Activity-Dependent Plasticity in Morphogenetically-Grown Recurrent Networks

Andrii Valenia; Mykola Glybovets; Sergii Medvid

arxiv: 2604.03386 · v1 · submitted 2026-04-03 · 💻 cs.RO · cs.NE

Activity-Dependent Plasticity in Morphogenetically-Grown Recurrent Networks

Sergii Medvid , Andrii Valenia , Mykola Glybovets This is my paper

Pith reviewed 2026-05-13 18:31 UTC · model grok-4.3

classification 💻 cs.RO cs.NE

keywords developmental roboticsmorphogenetic growthanti-Hebbian plasticityrecurrent controllersco-evolutionneural architecture searchactivity-dependent plasticity

0 comments

The pith

Morphogenetically grown recurrent networks perform substantially better when equipped with anti-Hebbian plasticity rather than Hebbian rules.

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

This paper examines how activity-dependent plasticity affects controllers grown from compact genomes through self-organisation. Across 50,000 such recurrent networks tested on CartPole and Acrobot, anti-Hebbian plasticity consistently outperforms Hebbian plasticity with moderate to large effect sizes. Fixed post-growth weights leave substantial performance on the table, with regret reaching 52 to 100 percent. Co-evolutionary experiments, where plasticity parameters evolve with the architecture, independently recover the anti-Hebbian pattern, especially on CartPole where 70 percent of runs do so. The advantage appears generic to small recurrent networks but is amplified by the specific topologies produced by morphogenetic growth.

Core claim

Characterisation of over 50,000 morphogenetically grown recurrent controllers reveals that anti-Hebbian plasticity significantly outperforms Hebbian plasticity for competent networks, with Cohen's d between 0.53 and 0.64, and that regret for the best fixed weight setting reaches 52-100 percent. Plasticity's role changes from fine-tuning to genuine adaptation when environments are non-stationary. Co-evolution of plasticity parameters alongside developmental architecture independently discovers these patterns, evolving anti-Hebbian plasticity in 70 percent of CartPole runs and near-zero eta with mixed signs on Acrobot. A random-RNN control confirms anti-Hebbian dominance is generic, yet the高拓扑

What carries the argument

Activity-dependent plasticity rules (Hebbian and anti-Hebbian) applied to weights in morphogenetically grown recurrent networks, with parameters that can be co-evolved with the developmental architecture.

If this is right

Anti-Hebbian plasticity enables better performance in competent developmental controllers compared to Hebbian.
High regret values indicate that fixed weights after growth miss a large fraction of possible improvement.
Co-evolution can discover effective plasticity rules without prior specification.
The performance gap between fixed and plastic is larger in morphogenetically grown networks than in random recurrent graphs with similar statistics.
Plasticity becomes essential for adaptation rather than just fine-tuning when task conditions change.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Evolving plasticity rules together with the network architecture may be a general strategy for building adaptable developmental systems.
The finding that anti-Hebbian is preferred suggests reconsidering standard Hebbian assumptions in growing neural controllers.
Similar experiments on other developmental growth models could test if anti-Hebbian dominance is widespread.
High regret in developmental networks points to potential benefits from incorporating plasticity in real-world robotic controllers.

Load-bearing premise

The 50,000 morphogenetically grown networks and the two specific control tasks sufficiently represent the space of developmental controllers and that co-evolution truly discovers plasticity rules independently.

What would settle it

If co-evolutionary runs on additional tasks such as more complex locomotion benchmarks evolve Hebbian plasticity instead or show no preference, this would falsify the independent discovery of anti-Hebbian patterns.

Figures

Figures reproduced from arXiv: 2604.03386 by Andrii Valenia, Mykola Glybovets, Sergii Medvid.

**Figure 2.** Figure 2: CartPole High-mid stratum: mean Δ𝑟 across the 𝜂×𝜆 grid (248-point extended grid). Anti-Hebbian (𝜂 < 0) with moderate decay (𝜆 = 0.01) yields the best results. Hebbian (𝜂 > 0) is consistently harmful. −0.75 −0.50 −0.25 0.00 0.25 0.50 0.75 1.00 Cohen's d (positive = anti-Hebbian advantage) Weak Low-mid High-mid Near-perf. Perfect -0.14 0.53 0.64 0.56 -0.07 [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: CartPole: Cohen’s 𝑑 (anti-Hebbian − Hebbian) by stratum. Anti-Hebbian significantly outperforms Hebbian for competent strata (𝑑 = 0.53–0.64, all 𝑝 < 0.001). The effect reverses for Weak and Perfect networks with small magnitudes. the highest quintile. While not dispositive, this pattern is consistent with magnitude moderation—anti-Hebbian updates counteract co-activation rather than amplifying it, preven… view at source ↗

**Figure 4.** Figure 4: CartPole adaptation premium (non-stationary plas [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

**Figure 6.** Figure 6: Co-evolution (Condition C) evolved 𝜂 over 200 generations on CartPole (top) and Acrobot (bottom). Left: best individual’s 𝜂; right: population mean𝜂 (median ± IQR across 30 runs). CartPole: strong anti-Hebbian drift (𝜂 < 0). Acrobot: 𝜂 hovers near zero with mixed signs, matching the characterisation finding. anti-Hebbian dominance is a property of small recurrent networks on CartPole, not of developmenta… view at source ↗

read the original abstract

Developmental approaches to neural architecture search grow functional networks from compact genomes through self-organisation, but the resulting networks operate with fixed post-growth weights. We characterise Hebbian and anti-Hebbian plasticity across 50,000 morphogenetically grown recurrent controllers (5M+ configurations on CartPole and Acrobot), then test whether co-evolutionary experiments -- where plasticity parameters are encoded in the genome and evolved alongside the developmental architecture -- recover these patterns independently. Our characterisation reveals that (1) anti-Hebbian plasticity significantly outperforms Hebbian for competent networks (Cohen's d = 0.53-0.64), (2) regret (fraction of oracle improvement lost under the best fixed setting) reaches 52-100%, and (3) plasticity's role shifts from fine-tuning to genuine adaptation under non-stationarity. Co-evolution independently discovers these patterns: on CartPole, 70% of runs evolve anti-Hebbian plasticity (p = 0.043); on Acrobot, evolution finds near-zero eta with mixed signs -- exactly matching the characterisation. A random-RNN control shows that anti-Hebbian dominance is generic to small recurrent networks, but the degree of topology-dependence is developmental-specific: regret is 2-6x higher for morphogenetically grown networks than for random graphs with matched topology statistics.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper shows anti-Hebbian plasticity beats Hebbian in morphogenetically grown recurrent controllers with high regret, and co-evolution recovers matching patterns, but the methods leave too many gaps to verify the independence claim.

read the letter

The main thing to know is that anti-Hebbian plasticity gives a moderate edge over Hebbian rules for competent networks on CartPole and Acrobot, with regret between 52 and 100 percent, and that co-evolutionary runs recover anti-Hebbian dominance on CartPole at 70 percent of trials. The random-graph control is useful because it shows the high topology dependence is stronger in the developmental case than in matched random RNNs. That comparison, plus the scale of the 50k-network characterization, is the clearest new piece beyond earlier fixed-weight developmental work. The co-evo recovery rates line up with the characterization, which is a reasonable empirical step for adaptive neuroevolution. The paper does a decent job laying out the quantitative patterns and effect sizes in the abstract. The soft spots sit in the methods. Exact growth mechanics, the plasticity update equations, data exclusion rules, and full statistical controls are not spelled out enough to check the central claims. The stress-test point about co-evolution independence is fair: without an ablation that changes the evolutionary operators or fitness while keeping the genome mapping, the match could still be an artifact of the search setup rather than a true interaction between development and task. The weakest assumption is that these two tasks and 50k networks represent the broader space of developmental controllers. This work is for people in evolutionary robotics and neuroevolution who already care about developmental NAS and plasticity. A reader building adaptive controllers would get some concrete patterns to test against. It deserves peer review because the scale and the control experiment give it enough substance to warrant referee time, even if revisions will need to fill in the method details and add the missing ablations.

Referee Report

2 major / 2 minor

Summary. The manuscript characterizes Hebbian and anti-Hebbian plasticity across 50,000 morphogenetically grown recurrent controllers (5M+ configurations) on CartPole and Acrobot, reporting that anti-Hebbian plasticity significantly outperforms Hebbian for competent networks (Cohen's d = 0.53-0.64), regret reaches 52-100%, and plasticity's role shifts under non-stationarity. Co-evolutionary experiments, encoding plasticity parameters in the genome, recover these patterns independently: 70% of CartPole runs evolve anti-Hebbian plasticity (p = 0.043) and Acrobot yields near-zero eta with mixed signs. A random-RNN control shows anti-Hebbian dominance is generic to small recurrent networks, but topology-dependence is developmental-specific (regret 2-6x higher for morphogenetically grown networks).

Significance. If the empirical results hold, the work provides large-scale evidence that activity-dependent plasticity enhances performance in developmentally grown controllers and that co-evolution can discover appropriate rules without direct supervision, advancing neuroevolution and adaptive robotics. The scale of the characterization, quantitative effect sizes, and random-network control are strengths that support falsifiable claims about generic vs. developmental-specific effects.

major comments (2)

[Abstract and Methods] Abstract and Methods: the abstract reports quantitative results including effect sizes and percentages, yet lacks details on network growth mechanics, exact plasticity update equations, data exclusion criteria, or statistical controls, leaving major gaps that prevent full verification of the central empirical claims about performance differences and regret.
[Co-evolutionary experiments] Co-evolutionary experiments: the claim that co-evolution independently discovers the patterns (70% anti-Hebbian on CartPole, p=0.043) requires that the reported patterns arise from developmental growth and task demands rather than the particular genetic encoding, selection operator, or fitness function; the manuscript provides no ablation that replaces the evolutionary algorithm while keeping the genome-to-network mapping, which is load-bearing for the independence assertion.

minor comments (2)

[Abstract] Abstract: the term 'regret' is introduced as the fraction of oracle improvement lost under the best fixed setting; provide an explicit equation or definition for how the oracle is computed and how regret is aggregated across networks.
[Results] Results: ensure all statistical tests (e.g., p = 0.043) are accompanied by the exact test used, sample sizes, and correction for multiple comparisons in the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive review and recommendation for major revision. We address each major comment point by point below, providing the strongest honest responses based on the manuscript content. We have revised the manuscript to improve clarity on methods and to discuss limitations of the co-evolutionary claims.

read point-by-point responses

Referee: [Abstract and Methods] Abstract and Methods: the abstract reports quantitative results including effect sizes and percentages, yet lacks details on network growth mechanics, exact plasticity update equations, data exclusion criteria, or statistical controls, leaving major gaps that prevent full verification of the central empirical claims about performance differences and regret.

Authors: We acknowledge that the abstract, constrained by length, omits full methodological specifics. The full Methods section details the morphogenetic growth mechanics (self-organizing rules from compact genomes, Section 3.1), exact plasticity equations (Hebbian and anti-Hebbian updates based on pre/post-synaptic activity, Equations 1-3), data exclusion criteria (e.g., excluding non-competent networks below performance thresholds, Section 4.2), and statistical controls (Cohen's d, regret calculations, binomial p-values). To address verification concerns, we have revised the abstract to briefly reference the setup and growth process, added explicit cross-references to equations and sections in the main text, and included a note on statistical methods. These changes ensure the quantitative claims can be traced without altering the reported results. revision: yes
Referee: [Co-evolutionary experiments] Co-evolutionary experiments: the claim that co-evolution independently discovers the patterns (70% anti-Hebbian on CartPole, p=0.043) requires that the reported patterns arise from developmental growth and task demands rather than the particular genetic encoding, selection operator, or fitness function; the manuscript provides no ablation that replaces the evolutionary algorithm while keeping the genome-to-network mapping, which is load-bearing for the independence assertion.

Authors: We agree that an ablation varying the evolutionary algorithm (while preserving the genome-to-network mapping) would strengthen the independence claim. The current setup uses a direct encoding of plasticity parameters (eta and sign) in the genome with standard fitness-based selection, and the evolved outcomes align precisely with the exhaustive characterization across 50,000 networks. This alignment indicates the patterns are driven by task demands and developmental constraints rather than arbitrary GA choices. In revision, we have added a dedicated limitations paragraph discussing potential influences of the encoding and selection operator, along with the p-value and percentage results from the performed runs. A full ablation is beyond the current scope due to computational cost but is noted as valuable future work; the existing evidence from pattern matching remains supportive of the claim. revision: partial

Circularity Check

0 steps flagged

No significant circularity; claims rest on separate empirical characterization and co-evolutionary runs

full rationale

The paper first characterizes plasticity performance across 50,000 independently grown networks (yielding Cohen's d values, regret figures, and topology comparisons) and then runs separate co-evolutionary experiments that encode plasticity parameters in the genome. No derivation step reduces by the paper's own equations or definitions to a fitted parameter renamed as a prediction, nor does any load-bearing claim collapse to a self-citation chain. The observation that co-evolution recovers anti-Hebbian dominance on CartPole is an empirical outcome of the evolutionary dynamics, not a definitional equivalence to the prior characterization. The work is therefore self-contained against its experimental benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claims rest on standard domain assumptions about plasticity rules and evolutionary search; no new entities are postulated. Free parameters include the plasticity rate and any evolutionary hyperparameters that were tuned or evolved to produce the reported outcomes.

free parameters (1)

plasticity rate eta
The learning rate parameter in the Hebbian and anti-Hebbian update rules is either fixed during characterization or evolved during co-evolution and directly affects measured performance differences.

axioms (1)

domain assumption Hebbian and anti-Hebbian rules constitute appropriate and sufficient models of activity-dependent plasticity for these recurrent controllers.
Invoked throughout the characterization and co-evolution experiments as the basis for testing plasticity effects.

pith-pipeline@v0.9.0 · 5548 in / 1484 out tokens · 59543 ms · 2026-05-13T18:31:43.223199+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

[1]

Joaquín Derrac, Salvador García, Daniel Molina, and Francisco Herrera. 2011. A practical tutorial on the use of nonparametric statistical tests as a methodology for comparing evolutionary and swarm intelligence algorithms.Swarm and Evolutionary Computation1, 1 (2011), 3–18. doi:10.1016/j.swevo.2011.02.002

work page doi:10.1016/j.swevo.2011.02.002 2011
[2]

Daniel García Núñez, Fergal Stapleton, and Edgar Galván. 2025. Time-Modulated Hebbian Learning for Classic Control Tasks. InGECCO ’25 Companion: Proceed- ings of the Genetic and Evolutionary Computation Conference Companion. ACM, 2119–2126. doi:10.1145/3712255.3734322

work page doi:10.1145/3712255.3734322 2025
[3]

Mykola Glybovets and Sergii Medvid. 2026. MorphoNAS: Embryogenic Neural Architecture Search Through Morphogen-Guided Development.Kibernetyka i Systemnyj Analiz (Cybernetics and Systems Analysis)62, 2 (2026). doi:10.34229/ KCA2522-9664.26.2.3 Preprint: arXiv:2507.13785

work page arXiv 2026
[4]

1949.The Organization of Behavior: A Neuropsychological Theory

Donald O Hebb. 1949.The Organization of Behavior: A Neuropsychological Theory. Wiley

work page 1949
[5]

Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson, and Michael Levin

work page
[6]

doi:10.23915/ distill.00023

Growing Neural Cellular Automata.Distill5, 2 (2020), e23. doi:10.23915/ distill.00023

work page 2020
[7]

Elias Najarro and Sebastian Risi. 2020. Meta-Learning through Hebbian Plasticity in Random Networks. InAdvances in Neural Information Processing Systems, Vol. 33. 20719–20731

work page 2020
[8]

Elias Najarro, Shyam Sudhakaran, Claire Glanois, and Sebastian Risi. 2022. Hy- perNCA: Growing Developmental Networks with Neural Cellular Automata. In From Cells to Societies: Collective Learning Across Scales (ICLR 2022 Workshop)

work page 2022
[9]

Elias Najarro, Shyam Sudhakaran, and Sebastian Risi. 2023. Towards Self- Assembling Artificial Neural Networks through Neural Developmental Programs. InProceedings of the 2023 Artificial Life Conference (ALIFE), Vol. 35. MIT Press,

work page 2023
[10]

doi:10.1162/isal_a_00693

work page doi:10.1162/isal_a_00693
[11]

Eleni Nisioti, Erwan Plantec, Milton Montero, Joachim Winther Pedersen, and Sebastian Risi. 2024. Growing Artificial Neural Networks for Control: the Role of Neuronal Diversity. InGECCO ’24 Companion: Proceedings of the Genetic and Evolutionary Computation Conference Companion. ACM, 175–178. doi:10.1145/ 3638530.3654356

work page arXiv 2024
[12]

Rasmus Berg Palm, Elias Najarro, and Sebastian Risi. 2021. Testing the Genomic Bottleneck Hypothesis in Hebbian Meta-Learning. InProceedings of Machine Learning Research, Vol. 148. 100–110

work page 2021
[13]

Erwan Plantec, Joachim Winther Pedersen, Milton Llera Montero, Eleni Nisioti, and Sebastian Risi. 2024. Evolving Self-Assembling Neural Networks: From Spontaneous Activity to Experience-Dependent Learning. InProceedings of the 2024 Conference on Artificial Life. MIT Press

work page 2024
[14]

Sergey Shuvaev, Divyansha Lachi, Alexei Koulakov, and Anthony Zador. 2024. Encoding innate ability through a genomic bottleneck.Proceedings of the National Academy of Sciences121, 38 (2024), e2409160121. doi:10.1073/pnas.2409160121

work page doi:10.1073/pnas.2409160121 2024
[15]

Andrea Soltoggio, Kenneth O Stanley, and Sebastian Risi. 2018. Born to learn: the inspiration, progress, and future of evolved plastic artificial neural networks. Neural Networks108 (2018), 48–67. doi:10.1016/j.neunet.2018.07.013

work page doi:10.1016/j.neunet.2018.07.013 2018
[16]

Kenneth O Stanley, David B D’Ambrosio, and Jason Gauci. 2009. A Hypercube- Based Encoding for Evolving Large-Scale Neural Networks.Artificial Life15, 2 (2009), 185–212

work page 2009
[17]

Kenneth O Stanley and Risto Miikkulainen. 2002. Evolving Neural Networks through Augmenting Topologies.Evolutionary Computation10, 2 (2002), 99–127

work page 2002
[18]

2024.Gymnasium

Mark Towers, Ariel Kwiatkowski, Jordan Terry, John U Balis, Gianluca De Cola, Tristan Deleu, Manuel Goulão, Andreas Kallinteris, Arjun KG, Markus Krimmel, Rodrigo Perez-Vicente, Andrea Pierré, Sander Schulhoff, Jun Jet Tai, Andrew Tan, and Omar G Younis. 2024.Gymnasium. https://github.com/Farama-Foundation/ Gymnasium

work page 2024
[19]

Alan M Turing. 1952. The Chemical Basis of Morphogenesis.Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences237, 641 (1952), 37–72. doi:10.1098/rstb.1952.0012

work page doi:10.1098/rstb.1952.0012 1952
[20]

Anthony M Zador. 2019. A critique of pure learning and what artificial neural networks can learn from animal brains.Nature Communications10, 1 (2019),

work page 2019
[21]

doi:10.1038/s41467-019-11786-6

work page doi:10.1038/s41467-019-11786-6

[1] [1]

Joaquín Derrac, Salvador García, Daniel Molina, and Francisco Herrera. 2011. A practical tutorial on the use of nonparametric statistical tests as a methodology for comparing evolutionary and swarm intelligence algorithms.Swarm and Evolutionary Computation1, 1 (2011), 3–18. doi:10.1016/j.swevo.2011.02.002

work page doi:10.1016/j.swevo.2011.02.002 2011

[2] [2]

Daniel García Núñez, Fergal Stapleton, and Edgar Galván. 2025. Time-Modulated Hebbian Learning for Classic Control Tasks. InGECCO ’25 Companion: Proceed- ings of the Genetic and Evolutionary Computation Conference Companion. ACM, 2119–2126. doi:10.1145/3712255.3734322

work page doi:10.1145/3712255.3734322 2025

[3] [3]

Mykola Glybovets and Sergii Medvid. 2026. MorphoNAS: Embryogenic Neural Architecture Search Through Morphogen-Guided Development.Kibernetyka i Systemnyj Analiz (Cybernetics and Systems Analysis)62, 2 (2026). doi:10.34229/ KCA2522-9664.26.2.3 Preprint: arXiv:2507.13785

work page arXiv 2026

[4] [4]

1949.The Organization of Behavior: A Neuropsychological Theory

Donald O Hebb. 1949.The Organization of Behavior: A Neuropsychological Theory. Wiley

work page 1949

[5] [5]

Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson, and Michael Levin

work page

[6] [6]

doi:10.23915/ distill.00023

Growing Neural Cellular Automata.Distill5, 2 (2020), e23. doi:10.23915/ distill.00023

work page 2020

[7] [7]

Elias Najarro and Sebastian Risi. 2020. Meta-Learning through Hebbian Plasticity in Random Networks. InAdvances in Neural Information Processing Systems, Vol. 33. 20719–20731

work page 2020

[8] [8]

Elias Najarro, Shyam Sudhakaran, Claire Glanois, and Sebastian Risi. 2022. Hy- perNCA: Growing Developmental Networks with Neural Cellular Automata. In From Cells to Societies: Collective Learning Across Scales (ICLR 2022 Workshop)

work page 2022

[9] [9]

Elias Najarro, Shyam Sudhakaran, and Sebastian Risi. 2023. Towards Self- Assembling Artificial Neural Networks through Neural Developmental Programs. InProceedings of the 2023 Artificial Life Conference (ALIFE), Vol. 35. MIT Press,

work page 2023

[10] [10]

doi:10.1162/isal_a_00693

work page doi:10.1162/isal_a_00693

[11] [11]

Eleni Nisioti, Erwan Plantec, Milton Montero, Joachim Winther Pedersen, and Sebastian Risi. 2024. Growing Artificial Neural Networks for Control: the Role of Neuronal Diversity. InGECCO ’24 Companion: Proceedings of the Genetic and Evolutionary Computation Conference Companion. ACM, 175–178. doi:10.1145/ 3638530.3654356

work page arXiv 2024

[12] [12]

Rasmus Berg Palm, Elias Najarro, and Sebastian Risi. 2021. Testing the Genomic Bottleneck Hypothesis in Hebbian Meta-Learning. InProceedings of Machine Learning Research, Vol. 148. 100–110

work page 2021

[13] [13]

Erwan Plantec, Joachim Winther Pedersen, Milton Llera Montero, Eleni Nisioti, and Sebastian Risi. 2024. Evolving Self-Assembling Neural Networks: From Spontaneous Activity to Experience-Dependent Learning. InProceedings of the 2024 Conference on Artificial Life. MIT Press

work page 2024

[14] [14]

Sergey Shuvaev, Divyansha Lachi, Alexei Koulakov, and Anthony Zador. 2024. Encoding innate ability through a genomic bottleneck.Proceedings of the National Academy of Sciences121, 38 (2024), e2409160121. doi:10.1073/pnas.2409160121

work page doi:10.1073/pnas.2409160121 2024

[15] [15]

Andrea Soltoggio, Kenneth O Stanley, and Sebastian Risi. 2018. Born to learn: the inspiration, progress, and future of evolved plastic artificial neural networks. Neural Networks108 (2018), 48–67. doi:10.1016/j.neunet.2018.07.013

work page doi:10.1016/j.neunet.2018.07.013 2018

[16] [16]

Kenneth O Stanley, David B D’Ambrosio, and Jason Gauci. 2009. A Hypercube- Based Encoding for Evolving Large-Scale Neural Networks.Artificial Life15, 2 (2009), 185–212

work page 2009

[17] [17]

Kenneth O Stanley and Risto Miikkulainen. 2002. Evolving Neural Networks through Augmenting Topologies.Evolutionary Computation10, 2 (2002), 99–127

work page 2002

[18] [18]

2024.Gymnasium

Mark Towers, Ariel Kwiatkowski, Jordan Terry, John U Balis, Gianluca De Cola, Tristan Deleu, Manuel Goulão, Andreas Kallinteris, Arjun KG, Markus Krimmel, Rodrigo Perez-Vicente, Andrea Pierré, Sander Schulhoff, Jun Jet Tai, Andrew Tan, and Omar G Younis. 2024.Gymnasium. https://github.com/Farama-Foundation/ Gymnasium

work page 2024

[19] [19]

Alan M Turing. 1952. The Chemical Basis of Morphogenesis.Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences237, 641 (1952), 37–72. doi:10.1098/rstb.1952.0012

work page doi:10.1098/rstb.1952.0012 1952

[20] [20]

Anthony M Zador. 2019. A critique of pure learning and what artificial neural networks can learn from animal brains.Nature Communications10, 1 (2019),

work page 2019

[21] [21]

doi:10.1038/s41467-019-11786-6

work page doi:10.1038/s41467-019-11786-6