sim01 — Pheromone Trails

Open the interactive visualization — renders the grid heatmap (play/pause, step slider) and all metric charts from results.json.

Mini simulations that build foundational algorithms for the artificial life simulator. Each solves one sub-problem at a time. These are building blocks — not tests of the full hypotheses, but code we'll need later.

sim01_pheromone_trails.py — Stigmergic Coordination

Status: Complete, runs successfully

What it tests: Basic stigmergy — ants wander, find food, return to nest leaving pheromone trails. Tests whether trails form and how decay rate affects trail stability.

What it teaches us:

• Environmental trace deposition and decay
• Agent-trace interaction (sensing, following, reinforcing)
• The transient/persistent trade-off in stigmergic traces
• Decay rate sweep reveals an optimal zone — too fast (0.2) and trails can't form, too slow (0.001) and everything is covered in pheromone

Control condition (added 2026-07-27). Until now sim01 had no control, so it could not show that any structure in the field was caused by stigmergic feedback rather than by the ants' movement statistics. run now contrasts sensing ants against a pheromone-blind control — ants that still deposit, with the field still decaying, but that cannot read it:

Two things follow, and both revise earlier readings of this simulation:

1. trail_cells runs opposite to trail formation. The blind control scores nearly three times higher on it. A laden ant deposits 100 units every step and decay is 2%/step, so a visited cell stays above threshold ~230 steps — the count measures coverage, and blind ants that wander widely cover more ground. Trail structure needs trail_concentration (share of pheromone in the densest 5% of cells; a uniform field scores 0.05). By that measure sensing genuinely does form trails: 0.786 vs 0.270.
2. Trail formation does not improve foraging here. The blind control collected more food (68 units vs 57). Pheromone following concentrates the ants onto shared paths but does not, in this model, feed them better. Worth stating plainly since the ant-colony framing invites the opposite assumption.

Key results from decay rate sweep (sensing condition):

Observations:

• At very low decay (0.001) pheromone saturates the grid — high coverage, low concentration (0.404). This is the "memory without adaptation" problem, and it is the one reading the old coverage metric got right.
• The claim that high decay (0.2) means "no stable trails" was wrong. Concentration is highest there (0.847); what falls is coverage. High decay produces few, sharply defined trails rather than none. The earlier interpretation followed from reading trail_cells as trail quality.
• Foraging varies little across the sweep (419–443 remaining of 500), so the decay rate has far less functional consequence in this model than the coverage numbers suggest.
• Concentration is not monotonic in decay (0.404 → 0.786 → 0.523 → 0.847), so there is no clean "optimal window" on this measure — the earlier ~0.01–0.05 window was an artifact of reading coverage.

Building blocks provided:

• 2D grid with pheromone field (deposition, decay, sensing)
• Agent movement and sensing
• Decay rate as a sweepable parameter
• Trail measurement (cell count above threshold)

Next steps:

• Add multiple food types to test trail competition
• Add pheromone evaporation to test trace differentiation
• Extend to test the trace→actor crossing hypothesis (H7): do accumulated traces ever become self-maintaining?