Lamarckian Evolution and the Baldwin Effect in Evolutionary Neural Networks

A. Prieto; G. Romero; J.G. Castellano; J.J. Merelo; M.G. Arenas; P.A. Castillo; V. Rivas

REVIEW

Lamarckian Evolution and the Baldwin Effect in Evolutionary Neural Networks

P.A. Castillo, M.G. Arenas, J.G. Castellano, J.J. Merelo, A. Prieto, V. Rivas, G. Romero

Neural and Evolutionary Computing · Submitted 1 Mar 2006 · 35 references

Pith's verdict: Honest but statistically weak comparison of Lamarckian and Baldwin learning in G-Prop; the trade-off is plausible but the accuracy advantage is within noise.

T0 · AI review

How this review was made: An AI referee (deepseek-v4-flash) read the paper on August 28, 2026 and answered a fixed list of questions about it. T0 is the starting level every paper gets. The levels describe how deeply Pith has checked the paper, never the authors.

T0 AI review An AI read the paper and answered a fixed list of questions.
T1 Audit Automatic checks that the references and files the paper relies on exist and match.
T2 Compute The paper's code or calculations re-run with the same results.
T3 Formal statement The claims written out in a formal math language a computer can read.
T4 Proof Those claims proved, and the proof checked by computer.

How Pith checks papers →

· 3 major objections · 5 minor · Ask about this paper · The compiled record · Read the full review →

Challenge this review Comment (0) Read on arXiv ↗

Cite ▾

arXiv IDcs/0603004 (version 1)

Pith ID7DQTLTET, Pith's permanent record of this paper and its review

CitationA. Prieto, G. Romero, J.G. Castellano, J.J. Merelo, M.G. Arenas, P.A. Castillo, V. Rivas. "Lamarckian Evolution and the Baldwin Effect in Evolutionary Neural Networks." arXiv:cs/0603004 (2006). https://pith.science/paper/7DQTLTET

Cite this reviewBibTeX for Pith's review of this paper

Claim paper

Why Pith says this

Highlight any text to comment on it.

Editor's note

read the letter

Quick take: this is a small, honest empirical paper about G-Prop, but the headline trade-off (Lamarckian = smaller/faster, Baldwin = more accurate) rests on differences that are within noise. Read it as a pointer, not a proof.

What's actually new: a concrete Baldwin-effect fitness function for G-Prop that ranks individuals by post-training accuracy, then pre-training accuracy, then size. That specific variant isn't in the earlier G-Prop papers, and the paper reports its behavior on two benchmarks (Glass1a and f2). The authors also lay out the comparison with Lamarckian weight-injection clearly. That is a legitimate extension, and the paper is readable.

Where it holds up: the direction of the effects matches prior work (Whitley et al., Gruau and Whitley): Lamarckian finds solutions faster and with smaller networks, Baldwin has a slower, more progressive search. The size differences are large (59 vs 112 parameters on Glass; 18 vs 85 on f2) and likely robust even without significance tests. The conclusion that errors are 'similar' is honestly stated in Section IV.

Soft spots: the abstract says Baldwin 'obtains MLPs with the smallest error rate,' but Tables I and II show 31 ± 2 vs 32 ± 2 on Glass and 0.086 ± 0.004 vs 0.09 ± 0.01 on f2. Those overlap within one standard deviation, and no run counts or significance tests are given. The paper itself says the error is similar, so the accuracy advantage is not established. Same problem for the external comparisons: Prechelt's 33 ± 5 and Grönroos's 32 ± 5 come from protocols that are not shown to match (partition, epochs, runs). The size comparison with Pomares (6 vs 18 parameters) is also apples-to-oranges: a fuzzy system with 4 rules is not the same model class. These are protocol-matching problems, not circular reasoning—the fitness functions are not fitted to the target results.

Who's this for: someone familiar with G-Prop who wants one more data point on learning–evolution couplings. The paper doesn't introduce a new mechanism or theory, so general readers get less.

Recommendation: deserving of peer review? Yes, as a modest empirical contribution, but a referee should ask for run counts, significance tests, and a matched protocol before the central claim can stand. My own verdict would be conditional.

Referee report

Significance

If the central trade-off were well supported, the paper would offer a practical design rule for choosing between Lamarckian and Baldwinian integration of local learning in neuro-evolution, and it would provide a useful comparison on two different problem types. The strengths of the manuscript are its controlled within-framework comparison of the two mechanisms, its use of both a real-world classification benchmark and a function-approximation problem, and the fact that no parameters are fitted to the target results; the authors also report their evolutionary configuration (300 generations, 200 individuals, 200 training epochs). However, the statistical support is currently too weak for the main quantitative claims: the error differences between the two variants are within the reported standard deviations, the number of runs is not stated, and the external comparisons require protocol matching that is not documented.

Major objections (3)

[Section IV, Tables I and II; Sections IV and V] The central claim that the Baldwinian variant obtains the smallest error rate is not supported by the reported statistics. For Glass1a the reported means are 32 ± 2 (Lamarckian) and 31 ± 2 (Baldwinian), and for f2 they are 0.09 ± 0.01 and 0.086 ± 0.004. These intervals overlap, the number of independent runs is not given, and no significance test is provided. The paper itself states in Section IV that the unplotted standard deviations 'show that error achieved is similar' and in Section V that 'errors are not significantly worse.' Consequently, the abstract's claim that the Baldwin effect obtains MLPs with the smallest error rate, and the design rule in Section IV recommending the Baldwinian strategy when accuracy matters, rest on differences that may be within experimental noise. The authors should report run counts and perform appropriate statistical tests, or explicitly weaken the accuracy-based part of the claim.
[Section IV, Tables I and II] The comparisons with Prechelt, Grönroos, and Pomares are not interpretable without protocol-matching evidence. The manuscript does not say how many runs produced its averages, which Proben1 partition(s) of Glass1a were used, how test error was computed, or whether the cited methods used the same training/test split and evaluation criterion. The differences to Prechelt (33 ± 5) and Grönroos (32 ± 5) are comparable to the reported standard deviations, and the G-Prop Lamarckian average (32 ± 2) is numerically tied with Grönroos's average (32 ± 5) on Glass1a. The statement that G-Prop 'outperforms' these methods therefore needs either a common benchmark protocol or a statistical comparison; otherwise only a 'comparable to' claim is supported.
[Section III, Experiments; Figure 1] The experimental set-up is under-specified in ways that are load-bearing for reproducibility. The paper states that Figure 1 shows average results over all the runs and Tables I and II report mean ± standard deviation, but no number of runs is given anywhere. It also does not state the data-partition scheme, such as how the 214 Glass1a instances were split into training, validation, and test sets and whether multiple Proben1 splits were averaged. Add the number of independent runs, the partition details, the random-seed or initialization policy, and the selection criterion used for the entries in Tables I and II. Without these, the size and generation comparisons, for example 59 ± 28 versus 112 ± 62 weights and 52 ± 54 versus 119 ± 55 generations on Glass1a, cannot be fully assessed.

Minor objections (5)

[Section II, Eq. (1)] Equation (1), the NMSE formula, is garbled by the typesetting with radical signs, and the variables N, s_i, o_i, and \bar{s} should be explicitly defined. Please provide a clean equation and define all symbols.
[Figure 1] The y-axis label 'Error (x10^-3)' is inconsistent with the vertical log scale and with the error values reported in the text; clarify whether plotted values are already multiplied by 10^3 and label the axes consistently.
[Section IV] The phrase 'It is evident that G-Prop outperforms other methods' is too strong in view of the ties and protocol-matching issues discussed above; please replace it with a more cautious statement, such as saying that results are comparable or better under the conditions actually documented.
[Section V and references] There are several typographical errors: 'due to to the use' in Section V, 'Baldwin efect' in reference [13], 'TABLA I/II' in the table captions, and Spanish abstract labels ('Resumen', 'Palabras clave') that should be converted to the journal's language. These should be corrected.
[Sections II and III] The algorithm encoding, genetic operators, and their probabilities are only described by reference to previous G-Prop papers; a self-contained description, even if brief, would help readers assess the method. The tables should also state whether the 'Size' values for Prechelt, Grönroos, and Pomares refer to the same quantity as the G-Prop size values.

How Pith checks papers → · Full audit of this paper →

The paper

The paper claims

G-Prop shows Lamarckian evolution makes smaller, faster nets; Baldwin-effect learning makes the most accurate ones.

The reading

G-Prop is an evolutionary trainer for multilayer perceptrons that uses a local-search learning operator inside a genetic search. This paper asks how the way learning is coupled to evolution changes the result: writing the learned weights back into the genome (Lamarckian) versus using the learned performance only to shape fitness (Baldwin effect). On a glass-classification problem and a function-approximation problem, the Lamarckian variant finds low-error networks fastest and with the smallest hidden layers, while the Baldwin variant takes more generations and builds larger networks but achieves the lowest average error overall. Both variants report lower average error than RPROP, an earlier evolutionary hybrid, and a fuzzy-logic approximator. The practical takeaway is a design rule: choose Lamarckian when speed and small networks matter, and Baldwin-effect learning when final accuracy is the priority.

What carries the argument

The load-bearing mechanism is the coupling between Quick-Propagation training and the genetic representation. In the Lamarckian version, the local-search operator trains each individual and then copies the trained weights back into the chromosome, so offspring inherit the learned solution. In the Baldwin version, the same training runs during fitness evaluation, but only the resulting validation-set ability is used for selection; the chromosome keeps the untrained weights. The Baldwin fitness function ranks individuals first by post-training ability, then by pre-training ability as a tie-breaker, and finally by network size. This coupling is what lets the paper isolate the effect of information flow from learning to evolution.

Extended reading notes

Core claim

The central discovery is a concrete trade-off between the two classic mechanisms when embedded in G-Prop. On Glass1a, the Lamarckian approach reaches 32 ± 2 percent test error with networks of about 59 weights in about 52 generations, while the Baldwinian approach reaches 31 ± 2 percent error with about 112 weights in about 119 generations. On the f2 function approximation problem, the Lamarckian variant reaches 0.09 ± 0.01 NMSE with 18 ± 8 weights, and the Baldwinian variant reaches 0.086 ± 0.004 NMSE with 85 ± 27 weights; the published fuzzy-logic result is 0.125 with 6 parameters. The Baldwin strategy is more progressive, keeps searching longer, and avoids the early stagnation seen under the elitist Lamarckian run, which tends to let one dominant individual stop evolution. The paper presents this as evidence that learning which only shapes fitness can match or beat learning that writes back into the genome, at the price of more computation and larger networks.

Load-bearing premise

The comparison with previously published results assumes those runs used the same data partitions, number of runs, training epochs, and error metric as the authors' runs, and the paper does not demonstrate that equivalence.

Editorial extensions

If this is right

A designer who needs a deployable network quickly should use the Lamarckian operator, because it reaches a good error rate in far fewer generations and with a much smaller network.
A designer who can afford longer runs should use the Baldwin-effect fitness function, because it produced the lowest error on both benchmarks in these experiments.
The Lamarckian strategy's early convergence under elitism implies that its reported solution is an early local optimum; problems with many local minima are the natural territory for the Baldwin variant.
The paper's comparison tables place both variants below the reported errors of RPROP, an earlier evolutionary method, and a fuzzy-logic approximator.

Reading between the lines

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

A natural extension is a mixed strategy that first lets Baldwin-effect learning explore for a fixed number of generations and then switches to writing weights back; this could combine the low error of the Baldwin variant with the small networks of the Lamarckian variant.
The large size gap between variants suggests the Baldwin search is exploring a wider range of architectures rather than refining one, a hypothesis that could be tested by measuring topology diversity over generations.
Re-running the comparison under a single shared protocol, with identical data splits and run counts, would turn the reported baseline differences from suggestive into directly interpretable.
If the size difference is robust, it implies that writing learned weights back biases evolution toward overfitting the current solution, while withholding them keeps the population exploring; this could be tested by tracking training versus validation error per generation.

Original abstract

Hybrid neuro-evolutionary algorithms may be inspired on Darwinian or Lamarckian evolu- tion. In the case of Darwinian evolution, the Baldwin effect, that is, the progressive incorporation of learned characteristics to the genotypes, can be observed and leveraged to improve the search. The purpose of this paper is to carry out an exper- imental study into how learning can improve G-Prop genetic search. Two ways of combining learning and genetic search are explored: one exploits the Baldwin effect, while the other uses a Lamarckian strategy. Our experiments show that using a Lamarckian op- erator makes the algorithm find networks with a low error rate, and the smallest size, while using the Bald- win effect obtains MLPs with the smallest error rate, and a larger size, taking longer to reach a solution. Both approaches obtain a lower average error than other BP-based algorithms like RPROP, other evolu- tionary methods and fuzzy logic based methods

The compiled record

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. Numbers in the tables are the authors', checked by code. No sentence in this record was written by a model.

Compiled by classifier-1.13-20260917 via OpenRouter, question bank abstract-p1-20260923, 2026-09-23: 1 request, 5,002 input tokens, $0.0002 reported.

What the abstract itself states, as probabilities: claims first 0.03 · states limitation 0.25 · human participants 0.02 · controlled comparison 0.70 · randomized assignment 0.02 · states quantitative result 0.15 · compares with prior results 0.96 · mentions code or data availability 0.02.

The abstract, sentence by sentence

arXiv listing abstract v1. Results and the main claim are underlined; show every role.

Hybrid neuro-evolutionary algorithms may be inspired on Darwinian or Lamarckian evolu- tion. In the case of Darwinian evolution, the Baldwin effect, that is, the progressive incorporation of learned characteristics to the genotypes, can be observed and leveraged to improve the search. The purpose of this paper is to carry out an exper- imental study into how learning can improve G-Prop genetic search. Two ways of combining learning and genetic search are explored: one exploits the Baldwin effect, while the other uses a Lamarckian strategy. Our experiments show that using a Lamarckian op- erator makes the algorithm find networks with a low error rate, and the smallest size, while using the Bald- win effect obtains MLPs with the smallest error rate, and a larger size, taking longer to reach a solution. Both approaches obtain a lower average error than other BP-based algorithms like RPROP, other evolu- tionary methods and fuzzy logic based methods

background contribution method result interpretation main claim

What the classifier reports for each sentence. The number is the probability it gave the label.
#	sentence	role	the authors' own work	states the main conclusion	strength
s0	Hybrid neuro-evolutionary algorithms may be inspired on Darwinian or Lamarckian evolu- tion.	background 1.00	0.04	0.02	hedged possibility 1.00
s1	In the case of Darwinian evolution, the Baldwin effect, that is, the progressive incorporation of learned characteristics to the genotypes, can be observed and leveraged to improve the search.	background 1.00	0.06	0.03	asserted 0.91
s2	The purpose of this paper is to carry out an exper- imental study into how learning can improve G-Prop genetic search.	contribution 1.00	0.42	0.03	asserted 0.71
s3	Two ways of combining learning and genetic search are explored: one exploits the Baldwin effect, while the other uses a Lamarckian strategy.	method 0.86	0.58	0.03	asserted 0.90
s4	Our experiments show that using a Lamarckian op- erator makes the algorithm find networks with a low error rate, and the smallest size, while using the Bald- win effect obtains MLPs with the smallest error rate, and a larger size, taking longer to reach a solution.	result 1.00	0.97	0.83	asserted 1.00
s5	Both approaches obtain a lower average error than other BP-based algorithms like RPROP, other evolu- tionary methods and fuzzy logic based methods	result 1.00	0.89	0.25	asserted 1.00

What kind of paper the classifier reports this is: empirical study 0.95 · evidence from observation or experiment 0.77.

What was examined, and what was not

sources read

arXiv listing abstract v1

not examined

full text
figures and tables
references and citing sentences
relations to other papers

levels run

P abstract

not run

S sections
T body sentences
C citations
R cross-paper relations

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. Every number in the measurements table is the authors', checked by code. Record as JSON

References

Works cited (35)

[1]

J.B. Lamarck, “Philosophie zoologique,” 1809

View on Pith
[2]

J.M. Baldwin, “A new factor in evolution,” American Naturalist 30, 441-451 , 1896

View on Pith
[3]

C.H. W addington, “Canalization of development and the inheritance of acquired characteristics,” Nature, 3811, pp. 563-565 , 1942

View on Pith 1942
[4]

M. Gorges-Schleuter, “Asparagos96 and the traveling salesman problem,” In Proceedings of 1997 IEEE Inter- national Conference on Evolutionary Computation, pp. 171-174. IEEE , 1997

View on Pith 1997
[5]

P. Merz; B. Freisleben, “Genetic local search for the TSP : New results,” In Proceedings of 1997 IEEE International Conference on Evolutionary Computation, pp. 159-163. IEEE, 1997

View on Pith 1997
[6]

B.J. Ross, “A Lamarckian evolution strategy for genetic algorithms,” In Lance D. Chambers, editor, Practical Handbook of Genetic Algorithms: Complex Coding Sys- tems, volume III, pp. 1-16. Boca Raton, FL: CRC Press , 1999

View on Pith 1999
[7]

G.E. Hinton and S.J. Nowlan, “How learning can guide evolution,” Complex Systems, 1, 495-502 , 1987

View on Pith 1987
[8]

R.K. Belew, “When both individuals and populations search: Adding simple learning to the genetic algo- rithm.,” In 3th Intern. Conf. on Genetic Algorithms, D. Schaﬀer, ed., Morgan Kaufmann , 1989

View on Pith 1989
[9]

I. Harvey, “The puzzle of the persistent question marks: a case study of genetic drift,” In 5th International Con- ference on Genetic Algorithms, pp. 15-22, S. Forrest, ed. Morgan Kaufmann , 1993

View on Pith 1993
[10]

D.H. Ackley and M. Littman, “Interactions between learning and evolution,” In C.G. Langton, C. Taylor, J.D. Garmer, and S. Rasmussen (Editors), Artiﬁcial Life II, 487-507, Addison-Wesley, Reading, MA , 1992

View on Pith 1992
[11]

E.J.W. Boers; M.V. Borst; I.G. Sprinkhuizen-Kuyper, “Evolving Artiﬁcial Neural Networks using the ”Bald- win Eﬀect”,” In D.W. Pearson, N.C. Steele and R.F. Albrecht (eds.); Artiﬁcial Neural Nets and Genetic Algo- rithms. Proceedings of the International Conference in Al` es, France, 333-336, Springer Verlag Wien New York , 1995

View on Pith 1995
[12]

M. Huesken; J.E. Gayko; b. Sendhoﬀ, “Optimization for Problem Classes - Neural Networks that Learn to Learn,” Proceedings of the First IEEE Symposium on Combina- tions of Evolutionary Computation and Neural Networks (ECNN 2000), IEEE Press , 2000

View on Pith 2000
[13]

F. Gruau and D. Whitley, “Adding learning to the cel- lular development of neural networks: Evolution and the Baldwin efect,” Evolutionary Computation, Volume I, No. 3, pp. 213-233 , 1993

View on Pith 1993
[14]

show that taking advantage of the Bald- win eﬀect can ﬁnd the global optimum, while a Lamarckian strategy, although faster, usually converges to a local optimum. On the other hand, results obtained by Ku and Mak [15] with a GA designed to evolve recurrent neural networks, show that the use of a Lamar- ckian strategy implies an improvement of the algorithm...

View on Pith
[15]

Springer-Verlag , 1994

View on Pith 1994
[16]

D. Whitley; V.S. Gordon; K. Mathias, “Lamarckian Evolution, The Baldwin Eﬀect and Function Optimiza- tion,” Parallel Problem Solving from Nature-PPSN III. Y. Davidor, H.P. Schwefel and R. Manner, eds. pp. 6-

View on Pith
[17]

B.A. Julstrom, “Comparing Darwinian, Baldwinian and Lamarckian Search in a Genetic Algorithm for the 4- Cycle Problem,” In Congress on Evolutionary Computa- tion,In Genetic and Evolutionary Computation Confer- ence, Late Breaking Papers, pp. 134-138, Orlando, USA , 1999

View on Pith 1999
[18]

K.W.C. Ku and M.W. Mak, “Exploring the eﬀects of Lamarckian and Baldwinian learning in evolving recur- rent neural networks,” In Proceedings of 1997 IEEE International Conference on Evolutionary Computation, pp. 159-163. IEEE , 1997

View on Pith 1997
[19]

C. Houck; J.A. Joines; M.G. Kay; J.R. Wilson, “Em- pirical investigation of the beneﬁts of partial Lamarck- ianism,” Evolutionary Computation, v.5, n.1, pp. 31-60 , 1997

View on Pith 1997
[20]

P.A. Castillo, J.J. Merelo, V. Rivas, G. Romero, and A. Prieto, “G-Prop: Global Optimization of Multilayer Perceptrons using GAs,” Neurocomputing, Vol.35/1-4, pp.149-163, 2000

View on Pith 2000
[21]

S.E. Fahlman, “Faster-Learning Variations on Back- Propagation: An Empirical Study,” Proceedings of the 1988 Connectionist Models Summer School, Morgan Kaufmann, 1988

View on Pith 1988
[22]

P.A. Castillo, J. Gonz´ alez, J.J. Merelo, V. Rivas, G. Romero, and A. Prieto, “SA-Prop: Optimization of Multilayer Perceptron Parameters using Simulated An- nealing,” Lecture Notes in Computer Science, ISBN:3- 540-66069-0, Vol. 1606, pp. 661-670, Springer-Verlag , 1999

View on Pith 1999
[23]

M. Riedmiller and H. Braun, “A Direct Adaptive Method for Faster Backpropagation Learning: The RPROP Al- gorithm,” In Ruspini, H., (Ed.) Proc. of the ICNN93, San Francisco, pp. 586-591 , 1993

View on Pith 1993
[24]

P.A. Castillo, J. Carpio, J.J. Merelo, V. Rivas, G. Romero, and A. Prieto, “Evolving Multilayer Percep- trons,” Neural Processing Letters, vol. 12, no. 2, pp.115-

View on Pith
[25]

P.A. Castillo, M.G. Arenas, J.G. Castellano, M.Ciller o, J.J. Merelo, A. Prieto, V. Rivas, and G. Romero, “Func- tion Approximation with Evolved Multilayer Percep- trons,” Advances in Neural Networks and Applica- tions. Artiﬁcial Intelligence Series. Nikos E. Mastorakis Editor. ISBN:960-8052-26-2, pp.195-200, Published by World Scientiﬁc and Engineering S...

View on Pith 2001
[26]

M.A. Gr¨ onroos, “Evolutionary Design of Neural Net- works,” Master of Science Thesis in Computer Sci- ence. Department of Mathematical Sciences. University of Turku. , 1998

View on Pith 1998
[27]

S. Fahlman, “An empirical study of learning speed in back-propagation networks,” Tech. Rep., Carnegie Mel- lon University, 1988

View on Pith 1988
[28]

Lutz Prechelt, “PROBEN1 — A set of benchmarks and benchmarking rules for neural network training al- gorithms,” Technical Report 21/94, Fakult¨ at f¨ ur Infor- matik, Universit¨ at Karlsruhe, D-76128 Karlsruhe, Ger- many. (Also in: http://wwwipd.ira.uka.de/˜prechelt/) , 1994

View on Pith 1994
[29]

in Table II. Approach Error Size Generations Lamarckian 0.09 ± 0.01 18 ± 8 34 ± 28 Baldwinian 0.086 ± 0.004 85 ± 27 97 ± 81 Pomares [29] 0.125 6 (4 rules) - TABLA II Results for the f2 problem obtained with G-Prop taking advantage of the Baldwin eﬀect and for the Lamarckian approach, as well as those obtained by Pomares, which are included for the sake of...

View on Pith 2000
[30]

V. Cherkassky; D. Gehring; F. Mulier, “Comparison of adaptive methods for function estimation from samples,” IEEE Trans. Neural Networks, vol. 7, no. 4, pp. 969-984 , 1996

View on Pith 1996
[31]

M. Sugeno and T. Yasukawa, “A fuzzy-logic based ap- proach to qualitative modeling,” IEEE Fuzzy Sets and Systems, vol. 1, no. 1 , 1993

View on Pith 1993
[32]

H. Pomares Cintas, “Nueva metodolog ´ ıa para el dise˜ no autom´ atico de sistemas difusos,” Tesis Doctoral. Depar- tamento de Arquitectura y Tecnolog ´ ıa de Computadores. Universidad de Granada , 1999

View on Pith 1999
[33]

M. Riedmiller, “Description and Implementation De- tails,” Tech. Rep., University of Karlsruhe, 1994

View on Pith 1994
[34]

M. Oliveira; J. Barreiros; E. Costa; F. Pereira, “Lam- BaDa: An Artiﬁcial Environment to Study the Inter- action beween Evolution and Learning,” In Congress on Evolutionary Computation, Volume I, pp. 145-152, Washington D.C., USA , 1999

View on Pith 1999
[35]

G. Romero, M.G. Arenas, J. Carpio, J.G. Castellano, P.A. Castillo, J.J. Merelo, A. Prieto, and V. Rivas, “Evo- lutionary Computation Visualization: Application to G- Prop,” Lecture Notes in Computer Science, ISBN:3-540- 41056-2, Vol.1917, pp.902-912. September , 2000

View on Pith 1917

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Population Based Training of Neural Networks
cs.LG 2017-11 conditional novelty 6.0 of 10

Running a population of models that copy weights from better performers and perturb their hyperparameters mid-training yields automatically discovered hyperparameter schedules and better final models than equal-budget...
A Fresh Look at Lamarckian Evolution and the Baldwin Effect
cs.NE 2026-05 unverdicted novelty 5.0 of 10

Baldwinian and Lamarckian evolution outperform Darwinian evolution empirically on Max Independent Set and Max Cut tasks and show better asymptotic runtime on the extended Deceptive Leading Block benchmark.

How the literature uses it

1 arXiv paper cite this in 2 sentences

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

Discussion

Disagree with the review, or know something it missed? Leave a comment, highlight any passage above to comment on it, or challenge this review formally.

[1] [1]

J.B. Lamarck, “Philosophie zoologique,” 1809

View on Pith

[2] [2]

J.M. Baldwin, “A new factor in evolution,” American Naturalist 30, 441-451 , 1896

View on Pith

[3] [3]

C.H. W addington, “Canalization of development and the inheritance of acquired characteristics,” Nature, 3811, pp. 563-565 , 1942

View on Pith 1942

[4] [4]

M. Gorges-Schleuter, “Asparagos96 and the traveling salesman problem,” In Proceedings of 1997 IEEE Inter- national Conference on Evolutionary Computation, pp. 171-174. IEEE , 1997

View on Pith 1997

[5] [5]

P. Merz; B. Freisleben, “Genetic local search for the TSP : New results,” In Proceedings of 1997 IEEE International Conference on Evolutionary Computation, pp. 159-163. IEEE, 1997

View on Pith 1997

[6] [6]

B.J. Ross, “A Lamarckian evolution strategy for genetic algorithms,” In Lance D. Chambers, editor, Practical Handbook of Genetic Algorithms: Complex Coding Sys- tems, volume III, pp. 1-16. Boca Raton, FL: CRC Press , 1999

View on Pith 1999

[7] [7]

G.E. Hinton and S.J. Nowlan, “How learning can guide evolution,” Complex Systems, 1, 495-502 , 1987

View on Pith 1987

[8] [8]

R.K. Belew, “When both individuals and populations search: Adding simple learning to the genetic algo- rithm.,” In 3th Intern. Conf. on Genetic Algorithms, D. Schaﬀer, ed., Morgan Kaufmann , 1989

View on Pith 1989

[9] [9]

I. Harvey, “The puzzle of the persistent question marks: a case study of genetic drift,” In 5th International Con- ference on Genetic Algorithms, pp. 15-22, S. Forrest, ed. Morgan Kaufmann , 1993

View on Pith 1993

[10] [10]

D.H. Ackley and M. Littman, “Interactions between learning and evolution,” In C.G. Langton, C. Taylor, J.D. Garmer, and S. Rasmussen (Editors), Artiﬁcial Life II, 487-507, Addison-Wesley, Reading, MA , 1992

View on Pith 1992

[11] [11]

E.J.W. Boers; M.V. Borst; I.G. Sprinkhuizen-Kuyper, “Evolving Artiﬁcial Neural Networks using the ”Bald- win Eﬀect”,” In D.W. Pearson, N.C. Steele and R.F. Albrecht (eds.); Artiﬁcial Neural Nets and Genetic Algo- rithms. Proceedings of the International Conference in Al` es, France, 333-336, Springer Verlag Wien New York , 1995

View on Pith 1995

[12] [12]

M. Huesken; J.E. Gayko; b. Sendhoﬀ, “Optimization for Problem Classes - Neural Networks that Learn to Learn,” Proceedings of the First IEEE Symposium on Combina- tions of Evolutionary Computation and Neural Networks (ECNN 2000), IEEE Press , 2000

View on Pith 2000

[13] [13]

F. Gruau and D. Whitley, “Adding learning to the cel- lular development of neural networks: Evolution and the Baldwin efect,” Evolutionary Computation, Volume I, No. 3, pp. 213-233 , 1993

View on Pith 1993

[14] [14]

show that taking advantage of the Bald- win eﬀect can ﬁnd the global optimum, while a Lamarckian strategy, although faster, usually converges to a local optimum. On the other hand, results obtained by Ku and Mak [15] with a GA designed to evolve recurrent neural networks, show that the use of a Lamar- ckian strategy implies an improvement of the algorithm...

View on Pith

[15] [15]

Springer-Verlag , 1994

View on Pith 1994

[16] [16]

D. Whitley; V.S. Gordon; K. Mathias, “Lamarckian Evolution, The Baldwin Eﬀect and Function Optimiza- tion,” Parallel Problem Solving from Nature-PPSN III. Y. Davidor, H.P. Schwefel and R. Manner, eds. pp. 6-

View on Pith

[17] [17]

B.A. Julstrom, “Comparing Darwinian, Baldwinian and Lamarckian Search in a Genetic Algorithm for the 4- Cycle Problem,” In Congress on Evolutionary Computa- tion,In Genetic and Evolutionary Computation Confer- ence, Late Breaking Papers, pp. 134-138, Orlando, USA , 1999

View on Pith 1999

[18] [18]

K.W.C. Ku and M.W. Mak, “Exploring the eﬀects of Lamarckian and Baldwinian learning in evolving recur- rent neural networks,” In Proceedings of 1997 IEEE International Conference on Evolutionary Computation, pp. 159-163. IEEE , 1997

View on Pith 1997

[19] [19]

C. Houck; J.A. Joines; M.G. Kay; J.R. Wilson, “Em- pirical investigation of the beneﬁts of partial Lamarck- ianism,” Evolutionary Computation, v.5, n.1, pp. 31-60 , 1997

View on Pith 1997

[20] [20]

P.A. Castillo, J.J. Merelo, V. Rivas, G. Romero, and A. Prieto, “G-Prop: Global Optimization of Multilayer Perceptrons using GAs,” Neurocomputing, Vol.35/1-4, pp.149-163, 2000

View on Pith 2000

[21] [21]

S.E. Fahlman, “Faster-Learning Variations on Back- Propagation: An Empirical Study,” Proceedings of the 1988 Connectionist Models Summer School, Morgan Kaufmann, 1988

View on Pith 1988

[22] [22]

P.A. Castillo, J. Gonz´ alez, J.J. Merelo, V. Rivas, G. Romero, and A. Prieto, “SA-Prop: Optimization of Multilayer Perceptron Parameters using Simulated An- nealing,” Lecture Notes in Computer Science, ISBN:3- 540-66069-0, Vol. 1606, pp. 661-670, Springer-Verlag , 1999

View on Pith 1999

[23] [23]

M. Riedmiller and H. Braun, “A Direct Adaptive Method for Faster Backpropagation Learning: The RPROP Al- gorithm,” In Ruspini, H., (Ed.) Proc. of the ICNN93, San Francisco, pp. 586-591 , 1993

View on Pith 1993

[24] [24]

P.A. Castillo, J. Carpio, J.J. Merelo, V. Rivas, G. Romero, and A. Prieto, “Evolving Multilayer Percep- trons,” Neural Processing Letters, vol. 12, no. 2, pp.115-

View on Pith

[25] [25]

P.A. Castillo, M.G. Arenas, J.G. Castellano, M.Ciller o, J.J. Merelo, A. Prieto, V. Rivas, and G. Romero, “Func- tion Approximation with Evolved Multilayer Percep- trons,” Advances in Neural Networks and Applica- tions. Artiﬁcial Intelligence Series. Nikos E. Mastorakis Editor. ISBN:960-8052-26-2, pp.195-200, Published by World Scientiﬁc and Engineering S...

View on Pith 2001

[26] [26]

M.A. Gr¨ onroos, “Evolutionary Design of Neural Net- works,” Master of Science Thesis in Computer Sci- ence. Department of Mathematical Sciences. University of Turku. , 1998

View on Pith 1998

[27] [27]

S. Fahlman, “An empirical study of learning speed in back-propagation networks,” Tech. Rep., Carnegie Mel- lon University, 1988

View on Pith 1988

[28] [28]

Lutz Prechelt, “PROBEN1 — A set of benchmarks and benchmarking rules for neural network training al- gorithms,” Technical Report 21/94, Fakult¨ at f¨ ur Infor- matik, Universit¨ at Karlsruhe, D-76128 Karlsruhe, Ger- many. (Also in: http://wwwipd.ira.uka.de/˜prechelt/) , 1994

View on Pith 1994

[29] [29]

in Table II. Approach Error Size Generations Lamarckian 0.09 ± 0.01 18 ± 8 34 ± 28 Baldwinian 0.086 ± 0.004 85 ± 27 97 ± 81 Pomares [29] 0.125 6 (4 rules) - TABLA II Results for the f2 problem obtained with G-Prop taking advantage of the Baldwin eﬀect and for the Lamarckian approach, as well as those obtained by Pomares, which are included for the sake of...

View on Pith 2000

[30] [30]

V. Cherkassky; D. Gehring; F. Mulier, “Comparison of adaptive methods for function estimation from samples,” IEEE Trans. Neural Networks, vol. 7, no. 4, pp. 969-984 , 1996

View on Pith 1996

[31] [31]

M. Sugeno and T. Yasukawa, “A fuzzy-logic based ap- proach to qualitative modeling,” IEEE Fuzzy Sets and Systems, vol. 1, no. 1 , 1993

View on Pith 1993

[32] [32]

H. Pomares Cintas, “Nueva metodolog ´ ıa para el dise˜ no autom´ atico de sistemas difusos,” Tesis Doctoral. Depar- tamento de Arquitectura y Tecnolog ´ ıa de Computadores. Universidad de Granada , 1999

View on Pith 1999

[33] [33]

M. Riedmiller, “Description and Implementation De- tails,” Tech. Rep., University of Karlsruhe, 1994

View on Pith 1994

[34] [34]

M. Oliveira; J. Barreiros; E. Costa; F. Pereira, “Lam- BaDa: An Artiﬁcial Environment to Study the Inter- action beween Evolution and Learning,” In Congress on Evolutionary Computation, Volume I, pp. 145-152, Washington D.C., USA , 1999

View on Pith 1999

[35] [35]

G. Romero, M.G. Arenas, J. Carpio, J.G. Castellano, P.A. Castillo, J.J. Merelo, A. Prieto, and V. Rivas, “Evo- lutionary Computation Visualization: Application to G- Prop,” Lecture Notes in Computer Science, ISBN:3-540- 41056-2, Vol.1917, pp.902-912. September , 2000

View on Pith 1917