Illuminating Mario Scenes in the Latent Space of a Generative Adversarial Network

Fontaine, Matthew C.; Liu, Ruilin; Khalifa, Ahmed; Modi, Jignesh; Togelius, Julian; Hoover, Amy K.; Nikolaidis, Stefanos

doi:10.1609/aaai.v35i7.16740

Cited by 34 publications

(25 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…There have also been applications to creative domains, most notably procedural content generation. For example, Map-Elites has been applied to Mario-like platformer game stage design [18], [19].…”

Section: Quality-diversity Algorithmsmentioning

confidence: 99%

Quality-diversity for Synthesizer Sound Matching

Masuda

Saito

2023

Journal of Information Processing

View full text Add to dashboard Cite

Synthesizers have become commonplace in music production due to their ability to create a diverse and unique timbre. However, many users find it difficult to control a synthesizer due to the complexity of the relationship between synthesis parameters and output sound. As such, the sound matching task focuses on the estimation of synthesis parameters that recreate the target sound as closely as possible within the capabilities of the synthesizer. In this paper, we introduce a quality-diversity approach to the problem of sound matching by introducing a novelty objective to genetic algorithm-based sound matching. We show that this QD method allows for the discovery of a collection of matches that are more diverse compared to matches found by standard genetic algorithms. These matches are spread across a space of interpretable spectral features, making it useful for the intuitive discovery of unique and useful matches.

show abstract

Section: Quality-diversity Algorithmsmentioning

confidence: 99%

Quality-diversity for Synthesizer Sound Matching

Masuda

Saito

2023

Journal of Information Processing

View full text Add to dashboard Cite

show abstract

“…To automatically evaluate the levels we make use of Robin Baumgarten's A*-based Mario level playing agent, the winning agent in the first Mario AI competition [24], which is included in the benchmark and is often used in PCG research to confirm that a level is completable [3,7].…”

Section: Mario Ai Benchmarkmentioning

confidence: 99%

“…Also extracted from the agent simulations was the amount of horizontal progress the agent was able to make as a proportion of the whole level width. This feature is often referred to as 'playability' or 'fitness' in the prior work using the Mario AI benchmark [3,7], though we refer to it as playability as it is the more specific term. Playability is calculated as a score from 0 to 1, where 1 indicates the agent was able to reach the end of the level.…”

Section: Metrics Assessedmentioning

confidence: 99%

The Right Variety: Improving Expressive Range Analysis with Metric Selection Methods

Withington

Tokarchuk

2023

Proceedings of the 18th International Conference on the Foundations of Digital Games

View full text Add to dashboard Cite

Expressive Range Analysis (ERA), an approach for visualising the output of Procedural Content Generation (PCG) systems, is widely used within PCG research to evaluate and compare generators, often to make comparative statements about their relative performance in terms of output diversity and search space exploration. Producing a standard ERA visualisation requires the selection of two metrics which can be calculated for all generated artefacts to be visualised. However, to our knowledge there are no methodologies or heuristics for justifying the selection of a specific metric pair over alternatives. Prior work has typically either made a selection based on established but unjustified norms, designer intuition, or has produced multiple visualisations across all possible pairs. This work aims to contribute to this area by identifying valuable characteristics of metric pairings, and by demonstrating that pairings that have these characteristics have an increased probability of producing an informative ERA projection of the underlying generator. We introduce and investigate three quantifiable selection criteria for assessing metric pairs, and demonstrate how these criteria can be operationalized to rank those available. Though this is an early exploration of the concept of quantifying the utility of ERA metric pairs, we argue that the approach explored in this paper can make ERA more useful and usable for both researchers and game designers. INTRODUCTIONProcedural Content Generation (PCG) systems for games can often produce huge amounts of diverse artefacts and experiences. Whether it is the billions of possible maps of Minecraft [18] or the quintillions of planets in No Man Sky [6], the total volume of a PCG system's possible output, commonly referred to as its 'generative space' can be huge and complex. The challenge of understanding and usefully comparing generative spaces is a vital one to both

show abstract

“…In this way CMA-ES is used to learn the distribution of more fit output samples. LVE has also been extended to Latent Variable Illumination by using CMA-ME, a quality-diversity version of CMA-ES (Fontaine et al 2020).…”

Section: Latent Space Examinationmentioning

confidence: 99%

Search-Based Exploration and Diagnosis of TOAD-GAN

Edwards

Jiang

Togelius

2021

AIIDE

View full text Add to dashboard Cite

Generative Adversarial Networks (GANs) have been used with great success to generate images. They have also been applied to the task of Procedural Content Generation (PCG) in games, particularly for level generation, with various approaches taken to solving the problem of training data. One of those approaches, TOAD-GAN (Token-Based One-Shot Arbitrary Dimension Generative Adversarial Network) (Awiszus, Schubert, and Rosenhahn 2020), can generate levels based on a single training example and has been able to closely reproduce token patterns found in the training sample. While TOAD-GAN is an impressive achievement, questions remain about what exactly it has learned. Can the generator be made to produce levels that are substantially different from the level it has been trained on? Can it reproduce specific level segments? How different are the generated levels? We investigate these questions and others by using the CMA-ES algorithm for Latent Space Evolution. To make the search space feasible, we use a random projection in latent space. We propose the investigation undertaken here as a paradigm for studies into what machine-learned generators have actually learned, and also as a test of a new method for projecting from a smaller search space to a larger latent space.

show abstract

Illuminating Mario Scenes in the Latent Space of a Generative Adversarial Network

Cited by 34 publications

References 22 publications

Quality-diversity for Synthesizer Sound Matching

Quality-diversity for Synthesizer Sound Matching

The Right Variety: Improving Expressive Range Analysis with Metric Selection Methods

Search-Based Exploration and Diagnosis of TOAD-GAN

Contact Info

Product

Resources

About