BioelectricaAlterspan Labs
Evolution and evolvability  ·  Article

Evolutionary algorithms and evolved neural network controllers

An evolutionary algorithm is a heuristic optimization method that keeps a population of candidate solutions and raises a fitness score over generations.8 It has been used to find regulatory networks, chemical networks, robot designs and neural cellular automaton parameters.217 In one study of chemical networks, most evolved reactions were superfluous, and smaller networks emerged even though simplicity was not selected for explicitly.1

Earliest held
2012, McGregor S, Vasas V, Husbands…
Most discussed in
Evolutionary Implications of Self-Assembling…, 2024
In the library
79 passages in 14 works
Rewritten
2026-10-30
18501875190019251950197520002025McGregor S, Vasas V, Husbands P…, 20121Lobo D, Levin M, 20152Lobo D, Levin M, 20153Kriegman S, Cheney N, Bongard J, 20184Manicka S, Levin M, 20195Kriegman S, Blackiston D, Levin M…, 20206Hartl B, Risi S, Levin M, 20247Pio-Lopez L, Hartl B, Levin M, 20258Pio-Lopez L, Hartl B, Levin M, 20259
01

How the method works

Pio-Lopez, Hartl and Levin (2025) describe evolutionary algorithms as heuristic techniques that maintain a population over successive generations to maximize an objective, called a fitness score.8 They note that the approach draws on natural selection and on DNA-based reproduction. High-fitness individuals are chosen for reproduction, and offspring arise by crossover and by occasional mutation with added noise.8 Manicka and Levin (2019) used a simpler variant, a microbial genetic algorithm. Two individuals are compared on fitness, the winner's genome passes to the loser at a recombination rate of 0.1, and the loser is then mutated and returned to the population.5

02

Evolving networks, not just weights

McGregor, Vasas, Husbands and Fernando (2012) let the algorithm add or remove chemical species and reactions. They evolved about ten networks per task. The typical evolved network had 12 reactions, but pruning showed that about 5 were enough to stay within 10% of the full network's fitness.1 They did not select explicitly for simplicity, yet smaller networks emerged. Lobo and Levin (2015) took a different course. Their search ran until a network with zero error was found, and it then continued to simplify the result because the mutation operators were biased toward simpler networks.23

03

Controllers, bodies and genomes

Kriegman, Cheney and Bongard (2018) found that the algorithm could quickly discover an actuation pattern giving slight forward movement in soft robots of any morphology. Selection could then climb an incremental path of growing legs and refining the controller.4 Kriegman, Blackiston, Levin and Bongard (2020) fed patterns common to successful deployed organisms back to the algorithm as constraints on later designs.6 Hartl, Risi and Levin (2024) split the genome of a neural cellular automaton into structural parameters, which set the initial cell state, and functional parameters, the weights and biases of each cell's network.7 Pio-Lopez and colleagues treat these parameters as the genotype and the final grid of cell types as the phenotype that selection sees.9

SourcesEach quotation was checked word for word against the passage it opens.
  1. typically only 5 reactions (mean 4.7, median 5) were necessary to achieve a fitness score within 10% of the entire network's fitness.McGregor S, Vasas V, Husbands P, Fernando C, 2012 · Evolution of associative learning in chemical networks · open at passage 17
  2. We employed an evolutionary algorithm [114] approach to search for regulatory networks, where a population of candidate networks evolve in parallel until a network with zero error is found.Lobo D, Levin M, 2015 · Inferring regulatory networks from experimental morphological phenotypes: a… · open at passage 51
  3. This extra evolutionary time is used to simplify the best network found, since the mutation operators are biased towards simpler networks.Lobo D, Levin M, 2015 · Inferring regulatory networks from experimental morphological phenotypes: a… · open at passage 54
  4. The evolutionary algorithm can rapidly discover an actuation pattern that elicits a very small amount of forward movement in these soft robots regardless of the morphology.Kriegman S, Cheney N, Bongard J, 2018 · How morphological development can guide evolution · open at passage 26
  5. The genome of the winner is transmitted to the loser at a certain “recombination rate”, which was set to 0.1.Manicka S, Levin M, 2019 · Modeling somatic computation with non-neural bioelectric networks · open at passage 59
  6. Common patterns among the successful systems are distilled down into constraints and supplied back to the evolutionary algorithmKriegman S, Blackiston D, Levin M, Bongard J, 2020 · A scalable pipeline for designing reconfigurable organisms · open at passage 10
  7. We explicitly separate the NCA parameters into a structural and a functional part. The structural parameters describe the initial cell state, and the functional parameters the weights and biases of the ANN of each agentHartl B, Risi S, Levin M, 2024 · Evolutionary Implications of Self-Assembling Cybernetic Materials with… · open at passage 11
  8. Evolutionary Algorithms (EAs) were heuristic optimization techniques designed to maintain and optimize a set, or populationPio-Lopez L, Hartl B, Levin M, 2025 · Aging as a Loss of Goal-Directedness: An Evolutionary Simulation and Analysis… · open at passage 33
  9. In this context, the NCA parameters thus correspond to the virtual organism's genotype, while the grid of final cell types represent the systemPio-Lopez L, Hartl B, Levin M, 2025 · Aging as a Loss of Goal-Directedness: An Evolutionary Simulation and Analysis… · open at passage 32
Linked ideas
Evolutionary algorithms supply the parameter search that trains the neural cellular automata described by Hartl and Pio-Lopez.
Kriegman and colleagues use an evolutionary algorithm to design reconfigurable organisms, with deployment results returned as constraints.
Lobo and Levin evolve regulatory networks to fit experimental phenotypes, so the search target is a network of this kind.
The algorithm borrows selection, reproduction and mutation from natural selection, but runs them as a simulated process.
Evolvabilityrelated to
What evolves quickly in these simulations, such as simple networks or growing legs, bears on how readily a system can be improved.
Designs for reconfigurable organisms come from an evolutionary algorithm before being built and observed.
Hartl et al. use an evolutionary algorithm on neural cellular automata to test how cells' competencies affect evolution itself; Levin argues evolution searches behavior-shaping signals, exploiting cellular intelligence.
Where it is discussedPassages matching evolutionary algorithm, evolution strategy, evolved ANN, genome comprising the ANN
2024Evolutionary Implications of Self-Assembling Cybernetic Materials with… · Hartl B, Risi S, Levin M26
2025Aging as a Loss of Goal-Directedness: An Evolutionary Simulation and Analysis… · Pio-Lopez L, Hartl B, Levin M21
2020A scalable pipeline for designing reconfigurable organisms · Kriegman S, Blackiston D, Levin M…6
2018How morphological development can guide evolution · Kriegman S, Cheney N, Bongard J5
2023The scaling of goals from cellular to anatomical homeostasis: an evolutionary… · Pio-Lopez L, Bischof J, LaPalme JV…4
2015Inferring regulatory networks from experimental morphological phenotypes: a… · Lobo D, Levin M4
2019Modeling somatic computation with non-neural bioelectric networks · Manicka S, Levin M3
2025AI-driven automated discovery tools reveal diverse behavioral competencies of… · Etcheverry M, Moulin-Frier C, Oudeyer…3
2020Phenotypes to remember: Evolutionary developmental memory capacity and… · Szilágyi A, Szabó P, Santos M…2
2024Stress sharing as cognitive glue for collective intelligences: A computational… · Shreesha L, Levin M1