Pith. sign in

REVIEW 4 major objections 6 minor 51 references

CR-BLEA: Contrastive Ranking for Adaptive Resource Allocation in Bilevel Evolutionary Algorithms

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read CR-BLEA claims that a contrastive ranking network trained online on pairwise comparisons can identify which lower-level tasks deserve computation, allowing any nested bilevel evolutionary algorithm to skip unpromising lower-level…

desk verdict CR-BLEA is a useful empirical wrapper for cutting lower-level FE waste in bilevel EAs, but the abstract overclaims: on problems with multiple lower-level optima (SMD6/SMD12) the savings collapse and accuracy can degrade. read the letter →

arxiv 2506.06362 v1 pith:4QOKOUDN submitted 2025-06-03 cs.NE cs.AIcs.LG

Dejun Xu , Jijia Chen , Gary G. Yen , Min Jiang This is my paper
classification cs.NEcs.AIcs.LG MSC 90C59
keywords bileveloptimizationevolutionarycomputationcontrastiverankingmodelresourceallocationlower-leveltaskselectionfunctionevaluationsalgorithmsreference-based
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

Bilevel optimization is expensive because every candidate upper-level solution requires solving a full lower-level problem, and most of those candidates are discarded anyway. This paper tries to break that bottleneck by deciding, before any lower-level work, which candidates are promising enough to deserve a lower-level solve. The proposed framework, CR-BLEA, trains a contrastive ranking network online from previously evaluated solutions and uses the learned relative rankings to allocate lower-level resources to the top half of each new population. The authors report that wrapping five existing bilevel evolutionary algorithms in this framework reduces total function evaluations by 28.8–36.3% on standard test suites while preserving or improving accuracy, with even larger savings on two real-world problems. The practical payoff is that expensive nested optimization could be run at a fraction of the cost without giving up solution quality.

What carries the argument

The load-bearing object is the contrastive ranking network $N_{\mathrm{cr}}$, a siamese architecture with two fully connected sub-networks that share weights: each maps an upper-level candidate $x_u$ to a representation $S$, and the model outputs $\sigma(S_i - S_j)$, the probability that candidate $i$ beats candidate $j$. A first quasi-residual layer maps the $m$ upper-level variables to $n$ neurons to mimic the reaction-set mapping $\psi(x_u) \approx x^*_l$ before joining the original features, which is how the model ties rankings to the lower-level response. Training uses binary cross-entropy on $N(N-1)$ paired labels derived from the pool of evaluated solutions; inference uses a reference-based variant $N^R_{\mathrm{cr}}$ with the second branch fixed to 0 so all candidates are scored against the same reference, avoiding pairwise re-ranking and transitivity cycles. The same scores drive a resampling rule that regenerates the offspring when the best new candidate ranks below the best parent candidate. This machinery together converts "which lower-level tasks to run" into a ranking problem solved from data already produced by the evolutionary process.

What would settle it

Give CR-BLEA a bilevel benchmark in which, for the same upper-level choice, the lower level has two or more equally optimal responses, and track both the ranking model's pair accuracy and the total function-evaluation saving across 21 runs; if the pair accuracy hovers near chance and the saving rate collapses toward zero while upper-level error grows, the central claim is refuted. The paper's own SMD6 and SMD12 results (58.6% and 62.3% model accuracy, roughly 7–9% savings) show the expected signature.

Watch

Extended reading notes

Core claim

The paper's central claim is that most of the computing spent in nested bilevel evolutionary search is wasted on lower-level tasks derived from upper-level candidates that later get eliminated, and that this waste can be removed by learning to rank candidates before optimizing them. Because the lower-level optimal response $x^*_l$ is treated as a function of the upper-level variable $x_u$, the upper-level fitness $F(x_u)$ is learnable from $x_u$ alone, so a network trained on pairs of evaluated solutions can say which of two candidates is better. CR-BLEA uses this pairwise knowledge through a reference-based score to keep the best half of each offspring population and allocate lower-level function evaluations only to those survivors, with a resampling step when the best offspring score trails the parent best. Across five base bilevel evolutionary algorithms, the framework reports equivalent or better accuracy on nearly all SMD and TP benchmark problems while cutting total evaluations by 28.8% to 36.3% on average, and large reductions on two real-world problems. The authors acknowledge the boundary of the claim: when the lower-level problem has multiple or infinitely many optima for the same $x_u$ (SMD6 and SMD12), model accuracy drops to about 59–62% and the savings shrink accordingly.

Load-bearing premise

Everything hinges on the assumption that a candidate's best lower-level response is a single, predictable function of the candidate itself; when a problem has several equally good lower-level answers, the ranking signal degrades and the savings shrink.

Editorial extensions

If this is right

  • Five base bilevel evolutionary algorithms (including transfer-learning, multi-objective-transformation, and collaboration-based variants) each consume 28.8–36.3% fewer total function evaluations on the SMD and TP suites when wrapped in CR-BLEA, with accuracy statistically equivalent or better on almost every problem.
  • The savings come mostly from the lower level: because only the top half of each offspring population is sent to the lower-level optimizer, lower-level evaluation counts fall by a similar margin while upper-level evaluations also drop because the search converges faster.
  • On problems where the lower level has multiple global optima for a given upper-level choice (SMD6, SMD12), the ranking model's accuracy falls to 58.6% and 62.3%, and the total-evaluation savings shrink to roughly 7–9%, marking the known boundary of the approach.
  • The ablation study shows both components matter: selecting tasks at random instead of by the model degrades accuracy and saves far fewer evaluations, while removing the resampling step mostly increases evaluation counts.
  • On two real-world bilevel problems, the wrapped algorithm reduces total function evaluations by 88.5% and 48.6% while reaching a similar upper-level objective and better lower-level objective.

Reading between the lines

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

  • A testable extension is to apply the same rank-then-allocate recipe to other nested expensive optimizations, such as neural architecture search or hyperparameter tuning, where the inner training run is the analogue of the lower-level solve; the paper does not test these settings.
  • Because the model is trained on relative order rather than absolute fitness, its usefulness may degrade gracefully as the population distribution shifts; one could measure online rank correlation between predicted scores and true fitness each generation to see when retraining is needed.
  • The reference-based score implicitly defines a scalar promise function on upper-level variables; if that is the real object being learned, then cheaper surrogate rankers such as isotonic regression over a few features might achieve similar resource savings, which would be a useful comparison the paper does not make.
  • The framework's own failure cases suggest a natural stress test: deliberately construct bilevel problems with multimodal reaction sets and count how often the model misranks pairs; the paper's SMD6 and SMD12 numbers imply that the saving rate would degrade continuously with the degree of multimodality.
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 / 6 minor

Summary. The manuscript proposes CR-BLEA, a framework that wraps a nested bilevel evolutionary algorithm and uses an online-trained contrastive ranking network to score newly generated upper-level solutions. Only the top half of the offspring is sent to the lower-level optimizer, and a reference-based ranking score is also used to trigger resampling when the offspring population appears inferior to the parents. The framework is integrated with five existing BLEAs (TLEA-CMA-ES, TLEA-DE, BL-CMA-ES, MOTEA, and BOC). Experiments on the SMD suite (12 problems, all five variants), the TP suite (only the TLEA-CMA-ES variant), and two real-world problems report average total function-evaluation reductions of roughly 28.8%–36.3%, with largely equivalent or better accuracy, and an ablation shows that the learned ranking outperforms random task selection and that resampling improves efficiency.

Significance. If the results hold, the paper offers a practical, algorithm-agnostic way to reduce the dominant lower-level cost in nested BLEAs. The empirical design has real strengths: standardized SMD/TP benchmarks, 21 independent runs per instance, Wilcoxon tests at 0.05, and an ablation that separates the contribution of the contrastive ranking model from the resampling heuristic. The paper also honestly identifies the failure cases on SMD6 and SMD12. However, the headline claim is stated without the single-valued-reaction-set caveat that the paper's own Section 4.4 shows to be necessary, and the evaluation does not compare against the closest existing resource-preselection methods. The contribution is therefore plausible but currently supported for a narrower scope than the abstract claims.

major comments (4)
  1. [Abstract, §4.3, §4.4, Tables 1–5] The headline claim is not supported on problems where the lower-level reaction set is multi-valued. Section 3.2(1) assumes x_l* is essentially a function of x_u, and Section 4.4 reports that on SMD6 (infinitely many lower-level optima) and SMD12 (multiple global lower-level optima) the contrastive model accuracy drops to 58.6% and 62.3%. The consequences are visible in Tables 1–5: FE savings on SMD6 range from -0.6% to 11.2%, and on SMD12 from 8.6% to 26.3%; Table 5 shows CR-BOC is significantly worse on SMD6 upper-level accuracy (3.37e-01 vs 1.00e-06, marked '-'), and Table 4 shows CR-MOTEA requires significantly more upper-level FEs on SMD6 (5.42e+02 vs 4.76e+02, marked '-'). The abstract's unqualified statement that CR-BLEA 'significantly reduces computational cost while preserving, or even enhancing, solution accuracy' should be restricted to problems where the reaction-set mapping is single-valued, or the framework should be modified to degrade gracefully when the ranking premise fails.
  2. [§4.3, §4.5] The evaluation does not compare CR-BLEA with existing resource-preselection or resource-allocation methods. The only ablation baseline is random task selection (v1), and the five comparison algorithms are the underlying BLEAs without any preselection mechanism. Related approaches such as classification model-based assisted preselection [31], multifidelity lower-level allocation [39], and surrogate-assisted reaction-set approximation [33–36] are discussed in Section 2.2 but are never used as baselines. Consequently, the paper has not established that the proposed contrastive-ranking mechanism is better than the closest data-driven preselection strategies; it only shows that it is better than solving all tasks and better than random selection. Add at least one representative resource-preselection baseline, or explicitly limit the claim to improving nested BLEAs that do not already use data-driven preselection.
  3. [§4.3, Table 6] The TP benchmark results are reported for only one base algorithm, CR-TLEA-CMA-ES. The text states that 'similar trends were observed across other variants' without presenting the corresponding data or statistical results. Since the abstract claims comprehensive experiments across five algorithms and the TP suite has a different problem structure from SMD, the missing per-algorithm TP results prevent verification of generality on this suite. Provide the full TP results for all five variants in the main text or an appendix, or revise the claim to state that only one representative integration was tested on TP.
  4. [§3.2, §4.2] The contrastive network training setup is not specified enough to reproduce. Section 3.2 gives the architecture sketch (shared sub-networks, a quasi-residual layer, BCE loss, Adam with learning rate 0.1) and Section 4.2 gives population and termination settings, but the number of hidden layers and neurons per layer (beyond Fig. 2(b)), the number of training epochs or gradient steps per update, the batch size, and the actual value of Np implied by 'ten times the size of the network's trainable parameters' are not reported. Because the method's effectiveness depends on this online-trained model, these details, and ideally a sensitivity analysis of the learning rate and Np, should be added.
minor comments (6)
  1. [Introduction] The first sentence contains the typo 'Bilvel'; it should be 'Bilevel'.
  2. [Tables 1–5] The column header 'Probelm' should be 'Problem' in all five comparison tables.
  3. [§2.1] The phrase 'x_l^* indicates the implicit constraints of BLOPs' is imprecise; x_l^* is the lower-level optimal solution, and the implicit constraint is that the upper level must evaluate F only at such solutions.
  4. [Algorithm 2] The symbol l_{i,j} is used both as a sign in line 3 and as a 0/1 label in lines 4–5; define the conversion between these two forms explicitly to avoid ambiguity.
  5. [§4.4, Fig. 4] The model accuracy in Fig. 4 is reported as a single average without variance or confidence intervals; report standard deviations or use box plots, since the average mixes models trained on different amounts of data.
  6. [General] No code repository is mentioned; given the number of tunable components (network architecture, Np, resampling rule), releasing code would substantially aid reproducibility.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central FE-savings and accuracy claims are benchmarked against external SMD/TP/real-world problems and five external base algorithms; the SMD6/SMD12 caveat is a correctness limitation, not a circular derivation.

full rationale

The paper's central empirical claims (28.8% to 36.3% average total-FE savings while preserving or improving accuracy) are measured against the external SMD and TP test suites plus two real-world problems, with CR- variants compared against five base algorithms (TLEA-CMA-ES, TLEA-DE, BL-CMA-ES, MOTEA, BOC) that are not authored by this group, so the headline results are externally anchored rather than reducible to the method's own outputs. The FE savings are not forced by construction: the ablation variant v1 (random task selection) also skips half of the lower-level tasks yet consumes more FEs and suffers accuracy deterioration, showing the contrastive model contributes real selection information beyond the trivial act of evaluating fewer tasks. No load-bearing step reduces to its own inputs by the paper's equations: the ranking network is trained on pairwise labels derived from true F(xu, x*l) evaluations, and tested on subsequently generated populations under a sequential holdout protocol; the reference-based ranking and resampling are heuristic uses of the model, with final accuracy always checked against known benchmark optima. The 'x*l is a function of xu' premise (Sec. 3.2, item 1) is a stated modeling assumption that the paper itself falsifies on SMD6 and SMD12 (Sec. 4.4 reports model accuracy of 58.6% and 62.3%, with ψ(xu) multimodal), which is an acknowledged limitation and an abstract-level overclaim, not a circular step. Two minor, non-circular blemishes keep the score above zero: (i) the model accuracy reported in Fig. 4 is computed on data produced within the same algorithm loop, so its test points are partially self-selected by the model's own rankings; and (ii) self-citations [41] and [42] appear only in the related-work survey and are not load-bearing for the framework's claims. Neither issue makes the derivation circular, so the honest verdict is no significant circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the domain assumption that the lower-level optimal solution is a deterministic function of the upper-level variable, so that a network can rank upper-level candidates from xu alone. The method also relies on several hand-chosen hyperparameters (selection fraction, learning rate, pool threshold, resampling trigger) that are not fitted to the benchmark data but affect the reported savings. No new physical entities are postulated.

free parameters (4)
  • Learning rate of Adam optimizer = 0.1
    Set in Section 3.2. This is a hand-chosen hyperparameter that affects network training stability and the quality of the learned ranking scores.
  • Selection fraction of offspring = Nu/2 (top half)
    The framework always performs lower-level optimization for the top half of the newly generated upper-level population. This design choice directly determines the maximum possible FE savings and is not derived from data.
  • Pool threshold Np = Np(Np-1) >= 10 * trainable parameters
    This threshold controls when the contrastive network is first trained and how often it is updated. It is tied to the network size and chosen by hand.
  • Resampling trigger condition = max offspring ranking score < max parent ranking score
    Resampling is triggered when the best predicted score of the offspring is below that of the parents. This heuristic condition controls how often extra offspring are generated and is chosen by hand.
assumptions (3)
  • domain assumption The lower-level optimal solution x*l is a deterministic function of the upper-level variable xu (the reaction-set mapping is single-valued).
    Invoked in Section 3.2 'Model Preference' when the paper equates data (xu, x*l) with (xu, psi(xu)) and then with (xu, F(xu)). This assumption underpins the claim that a network can rank upper-level solutions using only xu. The paper itself notes it breaks on SMD6 and SMD12, where model accuracy drops to 58.6% and 62.3%.
  • domain assumption The upper-level fitness F(xu) is learnable and rankable by a neural network from the limited online samples collected during the search.
    The whole CR-BLEA framework depends on the contrastive network generalizing from paired samples in the pool Pe to rank new offspring. No formal generalization guarantee is provided; the paper relies on empirical validation.
  • domain assumption The benchmark termination criteria and FE budgets provide a fair comparison between base algorithms and CR variants.
    Section 4.2 sets identical termination conditions for both, but CR-BLEA often terminates earlier, and accuracy is compared at termination rather than at equal FE budgets. This choice affects the interpretation of the accuracy comparisons.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CR-BLEA: Contrastive Ranking for Adaptive Resource Allocation in Bilevel Evolutionary Algorithms." pith.science (2026). https://pith.science/paper/4QOKOUDN

@misc{pith2026250606362,
  author       = {Pith},
  title        = {Pith review of: CR-BLEA: Contrastive Ranking for Adaptive Resource Allocation in Bilevel Evolutionary Algorithms},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4QOKOUDN}},
  note         = {Machine review of arXiv:2506.06362}
}
read the original abstract

Bilevel optimization poses a significant computational challenge due to its nested structure, where each upper-level candidate solution requires solving a corresponding lower-level problem. While evolutionary algorithms (EAs) are effective at navigating such complex landscapes, their high resource demands remain a key bottleneck -- particularly the redundant evaluation of numerous unpromising lower-level tasks. Despite recent advances in multitasking and transfer learning, resource waste persists. To address this issue, we propose a novel resource allocation framework for bilevel EAs that selectively identifies and focuses on promising lower-level tasks. Central to our approach is a contrastive ranking network that learns relational patterns between paired upper- and lower-level solutions online. This knowledge guides a reference-based ranking strategy that prioritizes tasks for optimization and adaptively controls resampling based on estimated population quality. Comprehensive experiments across five state-of-the-art bilevel algorithms show that our framework significantly reduces computational cost while preserving -- or even enhancing -- solution accuracy. This work offers a generalizable strategy to improve the efficiency of bilevel EAs, paving the way for more scalable bilevel optimization.

Figures

Figures reproduced from arXiv: 2506.06362 by the authors.

Figure 1
Figure 1. Flowchart of the proposed CR-BLEA. The middle section with a gray back [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Structure of the proposed contrastive ranking network. [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Convergence curve of TLEA-CMA-ES and CR-TLEA-CMA-ES on SMD prob [PITH_FULL_IMAGE:figures/full_fig_p022_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Average Model Accuracy of CR-TLEA-CMA-ES on Different SMD Problems. [PITH_FULL_IMAGE:figures/full_fig_p023_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages

  1. [31]

    L. Lin, T. Liu, J. Leng, S. Yao, H. Zhang, L. Wei, Q. Liu, Classifi- cation model-based assisted preselection and environment selection ap- proach for evolutionary expensive bilevel optimization, Applied Intelli- gence 53 (23) (2023) 28377–28400

    work page 2023

  2. [39]

    M. M. Mamun, H. K. Singh, T. Ray, A multifidelity approach for bilevel optimization with limited computing budget, IEEE Transactions on Evolutionary Computation 26 (2) (2021) 392–399

    work page 2021

  3. [1]

    Ghasemi, F

    P. Ghasemi, F. Goodarzian, A. Gunasekaran, A. Abraham, A bi-level mathematical model for logistic management considering the evolution- ary game with environmental feedbacks, The international journal of logistics management 34 (4) (2023) 1077–1100

    work page 2023

  4. [2]

    Zhang, P

    Y. Zhang, P. Khanduri, I. Tsaknakis, Y. Yao, M. Hong, S. Liu, An introduction to bilevel optimization: Foundations and applications in signal processing and machine learning, IEEE Signal Processing Maga- zine 41 (1) (2024) 38–59

    work page 2024

  5. [3]

    Zhang, G

    Y. Zhang, G. Zhang, P. Khanduri, M. Hong, S. Chang, S. Liu, Revisiting and advancing fast adversarial training through the lens of bi-level op- timization, in: International Conference on Machine Learning, PMLR, 2022, pp. 26693–26712

    work page 2022

  6. [4]

    J. F. Bard, Some properties of the bilevel programming problem, Journal of optimization theory and applications 68 (2) (1991) 371–378

    work page 1991

  7. [5]

    Camacho-Vallejo, C

    J.-F. Camacho-Vallejo, C. Corpus, J. G. Villegas, Metaheuristics for bilevel optimization: A comprehensive review, Computers & Operations Research (2023) 106410

    work page 2023

  8. [6]

    Chaabani, L

    A. Chaabani, L. Ben Said, Solving hierarchical production–distribution problem based on MDVRP under flexibility depot resources in supply chain management, in: Advances in Computational Logistics and Supply Chain Analytics, Springer, 2023, pp. 129–147. 27

    work page 2023

Show all 51 references
  1. [7]

    X. He, Y. Zhou, Z. Chen, Evolutionary bilevel optimization based on covariance matrix adaptation, IEEE Transactions on Evolutionary Com- putation 23 (2) (2018) 258–272

    2018

  2. [8]

    Chen, H.-L

    L. Chen, H.-L. Liu, K. C. Tan, K. Li, Transfer learning-based parallel evolutionary algorithm framework for bilevel optimization, IEEE Trans- actions on Evolutionary Computation 26 (1) (2021) 115–129

    2021

  3. [9]

    Chen, H.-L

    L. Chen, H.-L. Liu, K. Li, K. C. Tan, Evolutionary bi-level optimiza- tion via multi-objective transformation-based lower level search, IEEE Transactions on Evolutionary Computation (2023)

    2023

  4. [10]

    Huang, Q

    P.-Q. Huang, Q. Zhang, Y. Wang, Bilevel optimization via collabora- tions among lower-level optimization tasks, IEEE Transactions on Evo- lutionary Computation (2023)

    2023

  5. [11]

    Gupta, J

    A. Gupta, J. Ma´ ndziuk, Y.-S. Ong, Evolutionary multitasking in bi-level optimization, Complex & Intelligent Systems 1 (2015) 83–95

    2015

  6. [12]

    R. Liu, J. Gao, J. Zhang, D. Meng, Z. Lin, Investigating bi-level opti- mization for learning and vision from a unified perspective: A survey and beyond, IEEE Transactions on Pattern Analysis and Machine In- telligence 44 (12) (2021) 10045–10067

    2021

  7. [13]

    Roghanian, M.-B

    E. Roghanian, M.-B. Aryanezhad, S. J. Sadjadi, Integrating goal pro- gramming, Kuhn–Tucker conditions, and penalty function approaches to solve linear bi-level programming problems, Applied Mathematics and Computation 195 (2) (2008) 585–590

    2008

  8. [14]

    H. Tuy, A. Migdalas, P. V¨ arbrand, A global optimization approach for the linear two-level program, Journal of Global Optimization 3 (1993) 1–23

    1993

  9. [15]

    J. F. Bard, J. T. Moore, A branch and bound algorithm for the bilevel programming problem, SIAM Journal on Scientific and Statistical Com- puting 11 (2) (1990) 281–292

    1990

  10. [16]

    T. A. Edmunds, J. F. Bard, Algorithms for nonlinear bilevel mathemat- ical programs, IEEE transactions on Systems, Man, and Cybernetics 21 (1) (1991) 83–89. 28

    1991

  11. [17]

    Y. Lv, T. Hu, G. Wang, Z. Wan, A penalty function method based on Kuhn–Tucker condition for solving linear bilevel programming, Applied mathematics and computation 188 (1) (2007) 808–813

    2007

  12. [18]

    Colson, P

    B. Colson, P. Marcotte, G. Savard, A trust-region method for nonlinear bilevel programming: algorithm and computational experience, Compu- tational Optimization and Applications 30 (2005) 211–227

    2005

  13. [19]

    Vicente, G

    L. Vicente, G. Savard, J. J´ udice, Descent approaches for quadratic bilevel programming, Journal of Optimization theory and applications 81 (2) (1994) 379–399

    1994

  14. [20]

    Li, A genetic algorithm using a finite search space for solving non- linear/linear fractional bilevel programming problems, Annals of Oper- ations Research 235 (1) (2015) 543–558

    H. Li, A genetic algorithm using a finite search space for solving non- linear/linear fractional bilevel programming problems, Annals of Oper- ations Research 235 (1) (2015) 543–558

    2015

  15. [21]

    Jiang, X

    Y. Jiang, X. Li, C. Huang, X. Wu, Application of particle swarm op- timization based on CHKS smoothing function for solving nonlinear bilevel programming problem, Applied Mathematics and Computation 219 (9) (2013) 4332–4339

    2013

  16. [22]

    H. Li, L. Zhang, Y.-C. Jiao, An interactive approach based on a discrete differential evolution algorithm for a class of integer bilevel programming problems, International Journal of Systems Science 47 (10) (2016) 2330– 2341

    2016

  17. [23]

    Z. Wan, L. Mao, G. Wang, Estimation of distribution algorithm for a class of nonlinear bilevel programming problems, Information Sciences 256 (2014) 184–196

    2014

  18. [24]

    Y. Wang, H. Li, C. Dang, A new evolutionary algorithm for a class of nonlinear bilevel programming problems and its global convergence, INFORMS Journal on Computing 23 (4) (2011) 618–629

    2011

  19. [25]

    M. M. Islam, H. K. Singh, T. Ray, A. Sinha, An enhanced memetic al- gorithm for single-objective bilevel optimization problems, Evolutionary computation 25 (4) (2017) 607–642

    2017

  20. [26]

    L. Zhao, J. Wei, A nested particle swarm algorithm based on sphere mutation to solve bi-level optimization, Soft Computing 23 (21) (2019) 11331–11341. 29

    2019

  21. [27]

    H. I. Calvete, C. Gal´ e, M.-J. Oliveros, Bilevel model for production– distribution planning solved by using ant colony optimization, Comput- ers & operations research 38 (1) (2011) 320–327

    2011

  22. [28]

    Huang, K

    Y. Huang, K. Wang, K. Gao, T. Qu, H. Liu, Jointly optimizing micro- grid configuration and energy consumption scheduling of smart homes, Swarm and Evolutionary Computation 48 (2019) 251–261

    2019

  23. [29]

    J. S. Angelo, H. J. Barbosa, A study on the use of heuristics to solve a bilevel programming problem, International Transactions in Operational Research 22 (5) (2015) 861–882

    2015

  24. [30]

    Sinha, V

    A. Sinha, V. Shaikh, Solving bilevel optimization problems using krig- ing approximations, IEEE Transactions on Cybernetics 52 (10) (2021) 10639–10654

    2021

  25. [32]

    H. K. Singh, M. M. Islam, T. Ray, M. Ryan, Nested evolutionary al- gorithms for computationally expensive bilevel optimization problems: Variants and their systematic analysis, Swarm and Evolutionary Com- putation 48 (2019) 329–344

    2019

  26. [33]

    Sinha, P

    A. Sinha, P. Malo, K. Deb, Evolutionary algorithm for bilevel optimiza- tion using approximations of the lower level optimal solution mapping, European Journal of Operational Research 257 (2) (2017) 395–411

    2017

  27. [34]

    Sinha, P

    A. Sinha, P. Malo, K. Deb, An improved bilevel evolutionary algorithm based on quadratic approximations, in: 2014 IEEE congress on evolu- tionary computation (CEC), IEEE, 2014, pp. 1870–1877

    2014

  28. [35]

    Sinha, P

    A. Sinha, P. Malo, K. Deb, Solving optimistic bilevel programs by itera- tively approximating lower level optimal value function, in: 2016 IEEE Congress on Evolutionary Computation (CEC), IEEE, 2016, pp. 1877– 1884. 30

    2016

  29. [36]

    Sinha, Z

    A. Sinha, Z. Lu, K. Deb, P. Malo, Bilevel optimization based on iterative approximation of multiple mappings, Journal of Heuristics 26 (2) (2020) 151–185

    2020

  30. [37]

    M. M. Islam, H. K. Singh, T. Ray, A surrogate assisted approach for single-objective bilevel optimization, IEEE Transactions on Evolution- ary Computation 21 (5) (2017) 681–696

    2017

  31. [38]

    Kieffer, G

    E. Kieffer, G. Danoy, M. R. Brust, P. Bouvry, A. Nagih, Tackling large- scale and combinatorial bi-level problems with a genetic programming hyper-heuristic, IEEE Transactions on Evolutionary Computation 24 (1) (2019) 44–56

    2019

  32. [40]

    Y. Feng, L. Feng, S. Kwong, K. C. Tan, A multi-form evolutionary search paradigm for bi-level multi-objective optimization, IEEE Transactions on Evolutionary Computation (2023)

    2023

  33. [41]

    Jiang, Z

    M. Jiang, Z. Huang, L. Qiu, W. Huang, G. G. Yen, Transfer learning- based dynamic multiobjective optimization algorithms, IEEE Transac- tions on Evolutionary Computation 22 (4) (2017) 501–514

    2017

  34. [42]

    Jiang, Z

    M. Jiang, Z. Wang, H. Hong, G. G. Yen, Knee point-based imbalanced transfer learning for dynamic multiobjective optimization, IEEE Trans- actions on Evolutionary Computation 25 (1) (2020) 117–129

    2020

  35. [43]

    Chaabani, L

    A. Chaabani, L. B. Said, Transfer of learning with the co- evolutionary decomposition-based algorithm-II: a realization on the bi- level production-distribution planning system, Applied Intelligence 49 (2019) 963–982

    2019

  36. [44]

    Chaabani, S

    A. Chaabani, S. Bechikh, L. Ben Said, A co-evolutionary hybrid decomposition-based algorithm for bi-level combinatorial optimization problems, Soft Computing 24 (10) (2020) 7211–7229

    2020

  37. [45]

    Gupta, Y.-S

    A. Gupta, Y.-S. Ong, L. Feng, Multifactorial evolution: toward evolu- tionary multitasking, IEEE Transactions on Evolutionary Computation 20 (3) (2015) 343–357. 31

    2015

  38. [46]

    K. K. Bali, A. Gupta, Y.-S. Ong, P. S. Tan, Cognizant multitasking in multiobjective multifactorial evolution: MO-MFEA-II, IEEE transac- tions on cybernetics 51 (4) (2020) 1784–1796

    2020

  39. [47]

    T. Chen, S. Kornblith, M. Norouzi, G. Hinton, A simple framework for contrastive learning of visual representations, in: International confer- ence on machine learning, PMLR, 2020, pp. 1597–1607

    2020

  40. [48]

    Khorram, J

    S. Khorram, J. Kim, A. Tripathi, H. Lu, Q. Zhang, H. Sak, Contrastive siamese network for semi-supervised speech recognition, in: ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Sig- nal Processing (ICASSP), IEEE, 2022, pp. 7207–7211

    2022

  41. [49]

    Huang, Y.-J

    H.-G. Huang, Y.-J. Gong, Contrastive learning: An alternative surro- gate for offline data-driven evolutionary computation, IEEE Transac- tions on Evolutionary Computation 27 (2) (2022) 370–384

    2022

  42. [50]

    Sinha, P

    A. Sinha, P. Malo, K. Deb, Test problem construction for single- objective bilevel optimization, Evolutionary computation 22 (3) (2014) 439–477

    2014

  43. [51]

    Huang, Y

    P.-Q. Huang, Y. Wang, A framework for scalable bilevel optimization: Identifying and utilizing the interactions between upper-level and lower- level variables, IEEE Transactions on Evolutionary Computation 24 (6) (2020) 1150–1163. 32

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.