Pith. sign in

REVIEW 3 major objections 5 minor 45 references

Runtime Analysis of Evolutionary NAS for Multiclass Classification

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 The paper proves that (1+1)-ENAS with either one-bit or bit-wise outer-level mutation finds the optimal architecture for a new multiclass benchmark MCC in expected time $O(rM\ln(rM))$, with a lower bound $\Omega(rM\ln M)$ when initial…

desk verdict First multiclass ENAS runtime bounds with mostly standard proofs, but the general-M behavior of the fixed output layer is asserted rather than proven, so the bounds may be for a synthetic fitness function rather than the stated accuracy. read the letter →

arxiv 2506.06019 v1 pith:3FPCJXZS submitted 2025-06-06 cs.NE

Zeqiong Lv , Chao Qian , Yun Liu , Jiahao Fan , Yanan Sun This is my paper
classification cs.NE MSC 68W5068Q2568T20
keywords evolutionaryneuralarchitecturesearchruntimeanalysismulticlassclassificationone-bitmutationbit-wisefitnessfunctionmultiplicativedriftbenchmarkproblem
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper gives the first runtime analysis of evolutionary neural architecture search (ENAS) on a multiclass classification problem. It defines a benchmark called MCC, in which the unit disk is split into $2rM$ equal sectors and each of $M$ classes occupies $r$ segments, $r$ sectors, and $r$ triangles; a neural architecture's fitness is the area of correctly classified regions, computed exactly from the counts of its A-, B-, and C-type blocks. The main theorems show that the (1+1)-ENAS algorithm reaches the optimal architecture in expected $O(rM\ln(rM))$ fitness evaluations when the outer-level mutation is one-bit or bit-wise, and needs $\Omega(rM\ln M)$ generations when initial block counts are sampled from $\{1,\dots,r\}$. Those bounds are the first of their kind for multiclass ENAS, and they imply that the simpler one-bit operator is asymptotically as efficient as bit-wise mutation on this benchmark. Experiments on $M$ up to 24 and $r$ up to 10 support the theoretical results.

What carries the argument

The argument is carried by the MCC fitness function $$F(x) = \frac{\mathrm{Ar}_{\mathrm{tri}}(I_x+2r) + \mathrm{Ar}_{\mathrm{seg}}(J_x+2r-\epsilon_x)}{\pi},$$ where $\mathrm{Ar}_{\mathrm{tri}}$ and $\mathrm{Ar}_{\mathrm{seg}}$ are the areas of one triangle and one segment of the $2rM$-sector partition. For cell $m$, $I_x^m = \min\{n_B^m+n_C^m, 2r\}$ counts how many triangle and sector regions can be covered, $J_x^m = \min\{n_B^m,r\}+\min\{n_A^m, r+\max\{r-n_B^m,0\}\}$ counts segment coverage, and $\epsilon_x$ is the number of class-$M$ segments misclassified by the last cell. The solution encoding is two-level: $x=\{(n_A^1,n_B^1,n_C^1),\dots,(n_A^{M-1},n_B^{M-1},n_C^{M-1})\}$, and mutation first selects a cell (outer level: one-bit or bit-wise) then adds, deletes, or modifies blocks inside it (inner level: local or global). The proof splits the search space into fitness levels by $I_x$, then by $J_x-\epsilon_x$, and applies multiplicative drift (or a fitness-level argument) using a bit-string $\mathbf{o}$ that marks cells already contributing their full share; the inner-level mutation probabilities give constant factors $2/9$ and $1/9$ for increasing these counters.

What would settle it

Enumerate all $2^{M-1}$ binary cell-output vectors for some $M\ge 4$; compute $h_1=c_1+0.4c_2$, $h_m=0.5c_{m-1}+c_m+0.4c_{m+1}$ with the appropriate boundary cases, and $h_M=0.5c_{M-1}+0.1$, then check whether the softmax argmax is the class the construction assigns to that output pattern. A single mismatch would break the fitness formula in Lemma 3.2 and hence the runtime bounds.

Watch

Extended reading notes

Core claim

The paper's central claim is that, on the MCC benchmark, the expected runtime of (1+1)-ENAS is asymptotically the same for one-bit and bit-wise outer-level mutation. Theorems 4.1 and 4.3 prove the upper bound $E[T] = O(rM\ln(rM))$ for both operators, and Theorems 4.2 and 4.4 prove the lower bound $E[T] = \Omega(rM\ln M)$ when the initialization bound $s$ equals $r$. The proof reduces fitness to two aggregate counters: $I_x$ counts the triangular and sector regions the cells can cover, and $J_x - \epsilon_x$ counts the segment regions correctly classified; increasing either counter by one happens with constant probability per useful mutation. The paper presents this as the first theoretical analysis of ENAS for multiclass problems and argues that the benchmark behaves like a OneMax analogue for architecture search.

Load-bearing premise

The runtime bounds rest on the assumption that the fixed hidden-layer weights ($w_{i,i}=1$, $w_{i,i-1}=0.5$, $w_{i,i+1}=0.4$, $b_M=0.1$) make the softmax layer choose the intended class for every possible pattern of binary cell outputs, for every number of classes $M$, while the paper illustrates this only for three-class examples.

Editorial extensions

If this is right

  • On MCC, choosing one-bit outer-level mutation instead of bit-wise mutation does not change the asymptotic expected runtime: both give $O(rM\ln(rM))$ generations to reach the optimum.
  • With the standard initialization $s=r$, neither operator can find the optimum in fewer than $\Omega(rM\ln M)$ expected generations, so the two are asymptotically equivalent on this benchmark.
  • The MCC benchmark provides a mathematically concrete fitness function for multiclass ENAS, giving later theoretical work a testbed that earlier binary-only benchmarks could not supply.
  • The paper's experiments report that local and global inner-level mutations behave similarly, and that one-bit mutation remains comparable to bit-wise in population-based and crossover-based ENAS variants, although those variants are not covered by the theorems.

Reading between the lines

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

  • Our inference: since both outer-level mutations share the same drift lower bounds, the bottleneck on MCC is the cell-level accumulation of B- and C-type blocks rather than the outer-level mutation distribution; recording the empirical distribution of per-cell deficits $2r - (n_B^m+n_C^m)$ over time would test this.
  • Our inference: because $F(x)$ is a sum of per-cell monotone region counts, MCC is a natural first candidate for a formal runtime analysis of population-based ENAS and crossover, which the paper currently supports only empirically.
  • Our inference: the fixed hidden-layer aggregation is what makes the closed-form fitness function possible; a natural stress test is to let the aggregation weights evolve, which would break the formula and require a new analytical handle.
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 MCC, a synthetic multiclass classification benchmark on the unit circle, together with a two-level cell/block search space, and analyzes the expected runtime of (1+1)-ENAS with one-bit and bit-wise outer-level mutations (each combined with local or global inner-level mutation). The main results, Theorems 4.1-4.4, claim upper bounds O(rM ln(rM)) and lower bounds Omega(rM ln M) for both mutation operators, and the authors conclude that the simpler one-bit mutation is essentially as efficient as bit-wise mutation on this benchmark. The paper also includes experiments and extended experiments with population and crossover variants. The framework is a natural extension of the authors' prior binary-classification runtime analyses, and the intended contribution is the first runtime analysis of evolutionary neural architecture search for a multiclass classification problem.

Significance. If the results are correct, this would be a useful first step in the runtime analysis of ENAS for multiclass classification. The proposed MCC benchmark is explicit and the fitness function is mathematically formulated, which allows standard drift and fitness-level arguments to be applied. The comparison of one-bit versus bit-wise mutation is a practically motivated question, and the conclusion that the simpler mutation is competitive is valuable. The paper is also careful to include both upper and lower bounds and to support the theory with experiments. However, the significance currently rests on several unproven or incorrect technical steps: the general-M behavior of the fixed output layer, an inconsistency in the definition of the bit-string o used in the drift bounds, and a flawed probability estimate in the lower-bound proofs. These points need to be resolved before the claimed runtime bounds can be accepted.

major comments (3)
  1. [Section 2.1 and Lemma 3.2] The claim that the fixed hidden layer with weights w_{i,i}=1, w_{i,i-1}=0.5, w_{i,i+1}=0.4 and bias b_M=0.1 implements the correct softmax aggregation for every M is asserted but never proved. Appendix B only verifies the M=3 case. Lemma 3.2's fitness formula, Eq. (6), is derived from Eq. (1), which presupposes that any point in class m<M is correctly classified exactly when the m-th cell fires, and that points in class M are correct exactly when no cell fires. This presupposition requires a general-M proof that the set of firing cells is always empty, a singleton, or a consecutive pair {m-1,m}, and that the softmax argmax is the rightmost firing cell (or class M when empty). For M>=4, a pattern such as cells 1 and 3 firing gives h_1=h_3=1, a tie, so the reachability constraint is essential. Without such a lemma, Eq. (6) is not shown to be the classification accuracy on MCC, and the runtime theorems are statements about a surrogate fitness rather than the stated problem.
  2. [Section 4.1, definition of o and Eq. (8)] The upper bound |o|1 <= I_x/(2r) in Eq. (8) is inconsistent with the definition of o_{M-1}. The text sets o_{M-1}=1 already when I_x<N and I_{M-1}^x=r, but for a solution in which only the last cell has I=r and all other cells have I=0, we have I_x=r, |o|1=1, and r/(2r)=1/2, violating the claimed inequality. Since the drift calculation in Eq. (9) uses |o|0 >= M-1 - I_x/(2r), and the bit-wise proof in Theorem 4.3 uses the same bound, this invalidates the phase-1 drift arguments as written. The threshold for the last cell should presumably be 2r, or a different bound that accounts for the special definition of o_{M-1} must be supplied.
  3. [Appendix C.1 and C.3 (Theorems 4.2 and 4.4)] The lower-bound proof contains a false probability estimate. For M=3 and r=2, the middle expression in the chain for P(G) evaluates to 1-(1-0.5^{2 ln 2}) = 1-(1-0.5^{1.386}) = 0.382, which is less than the claimed 1-e^{-1} = 0.632. Asymptotically, for fixed r, (1-1/(M-1))^{r(M-2)ln(M-1)} is approximately (M-1)^{-r}, so the expression claimed to be at least 1-e^{-1} is approximately (M-1)^{1-r}/r, which tends to 0 for r>=2. The event that a 0-bit cell is never increased has probability about e^{-gamma r ln(M-1)} per cell, so the expected number of such cells is ((M-1)/r)e^{-gamma r ln(M-1)}, which tends to 0 for fixed r>=5 rather than being large. The proof needs to be reworked, for example by considering cells that receive fewer than the required number of successful increments rather than zero increments. Since Theorems 4.2 and 4.4 both rely on this argument, the stated Omega(rM ln M) lower bounds are not established as written.
minor comments (5)
  1. [Section 3.1] There is a typo 'segmenets' in the definition of MCC, and the terminology for 'r segments, r sectors, and r triangles' should be aligned with the later use of '2r triangles' in the fitness derivation.
  2. [Section 4.1] The constants gamma_{i,i+1} and eta_{z,z+1} are imported from Lv et al. (2024a, 2024b) without stating the corresponding lemmas or re-deriving them. Please include the lemmas or a short derivation in the appendix so that the paper is self-contained.
  3. [Appendix B, Table 1] The caption of Table 1 uses '!' and '%' to denote correct and incorrect classifications but does not explain these symbols; please spell them out.
  4. [Figures 3 and Appendix D] The axis labels and legends in Figure 3 and the figures in Appendix D contain unicode replacement characters and are not legible; the figure files need to be regenerated.
  5. [Abstract] The phrase 'we step for the runtime analysis' should read 'we take a step towards the runtime analysis'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the runtime analysis optimizes an explicitly defined synthetic fitness function, with self-cited mutation-probability lemmas serving as independent parameter-free support; the general-M hidden-layer classification gap is a correctness risk, not a circular reduction.

full rationale

The paper's derivation is a normal benchmark-and-analysis chain rather than a circular one. MCC is defined geometrically (Definition 3.1), and Eq. (1) defines accuracy by coverage sets; Lemma 3.2 then derives the closed-form fitness F(x) from that definition. The runtime theorems optimize this F through the Ix/Jx partition, so the bounds are genuine mathematical consequences of the model, not re-statements of fitted parameters. The proof of Lemma 3.2 depends on the unproven assertion (Section 2.1) that the fixed hidden-layer weights and bias make the softmax layer select the correct class for every reachable cell-output pattern for general M; only M=3 examples are shown (Appendix B). This is a load-bearing missing proof and a correctness risk, but it is not circular: F is not defined as the drift objective, and the runtime analysis would still be a valid analysis of F even if F failed to equal the network's true accuracy. The upper and lower bound proofs invoke probability bounds gamma >= 2/9 and eta >= 1/9 from Lv et al. 2024a/b, which are self-citations by overlapping authors; under the stated rules these count as independent support because they are parameter-free, single-cell mutation lemmas whose assumptions do not include the target runtime bound. The benchmark is author-designed and OneMax-like, but that is standard practice for runtime analysis; the one-bit vs bit-wise comparison follows from the outer-level selection probabilities, not from a fitted or definitionally forced equivalence. No step meets the quoted-reduction bar for circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The runtime bounds rest on the hand-built MCC geometry, the hidden-layer parameter choices, and two probability constants imported from the authors' earlier papers. No parameters are fitted to data, but several design choices are made ad hoc to make the fitness landscape behave like a product of ONEMAX-like functions.

free parameters (3)
  • Hidden layer weights and bias = w_{i,i-1}=0.5, w_{i,i+1}=0.4, b_M=0.1
    Hand-chosen in Section 2.1 so that the softmax layer maps cell outputs to the intended class; no proof is given that these exact values work for all M and all cell-output combinations.
  • Block geometry parameters = biases cos(pi/n); angle differences pi +/- 2pi/n, pi/2 +/- pi/n, 3pi/2 - pi/n
    Set in Appendix A.1 so that A-type, B-type, and C-type blocks cover exactly the intended segment, sector, and triangle regions of MCC; these are design choices with no independent justification.
  • Initial block-count upper bound s = s = r in Theorems 4.2 and 4.4
    The lower-bound theorems assume initial block counts are sampled from U[1, s] with s = r; the stated lower bounds are conditional on this choice.
assumptions (5)
  • standard math Multiplicative drift theorem and fitness-level technique give expected runtime bounds from per-step drift or progress probabilities.
    Used in Theorems 4.1 and 4.3; these are standard tools in runtime analysis of evolutionary algorithms.
  • domain assumption The optimal parameters of every evolved architecture can be achieved during fitness evaluation.
    Stated in Section 3.2; this idealization isolates architecture search from weight optimization and is common in the ENAS theory literature.
  • ad hoc to paper The hand-set hidden layer weights and biases implement the claimed class aggregation for all M and all cell-output combinations.
    Assumed in Section 2.1 and used in the proof of Lemma 3.2; only example-based evidence is provided for M=3.
  • ad hoc to paper The probability constants gamma_{i,i+1} >= 2/9 and eta_{z,z+1} >= 1/9 from Lv et al. 2024a and 2024b transfer to the two-level search space.
    Invoked in the proof of Theorem 4.1; the present paper does not re-derive these constants for the new cell-based encoding.
  • domain assumption MCC's geometry, including the unit circle, sectors, triangles, segments, and interleaving, is exactly as defined in Definition 3.1.
    Defines the benchmark; all subsequent fitness and runtime statements are relative to this specific problem definition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Runtime Analysis of Evolutionary NAS for Multiclass Classification." pith.science (2026). https://pith.science/paper/3FPCJXZS

@misc{pith2026250606019,
  author       = {Pith},
  title        = {Pith review of: Runtime Analysis of Evolutionary NAS for Multiclass Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3FPCJXZS}},
  note         = {Machine review of arXiv:2506.06019}
}
abstract

Evolutionary neural architecture search (ENAS) is a key part of evolutionary machine learning, which commonly utilizes evolutionary algorithms (EAs) to automatically design high-performing deep neural architectures. During past years, various ENAS methods have been proposed with exceptional performance. However, the theory research of ENAS is still in the infant. In this work, we step for the runtime analysis, which is an essential theory aspect of EAs, of ENAS upon multiclass classification problems. Specifically, we first propose a benchmark to lay the groundwork for the analysis. Furthermore, we design a two-level search space, making it suitable for multiclass classification problems and consistent with the common settings of ENAS. Based on both designs, we consider (1+1)-ENAS algorithms with one-bit and bit-wise mutations, and analyze their upper and lower bounds on the expected runtime. We prove that the algorithm using both mutations can find the optimum with the expected runtime upper bound of $O(rM\ln{rM})$ and lower bound of $\Omega(rM\ln{M})$. This suggests that a simple one-bit mutation may be greatly considered, given that most state-of-the-art ENAS methods are laboriously designed with the bit-wise mutation. Empirical studies also support our theoretical proof.

Figures

Figures reproduced from arXiv: 2506.06019 by the authors.

Figure 1
Figure 1. The neural architecture skeleton, including a set of cells where each of them consists of a set of blocks. There are three types of blocks: A-type, B-type, and C-type. Each block receives the input x and outputs a binary bit (1 or 0). Thus, M − 1 cells will output a binary set for each neuron in {N1, N2, . . . , NM}. Each neuron Nm computes a value hm, which is then normalized by the Softmax layer to yield the proba… view at source ↗
Figure 2
Figure 2. Illustration of neural architecture solving MCC. (a) MCC with M = 3 and r = 2, depicting a 3-class classification problem. The points labeled as class 1, 2, and 3 are colored off-white, yellow, and khaki, respectively. Each class (e.g., class 1) has two segments (e.g., S 9 seg ∪ S 12 seg), two sectors (e.g., S 1 sec ∪ S 4 sec), and two triangles (e.g., S 7 tri ∪ S 10 tri). (b) Decision regions produced by neural arc… view at source ↗
Figure 3
Figure 3. Average number of generations of the (1+1)-ENAS algo￾rithm with different mutations for solving the MCC problem. The legend identifies mutation types: the term before the symbol “+” denotes the outer-level mutation, and the term after the symbol represents the inner-level mutation. The above experiment results also offer practical guidance for mutation design in ENAS. In particular, many ENAS algorithms adopt bit-wi… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Visualization of the parameter settings for A-type, B-type, and C-type blocks, along with the corresponding decision regions generated by each block type, upon the MCC problem (M = 3, r = 2, n = 12). The red arrow represents the unit normal vector of a hyperplane (line…
Figure 5
Figure 5. Figure 5: Parameters (weight and bias) of the neurons in {N1, N2, . . . , NM}. A.2. Decision Regions Based on the above parameter settings, it is clear that when the dataset is on the unit circle, three types of blocks in the neural architecture generate decision regions with di…
Figure 6
Figure 6. Figure 6: Average number of generations of the ENAS algorithms with population and crossover for solving the MCC problem (r = 10, varying M). The ENAS algorithms include (λ+λ)-ENAS algorithm (with mutation only), (λ+λ)-ENAS algorithm with one-point crossover, and (λ+λ)-ENAS algo…
Figure 7
Figure 7. Figure 7: Average number of generations of the ENAS algorithms with population and crossover for solving the MCC problem (M = 20, varying r). The ENAS algorithms include (λ+λ)-ENAS algorithm (with mutation only), (λ+λ)-ENAS algorithm with one-point crossover, and (λ+λ)-ENAS algo…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

45 extracted references · 42 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Survey on multiclass classification methods

    Aly, M. Survey on multiclass classification methods. Neural Netw, 19 0 (1-9): 0 2, 2005

    work page 2005

  3. [3]

    and Doerr, B

    Auger, A. and Doerr, B. Theory of Randomized Search Heuristics - Foundations and Recent Developments. World Scientific, Singapore, 2011

    work page 2011

  4. [4]

    and Tsitsiklis, J

    Bertsimas, D. and Tsitsiklis, J. N. Introduction to Linear Optimization. Athena Scientific, 1997

    work page 1997

  5. [5]

    and Goldberg, L

    Doerr, B. and Goldberg, L. A. Adaptive drift analysis. Algorithmica, 65: 0 224--250, 2013

    work page 2013

  6. [6]

    and Neumann, F

    Doerr, B. and Neumann, F. Theory of Evolutionary Computation: Recent Developments in Discrete Optimization. Springer, Cham, Switzerland, 2020

    work page 2020

  7. [7]

    Comparing global and local mutations on bit strings

    Doerr, B., Jansen, T., and Klein, C. Comparing global and local mutations on bit strings. In Proceedings of the 10th Annual Conference on Genetic and Evolutionary Computation (GECCO'08), pp.\ 929--936, Atlanta, GA, 2008

    work page 2008

  8. [8]

    Multiplicative drift analysis

    Doerr, B., Johannsen, D., and Winzen, C. Multiplicative drift analysis. Algorithmica, 64 0 (4): 0 673–697, 2012. ISSN 0178-4617. doi:10.1007/s00453-012-9622-x. URL https://doi.org/10.1007/s00453-012-9622-x

    doi:10.1007/s00453-012-9622-x 2012

Show all 45 references
  1. [9]

    On the analysis of the (1+1) evolutionary algorithm

    Droste, S., Jansen, T., and Wegener, I. On the analysis of the (1+1) evolutionary algorithm. Theoretical Computer Science, 276 0 (1-2): 0 51--81, 2002

    2002

  2. [10]

    Computational complexity analysis of simple genetic programming on two problems modeling isolated program semantics

    Durrett, G., Neumann, F., and O'Reilly, U.-M. Computational complexity analysis of simple genetic programming on two problems modeling isolated program semantics. In Proceedings of the 11th Workshop on Foundations of Genetic Algorithms (FOGA'11), pp.\ 69--80, Schwarzenberg, Au...

    2011

  3. [11]

    H., and Hutter, F

    Elsken, T., Metzen, J. H., and Hutter, F. Neural architecture search: A survey. Journal of Machine Learning Research, 20 0 (1): 0 1997--–2017, 2019

    1997

  4. [12]

    L., and Witt, C

    Fischer, P., Larsen, E. L., and Witt, C. First steps towards a runtime analysis of neuroevolution. In Proceedings of the 17th International Workshop on Foundations of Genetic Algorithms (FOGA'23), pp.\ 61--72, Potsdam, Germany, 2023

    2023

  5. [13]

    A., and Witt, C

    Fischer, P., Warwicker, J. A., and Witt, C. A runtime analysis of bias-invariant neuroevolution and dynamic fitness evaluation. In Proceedings of the 26th Annual Conference on Genetic and Evolutionary Computation (GECCO'24), pp.\ 1560--1568, Melbourne, VIC, Australia, 2024

    2024

  6. [14]

    Additive logistic regression: A statistical view of boosting (with discussion and a rejoinder by the authors)

    Friedman, J., Hastie, T., and Tibshirani, R. Additive logistic regression: A statistical view of boosting (with discussion and a rejoinder by the authors). The Annals of Statistics, 28 0 (2): 0 337--407, 2000

    2000

  7. [15]

    Gibson, G. J. and Cowan, C. F. On the decision regions of multilayer perceptrons. Proceedings of the IEEE, 78 0 (10): 0 1590--1594, 2002

    2002

  8. [16]

    Metrics for multi-class classification: An overview

    Grandini, M., Bagli, E., and Visani, G. Metrics for multi-class classification: An overview. arXiv preprint arXiv:2008.05756, 2020

    2008 arXiv

  9. [17]

    Automl: A survey of the state-of-the-art

    He, X., Zhao, K., and Chu, X. Automl: A survey of the state-of-the-art. Knowledge-Based Systems, 212: 0 106622, 2021

    2021

  10. [18]

    K., Neumann, F., and Oliveto, P

    Kratsch, S., Lehre, P. K., Neumann, F., and Oliveto, P. S. Fixed parameter evolutionary algorithms and maximum leaf spanning trees: A matter of mutation. In Proceedings of the 11th International Conference on Parallel Problem Solving from Nature (PPSN'10), pp.\ 204--213, Krak\...

    2010

  11. [19]

    Hierarchical representations for efficient architecture search

    Liu, H., Simonyan, K., Vinyals, O., Fernando, C., and Kavukcuoglu, K. Hierarchical representations for efficient architecture search. In Proceedings of the 6th International Conference of Learning Representation (ICLR'18), Vancouver, BC, Canada, 2018

    2018

  12. [20]

    G., and Tan, K

    Liu, Y., Sun, Y., Xue, B., Zhang, M., Yen, G. G., and Tan, K. C. A survey on evolutionary neural architecture search. IEEE Transactions on Neural Networks and Learning Systems, 34 0 (2): 0 550--570, 2023

    2023

  13. [21]

    Runtime analysis of population-based evolutionary neural architecture search for a binary classification problem

    Lv, Z., Bian, C., Qian, C., and Sun, Y. Runtime analysis of population-based evolutionary neural architecture search for a binary classification problem. In Proceedings of the 26th Annual Conference on Genetic and Evolutionary Computation (GECCO'24), pp.\ 358--366, Melbourne, ...

    2024

  14. [22]

    A first step towards runtime analysis of evolutionary neural architecture search

    Lv, Z., Qian, C., and Sun, Y. A first step towards runtime analysis of evolutionary neural architecture search. arXiv:2401.11712, 2024 b . URL https://arxiv.org/abs/2401.11712

    2024 arXiv

  15. [23]

    Evolving deep neural networks

    Miikkulainen, R., Liang, J., Meyerson, E., Rawal, A., Fink, D., Francon, O., Raju, B., Shahrzad, H., Navruzyan, A., Duffy, N., et al. Evolving deep neural networks. In Artificial Intelligence in the Age of Neural Networks and Brain Computing (Second Edition), pp.\ 269--287. Ac...

    2024

  16. [24]

    and Witt, C

    Neumann, F. and Witt, C. Bioinspired Computation in Combinatorial Optimization - Algorithms and Their Computational Complexity. Springer, Berlin, Germany, 2010

    2010

  17. [25]

    and Witt, C

    Neumann, F. and Witt, C. On the runtime of randomized local search and simple evolutionary algorithms for dynamic makespan scheduling. In Proceedings of the 24th International Joint Conference on Artificial Intelligence (IJCAI'15), pp.\ 3742--3748, Buenos Aires, Argentina, 2015

    2015

  18. [26]

    C., and Hein, M

    Nguyen, Q., Mukkamala, M. C., and Hein, M. Neural networks should be wide enough to learn disconnected decision regions. In Proceedings of the 35th International Conference on Machine Learning (ICML'18), pp.\ 3737--3746, Stockholm, Sweden, 2018. PMLR

    2018

  19. [27]

    Brain-inspired multi-scale evolutionary neural architecture search for deep spiking neural networks

    Pan, W., Zhao, F., Shen, G., Han, B., and Zeng, Y. Brain-inspired multi-scale evolutionary neural architecture search for deep spiking neural networks. IEEE Transactions on Evolutionary Computation, 2024

    2024

  20. [28]

    Variable solution structure can be helpful in evolutionary optimization

    Qian, C., Yu, Y., and Zhou, Z.-H. Variable solution structure can be helpful in evolutionary optimization. Science China: Information Sciences, 58 0 (11): 0 1--17, 2015

    2015

  21. [29]

    Multi-objective evolutionary algorithms are generally good: Maximizing monotone submodular functions over sequences

    Qian, C., Liu, D.-X., Feng, C., and Tang, K. Multi-objective evolutionary algorithms are generally good: Maximizing monotone submodular functions over sequences. Theoretical Computer Science, 943: 0 241--266, 2023

    2023

  22. [30]

    L., Tan, J., Le, Q

    Real, E., Moore, S., Selle, A., Saxena, S., Suematsu, Y. L., Tan, J., Le, Q. V., and Kurakin, A. Large-scale evolution of image classifiers. In Proceedings of the 34th International Conference on Machine Learning (ICML'17), pp.\ 2902--2911, Sydney, Australia, 2017. PMLR

    2017

  23. [31]

    Real, E., Aggarwal, A., Huang, Y., and Le, Q. V. Regularized evolution for image classifier architecture search. In Proceedings of the 33th AAAI Conference on Artificial Intelligence (AAAI'19), pp.\ 4780--4789, Honolulu, Hawaii, USA, 2019

    2019

  24. [32]

    Activation functions in neural networks

    Sharma, S., Sharma, S., and Athaiya, A. Activation functions in neural networks. International Journal of Engineering Applied Sciences and Technology, 4 0 (12): 0 310--316, 2017

    2017

  25. [33]

    So, D., Ma \'n ke, W., Liu, H., Dai, Z., Shazeer, N., and Le, Q. V. Searching for efficient transformers for language modeling. In Advances in Neural Information Processing Systems 34 (NeurIPS'21), pp.\ 6010--6022, virtual, 2021

    2021

  26. [34]

    A new method for lower bounds on the running time of evolutionary algorithms

    Sudholt, D. A new method for lower bounds on the running time of evolutionary algorithms. IEEE Transactions on Evolutionary Computation, 17 0 (3): 0 418--435, 2013

    2013

  27. [35]

    Sun, Y., Xue, B., Zhang, M., and Yen, G. G. Completely automated CNN architecture design based on blocks. IEEE Transactions on Neural Networks and Learning Systems, 31 0 (4): 0 1242--1254, 2019 a

    2019

  28. [36]

    Sun, Y., Xue, B., Zhang, M., and Yen, G. G. Evolving deep convolutional neural networks for image classification. IEEE Transactions on Evolutionary Computation, 24 0 (2): 0 394--407, 2019 b

    2019

  29. [37]

    \"U nal, H. T. and Ba s c ift c i, F. Evolutionary design of neural network architectures: a review of three decades of research. Artificial Intelligence Review, 55 0 (3): 0 1723--1802, 2022

    2022

  30. [38]

    Methods for the analysis of evolutionary algorithms on pseudo- Boolean functions

    Wegener, I. Methods for the analysis of evolutionary algorithms on pseudo- Boolean functions. In Evolutionary Optimization, pp.\ 349--369. Springer, 2003

    2003

  31. [39]

    Tight bounds on the optimization time of a randomized search heuristic on linear functions

    Witt, C. Tight bounds on the optimization time of a randomized search heuristic on linear functions. Combinatorics, Probability and Computing, 22: 0 294--318, 2013

    2013

  32. [40]

    and Yuille, A

    Xie, L. and Yuille, A. Genetic CNN . In Proceedings of the IEEE International Conference on Computer Vision (ICCV'17), pp.\ 1379--1388, Venice, Italy, 2017

    2017

  33. [41]

    Neural architecture search based on bipartite graphs for text classification

    Yan, X., Huang, H., Jin, Y., Wang, Z., and Hao, Z. Neural architecture search based on bipartite graphs for text classification. IEEE Transactions on Neural Networks and Learning Systems, 2024

    2024

  34. [42]

    Evolutionary neural architecture search for transformer in knowledge tracing

    Yang, S., Yu, X., Tian, Y., Yan, X., Ma, H., and Zhang, X. Evolutionary neural architecture search for transformer in knowledge tracing. In Advances in Neural Information Processing Systems 36 (NeurIPS'23), pp.\ 19520--19539, New Orleans, LA, USA, 2023

    2023

  35. [43]

    Practical block-wise neural network architecture generation

    Zhong, Z., Yan, J., Wu, W., Shao, J., and Liu, C.-L. Practical block-wise neural network architecture generation. In Proceedings of the 35th IEEE Conference on Computer Vision and Pattern Recognition (CVPR'18), pp.\ 2423--2432, 2018

    2018

  36. [44]

    Evolutionary Learning: Advances in Theories and Algorithms

    Zhou, Z.-H., Yu, Y., and Qian, C. Evolutionary Learning: Advances in Theories and Algorithms. Springer, Singapore, 2019

    2019

  37. [45]

    Zoph, B., Vasudevan, V., Shlens, J., and Le, Q. V. Learning transferable architectures for scalable image recognition. In Proceedings of the 35th IEEE Conference on Computer Vision and Pattern Recognition (CVPR'18), pp.\ 8697--8710, Salt Lake City, UT, USA, 2018

    2018

Pith tools

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