Evolving Mario Levels in the Latent Space of a Deep Convolutional Generative Adversarial Network (1805.00728v1)

Abstract: Generative Adversarial Networks (GANs) are a machine learning approach capable of generating novel example outputs across a space of provided training examples. Procedural Content Generation (PCG) of levels for video games could benefit from such models, especially for games where there is a pre-existing corpus of levels to emulate. This paper trains a GAN to generate levels for Super Mario Bros using a level from the Video Game Level Corpus. The approach successfully generates a variety of levels similar to one in the original corpus, but is further improved by application of the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). Specifically, various fitness functions are used to discover levels within the latent space of the GAN that maximize desired properties. Simple static properties are optimized, such as a given distribution of tile types. Additionally, the champion A* agent from the 2009 Mario AI competition is used to assess whether a level is playable, and how many jumping actions are required to beat it. These fitness functions allow for the discovery of levels that exist within the space of examples designed by experts, and also guide the search towards levels that fulfill one or more specified objectives.

Evolving Mario Levels in the Latent Space of a DCGAN

The paper "Evolving Mario Levels in the Latent Space of a Deep Convolutional Generative Adversarial Network" presents a method for procedurally generating video game levels by integrating a Deep Convolutional Generative Adversarial Network (DCGAN) with an evolutionary algorithm. This research focuses on creating new levels for the classic game, Super Mario Bros, utilizing the capabilities of GANs to emulate expert-designed levels while introducing variability and novelty.

Overview

The paper leverages the strength of Generative Adversarial Networks (GANs), specifically a deep convolutional variant, to autonomously produce playable levels that capture the essence of human-designed examples from existing level corpora. By training a DCGAN on a single Super Mario Bros level from the Video Game Level Corpus, the authors demonstrate the ability to generate new level sections that exhibit similar styles and characteristics to those within the training data.

To further refine and meet the desired design objectives, the authors employ the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). This approach allows the exploration of the latent space of the DCGAN, utilizing different fitness functions to optimize certain level properties, such as playability and tile distribution, including the ability to maximize or minimize enemy presence and ground tiles.

Numerical and Experimental Results

The research presents several exemplary outputs from the DCGAN, illustrating the diversity and coherence of generated level sections in relation to human-designed levels. The numerical analysis includes fitness evaluations for evolved levels targeting specific design goals, such as varying the percentage of ground tiles and the number of enemies, showcasing the model's ability to meet these structured objectives.

Notably, the optimization via CMA-ES helps ensure the levels are both creative and playable by employing the results of game simulations through an agent, specifically the champion A* pathfinding agent from the Mario AI Competition. This added layer of validation assesses the levels beyond static properties by considering dynamic elements such as the number of jumps required to complete a level, thereby correlating with perceived gameplay difficulty.

Implications and Future Directions

The implications of this work are significant in both the academic and commercial avenues of Procedural Content Generation (PCG). By using a data-driven approach, designers can automate content creation, potentially leading to more dynamic and personalized gaming experiences. This research also sets a foundation for future exploration of machine learning techniques in creative domains beyond traditional image and text applications.

For the academic community, this paper opens new pathways in the intersection of generative models and evolution strategies, suggesting further exploration of richer datasets and multi-objective optimization to capture a wider variety of design characteristics in games. Future research might focus on enhancing GAN architectures to address discrete and structured datasets typical of game environments, as mentioned regarding the issue of broken structures like incomplete pipes.

The possibility of extending this model to other games and genres also suggests a fruitful area of paper, as the technique theoretically generalizes across games with existing level datasets. Additionally, integrating these methods with real-time game adaptation elements could revolutionize how games evolve in response to player interaction, potentially leading to bespoke gameplay experiences.

In conclusion, this paper contributes a novel approach to automated game level design by synthesizing advanced GAN architectures with evolutionary algorithms, providing a compelling example of machine learning's expanding role in creative and entertainment technology.