pith. sign in

arxiv: 1907.09209 · v1 · pith:A5MIQY52new · submitted 2019-07-22 · 💻 cs.NE · cs.AI· q-bio.NC

Automatic Calibration of Artificial Neural Networks for Zebrafish Collective Behaviours using a Quality Diversity Algorithm

Leo Cazenille , Nicolas Bredeche , Jos\'e Halloy This is my paper

Pith reviewed 2026-05-24 17:57 UTC · model grok-4.3

classification 💻 cs.NE cs.AIq-bio.NC
keywords zebrafishcollective motionquality diversityMAP-Elitesneural network controllersagent-based modelbiomimetic simulationevolutionary calibration
0
0 comments X

The pith

Quality Diversity algorithms calibrate neural networks to match real zebrafish collective motion better than standard evolutionary methods.

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

The work seeks an agent-based model of zebrafish groups that can be realized on robotic fish capable of mixing undetected with live schools. It tunes the parameters of artificial neural network controllers for each simulated fish by treating the calibration task as a quality-diversity search rather than a pure optimization problem. The chosen algorithm, CVT-MAP-Elites, maintains an archive of solutions that differ in behavioral descriptors yet all achieve high fidelity to recorded trajectories. When compared with classic evolutionary reinforcement learning, the quality-diversity approach yields closer agreement with empirical data at both the scale of entire groups and the scale of single-fish paths.

Core claim

Applying CVT-MAP-Elites to evolve the weights of neural controllers for simulated zebrafish produces collective trajectories whose macroscopic statistics and whose microscopic individual statistics both align more closely with real tracking data than the trajectories obtained by conventional evolutionary reinforcement learning.

What carries the argument

CVT-MAP-Elites, a quality-diversity algorithm that fills a tessellated behavior space with locally optimal neural-network controllers whose performance is measured against real zebrafish trajectories.

If this is right

  • The calibrated models become candidates for direct transfer onto robotic fish intended to blend into live collectives.
  • Improved individual-level realism supplies a more faithful test bed for hypotheses about the biological rules of social interaction.
  • The same calibration pipeline can be applied to larger group sizes or to environments with obstacles once the behavior space is suitably redefined.
  • Quality-diversity archives supply a library of controllers that can be swapped or combined to explore how different individual rules affect group outcomes.

Where Pith is reading between the lines

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

  • The same quality-diversity calibration loop could be reused for other schooling species whose motion is captured by comparable tracking systems.
  • If the resulting controllers prove robust on robots, hybrid experiments become feasible in which real and simulated fish interact under controlled perturbations.
  • The approach suggests that many other parameter-fitting tasks in collective-behavior modeling may benefit from illumination of diverse solutions rather than convergence to a single optimum.

Load-bearing premise

The chosen behavioral descriptors and fitness functions remain valid when the same neural controllers are placed in new environments or on physical robots.

What would settle it

A new, independent set of zebrafish tracking recordings on which the best QD-generated models produce statistically worse matches than the best models from classic evolutionary reinforcement learning.

Figures

Figures reproduced from arXiv: 1907.09209 by Jos\'e Halloy, Leo Cazenille, Nicolas Bredeche.

Figure 1
Figure 1. Figure 1: Description of the presented methodology to calibrate artificial neural networks to generate fish trajectories. We apply CVT-MAP [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Agent trajectories in the square (1m) experimental arena after 30-minute trials, for all considered cases: Control reference experimental fish data obtained as in [9], [29], CVT-MAP-Elites and CMA-ES corresponding to simulated MLP-driven agents. A Examples of an individual trajectory of one agent among the 5 making the group (fish or simulated agent) during 1 minute out of a 30-minute trial. B Presence pro… view at source ↗
Figure 3
Figure 3. Figure 3: The distributions of angular speed (Fig. 4C) of [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 3
Figure 3. Figure 3: Similarity scores between the trajectories of the experimental fish groups (Control) and those of the best-performing simulated [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Behavioural comparison between ten 30-minute trials of experimental fish in groups of 5 and the 5-sized simulated fish groups for both [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
read the original abstract

During the last two decades, various models have been proposed for fish collective motion. These models are mainly developed to decipher the biological mechanisms of social interaction between animals. They consider very simple homogeneous unbounded environments and it is not clear that they can simulate accurately the collective trajectories. Moreover when the models are more accurate, the question of their scalability to either larger groups or more elaborate environments remains open. This study deals with learning how to simulate realistic collective motion of collective of zebrafish, using real-world tracking data. The objective is to devise an agent-based model that can be implemented on an artificial robotic fish that can blend into a collective of real fish. We present a novel approach that uses Quality Diversity algorithms, a class of algorithms that emphasise exploration over pure optimisation. In particular, we use CVT-MAP-Elites, a variant of the state-of-the-art MAP-Elites algorithm for high dimensional search space. Results show that Quality Diversity algorithms not only outperform classic evolutionary reinforcement learning methods at the macroscopic level (i.e. group behaviour), but are also able to generate more realistic biomimetic behaviours at the microscopic level (i.e. individual behaviour).

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.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes using Quality Diversity algorithms (specifically CVT-MAP-Elites) to calibrate an agent-based model of zebrafish collective motion against real-world tracking data. The goal is to produce simulations realistic enough for implementation on robotic fish that can blend into live groups. The central claim is that QD methods outperform classic evolutionary reinforcement learning approaches at both the macroscopic (group) level and the microscopic (individual) level.

Significance. If the performance claims hold under independent validation, the work would demonstrate a practical advantage of quality-diversity search over pure optimization for high-dimensional calibration of collective-behavior models, with direct relevance to scalable biomimetic robotics. The data-driven calibration against external recordings is a positive feature.

major comments (2)
  1. [Abstract] Abstract: the claim of comparative superiority at both macroscopic and microscopic levels is stated without any quantitative metrics, error bars, dataset sizes, or validation protocol, preventing verification of the central claim against the paper's own data.
  2. [Methods/Results] Methods/Results: the manuscript does not demonstrate that the microscopic evaluation metrics (individual speed distributions, turning angles, nearest-neighbor statistics) are computed from features disjoint from the behavioral descriptors used to populate the CVT-MAP-Elites archive. If the same statistics are reused, reported micro-level gains become an artifact of archive construction rather than independent evidence of improved biomimetic fidelity.
minor comments (1)
  1. [Abstract] Abstract: phrasing such as 'collective of zebrafish' and 'collective trajectories' should be revised for grammatical clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the opportunity to respond to the referee's report. We address each major comment below, indicating where we agree revisions are warranted and providing clarifications where the manuscript already supports the claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim of comparative superiority at both macroscopic and microscopic levels is stated without any quantitative metrics, error bars, dataset sizes, or validation protocol, preventing verification of the central claim against the paper's own data.

    Authors: We agree that the abstract presents the comparative claims in qualitative terms only. In the revised manuscript we will expand the abstract to include key quantitative results drawn from the experiments (e.g., mean and standard-deviation improvements on group-level metrics such as polarization and nearest-neighbor distance error, plus individual-level distribution matches), the number of tracked trajectories in the dataset, and a concise statement of the validation protocol (training on a subset of recordings with evaluation on held-out segments). This will allow direct verification of the central claim from the abstract itself. revision: yes

  2. Referee: [Methods/Results] Methods/Results: the manuscript does not demonstrate that the microscopic evaluation metrics (individual speed distributions, turning angles, nearest-neighbor statistics) are computed from features disjoint from the behavioral descriptors used to populate the CVT-MAP-Elites archive. If the same statistics are reused, reported micro-level gains become an artifact of archive construction rather than independent evidence of improved biomimetic fidelity.

    Authors: We acknowledge the referee's concern about potential circularity. The CVT-MAP-Elites behavioral descriptors are constructed from binned statistics of speed, turning angle and nearest-neighbor distance precisely to encourage coverage of the feature space. The microscopic evaluation, however, consists of a separate post-selection comparison of the full empirical distributions produced by the calibrated agents against the real tracking data, using quantitative discrepancy measures (e.g., Earth Mover's Distance) that are not part of the archive-filling objective. To remove any ambiguity we will add an explicit paragraph in the Methods section that (i) distinguishes descriptor computation for diversity maintenance from the fidelity assessment against held-out real data and (ii) reports results on at least one additional microscopic metric (e.g., long-term inter-individual distance autocorrelation) that is not used as a descriptor, thereby providing independent corroboration of the micro-level improvements. revision: yes

Circularity Check

0 steps flagged

No circularity: data-driven calibration against external tracking data

full rationale

The paper performs calibration of neural network controllers for zebrafish simulation by optimizing against real-world tracking recordings using CVT-MAP-Elites. All reported performance metrics (macroscopic group behavior and microscopic individual trajectories) are computed directly from comparison to this independent external dataset. No derivation step reduces a claimed result to a fitted parameter or self-citation by construction; the behavioral descriptors and fitness functions are applied to held-out or validation trajectories rather than being tautological with the optimization targets. The central claims therefore rest on external empirical benchmarks rather than internal redefinition.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract provides no explicit free parameters, invented entities, or non-standard axioms; the work rests on the standard domain assumption that collective motion can be approximated by local agent rules whose parameters can be tuned against trajectory data.

axioms (1)
  • domain assumption Agent-based models with local neural controllers can reproduce observed collective trajectories when parameters are appropriately chosen.
    Invoked implicitly by the decision to calibrate an ANN controller to real tracking data.

pith-pipeline@v0.9.0 · 5745 in / 1305 out tokens · 28670 ms · 2026-05-24T17:57:17.907694+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 36 canonical work pages · 4 internal anchors

  1. [1]

    In: Evolutionary Computation,

    Auger, A., Hansen, N.: A restart CMA evolution strategy with increasing population size. In: Evolutionary Computation,

    work page

  2. [2]

    The 2005 IEEE Congress on. vol. 2, pp. 1769–1776. IEEE (2005)

    work page 2005

  3. [3]

    Springer-Verlag, Berlin, Heidel- berg (2006)

    Bishop, C.: Pattern Recognition and Machine Learning (Infor- mation Science and Statistics). Springer-Verlag, Berlin, Heidel- berg (2006)

    work page 2006

  4. [4]

    PLoS com- putational biology 14(1), e1005933 (2018)

    Calovi, D.S., Litchinko, A., Lecheval, V ., Lopez, U., Escudero, A.P., Chat ´e, H., Sire, C., Theraulaz, G.: Disentangling and modeling interactions in fish with burst-and-coast swimming reveal distinct alignment and attraction behaviors. PLoS com- putational biology 14(1), e1005933 (2018)

    work page 2018

  5. [5]

    https://gitlab.com/leo.cazenille/qdpy (2018)

    Cazenille, L.: Qdpy: A python framework for quality-diversity. https://gitlab.com/leo.cazenille/qdpy (2018)

    work page 2018

  6. [6]

    In: Conference on biomimetic and biohybrid systems

    Cazenille, L., Chemtob, Y ., Bonnet, F., Gribovskiy, A., Mon- dada, F., Bredeche, N., Halloy, J.: Automated calibration of a biomimetic space-dependent model for zebrafish and robot col- lective behaviour in a structured environment. In: Conference on biomimetic and biohybrid systems. pp. 107–118. Springer (2017)

    work page 2017

  7. [7]

    Cazenille, L., Collignon, B., Bonnet, F., Gribovskiy, A., Mon- dada, F., Bredeche, N., Halloy, J.: How mimetic should a robotic fish be to socially integrate into zebrafish groups ? Bioinspiration & biomimetics (2017)

    work page 2017

  8. [8]

    In: Conference on Biomimetic and Biohybrid Systems

    Cazenille, L., Bredeche, N., Halloy, J.: Evolutionary optimi- sation of neural network models for fish collective behaviours in mixed groups of robots and zebrafish. In: Conference on Biomimetic and Biohybrid Systems. pp. 85–96. Springer (2018)

    work page 2018

  9. [9]

    Modelling zebrafish collective behaviours with multilayer perceptrons optimised by evolutionary algorithms

    Cazenille, L., Bredeche, N., Halloy, J.: Modelling zebrafish collective behaviours with multilayer perceptrons optimised by evolutionary algorithms. arXiv preprint arXiv:1811.11040 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  10. [10]

    Royal Society Open Science 3(1), 150473 (2016)

    Collignon, B., S ´eguret, A., Halloy, J.: A stochastic vision- based model inspired by zebrafish collective behaviour in heterogeneous environments. Royal Society Open Science 3(1), 150473 (2016)

    work page 2016

  11. [11]

    Collective departures in zebrafish: profiling the initiators

    Collignon, B., S ´eguret, A., Chemtob, Y ., Cazenille, L., Halloy, J.: Collective departures in zebrafish: profiling the initiators. arXiv preprint arXiv:1701.03611 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  12. [12]

    Nature 521(7553), 503 (2015)

    Cully, A., Clune, J., Tarapore, D., Mouret, J.B.: Robots that can adapt like animals. Nature 521(7553), 503 (2015)

    work page 2015

  13. [13]

    IEEE Transactions on Evolution- ary Computation 22(2), 245–259 (2018)

    Cully, A., Demiris, Y .: Quality and diversity optimization: A unifying modular framework. IEEE Transactions on Evolution- ary Computation 22(2), 245–259 (2018)

    work page 2018

  14. [14]

    Interface Focus 2(6), 689 (2012)

    Deutsch, A., Theraulaz, G., Vicsek, T.: Collective motion in biological systems. Interface Focus 2(6), 689 (2012)

    work page 2012

  15. [15]

    Elsevier (2006)

    Deza, M., Deza, E.: Dictionary of distances. Elsevier (2006)

    work page 2006

  16. [16]

    IEEE Transactions on Evolutionary Com- putation 22(2), 314–328 (2018)

    Duarte, M., Gomes, J., Oliveira, S.M., Christensen, A.L.: Evolution of repertoire-based control for robots with complex locomotor systems. IEEE Transactions on Evolutionary Com- putation 22(2), 314–328 (2018)

    work page 2018

  17. [17]

    Journal of Machine Learning Research 13(Jul), 2171–2175 (2012)

    Fortin, F.A., De Rainville, F.M., Gardner, M.A., Parizeau, M., Gagn´e, C.: Deap: Evolutionary algorithms made easy. Journal of Machine Learning Research 13(Jul), 2171–2175 (2012)

    work page 2012

  18. [18]

    bioRxiv p

    Heras, F.J., Romero-Ferrero, F., Hinz, R.C., de Polavieja, G.G.: Aggregation rule in animal collectives dynamically changes between majority and minority influence. bioRxiv p. 400747 (2018)

    work page 2018

  19. [19]

    bioRxiv p

    Heras, F.J., Romero-Ferrero, F., Hinz, R.C., de Polavieja, G.G.: Deep attention networks reveal the rules of collective motion in zebrafish. bioRxiv p. 400747 (2018)

    work page 2018

  20. [20]

    Biology letters 11(12), 20150674 (2015)

    Herbert-Read, J.E., Romenskyy, M., Sumpter, D.J.: A turing test for collective motion. Biology letters 11(12), 20150674 (2015)

    work page 2015

  21. [21]

    In: Artificial Life Conference Proceedings

    Iizuka, H., Nakamoto, Y ., Yamamoto, M.: Learning of individ- ual sensorimotor mapping to form swarm behavior from real fish data. In: Artificial Life Conference Proceedings. pp. 179–

    work page

  22. [22]

    Journal of Theoretical Biology 225(4), 443–451 (2003)

    Jeanson, R., Blanco, S., Fournier, R., Deneubourg, J., Four- cassi´e, V ., Theraulaz, G.: A model of animal movements in a bounded space. Journal of Theoretical Biology 225(4), 443–451 (2003)

    work page 2003

  23. [23]

    PLoS computational biology 13(11), e1005822 (2017)

    Jiang, L., Giuggioli, L., Perna, A., Escobedo, R., Lecheval, V ., Sire, C., Han, Z., Theraulaz, G.: Identifying influential neighbors in animal flocking. PLoS computational biology 13(11), e1005822 (2017)

    work page 2017

  24. [24]

    In: Proceedings of the 15th annual conference on Genetic and evolutionary computation

    Lehman, J., Stanley, K.O., Miikkulainen, R.: Effective diversity maintenance in deceptive domains. In: Proceedings of the 15th annual conference on Genetic and evolutionary computation. pp. 215–222. ACM (2013)

    work page 2013

  25. [25]

    Interface focus 2(6), 693–707 (2012)

    Lopez, U., Gautrais, J., Couzin, I.D., Theraulaz, G.: From behavioural analyses to models of collective motion in fish schools. Interface focus 2(6), 693–707 (2012)

    work page 2012

  26. [26]

    Illuminating search spaces by mapping elites

    Mouret, J.B., Clune, J.: Illuminating search spaces by mapping elites. arXiv preprint arXiv:1504.04909 (2015)

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  27. [27]

    Advanced textbooks in control and signal processing

    Norgaard, M., Ravn, O., Poulsen, N., Hansen, L.: Neural networks for modelling and control of dynamic systems: a practitioner’s handbook. Advanced textbooks in control and signal processing. Springer, Berlin (2000)

    work page 2000

  28. [28]

    Frontiers in Robotics and AI 3, 40 (2016)

    Pugh, J.K., Soros, L.B., Stanley, K.O.: Quality diversity: A new frontier for evolutionary computation. Frontiers in Robotics and AI 3, 40 (2016)

    work page 2016

  29. [29]

    Computer Graphics 21, 25–34 (1987)

    Reynolds, C.: Flocks, herds, and schools: a distributed behav- ioral model. Computer Graphics 21, 25–34 (1987)

    work page 1987

  30. [30]

    Loose social organisation of AB strain zebrafish groups in a two-patch environment

    S ´eguret, A., Collignon, B., Cazenille, L., Chemtob, Y ., Halloy, J.: Loose social organisation of ab strain zebrafish groups in a two-patch environment. arXiv preprint arXiv:1701.02572 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  31. [31]

    Interface focus 2(6), 764–773 (2012)

    Sumpter, D.J., Mann, R.P., Perna, A.: The modelling cycle for collective animal behaviour. Interface focus 2(6), 764–773 (2012)

    work page 2012

  32. [32]

    Sutton, R.S., Barto, A.G.: Reinforcement learning: An intro- duction (2018)

    work page 2018

  33. [33]

    IEEE Transactions on Evolutionary Computation 22(4), 623–630 (2018)

    Vassiliades, V ., Chatzilygeroudis, K., Mouret, J.B.: Using cen- troidal voronoi tessellations to scale up the multidimensional archive of phenotypic elites algorithm. IEEE Transactions on Evolutionary Computation 22(4), 623–630 (2018)

    work page 2018

  34. [34]

    Physical review letters 75(6), 1226 (1995)

    Vicsek, T., Czir ´ok, A., Ben-Jacob, E., Cohen, I., Shochet, O.: Novel type of phase transition in a system of self-driven particles. Physical review letters 75(6), 1226 (1995)

    work page 1995

  35. [35]

    In: Reinforcement learning, pp

    Whiteson, S.: Evolutionary computation for reinforcement learning. In: Reinforcement learning, pp. 325–355. Springer (2012)

    work page 2012

  36. [36]

    In: Proceedings of the 2014 Annual Conference on Genetic and Evolutionary Computation

    Yuan, Y ., Xu, H., Wang, B.: An improved nsga-iii procedure for evolutionary many-objective optimization. In: Proceedings of the 2014 Annual Conference on Genetic and Evolutionary Computation. pp. 661–668. ACM (2014)

    work page 2014