Skip to content

Results Overview

Jump to bottom
anonymous-author-submissions edited this page Feb 3, 2026 · 2 revisions

This page summarises how task difficulty, convergence behaviour, and evolved architecture complexity scale across the neat-dnfs benchmark suite.

The tasks progress from reactive instabilities to memory-guided and attention-driven cognition, allowing direct comparison of evolutionary demands.

Task hierarchy

The benchmark tasks form a natural hierarchy of increasing cognitive requirements:

  1. Detection Instability – stimulus-driven activation
  2. Memory Instability – persistent working memory
  3. Selection Instability – competitive decision-making
  4. Delayed Match-to-Sample (DMTS) – memory-guided comparison
  5. Inhibition of Return (IOR) – memory-mediated inhibitory control

This ordering is reflected consistently in convergence speed, architectural complexity, and variability across runs.

Success rates

All evaluated tasks achieve 100% success across runs, demonstrating that neat-dnfs is robust across a wide range of dynamical demands.

Task Runs Success rate
Detection 100 100.0%
Memory 100 100.0%
Selection 100 100.0%
DMTS 100 100.0%
IOR 100 100.0%

Convergence behaviour

Generations to fitness threshold

Task Mean Median Std
Detection 1.00 1.00 0.00
Memory 2.43 2.00 1.23
Selection 13.00 9.00 12.26
DMTS 48.44 41.00 32.01
IOR 76.03 57.50 54.53

Key observation: Each additional cognitive requirement (persistence → competition → memory-conditioned selection → inhibition) introduces an approximately order-of-magnitude increase in evolutionary effort.

Convergence speed

Task Mean fitness gain / gen
Detection 0.2267
Memory 0.2241
Selection 0.0357
DMTS 0.0201
IOR 0.0152

Early tasks benefit from strong gradients toward trivial solutions, while later tasks require structural innovation, slowing convergence.

Architectural complexity

Hidden fields

Task Mean ± Std
Detection 0.00 ± 0.00
Memory 0.00 ± 0.00
Selection 0.27 ± 0.56
DMTS 1.30 ± 0.50
IOR 1.19 ± 0.39

Enabled connections

Task Mean ± Std
Detection 1.00 ± 0.00
Memory 1.00 ± 0.00
Selection 1.45 ± 1.07
DMTS 3.46 ± 1.23
IOR 3.43 ± 0.96

Key observation: Complexity increases only when functionally necessary:

  • Detection and Memory are solved with strictly minimal topologies.
  • Selection sometimes requires additional structure.
  • DMTS and IOR reliably induce multi-field, multi-connection architectures.

Emergent principles

Across all tasks, several consistent principles emerge:

  • Minimal sufficiency: Evolution prefers the smallest architecture that satisfies task demands.
  • Functional modularisation: Memory and inhibition reliably produce dedicated fields.
  • Interpretability: Evolved solutions closely match established Dynamic Field Theory models.
  • Graceful scaling: Increased difficulty leads to gradual, not explosive, increases in complexity.

Clone this wiki locally