mlsanityData to 14 Oct 2026

Learning Local Constraints for Reinforcement-Learned Content Generators

arXiv 2605.13570 ↗First posted 13 May 2026cs.AI cs.LGcited by 0

Not on the map. Why?

Abstract (arXiv)

Constraint-based game content generators that learn local constraints from existing content, such as Wave Function Collapse (WFC), can generate visually satisfying game levels but face challenges in optimizing global properties, such as playability. On the other hand, reinforcement-learning-trained generators can optimize global properties---because such properties can easily be included in reward functions---but the results can be visually dissatisfying. In this paper, we explore ways to combine these methods. Specifically, we constrain the action space of a PCGRL generator with constraints learned by WFC, effectively allowing the PCGRL generator to achieve global properties while being forced to adhere to local constraints. To better analyze how this hybrid content generation method operates, we vary the number and type of inputs, and we test whether to randomly collapse the starting state and exclude rare patterns. While the method is sensitive to hyperparameter tuning, the best of our trained generators produce visually satisfying and playable puzzle-platform game levels---such as Lode Runner levels---with desired global properties.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 21 of its 41 references that are in the map. First posted on arXiv 13 May 2026.

Builds on

21 of 41 references.

References · 21

  1. 2016OpenAI Gymcited by 1,736
  2. 2020A Closer Look at Invalid Action Masking in Policy Gradient Algorithmscited by 171
  3. 2017Procedural Content Generation via Machine Learning (PCGML)cited by 137
  4. 2020PCGRL: Procedural Content Generation via Reinforcement Learningcited by 90
  5. 2018Evolving Mario Levels in the Latent Space of a Deep Convolutional Generative Adversarial Networkcited by 76
  6. 2023Level Generation Through Large Language Modelscited by 68
  7. 2016The VGLC: The Video Game Level Corpuscited by 41
  8. 2016Super Mario as a String: Platformer Level Generation Via LSTMscited by 40
  9. 2021Learning Controllable Content Generatorscited by 40
  10. 2021Growing 3D Artefacts and Functional Machines with Neural Cellular Automatacited by 38
  11. 2021Adversarial Reinforcement Learning for Procedural Content Generationcited by 36
  12. 2021Experience-Driven PCG via Reinforcement Learning: A Super Mario Bros Studycited by 25
9 more
  1. 2020Interactive Evolution and Exploration Within Latent Level-Design Space of Generative Adversarial Networkscited by 22
  2. 2022Learning Controllable 3D Level Generatorscited by 19
  3. 2019Tile Pattern KL-Divergence for Analysing and Evolving Game Levelscited by 16
  4. 2020Controllable Level Blending between Games using Variational Autoencoderscited by 14
  5. 2021Illuminating the Space of Beatable Lode Runner Levels Produced By Various Generative Adversarial Networkscited by 14
  6. 2022Path of Destruction: Learning an Iterative Level Generator Using a Small Datasetcited by 14
  7. 2020Multi-Domain Level Generation and Blending with Sketches via Example-Driven BSP and Variational Autoencoderscited by 11
  8. 2018Addressing the Fundamental Tension of PCGML with Discriminative Learningcited by 8
  9. 2021Generating Lode Runner Levels by Learning Player Paths with LSTMscited by 3

Built on by

Not cited yet.