Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Controllable Game Level Generation: Assessing the Effect of Negative Examples in GAN Models

Mahsa Bazzaz, Seth Cooper

TL;DR

This work evaluates the performance of two controllable GAN variants, CGAN and Rumi-GAN, in generating game levels targeting specific constraints of interest: playability and controllability.

Abstract

Generative Adversarial Networks (GANs) are unsupervised models designed to learn and replicate a target distribution. The vanilla versions of these models can be extended to more controllable models. Conditional Generative Adversarial Networks (CGANs) extend vanilla GANs by conditioning both the generator and discriminator on some additional information (labels). Controllable models based on complementary learning, such as Rumi-GAN, have been introduced. Rumi-GANs leverage negative examples to enhance the generator's ability to learn positive examples. We evaluate the performance of two controllable GAN variants, CGAN and Rumi-GAN, in generating game levels targeting specific constraints of interest: playability and controllability. This evaluation is conducted under two scenarios: with and without the inclusion of negative examples. The goal is to determine whether incorporating negative examples helps the GAN models avoid generating undesirable outputs. Our findings highlight the strengths and weaknesses of each method in enforcing the generation of specific conditions when generating outputs based on given positive and negative examples.

Controllable Game Level Generation: Assessing the Effect of Negative Examples in GAN Models

TL;DR

This work evaluates the performance of two controllable GAN variants, CGAN and Rumi-GAN, in generating game levels targeting specific constraints of interest: playability and controllability.

Abstract

Generative Adversarial Networks (GANs) are unsupervised models designed to learn and replicate a target distribution. The vanilla versions of these models can be extended to more controllable models. Conditional Generative Adversarial Networks (CGANs) extend vanilla GANs by conditioning both the generator and discriminator on some additional information (labels). Controllable models based on complementary learning, such as Rumi-GAN, have been introduced. Rumi-GANs leverage negative examples to enhance the generator's ability to learn positive examples. We evaluate the performance of two controllable GAN variants, CGAN and Rumi-GAN, in generating game levels targeting specific constraints of interest: playability and controllability. This evaluation is conducted under two scenarios: with and without the inclusion of negative examples. The goal is to determine whether incorporating negative examples helps the GAN models avoid generating undesirable outputs. Our findings highlight the strengths and weaknesses of each method in enforcing the generation of specific conditions when generating outputs based on given positive and negative examples.

Paper Structure

This paper contains 16 sections, 3 equations, 3 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Domains
  4. System Overview
  5. Vanilla Generative Adversarial Nets
  6. Conditional Generative Adversarial Networks
  7. Rumi Generative Adversarial Nets
  8. Experiments
  9. Experiment One: Playability
  10. Experiment Two: Playability and Controllability
  11. Evaluation
  12. Experiment One: Playability
  13. Experiment Two: Playability and Controllability
  14. Discussion
  15. Conclusion
  16. ...and 1 more sections

Figures (3)

  • Figure 1: Venn diagrams of the sample space. Green plus signs represent positive samples and red minus signs represent negative samples provided to models.
  • Figure 2: Examples of successful and failed levels generated in Experiment 1. A successful level is defined as a playable level. In contrast, unplayable levels either lack start or end positions or have blocked player paths.
  • Figure 3: Examples of successful and failed levels generated in Experiment 2. A successful level must have the exact correct features (pipes/treasures) and be playable. Failed levels may be playable but with an incorrect number of features, or unplayable but with the correct number of features.