Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 21 of the 21 references it cites were found in a registry or an index; 10 of the 13 sources its claims cite were checked for support.

Level the Level: Balancing Game Levels for Asymmetric Player Archetypes With Reinforcement Learning

Open access
Rupp, Florian, Eckert, Kai · 2025 · Proceedings of the 20th International Conference on the Foundations of Digital Games

Abstract

Balancing games, especially those with asymmetric multiplayer content, requires significant manual effort and extensive human playtesting during development.For this reason, this work focuses on generating balanced levels tailored to asymmetric player archetypes, where the disparity in abilities is balanced entirely through the level design.For instance, while one archetype may have an advantage over another, both should have an equal chance of winning.We therefore conceptualize game balancing as a procedural content generation problem and build on and extend a recently introduced method that uses reinforcement learning to balance tile-based game levels.We evaluate the method on four different player archetypes and demonstrate its ability to balance a larger proportion of levels compared to two baseline approaches.Furthermore, our results indicate that as the disparity between player archetypes increases, the required number of training steps grows, while the model's accuracy in achieving balance decreases.

Overview

CoChat Summary

TL;DR

This paper extends PCGRL to balance tile-based game levels for asymmetric player archetypes entirely through level design, using reinforcement learning to swap tiles. It reports that PCGRL balances more levels than random search and hill-climbing, but performance drops as archetype disparity grows and the model can exploit unwinnable draws.

Key findings

  • PCGRL achieved higher balanced-level proportions than random search and hill-climbing across all tested setups: A vs. B 80.4% vs. 46.1% and 27.6%; A vs. C 56.5% vs. 24.6% and 16.2%; A vs. D1 72.3% vs. 46.6% and 28.8%; A vs. D2 57.9% vs. 35.2% and 26.8%; A vs. A 89.7% vs. 59.6% and 55.8% (Table 1).
  • Halving the action space from `[h, w, h, w, 2]` (2592 actions for a 6x6 level) to `[h, w, h, w]` (1296 actions) sped up convergence and allowed higher rewards.
  • Greater initial disparity between archetypes required more training steps and reduced the model’s accuracy in achieving balance; symmetric A vs. A initially favored one player in 50.1% of levels, while asymmetric setups strongly favored one player.
  • Performance varied by about 20 percentage points across archetypes, and greater initial unfairness made it harder for the model to compensate through level modification alone.

Methods

  • Balancing is formulated as a PCGRL Markov decision process on a 6x6 tile-based NMMO survival game; the RL agent swaps tiles, and balance is determined by simulated win frequencies of two heuristic agents, with reward 0.5 for equal win rates and 0 or 1 when one player wins every round.
  • Four new archetypes are introduced against base archetype A: B can cross rock tiles, C acts only every second turn, D1 wins with four food resources, and D2 wins with three food resources; A cannot cross rock or water and wins with five food resources.
  • Experiments use 500 generated levels, PPO with two 3-layered MLPs of sizes 64, 128, and 64, step size 512, 60 parallel environments, and 30,720 trajectories per update; results are compared against random search and hill-climbing baselines.

Limitations

  • Stated: the model can exploit a strategy where neither player can access food, achieving balance because both players never win; the reward function assumes draws are always balanced, including when both players lose, and future work aims to distinguish winnable from unwinnable draws.
  • Stated: as the initial disparity between archetypes increases, the required number of training steps grows and the model’s accuracy in achieving balance decreases.
  • Evident from the text: evaluation uses heuristic agents and simulated win rates rather than human playtesting in this study, and the experiments are limited to a 6x6 tile-based NMMO environment with four archetypes, so broader generalizability is not established.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3723498.3723747

Cites

21 of 21 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Training steps required for model convergence compared to the initial imbalance due to agent asymmetries per setup. A value of 0.5 indicates equal wins, while 1.0 means one agent always wins.
Figure 1 Training steps required for model convergence compared to the initial imbalance due to agent asymmetries per setup. A value of 0.5 indicates equal wins, while 1.0 means one agent always wins.CC-BY · From the PDF
Figure 2 Generated samples from different models and archetype setups in comparison with the initial unbalanced version. A player of archetype A (red) is paired with different archetypes (yellow).
Figure 2 Generated samples from different models and archetype setups in comparison with the initial unbalanced version. A player of archetype A (red) is paired with different archetypes (yellow).CC-BY · From the PDF

Verification

Claims the paper makes

8 Valid · 0 Not valid · 1 Not checked · 15 Own results and interpretations

10 more claims

Create a free account to see all 24 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 7
  • It may state two separate claims as one. 3

Create a free account to see these claims.

Citations

Cites

Show all 20 references

Research works printed but not linked: 1

  1. 10.1007/978-3-319-55849-325 · doi:10.1007/978-3-319-55849-325

Cited by

See all 1

Metadata

Authors
Rupp, F., & Eckert, K.
Affiliations
Technische Hochschule Mannheim (DE)
Venue
Proceedings of the 20th International Conference on the Foundations of Digital Games
Publisher
ACM
Published
2025-04-15
Pages
1-4
OpenAlex
W4410236929
Cited by
1
References
20
Citations by year
2026: 1
FWCI
0.95
Type
conference paper
Language
en
Open access
Locations
  • International Conference on the Foundations of Digital Games (FDG) (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University) (cc-by)
Funders
This research was supported by the Volkswagen Foundation (Project: Consequences of Artificial Intelligence on Urban Societies, Grant 98555). (98555)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Artificial Intelligence in Games
  • Social Sciences › Sociology and Political Science › Digital Games and Media
  • Physical Sciences › Computer Science › Artificial Intelligence › Reinforcement Learning in Robotics
Keywords
procedural content generation, reinforcement learning

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Level the Level: Balancing Game Levels for Asymmetric Player Archetypes With Reinforcement Learning. https://app.cochat.ai/papers/10.1145/3723498.3723747
The paper
Rupp, F., & Eckert, K. (2025). Level the Level: Balancing Game Levels for Asymmetric Player Archetypes With Reinforcement Learning. Proceedings of the 20th International Conference on the Foundations of Digital Games. https://doi.org/10.1145/3723498.3723747

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.