design, music composition, and digital art. The paper outlines the architecture, training methodology, and potential applications of this GAN-based PCG system.