Pith. sign in

REVIEW 4 major objections 3 minor 33 references

Evolutionary Optimization of Deep Learning Agents for Sparrow Mahjong

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Evolution-trained LSTM matches PPO at Sparrow Mahjong

desk verdict The abstract claims a decent niche result for evolving LSTM policies in Sparrow Mahjong, but the full text is an unrelated graph-theory manuscript, so there is no paper to review yet. read the letter →

arxiv 2508.07522 v1 pith:IEO3LMOJ submitted 2025-08-11 cs.NE

Jim O'Connor , Derin Gezgin , Gary B. Parker This is my paper · ORCID
classification cs.NE
keywords SparrowMahjongCMA-ESLSTMevolutionaryoptimizationdeepreinforcementlearningpartialobservabilitygameAIpolicysearch
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

This paper claims that a deep LSTM policy trained by the black-box evolutionary optimizer CMA-ES can play Sparrow Mahjong at a level that beats random and rule-based agents and matches a Proximal Policy Optimization (PPO) reinforcement-learning baseline. The point is that gradient-free evolution can substitute for backpropagation-based RL in a stochastic, partially observable game, at lower training cost. If the claim holds, it supports hybrid deep-learning-plus-evolution as a practical route for strategic games and adaptive decision-making.

What carries the argument

The machinery is CMA-ES operating as a black-box optimizer over the weights of an LSTM policy. The LSTM supplies memory of past observations, which matters in a partially observable game where the current board state is not the full state; CMA-ES proposes weight vectors, evaluates each candidate by simulating many games, and iteratively updates a covariance model to concentrate sampling in promising regions. This replaces the policy-gradient credit-assignment loop of PPO with direct population-based search.

What would settle it

Run Evo-Sparrow against a converged PPO agent and a strong human benchmark on a publicly available Sparrow Mahjong environment with fixed seeds and full tile-set legality checks; if Evo-Sparrow's win rate falls below chance or clearly below PPO under those conditions, the comparable-performance claim fails. Simpler: if removing the LSTM's memory (replacing it with a feedforward net) does not change performance, the claimed role of memory in partial observability is not supported.

Watch

Extended reading notes

Core claim

The central claim is that Evo-Sparrow—an LSTM network whose weights are optimized by CMA-ES—evaluates board states in Sparrow Mahjong and chooses actions well enough to outperform random and rule-based agents while matching a Proximal Policy Optimization baseline over many simulations. The paper presents this as evidence that evolutionary optimization of memory-based deep networks can yield strong play in non-deterministic, partially observable environments without relying on gradient signals.

Load-bearing premise

The load-bearing premise is that the simulation environment and the board-state encoding used in training and evaluation faithfully represent Sparrow Mahjong, so that the measured win rates against random, rule-based, and PPO agents reflect genuine playing strength rather than artifacts of the simulator.

Editorial extensions

If this is right

  • If the reported win rates hold, Evo-Sparrow beats random play and rule-based agents in Sparrow Mahjong without any gradient-based training.
  • The reported parity with PPO suggests that population-based search over LSTM weights can reach the ballpark of policy-gradient RL in a stochastic, partially observable game.
  • Training cost is the claimed virtue: CMA-ES needs only game-simulation evaluations and no backpropagated gradients, making it usable where reward signals are sparse or non-differentiable.
  • The same hybrid recipe is proposed as a template for adaptive decision-making agents beyond Mahjong.

Reading between the lines

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

  • Editorial note: the full text supplied with this record is a different manuscript—on Turán numbers for generalized book graphs—so the specifics of Evo-Sparrow's architecture, hyperparameters, and simulation counts could not be checked from the paper body; the summary above draws only on the title, abstract, and author metadata.
  • A sharper test than the abstract's comparison would pit Evo-Sparrow against a converged PPO agent on a standardized, third-party Sparrow Mahjong simulator with fixed random seeds; 'comparable to a PPO baseline' leaves open how strong that baseline is.
  • The method transfers naturally to other imperfect-information tile games (Rummy, Hanafuda, Mahjong variants): swapping the encoder while keeping CMA-ES over LSTM weights would test whether the reported success comes from the memory architecture, the evolutionary search, or the specific game.
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.

Referee Report

4 major / 3 minor

Summary. The submission, titled 'Evolutionary Optimization of Deep Learning Agents for Sparrow Mahjong', presents an abstract claiming that a CMA-ES-optimized LSTM agent, Evo-Sparrow, outperforms random and rule-based agents and achieves performance comparable to a PPO baseline in Sparrow Mahjong. The abstract further asserts that empirical analysis over a significant number of simulations supports these claims. However, the full text of the provided manuscript is an unrelated, garbled extremal graph theory paper (arXiv:2508.07533v1, math.CO) on Turán numbers of generalized book graphs. It contains no mention of Sparrow Mahjong, LSTM, CMA-ES, PPO, board-state encoding, simulation environment, or any experimental results. The central claim is therefore based entirely on the abstract and has no supporting evidence in the manuscript body.

Significance. If the abstract's claims were supported by a proper methods and evaluation section, the result could be a useful demonstration of evolutionary optimization of deep memory-based policies for a stochastic, partially observable game, potentially offering a computationally efficient alternative to gradient-based RL. However, the submitted manuscript provides no such evidence. There are no machine-checked proofs, reproducible code, parameter-free derivations, or falsifiable predictions in the document; the full text is a different mathematical paper. Consequently, the contribution is currently unassessable, and the claimed significance cannot be verified.

major comments (4)
  1. [Full text (entire manuscript body)] The full text is an unrelated math.CO manuscript on extremal numbers of generalized book graphs (Theorem 1.4). It does not mention Sparrow Mahjong, LSTM, CMA-ES, PPO, board-state encoding, simulation environment, or any experimental results. The abstract's central claim—that Evo-Sparrow outperforms random and rule-based agents and matches a PPO baseline—has zero supporting evidence in the document. This is load-bearing because every downstream assertion of strategic play, robust policy quality, and computational efficiency depends on experiments that are not reported.
  2. [Abstract] The abstract states 'Empirical analysis conducted over a significant number of simulations' but provides no simulation counts, variance, error bars, statistical significance tests, or evaluation protocol. Without these, even a well-formed abstract would be insufficient to verify the comparison; in this manuscript the gap is absolute because there is no methods or results section.
  3. [Abstract (baseline comparison)] The comparison against random, rule-based, and PPO agents cannot be assessed. The PPO baseline's architecture, training procedure, hyperparameters, and evaluation conditions are unspecified; the rule-based agent's strategy is not described. Since no definitions appear anywhere in the full text, the claimed parity with PPO is unfalsifiable and the comparative claim is unsupported.
  4. [Full text (garbled encoding)] Apart from the content mismatch, the full text is corrupted by encoding errors (e.g., 'Tur´ an', '�', missing diacritics), making it difficult to read even as a graph theory paper. This presentation issue is secondary, but it compounds the impossibility of evaluating the submission.
minor comments (3)
  1. [Full text (header)] The displayed arXiv identifier in the full-text header is 2508.07533v1, which differs from the submitted arXiv number 2508.07522; this should be rectified if the correct manuscript is uploaded.
  2. [References] The mathematical references (e.g., Bollobás, West) are unrelated to the claimed topic and are not cited in support of any Evo-Sparrow experiments.
  3. [Abstract] The abstract uses 'our model' and 'Evo-Sparrow' without a concrete system description; even if the intended full text were present, the abstract would need a more precise statement of the agent architecture and training setting.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reduction found: the abstract's claims are unsupported by the enclosed full text, but that is an evidence failure, not a circular derivation.

full rationale

The paper presents an abstract claiming that an LSTM agent optimized by CMA-ES outperforms random and rule-based agents and matches a PPO baseline in Sparrow Mahjong, based on 'Empirical analysis conducted over a significant number of simulations.' The full text supplied, however, is an unrelated extremal graph theory manuscript on Turán numbers of generalized book graphs (arXiv:2508.07533v1, math.CO) and contains no description of the Sparrow Mahjong environment, the LSTM/CMA-ES training procedure, the PPO baseline, the rule-based agent, the board-state encoding, or any experimental results. This is a severe missing-support problem: the central claim is unfalsifiable from the provided document and no derivation chain exists to inspect. But circularity in the technical sense requires a specific reduction of a claimed result to its own inputs—e.g., a fitted parameter later called a prediction, or a theorem invoked from the authors' own prior work that assumes the conclusion. No such reduction appears here. There are no equations, no fitted parameters, and no self-citations in the abstract; the mismatch between abstract and full text does not make the claim circular, it makes it unsupported. Accordingly, the circularity score is 0, with the caveat that the manuscript's evidentiary integrity is independently compromised by the content mismatch.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

The central claim rests on unreported hyperparameters and two unverified domain assumptions: simulator fidelity and optimizer suitability. No independent evidence for either is visible in the abstract, and the body provides none. Note the ledger is necessarily incomplete because the actual methods text is absent.

free parameters (1)
  • LSTM architecture and CMA-ES hyperparameters (population size, network dimensions, fitness weights, mutation settings) = not reported in abstract
    The claimed performance depends on these hand-chosen or tuned values, and none are stated in the abstract or anywhere in the provided text.
assumptions (2)
  • domain assumption The simulation environment accurately models Sparrow Mahjong's rules, randomness, and partial observability.
    The abstract evaluates the agent 'over a significant number of simulations' in 'a non-deterministic, partially observable game environment'; if the simulator is flawed, the claimed win rates do not reflect actual play.
  • domain assumption CMA-ES can effectively optimize an LSTM policy without gradient information in a sequential, partially observable setting.
    The method's premise is that evolutionary search over policy weights produces competent play; the abstract asserts this outcome without supporting analysis.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evolutionary Optimization of Deep Learning Agents for Sparrow Mahjong." pith.science (2026). https://pith.science/paper/IEO3LMOJ

@misc{pith2026250807522,
  author       = {Pith},
  title        = {Pith review of: Evolutionary Optimization of Deep Learning Agents for Sparrow Mahjong},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IEO3LMOJ}},
  note         = {Machine review of arXiv:2508.07522}
}
read the original abstract

We present Evo-Sparrow, a deep learning-based agent for AI decision-making in Sparrow Mahjong, trained by optimizing Long Short-Term Memory (LSTM) networks using Covariance Matrix Adaptation Evolution Strategy (CMA-ES). Our model evaluates board states and optimizes decision policies in a non-deterministic, partially observable game environment. Empirical analysis conducted over a significant number of simulations demonstrates that our model outperforms both random and rule-based agents, and achieves performance comparable to a Proximal Policy Optimization (PPO) baseline, indicating strong strategic play and robust policy quality. By combining deep learning with evolutionary optimization, our approach provides a computationally effective alternative to traditional reinforcement learning and gradient-based optimization methods. This research contributes to the broader field of AI game playing, demonstrating the viability of hybrid learning strategies for complex stochastic games. These findings also offer potential applications in adaptive decision-making and strategic AI development beyond Sparrow Mahjong.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 24 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Azizzadenesheli, K.; Yue, Y.; and Anandkumar, A. 2018. Policy gradient in partially observable environments: Approximation and convergence. arXiv preprint arXiv:1810.07900

    arXiv 2018

  4. [4]

    N.; Chandar, S.; Burch, N.; Lanctot, M.; Song, H

    Bard, N.; Foerster, J. N.; Chandar, S.; Burch, N.; Lanctot, M.; Song, H. F.; Parisotto, E.; Dumoulin, V.; Moitra, S.; Hughes, E.; Dunning, I.; Mourad, S.; Larochelle, H.; Bellemare, M. G.; and Bowling, M. 2020. The Hanabi challenge: A new frontier for AI research. Artif. Intell., 280(C)

    work page 2020

  5. [5]

    Brown, N.; and Sandholm, T. 2019. Superhuman AI for multiplayer poker. Science, 365: 885 -- 890

    work page 2019

  6. [6]

    Brown, N.; Sandholm, T.; and Amos, B. 2018. Depth-limited solving for imperfect-information games. In Proceedings of the 32nd International Conference on Neural Information Processing Systems, NIPS'18, 7674–7685. Red Hook, NY, USA: Curran Associates Inc

    work page 2018

  7. [7]

    J.; and Hsu, F.-h

    Campbell, M.; Hoane, A. J.; and Hsu, F.-h. 2002. Deep Blue. Artif. Intell., 134(1–2): 57–83

    work page 2002

  8. [8]

    Chen, J.; Tang, S.; and Wu, I. 2022. Monte-Carlo Simulation for Mahjong. Journal of Information Science and Engineering, 38(4): 775--790

    work page 2022

Show all 33 references
  1. [9]

    Coulom, R. 2006. Efficient Selectivity and Backup Operators in Monte-Carlo Tree Search. In Computers and Games, volume 4630, 72--83. Springer. ISBN 978-3-540-75537-1

    2006

  2. [10]

    Gao, S.; Okuya, F.; Kawahara, Y.; and Tsuruoka, Y. 2019. Building a Computer Mahjong Player via Deep Convolutional Neural Networks. ArXiv, abs/1906.02146

    2019 arXiv

  3. [11]

    Gelly, S.; Kocsis, L.; Schoenauer, M.; Sebag, M.; Silver, D.; Szepesv\' a ri, C.; and Teytaud, O. 2012. The grand challenge of computer Go: Monte Carlo tree search and extensions. Commun. ACM, 55(3): 106–113

    2012

  4. [12]

    Hochreiter, S.; and Schmidhuber, J. 1997. Long Short-Term Memory. Neural Comput., 9(8): 1735–1780

    1997

  5. [13]

    E.; and Moore, R

    Knuth, D. E.; and Moore, R. W. 1975. An Analysis of Alpha-Beta Pruning. Artificial Intelligence, 6: 293--326

    1975

  6. [14]

    Koyamada, S.; Okano, S.; Nishimori, S.; Murata, Y.; Habara, K.; Kita, H.; and Ishii, S. 2023. Pgx: Hardware-Accelerated Parallel Game Simulators for Reinforcement Learning. In Advances in Neural Information Processing Systems, volume 36, 45716--45743. Curran Associates, Inc

    2023

  7. [15]

    Li, J.; Koyamada, S.; Ye, Q.; Liu, G.; Wang, C.; Yang, R.; Zhao, L.; Qin, T.; Liu, T.; and Hon, H. 2020. Suphx: Mastering Mahjong with Deep Reinforcement Learning. CoRR, abs/2003.13590

    2020 arXiv

  8. [16]

    Li, J.; Wu, S.; Fu, H.; Fu, Q.; Zhao, E.; and Xing, J. 2022. Speedup Training Artificial Intelligence for Mahjong via Reward Variance Reduction. In 2022 IEEE Conference on Games (CoG), 345--352

    2022

  9. [17]

    Li, X.; Liu, B.; Wei, Z.; Wang, Z.; and Wu, L. 2024. Tjong: A transformer‐based Mahjong AI via hierarchical decision‐making and fan backward. CAAI Transactions on Intelligence Technology, 9(4): 982–995

    2024

  10. [18]

    Lu, Y.; Li, W.; and Li, W. 2023. Official International Mahjong: A New Playground for AI Research. Algorithms, 16: 235

    2023

  11. [19]

    Mizukami, N.; and Tsuruoka, Y. 2015. Building a computer Mahjong player based on Monte Carlo simulation and opponent models. In 2015 IEEE Conference on Computational Intelligence and Games (CIG), 275--283

    2015

  12. [20]

    Mnih, V.; Kavukcuoglu, K.; Silver, D.; Graves, A.; Antonoglou, I.; Wierstra, D.; and Riedmiller, M. A. 2013. Playing Atari with Deep Reinforcement Learning. CoRR, abs/1312.5602

    2013 arXiv

  13. [21]

    Ostermeier, A.; Gawelczyk, A.; and Hansen, N. 1994. A Derandomized Approach to Self-Adaptation of Evolution Strategies. Evolutionary Computation, 2(4): 369--380

    1994

  14. [22]

    Samuel, A. L. 1959. Some Studies in Machine Learning Using the Game of Checkers. IBM Journal of Research and Development , 3(3): 210--229

    1959

  15. [23]

    P.; and Silver, D

    Schrittwieser, J.; Antonoglou, I.; Hubert, T.; Simonyan, K.; Sifre, L.; Schmitt, S.; Guez, A.; Lockhart, E.; Hassabis, D.; Graepel, T.; Lillicrap, T. P.; and Silver, D. 2019. Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model. CoRR, abs/1911.08265

    2019 arXiv

  16. [24]

    Schulman, J.; Wolski, F.; Dhariwal, P.; Radford, A.; and Klimov, O. 2017. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347

    2017 arXiv

  17. [25]

    Shannon, C. E. 1950. Programming a Computer for Playing Chess. Philosophical Magazine, 41: 256--275

    1950

  18. [26]

    J.; Guez, A.; Sifre, L.; Van Den Driessche, G.; Schrittwieser, J.; Antonoglou, I.; Panneershelvam, V.; Lanctot, M.; et al

    Silver, D.; Huang, A.; Maddison, C. J.; Guez, A.; Sifre, L.; Van Den Driessche, G.; Schrittwieser, J.; Antonoglou, I.; Panneershelvam, V.; Lanctot, M.; et al. 2016. Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587): 484--489

    2016

  19. [27]

    P.; Simonyan, K.; and Hassabis, D

    Silver, D.; Hubert, T.; Schrittwieser, J.; Antonoglou, I.; Lai, M.; Guez, A.; Lanctot, M.; Sifre, L.; Kumaran, D.; Graepel, T.; Lillicrap, T. P.; Simonyan, K.; and Hassabis, D. 2017 a . Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm. CoR...

    2017 arXiv

  20. [28]

    Silver, D.; Schrittwieser, J.; Simonyan, K.; Antonoglou, I.; Huang, A.; Guez, A.; Hubert, T.; Baker, L.; Lai, M.; Bolton, A.; Chen, Y.; Lillicrap, T.; Hui, F.; Sifre, L.; van den Driessche, G.; Graepel, T.; and Hassabis, D. 2017 b . Mastering the game of Go without human knowl...

    2017

  21. [29]

    Tang, S.-C.; Chen, J.-C.; and Wu, I.-C. 2025. An Efficient Method for Assessing the Strength of Mahjong Programs. In Proceedings of the 17th International Conference on Agents and Artificial Intelligence - Volume 2: ICAART, 124--132. INSTICC, SciTePress. ISBN 978-989-758-737-5

    2025

  22. [30]

    Tesauro, G. 1995. Temporal difference learning and TD-Gammon. Commun. ACM, 38(3): 58–68

    1995

  23. [31]

    Truong, T.-D. 2021. A Supervised Attention-Based Multiclass Classifier for Tile Discarding in Japanese Mahjong. Master's thesis, University of Agder, Grimstad, Norway

    2021

  24. [32]

    W.; and van Rijswijck , J

    van den Herik , H.; Uiterwijk, J. W.; and van Rijswijck , J. 2002. Games solved: Now and in the future. Artificial Intelligence

    2002

  25. [33]

    Zhao, X.; and Holden, S. 2022. Building a 3-Player Mahjong AI using Deep Reinforcement Learning. ArXiv, abs/2202.12847

    2022 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.