arxiv: 2604.22000 · v2 · submitted 2026-04-23 · 💻 cs.NE

Recognition: unknown

L-System Genetic Encoding for Scalable Neural Network Evolution: A Comparison with Direct Matrix Encoding

Alexander Stuy, Nodin Weddington

Authors on Pith no claims yet

Pith reviewed 2026-05-08 12:49 UTC · model grok-4.3

classification 💻 cs.NE

keywords L-Systemgenetic encodingneuroevolutionneural network topologygenetic algorithmsHebbian learningartificial lifegeneralization

0 comments

The pith

L-System genetic encoding for neural networks produces faster convergence, higher performance, and better generalization than direct matrix encoding.

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

The paper introduces Lsys, an L-System formal grammar for representing neural network topologies, and tests it against direct matrix encoding when genetic algorithms evolve Hebbian networks to navigate an artificial world of barriers, plains, and food. In 24 runs, Lsys populations collected 2.74 times more food on average at generation 1000, succeeded in every run while matrix failed in half, and transferred to a new maze environment with a 5.82 times performance advantage. A control shows the gain comes from the compressed symbolic alphabet operating throughout evolution rather than from the starting population alone. A sympathetic reader would care because the result points to a way of scaling evolutionary design of neural controllers without hand-designed topologies or prior knowledge of the task.

Core claim

Lsys encoding is shown to provide faster convergence, higher peak performance, dramatically greater reliability, and superior generalization to novel environments compared to Matrix encoding across all experimental conditions tested. In the training world Lsys reached a mean maximum food count of 3802 with low variance while matrix reached 1388 with high variance; when moved to an unseen maze the gap widened to 2455 versus 422. The MatrixLSG control, which seeded matrix evolution with Lsys-generated initial populations, performed like standard matrix runs, confirming that the advantage lies in the L-System operators acting on the symbolic alphabet for the full duration of evolution.

What carries the argument

Lsys, a formal L-System based genetic alphabet that encodes neural network topologies as compressed symbolic strings on which genetic operators act directly.

If this is right

All Lsys populations learn to navigate successfully while half the matrix populations remain at low performance.
Lsys populations immediately collect several times more food when placed in a novel maze without further evolution.
The performance gap is produced by the genetic algorithm operating on the L-System alphabet throughout evolution rather than by the structure of the initial population.
Lsys encoding yields both a higher mean and an 8.5-fold smaller coefficient of variation in final food counts.

Where Pith is reading between the lines

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

If the compression and regularity-capturing properties of L-Systems extend to larger networks, they could make evolutionary search feasible for networks whose direct matrix representations would be too large to evolve directly.
The same symbolic alphabet might be applied to evolve controllers for physical robots or other sequential decision tasks where topology must be discovered rather than specified in advance.
Replacing the Hebbian rule with other local learning mechanisms could test whether the Lsys advantage is tied to the particular plasticity rule used here.

Load-bearing premise

The specific artificial world of barriers, plains, and food together with Hebbian learning rules forms a representative testbed whose performance differences will appear in other domains or network types.

What would settle it

Finding a different navigation task, learning rule, or network size in which Lsys populations no longer show both higher peak scores and lower failure rates than matrix populations after 1000 generations.

Figures

Figures reproduced from arXiv: 2604.22000 by Alexander Stuy, Nodin Weddington.

**Figure 1.** Figure 1: Biological Neuron. 2.1.2 Artificial Neural Networks Like a BNN an artificial neural network (ANN) consists of one or more interconnected neurons. The neurons in an ANN can be described mathematically view at source ↗

read the original abstract

An artificial world of barriers and plains scattered with food is used to test the feasibility of using genetic algorithms to optimize hebbian neural networks to perform on problems without apriori knowledge of the problem domain. A formal L-System based genetic alphabet for neural networks, titled Lsys, and a neural network genetic modeling tool titled Wp1hgn are introduced. Lsys and Matrix neural network topology genetic encoding methods are compared across 24 experimental runs. Lsys encoding achieved a mean maximum food count of 3802 +- 197 at generation 1000 across 8 runs with varied parameters, compared to 1388 +- 610 for Matrix encoding, a 2.74x performance advantage with an 8.5-fold improvement in consistency as measured by coefficient of variation (5.2% vs 44.0%). All 8 Lsys populations successfully learned to navigate the environment, while 4 of 8 Matrix populations failed to achieve competitive performance at any point during 1000 generations. When transferred to a novel maze environment, Lsys populations demonstrated immediate robust generalization, achieving a mean maximum food count of 2455 +- 176 compared to 422 +- 212 for Matrix populations, a 5.82x advantage that exceeded the training world performance gap. A MatrixLSG control condition, in which initial populations were generated using Lsys genotypes and then evolved using Matrix operators, demonstrated that the performance advantage of Lsys encoding derives primarily from the genetic algorithm operating on the compressed symbolic Lsys alphabet throughout evolution rather than from initial population structure. Lsys encoding is shown to provide faster convergence, higher peak performance, dramatically greater reliability, and superior generalization to novel environments compared to Matrix encoding across all experimental conditions tested.

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

Lsys encoding beats matrix on reliability and generalization in this navigation task thanks to a control that isolates the operators, but thin methods details limit how far the numbers can be trusted.

read the letter

The main takeaway is that Lsys encoding produces faster convergence, higher peak performance, far better run-to-run consistency, and stronger transfer to a new maze than direct matrix encoding in this food-gathering setup with Hebbian networks. All eight Lsys populations succeeded while four matrix ones failed entirely, and the generalization gap was even larger than the training gap. A MatrixLSG control shows the advantage comes from running the genetic algorithm on the compressed Lsys alphabet throughout evolution rather than from initial population structure alone. That control is a clear strength and addresses a reasonable alternative explanation. The work introduces a formal L-system genetic alphabet and the Wp1hgn modeling tool, then runs a direct comparison with quantitative reporting of means, standard deviations, coefficients of variation, and success counts across 24 total runs. The generalization test to a novel environment adds useful evidence that the evolved behaviors are not just overfit to the training world. The soft spots are the modest scale and missing specifics. Eight runs per condition is low for claims about dramatically greater reliability. No statistical tests are described to support the differences. The abstract and available summary give no numbers on network sizes, the exact L-system production rules, or how food counts were tallied, which makes independent verification difficult. Results are tied to one artificial world and Hebbian learning, so it is unclear whether the encoding advantage would appear in other tasks or with different network types. This is for researchers working on neuroevolution and alternative genetic representations for network topologies. A reader already thinking about encodings for evolving connectivity would get concrete value from the head-to-head and the control. It deserves a serious referee because the central empirical claim rests on a new method plus a control that isolates the key factor, even if the scope stays narrow and some protocol details need filling in. Send it for peer review.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces an L-System-based genetic encoding (Lsys) for neural network topologies, along with a modeling tool Wp1hgn, and compares it empirically to direct Matrix encoding for evolving Hebbian networks in an artificial foraging environment of barriers, plains, and food. Across 24 runs (8 per condition), it reports that Lsys yields a 2.74x higher mean maximum food count (3802 ± 197 vs 1388 ± 610 at generation 1000), an 8.5-fold better consistency (CV 5.2% vs 44.0%), 100% success rate versus 50% for Matrix, and a 5.82x advantage in generalization to a novel maze (2455 ± 176 vs 422 ± 212); a MatrixLSG control isolates the advantage to the Lsys operators during evolution rather than initial population structure.

Significance. If the performance differences are reproducible, the work offers a concrete demonstration that compressed symbolic encodings can mitigate the scalability limitations of direct matrix encodings in neuroevolution, with the control condition providing a useful isolation of operator effects. The reporting of means, standard deviations, coefficients of variation, and success/failure counts strengthens the empirical case within the described testbed.

major comments (3)

[Abstract and Results] Abstract and Results sections: the reported means, standard deviations, success rates, and generalization metrics are presented without any description of neural network sizes (number of neurons or connections), the precise L-system alphabet and production rules, or the exact procedure for measuring food counts, all of which are load-bearing for verifying the 2.74x and 5.82x performance claims.
[Methods and Experimental Setup] Methods and Experimental Setup: no details are given on the statistical tests used to compare conditions, the number of independent evaluations per run, or how the Hebbian learning rules and environment dynamics (barrier and food placement) were implemented, preventing assessment of whether the observed reliability and generalization differences are statistically supported.
[Generalization experiments] Generalization experiments: the novel maze is described only as 'novel' without specifying its structure, size, or differences from the training world, and it is unclear whether networks were evaluated zero-shot or with any continued learning, which directly affects the strength of the 'immediate robust generalization' conclusion.

minor comments (2)

[Abstract] The tool name 'Wp1hgn' is introduced in the abstract without any subsequent definition or reference; a brief description of its purpose and availability would improve clarity.
[Experimental Setup] The manuscript would benefit from a table summarizing the exact genetic algorithm hyperparameters (population size, mutation rates, generations) used across the 24 runs, as these are listed as free parameters but not enumerated.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments, which highlight areas where additional detail will strengthen the manuscript. We agree that the reported performance advantages require more supporting context for full verifiability and will revise the paper to address each point.

read point-by-point responses

Referee: [Abstract and Results] Abstract and Results sections: the reported means, standard deviations, success rates, and generalization metrics are presented without any description of neural network sizes (number of neurons or connections), the precise L-system alphabet and production rules, or the exact procedure for measuring food counts, all of which are load-bearing for verifying the 2.74x and 5.82x performance claims.

Authors: We agree that these elements are important for verifying the claims. The Methods section contains the neural network sizes, L-system alphabet and production rules, and food-count measurement procedure. In the revision we will add concise summaries of these items to both the Abstract and Results sections so that readers can assess the 2.74x and 5.82x figures without first consulting the full Methods. revision: yes
Referee: [Methods and Experimental Setup] Methods and Experimental Setup: no details are given on the statistical tests used to compare conditions, the number of independent evaluations per run, or how the Hebbian learning rules and environment dynamics (barrier and food placement) were implemented, preventing assessment of whether the observed reliability and generalization differences are statistically supported.

Authors: We acknowledge the omission. The revised Methods section will explicitly state the statistical tests used to compare conditions, the number of independent evaluations performed per run, and the precise implementation of the Hebbian learning rules together with the barrier and food placement mechanics. These additions will allow readers to evaluate the statistical support for the reported reliability and generalization differences. revision: yes
Referee: [Generalization experiments] Generalization experiments: the novel maze is described only as 'novel' without specifying its structure, size, or differences from the training world, and it is unclear whether networks were evaluated zero-shot or with any continued learning, which directly affects the strength of the 'immediate robust generalization' conclusion.

Authors: We agree that greater specificity is required. The revision will describe the novel maze's structure, size, and differences from the training environment, and will state that transfer was performed zero-shot with no additional learning or evolution. This clarification will support the interpretation of immediate robust generalization. revision: yes

Circularity Check

0 steps flagged

No significant circularity; results are direct empirical comparisons

full rationale

The manuscript reports performance metrics from 24 independent evolutionary runs (8 per condition) in a fixed artificial environment using Hebbian networks. All central claims—faster convergence, higher peak food counts, reliability, and generalization—are quantified directly from simulation outcomes (means, standard deviations, success rates) with an explicit MatrixLSG control that isolates the contribution of Lsys operators. No equations, derivations, fitted parameters renamed as predictions, or self-citation chains appear in the load-bearing argument; the advantages are measured outputs rather than constructed by definition or prior self-reference.

Axiom & Free-Parameter Ledger

1 free parameters · 0 axioms · 1 invented entities

The central claim rests on the validity of the simulated environment and the specific L-system rules used to generate networks; these are introduced in the paper without external benchmarks or independent verification.

free parameters (1)

Genetic algorithm hyperparameters (population size, mutation rates, generations)
Required to reproduce the 1000-generation runs and performance numbers but not specified in the abstract.

invented entities (1)

Lsys genetic alphabet no independent evidence
purpose: Compressed symbolic representation that grows into neural network topologies via L-system rewriting rules
New encoding introduced by the paper to enable more scalable evolution.

pith-pipeline@v0.9.0 · 5621 in / 1296 out tokens · 71507 ms · 2026-05-08T12:49:23.755837+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

17 extracted references · 1 canonical work pages

[1]

Anderson and E

J.A. Anderson and E. Rosenfeld, eds. Neurocomputing: Foundations of Research. Cambridge: MIT Press, 1988. 17

1988
[2]

A.K. Dewdney. Computer Recreations: Exploring the field of genetic algorithms in a primordial computer sea full of flibs. Scientific American (Vol. 253, pp. 21-32) November 1985

1985
[3]

Genetics-Based Machine Learning and Behavior-Based Robotics: A New Synthesis

Marco Dorigo and Uwe Schnepf. Genetics-Based Machine Learning and Behavior-Based Robotics: A New Synthesis. IEEE Transactions on Systems, Man, and Cybernetics, Vol. 23, No. 1, January/February 1993

1993
[4]

Fischbach

Gerald D. Fischbach. Mind and Brain. Scientific American, pages 24-33, September 1992

1992
[5]

Fogel, L.J

D.B. Fogel, L.J. Fogel, and V.W. Porto. Evolving Neural Networks. Biological Cybernetics, Sprint 1990., pages 487-493

1990
[6]

John Hertz, Anders Krough, Richard G. Palmer. Introduction to the Theory of Neural Computation. Addison-Wesley Publishing Company, 1991

1991
[7]

D.O. Hebb. The Organization of Behaviour. New York: Wiley. Partially reprinted in Anderson and Rosenfeld [1988]

1988
[8]

Goldberg

David E. Goldberg. Genetic Algorithms in Search, Optimization, and Machine Learning. Addison-Wesley Publishing Company, Inc
[9]

Hinton and Steven J

Geoffrey E. Hinton and Steven J. Nowlan. How Learning Can Guide Evolution. Complex Systems 1 (1987) pages 495-502

1987
[10]

Designing Neural Networks Using Genetic Algorithms with Graph Generation System

Hiroaki Kitano. Designing Neural Networks Using Genetic Algorithms with Graph Generation System. Complex Systems 4 (1990), pages 461-476

1990
[11]

Muhlenbein and J

H. Muhlenbein and J. Kindermann. The Dynamics of Evolution and Learning-Towards Genetic Neural Networks. Connectionism in Perspective, Elsevier Science Publishers B.V (North-Holland), 1989, pages 173-197

1989
[12]

Norfi and D

S. Norfi and D. Parisi. Auto-Teaching: Networks that Develop their own Teaching Input. Proceedings of the Second European Conference on Artificial Life (ECAL '93) May 24, 1993 - May 26, 1993

1993
[13]

Pettit and Michael J

Elaine J. Pettit and Michael J. Pettit. Analysis of the Performance of a Genetic Algorithm Based System for Message Classification in Noisy Environments. Int. J. Man-Machine Studies (1987) 27, pages 205-220

1987
[14]

Prusinkiewicz and A

P. Prusinkiewicz and A. Lindenmayer. The Algorithmic Beauty of Plants. Springer Verlaag Inc., New York 1990

1990
[15]

L-Systems

Grzegorz Rozenberg. L-Systems. Springer Verlaag Inc., New York 1974

1974
[16]

Genetically-trained deep neural networks

Krzysztof Pawełczyk, Michal Kawulok, Jakub Nalepa. Genetically-trained deep neural networks. GECCO '18: Proceedings of the Genetic and Evolutionary Computation Conference Companion Pages 63 – 64 https://doi.org/10.1145/3205651.3208763

work page doi:10.1145/3205651.3208763
[17]

A., & Krichmar, J

Ascoli, G. A., & Krichmar, J. L. (2000). L-neuron: A modeling tool for the efficient generation and parsimonious description of dendritic morphology. Neurocomputing, 32, 1013-1019

2000