The DOI matches its Crossref record; 58 of the 58 references it cites were found in a registry or an index; 4 of the 9 sources its claims cite were checked for support.
Simulation-Driven Balancing of Competitive Game Levels With Reinforcement Learning
Open accessAbstract
The balancing process for game levels in competitive two-player contexts involves a lot of manual work and testing, particularly for non-symmetrical game levels. In this work, we frame game balancing as a procedural content generation task and propose an architecture for automatically balancing of tile-based levels within the PCGRL framework (procedural content generation via reinforcement learning). Our architecture is divided into three parts: (1) a level generator, (2) a balancing agent, and (3) a reward modeling simulation. Through repeated simulations, the balancing agent receives rewards for adjusting the level towards a given balancing objective, such as equal win rates for all players. To this end, we propose new swap-based representations to improve the robustness of playability, thereby enabling agents to balance game levels more effectively and quickly compared to traditional PCGRL. By analyzing the agent's swapping behavior, we can infer which tile types have the most impact on the balance. We validate our approach in the Neural MMO (NMMO) environment in a competitive two-player scenario. In this extended conference paper, we present improved results, explore the applicability of the method to various forms of balancing beyond equal balancing, compare the performance to another search-based approach, and discuss the application of existing fairness metrics to game balancing.
Overview
CoChat Summary
TL;DR
The paper frames competitive level balancing as a PCGRL task and proposes a three-part architecture—level generator, balancing agent, and simulation-based reward—using new swap-based tile representations. In a simplified two-player Neural MMO forage-survival domain, the method improves balance more effectively than original PCGRL and hill climbing while keeping levels playable.
Key findings
- Swap representations improved balance over original PCGRL: S-narrow balanced 48.1% and improved 63.3%, S-turtle 42.5%/56.8%, S-wide 48.1%/64.1%, versus PCGRL narrow 30.4%/36.7%; original PCGRL left 38.8% of levels unplayable, while swap methods produced 100% playable levels in this domain.
- Extended training and separate spawn-tile encodings raised performance to 68.0% balanced and 88.9% improved for target balance b = 0.5; for varied targets b ∈ [0, 0.5], balanced-to-b rates ranged from 36.1% at b = 0.0 to 68.0% at b = 0.5, with improved-toward-b rates from 71.1% to 88.9%.
- PCGRL with S-wide outperformed hill climbing across b values; at b = 0.5 it balanced 68.0% versus 59.6% for S-narrow hill climbing and 14.3% for narrow hill climbing, while narrow hill climbing left 83.1%–92.4% of levels unplayable.
- Swap-frequency analysis indicated swapping forest for stone had the most impact on balance, followed by forest for water; swapping player spawn positions was not a favored action. Generated levels had 100% diversity and none were symmetric.
Methods
- Uses the PCGRL framework with PPO and introduces swap-narrow, swap-turtle, and swap-wide MDP representations; reward is derived from statistical parity, with balance state b_t = |wp1_t − b| and reward r_t = b_{t−1} − b_t + α.
- Level generator is trained separately to produce playable levels with exactly two players and a valid path; balancing agent then swaps tiles, and reward is computed from n = 14 simulations using the scripted Forage-agent in Neural MMO.
- Experiments use a 6×6 grid, two players, combat disabled, win conditions of collecting five food or being last standing; evaluation uses 1,000 generated levels, 200k training steps / 970 policy updates, 20% change percentage, and max 100 steps per episode.
Limitations
- Stated: the balance metric depends on the heuristic player agents used in simulation; different skill levels or agent types would change the measured balance, and human playtests are left for future work.
- Stated: training is computationally intensive because each swap requires multiple simulations; rewards are mostly sparse, and swapping may break playability in domains where multiple tiles have semantic relationships. The paper also notes fairness is a social/ethical concept and that residual uncertainty may require treating a range rather than an exact target as balanced.
- Evident: evaluation is limited to a simplified 6×6 two-player Neural MMO forage-survival setting with combat disabled and scripted agents; not all levels reach the target balance (e.g., 68.0% balanced at b = 0.5, 36.1% at b = 0.0), so the method improves but does not fully solve balancing.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
2 Valid · 5 Not valid · 5 Not checked · 71 Own results and interpretations
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“For this game with the configured number of players, this is n ≥ 14.”
plainstates a causen = 30- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“In contrast, the narrow hill climbing approach shows a weak performance, still clearly lagging behind the methods using swap representations.”
§ C. Comparison with Hill Climbing Approaches > 0.0 36.1 33.7 5.4 92.4
boostedstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“The performance, however, differs depending on b.”
§ VII. EXTENDED STUDY ON OVERALL IMPROVEMENTS > 0.0 36.1 71.1 8.6 -
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Splitting up the level generation and balancing process into separate units yielded much better results than doing the balancing in a single step as in the original PCGRL.”
plainstates a cause- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper states — it cites no source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“For representations with a higher proportion of balanced levels, the average episode lengths are shorter.”
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Rests on an assumption not shown here.
32 more claims
Create a free account to see all 83 claims and to verify the 1 reference still open.
- Probably not a substantial claim. 29
- It may state two separate claims as one. 27
- The quote only partly states this claim. 1
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“In a direct comparison, the training using the smaller action space converges faster for all models and allows for higher rewards, resulting in overall better performance, in particular for type A vs. A (cf. Table 1 and also [12,13]).”
From Level the Level: Balancing Game Levels for Asymmetric Player Archetypes With Reinforcement Learning (2025) · cites this paper as [13]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“This approach is, however, limited by the fact that it only uses exactly the same archetype of agents playing against each other.”
From Level the Level: Balancing Game Levels for Asymmetric Player Archetypes With Reinforcement Learning (2025) · cites this paper as [13]
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“For this work, we use the swap-wide representation to define the action space of the MDP, as it gives the best results in [13].”
From Level the Level: Balancing Game Levels for Asymmetric Player Archetypes With Reinforcement Learning (2025) · cites this paper as [13]
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ C. Reward Design
“To reward the agent additionally, a reward α is given if bt is exactly b. Otherwise, α is 0. If α > 0, the episode always ends since bt = b.”
§ C. Reward Design
“bt = |wp1t − b | (3)”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ C. Reward Design
“Due to the use of the absolute value, wpit of both players could be used in Eq. 3, since their values describe probabilities that sum up to 1.”
§ C. Reward Design
“bt = |wp1t − b | (3)”
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ VI. EXPERIMENTS AND RESULTS
“The original PCGRL narrow representation also improved the share of balanced levels, however, with 36.7% this share is smaller compared to the swap represen- tations.”
§ VI. EXPERIMENTS AND RESULTS
“Balanced (%) 48.1 42.5 48.1 30.4 Improved (%) 63.3 56.8 64.1 36.7”
Citations
Cites
Golub, G. (2012) · doi:10.56021/9781421407944 · Cited by 34,505
Proximal Policy Optimization Algorithms
Schulman, J., Wolski, F., Dhariwal, P., Radford, A., & Klimov, O. (2017) · arXiv · doi:10.48550/arxiv.1707.06347 · Cited by 10,625 · cited 1×
“TheNMMOenvironmentusesPCGbydefaulttogenerate randomlevels,however,theplayers’spawnpositionsareran-domlydeterminedatthestartofthegame.”
A general reinforcement learning algorithm that masters chess, shogi, and Go through self-play
Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., Lanctot, M., Sifre, L., Kumaran, D., Graepel, T., et al. (2018) · Science · doi:10.1126/science.aar6404 · Cited by 3,562 · cited 1×
“In the games community, RL has been widely applied to play games [29], [30], but recently, it has also been applied to PCG [12], [31], [32], [33].”
Dwork, C., Hardt, M., Pitassi, T., Reingold, O., & Zemel, R. (2012) · Proceedings of the 3rd Innovations in Theoretical Computer Science Conference · doi:10.1145/2090236.2090255 · Cited by 3,501 · cited 3×
“The statistical parity metric [15] is one of them and is often used to measure fairness between two groups.”
InfluentialGrandmaster level in StarCraft II using multi-agent reinforcement learning
Vinyals, O., Babuschkin, I., Czarnecki, W. M., Mathieu, M., Dudzik, A., Chung, J., Choi, D. H., Powell, R., Ewalds, T., Georgiev, P., et al. (2019) · Nature · doi:10.1038/s41586-019-1724-z · Cited by 3,455 · cited 1×
“In the games community, RL has been widely applied to play games [29], [30], but recently, it has also been applied to PCG [12], [31], [32], [33].”
Fair prediction with disparate impact: a study of bias in recidivism prediction instruments
Chouldechova, A. (2017) · Big Data · doi:10.1089/big.2016.0047 · Cited by 2,159 · cited 1×
“At this point, it is important to note that the concept of fairness is a social and ethical concept, and not a statistical one [45].”
Certifying and Removing Disparate Impact
Feldman, M., Friedler, S. A., Moeller, J., Scheidegger, C., & Venkatasubramanian, S. (2015) · Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining · doi:10.1145/2783258.2783311 · Cited by 1,827 · cited 1×
“The Disparate Impact metric [46], [47] expresses fairness with the ratio of a predicted outcome of two groups.”
Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm
Silver, D., Hubert, T., Schrittwieser, J., Antonoglou, I., Lai, M., Guez, A., Lanctot, M., Sifre, L., Kumaran, D., Graepel, T., et al. (2017) · arXiv · doi:10.48550/arxiv.1712.01815 · Cited by 1,102
Search-Based Procedural Content Generation: A Taxonomy and Survey
Togelius, J., Yannakakis, G. N., Stanley, K. O., & Browne, C. (2011) · IEEE Transactions on Computational Intelligence and AI in Games · doi:10.1109/tciaig.2011.2148116 · Cited by 706 · cited 3×
“PCG methods span over from dedicated algorithms [23], search-based methods [4], [7], [8] up to the use of machine learning [24] and deep learning [25], [26], [27], [28].”
Aequitas: A Bias and Fairness Audit Toolkit
Saleiro, P., Kuester, B., Hinkson, L., London, J., Stevens, A., Anisfeld, A., Rodolfa, K. T., & Ghani, R. (2018) · arXiv · doi:10.48550/arxiv.1811.05577 · Cited by 496
Research works printed but not linked: 1
- Elo, 1978, The Rating of Chessplayers, Past and Present
Cited by
G-PCGRL: Procedural Graph Data Generation via Reinforcement Learning
Rupp, F., & Eckert, K. (2024) · IEEE Conference on Games (CoG) · doi:10.1109/cog60054.2024.10645633 · Cited by 6
“…use tools like Machinations 1 [3] or are supported by works on procedural content generation (PCG) such as game levels [4]–[6], balancing [7], [8], or game This research was supported by the Volkswagen Foundation (Project: Consequences of Artificial Intelligence on Urban Societies, Grant…”
MethodSelf-citationIt might be balanced, but is it actually good? An Empirical Evaluation of Game Level Balancing
Rupp, F., Puddu, A., Becker-Asano, C., & Eckert, K. (2024) · 2024 IEEE Conference on Games (CoG) · doi:10.1109/cog60054.2024.10645642 · Cited by 4
“Similar to our work [10], [11], Karavolos et al. [5] use a PCG method for automated game level design using artificial agents to simulate game play.”
MethodSelf-citationGenerative AI in Procedural Content Generation for Computer Games: Key Contributions and Trends
Alyaseri, S. (2025) · IEEE Region 10 Symposium (TENSYMP) · doi:10.1109/tensymp63728.2025.11144949 · Cited by 3
“Future work includes enhancing aesthetics via Stable Diffusion and balancing for varied player types [34].”
Horst, R., Dörner, R., Sinha, S. N., & Kühn, P. J. (2026) · Computers & Graphics · doi:10.1016/j.cag.2026.104683 · Cited by 2
Level the Level: Balancing Game Levels for Asymmetric Player Archetypes With Reinforcement Learning
Rupp, F., & Eckert, K. (2025) · Proceedings of the 20th International Conference on the Foundations of Digital Games · doi:10.1145/3723498.3723747 · Cited by 1
Self-citation
Metadata
- Authors
- Rupp, F., Eberhardinger, M., & Eckert, K.
- Affiliations
- Technische Hochschule Mannheim (DE)
- Stuttgart Media University (DE)
- Venue
- IEEE Transactions on Games
- Published
- 2024-12
- Volume
- 16
- Issue
- 4
- Pages
- 903-913
- ISSN
- 2475-1502 (ISSN-L)
- 2475-1510
- OpenAlex
- W4396790128
- Cited by
- 11
- References
- 52
- Citations by year
- 2026: 6 · 2025: 3 · 2024: 2
- FWCI
- 1.62
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- Locations
- IEEE Transactions on Games (publishedVersion)
- arXiv (Cornell University) (submittedVersion)
- arXiv (Cornell University) (cc-by)
- Funders
- Volkswagen Foundation through the Project: Consequences of Artificial Intelligence on Urban Societies (98555)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Artificial Intelligence in Games
- Physical Sciences › Computer Science › Artificial Intelligence › Reinforcement Learning in Robotics
- Social Sciences › Sociology and Political Science › Digital Games and Media
- Keywords
- procedural content generation, reinforcement learning, reward modeling, fairness metrics
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).
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: Simulation-Driven Balancing of Competitive Game Levels With Reinforcement Learning. https://app.cochat.ai/papers/10.1109/tg.2024.3399536
- The paper
- Rupp, F., Eberhardinger, M., & Eckert, K. (2024). Simulation-Driven Balancing of Competitive Game Levels With Reinforcement Learning. IEEE Transactions on Games. https://doi.org/10.1109/tg.2024.3399536
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.