Search-Based Exploration and Diagnosis of TOAD-GAN

Abstract

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.

Procedural Level Generation with Diffusion Models from a Single Example

“…This kind of incoherence and discontinuities affect the playability and aesthetics of level. In order to find samples that satisfy specified requirements, evolutionary search is used to explore the latent space of a GAN (Volz et al 2018;Edwards, Jiang, and Togelius 2021;Fontaine et al 2021). Previous works have demonstrated that GAN-based generative models have a vast latent space, requiring conditional constraints or search algorithms to improve the quality of generated levels.…”