Pith. sign in

REVIEW 3 major objections 5 minor 60 references

Global optimization of graph acquisition functions for neural architecture search

T0 review · 3 major / 5 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 This paper claims that graph acquisition functions in Bayesian optimization for neural architecture search can be globally optimized by encoding the graph space into a mixed-integer program whose feasible region is in exact bijection with…

desk verdict Solid extension of BoGrape to arbitrary graphs with a sound encoding theorem; the empirical 'global optimization' claim needs solver-gap reporting but the paper deserves review. read the letter →

arxiv 2505.23640 v1 pith:QKEIU7VE submitted 2025-05-29 cs.LG math.OC

Yilin Xie , Shiqiang Zhang , Jixiang Qing , Ruth Misener , Calvin Tsay This is my paper
classification cs.LGmath.OC MSC 90C1168R1068T0790C27
keywords neuralarchitecturesearchgraphBayesianoptimizationacquisitionfunctionmixed-integerprogrammingshortest-pathkernelNAS-Bench-101NAS-Bench-201encoding
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

The paper introduces NAS-GOAT, a Bayesian-optimization method for neural architecture search in which the step that proposes the next architecture, normally done by mutation or sampling, is replaced by an exact mixed-integer program. The core claim is that the space of valid cell graphs, including reachability and shortest-path information, can be encoded into variables and linear constraints so that every architecture corresponds to exactly one feasible solution and every feasible solution to one architecture. With this encoding, the lower-confidence-bound acquisition function and the shortest-path graph kernel become part of the mixed-integer program, so the next candidate is the global optimum of the acquisition function rather than a heuristic guess. The paper demonstrates this on NAS-Bench-101 and NAS-Bench-201, where the method finds near-optimal architectures in most tested settings. If the claim holds, graph Bayesian optimization for NAS gains an exactness that sampling- and mutation-based acquisition optimization lacks.

What carries the argument

The load-bearing object is the graph encoding of Eq. (Graph-Encoding): a system of linear constraints over node-existence variables $A_{v,v}$, edge variables $A_{u,v}$, reachability variables $r_{u,v}$, shortest distances $d_{u,v}\in[n+1]$, and shortest-path membership variables $\delta^w_{u,v}$. Conditions (C1)-(C8) force these variables to take exactly the values they would take for some graph, and Theorem 1 proves the feasible set is in bijection with the graph space with $n_0$ to $n$ nodes. This encoding is what lets kernel values and the LCB acquisition function be written as mixed-integer programming expressions, so the acquisition subproblem becomes a finite mixed-integer program that a solver can optimize with a global optimality certificate.

What would settle it

Run the NAS-GOAT MIP for NAS-Bench-101 with $n=7$, $E=9$ and record the solver's reported optimality gap and solve time; if any Bayesian-optimization iteration terminates at the 1800-second time limit with a nonzero gap, the claim of global optimization at each iteration is not realized in practice. A second check: on NAS-Bench-201, enumerate all 15,625 architectures, evaluate the trained Gaussian process's LCB value for each, and compare the best true value with the MIP's returned candidate; any mismatch would show the encoded acquisition optimization is not equivalent to a true global search.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the graph search space of cell-based neural architecture search can be represented exactly as the feasible region of a mixed-integer linear system, without assuming strong connectivity. The encoding introduces node-existence, edge, reachability, shortest-distance, and shortest-path-membership variables constrained by conditions (C1)-(C8); Theorem 1 proves a bijection between the feasible domain and the whole graph space with node counts in $[n_0, n]$. Adding DAG, single-source, single-sink, and node- or edge-label constraints restricts the encoding to the NAS-Bench-101 and NAS-Bench-201 search spaces. Because the shortest-path kernel and the lower-confidence-bound acquisition function are written in these same variables, the acquisition-optimization subproblem is a mixed-integer program that can be solved to global optimality, and the numerical results show that full Bayesian-optimization loops using this exact acquisition optimization find near-optimal architectures across the tested benchmarks.

Load-bearing premise

The promised global-optimality guarantee rests on the MIP solver certifying optimality before the 1800-second time limit; the paper does not report optimality gaps or solve times, so that guarantee may not be met for the largest NAS-Bench-101 cases.

Editorial extensions

If this is right

  • At each Bayesian-optimization iteration, the next architecture is the global minimizer of the LCB acquisition function over the entire search space, so the invalid-candidate rejection that mutation and sampling solvers must handle is avoided.
  • The encoding covers arbitrary weakly connected directed acyclic graphs with any node count in $[n_0,n]$, so the method transfers from NAS-Bench-101 and NAS-Bench-201 to other cell-based search spaces with one source and one sink.
  • Both node-labeled and edge-labeled architectures are handled by the same kernel and acquisition formulation through the linear kernel form combining graph-structure, node-label, and edge-label terms.
  • If each MIP is solved to proven optimality, the quality of proposed architectures is limited by the Gaussian-process surrogate and kernel rather than by the acquisition search procedure.

Reading between the lines

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

  • An untested but natural control experiment would compare NAS-GOAT against the same graph-GP surrogate and shortest-path kernel with mutation-based acquisition optimization, isolating the value of exact global acquisition optimization from the choice of surrogate.
  • The same bijective graph encoding may extend to other graph black-box optimization problems, such as molecular design or network synthesis, whenever reachability and shortest-path properties define feasibility.
  • Scalability beyond the small benchmark graphs is open: the experiments cap MIP solve time at 1800 seconds and restrict NAS-Bench-101 to sizes 6 and 7, so the tractability of exact acquisition optimization on larger architecture spaces is not established by this paper.
  • The kernel comparison indicates that WL kernels predict better than the SP and ESP kernels used here, yet NAS-GOAT still optimizes well; this hints that exact acquisition optimization can compensate for a weaker surrogate, a trade-off that would be worth testing directly.
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 / 5 minor

Summary. The paper proposes NAS-GOAT, a graph Bayesian optimization framework for neural architecture search in which the graph search space and the shortest-path kernel are encoded as a mixed-integer program, enabling the LCB acquisition function to be optimized by solving a MIP at each BO iteration. The encoding introduces variables for node existence, edges, reachability, shortest distances, and shortest-path membership, and Theorem 1 claims a bijection between the feasible set of the constraint system and the space of all graphs with n0 to n nodes. The framework is specialized to cell-based NAS via node-labeled (NAS-Bench-101) and edge-labeled (NAS-Bench-201) DAG constraints. Experiments compare NAS-GOAT with eleven baselines on four benchmark tasks, reporting competitive or superior validation/test error, and kernel comparisons show the exponential form of the proposed kernel to be competitive with WL kernels.

Significance. If the central claims hold, the paper makes a valuable contribution to graph BO for NAS: it provides a principled alternative to mutation- and sampling-based acquisition optimization, with a theoretical encoding theorem that is proved in the appendix and appears sound. The ability to handle both node and edge labels, to enforce NAS-specific structural constraints, and to formulate acquisition optimization as a discrete global optimization problem is a genuine advance over prior sample-based approaches. The paper also builds on the authors' earlier BoGrape framework and extends it to the weakly connected, acyclic graphs typical of NAS, which is a nontrivial generalization. The main caveat is that the experimental support for the 'global optimization' claim is incomplete: the MIP solver time limit is reported, but no optimality gaps, solver statuses, or solve times are given, and the NAS-Bench-101 search is restricted to 6- and 7-node graphs without empirical justification.

major comments (3)
  1. [4.4, Appendix C.1] The central claim (Section 5) that NAS-GOAT 'globally optimizes the acquisition function at each BO iteration' is not supported by the reported experiments. Appendix C.1 sets Gurobi TimeLimit=1800 s and uses PoolSearchMode=2 to return 5 candidates, but the paper reports no solver statuses, no optimality gaps, and no solve times for any MIP instance. If Gurobi terminates at the time limit on the larger NAS-Bench-101 instances, the returned candidates are feasible incumbents rather than certified global optima, and the claimed advantage over mutation- and sampling-based acquisition optimization is unverified. Please report per-iteration solver status and optimality gap, or restrict the global optimality claim to instances solved to proven optimality.
  2. [Appendix C.1] The NAS-Bench-101 experiments are restricted to graph sizes N=6 and N=7, with the statement that 'most high-quality architectures in NAS-Bench-101 have either 6 or 7 nodes.' This is an unsupported assumption about the benchmark. If the optimal or near-optimal architectures have fewer nodes, the method cannot find them, and the comparison with baselines that search the full NAS-Bench-101 space is not fair. Please provide evidence for this claim (e.g., the distribution of top-performing architectures by node count in NAS-Bench-101) or explicitly state that the global optimization claim applies only to the 6/7-node subset.
  3. [3.3, 4.4] The paper repeatedly describes the final acquisition optimization as a 'MIP' (Sections 1 and 4.4), but NAS-GOAT-E uses the exponential kernel kexp(X1,X2)=σ_k^2 exp(k_lin(X1,X2)), which makes the LCB acquisition optimization a mixed-integer nonlinear program involving exp and nonlinear covariance terms. The paper does not describe how this problem is solved to global optimality with Gurobi (e.g., piecewise-linear approximation, spatial branching, or a solver-specific global method), and no optimality certificates are reported. Please clarify the exact mathematical programming formulation for the exponential kernel and state what global optimality guarantees, if any, Gurobi provides for it.
minor comments (5)
  1. [Table 3] The linear SP kernel used by NAS-GOAT-L has substantially worse MNLL than the exponential form (e.g., 227.67 vs. 28.83 on NAS-Bench-101) and also underperforms WL in Spearman correlation; the paper should discuss the implication of this poor uncertainty quantification for the BO loop, given that the linear kernel is the version whose MIP formulation is straightforward.
  2. [Figure 3] The captions state 'Median with one standard deviation over 20 replications,' but it is unclear whether the shaded region is the standard deviation of the median or of the raw values; please clarify the convention.
  3. [Eq. (1a)] The notation 'w≠u,v' should be written as 'w ∉ {u,v}' to avoid ambiguity regarding the intended scope of the universal quantifier.
  4. [Table 2] The baseline name 'Local seach' contains a typo and should read 'Local search'.
  5. [4.4] The paper reports no wall-clock time for the full BO loop; given the 1800 s MIP time limit, a comparison of total runtime against the mutation- and sampling-based baselines would be informative for assessing practical efficiency.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation found; the graph encoding is proved from first principles, and the self-cited kernel and MIP encodings are restated in Appendix B rather than assumed.

full rationale

The central claim of the paper is that Eq. (Graph-Encoding) is a bijection onto the graph space and that the resulting MIP globally optimizes the acquisition function. This claim is not circular: the encoding is derived from the eight necessary conditions C1-C8 in Appendix A.1, and Theorem 1 is proved directly in Appendix A.2 by induction on shortest distances. The NAS restrictions in Section 3.2 are additional constraints on that encoding and do not presuppose the optimal architecture. The kernel and acquisition MIP encodings in Section 3.3 and Appendix B are taken from the authors' prior BoGrape paper (Xie et al., 2025), and the text explicitly says they are 'given in (Xie et al., 2025)' and 'proposed in (Xie et al., 2025)'. However, the paper restates the full formulations in Appendix B, including linearizations of the quadratic kernel terms, so the derivation does not reduce to a black-box self-citation. No fitted parameter is renamed as a prediction: GP kernel parameters are trained in the standard way, and the reported results are evaluated on held-out benchmark data. The only substantive weakness is a verification gap, not circularity: Appendix C.1 sets Gurobi TimeLimit=1800s and uses PoolSearchMode=2, but the paper reports no optimality gaps, solution times, or solver statuses, so the empirical claim of certified global optimality at each BO iteration is not verified. That concern is about experimental evidence, not about the derivation reducing to its inputs.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method rests on the correctness of the graph encoding (proved), the chosen kernel form and its MIP encoding (largely from prior work by the same authors), the tractability of the resulting MIPs, and an experimental restriction of NAS-Bench-101 to 6-7 node graphs. No new physical or conceptual entities are introduced.

free parameters (6)
  • alpha (kernel weight on graph structure) = optimized by GPflow, bounds [0.01,100]
    Used in kernel (linear) as coefficient for k_g; trained on observed architecture data.
  • beta (kernel weight on node labels) = optimized by GPflow, bounds [0.01,100]
    Used in kernel (linear) as coefficient for k_n; trained on observed data.
  • gamma (kernel weight on edge labels) = optimized by GPflow, bounds [0.01,100]
    Used in kernel (linear) as coefficient for k_e; trained on observed data.
  • sigma_k^2 (exponential kernel variance) = optimized by GPflow, bounds [0.01,100]
    Controls magnitude of exponential kernel (exponential); trained on data.
  • beta_t^{1/2} (LCB exploration weight) = 3
    Hand-set to balance exploration and exploitation; not tuned. Affects acquisition optimization.
  • Gurobi TimeLimit = 1800 seconds
    Hand-set solver budget per MIP. If insufficient for proof of optimality, the 'global' claim is put in doubt.
assumptions (4)
  • domain assumption Shortest-path kernel (Eq. (kg)) is a valid graph kernel and its MIP encoding is faithful.
    The paper borrows the SP kernel from Borgwardt and Kriegel (2005) and the encoding from Xie et al. (2025); validity is assumed from prior literature.
  • domain assumption The GP with LCB acquisition (beta_t^{1/2}=3) is a suitable surrogate for NAS performance.
    Standard BO practice; suitability for architecture accuracy is assumed rather than derived.
  • ad hoc to paper Gurobi can solve the resulting MIP to global optimality within practical time limits.
    The method's central advantage depends on solver tractability; the paper imposes a time limit but does not verify optimality gaps.
  • ad hoc to paper For NAS-Bench-101, restricting the search to 6 or 7 node graphs does not exclude the optimal architecture.
    Appendix C.1 uses this restriction to avoid variable-size kernel normalization; if the best architecture had fewer nodes, the method could miss it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Global optimization of graph acquisition functions for neural architecture search." pith.science (2026). https://pith.science/paper/QKEIU7VE

@misc{pith2026250523640,
  author       = {Pith},
  title        = {Pith review of: Global optimization of graph acquisition functions for neural architecture search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QKEIU7VE}},
  note         = {Machine review of arXiv:2505.23640}
}
read the original abstract

Graph Bayesian optimization (BO) has shown potential as a powerful and data-efficient tool for neural architecture search (NAS). Most existing graph BO works focus on developing graph surrogates models, i.e., metrics of networks and/or different kernels to quantify the similarity between networks. However, the acquisition optimization, as a discrete optimization task over graph structures, is not well studied due to the complexity of formulating the graph search space and acquisition functions. This paper presents explicit optimization formulations for graph input space including properties such as reachability and shortest paths, which are used later to formulate graph kernels and the acquisition function. We theoretically prove that the proposed encoding is an equivalent representation of the graph space and provide restrictions for the NAS domain with either node or edge labels. Numerical results over several NAS benchmarks show that our method efficiently finds the optimal architecture for most cases, highlighting its efficacy.

Figures

Figures reproduced from arXiv: 2505.23640 by the authors.

Figure 1
Figure 1. Illustration of NAS-GOAT. The main idea is to represent graphs in variable space and introduce constraints to [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Predictive performance of graph GPs with different kernels. 50 and 400 architectures are randomly sampled [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Numerical results of Graph BO on NAS-Bench-101 (N101) and NAS-Bench-201 (N201). ( [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Predictive performance of graph GPs with different kernels. 50 and 400 architectures are randomly sampled [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: Comparison NAS-GOAT with the remaining baselines. Numerical results of Graph BO on NAS-Bench-101 [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: Numerical results of Graph BO on NAS-Bench-101 (N101) and NAS-Bench-201 (N201). ( [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 53 canonical work pages

  1. [1]

    B. L. Ammari, E. S. Johnson, G. Stinchfield, T. Kim, M. Bynum, W. E. Hart, J. Pulsipher, and C. D. Laird. Linear model decision trees as surrogates in optimization of engineering applications. Computers & Chemical Engineering, 178, 2023

    work page 2023

  2. [2]

    Anderson, J

    R. Anderson, J. Huchette, W. Ma, C. Tjandraatmadja, and J. P. Vielma. Strong mixed-integer programming formulations for trained neural networks. Mathematical Programming, 183 0 (1): 0 3--39, 2020

    work page 2020

  3. [3]

    Borgwardt, E

    K. Borgwardt, E. Ghisu, F. Llinares-L \'o pez, L. O’Bray, B. Rieck, et al. Graph kernels: State-of-the-art and future challenges. Foundations and Trends in Machine Learning , 13 0 (5-6): 0 531--712, 2020

    work page 2020

  4. [4]

    K. M. Borgwardt and H.-P. Kriegel. Shortest-path kernels on graphs. In International Conference on Data Mining, 2005

    work page 2005

  5. [5]

    Cheng, J

    A. Cheng, J. Wang, X. Zhang, Q. Chen, P. Wang, and J. Cheng. DPNAS : Neural architecture search for deep learning with differential privacy. AAAI, 2022

    work page 2022

  6. [6]

    Dong and Y

    X. Dong and Y. Yang. NAS-Bench-201 : Extending the scope of reproducible neural architecture search. In ICLR, 2020

    work page 2020

  7. [7]

    Elsken, J

    T. Elsken, J. H. Metzen, and F. Hutter. Neural architecture search: a survey. Journal of Machine Learning Research, 2019

    work page 2019

  8. [8]

    Fischetti and J

    M. Fischetti and J. Jo. Deep neural networks and mixed integer linear optimization. Constraints, 23 0 (3): 0 296--309, 2018

    work page 2018

Show all 60 references
  1. [9]

    R. W. Floyd. Algorithm 97: Shortest path. Communications of the ACM, 5 0 (6): 0 345--345, 1962

    1962

  2. [10]

    P. I. Frazier. A tutorial on B ayesian optimization. arXiv preprint arXiv:1807.02811, 2018

    2018 arXiv

  3. [11]

    R. Garnett. Bayesian Optimization. Cambridge University Press, 2023

    2023

  4. [12]

    G \"a rtner, P

    T. G \"a rtner, P. Flach, and S. Wrobel. On graph kernels: Hardness results and efficient alternatives. In Learning Theory and Kernel Machines, 2003

    2003

  5. [13]

    Gurobi optimizer reference manual , 2024

    Gurobi Optimization, LLC . Gurobi optimizer reference manual , 2024. URL https://www.gurobi.com

    2024

  6. [14]

    Hojny, S

    C. Hojny, S. Zhang, J. S. Campos, and R. Misener. Verifying message-passing neural networks via topology-based bounds tightening. In ICML, 2024

    2024

  7. [15]

    Huchette, G

    J. Huchette, G. Mu \ n oz, T. Serra, and C. Tsay. When deep learning meets polyhedral theory: A survey. arXiv preprint arXiv:2305.00241, 2023

    2023 arXiv

  8. [16]

    Jaafra, J

    Y. Jaafra, J. L. Laurent, A. Deruyver, and M. S. Naceur. Reinforcement learning for neural architecture search: A review. Image and Vision Computing, 89: 0 57--66, 2019

    2019

  9. [17]

    Kandasamy, W

    K. Kandasamy, W. Neiswanger, J. Schneider, B. Poczos, and E. P. Xing. Neural architecture search with B ayesian optimisation and optimal transport. NeurIPS, 31, 2018

    2018

  10. [18]

    Kriege and P

    N. Kriege and P. Mutzel. Subgraph matching kernels for attributed graphs. In ICML, 2012

    2012

  11. [19]

    N. M. Kriege, P.-L. Giscard, and R. Wilson. On valid optimal assignment kernels and applications to graph classification. NeurIPS, 2016

    2016

  12. [20]

    N. M. Kriege, F. D. Johansson, and C. Morris. A survey on graph kernels. Applied Network Science, 5: 0 1--42, 2020

    2020

  13. [21]

    C. Liu, B. Zoph, M. Neumann, J. Shlens, W. Hua, L.-J. Li, L. Fei-Fei, A. Yuille, J. Huang, and K. Murphy. Progressive neural architecture search. In ECCV, 2018

    2018

  14. [22]

    H. Liu, K. Simonyan, and Y. Yang. DARTS : Differentiable architecture search. In ICLR, 2019

    2019

  15. [23]

    A. G. d. G. Matthews, M. van der Wilk , T. Nickson, K. Fujii, A. Boukouvalas , P. Le \'o n-Villagr \'a , Z. Ghahramani, and J. Hensman. GP flow: A G aussian process library using T ensor F low . Journal of Machine Learning Research, 18 0 (40): 0 1--6, 2017

    2017

  16. [24]

    McDonald, C

    T. McDonald, C. Tsay, A. M. Schweidtmann, and N. Yorke-Smith. Mixed-integer optimisation of graph neural networks for computer-aided molecular design. Computers & Chemical Engineering, 185: 0 108660, 2024

    2024

  17. [25]

    V. V. Mi s i \'c . Optimization of tree ensembles. Operations Research, 68 0 (5): 0 1605--1624, 2020

    2020

  18. [26]

    Mistry, D

    M. Mistry, D. Letsios, G. Krennrich, R. M. Lee, and R. Misener. Mixed-integer convex nonlinear optimization with gradient-boosted trees embedded. INFORMS Journal on Computing, 33 0 (3): 0 1103--1119, 2021

    2021

  19. [27]

    Neiswanger, K

    W. Neiswanger, K. Kandasamy, B. Poczos, J. Schneider, and E. Xing. ProBo : a framework for using probabilistic programming in B ayesian optimization. arXiv preprint arXiv:1901.11515, 2019

    1901 arXiv

  20. [28]

    Nikolentzos, G

    G. Nikolentzos, G. Siglidis, and M. Vazirgiannis. Graph kernels: A survey. Journal of Artificial Intelligence Research, 72: 0 943--1027, 2021

    2021

  21. [29]

    Z. Qiu, W. Bi, D. Xu, H. Guo, H. Ge, Y. Liang, H. P. Lee, and C. Wu. Efficient self-learning evolutionary neural architecture search. Applied Soft Computing, 146: 0 110671, 2023

    2023

  22. [30]

    E. Real, A. Aggarwal, Y. Huang, and Q. V. Le. Regularized evolution for image classifier architecture search. In AAAI, 2019

    2019

  23. [31]

    P. Ren, Y. Xiao, X. Chang, P.-Y. Huang, Z. Li, X. Chen, and X. Wang. A comprehensive survey of neural architecture search: Challenges and solutions. ACM Computing Surveys, 54 0 (4): 0 1--34, 2021

    2021

  24. [32]

    B. Ru, X. Wan, X. Dong, and M. Osborne. Interpretable neural architecture search via B ayesian optimisation with W eisfeiler- L ehman kernels. In ICLR, 2021

    2021

  25. [33]

    Salmani Pour Avval, N

    S. Salmani Pour Avval, N. D. Eskue, R. M. Groves, and V. Yaghoubi. Systematic review on neural architecture search. Artificial Intelligence Review, 58 0 (3): 0 73, 2025

    2025

  26. [34]

    Schulz, M

    E. Schulz, M. Speekenbrink, and A. Krause. A tutorial on G aussian process regression: Modelling, exploring, and exploiting functions. Journal of mathematical psychology, 85, 2018

    2018

  27. [35]

    A. M. Schweidtmann, D. Bongartz, D. Grothe, T. Kerkenhoff, X. Lin, J. Najman, and A. Mitsos. Deterministic global optimization with G aussian processes embedded. Mathematical Programming Computation, 13 0 (3): 0 553--581, 2021

    2021

  28. [36]

    Shervashidze, P

    N. Shervashidze, P. Schweitzer, E. J. Van Leeuwen, K. Mehlhorn, and K. M. Borgwardt. Weisfeiler- l ehman graph kernels. Journal of Machine Learning Research, 12 0 (9), 2011

    2011

  29. [37]

    H. Shi, R. Pi, H. Xu, Z. Li, J. Kwok, and T. Zhang. Bridging the gap between sample-based and one-shot neural architecture search with BONAS . NeurIPS, 2020

    2020

  30. [38]

    Siglidis, G

    G. Siglidis, G. Nikolentzos, S. Limnios, C. Giatsidis, K. Skianis, and M. Vazirgiannis. Gra K el: A graph kernel library in P ython. Journal of Machine Learning Research, 21 0 (54): 0 1--5, 2020

    2020

  31. [39]

    Snoek, O

    J. Snoek, O. Rippel, K. Swersky, R. Kiros, N. Satish, N. Sundaram, M. Patwary, M. Prabhat, and R. Adams. Scalable B ayesian optimization using deep neural networks. In ICML, 2015

    2015

  32. [40]

    J. T. Springenberg, A. Klein, S. Falkner, and F. Hutter. Bayesian optimization with robust B ayesian neural networks. NeurIPS, 2016

    2016

  33. [41]

    Srinivas, A

    N. Srinivas, A. Krause, S. Kakade, and M. Seeger. G aussian process optimization in the bandit setting: No regret and experimental design. In ICML, 2010

    2010

  34. [42]

    Thebelt, J

    A. Thebelt, J. Kronqvist, M. Mistry, R. M. Lee, N. Sudermann-Merx, and R. Misener. ENTMOOT : A framework for optimization over ensemble tree models. Computers & Chemical Engineering, 151: 0 107343, 2021

    2021

  35. [43]

    Thebelt, J

    A. Thebelt, J. Wiebe, J. Kronqvist, C. Tsay, and R. Misener. Maximizing information from chemical engineering data sets: Applications to machine learning. Chemical Engineering Science, 252: 0 117469, 2022

    2022

  36. [44]

    C. Tsay, J. Kronqvist, A. Thebelt, and R. Misener. Partition-based formulations for mixed-integer optimization of trained ReLU neural networks. In NeurIPS, 2021

    2021

  37. [45]

    S. V. N. Vishwanathan, N. N. Schraudolph, R. Kondor, and K. M. Borgwardt. Graph kernels. The Journal of Machine Learning Research, 11: 0 1201--1242, 2010

    2010

  38. [46]

    X. Wan, H. Kenlay, B. Ru, A. Blaas, M. Osborne, and X. Dong. Adversarial attacks on graph classifiers via B ayesian optimisation. In NeurIPS, 2021

    2021

  39. [47]

    X. Wan, P. Osselin, H. Kenlay, B. Ru, M. A. Osborne, and X. Dong. Bayesian optimisation of functions on graphs. NeurIPS, 2023

    2023

  40. [48]

    K. Wang, L. Lozano, C. Cardonha, and D. Bergman. Optimizing over an ensemble of trained neural networks. INFORMS Journal on Computing, 2023

    2023

  41. [49]

    W. Wen, H. Liu, Y. Chen, H. Li, G. Bender, and P.-J. Kindermans. Neural predictor for neural architecture search. In ECCV, 2020

    2020

  42. [50]

    White, W

    C. White, W. Neiswanger, S. Nolen, and Y. Savani. A study on encodings for neural architecture search. In NeurIPS, 2020

    2020

  43. [51]

    White, W

    C. White, W. Neiswanger, and Y. Savani. BANANAS : Bayesian optimization with neural architectures for neural architecture search. In AAAI, 2021 a

    2021

  44. [52]

    White, S

    C. White, S. Nolen, and Y. Savani. Exploring the loss landscape in neural architecture search. In UAI, 2021 b

    2021

  45. [53]

    White, M

    C. White, M. Safari, R. Sukthanker, B. Ru, T. Elsken, A. Zela, D. Dey, and F. Hutter. Neural architecture search: insights from 1000 papers. arXiv preprint arXiv:2301.08727, 2023

    2023 arXiv

  46. [54]

    B. Wu, X. Dai, P. Zhang, Y. Wang, F. Sun, Y. Wu, Y. Tian, P. Vajda, Y. Jia, and K. Keutzer. FBNet : Hardware-aware efficient convnet design via differentiable neural architecture search. In CVPR, 2019

    2019

  47. [55]

    Y. Xie, S. Zhang, J. Paulson, and C. Tsay. Global optimization of G aussian process acquisition functions using a piecewise-linear kernel approximation. arXiv preprint arXiv:2410.16893, 2024

    2024 arXiv

  48. [56]

    Y. Xie, S. Zhang, J. Qing, R. Misener, and C. Tsay. BoGrape : Bayesian optimization over graphs with shortest-path encoded. arXiv preprint arXiv:2503.05642, 2025

    2025 arXiv

  49. [57]

    C. Ying, A. Klein, E. Christiansen, E. Real, K. Murphy, and F. Hutter. NAS-Bench-101 : Towards reproducible neural architecture search. In ICML, 2019

    2019

  50. [58]

    Zhang, J

    S. Zhang, J. S. Campos, C. Feldmann, D. Walz, F. Sandfort, M. Mathea, C. Tsay, and R. Misener. Optimizing over trained GNNs via symmetry breaking. In NeurIPS, 2023

    2023

  51. [59]

    Zhang, J

    S. Zhang, J. S. Campos, C. Feldmann, F. Sandfort, M. Mathea, and R. Misener. Augmenting optimization-based molecular design with graph neural networks. Computers & Chemical Engineering, 186: 0 108684, 2024

    2024

  52. [60]

    Zoph and Q

    B. Zoph and Q. Le. Neural architecture search with reinforcement learning. In ICLR, 2017

    2017

Pith tools

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