Playtesting: What is Beyond Personas
Open accessThe DOI matches its Crossref record; 33 of the 53 references it cites were found in a registry or an index; 17 of the 28 sources its claims cite were checked for support.
Abstract
Playtesting is an essential step in the game design process. Game designers use the feedback from playtests to refine their designs. Game designers may employprocedural personasto automate the playtesting process. In this article, we present two approaches to improve automated playtesting. First, we proposedeveloping persona, which allows a persona to progress to different goals. In contrast, the procedural persona is fixed to a single goal. Second, a human playtester knows which paths she has tested before, and during the consequent tests, she may test different paths. However, reinforcement learning (RL) agents disregard these previous paths. We propose a novel methodology that we refer to as alternative path finder (APF). We train APF with previous paths and employ APF during the training of an RL agent. APF modulates the reward structure of the environment, while preserving the agent’s goal. When evaluated, the agent generates a different trajectory that achieves the same goal. We use the general video game artificial intelligence and VizDoom frameworks to test our proposed methodologies. We use proximal policy optimization RL agent during experiments. First, we compare the playtest data generated by developing and procedural persona. Our experiments show that developing persona provides better insight into the game and how different players would play. Second, we present the alternative paths found using APF and argue why traditional RL agents cannot learn those paths.
Verifying this paper
8 of 9 steps done · this page updates itself — you can leave and come back
- Registry recordVerified
- Full textRead the full text
- Reference list33 references found
- Reference checkChecking each cited work — 33 of 33
- StatisticsRecomputed
- Claims134 claims anchored to a quote
- Cited sourcesChecked
- Self-checkChecked
- VerificationFindings verified
Overview
CoChat Summary
TL;DR
The paper proposes a “developing persona” that links multiple goals with criteria and utility functions, and an Alternative Path Finder (APF) that modulates RL rewards to discover trajectories different from previously seen paths. Experiments in GVG-AI and VizDoom show developing personas give broader/more controllable playtest behavior than procedural personas, and APF finds alternative paths that traditional RL agents overlook.
Key findings
- In GVG-AI Experiment I, procedural personas interacted with a small region and often produced the same behavior: Exit, Monster Killer, Treasure Collector, and Completionist each killed 1 monster, collected 1 treasure, and reached 1 door. Developing personas interacted more broadly: Dev. Killer killed 3 monsters, Dev. Collector collected 4 treasures, Dev. Raider killed 3 and collected 4, and Dev. Completionist killed 5 and collected 8.
- With PPO+CTS, procedural personas fit their decision models better: Monster Killer killed 2 monsters, Treasure Collector collected 3 treasures, and Completionist killed 2 and collected 3.
- APFCTS found alternative paths in GVG-AI: Path 1 had four alternative paths (Paths 2–6); using APFCTS decreased reward for playing the same path by at least 0.1 and increased reward for space-disjoint paths by at least 0.1. APFCTS clustered paths into equivalence classes {1,2} and {3,4,5,6}.
- In Doom Experiment III over 1000 evaluations, personas behaved per specification: Exit reached the door 1.00±0.00; Monster Killer killed 5.79±0.91 monsters; Dev. Killer killed 3.54±0.98; Treasure Collector killed 1.94±0.70 and collected 0.94±0.24 treasures; Dev. Collector killed 2.00±0.65 and collected 0.95±0.22; Dev. Raider killed 3.52±0.73 and collected 0.98±0.15; Completionist killed 5.76±1.06 and collected 1.91±0.38; Dev. Completionist killed 5.81±0.92 and collected 1.91±0.36; Dev. Casual Completionist killed 5.83±0.53 and collected 0.98±0.13.
- In Doom Experiment IV, without APFICM the Exit persona took the first path (52 steps) with reward 0.80±0.02 versus second path (77 steps) 0.68±0.01; after training APFICM on path 1, path 1 reward dropped to 0.51±0.02 and path 2 rose to 0.78±0.01.
Methods
- The paper introduces developing persona as a procedural persona extended to a linked sequence of goals, each with a utility function and transition criteria; it also introduces APF, with APFCTS using Context Tree Switching recoding probability and APFICM using Intrinsic Curiosity Module prediction error to add reward/penalty for novel versus similar states/transitions.
- Experiments use GVG-AI (14×20 grid, 200 timesteps, six actions) and two VizDoom levels (1600×832 and 1664×704, 2000 timesteps, seven actions), with PPO agents from StableBaselines; the deterministic first game was evaluated once, while stochastic Doom games were evaluated 1000 times, and the APF experiment used the same random seed.
- Four procedural personas (Exit, Monster Killer, Treasure Collector, Completionist) and five developing personas (Dev. Killer, Dev. Collector, Dev. Raider, Dev. Completionist, Dev. Casual Completionist) were tested, with utility weights and criteria specified in tables.
Limitations
- Stated: developing and procedural persona performance depends on the RL algorithm; if the RL algorithm cannot play a game, the designer cannot benefit from these automated playtesters. APF performance is linked to how well the underlying exploration algorithm performs in the environment.
- Stated: APF feedback is unbounded, so agents may loop over novel states or get stuck; strict game portions with no alternative paths may be penalized; APFCTS can be sensitive to slight frame noise; APFICM is more data-intensive and requires a previously trained agent for transfer learning. Alternative paths are also subjective for game designers even if APF is objective.
- Evident from the text: utility functions were handcrafted rather than extracted from human playtest data, and the evaluation is limited to GVG-AI and VizDoom testbeds without direct human playtest comparison.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
12 Valid · 2 Not valid · 7 Not checked · 113 Own results and interpretations
Some cited sources support this claim and some don’t — read each source’s verdict below.
In the paper
“RL algorithms such as Deep Q-Network (DQN) [9], Proximal Policy Optimization (PPO) [10], and Monte Carlo Tree Search (MCTS) [11] disregard the previous trajectories.”
What the sources say
- Not checked — no full text of the sourcecites [9] [9] V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. A. Riedmiller, A. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis, 'Human-level control through deep reinforcement learning,' Nature , vol. 518, pp. 529-533, 2015. - source
- Supportedcites [10] [10] J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, 'Proximal policy optimization algorithms,' CoRR , vol. abs/1707.06347, 2017. - source
The citation to this paper is attached to PPO, not DQN or MCTS. Section 5 (p. 5) describes collecting fresh trajectory segments for each iteration and optimizing on that iteration’s data, which supports the claim that PPO does not carry earlier iterations’ trajectories forward. PPO does reuse the current iteration’s data for multiple optimization epochs. (LLM, conf=0.82)
- Disputedcites [11] [11] C. B. Browne, E. Powley, D. Whitehouse, S. M. Lucas, P. I. Cowling, P. Rohlfshagen, S. Tavener, D. Perez, S. Samothrakis, and S. Colton, 'A survey of monte carlo tree search methods,' IEEE Transactions on Computational Intelligence and AI in Games , vol. 4, no. 1, pp. 1-43, March 2012. - sourceExperimental — accuracy not yet measured
For the MCTS portion attributed to this source, Section 3 (p. 5) says the search tree is guided by previous exploration, rather than disregarding it. Section 1.1 (p. 1) also says each simulation updates the tree and its statistics. This paper does not establish the sentence’s claims about DQN or PPO. (LLM, conf=0.88)
Checked by a reasoning model — may be wrong.
plain- 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.
Every source the paper cites for this claim was checked and supports it.
In the paper
“The second drawback is that some portion of the game may be strict, offering no alternative paths such as Super Mario Bros. [14].”
What the sources say
- Supportedcites [14] [14] D. Pathak, P. Agrawal, A. A. Efros, and T. Darrell, 'Curiosity-driven exploration by self-supervised prediction,' in Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6-11 August 2017 , 2017, pp. 2778-2787. - source
The Discussion (page unavailable) describes a pit 38% into Super Mario Bros. that requires a specific sequence of key presses to cross. If the agent fails, it dies rather than progressing beyond the pit, supporting the claim that this portion offers no alternative path forward. (LLM, conf=0.88)
hedged- 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.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Procedural personas [4] [5] accomplish this by rewiring the feedback mechanism by a utility function.”
What the sources say
- Supportedcites [4] [4] C. Holmgard, M. C. Green, A. Liapis, and J. Togelius, 'Automated playtesting with procedural personas with evolved heuristics,' IEEE Transactions on Games , pp. 1-1, 2018. - source
The Introduction (p. 1) says personas use different utility functions to guide decision-making. The MCTS description (p. 3) explains that utility scores are fed back into future policy decisions, and the Discussion (p. 9) reports different behaviors driven by those functions. Together, these support the paraphrase. (LLM, conf=0.90)
- Supportedcites [5] [5] L. Mugrai, F. Silva, C. Holmgård, and J. Togelius, 'Automated playtesting of matching tile games,' in 2019 IEEE Conference on Games (CoG) . IEEE, 2019, pp. 1-7. - source
The Background defines procedural personas in terms of utility functions. The Procedural Personas and Backpropogation sections describe agents with different goals whose evolved formulas replace the standard MCTS selection equation, supporting the paraphrase that changing the utility mechanism produces varied play strategies. (LLM, conf=0.86)
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.
- Supported
Every source the paper cites for this claim was checked and supports it.
In the paper
“Second, depending on the level layout, personas can execute a similar sequence [4].”
What the sources say
- Supportedcites [4] [4] C. Holmgard, M. C. Green, A. Liapis, and J. Togelius, 'Automated playtesting with procedural personas with evolved heuristics,' IEEE Transactions on Games , pp. 1-1, 2018. - source
In “Evaluating Levels with Personas” (p. 8), the paper examines how map layout affects persona playthroughs and reports that the Runner and Completionist follow a similar path to the exit on map 6. This supports the claim that personas can execute similar sequences depending on the level layout. (LLM, conf=0.90)
plainstates a possible 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.
- Supported
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
“When evaluated, the agent generates a different trajectory that achieves the same goal.”
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.
63 more claims
Create a free account to see all 134 claims and to verify the 1 reference still open.
- Probably not a substantial claim. 37
- It may state two separate claims as one. 35
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 major: it would touch a result — a value a result rests on, reported and recomputed p-values on opposite sides of the significance level, two passages that conflict with no sign which is right, or a cited work retracted before this paper was published.
In the paper
§ B. Alternative Path Finder · p. 6
“This formula yields maximum β reward when q new → 0 and minimum -β when q new → ∞ . We use this additional feedback signal to reward the novel transitions and to penalize similar transitions.”
§ B. Alternative Path Finder · p. 6
“If q new is greater than q mean , this indicates that this transition is less likely to exist in the previous trajectories. If q new is less than q mean , this indicates that this transition is likely to exist in the previous trajectories.”
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
§ B. Experiment II: Alternative paths found in GVG-AI: · p. 9
“Path 1 has four alternative paths-Paths 2 to 6.”
§ B. Experiment II: Alternative paths found in GVG-AI: · p. 9
“APFCTS clusters the paths in Experiment II into two equivalence classes, which are { 1 , 2 } and { 3 , 4 , 5 , 6 } .”
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. Experiment III: Personas in Doom: · p. 9
“The Completionist and Developing Completionist regularly collect the second Treasure , but in doing so, rarely die and cannot finish the level.”
§ C. Experiment III: Personas in Doom: · p. 10
“| Comp. | 5.76 ± 1.06 | 1.91 ± 0.38 | 0.95 ± 0.22 | 0.01 ± 0.11 | | Dev. Comp. | 5.81 ± 0.92 | 1.91 ± 0.36 | 0.96 ± 0.19 | 0.01 ± 0.09 |”
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. Experiment III: Personas in Doom: · p. 9
“The Developing Casual Completionist always finishes the level but usually cannot collect the second Treasure .”
§ C. Experiment III: Personas in Doom: · p. 10
“| Dev. Cas. Comp. | 5.83 ± 0.53 | 0.98 ± 0.13 | 0.98 ± 0.14 | 0.00 ± 0.00 |”
Citations
Cites
Human-level control through deep reinforcement learning
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmiller, M., Fidjeland, A. K., Ostrovski, G., et al. (2015) · Nature · doi:10.1038/nature14236 · Cited by 30,357 · cited 3×
“Researchers measured the uncertainty using bootstrapped DQN [24], statespace features [14], and error of a neural function [25].”
MethodReinforcement Learning: An Introduction
Sutton, R. S., & Barto, A. G. (1998) · IEEE Transactions on Neural Networks · doi:10.1109/tnn.1998.712192 · Cited by 27,225 · cited 1×
“MDP formulates the interaction between an actor and the environment [27].”
Reinforcement Learning: An Introduction
(2005) · IEEE Transactions on Neural Networks · doi:10.1109/tnn.2004.842673 · Cited by 24,496
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
A Survey of Monte Carlo Tree Search Methods
Browne, C. B., Powley, E., Whitehouse, D., Lucas, S. M., Cowling, P. I., Rohlfshagen, P., Tavener, S., Perez, D., Samothrakis, S., & Colton, S. (2012) · IEEE Transactions on Computational Intelligence and AI in Games · doi:10.1109/tciaig.2012.2186810 · Cited by 2,970 · cited 8×
“Their reasoning for using MCTS, especially Vanilla MCTS [11], was to provide faster data to the game designer.”
MethodInfluentialCuriosity-Driven Exploration by Self-Supervised Prediction
Pathak, D., Agrawal, P., Efros, A. A., & Darrell, T. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) · doi:10.1109/cvprw.2017.70 · Cited by 1,815 · cited 3×
“Researchers measured the uncertainty using bootstrapped DQN [24], statespace features [14], and error of a neural function [25].”
BackgroundMethod(2016) · neural information processing systems · Cited by 735
ViZDoom: A Doom-based AI research platform for visual reinforcement learning
Kempka, M., Wydmuch, M., Runc, G., Toczek, J., & Jaskowski, W. (2016) · 2016 IEEE Conference on Computational Intelligence and Games (CIG) · doi:10.1109/cig.2016.7860433 · Cited by 541 · cited 4×
“We use the GVG-AI [15] and VizDoom [16] environments to demonstrate our proposed methodologies.”
BackgroundMethodInfluentialExploration by Random Network Distillation
Burda, Y., Edwards, H., Storkey, A., & Klimov, O. (2018) · arXiv · doi:10.48550/arxiv.1810.12894 · Cited by 527 · cited 2×
“Researchers measured the uncertainty using bootstrapped DQN [24], statespace features [14], and error of a neural function [25].”
BackgroundMethodDeep Exploration via Bootstrapped DQN
Osband, I., Blundell, C., Pritzel, A., & Roy, B. V. (2016) · arXiv (Cornell University) · doi:10.48550/arxiv.1602.04621 · Cited by 434 · cited 3×
“Researchers measured the uncertainty using bootstrapped DQN [24], statespace features [14], and error of a neural function [25].”
Method
Research works printed but not linked: 1
- bartle, 2005, Virtual worlds: Why people play, Massively Multiplayer Game Development, 2, 3
Cited by
An Industrial Experience Leveraging the iv4XR Framework for BDD Testing of a 3D Sandbox Game
Ricós, F. P., Marín, B., Prasetya, I. S. W. B., Vos, T. E. J., Davidson, J., & Hovorka, K. (2024) · Lecture notes in business information processing · doi:10.1007/978-3-031-59465-6_24 · Cited by 4
A Literature Review of Software Testing Practices and Frameworks in the Video Gaming Industry
Roque, A., Sotomayor, J. P., Santiago, D., & Clarke, P. (2025) · Software Testing Verification and Reliability · doi:10.1002/stvr.70001 · Cited by 8
“Taesiri et al. [18] implement this by introducing an alternative pathfinder, which allows the agent to leverage previous gameplay trajectories in combination with altered rewards that fit the unique persona to create an agent capable of playing the same level differently.”
Behavior Driven Development for 3D games
Ricós, F. P., Marín, B., Prasetya, I. S. W. B., Vos, T. E. J., Davidson, J., & Hovorka, K. (2025) · Data & Knowledge Engineering · doi:10.1016/j.datak.2025.102486 · Cited by 3
Comparative Analysis of GAT and BERT for Human-Like Playtesting
Fragkedaki, K., Panagiotakopoulos, T., Biasielli, M., & Wang, H. (2025) · IEEE Conference on Games (CoG) · doi:10.1109/cog64752.2025.11114390 · Cited by 1
UX Design in Digital Game Development: A Systematic Literature Review and Research Agenda
da Silva, F. F., Machado, L. D. S., Leal, G. C. L., & Balancieri, R. (2026) · IEEE Access · doi:10.1109/access.2026.3652449 · Cited by 6
“During the evaluation phase, techniques such as playtest-ing [93] are applied, where users interact with the product and provide feedback on gameplay, usability, and errors.”
Metadata
- Authors
- Ariyurek, S., Surer, E., & Betin-Can, A.
- Affiliations
- Middle East Technical University (TR)
- Graduate School of Informatics, Middle East Technical University, Ankara, Turkey
- Venue
- IEEE Transactions on Games
- Published
- 2023-09
- Volume
- 15
- Issue
- 3
- Pages
- 348-359
- ISSN
- 2475-1502 (ISSN-L)
- 2475-1510
- OpenAlex
- W3185235798
- Cited by
- 7
- References
- 53
- Citations by year
- 2026: 2 · 2025: 3 · 2024: 1 · 2023: 1
- FWCI
- 0.41
- Type
- journal article
- Language
- en
- Open access
- green
- arxiv.org
- Licence
- Locations
- IEEE Transactions on Games (publishedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by)
- OpenMETU (Middle East Technical University) (submittedVersion, cc-by-nc-nd)
- arXiv (Cornell University) (cc-by)
- 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
- game testing, game design, reinforcement learning agents, player behavior modeling, ViZDoom, proximal policy optimization
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 2026-11-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-11-01 (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: Playtesting: What is Beyond Personas. https://app.cochat.ai/papers/10.1109/tg.2022.3165882
- The paper
- Ariyurek, S., Surer, E., & Betin-Can, A. (2023). Playtesting: What is Beyond Personas. IEEE Transactions on Games. https://doi.org/10.1109/tg.2022.3165882
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.