Delayed Match‐to‐Sample (DMTS)
The Delayed Match-to-Sample (DMTS) task evaluates whether an evolved Dynamic Neural Field (DNF) architecture can perform memory-guided decision-making.
The desired behaviour is:
- A sample stimulus is presented at a particular feature/location.
- The system must encode and maintain this sample internally during a delay period.
- Two or more test stimuli are then presented.
- The system should select the stimulus matching the memorised sample, while suppressing non-matching alternatives.
This task is a canonical paradigm in cognitive neuroscience for studying working memory and comparison-based decision processes.
DMTS consistently requires a three-field architecture, with explicit separation between perception, memory, and decision:
-
Fields: 3 (input, memory, output)
-
Hidden fields: typically 1–3
-
Connections: multiple feedforward couplings
-
Dynamics:
- Input field: stimulus-driven encoding of samples and test cues
- Memory field: sustained representation of the sample across the delay
- Output field: competitive selection biased toward the matching stimulus
The figure above illustrates:
- Encoding of the sample in a dedicated memory field
- Maintenance of the sample representation during the delay
- Matching-based selection in the output field
The evolved solution closely mirrors classical Dynamic Field Theory accounts of DMTS.
Analysed 100 runs; 100 reached the fitness threshold (100.0% success, threshold = 0.950)
- Mean: 48.44
- Median: 41.00
- Std: 32.01
- Mean convergence rate (fitness gain/gen): 0.0177
- Mean fitness improvement/gen: 0.0201
- Hidden fields (mean ± std): 1.30 ± 0.50
- Enabled connections (mean ± std): 3.46 ± 1.23
| label | run | value |
|---|---|---|
| Highest max fitness | 2026-01-22 12h47m03s | 0.987305 |
| Lowest max fitness | 2026-01-28 05h08m57s | 0.950243 |
| Fastest to threshold | 2026-01-23 00h27m03s | 10.0 |
| Slowest to threshold | 2026-01-23 21h18m55s | 159.0 |
| Most hidden fields | 2026-01-23 21h18m55s | 3.0 |
| Most enabled connections | 2026-01-23 21h18m55s | 8.0 |
This section reports detailed statistics from the evolutionary run that produced the highest-fitness solution for the Delayed Match-to-Sample task.
Final generation: g = 30
- Best fitness: 0.9873
- Target fitness: 0.950
- Average fitness: 0.4851
Overall run statistics
- Max best fitness: 0.9873 (reached at generation 30)
- Mean best fitness over run (AUC): 0.6965
- Mean average fitness over run (AUC): 0.4010
- Longest stagnation period: 3 generations
- Target fitness first reached: generation 30 (best fitness ≈ 0.9873)
Final generation (g = 30)
- Species: 92
- Active species: 56
Across run
- Total distinct species created: 91
- Species extinct by final generation: 29
- Average species per generation: 30.32
- Average active species per generation: 17.94
- Max active species in a generation: 62 (at g = 28)
Species lifetime & size
- Average species lifespan: 9.32 generations
- Longest-lived species: 0 (lifespan 30)
- Average max members per species: 73.02
- Average offspring per species: 318.68
Final generation (g = 30)
- Avg genome size: 6.00
- Avg field genes: 3.00
- Avg connection genes: 3.00
Growth over run
- Genome size change: +4.00 (≈ +0.133 / gen)
- Field genes change: +1.00 (≈ +0.033 / gen)
- Connection genes change: +3.00 (≈ +0.100 / gen)
Ratios
- Avg connections per field (final gen): 1.00
-
Field kernels:
- Gaussian: 56 544 (85.2%)
- Mexican-hat: 9 835 (14.8%)
-
Interaction kernels:
- Gaussian: 10 627 (36.2%)
- Mexican-hat: 18 746 (63.8%)
Generation: 30 Fitness: 0.9873
| Field | Role | Kernel type | Field parameters | Kernel parameters |
|---|---|---|---|---|
| nf 1 | Input | Gaussian | h = -11.85, τ = 19.20 | A = 8.33, σ = 4.20, A_glob = -0.02 |
| nf 2 | Output | Gaussian | h = -6.88, τ = 18.27 | A = 4.18, σ = 4.93, A_glob = -0.13 |
| nf 3 | Hidden | Gaussian | h = -11.85, τ = 19.20 | A = 23.95, σ = 1.00, A_glob = -0.20 |
| Interaction gene | From → To | Kernel parameters |
|---|---|---|
| gk cg 1-2-0 | nf 1 → nf 2 | A = 11.53, σ = 5.42, A_glob = -0.05 |
| gk cg 1-3-18 | nf 1 → nf 3 | A = 15.73, σ = 5.57, A_glob = 0.00 |
| gk cg 3-2-19 | nf 3 → nf 2 | A = 3.28, σ = 2.20, A_glob = 0.00 |
dmts-cut-compressed.mp4
- Nearly every generation introduces new species
- Many species go extinct, indicating strong selection pressure
- A dedicated hidden memory field reliably emerges
- Interaction kernels shift toward Mexican-hat dominance
The best-performing solution is interpretable, implementing perception → memory → decision as distinct dynamical components.
DMTS represents a qualitative leap in cognitive complexity:
- Evolution must coordinate memory maintenance and stimulus comparison.
- Convergence is slower and architectures are more elaborate.
- Nevertheless, neat-dnfs reliably evolves interpretable, modular solutions aligned with neurocognitive theory.
This task demonstrates that the framework supports compositional cognition, not just isolated instabilities.