Pith. sign in

REVIEW 3 major objections 6 minor 73 references

OpenSTARLab: Open Approach for Spatio-Temporal Agent Data Analysis in Soccer

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read OpenSTARLab claims that standardizing soccer event and tracking data into unified formats enables fair model comparisons, and under that pipeline a lightweight three-event model (LEM 3) beats transformer-based baselines on action and time…

desk verdict A genuinely useful open-source framework for soccer data pipelines, with a benchmark that needs stronger evaluation before its headline claim holds. read the letter →

arxiv 2502.02785 v2 pith:G5GLMU27 submitted 2025-02-05 cs.LG

Calvin Yeung , Kenjiro Ide , Taiga Someya , Keisuke Fujii This is my paper · ORCID
classification cs.LG
keywords socceranalyticsspatio-temporaldataeventpredictionstandardizationreinforcementlearningmulti-agentRLopen-sourcesoftwareUIEDformat
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 the main barrier to soccer analytics is not model quality but data plumbing: event and tracking data arrive in incompatible vendor formats, and annotation tools are scarce. The authors build an open-source framework that standardizes event data into the Unified and Integrated Event Data (UIED) format, converts frame-level tracking plus event data into a State-Action-Reward (SAR) format for reinforcement learning, and provides a video labeling tool. Benchmarking four published deep-learning event predictors plus a majority-class baseline under this standardized pipeline on two large event datasets, they report that LEM 3, a lightweight model conditioned on the three most recent events, achieves the best action accuracy, F1 score, and time prediction error, while requiring orders of magnitude less compute than transformer-based rivals. In the reinforcement-learning experiments, they find that increasing the weight on imitating observed actions improves action accuracy but raises temporal-difference loss, destabilizing reward prediction. A sympathetic reader would take away that standardized data formats can make model comparisons meaningful, that context length can matter more than model scale for next-event prediction, and that the action-imitation/reward-prediction trade-off is a real constraint in soccer RL.

What carries the argument

The central mechanism is the standardized UIED schema, which maps event types from four providers to a common action taxonomy on a 105-by-68 pitch with derived features such as distance and angle to goal, deltas, and possession-terminating markers, plus the SAR schema that aligns frame-level tracking with actions and rewards. On top of that, the paper's headline model, LEM 3, is a non-end-to-end stack of three multilayer perceptrons fed the three most recent events; comparing it against transformer-encoder and transformer-decoder baselines isolates the effect of context length versus architectural scale. In the reinforcement-learning arm, the load-bearing construction is the composite loss $L_{\mathrm{total}} = L_{\mathrm{td}} + \lambda_1 L_{\mathrm{L1}} + \lambda_2 L_{\mathrm{as}}$, whose sweep over $\lambda_1$ produces the reported trade-off between action accuracy and temporal-difference loss.

What would settle it

Re-benchmark the same models with compute-matched or convergence-based tuning—for instance, allowing the transformer models substantially more tuning trials or early stopping on validation loss until all models have converged—and check whether LEM 3 still tops action accuracy and F1; the ranking changing would show that the headline result depends on the tuning budget. A second check: train the RL models with a dense per-step reward instead of a sparse terminal reward and test whether the action-accuracy/temporal-difference trade-off persists; if it collapses, the reported trade-off is an artifact of the sparse reward design.

Watch

Extended reading notes

Core claim

Under the UIED-standardized preprocessing, on the Wyscout and StatsBomb datasets, the LEM 3 model—three independently trained multilayer perceptrons that predict next-event action, time, and coordinates from the previous three events—reaches the top action accuracy (0.67 on Wyscout, 0.66 on StatsBomb), top F1 score (0.20 and 0.25), and lowest time MAE (2.69 seconds and 2.07 seconds), while using roughly 19-20 million FLOPs compared with hundreds of millions to billions for the transformer-based baselines. The paper also reports that NMSTPP remains the best at spatial coordinate prediction overall, and that LEM 3's event simulation stays competitive for about ten timesteps and then decays relative to the majority-class baseline. In the RLearn experiments, the paper claims that with the combined temporal-difference, L1, and action-supervised loss, raising the action-supervision coefficient improves action accuracy but increases the temporal-difference loss, so the two objectives trade off; at the lowest coefficient tested, all models' action accuracy collapses toward the random level. These results are presented as evidence that the framework enables reproducible benchmarking and that a lightweight, context-limited model can outperform much larger models on this task.

Load-bearing premise

The load-bearing premise is that giving every model the same 100-trial hyperparameter search produces fairly tuned comparators even though the models differ by roughly 190-fold in compute; if the larger transformer models are under-tuned, the reported superiority of the lightweight LEM 3 could be an artifact of the tuning budget rather than model quality.

Editorial extensions

If this is right

  • A lightweight model conditioned on three previous events can outperform transformer-based models on standardized next-event prediction, which would make high-quality event prediction cheap enough for amateur and lower-league teams.
  • The UIED format turns multiple vendor schemas into one benchmarkable stream, so future event-prediction studies can be compared directly instead of through bespoke pipelines.
  • The SAR format lets researchers combine event and tracking data at every frame, which is a precondition for multi-agent deep reinforcement learning on real matches.
  • The RL results imply that action imitation and reward prediction are in tension under the composite loss, so practitioners must choose a point on that frontier rather than optimizing action accuracy alone.
  • Event simulation with LEM 3 degrades after roughly the tenth timestep, meaning short-horizon simulation is reliable but long-horizon match simulation still needs new machinery.

Reading between the lines

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

  • Inference: If LEM 3's advantage persists under compute-matched tuning, the field's recent move toward billion-parameter transformers for event prediction may be buying little; a three-event context window may already capture most of the information relevant to the next action, and the next gains would come from richer input such as freeze-frame positions rather than bigger decoders.
  • Inference: The UIED standardization makes transfer learning across providers testable—for example, training on a public provider and fine-tuning on a private amateur dataset created with the label tool; the paper supplies all the pieces but does not run that experiment.
  • Inference: The RL trade-off suggests an automatic schedule for $\lambda_1$ that ramps up action supervision only after the temporal-difference loss stabilizes could push past the reported frontier; the paper reports the frontier but does not try dynamic balancing.
  • Inference: The possession metrics and heatmaps could be combined into a counterfactual tool such as 'what if this pass were a shot,' which the paper's own simulation function already enables but does not demonstrate.
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

3 major / 6 minor

Summary. The paper presents OpenSTARLab, an open-source framework for soccer spatio-temporal agent data analysis that includes a video annotation tool (STE), standardized data formats (UIED and SAR), preprocessing packages, an event modeling package, and an RL package (RLearn). It benchmarks several event prediction models (MAJ, Seq2Event, NMSTPP, LEM 1, LEM 3, FMS) on Wyscout and StatsBomb data and RL models (MLP, GRU, LSTM) on J-League data. The headline empirical claims are that LEM 3 achieves the best action accuracy, F1, and time-prediction error on both datasets, and that the RL experiments reveal a trade-off between action accuracy and temporal-difference loss.

Significance. If the empirical claims hold, the paper would provide a useful open and standardized evaluation platform for soccer event prediction, with the notable finding that a small MLP-based model (LEM 3) outperforms much larger transformer-based models. The availability of open-source code, pretraining configurations, and standardized formats is a genuine strength and supports reproducibility. The RL trade-off finding, if correctly established, would be informative for using TD losses with action supervision in sports RL. However, the evaluation protocol currently does not support the strength of the headline claims, and the RL loss notation is internally inconsistent, so the significance is conditional on fixing these issues.

major comments (3)
  1. [§4.2, Table 6] The equal 100-trial Optuna budget is not a fair optimization budget across models that differ by up to roughly 190x in FLOPs (LEM 3 at 19-20M vs FMS at 930M-3.66B, Seq2Event at 112M-4.03B). A fixed number of trials gives the large transformer models far fewer effective training steps per trial, so their results in Table 6 may reflect under-training rather than architectural inferiority. Because the headline claim that LEM 3 is superior is read directly from Table 6, the authors should either give the larger models a comparable effective compute budget, report compute-normalized results, or justify why 100 trials is sufficient for all models.
  2. [Tables 6 and 7] Tables 6 and 7 report single runs without repeated-seed statistics or error bars. The margins on the key comparison are very small: on StatsBomb, LEM 3's action accuracy is 0.66 versus 0.65 for the other learned models, and F1 is 0.25 versus 0.23-0.24. On Wyscout, all five learned models report exactly 0.67 action accuracy to two decimals, which suggests either rounding or run-to-run noise. Without multiple seeds and a variance estimate, the reported advantages are not distinguishable from stochastic variation. Please report means with standard deviations or confidence intervals, and state the number of seeds used.
  3. [§4.6, Eq. (3), §4.7, Table 7] Equation (3) defines Ltotal = Ltd + λ1 LL1 + λ2 Las, so λ1 is the coefficient on the L1 regularization term. However, Section 4.7's discussion and Table 7 treat λ1 as the weight for the action-supervised loss: for example, the text says that when λ1 is small, the action loss is under-optimized, and that increasing λ1 improves action accuracy. This is inconsistent with Eq. (3) and makes the reported trade-off between action accuracy and TD loss uninterpretable. The authors must correct either the loss definition or the notation in the text and table, and then re-state the results consistently.
minor comments (6)
  1. [§3.2, Table 3; §3.3, Table 4] The table headers read "Avalible variables" and should read "Available variables."
  2. [§2.2] The bullet points contain stray spaces in the words "T erminology" and "F ragmented"; please fix these formatting artifacts.
  3. [§4.8] The text states that Figure 9 uses the model with λ1 = 0.05 for the RNN, but Table 7 only reports λ1 values of 0.01, 0.005, and 0.001. Please clarify which setting was actually used or correct the text.
  4. [§3.4] In the SAR format option bullet, "conputing" should be "computing."
  5. [§3.4.1, §3.7] Section 3.4.1 says the package supports many providers including RoboCup 2D and Sportec, but Section 3.7 states that full mapping to UIED is only supported for GRF, StatsBomb, Wyscout, and DataStadium. Please harmonize these statements so readers know which providers currently have end-to-end UIED mapping.
  6. [§4.4] The claim that the model "maintained robust event prediction performance up to timestep 10" is not directly supported by Figure 5, since ACC action drops from 0.66 to 0.5 and X/Y MAE increase over time. Please provide a quantitative threshold or a comparison to a baseline at each timestep to justify this characterization.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the central empirical claims are direct measurements of third-party and baseline models, and the RL trade-off is a data-dependent observation rather than a definitional consequence.

full rationale

Walking the derivation chain, I found no step where an output claim is equivalent to its input by construction. The headline result (LEM 3 achieving the best action accuracy, F1, and time MAE) is an empirical comparison of models from other groups under the authors' UIED standardization pipeline; the UIED format standardizes inputs but does not mathematically force any model's ranking. The RL 'trade-off between action accuracy and temporal difference loss' is presented in Section 4.6 as a design rationale ('This difference creates a trade-off') and then tested in Section 4.7 by varying hyperparameters; because the observed accuracies and TD losses depend on the optimization outcomes and on the specific data/loss landscape, the reported relationship is not a tautology. The paper's self-citations ([11], [37], [39], etc.) are used for baselines, methodological continuity, and feature provenance, but the load-bearing comparisons and the trade-off measurements do not reduce to those citations. The equal-tuning-budget and FLOPs-disparity concerns raised by the skeptic are about evaluation fairness and correctness, not circularity. Accordingly, the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 2 invented entities

The central claims rest on standardization choices (45 m threshold, event mappings, pitch rescaling), the reward and sequence-segmentation assumptions inherited from [11], and a single-run evaluation protocol. None of these are derived from first principles, and several are hand-chosen.

free parameters (3)
  • pass length threshold for short/long pass classification = 45 meters
    Section 3.2 splits passes into Short and Long at 45 m with no justification from provider conventions; this assignment changes the action labels that every event model trains on, and the benchmark rankings could shift with a different threshold.
  • attack sequence length bounds for SAR = 50 to 300 frames
    Section 3.3 keeps only sequences with 50 to 300 frames for the SAR/RL dataset, following [11]; this filters the data and affects the RL trade-off results.
  • simulation maximum timestep = 26
    Section 4.4 caps simulations at 26 events because about 98 percent of possessions are shorter; this truncation shapes the reported simulation accuracy and MAE curves.
assumptions (4)
  • domain assumption UIED event-type mapping preserves semantic equivalence across providers
    Section 3.2 maps StatsBomb Ground Pass and Wyscout Simple Pass to the same UIED Short Pass; if providers label semantics differently, models trained on different providers are not solving the same task.
  • domain assumption Reward function from Nakahara et al. (2023) is a valid proxy for action value
    Section 4.6 sets reward to 0 except for +1 for a goal, -1 for a conceded goal, or EPV at the sequence end; the paper itself says the trade-off may stem from the reward function's incomplete representation of player behavior.
  • domain assumption Single-run evaluation with fixed train/validation/test splits is sufficient for the performance claims
    Sections 4.2 and 4.6 describe single training runs with no repeated-seed variance, and Tables 6 and 7 report no confidence intervals.
  • domain assumption Pitch coordinate standardization to 105 by 68 meters is valid for all providers
    Section 3.2 assumes FIFA-recommended dimensions and rescales all providers accordingly; tracking data from nonstandard pitches would be distorted.
invented entities (2)
  • UIED (Unified and Integrated Event Data) format independent evidence
    purpose: Standardized event-data schema mapping GRF, StatsBomb, Wyscout, and DataStadium into one format for model training.
    A new data standard introduced by this paper; implemented in the public Pre-processing package, so third parties can apply it and falsify the claimed mappings.
  • SAR (State-Action-Reward) format independent evidence
    purpose: Standardized per-frame state, action, and reward schema for deep RL on event plus tracking data.
    A new format introduced in Section 3.3; it is implemented in public code and can be applied to any compatible dataset.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OpenSTARLab: Open Approach for Spatio-Temporal Agent Data Analysis in Soccer." pith.science (2026). https://pith.science/paper/G5GLMU27

@misc{pith2026250202785,
  author       = {Pith},
  title        = {Pith review of: OpenSTARLab: Open Approach for Spatio-Temporal Agent Data Analysis in Soccer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G5GLMU27}},
  note         = {Machine review of arXiv:2502.02785}
}
read the original abstract

Sports analytics has become both more professional and sophisticated, driven by the growing availability of detailed performance data. This progress enables applications such as match outcome prediction, player scouting, and tactical analysis. In soccer, the effective utilization of event and tracking data is fundamental for capturing and analyzing the dynamics of the game. However, there are two primary challenges: the limited availability of event data, primarily restricted to top-tier teams and leagues, and the scarcity and high cost of tracking data, which complicates its integration with event data for comprehensive analysis. Here we propose OpenSTARLab, an open-source framework designed to democratize spatio-temporal agent data analysis in sports by addressing these key challenges. OpenSTARLab includes the Pre-processing Package that standardizes event and tracking data through Unified and Integrated Event Data and State-Action-Reward formats, the Event Modeling Package that implements deep learning-based event prediction, alongside the RLearn Package for reinforcement learning tasks. These technical components facilitate the handling of diverse data sources and support advanced analytical tasks, thereby enhancing the overall functionality and usability of the framework. To assess OpenSTARLab's effectiveness, we conducted several experimental evaluations. These demonstrate the superior performance of the specific event prediction model in terms of action and time prediction accuracies and maintained its robust event simulation performance. Furthermore, reinforcement learning experiments reveal a trade-off between action accuracy and temporal difference loss and show comprehensive visualization. Overall, OpenSTARLab serves as a robust platform for researchers and practitioners, enhancing innovation and collaboration in the field of soccer data analytics.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 70 canonical work pages

  1. [1]

    & Dubitzky, W

    Berrar, D., Lopes, P. & Dubitzky, W. A data-and knowledge-driven framework for developing machine learning models to predict soccer match outcomes. Machine Learning 113, 8165–8204 (2024)

    work page 2024

  2. [2]

    & Fujii, K

    Yeung, C., Bunker, R., Umemoto, R. & Fujii, K. Evaluating soccer match prediction models: a deep learning approach and feature optimization for gradient-boosted trees. Machine Learning 113, 7541–7564 (2024)

    work page 2024

  3. [3]

    C., Bunker, R

    Yeung, C. C., Bunker, R. & Fujii, K. A framework of interpretable match results prediction in football with fifa ratings and team formation. Plos one 18, e0284318 (2023)

    work page 2023

  4. [4]

    & Davis, J

    Decroos, T., Bransen, L., Van Haaren, J. & Davis, J. Actions speak louder than goals: Valuing player actions in soccer. Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining 1851–1861 (2019)

    work page 2019

  5. [5]

    Beyond expected goals

    Spearman, W. Beyond expected goals. Proceedings of the 12th MIT sloan sports analytics conference 1–17 (2018)

    work page 2018

  6. [7]

    J., Locke, D

    Simpson, I., Beal, R. J., Locke, D. & Norman, T. J. Seq2event: learning the language of soccer using transformer-based match event prediction. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining 3898–3908 (2022)

    work page 2022

  7. [8]

    A Foundation Model for Soccer

    Baron, E., Hocevar, D. & Salehe, Z. A foundation model for soccer. arXiv preprint arXiv:2407.14558 (2024)

    work page Pith review arXiv 2024

  8. [9]

    & Bunker, R

    Yeung, C. & Bunker, R. An events and 360 data-driven approach for extract- ing team tactics and evaluating performance in football. Statsbomb Conference Proceedings (2023)

    work page 2023

Show all 73 references
  1. [10]

    & Kharrat, T

    Liu, G., Luo, Y., Schulte, O. & Kharrat, T. Deep soccer analytics: learning an action-value function for evaluating soccer players. Data Mining and Knowledge Discovery 34, 1531–1559 (2020)

    2020

  2. [11]

    & Fujii, K

    Nakahara, H., Tsutsui, K., Takeda, K. & Fujii, K. Action valuation of on-and off-ball soccer players based on multi-agent deep reinforcement learning. IEEE Access 11, 131237–131244 (2023)

    2023

  3. [12]

    & Ghanem, B

    Giancola, S., Amine, M., Dghaily, T. & Ghanem, B. Soccernet: A scalable dataset for action spotting in soccer videos. Proceedings of the IEEE Conference on 32 Computer Vision and Pattern Recognition Workshops 1711–1721 (2018)

    2018

  4. [13]

    Deliege, A. et al. Soccernet-v2: A dataset and benchmarks for holistic under- standing of broadcast soccer videos. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition 4508–4519 (2021)

    2021

  5. [14]

    Cioppa, A. et al. Soccernet 2023 challenges results. Sports Engineering 27, 24 (2024)

    2024

  6. [15]

    Cui, Y. et al. SportsMOT: A large multi-object tracking dataset in multiple sports scenes. Proceedings of the IEEE/CVF International Conference on Computer Vision 9921–9931 (2023)

    2023

  7. [16]

    Scott, A. et al. Soccertrack: A dataset and tracking algorithm for soccer with fish- eye and drone videos. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 3569–3579 (2022)

    2022

  8. [17]

    Scott, A. et al. Teamtrack: A dataset for multi-sport multi-object tracking in full-pitch videos. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 3357–3366 (2024)

    2024

  9. [18]

    & Wang, L

    Wu, T., He, R., Wu, G. & Wang, L. Sportshhi: A dataset for human-human interaction detection in sports videos. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 18537–18546 (2024)

    2024

  10. [19]

    & Fujii, K

    Yeung, C., Ide, K. & Fujii, K. Autosoccerpose: Automated 3d posture analysis of soccer shot movements. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshop 3214–3224 (2024)

    2024

  11. [20]

    & Davis, J

    Van Roy, M., Cascioli, L. & Davis, J. Etsy: A rule-based approach to event and tracking data synchronization. International Workshop on Machine Learning and Data Mining for Sports Analytics 11–23 (2023)

    2023

  12. [21]

    & Serra, G

    Ballan, L., Bertini, M., Del Bimbo, A. & Serra, G. Semantic annotation of soccer videos by visual instance clustering and spatial/temporal reasoning in ontologies. Multimedia Tools and Applications 48, 313–337 (2010)

    2010

  13. [22]

    Stein, M. et al. Director’s cut: Analysis and annotation of soccer matches. IEEE computer graphics and applications 36, 50–60 (2016)

    2016

  14. [23]

    & Alpaslan, F

    Alan, O., Akpinar, S., Sabuncu, O., Cicekli, N. & Alpaslan, F. Ontological video annotation and querying system for soccer games. 2008 23rd International Symposium on Computer and Information Sciences 1–6 (2008)

    2008

  15. [24]

    & Eftekhari-Moghadam, A.-M

    Hosseini, M.-S. & Eftekhari-Moghadam, A.-M. Fuzzy rule-based reasoning approach for event detection and annotation of broadcast soccer video. Applied Soft Computing 13, 846–866 (2013). 33

    2013

  16. [25]

    & Nunziati, W

    Assfalg, J., Bertini, M., Colombo, C., Del Bimbo, A. & Nunziati, W. Semantic annotation of soccer videos: automatic highlights identification. Computer vision and image understanding 92, 285–305 (2003)

    2003

  17. [26]

    Wang, Z., Yu, J. & He, Y. Soccer video event annotation by synchronization of attack–defense clips and match reports with coarse-grained time information. IEEE Transactions on Circuits and Systems for Video Technology 27, 1104–1117 (2016)

    2016

  18. [27]

    & Bhanu, B

    Theagarajan, R. & Bhanu, B. An automated system for generating tactical per- formance statistics for individual soccer players from videos. IEEE Transactions on Circuits and Systems for Video Technology 31, 632–646 (2020)

    2020

  19. [28]

    Cioppa, A. et al. A context-aware loss function for action spotting in soc- cer videos. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 13126–13136 (2020)

    2020

  20. [29]

    Barra, S. et al. Footapp: An ai-powered system for football match annotation. Multimedia Tools and Applications 82, 5547–5567 (2023)

    2023

  21. [30]

    & Liu, W

    Ma, S., Shao, E., Xie, X. & Liu, W. Event detection in soccer video based on self-attention. 2020 IEEE 6th International Conference on Computer and Communications (ICCC) 1852–1856 (2020)

    2020

  22. [31]

    & Pappalardo, L

    Sorano, D., Carrara, F., Cintia, P., Falchi, F. & Pappalardo, L. Automatic pass annotation from soccer video streams based on object detection and lstm. Machine Learning and Knowledge Discovery in Databases. Applied Data Science and Demo Track: European Conference, ECML PKDD 2...

    2021

  23. [32]

    Cioppa, A. et al. Soccernet 2024 challenges results. arXiv preprint arXiv:2409.10587 (2024)

    2024 arXiv

  24. [33]

    & Lamberti, F

    Morra, L., Manigrasso, F. & Lamberti, F. Soccer: Computer graphics meets sports analytics for soccer event recognition. SoftwareX 12, 100612 (2020)

    2020

  25. [34]

    & Davis, J

    Decroos, T., Bransen, L., Van Haaren, J. & Davis, J. Vaep: An objective approach to valuing on-the-ball actions in soccer. IJCAI 4696–4700 (2020)

    2020

  26. [35]

    & Fujii, K

    Yeung, C., Bunker, R. & Fujii, K. Unveiling multi-agent strategies: A data- driven approach for extracting and evaluating team tactics from football event and freeze-frame data. Journal of Robotics and Mechatronics 36, 603–617 (2024)

    2024

  27. [36]

    Kurach, K. et al. Google research football: A novel reinforcement learning environ- ment. Proceedings of the AAAI conference on artificial intelligence 34, 4501–4510 (2020). 34

    2020

  28. [37]

    & Fujii, K

    Yeung, C., Sit, T. & Fujii, K. Transformer-based neural marked spatio tem- poral point process model for football match events analysis. arXiv preprint arXiv:2302.09276 (2023)

    2023 arXiv

  29. [38]

    Pappalardo, L. et al. A public data set of spatio-temporal match events in soccer competitions. Scientific data 6, 1–15 (2019)

    2019

  30. [39]

    & Mendes-Moreira, J

    Mendes-Neves, T., Meireles, L. & Mendes-Moreira, J. Towards a foundation large events model for soccer. Machine Learning 1–23 (2024)

    2024

  31. [40]

    & Mendes-Moreira, J

    Mendes-Neves, T., Meireles, L. & Mendes-Moreira, J. Forecasting events in soccer matches through language. arXiv preprint arXiv:2402.06820 (2024)

    2024 arXiv

  32. [41]

    & Pechenizkiy, M

    Eggels, H., van Elk, R. & Pechenizkiy, M. Expected goals in soccer: Explaining match results using predictive analytics. The machine learning and data mining for sports analytics workshop 16 (2016)

    2016

  33. [42]

    & Fujii, K

    Toda, K., Teranishi, M., Kushiro, K. & Fujii, K. Evaluation of soccer team defense based on prediction models of ball recovery and being attacked: A pilot study. Plos one 17, e0263051 (2022)

    2022

  34. [43]

    & Fujii, K

    Umemoto, R., Tsutsui, K. & Fujii, K. Location analysis of players in uefa euro 2020 and 2022 using generalized valuation of defense by estimating probabilities. arXiv preprint arXiv:2212.00021 (2022)

    2022 arXiv

  35. [44]

    Attention is all you need

    Vaswani, A. Attention is all you need. Advances in Neural Information Processing Systems (2017)

    2017

  36. [45]

    C., Januschowski, T

    Shchur, O., Turkmen, A. C., Januschowski, T. & G¨ unnemann, S. Neural temporal point processes: A review. IJCAI 2021 (2021)

    2021

  37. [46]

    Fujii, K. et al. Adaptive action supervision in reinforcement learning from real- world multi-agent demonstrations. 16th International Conference on Agents and Artificial Intelligence (ICAART’ 24) 2, 27–39 (2024)

    2024

  38. [47]

    Data-driven analysis for understanding team sports behaviors

    Fujii, K. Data-driven analysis for understanding team sports behaviors. Journal of Robotics and Mechatronics 33, 505–514 (2021)

    2021

  39. [48]

    Machine Learning in Sports: Open Approach for Next Play Analytics (Springer, 2025)

    Fujii, K. Machine Learning in Sports: Open Approach for Next Play Analytics (Springer, 2025)

    2025

  40. [49]

    & Davis, J

    Van Roy, M., Robberechts, P., Yang, W.-C., De Raedt, L. & Davis, J. Leaving goals on the pitch: Evaluating decision making in soccer. Proceedings of the MIT Sloan Sports Analytics Conference (2021)

    2021

  41. [50]

    & Davis, J

    Van Roy, M., Robberechts, P., Yang, W.-C., De Raedt, L. & Davis, J. A markov framework for learning and reasoning about strategies in professional soccer. Journal of Artificial Intelligence Research 77, 517–562 (2023). 35

    2023

  42. [51]

    & Toka, L

    Rahimian, P., Oroojlooy, A. & Toka, L. Towards optimized actions in critical situations of soccer games with deep reinforcement learning. 2021 IEEE 8th International Conference on Data Science and Advanced Analytics (DSAA) 1–12 (2021)

    2021

  43. [52]

    & Toka, T

    Rahimian, P., Van Haaren, J. & Toka, T. A.-L. Beyond action valuation: A deep reinforcement learning framework for optimizing player decisions in soccer. 16th Annual MIT Sloan Sports Analytics Conference (2022)

    2022

  44. [53]

    & Toka, L

    Rahimian, P., Van Haaren, J. & Toka, L. Towards maximizing expected possession outcome in soccer. International Journal of Sports Science & Coaching 19, 230– 244 (2024)

    2024

  45. [54]

    & Schulte, O

    Liu, G. & Schulte, O. Deep reinforcement learning in ice hockey for context- aware player evaluation. Proceedings of the 27th International Joint Conference on Artificial Intelligence 3442–3448 (2018)

    2018

  46. [55]

    Schulte, O. et al. A markov game model for valuing actions, locations, and team performance in ice hockey. Data Mining and Knowledge Discovery 31, 1735–1757 (2017)

    2017

  47. [56]

    & Fujii, K

    Ding, N., Takeda, K. & Fujii, K. Deep reinforcement learning in a racket sport for player evaluation with technical and tactical contexts. IEEE Access 10, 54764– 54772 (2022)

    2022

  48. [57]

    & Rokach, L

    Yanai, C., Solomon, A., Katz, G., Shapira, B. & Rokach, L. Q-ball: Modeling basketball games using deep reinforcement learning. Proceedings of the AAAI Conference on Artificial Intelligence 36, 8806–8813 (2022)

    2022

  49. [58]

    Chen, X. et al. Reliable: Offline reinforcement learning for tactical strategies in professional basketball games. Proceedings of the 31st ACM International Conference on Information & Knowledge Management 3023–3032 (2022)

    2022

  50. [59]

    & Poupart, P

    Luo, Y., Schulte, O. & Poupart, P. Inverse reinforcement learning for team sports: Valuing actions and players. Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, IJCAI-20 3356–3363 (2020)

    2020

  51. [60]

    & Toka, L

    Rahimian, P. & Toka, L. Inferring the strategy of offensive and defensive play in soccer with inverse reinforcement learning. Machine Learning and Data Mining for Sports Analytics 26–38 (2022)

    2022

  52. [61]

    M., Yue, Y., Carr, P

    Le, H. M., Yue, Y., Carr, P. & Lucey, P. Coordinated multi-agent imitation learning. Proceedings of the 34th International Conference on Machine Learning- Volume 70 1995–2003 (2017)

    2017

  53. [62]

    M., Carr, P., Yue, Y

    Le, H. M., Carr, P., Yue, Y. & Lucey, P. Data-driven ghosting using deep imitation learning. Proceedings of MIT Sloan Sports Analytics Conference (2017). 36

    2017

  54. [63]

    & Takeda, K

    Teranishi, M., Fujii, K. & Takeda, K. Trajectory prediction with imitation learning reflecting defensive evaluation in team sports. 2020 IEEE 9th Global Conference on Consumer Electronics (GCCE) 124–125 (2020)

    2020

  55. [64]

    & Takeda, K

    Fujii, K., Takeishi, N., Kawahara, Y. & Takeda, K. Decentralized policy learning with partial observation and mechanical constraints for multiperson modeling. Neural Networks 171, 40–52 (2024)

    2024

  56. [65]

    & Lucey, P

    Zhan, E., Zheng, S., Yue, Y., Sha, L. & Lucey, P. Generating multi-agent trajecto- ries using programmatic weak supervision. International Conference on Learning Representations (2019)

    2019

  57. [66]

    A., Schwing, A

    Yeh, R. A., Schwing, A. G., Huang, J. & Murphy, K. Diverse generation for multi- agent sports games. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 4610–4619 (2019)

    2019

  58. [67]

    Li, L. et al. GRIN: Generative relation and intention network for multi-agent trajectory prediction. Advances in Neural Information Processing Systems 34, 27107–27118 (2021)

    2021

  59. [68]

    Fujii, K. et al. Estimating counterfactual treatment outcomes over time in com- plex multi-agent scenarios. IEEE Transactions on Neural Networks and Learning Systems 1–15 (2024)

    2024

  60. [69]

    & Koyama, M

    Akiba, T., Sano, S., Yanase, T., Ohta, T. & Koyama, M. Optuna: A next- generation hyperparameter optimization framework. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (2019)

    2019

  61. [70]

    & Little, J

    Chen, J. & Little, J. J. Sports camera calibration via synthetic data. Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition workshops (2019)

    2019

  62. [71]

    Sha, L. et al. End-to-end camera calibration for broadcast videos. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition 13627– 13636 (2020)

    2020

  63. [72]

    Chu, Y.-J. et al. Sports field registration via keypoints-aware label condition. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 3523–3530 (2022)

    2022

  64. [73]

    & Ewerth, R

    Theiner, J. & Ewerth, R. Tvcalib: Camera calibration for sports field registration in soccer. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision 1166–1175 (2023)

    2023

  65. [74]

    & Agudo, A

    Guti´ errez-P´ erez, M. & Agudo, A. No bells just whistles: Sports field registration by leveraging geometric properties. Proceedings of the IEEE/CVF Conference on 37 Computer Vision and Pattern Recognition 3325–3334 (2024). 38

    2024

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.