Pith. sign in

REVIEW 3 major objections 6 minor 33 references

Searching Efficient Deep Architectures for Radar Target Detection using Monte-Carlo Tree Search

T0 review · 3 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 Monte-Carlo tree search, scored by a training-free metric, finds a radar target detector that matches an expert U-Net's detection probability on endoclutter maps with only 40% of its parameters.

desk verdict Useful NAS-for-radar result, but the central comparison rests on an unvalidated proxy and a single run per algorithm. read the letter →

arxiv 2506.21772 v1 pith:T4BG5XBL submitted 2025-06-11 eess.SP cs.LG

Noé Lallouet , Tristan Cazenave , Cyrille Enderli , Stéphanie Gourdin This is my paper
classification eess.SPcs.LG
keywords radartargetdetectionneuralarchitecturesearchMonte-Carlotreenestedtraining-freemetricrange-DopplermapU-Netembedded
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

Radar target detectors built from deep networks work well in clutter, but their size keeps them off embedded platforms, which often have only a CPU and strict real-time limits. This paper tries to automate the design of lightweight detectors by searching a cell-based NASNet space, extended with an upsampling cell so U-Net-like segmentation architectures can be generated, and by scoring every candidate at initialization with a training-free metric instead of training it. The central claim is that the architecture found by Nested Monte-Carlo Search matches the expert-designed U-Net baseline's detection probability on endoclutter range-Doppler maps while using 48,209 parameters versus the baseline's 120,441, with a false-alarm probability of 0.29e-4 versus 0.30e-4. If the claim holds, the main obstacle to deploying deep radar detectors, their computational cost, can be attacked at the architecture-design stage rather than by compressing already-heavy networks.

What carries the argument

The machinery is a Monte-Carlo tree search over a NASNet-style cell space, extended with a third Upsample Cell type that doubles feature-map size so the search can assemble U-Net-like segmentation detectors from normal, reduction, and upsampling blocks. Candidates are scored without training by the metric $\log|K_H|$, where $K_H$ is a kernel matrix of Hamming distances between the binary ReLU activation patterns of the inputs in a minibatch at initialization; a high score means the untrained network already separates its inputs. A hard complexity constraint returns reward zero during a playout whenever a sampled network exceeds the baseline U-Net's parameter count, pushing the search toward light cells, and eight parallel leaf playouts stabilize the value estimates. The winning architecture comes from Nested Monte-Carlo Search, which selects moves by recursively running lower-level searches rather than by bandit-style statistics.

What would settle it

Compute the rank correlation between the training-free score $\log|K_H|$ and the post-training detection probability across the architectures returned by the four search algorithms; if the score does not order candidates the way their trained detection probability does on the held-out set, the search was not guided by the claimed proxy. A second check: retrain the 48,209-parameter NMCS network and the 120,441-parameter U-Net from matched initializations with identical training budgets, and compare detection probability at a fixed false-alarm rate separately on each of the nine clutter scenarios; the headline comparison fails if the lighter network loses to the baseline on any scenario.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that Nested Monte-Carlo Search, guided by a training-free initialization score, discovers a radar target detector whose detection performance is comparable to the current state of the art on endoclutter environments while using only 40% of the baseline model's parameters. The numbers behind that claim: the NMCS-searched network has 48,209 parameters versus 120,441 for the expert U-Net, a test loss of 0.54 versus 0.57, a false-alarm probability of 0.29e-4 versus 0.30e-4, and a detection probability on par with the baseline, which the paper describes as slightly higher (Figure 3). The paper also claims the first applications of GRAVE and Nested Monte-Carlo Search to neural architecture search, and argues that the winning network's cell design, not its parameter count, is what lets a light architecture perform like a heavy one.

Load-bearing premise

The load-bearing premise is that an untrained network's ability to separate inputs, measured by the initialization score $\log|K_H|$ over Hamming distances of ReLU activations, reliably ranks how well the architecture will detect targets after real training, a premise the paper does not validate and explicitly calls 'possibly ill-suited to evaluating the false alarm probability.'

Editorial extensions

If this is right

  • The 48,209-parameter detector is a concrete, hardware-friendlier alternative to the 120,441-parameter expert U-Net for endoclutter radar detection, with on-par detection probability and false-alarm rate.
  • Because the winning network is not the largest one, detection performance on this task is driven more by cell design than by parameter count, so searching directly for efficiency is a viable route to embedded deployment.
  • Monte-Carlo NAS with a training-free initialization score becomes a practical pipeline: a 25-minute single-GPU search followed by one 3-hour training run produces the reported detector, without training every candidate.
  • The added Upsample Cell extends the NASNet search space from classification to segmentation-style detectors, so the same search machinery applies to other dense-prediction radar tasks.

Reading between the lines

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

  • Because the score that guided the search is suspected of mis-ranking false-alarm behaviour, a cheap extension would be to score candidates by detection probability at a fixed false-alarm rate rather than by the single $\log|K_H|$ value; the authors list exactly this as future work.
  • The reported failure mode, searches undershooting the parameter budget and settling into shallow local minima, suggests that a softened complexity penalty in the reward, rather than a hard cutoff, could let Monte-Carlo methods explore more expressive spaces.
  • A benchmark the paper does not run: train the NMCS cell design at the same parameter budget as the expert U-Net by scaling channel width, to see whether the searched architecture also dominates the baseline on the performance-versus-size trade-off curve.
  • The same search protocol could transfer to other radar representations such as micro-Doppler signatures or raw I/Q data, where the cell space and the training-free score would need re-validation but the pipeline itself carries over.
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. This paper applies Monte-Carlo Tree Search (MCTS) variants - UCT, MC-RAVE, GRAVE, and Nested Monte-Carlo Search (NMCS) - to design convolutional neural network architectures for radar target detection on range-Doppler maps. The search space extends the NASNet cell-based space with a new Upsample Cell, allowing U-Net-like segmentation networks. Candidate architectures are scored at initialization using a training-free proxy (Mellor et al.'s log|K_H| score) under a parameter budget constraint. The authors report that the NMCS-searched architecture achieves detection probability comparable to an expert-designed U-Net baseline while using only 40% of its parameters (48,209 vs. 120,441). The models are evaluated on a held-out test set of simulated endoclutter radar signals. The paper also discusses limitations, including the need for additional clutter scenarios and longer training, and acknowledges that the training-free metric may be ill-suited to false alarm probability estimation.

Significance. If fully validated, the paper would make a useful contribution: it introduces GRAVE and NMCS to neural architecture search, extends the NASNet search space to segmentation tasks with an Upsample Cell, and demonstrates that an automatically searched lightweight network can match an expert-designed baseline on a radar detection task, which is practically relevant for embedded systems. The use of a held-out test set and the explicit statement of limitations are positive features. However, the central efficiency claim currently rests on an unvalidated training-free proxy and single-run comparisons, so the significance is conditional on additional empirical support.

major comments (3)
  1. [Section III-C and Section IV] The search reward at every terminal node is the training-free score of Eq. (3), computed from Hamming distances of ReLU activations at initialization. The paper provides no evidence that this score correlates with trained detection probability (PD) or false alarm probability (PFA) on the radar range-Doppler task. Section IV even states that the metric is 'possibly ill-suited to evaluating the false alarm probability at initialization.' Since this proxy is the sole selection signal, the NMCS result in Table II could be the lucky tail of a random search. Please add a validation study (e.g., a scatter plot or Spearman rank correlation between proxy scores and trained PD/PFA for a sample of architectures) or an ablation that compares proxy-driven MCTS against random search with the same search-time budget.
  2. [Table II and Figure 3] All performance comparisons are based on a single training run per architecture, with no error bars, no multiple seeds, and no significance tests. The PFA values for NMCS (0.29e-4) and the baseline U-Net (0.30e-4) are extremely close, and the PD curves in Figure 3 lack confidence bands. The claimed parity or slight superiority of NMCS is therefore not statistically established. Please report means and standard deviations over several training seeds, or provide confidence intervals for PD and PFA.
  3. [Section III-A] The search space restricts each cell to a single block (N=1), whereas NASNet recommends N=5. The paper justifies this choice only by 'search efficiency purposes' and does not analyze whether this restricted space is expressive enough to contain competitive radar detectors. The conclusion that the discovered architecture is 'efficient' is meaningful only within this narrow family. Please add a discussion of the expressiveness of the single-block search space, and ideally a comparison with a larger N for at least one algorithm.
minor comments (6)
  1. [Section III-C] The MC-RAVE selection formula and the definition of β contain typographical errors (e.g., '˜ni+ni+4ni ˜ni˜b2' appears malformed) and missing parentheses. Please rewrite the equations with clear notation and define all variables.
  2. [Table II] The column 'Test loss' is not defined; specify whether it is binary cross-entropy and how it is computed on the test set. Also, the PFA values are given as x10^-4 in the header but referred to as 0.30 in the text without units; make the units consistent.
  3. [Section IV] The sentence 'The detection probabilities associated with these architectures... can be appreciated in Figure 3' should be supplemented with numerical PD values in a table or in the text, since the figure alone does not support quantitative comparison of the architectures.
  4. [Section III-B] Please clarify the role of the validation set: is it used for early stopping or model selection during training, and were any of the search algorithms given access to validation labels? The text mentions an 80%-20% train-validation split but does not state how the validation set is used.
  5. [Section V] The paragraph on the extended search space (additional hyperparameters such as number of blocks or initial channels) is too vague to be reproducible. Either provide the details of the extended space and the failure mode, or move this discussion entirely to future work.
  6. [Introduction] There are grammatical and typographical errors throughout, e.g., 'prohibitionally' should be 'prohibitively.' A careful proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the search-time proxy and the final trained evaluation are independent.

full rationale

The paper's selection of architectures is driven by the external training-free score log|K_H| from Mellor et al. [13], computed at initialization from Hamming distances of ReLU activations. The final comparison in Table II and Figure 3 is obtained only after training the selected architectures and evaluating them on a held-out test set of 2000 novel range-Doppler maps. The proxy is not fitted to the radar task, and no fitted parameter is renamed as a prediction: the reported PD and PFA values come from trained networks on separate test data. The complexity bound alpha is set to the baseline U-Net parameter count as a constraint, not derived from the target detection metrics; the fact that NMCS returns a smaller network is an unforced search outcome. The self-citations present ([4] for the expert U-Net baseline, [30] and [31] for GRAVE and NMCS, [32] for leaf parallelization) are not load-bearing: the baseline is an independent comparison point, and the search algorithms are general methods rather than results that presuppose the radar outcome. The paper's own admission that the Mellor metric is 'possibly ill-suited to evaluating the false alarm probability at initialization' (Section IV) is an evidence-quality weakness about proxy validity, not circularity, because the central efficiency claim is not defined in terms of the proxy. No equation in the paper reduces by construction to its own input, so no circular step is present.

Assumptions & free parameters 5 free parameters · 6 assumptions · 1 invented entities

The ledger is modest for an applied NAS paper: the only numeric free parameters are search hyperparameters (alpha, k, tref, RAVE bias) and the one-block design choice; none are fit to the test set. The central assumptions are that range-Doppler detection is a segmentation problem, that parameter count proxies latency, and that the external training-free score ranks architectures that will train well. The main unverified load-bearing assumption is proxy validity for radar data. No new physical entities are introduced; the Upsample Cell is a search-space component, not a physical postulate.

free parameters (5)
  • alpha (parameter budget) = 120441 (parameter count of baseline U-Net)
    Hand-set complexity cap. During MCTS playout, any sampled architecture with more parameters receives reward 0, which strongly shapes the search. Section III-C.
  • exploration constant k = not reported
    Tunable constant in UCT, MC-RAVE, and GRAVE node-selection formulas; no value or sensitivity analysis is given. Section III-C.
  • tref = 30
    GRAVE switches to ancestors' AMAF statistics when node visits are below 30; fixed without ablation. Section III-C.
  • RAVE bias constant b = not reported
    Appears in the beta weighting formula of MC-RAVE; the value is not given. Section III-C.
  • number of blocks per cell N = 1
    NASNet defaults to N=5 but the search uses a single block for efficiency; the adequacy of this reduction is not validated. Section III-A.
assumptions (6)
  • domain assumption Radar target detection reduces to binary pixel-wise segmentation of range-Doppler maps under H0/H1.
    Frames the whole problem in Section I; if false, the segmentation formulation and all searched architectures may be mismatched.
  • domain assumption The number of network parameters is a good hardware-agnostic proxy for latency.
    Used to enforce the complexity bound alpha and to compare frugality; stated in Section III-C without benchmarks.
  • domain assumption The training-free metric score = log|K_H| computed at initialization ranks final trained detection performance on radar data.
    Borrowed from [13]; no correlation analysis on this radar task; Section IV acknowledges poor alignment with PFA.
  • domain assumption RAVE's assumption that move order does not matter holds for NAS move sequences.
    Needed for MC-RAVE and GRAVE to be valid; stated in Section III-C.
  • ad hoc to paper A one-block NASNet cell search space is expressive enough to contain competitive radar detectors.
    Chosen for search efficiency rather than demonstrated sufficiency; Section III-A.
  • domain assumption Zero padding all maps to 128x128 preserves all detection-relevant information.
    Applied to all examples in Section III-B; no analysis of resolution loss.
invented entities (1)
  • Upsample Cell
    purpose: Doubles the spatial dimension inside NASNet-style cells so the search can generate U-Net-like segmentation networks.
    Introduced in Section III-A and used in the searched architectures; its benefits are evidenced only by the single best architecture found here, with no independent benchmark.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Searching Efficient Deep Architectures for Radar Target Detection using Monte-Carlo Tree Search." pith.science (2026). https://pith.science/paper/T4BG5XBL

@misc{pith2026250621772,
  author       = {Pith},
  title        = {Pith review of: Searching Efficient Deep Architectures for Radar Target Detection using Monte-Carlo Tree Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T4BG5XBL}},
  note         = {Machine review of arXiv:2506.21772}
}
read the original abstract

Recent research works establish deep neural networks as high performing tools for radar target detection, especially on challenging environments (presence of clutter or interferences, multi-target scenarii...). However, the usually large computational complexity of these networks is one of the factors preventing them from being widely implemented in embedded radar systems. We propose to investigate novel neural architecture search (NAS) methods, based on Monte-Carlo Tree Search (MCTS), for finding neural networks achieving the required detection performance and striving towards a lower computational complexity. We evaluate the searched architectures on endoclutter radar signals, in order to compare their respective performance metrics and generalization properties. A novel network satisfying the required detection probability while being significantly lighter than the expert-designed baseline is proposed.

Figures

Figures reproduced from arXiv: 2506.21772 by the authors.

Figure 1
Figure 1. Monte-Carlo tree search those long training times by training a very large network only once, but they also suffer from some drawbacks (e.g. deep coupling between architecture parameters and supernetwork weights). Recently, drawing from the neural network pruning literature [10] [11], NAS research efforts investigate training￾free metrics in order to score architecture at initialization, without training the candida… view at source ↗
Figure 2
Figure 2. A training dataset range-Doppler map architecture to the search of cells, which are then stacked to produce the final architecture. Each cell is composed of N blocks, which have searchable inputs and operations. Even though [27] recommends N = 5, we choose to search a single block, for search efficiency purposes. The NASNet search space has been designed for the search of image classification networks, and thus only… view at source ↗
Figure 3
Figure 3. Detection probability PD The normal, reduction and upsample cells returned by the most effective search algorithm, Nested Monte-Carlo Search, are shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Best performing architecture cells to produce high-performing networks with the severe network complexity constraint introduced in Section III. Indeed, the MCTS-based algorithms fell in local minima, represented by hyperparameters choices associated to overly shallow n…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 23 canonical work pages

  1. [1]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” 2015. [Online]. Available: https://arxiv.org/abs/1512.03385

    arXiv 2015

  2. [2]

    Deep Radar Detector,

    D. Brodeski, I. Bilik, and R. Giryes, “Deep Radar Detector,” in2019 IEEE Radar Conference (RadarConf). Boston, MA, USA: IEEE, Apr. 2019, pp. 1–6. [Online]. Available: https://ieeexplore.ieee.org/document/8835792/

    work page arXiv 2019

  3. [3]

    A Neyman-Pearson Criterion-Based Neural Network Detector for Maritime Radar,

    Z. Baird, M. K. McDonald, S. Rajan, and S. Lee, “A Neyman-Pearson Criterion-Based Neural Network Detector for Maritime Radar,” in 2021 IEEE 24th International Conference on Information Fusion (FUSION) . Sun City, South Africa: IEEE, Nov. 2021, pp. 1–8. [Online]. Available: https://ieeexplore.ieee.org/document/9626944/

    work page arXiv 2021

  4. [4]

    Loss Function Design For Training Robust Radar Detectors Using Deep Learning,

    N. Lallouet, T. Cazenave, C. Enderli, and S. Gourdin, “Loss Function Design For Training Robust Radar Detectors Using Deep Learning,” in Conference on Artificial Intelligence for Defense . Rennes, France: DGA Ma ˆıtrise de l’Information, Nov. 2023. [Online]. Available: https://hal.science/hal-04328554 5 Published as a conference paper at RADAR 2024

    work page 2023

  5. [5]

    Deep Learning-Based UA V Detection in Pulse-Doppler Radar,

    C. Wang, J. Tian, J. Cao, and X. Wang, “Deep Learning-Based UA V Detection in Pulse-Doppler Radar,” IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp. 1–12, 2022

    work page 2022

  6. [6]

    Radar Target Detection with CNN,

    F. Yavuz, “Radar Target Detection with CNN,” in 2021 29th European Signal Processing Conference (EUSIPCO) , 2021, pp. 1581–1585

    work page 2021

  7. [7]

    Neural Architecture Search with Reinforcement Learning,

    B. Zoph and Q. V . Le, “Neural Architecture Search with Reinforcement Learning,” 2016. [Online]. Available: https://arxiv.org/abs/1611.01578

    arXiv 2016

  8. [8]

    DARTS: Differentiable Architecture Search,

    H. Liu, K. Simonyan, and Y . Yang, “DARTS: Differentiable Architecture Search,” arXiv, Tech. Rep. arXiv:1806.09055, Apr. 2019, arXiv:1806.09055 [cs, stat] type: article. [Online]. Available: http://arxiv.org/abs/1806.09055

    arXiv 2019

Show all 33 references
  1. [9]

    Single Path One-Shot Neural Architecture Search with Uniform Sampling,

    Z. Guo, X. Zhang, H. Mu, W. Heng, Z. Liu, Y . Wei, and J. Sun, “Single Path One-Shot Neural Architecture Search with Uniform Sampling,” arXiv, Tech. Rep. arXiv:1904.00420, Jul. 2020, arXiv:1904.00420 [cs] type: article. [Online]. Available: http://arxiv.org/abs/1904.00420

    1904 arXiv

  2. [10]

    Pruning neural networks without any data by iteratively conserving synaptic flow,

    H. Tanaka, D. Kunin, D. L. K. Yamins, and S. Ganguli, “Pruning neural networks without any data by iteratively conserving synaptic flow,” arXiv, Tech. Rep. arXiv:2006.05467, Nov. 2020, arXiv:2006.05467 [cond-mat, q-bio, stat] type: article. [Online]. Available: http://arxiv.or...

    2006 arXiv

  3. [11]

    Picking Winning Tickets Before Training by Preserving Gradient Flow,

    C. Wang, G. Zhang, and R. Grosse, “Picking Winning Tickets Before Training by Preserving Gradient Flow,” arXiv, Tech. Rep. arXiv:2002.07376, Aug. 2020, arXiv:2002.07376 [cs, stat] type: article. [Online]. Available: http://arxiv.org/abs/2002.07376

    2002 arXiv

  4. [12]

    Zero-Cost Proxies for Lightweight NAS,

    M. S. Abdelfattah, A. Mehrotra, L. Dudziak, and N. D. Lane, “Zero-Cost Proxies for Lightweight NAS,” arXiv, Tech. Rep. arXiv:2101.08134, Mar. 2021, arXiv:2101.08134 [cs] type: article. [Online]. Available: http://arxiv.org/abs/2101.08134

    2021 arXiv

  5. [13]

    Neural Architecture Search without Training,

    J. Mellor, J. Turner, A. Storkey, and E. J. Crowley, “Neural Architecture Search without Training,” arXiv, Tech. Rep. arXiv:2006.04647, Jun. 2021, arXiv:2006.04647 [cs, stat] type: article. [Online]. Available: http://arxiv.org/abs/2006.04647

    2006 arXiv

  6. [14]

    NASI: Label- and Data-agnostic Neural Architecture Search at Initialization,

    Y . Shu, S. Cai, Z. Dai, B. C. Ooi, and B. K. H. Low, “NASI: Label- and Data-agnostic Neural Architecture Search at Initialization,” arXiv, Tech. Rep. arXiv:2109.00817, Apr. 2022, arXiv:2109.00817 [cs] type: article. [Online]. Available: http://arxiv.org/abs/2109.00817

    2022 arXiv

  7. [15]

    DeepHybrid: Deep Learning on Automotive Radar Spectra and Reflec- tions for Object Classification,

    A.-E. Cozma, L. Morgan, M. Stolz, D. Stoeckel, and K. Rambach, “DeepHybrid: Deep Learning on Automotive Radar Spectra and Reflec- tions for Object Classification,” in 2021 IEEE International Intelligent Transportation Systems Conference (ITSC) , 2021, pp. 2682–2687

    2021

  8. [16]

    Exploration d’architectures de r ´eseaux de neurones pour la segmentation s ´emantique d’images a ´eriennes,

    A. Archet, F. Orieux, N. Ventroux, and N. Gac, “Exploration d’architectures de r ´eseaux de neurones pour la segmentation s ´emantique d’images a ´eriennes,” Aug. 2023

    2023

  9. [17]

    3D Object Detection of Cars and Pedestrians by Deep Neural Networks from Unit-Sharing One-Shot NAS,

    O. T.-C. Chen, Y .-X. Chang, Y .-W. Jhao, C.-Y . Chung, Y .-L. Chang, and W.-H. Huang, “3D Object Detection of Cars and Pedestrians by Deep Neural Networks from Unit-Sharing One-Shot NAS,” in 2022 18th IEEE International Conference on Advanced Video and Signal Based Surveillan...

    2022

  10. [18]

    SCM: A Searched Convolutional Metaformer for SAR Ship Classification,

    H. Zhu, S. Guo, W. Sheng, and L. Xiao, “SCM: A Searched Convolutional Metaformer for SAR Ship Classification,” Remote Sensing, vol. 15, no. 11,

  11. [19]

    MCTS/UCT in solving real-life problems,

    J. Ma ´ndziuk, “MCTS/UCT in solving real-life problems,” in Studies in Computational Intelligence, Jan. 2018, pp. 277–292

    2018

  12. [20]

    Modification of UCT with Patterns in Monte-Carlo Go,

    S. Gelly, Y . Wang, and O. Teytaud, “Modification of UCT with Patterns in Monte-Carlo Go,” Jan. 2006

    2006

  13. [21]

    Dec-MCTS: Decen- tralized planning for multi-robot active perception,

    G. Best, O. Cliff, T. Patten, R. Mettu, and R. Fitch, “Dec-MCTS: Decen- tralized planning for multi-robot active perception,” The International Journal of Robotics Research , vol. 38, p. 027836491875592, Mar. 2018

    2018

  14. [22]

    Optimization of the Nested Monte-Carlo Algorithm on the Traveling Salesman Problem with Time Windows,

    A. Rimmel, F. Teytaud, and T. Cazenave, “Optimization of the Nested Monte-Carlo Algorithm on the Traveling Salesman Problem with Time Windows,” in Evostar, Turin, Italy, Apr. 2011. [Online]. Available: https://inria.hal.science/inria-00563668

    2011

  15. [23]

    DeepArchitect: Automatically Designing and Training Deep Architectures,

    R. Negrinho and G. Gordon, “DeepArchitect: Automatically Designing and Training Deep Architectures,” arXiv, Tech. Rep. arXiv:1704.08792, Apr. 2017, arXiv:1704.08792 [cs, stat] type: article. [Online]. Available: http://arxiv.org/abs/1704.08792

    2017 arXiv

  16. [24]

    Practical Deep Learning Architecture Optimization,

    M. Wistuba, “Practical Deep Learning Architecture Optimization,” in 2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA) . IEEE, Oct. 2018. [Online]. Available: http://dx.doi.org/10.1109/DSAA.2018.00037

    2018

  17. [25]

    Monte-Carlo tree search and rapid action value estimation in computer Go,

    S. Gelly and D. Silver, “Monte-Carlo tree search and rapid action value estimation in computer Go,” Artificial Intelligence, vol. 175, pp. 1856–1875, 2011. [Online]. Available: http://www.sciencedirect.com/ science/article/pii/S000437021100052X

    2011

  18. [26]

    AlphaX: eXploring Neural Architectures with Deep Neural Networks and Monte Carlo Tree Search,

    L. Wang, Y . Zhao, Y . Jinnai, Y . Tian, and R. Fonseca, “AlphaX: eXploring Neural Architectures with Deep Neural Networks and Monte Carlo Tree Search,” arXiv, Tech. Rep. arXiv:1903.11059, Oct. 2019, arXiv:1903.11059 [cs] type: article. [Online]. Available: http://arxiv.org/ab...

    1903 arXiv

  19. [27]

    Learning Transferable Architectures for Scalable Image Recognition,

    B. Zoph, V . Vasudevan, J. Shlens, and Q. V . Le, “Learning Transferable Architectures for Scalable Image Recognition,” arXiv, Tech. Rep. arXiv:1707.07012, Apr. 2018, arXiv:1707.07012 [cs, stat] type: article. [Online]. Available: http://arxiv.org/abs/1707.07012

    2018 arXiv

  20. [28]

    U-Net: Convolutional Networks for Biomedical Image Segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional Networks for Biomedical Image Segmentation,” 2015. [Online]. Available: https://arxiv.org/abs/1505.04597

    2015 arXiv

  21. [29]

    Bandit Based Monte-Carlo Planning,

    L. Kocsis and C. Szepesv ´ari, “Bandit Based Monte-Carlo Planning,” in Machine Learning: ECML 2006 , J. F ¨urnkranz, T. Scheffer, and M. Spiliopoulou, Eds. Berlin, Heidelberg: Springer Berlin Heidelberg, 2006, pp. 282–293

    2006

  22. [30]

    Generalized Rapid Action Value Estimation,

    T. Cazenave, “Generalized Rapid Action Value Estimation,” in 24th International Conference on Artificial Intelligence , Buenos Aires, Argentina, Jul. 2015, pp. 754–760. [Online]. Available: https://hal.science/hal-01436522

    2015

  23. [31]

    Nested Monte-Carlo Search,

    ——, “Nested Monte-Carlo Search,” 2009

    2009

  24. [32]

    On the Parallelization of UCT,

    T. Cazenave and N. Jouandeau, “On the Parallelization of UCT,” in Computer Games Workshop , Amsterdam, Netherlands, Jun. 2007. [Online]. Available: https://hal.science/hal-02310186 6

    2007

  25. [2023]

    Available: https://www.mdpi.com/2072-4292/15/11/2904

    [Online]. Available: https://www.mdpi.com/2072-4292/15/11/2904

Pith tools

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