Pith. sign in

REVIEW 4 major objections 4 minor 9 references

Multi-Strategy Improved Snake Optimizer Accelerated CNN-LSTM-Attention-Adaboost for Trajectory Prediction

T0 review · 4 major / 4 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 Tuned snake optimizer predicts flight trajectories with 0.9978 R2

desk verdict A competent assembly of known optimizer tricks and hybrid networks, but the evaluation uses the tuning set as the test set and the headline 39.89% 'accuracy' is actually a training loss reduction—so the central claims do not hold. read the letter →

arxiv 2507.15832 v1 pith:5LZURIUK submitted 2025-07-21 cs.LG

Shiyang Li This is my paper · ORCID
classification cs.LG
keywords four-dimensionaltrajectorypredictionsnakeoptimizerCNN-LSTM-attentionAdaBoostensemblehyperparameteroptimizationADS-Bdatametaheuristictuning
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 hybrid predictor — CNN for spatial features, LSTM for temporal dynamics, attention for global context, and AdaBoost to combine weak learners — becomes more accurate when its hyperparameters are tuned by a multi-strategy improved Snake Optimizer (SO). On a real ADS-B trajectory dataset from Tianjin to Xi'an, the tuned model reaches RMSE 125.43, MAPE 1.35%, and $R^2=0.9978$, beating eight other optimizer-tuned variants including particle swarm, whale, and gray wolf optimizers. The paper also reports an ablation study attributing a 39.89% drop in training loss to the addition of the improved SO module. If these results hold, metaheuristic hyperparameter tuning would be a strong lever for spatio-temporal trajectory prediction.

What carries the argument

The load-bearing mechanism is the multi-strategy improved Snake Optimizer, which keeps the original SO's male/female sub-population exploration–exploitation phases and adds four enhancements: good-point-set initialization built on number-theoretic uniform distribution; periodic cosine/sine adaptive updates for the C1, C3, food-threshold, and temperature-threshold parameters; a dual mutation scheme (Cauchy then Gaussian as primary, plus head-chaotic, body-fusion, and tail-splicing mutations as auxiliary); and an adaptive flight function that shifts from Levy flight to random walk as iterations proceed. These mechanisms govern how the optimizer searches for the three hyperparameters — batch size, learning rate, and neuron count — that configure the CNN-LSTM-attention-Adaboost ensemble. The improved SO is what the paper credits for the reported accuracy gains.

What would settle it

Run the identical pipeline with a strictly held-out test set — for example, a different flight or a different day on the same Tianjin–Xi'an route — and recompute RMSE and MAPE. If SO-CLA-adaboost's advantage over the best competing optimizer narrows below the reported 19–57% margins, or if the 39.89% ablation gain disappears, the central claim fails.

Watch

Extended reading notes

Core claim

The central claim is that SO-CLA-adaboost, the proposed combination, outperforms all eight compared optimizer-tuned variants on every reported metric: RMSE 125.4342 versus the next-best 155.7012 (WOA), MAPE 1.3504% versus 1.9262%, MAE 51.9559 versus 64.0093, MAXAE 609.4213 versus 1274.5740, and $R^2=0.9978$ versus 0.9962. The paper further claims that the full multi-strategy improvement contributes a 39.89% loss reduction over the non-optimized CLA-adaboost, and that the improved SO stabilizes faster (16 iterations) than the comparison optimizers. On the CEC2022 benchmark suite, the improved SO ranks first on ten of the twelve tested functions.

Load-bearing premise

The reported accuracy is measured on the same validation set that the optimizer used to choose hyperparameters, so there is no independent test set confirming the numbers generalize.

Editorial extensions

If this is right

  • If the reported numbers hold, metaheuristic hyperparameter tuning of hybrid spatio-temporal networks becomes a practical route to high-accuracy 4D trajectory prediction.
  • The 39.89% loss reduction in the ablation implies that the multi-strategy improvement to SO, rather than the ensemble architecture alone, is the decisive factor in the reported accuracy.
  • The convergence results imply SO needs fewer iterations (16) than PSO (27), WOA (38), and GWO (43), which matters for near-real-time hyperparameter tuning.
  • The MAXAE figures indicate that even the best model still lags at rapid altitude changes, so the reported gains apply mainly to normal flight phases rather than abrupt maneuvers.

Reading between the lines

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

  • A consequence the paper leaves implicit: because the evaluation uses the validation set for both tuning and reporting, the true generalization gap is unknown; a held-out test across different flights, dates, or weather conditions would be the natural next check.
  • The improved SO's components (good-point-set initialization, adaptive thresholds, Cauchy–Gaussian and head/body/tail mutations, Levy-and-random-walk flight) are generic and could transfer to other hybrid architectures, though the paper only demonstrates them on this one dataset and task.
  • The MAXAE spikes at altitude changes suggest that the model class itself, not just the optimizer, may need a mechanism for detecting regime shifts such as climb/descent transitions to close the remaining error.
  • A simple experiment could isolate whether the 39.89% ablation gain comes from the optimizer's search or from the particular hyperparameters it lands on, by comparing against random search with the same budget.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The manuscript proposes a multi-strategy improved Snake Optimizer (SO) for tuning the hyperparameters of a CNN-LSTM-attention-Adaboost hybrid network for four-dimensional aircraft trajectory prediction from ADS-B data. The authors describe several SO modifications (good-point-set initialization, adaptive thresholds, dual mutation, and Levy/random-walk flight), benchmark the improved optimizer against eight other metaheuristics on CEC2022 functions, and then compare SO-CLA-adaboost against variants tuned by the other optimizers on a Tianjin-Xi'an ADS-B dataset. The headline results are RMSE 125.4342, MAPE 1.3504%, MAE 51.9559, MAXAE 609.4213, and R² 0.9978, with an advertised 39.89% improvement attributable to the SO module.

Significance. If the reported results had been obtained on a properly held-out test set, the paper would be a useful demonstration of metaheuristic hyperparameter tuning for a hybrid spatio-temporal network in trajectory prediction. The CEC2022 benchmark comparison, including Wilcoxon rank-sum tests, and the ablation study are constructive components. However, the central quantitative claims are undermined by an evaluation protocol that uses the validation set for both hyperparameter optimization and final performance reporting, by a headline metric that is a loss reduction mislabeled as accuracy, and by the absence of any repeated-run or held-out evaluation. These issues make the claimed performance gains unsubstantiated and not yet credible.

major comments (4)
  1. [§4.1–4.2, Tables 6 and 8] Section 4.1 states that the model output is the validation set's true values, and Section 4.2 tunes batch size, learning rate, and LSTM neuron count on that same split. Tables 6 and 8 then report RMSE, MAPE, MAE, MAXAE, and R² on this set. Because the hyperparameters are chosen by minimizing the loss on the very set used for final evaluation, the reported errors and the ranking among optimizers reflect tuning-set fit rather than generalization. No independent test split or cross-validation is described. This is a load-bearing flaw for the abstract's headline RMSE, MAPE, and R² values and for the claimed margin over the eight other optimizers; the reported numbers cannot be distinguished from tuning-set artifacts.
  2. [Abstract and Table 7] The abstract's claim that 'introducing the full-strategy collaborative improvement SO algorithm improves the model's prediction accuracy by 39.89%' is not supported by the metric cited. According to Table 7, the 39.89% figure is the relative decrease in training loss from CNN-LSTM-attention-adaboost (0.000905) to SO-CNN-LSTM-attention-adaboost (0.000544), not an improvement in prediction accuracy. In addition, Table 8 mixes formats for MAPE: the LSTM, CNN-LSTM, CNN-LSTM-attention, and CLA-adaboost rows report 0.0536, 0.0478, 0.0451, and 0.0344, while the SO row reports 1.3504%, so the apparent improvement in MAPE across ablation stages is computed on incompatible scales.
  3. [Tables 5, 6, and 8] All trajectory-prediction results are single runs. The paper reports no standard deviations, confidence intervals, or significance tests for the RMSE/MAPE comparisons across optimizers or across ablation stages. Given the stochasticity of neural-network training and of metaheuristic optimization, the observed differences (for example, SO RMSE of 125.4 versus WOA RMSE of 155.7) may fall within run-to-run variability. Without multiple independent runs, the claim that SO statistically outperforms the other optimizers is not established.
  4. [§2.2–2.3 and §4.2] The paper never compares the proposed multi-strategy improved SO against the original Snake Optimizer. Section 2.2 introduces good-point-set initialization, adaptive thresholds, dual mutation, and Levy/random-walk flight, but Table 2 in Section 2.3 only compares 'SO' with WOA, GWO, BSO, PSO, DE, GA, BBO, and GCO; Section 4.2 compares SO-CLA-adaboost only with other optimizer-tuned CLA-adaboost variants. Consequently, the individual and joint contributions of the proposed improvements to SO cannot be isolated. Furthermore, Section 2.3.1 states that 12 CEC2022 functions were used, while Table 1 and Table 2 list only 10 functions (F1–F10), and the Table 2 caption contains an unexplained 'PLORCS' label.
minor comments (4)
  1. [§4.2] The text contains placeholder words '表几' and '图几' which should be replaced with the actual table and figure numbers.
  2. [Eq. (32)] The logistic chaotic map X(1−X) is applied to an arbitrary position vector without specifying how variables outside the [0,1] interval are handled; the authors should either restrict the mutation to normalized coordinates or justify the operation.
  3. [Throughout] There are spelling and caption inconsistencies, such as 'Atention' for 'Attention' and missing spaces in figure captions (e.g., 'Figure4 CNN-LSTMattentionadaboosthybridmodelframework').
  4. [References] Reference [28] is cited as a CNN background reference, but the cited article is on vehicle trajectory extraction with convolutional neural networks; the connection should be clarified or a more appropriate CNN reference should be used.

Circularity Check

2 steps flagged · score 6.0 of 10

The reported 39.89% improvement is a training-loss decrease, and the RMSE/MAPE/R2 claims are evaluated on the same validation set used for hyperparameter tuning, with no held-out test described.

  1. fitted input called prediction [Section 4.1 and Section 4.2; Tables 5, 6, and 8]
    "输入数据为已有数据划分的训练集的高度、经度、纬度真实值，输出数据为已有数据划分的验证集的高度、经度、纬度预测值。 / 批大小、学习率和神经元数量三个超参数的寻优范围分别设置为[16，128]、[0.0001，0.02]、[50，200]。"

    The only evaluation set named in the paper is the validation set in Section 4.1, and Section 4.2 uses SO to minimize the model's loss while searching batch size, learning rate, and neuron count; Table 5 reports the minimized loss and selected hyperparameters. Tables 6 and 8 then report RMSE 125.4342, MAPE 1.3504%, MAE 51.9559, MAXAE 609.4213, and R2 0.9978 on that same experimental split, with no held-out test or cross-validation described. The final error numbers are therefore the fitting procedure's own objective evaluated on the selection set, and the margin over eight optimizer-tuned variants is a comparison of tuning-set fits rather than an independent generalization result.

  2. self definitional [Abstract and Section 4.4 / Table 7]
    "此外，全策略协同改进SO算法的引入，使模型的预测精度提高了39.89%。 / 加入对应模块后loss 值的下降幅度分别为7.14%、16.01%、24.71%、39.89%"

    The abstract's 39.89% 'prediction accuracy improvement' is, by the paper's own Table 7, the relative decrease in loss from 0.000905 (CNN-LSTM-attention-adaboost) to 0.000544 (SO-CNN-LSTM-attention-adaboost). That loss is the objective the SO-tuned model minimizes, so the headline improvement is an arithmetic restatement of a fitted training quantity, not an independently measured accuracy gain. The claimed 39.89% improvement thus reduces by definition to the loss decrease it is computed from.

full rationale

The SO algorithm itself is benchmarked on CEC2022 functions with a Wilcoxon test, so that component is self-contained and not circular; no load-bearing self-citation chain was found. However, the central trajectory-prediction claims exhibit specific reductions that can be quoted and traced to the paper's own tables: the 39.89% improvement in the abstract is explicitly the training-loss decrease reported in Table 7, and the final RMSE/MAPE/MAE/MAXAE/R2 numbers are computed on the validation set named in Section 4.1 after SO has minimized the model's loss during hyperparameter search, with no held-out test described. These are not merely suspicions of overfitting; they are cases where the reported 'prediction' is equivalent to the fitted quantity by construction. The independent optimizer benchmark keeps the circularity partial rather than total, so the score is 6 rather than higher.

Assumptions & free parameters 8 free parameters · 4 assumptions · 0 invented entities

The central claim depends on three tuned hyperparameters and on several hand-chosen constants in the improved optimizer. The most fragile assumptions are that the validation set is acceptable as a test set and that benchmark function results transfer to the tuning problem. The logistic mutation introduces an additional unjustified assumption.

free parameters (8)
  • batch_size = 48
    Tuned by SO on the validation set (Table 5); directly affects final prediction metrics.
  • learning_rate = 0.0076
    Tuned by SO on the validation set (Table 5).
  • lstm_neurons = 71
    Tuned by SO on the validation set (Table 5).
  • cauchy_scale_gamma = 0.05
    Hand-chosen scale in Eq. (30); affects mutation step size in the optimizer.
  • gaussian_std_sigma = 0.1
    Hand-chosen std in Eq. (31); controls fine-grained mutation.
  • levy_exponent_beta = unspecified
    Appears in Eq. (35); no value is given in the paper.
  • chaos_factor_alpha = unspecified
    Appears in Eq. (32); no value is given.
  • exploration_coefficient_C2 = 0.05
    Constant in Eqs. (6) and (8); hand-chosen.
assumptions (4)
  • domain assumption The validation set used for hyperparameter tuning can also serve as the test set for unbiased performance evaluation.
    Sections 4.1 and 4.2 use the same validation data for tuning and for reporting final errors; this is likely invalid.
  • domain assumption CEC2022 benchmark performance transfers to neural network hyperparameter optimization.
    The paper assumes SO's ranking on synthetic functions predicts its performance on model tuning, without evidence.
  • ad hoc to paper The logistic chaotic map can be applied directly to arbitrary position vectors.
    Eq. (32) computes X*(1-X) for positions that may be outside [0,1], which is not mathematically justified.
  • standard math Standard CNN, LSTM, attention, and Adaboost formulations are correct as stated.
    The model relies on textbook versions of these components.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Strategy Improved Snake Optimizer Accelerated CNN-LSTM-Attention-Adaboost for Trajectory Prediction." pith.science (2026). https://pith.science/paper/5LZURIUK

@misc{pith2026250715832,
  author       = {Pith},
  title        = {Pith review of: Multi-Strategy Improved Snake Optimizer Accelerated CNN-LSTM-Attention-Adaboost for Trajectory Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5LZURIUK}},
  note         = {Machine review of arXiv:2507.15832}
}
read the original abstract

To address the limitations of medium- and long-term four-dimensional (4D) trajectory prediction models, this paper proposes a hybrid CNN-LSTM-attention-adaboost neural network model incorporating a multi-strategy improved snake-herd optimization (SO) algorithm. The model applies the Adaboost algorithm to divide multiple weak learners, and each submodel utilizes CNN to extract spatial features, LSTM to capture temporal features, and attention mechanism to capture global features comprehensively. The strong learner model, combined with multiple sub-models, then optimizes the hyperparameters of the prediction model through the natural selection behavior pattern simulated by SO. In this study, based on the real ADS-B data from Xi'an to Tianjin, the comparison experiments and ablation studies of multiple optimizers are carried out, and a comprehensive test and evaluation analysis is carried out. The results show that SO-CLA-adaboost outperforms traditional optimizers such as particle swarm, whale, and gray wolf in handling large-scale high-dimensional trajectory data. In addition, introducing the full-strategy collaborative improvement SO algorithm improves the model's prediction accuracy by 39.89%.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 8 canonical work pages

  1. [1]

    MinistryofTransportationandCommunications,People'sRepublicofChina.Civil AviationofChina:StatisticsonMajorProductionIndicatorsforJanuary2024

    work page

  2. [2]

    Ruwantissa,A.AirNavigationLaw.Berlin:Springer,2012

    work page 2012

  3. [3]

    BenavidesJV,KaneshigeJ,SharmaS,etal.Implementationofatrajectoryprediction functionfortrajectorybasedoperations[C]//AIAAAtmosphericFlightMechanics Conference.2014:2198

    work page 2014

  4. [4]

    LinY,ZhangJ,LiuH.Analgorithmfortrajectorypredictionofflightplanbasedon relativemotionbetweenpositions[J].FrontiersofInformationTechnology&Electronic Engineering,2018,19(7):905-916

    work page 2018

  5. [5]

    AvanziniG.Frenet-basedalgorithmfortrajectoryprediction[J].Journalofguidance, control,anddynamics,2004,27(1):127-135

    work page 2004

  6. [6]

    Computer Applications2021, 41(S1):pp.277-282

    Lv,B.,Wang,C.ApplicationofimprovedextendedKalmanfilterinaircraft4D trajectorypredictionalgorithm. Computer Applications2021, 41(S1):pp.277-282

    work page

  7. [7]

    YepesJL,HwangI,RoteaM.Newalgorithmsforaircraftintentinferenceandtrajectory prediction[J].Journalofguidance,control,anddynamics,2007,30(2):370-382

    work page 2007

  8. [8]

    **L,JunZ,YanboZ,etal.Simulationstudyofalgorithmsforaircrafttrajectory predictionbasedonADS-Btechnology[C]//2008AsiaSimulationConference-7th InternationalConferenceonSystemSimulationandScientificComputing.IEEE,2008: 322-327

    work page 2008

Show all 9 references
  1. [9]

    [10]ShiZ,XuM,PanQ.4-DflighttrajectorypredictionwithconstrainedLSTMnetwork[J]

    HwangI,BalakrishnanH,TomlinC.Stateestimationforhybridsystems:applicationsto aircrafttracking[J].IEEProceedings-ControlTheoryandApplications,2006,153(5): 556-566. [10]ShiZ,XuM,PanQ.4-DflighttrajectorypredictionwithconstrainedLSTMnetwork[J]. IEEEtransactionsonintelligenttranspor...

    2006 arXiv

Pith tools

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