Skip to content

Selection Instability

Jump to bottom
Jgocunha edited this page Mar 31, 2026 · 8 revisions

Selection Instability

Task description

The Selection Instability task evaluates whether an evolved Dynamic Neural Field (DNF) architecture can perform decision-making through competitive dynamics.

The desired behaviour is:

  1. Multiple external stimuli are presented simultaneously, inducing multiple peaks in the input field.
  2. The input field maintains a multi-peak representation during stimulus presentation.
  3. Activity is transmitted to the output field.
  4. The output field resolves competition via winner-take-all dynamics, stabilizing a single dominant peak.
  5. After stimulus offset, both fields relax back to baseline.

This task tests the emergence of selection and decision-making as a dynamical property, rather than as an explicit symbolic operation.

Highest-fitness minimal phenotype

The highest-fitness solutions typically converge to a near-minimal topology, though with greater variability than detection or memory:

  • Fields: 2 (input → output)

  • Hidden fields: usually 0, occasionally 1–2

  • Connections: typically 1 feedforward coupling, with optional additional connections

  • Dynamics:

    • Input field: Mexican-hat kernel supporting stable multi-peak activation
    • Output field: Gaussian kernel enabling competitive suppression and selection
phenotype-selection-wbg

The figure above shows:

  • Blue: activity during stimulus presentation
  • Red: activity after stimulus removal

The output field selects a single stimulus location while suppressing competitors, implementing winner-take-all behavior.

Experiment-level statistics across runs

Analysed 100 runs; 100 reached the fitness threshold (100.0% success, threshold = 0.950)

Generations to threshold (successful runs)

  • Mean: 13.00
  • Median: 9.00
  • Std: 12.26

Convergence speed

  • Mean convergence rate (fitness gain/gen): 0.0347
  • Mean fitness improvement/gen: 0.0357

Architecture (successful solutions)

  • Hidden fields (mean ± std): 0.27 ± 0.56
  • Enabled connections (mean ± std): 1.45 ± 1.07

Notable runs

label run value
Highest max fitness 2026-01-24 19h08m33s 0.981938
Lowest max fitness 2026-01-24 18h55m02s 0.950192
Fastest to threshold 2026-01-24 23h07m48s 1.0
Slowest to threshold 2026-01-24 20h15m51s 73.0
Most hidden fields 2026-01-24 19h08m33s 2.0
Most enabled connections 2026-01-24 19h54m56s 7.0

Best-run evolutionary dynamics

This section reports detailed statistics from the evolutionary run that produced the highest-fitness solution for the Selection Instability task.

Fitness statistics

Final generation: g = 11

  • Best fitness: 0.9803
  • Target fitness: 0.950
  • Average fitness: 0.7336

Overall run statistics

  • Max best fitness: 0.9803 (reached at generation 11)
  • Mean best fitness over run (AUC): 0.7756
  • Mean average fitness over run (AUC): 0.7020
  • Longest stagnation period: 2 generations
  • Target fitness first reached: generation 11 (best fitness ≈ 0.9803)

Species statistics

Final generation (g = 11)

  • Species: 118
  • Active species: 93

Across run

  • Total distinct species created: 63
  • Species extinct by final generation: 2
  • Average species per generation: 26.25
  • Average active species per generation: 23.67
  • Max active species in a generation: 93 (at g = 11)

Species lifetime & size

  • Average species lifespan: 3.13 generations
  • Longest-lived species: 0 (lifespan 11)
  • Average max members per species: 42.65
  • Average offspring per species: 158.73

Topology statistics

Final generation (g = 11)

  • Avg genome size: 7.22
  • Avg field genes: 3.36
  • Avg connection genes: 3.86

Growth over run

  • Genome size change: +5.22 (≈ +0.475 / gen)
  • Field genes change: +1.36 (≈ +0.124 / gen)
  • Connection genes change: +3.86 (≈ +0.351 / gen)

Ratios

  • Avg connections per field (final gen): 1.15

Population-level kernel usage

  • Field kernels:

    • Gaussian: 22 497 (89.4%)
    • Mexican-hat: 2 674 (10.6%)

  • Interaction kernels:

    • Gaussian: 11 241 (97.9%)
    • Mexican-hat: 240 (2.1%)

Genome of the highest-performing solution

Generation: 11 Fitness: 0.9803

Field genes

Field Role Kernel type Field parameters Kernel parameters
nf 1 Input Mexican-hat h = -4.20, τ = 55.66 A_exc = 17.34, σ_exc = 5.00, A_inh = 28.55, σ_inh = 5.88, A_glob = -0.08
nf 2 Output Gaussian h = -3.42, τ = 31.69 A = 3.00, σ = 1.55, A_glob = -0.13

Interaction genes

Interaction gene From → To Kernel parameters
gk cg 1-2-1 nf 1 → nf 2 A = 3.00, σ = 2.09, A_glob = 0.00
selection.mp4

Interpretation

  • Winner-take-all behavior emerges from strong inhibitory Mexican-hat dynamics in the input field

Despite the population exploring highly complex genomes, the best solution collapses back to a compact, interpretable architecture, reinforcing neat-dnfs’ bias toward minimal sufficiency.

Takeaway

Selection Instability is substantially more demanding than detection or memory:

  • Convergence is slower and more variable.
  • Some (very little) runs require additional structure beyond the minimal two-field topology.
  • Nevertheless, winner-take-all decision-making reliably emerges from continuous DNF dynamics.

This task marks the transition from representation to choice.

Clone this wiki locally