Play Style Identification Using Low-Level Representations of Play Traces in MicroRTS

Jeremy Gow; Ruizhe Yu Xia; Simon Lucas

REVIEW 3 major objections 6 minor 16 references

Play Style Identification Using Low-Level Representations of Play Traces in MicroRTS

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Unsupervised CNN-LSTM autoencoders identify MicroRTS play styles from raw traces, beating handcrafted features.

desk verdict A modest but real contribution: unsupervised low-level autoencoder embeddings separate MicroRTS agents better than handcrafted features, with an honest qualitative analysis that also exposes the method's temporal-window limitation. read the letter →

arxiv 2507.10172 v1 pith:QLNVTYWV submitted 2025-07-14 cs.LG cs.AI

Ruizhe Yu Xia , Jeremy Gow , Simon Lucas This is my paper · ORCID

classification cs.LGcs.AI

keywords playstyleidentificationMicroRTSCNN-LSTMautoencoderlatentrepresentationunsupervisedclusteringgameanalyticsdeeplearning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that an unsupervised CNN-LSTM autoencoder can learn latent representations of low-level MicroRTS play traces—raw grid-encoded states and actions—that separate different AI agents' play styles better than handcrafted aggregate features. On the held-out test map, the joint state-action model reaches an adjusted mutual information of 0.715 against agent labels, versus 0.675 for handcrafted features. If this holds, play-style identification can skip the domain-expert feature design step and let styles emerge directly from data, improving portability across games. The latent space also serves as a qualitative analysis tool, surfacing within-agent behaviors such as Droplet's worker-rush tendencies and mayari's path-disruption glitches.

What carries the argument

The load-bearing component is the reconstruction-based CNN-LSTM autoencoder: a time-agnostic CNN frame encoder (two blocks of convolution, max pooling, and GELU) maps each grid frame to a compact vector, and two stacked bidirectional LSTM layers compress the sequence into a 1024-dimension latent code. The decoder, built from bidirectional LSTMs and a transposed-CNN frame decoder, reconstructs the original sequence from that code. Training the autoencoder to reconstruct low-level state and action tensors forces the latent code to preserve the information needed to regenerate the play trace, and clustering that code with k-means after PCA reduction to 64 dimensions is what yields the agent separation.

What would settle it

Hold out a map or agent set not used in training and compare the joint autoencoder's clustering AMI against the 18-feature handcrafted baseline; if the learned embedding fails to exceed the baseline (which the paper reports at 0.675 AMI on map L) or fails to separate a known behaviorally distinct agent, the generalizability claim would be refuted. A more direct test is to run the same pipeline on traces truncated to different starting windows, such as 16 and 64 steps, and check whether cluster quality changes drastically, which would confirm the dependency on the starting-subsequence assumption.

Watch

Extended reading notes

Core claim

The authors establish that encoding each frame with a shared CNN, summarizing the frame sequence with stacked bidirectional LSTMs, and reconstructing it through a BiLSTM plus transpose-CNN decoder produces a 1024-dimensional latent code from which k-means clusters recover agent identity as well as or better than handcrafted features do. Actions-only and joint state-action embeddings outperform states-only and handcrafted baselines on completeness, homogeneity, adjusted rand index, and adjusted mutual information, both averaged over maps A–K and on the held-out map L. The method is unsupervised in training: no agent labels or style annotations are used to train the autoencoder; labels are used only to evaluate the resulting clusters.

Load-bearing premise

The approach clusters only the first 32-step subsequence of each game, assuming that this opening segment reliably represents an agent's play style; the authors observe that some style-relevant behavior, like mayari's erratic worker pathing, corrects itself only after this window, so styles that manifest later in a match could be misattributed.

Editorial extensions

If this is right

If the learned embeddings are genuinely style-relevant, game analytics can drop handcrafted feature design and apply the same autoencoder pipeline to new games with only an observation-encoding step.
The latent space can slice a single agent into behavioral modes—Droplet splits between worker-rush and MCTS-like clusters—enabling data-driven discovery of sub-styles and failure states within one AI.
Because the approach is unsupervised, it can scale to unlabeled player data, producing style taxonomies from raw behavior rather than from designer-defined categories.
The actions-only and joint embeddings outperforming states-only suggests action history carries much of the style signal, guiding future representations toward action-conditioned encoders.
The clustering also flags anomalous traces, such as crashed agents clustering with PassiveAI, giving a cheap way to detect data-collection failures.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension the authors do not pursue is applying the same encoder to human play traces, which could make play-style identification in commercial games feasible from telemetry alone, though human style drift within a session may require segmenting traces at behavior-change points rather than fixed windows.
The starting-window limitation suggests a testable improvement: replace fixed 32-step subsequences with adaptive segmentation, such as change-point detection on the latent code, so that styles appearing mid-match like mayari's erratic pathing are captured.
The joint model's edge on map L hints that state information helps most when maps are harder to cluster; one could test this by varying map geometry while holding agents fixed.
Because the autoencoder must be retrained for each game's observation structure, the portability claim would be truly tested only by a shared cross-game observation encoder.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Worth a look if you work on play style or behavioral cloning from traces. The paper does something new in a modest way: it trains a CNN-LSTM autoencoder directly on raw MicroRTS observation and action tensors, then clusters the latent vectors, and shows that this separates agents better than an 18-feature handcrafted baseline (joint model AMI 0.715 vs 0.675 on map L). The method itself is an application of existing components, but the low-level representation and the honest qualitative analysis make it a useful data point.

The best part is Section IV-B. They actually inspect misclustered traces and find interpretable causes: MixedBot and Tiamat merge because Tiamat is in MixedBot's ensemble, traces from agents killed by WorkerRush never get to express a style, and mayari's 'erratic' worker path is corrected after the chosen 32-step window. That kind of self-criticism is rare and should count in the paper's favor.

The soft spots are real but not disqualifying. The main quantitative claim rests entirely on the first 32 action steps. The authors are explicit about this, and they even state that mayari's style-relevant behavior appears outside that window, so the abstract's 'meaningful separation of agents' play styles' is only demonstrated for an early-game slice. A reviewer should ask for stability across temporal windows or a clear scope statement. Also missing: error bars and significance tests (though Table I is an average over maps A-K, which helps a little), a principled choice of k (10/13/16 is arbitrary), and a comparison to the deep-embedding baselines in [12] and [13]. They compare only to their own 18 handcrafted features, which is a weak baseline. No code or data is released, which for a paper whose selling point is 'low-level, no domain expertise' is a missed opportunity.

The circularity worry does not land: the latent space is trained without agent labels, and labels are only used for evaluation. The weakest assumption is the temporal window, not the unsupervised setup.

Bottom line: this is a serious short paper with a modest but real contribution. It deserves peer review, conditional on the authors either widening the temporal analysis or narrowing the claim, and releasing code/data. I'd bring it to a reading group if the topic is on the table, and I'd cite it as an example of low-level unsupervised style embedding in RTS.

Referee Report

3 major / 6 minor

Summary. The paper proposes an unsupervised CNN-LSTM autoencoder that maps low-level MicroRTS play traces (state tensors and action tensors) into a 1024-dimensional latent space, followed by PCA and k-means clustering to identify play styles. Experiments use 13 MicroRTS agents on 12 maps, training on maps A-K and testing on map L, with subsequences of length 32. The authors compare states-only, actions-only, joint, and handcrafted feature embeddings using completeness, homogeneity, ARI, and AMI, and perform a qualitative t-SNE analysis. The main claim is that the learned latent representations yield meaningful separation of different agents' play styles without handcrafted features, with the joint model achieving AMI 0.715 on map L.

Significance. If the result holds, the paper makes a useful contribution: it demonstrates that a purely data-driven, low-level representation can support play style clustering in a real-time strategy game, reducing reliance on domain-specific feature engineering. The experimental design has clear strengths: a held-out map (L) for testing, multiple clustering metrics, and qualitative trace inspection that connects clusters to concrete behaviors (e.g., mayari's worker-path disruption, MixedBot/Tiamat merging, WorkerRush-victim traces). The authors also correctly note that agent identity is not identical to play style. However, the evidence is currently weakened by the lack of statistical robustness analysis and by the exclusive use of a single short temporal window, as detailed in the major comments.

major comments (3)

[Section IV-A and Table I] The quantitative claim of meaningful separation is based exclusively on the starting 32-step subsequence of each trace. Section IV-B itself states that mayari's style-relevant erratic behavior is 'later corrected outside of our chosen subsequence length' and that traces of agents defeated by WorkerRush-like strategies 'did not have the time to show their playstyle before being defeated.' Because no other temporal window or full-trace evaluation is reported, the headline AMI values (e.g., 0.715 for the joint model on map L) may be specific to this narrow early-game slice. Please report clustering metrics for multiple subsequence positions and lengths (early, mid, late, and full traces) and show whether the separation is stable; otherwise the abstract's general claim about play style separation is not fully supported.
[Table I and Section III-C] All clustering metrics are point estimates from a single training/clustering run. The autoencoder training is stochastic (random initialization, data augmentation) and k-means is initialization-dependent, yet no variance, confidence intervals, or significance tests are reported. Differences such as the joint model's AMI 0.715 versus the handcrafted baseline's 0.675 on map L may be within run-to-run noise. Please report means and standard deviations over multiple seeds and perform a significance test or bootstrap comparison for the central learned-versus-handcrafted comparisons.
[Sections III-D and IV-A] The quantitative evaluation uses agent names as ground-truth labels for clustering metrics. The paper correctly acknowledges that agent identity is not the same as play style, and the qualitative analysis adds useful context. Nevertheless, AMI/ARI measure recovery of agent labels, not recovery of play styles per se. A quantitative validation against actual style labels (e.g., human annotation of a sample of traces or a downstream style-based task) would strengthen the central claim. At minimum, the manuscript should explicitly state this limitation in the evaluation section rather than only in passing.

minor comments (6)

[Table I] The table header 'E' is not expanded in the caption; please spell out 'Embedding' and 'Handcrafted' for readability.
[Section III-B] Please clarify how the data augmentation (random vertical/horizontal mirroring) transforms the action direction parameters and the relative attack position features. A spatial flip changes absolute directions and relative coordinates, so the augmentation is only valid if these features are transformed consistently.
[Section III-B and Table I] The rows labeled 'A-K' are computed on maps used for training and validation of the autoencoder. These are training-domain diagnostics, not generalization results; please label them as such and emphasize the held-out map L results in the main text.
[Section III-C] For reproducibility, please report training hyperparameters such as learning rate, optimizer, number of epochs, loss weights, and hardware/software environment, and consider releasing code and trained models.
[Section IV-B] The t-SNE visualization is useful, but the claim that k-means separates NaiveMCTS from random variants 'quite successfully' would be better supported by reporting the confusion matrix or per-agent cluster assignment counts for map L.
[References] Reference [15] contains a typo: 'Inteligence' should be 'Intelligence'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the unsupervised embeddings are evaluated against held-out agent labels that were never used in training, and the cited prior work is external rather than self-referential.

full rationale

The paper's derivation chain is self-contained. Low-level play traces from maps A-K are used to train a CNN-LSTM autoencoder without any ground-truth labels, and the test set is explicitly a held-out map: 'The test set is comprised of play traces played in map L.' Clustering is then applied to the latent representations of these test traces, and agent names are used only as evaluation labels: 'Because different AI agents are being used to generate the play traces, we can assign the agent name as the ground truth label. Note how this is not a play style label.' The reported AMI/ARI scores therefore measure agreement between unsupervised clusters and external agent identities; nothing in the training objective or clustering procedure is fitted to those identities. The handcrafted baseline is independently constructed from aggregate action frequencies and is not derived from the learned latent space. The architecture follows external prior work (references [13] and [14]) rather than the authors' own prior results, so no load-bearing self-citation chain exists. The main weakness noted in Section IV-B, that mayari's erratic style-relevant behavior is 'corrected outside of our chosen subsequence length,' is a temporal-window validity concern about whether the starting subsequence fully represents play style; it is not circular because the subsequence choice was fixed before clustering and is not an output of the clustering procedure. Similarly, the observation that some traces 'did not have time to show their playstyle before being defeated' is an interpretive limitation, not a case of the target result being reused as an input. Overall, no prediction in the paper reduces by construction to a fitted parameter or to a self-citation, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim depends on a handful of manually chosen hyperparameters (cluster count, sequence length, latent size, PCA dimension) and on the unvalidated assumptions that play style is visible in short low-level traces and that agent labels are a good proxy for style. No new entities are introduced.

free parameters (6)

Number of clusters k = 10, 13, 16
Chosen manually to allow comparison; affects clustering metrics.
Subsequence length = 32
Fixed length for sequence extraction; longer behaviors may be missed.
Stride = 8 (training), 32 (test/validation)
Overlap controls data augmentation and coverage.
Latent vector size = 1024
Architecture hyperparameter, not tuned.
PCA dimension = 64
Dimensionality reduction before k-means.
Autoencoder training hyperparameters (learning rate, optimizer, epochs, loss weights)
Not reported in the paper; needed to replicate training.

assumptions (3)

domain assumption Play style is observable in the low-level state/action subsequences
The entire method assumes style information is present in the raw traces; the paper does not independently validate this beyond clustering against agent labels.
domain assumption Agent names are a valid proxy ground truth for play style
Used to evaluate clustering; the paper acknowledges agents may share styles or change behavior, but still uses completeness/homogeneity against these labels.
ad hoc to paper The CNN-LSTM autoencoder preserves style-relevant variation under reconstruction loss
The reconstruction objective does not explicitly encourage style separation; the paper assumes the latent space captures it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Play Style Identification Using Low-Level Representations of Play Traces in MicroRTS." pith.science (2026). https://pith.science/paper/QLNVTYWV

@misc{pith2026250710172,
  author       = {Pith},
  title        = {Pith review of: Play Style Identification Using Low-Level Representations of Play Traces in MicroRTS},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QLNVTYWV}},
  note         = {Machine review of arXiv:2507.10172}
}

read the original abstract

Play style identification can provide valuable game design insights and enable adaptive experiences, with the potential to improve game playing agents. Previous work relies on domain knowledge to construct play trace representations using handcrafted features. More recent approaches incorporate the sequential structure of play traces but still require some level of domain abstraction. In this study, we explore the use of unsupervised CNN-LSTM autoencoder models to obtain latent representations directly from low-level play trace data in MicroRTS. We demonstrate that this approach yields a meaningful separation of different game playing agents in the latent space, reducing reliance on domain expertise and its associated biases. This latent space is then used to guide the exploration of diverse play styles within studied AI players.

Figures

Figures reproduced from arXiv: 2507.10172 by the authors.

**Figure 1.** Example of MicroRTS map. Players are represented by red and blue borders and start with a worker (grey circle), a base (grey square), and two resource mines (green squares). III. METHODS A. Data Description Our data is represented like [14] with modifications: 1) We unify the direction parameters into a single direction parameter to reduce dimensionality without loss of information. 2) We use a two-dimensional repre… view at source ↗

**Figure 2.** Model architecture for the actions-only autoencoder with output sizes. Batch size is omitted for clarity. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** t-SNE for joint model embeddings of map L games with original labels (left) and 13 k-means clusters (right). Best visualized in the electronic version. V. CONCLUSION This work explores the use of CNN-LSTM autoencoders to learn embeddings of low-level representations of play traces to identify play styles, reducing reliance on domain expertise during feature creation. By using raw action and state sequences instead o… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 15 canonical work pages

[12]

An unsupervised video game playstyle metric via state discretization,

C.-C. Lin, W.-C. Chiu, and I.-C. Wu, “An unsupervised video game playstyle metric via state discretization,” in Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence . PMLR, Dec. 2021

work page 2021
[13]

Play-style Identification through Deep Unsupervised Clustering of Trajectories,

B. Ingram, B. Rosman, C. V . Alten, and R. Klein, “Play-style Identification through Deep Unsupervised Clustering of Trajectories,” in 2022 IEEE Conference on Games (CoG) . Beijing, China: IEEE, Aug. 2022

work page 2022
[1]

Counter-Strike Deathmatch with Large-Scale Behavioural Cloning

T. Pearce and J. Zhu, “Counter-Strike Deathmatch with Large-Scale Behavioural Cloning,” Dec. 2021, arXiv:2104.04258 [cs, stat]

work page Pith review arXiv 2021
[2]

Creating Diverse Play-Style-Centric Agents through Behavioural Cloning,

B. Ingram, C. V . Alten, R. Klein, and B. Rosman, “Creating Diverse Play-Style-Centric Agents through Behavioural Cloning,” Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, vol. 19, no. 1, Oct. 2023

work page 2023
[3]

Learning Models of Individual Behavior in Chess,

R. McIlroy-Young, R. Wang, S. Sen, J. Kleinberg, and A. Anderson, “Learning Models of Individual Behavior in Chess,” in Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Aug. 2022, arXiv:2008.10086 [cs]

arXiv 2022
[4]

Hearts, Clubs, Diamonds, Spades: Players Who Suit MUDs,

R. Bartle, “Hearts, Clubs, Diamonds, Spades: Players Who Suit MUDs,” Journal of MUD research , 1996

work page 1996
[5]

Motivations for Play in Online Games,

N. Yee, “Motivations for Play in Online Games,” CyberPsychology & Behavior, vol. 9, no. 6, Dec. 2006

work page 2006
[6]

Defining personas in games using metrics,

A. Tychsen and A. Canossa, “Defining personas in games using metrics,” in Proceedings of the 2008 Conference on Future Play: Research, Play, Share. Toronto Ontario Canada: ACM, Nov. 2008

work page 2008

Show all 16 references

[7]

Player behavioural modelling for video games,

S. C. J. Bakkes, P. H. M. Spronck, and G. van Lankveld, “Player behavioural modelling for video games,” Entertainment Computing , vol. 3, no. 3, Aug. 2012

2012
[8]

Player modeling using self-organization in Tomb Raider: Underworld,

A. Drachen, A. Canossa, and G. N. Yannakakis, “Player modeling using self-organization in Tomb Raider: Underworld,” in 2009 IEEE Symposium on Computational Intelligence and Games , Sep. 2009

2009
[9]

Guns, swords and data: Clustering of player behavior in computer games in the wild,

A. Drachen, R. Sifa, C. Bauckhage, and C. Thurau, “Guns, swords and data: Clustering of player behavior in computer games in the wild,” in 2012 IEEE Conference on Computational Intelligence and Games (CIG), Sep. 2012

2012
[10]

Guns and guardians: Comparative cluster analysis and behavioral profiling in Destiny,

A. Drachen, J. Green, C. Gray, E. Harik, P. Lu, R. Sifa, and D. Klabjan, “Guns and guardians: Comparative cluster analysis and behavioral profiling in Destiny,” in 2016 IEEE Conference on Computational Intelligence and Games (CIG) , Sep. 2016

2016
[11]

Like a DNA String: Sequence-Based Player Profiling in Tom Clancy’s The Division,

A. Canossa, S. Makarovych, J. Togelius, and A. Drachen, “Like a DNA String: Sequence-Based Player Profiling in Tom Clancy’s The Division,” Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment , vol. 14, no. 1, Sep. 2018

2018
[14]

Gym-µRTS: Toward Affordable Full Game Real-time Strategy Games Research with Deep Reinforcement Learning,

S. Huang, S. Onta ˜n´on, C. Bamford, and L. Grela, “Gym-µRTS: Toward Affordable Full Game Real-time Strategy Games Research with Deep Reinforcement Learning,” in 2021 IEEE Conference on Games (CoG), Copenhagen, Denmark, August 17-20, 2021 . IEEE, 2021

2021
[15]

Behavior evolution in Tomb Raider Underworld,

R. Sifa, A. Drachen, C. Bauckhage, C. Thurau, and A. Canossa, “Behavior evolution in Tomb Raider Underworld,” in 2013 IEEE Conference on Computational Inteligence in Games (CIG) , Aug. 2013

2013
[16]

Playing Styles in Starcraft,

Y . N. Ravari, S. Bakkes, and P. Spronck, “Playing Styles in Starcraft,” in European GAME-ON Conference on Simulation and AI in Computer Games, 2018

2018

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [12]

An unsupervised video game playstyle metric via state discretization,

C.-C. Lin, W.-C. Chiu, and I.-C. Wu, “An unsupervised video game playstyle metric via state discretization,” in Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence . PMLR, Dec. 2021

work page 2021

[2] [13]

Play-style Identification through Deep Unsupervised Clustering of Trajectories,

B. Ingram, B. Rosman, C. V . Alten, and R. Klein, “Play-style Identification through Deep Unsupervised Clustering of Trajectories,” in 2022 IEEE Conference on Games (CoG) . Beijing, China: IEEE, Aug. 2022

work page 2022

[3] [1]

Counter-Strike Deathmatch with Large-Scale Behavioural Cloning

T. Pearce and J. Zhu, “Counter-Strike Deathmatch with Large-Scale Behavioural Cloning,” Dec. 2021, arXiv:2104.04258 [cs, stat]

work page Pith review arXiv 2021

[4] [2]

Creating Diverse Play-Style-Centric Agents through Behavioural Cloning,

B. Ingram, C. V . Alten, R. Klein, and B. Rosman, “Creating Diverse Play-Style-Centric Agents through Behavioural Cloning,” Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, vol. 19, no. 1, Oct. 2023

work page 2023

[5] [3]

Learning Models of Individual Behavior in Chess,

R. McIlroy-Young, R. Wang, S. Sen, J. Kleinberg, and A. Anderson, “Learning Models of Individual Behavior in Chess,” in Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Aug. 2022, arXiv:2008.10086 [cs]

arXiv 2022

[6] [4]

Hearts, Clubs, Diamonds, Spades: Players Who Suit MUDs,

R. Bartle, “Hearts, Clubs, Diamonds, Spades: Players Who Suit MUDs,” Journal of MUD research , 1996

work page 1996

[7] [5]

Motivations for Play in Online Games,

N. Yee, “Motivations for Play in Online Games,” CyberPsychology & Behavior, vol. 9, no. 6, Dec. 2006

work page 2006

[8] [6]

Defining personas in games using metrics,

A. Tychsen and A. Canossa, “Defining personas in games using metrics,” in Proceedings of the 2008 Conference on Future Play: Research, Play, Share. Toronto Ontario Canada: ACM, Nov. 2008

work page 2008