Interactive Evolution and Exploration Within Latent Level-Design Space of Generative Adversarial Networks (2004.00151v1)

Abstract: Generative Adversarial Networks (GANs) are an emerging form of indirect encoding. The GAN is trained to induce a latent space on training data, and a real-valued evolutionary algorithm can search that latent space. Such Latent Variable Evolution (LVE) has recently been applied to game levels. However, it is hard for objective scores to capture level features that are appealing to players. Therefore, this paper introduces a tool for interactive LVE of tile-based levels for games. The tool also allows for direct exploration of the latent dimensions, and allows users to play discovered levels. The tool works for a variety of GAN models trained for both Super Mario Bros. and The Legend of Zelda, and is easily generalizable to other games. A user study shows that both the evolution and latent space exploration features are appreciated, with a slight preference for direct exploration, but combining these features allows users to discover even better levels. User feedback also indicates how this system could eventually grow into a commercial design tool, with the addition of a few enhancements.

• The paper presents an approach using GANs for interactive evolution and direct exploration of latent spaces to generate video game levels, addressing the challenge of evaluating subjective design quality.
• The system employs WGANs to create a latent space navigable via user selection for interactive evolution or direct slider manipulation for real-time design adjustment.
• A user study indicated a preference for direct latent space exploration, but combining both methods produced superior levels, highlighting the potential of AI-powered, user-driven content creation tools.

Interactive Evolution and Exploration Within Latent Level-Design Space of Generative Adversarial Networks

The paper presents a nuanced exploration into interactive evolution and latent space exploration using Generative Adversarial Networks (GANs) for generating video game levels, specifically targeting Super Mario Bros. and The Legend of Zelda. This research integrates GANs, a well-recognized indirect encoding method, which allows for Latent Variable Evolution (LVE). The primary focus is on facilitating user interaction to evolve or directly manipulate the latent space to create appealing, playable levels, addressing the inherent difficulty of quantifying subjective player satisfaction through objective metrics.

Methodology and System Design

The system employs WGANs, a variant known for stable training, trained on level data to induce a latent space, enabling the exploration and evolution of game levels. The highly compact architecture allows users to navigate through this space using two main approaches:

1. Interactive Evolution: Through a user selection mechanism, levels are iteratively evolved. The system presents a visual array of levels, allowing users to select preferred level designs to guide the evolutionary process. This method leverages human intuition and creativity to explore various level designs without preset objectives, accommodating dynamic user preferences over the design session.
2. Direct Latent Space Exploration: Users can manipulate level designs by directly adjusting the latent variables. This real-time adjustment, facilitated through a slider interface, gives users fine control over specific features without predefined mappings between latent variables and level elements.

User Study and Findings

A paper involving 22 participants found a slight user preference for direct exploration of latent space over evolutionary methods. However, the combination of both methods was generally regarded as leading to the creation of superior levels. This preference is indicative of the nuanced advantages provided by interactive design spaces, where both exploration and directed evolution can yield diverse and high-quality content.

Implications and Future Work

The system's current form highlights the potential of GANs to simplify the genotype-to-phenotype mapping for complex structures such as game levels. This interactive platform suggests a path forward where AI systems can power agile and user-driven content creation tools in game design. Such systems not only aid designers by offering vast exploratory spaces but can eventually streamline the creative process across various domains.

Future directions may include enhancing the intuitive control users have over specific level features—potentially through modified training regimes that more closely align latent dimensions with distinct level properties. Integrating direct level editing, potentially alongside latent vector adjustments, could further align the system with professional game design workflows.

Overall, this paper provides a compelling case for integrating user-driven exploration in AI-generated content, leveraging human intuition to traverse and harness the nuanced latent spaces of GANs for creative endeavors.